392 files changed, 55781 insertions, 74625 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/barrier.hpp b/src/core/NEON/kernels/arm_gemm/barrier.hpp
index 3b34e8089f..b7abd87c29 100644
--- a/src/core/NEON/kernels/arm_gemm/barrier.hpp
+++ b/src/core/NEON/kernels/arm_gemm/barrier.hpp
@@ -39,14 +39,6 @@ private:
 public:
     barrier(unsigned int threads) : m_threads(threads), m_waiters(0), m_leavers(0) { }
 
-    // Add a move constructor because these objects might be moved around at setup time.
-    // Moving while the barrier is active won't work.
-    barrier(barrier &&other) : m_threads(other.m_threads), m_waiters(0), m_leavers(0) {
-        // This doesn't make it safe, but will have a chance of firing if something odd is occurring.
-        assert(other.m_waiters==0);
-        assert(other.m_leavers==0);
-    }
-
     /* This isn't safe if any thread is waiting... */
     void set_nthreads(unsigned int nthreads) {
         m_threads = nthreads;
diff --git a/src/core/NEON/kernels/arm_gemm/convolver.hpp b/src/core/NEON/kernels/arm_gemm/convolver.hpp
index a9f3dd0ad8..b15f669132 100644
--- a/src/core/NEON/kernels/arm_gemm/convolver.hpp
+++ b/src/core/NEON/kernels/arm_gemm/convolver.hpp
@@ -231,8 +231,8 @@ public:
         for (unsigned int ky=0; ky<params.kernel_height; ky++) {
             for (unsigned int kx=0; kx<params.kernel_width; kx++) {
                 unsigned int n = (ky * params.kernel_width) + kx;
-                m_kernel_y[n] = (ky * params.dilation_h) - params.padding_top;
-                m_kernel_x[n] = (kx * params.dilation_w) - params.padding_left;
+                m_kernel_y[n] = ky - params.padding_top;
+                m_kernel_x[n] = kx - params.padding_left;
             }
         }
     }
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
index 7d746789e4..c8bd8fd658 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
@@ -46,7 +46,6 @@
 
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
 #include "kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp"
-#include "kernels/sve_ffinterleaved_bf16fp32_dot_8x3VL.hpp"
 #include "kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp"
 #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
 
@@ -68,10 +67,10 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<bfloat16, bfloat16, float> gemm_bf16_methods[] =
+static const GemmImplementation<bfloat16, float> gemm_bf16_methods[] =
 {
-#ifdef ARM_COMPUTE_ENABLE_BF16
 #ifdef __aarch64__
+#ifdef ARM_COMPUTE_ENABLE_BF16
 #ifdef ARM_COMPUTE_ENABLE_SVE
 #ifdef ARM_COMPUTE_ENABLE_SME2
 // SME kernels
@@ -107,36 +106,36 @@ static const GemmImplementation<bfloat16, bfloat16, float> gemm_bf16_methods[] =
 },
 #endif // ARM_COMPUTE_ENABLE_SME2
 // gemm_bf16_interleaved
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_bf16fp32_mmla_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_svebf16() && (args._Ksize>4); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_bf16fp32_mmla_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_svebf16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_bf16fp32_dot_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_svebf16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_bf16fp32_dot_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_svebf16() && (args._Ksize>2); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, float>(args); }
 ),
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_ffinterleaved_bf16fp32_mmla_8x3VL",
     KernelWeightFormat::VL2VL_BL64,
@@ -144,15 +143,7 @@ GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
-    GemmMethod::GEMM_INTERLEAVED,
-    "sve_ffinterleaved_bf16fp32_dot_8x3VL",
-    KernelWeightFormat::VL1VL_BL32,
-    [](const GemmArgs &args) { return args._ci->has_svebf16(); },
-    [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_dot_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_dot_8x3VL, bfloat16, float>(args); }
-),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_ffhybrid_bf16fp32_mmla_6x4VL",
     KernelWeightFormat::VL2VL_BL64,
@@ -162,36 +153,36 @@ GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
 ),
 #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
 #endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_bf16fp32_mmla_6x16",
     [](const GemmArgs &args) { return args._ci->has_bf16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_bf16fp32_mmla_8x12",
     [](const GemmArgs &args) { return args._ci->has_bf16() && (args._Ksize>4); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_bf16fp32_dot_6x16",
     [](const GemmArgs &args) { return args._ci->has_bf16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_bf16fp32_dot_8x12",
     [](const GemmArgs &args) { return args._ci->has_bf16() && (args._Ksize>2); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, float>(args); }
 ),
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_ffinterleaved_bf16fp32_mmla_8x12",
     KernelWeightFormat::VL256_BL64,
@@ -199,7 +190,7 @@ GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_ffhybrid_bf16fp32_mmla_6x16",
     KernelWeightFormat::VL256_BL64,
@@ -207,7 +198,7 @@ GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
     [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_bf16fp32_mmla_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
     [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_bf16fp32_mmla_6x16, bfloat16, float>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_ffinterleaved_bf16fp32_dot_8x12",
     KernelWeightFormat::VL128_BL32,
@@ -216,25 +207,15 @@ GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_dot_8x12, bfloat16, float>(args); }
 ),
 #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
-GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
+GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_sgemm_8x12",
     nullptr,
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, float>(args); }
 ),
-#elif defined(__arm__)
-{
-    GemmMethod::GEMM_INTERLEAVED,
-    "sgemm_8x6",
-    nullptr,
-    nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, bfloat16, bfloat16, float>(args); }
-},
-#else
-# error "Unknown Architecture"
-#endif
 #endif // ARM_COMPUTE_ENABLE_BF16
+#endif // __aarch64__
 {
     GemmMethod::DEFAULT,
     "",
@@ -245,14 +226,14 @@ GemmImplementation<bfloat16, bfloat16, float>::with_estimate(
 };
 
 template<>
-const GemmImplementation<bfloat16, bfloat16, float> *gemm_implementation_list<bfloat16, bfloat16, float>() {
+const GemmImplementation<bfloat16, float> *gemm_implementation_list<bfloat16, float>() {
     return gemm_bf16_methods;
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<bfloat16, bfloat16, float> gemm<bfloat16, bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
-template bool has_opt_gemm<bfloat16, bfloat16, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
-template KernelDescription get_gemm_method<bfloat16, bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
-template std::vector<KernelDescription> get_compatible_kernels<bfloat16, bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
+template UniqueGemmCommon<bfloat16, float> gemm<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<bfloat16, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
 
 } // namespace arm_gemm
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp
index 1e4de4a39e..aa761b46e4 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_bf16bf16.cpp
@@ -32,12 +32,12 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<bfloat16, bfloat16, bfloat16> gemm_bf16bf16_methods[] =
+static const GemmImplementation<bfloat16, bfloat16> gemm_bf16bf16_methods[] =
 {
 #ifdef __aarch64__
 #ifdef ARM_COMPUTE_ENABLE_BF16
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
-GemmImplementation<bfloat16, bfloat16, bfloat16>::with_estimate(
+GemmImplementation<bfloat16, bfloat16>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_ffinterleaved_bf16fp32_mmla_8x12",
     KernelWeightFormat::VL256_BL64,
@@ -45,7 +45,7 @@ GemmImplementation<bfloat16, bfloat16, bfloat16>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, bfloat16>::estimate_cycles<bfloat16>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, bfloat16>(args); }
 ),
-GemmImplementation<bfloat16, bfloat16, bfloat16>::with_estimate(
+GemmImplementation<bfloat16, bfloat16>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_ffinterleaved_bf16fp32_mmla_8x3VL",
     KernelWeightFormat::VL2VL_BL64,
@@ -66,14 +66,14 @@ GemmImplementation<bfloat16, bfloat16, bfloat16>::with_estimate(
 };
 
 template<>
-const GemmImplementation<bfloat16, bfloat16, bfloat16> *gemm_implementation_list<bfloat16, bfloat16, bfloat16>() {
+const GemmImplementation<bfloat16, bfloat16> *gemm_implementation_list<bfloat16, bfloat16>() {
     return gemm_bf16bf16_methods;
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<bfloat16, bfloat16, bfloat16> gemm<bfloat16, bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
-template bool has_opt_gemm<bfloat16, bfloat16, bfloat16, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
-template KernelDescription get_gemm_method<bfloat16, bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
-template std::vector<KernelDescription> get_compatible_kernels<bfloat16, bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
+template UniqueGemmCommon<bfloat16, bfloat16> gemm<bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<bfloat16, bfloat16, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<bfloat16, bfloat16, Nothing>(const GemmArgs &args, const Nothing &);
 
 } // namespace arm_gemm
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp
index 0d9f53b84d..12bddf15e1 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp
@@ -23,7 +23,7 @@
  */
 
 // This can only be built if the target/compiler supports FP16 arguments.
-#if defined(__aarch64__) && (defined(ENABLE_FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
+#if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
 
 #include "arm_gemm.hpp"
 
@@ -42,10 +42,12 @@
 #include "kernels/a64_hgemm_8x24.hpp"
 #include "kernels/a64_hybrid_fp16_mla_6x32.hpp"
 #include "kernels/a64_sgemm_8x12.hpp"
+#ifdef ARM_COMPUTE_ENABLE_SME2
 #include "kernels/sme2_gemv_fp16fp32fp16_dot_16VL.hpp"
 #include "kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL.hpp"
 #include "kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL.hpp"
 #include "kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_SME2
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
 #include "kernels/sve_ffhybrid_fp16_mla_6x4VL.hpp"
 #include "kernels/sve_ffinterleaved_fp16_mla_8x3VL.hpp"
@@ -55,7 +57,7 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<__fp16, __fp16, __fp16> gemm_fp16_methods[] = {
+static const GemmImplementation<__fp16, __fp16> gemm_fp16_methods[] = {
 #ifdef ARM_COMPUTE_ENABLE_SVE
 #ifdef ARM_COMPUTE_ENABLE_SME2
 {
@@ -71,7 +73,7 @@ static const GemmImplementation<__fp16, __fp16, __fp16> gemm_fp16_methods[] = {
     [](const GemmArgs &args) { return args._ci->has_sme2(); },
     [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
                                return args._Nsize >= 8*VL || args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL, __fp16, __fp16, __fp16, Nothing, false, false, false, true>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL, __fp16, __fp16, Nothing, false, false, false, true>(args); }
 },
 {
     GemmMethod::GEMM_INTERLEAVED,
@@ -79,32 +81,32 @@ static const GemmImplementation<__fp16, __fp16, __fp16> gemm_fp16_methods[] = {
     [](const GemmArgs &args) { return args._ci->has_sme2(); },
     [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
                                return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL, __fp16, __fp16, __fp16, Nothing, false, false, false, true>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL, __fp16, __fp16, Nothing, false, false, false, true>(args); }
 },
 {
     GemmMethod::GEMM_INTERLEAVED,
     "sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL",
     [](const GemmArgs &args) { return args._ci->has_sme2(); },
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL, __fp16, __fp16, __fp16, Nothing, false, false, false, true>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL, __fp16, __fp16, Nothing, false, false, false, true>(args); }
 },
 #endif // ARM_COMPUTE_ENABLE_SME2
-GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
+GemmImplementation<__fp16, __fp16>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_fp16_mla_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_sve(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp16_mla_6x4VL, __fp16, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp16_mla_6x4VL, __fp16, __fp16, __fp16>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp16_mla_6x4VL, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp16_mla_6x4VL, __fp16, __fp16>(args); }
 ),
-GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
+GemmImplementation<__fp16, __fp16>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_fp16_mla_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_sve() && (args._Ksize > 4); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_fp16_mla_8x3VL, __fp16, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp16_mla_8x3VL, __fp16, __fp16, __fp16>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_fp16_mla_8x3VL, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp16_mla_8x3VL, __fp16, __fp16>(args); }
 ),
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
-GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
+GemmImplementation<__fp16, __fp16>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_ffinterleaved_fp16_mla_8x3VL",
     KernelWeightFormat::VL1VL_BL16,
@@ -112,7 +114,7 @@ GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_fp16_mla_8x3VL, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_fp16_mla_8x3VL, __fp16, __fp16>(args); }
 ),
-GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
+GemmImplementation<__fp16, __fp16>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_ffhybrid_fp16_mla_6x4VL",
     KernelWeightFormat::VL1VL_BL16,
@@ -123,22 +125,22 @@ GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
 #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
 #endif // ARM_COMPUTE_ENABLE_SVE
 #if defined(__aarch64__)
-GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
+GemmImplementation<__fp16, __fp16>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_fp16_mla_6x32",
     [](const GemmArgs &args) { return args._ci->has_fp16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16, __fp16>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16>(args); }
 ),
-GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
+GemmImplementation<__fp16, __fp16>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_hgemm_8x24",
     [](const GemmArgs &args) { return args._ci->has_fp16(); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_hgemm_8x24, __fp16, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_hgemm_8x24, __fp16, __fp16, __fp16>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_hgemm_8x24, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_hgemm_8x24, __fp16, __fp16>(args); }
 ),
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
-GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
+GemmImplementation<__fp16, __fp16>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_ffinterleaved_fp16_mla_8x24",
     KernelWeightFormat::VL128_BL16,
@@ -146,7 +148,7 @@ GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_fp16_mla_8x24, __fp16, __fp16>::estimate_cycles<__fp16>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_fp16_mla_8x24, __fp16, __fp16>(args); }
 ),
-GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
+GemmImplementation<__fp16, __fp16>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_ffhybrid_fp16_mla_6x32",
     KernelWeightFormat::VL128_BL16,
@@ -160,7 +162,7 @@ GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
     "a64_sgemm_8x12",
     nullptr,
     [](const GemmArgs &args) { return !args._ci->has_fp16(); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, __fp16, __fp16, __fp16>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, __fp16, __fp16>(args); }
 },
 #elif defined(__arm__)
 {
@@ -168,7 +170,7 @@ GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
     "sgemm_8x6",
     nullptr,
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, __fp16, __fp16, __fp16>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, __fp16, __fp16>(args); }
 },
 #else // not AArch64 or AArch32
 # error Unknown Architecture
@@ -183,16 +185,16 @@ GemmImplementation<__fp16, __fp16, __fp16>::with_estimate(
 };
 
 template<>
-const GemmImplementation<__fp16, __fp16, __fp16> *gemm_implementation_list<__fp16, __fp16, __fp16>() {
+const GemmImplementation<__fp16, __fp16> *gemm_implementation_list<__fp16, __fp16>() {
     return gemm_fp16_methods;
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<__fp16, __fp16, __fp16> gemm<__fp16, __fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
-template bool has_opt_gemm<__fp16, __fp16, __fp16, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
-template KernelDescription get_gemm_method<__fp16, __fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
-template std::vector<KernelDescription> get_compatible_kernels<__fp16, __fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
+template UniqueGemmCommon<__fp16, __fp16> gemm<__fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<__fp16, __fp16, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<__fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<__fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
 
 } // namespace arm_gemm
 
-#endif // defined(__aarch64__) && (defined(ENABLE_FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
+#endif // defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
index 5da7161671..0c1d3a387b 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
@@ -82,7 +82,7 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<float, float, float> gemm_fp32_methods[] =
+static const GemmImplementation<float, float> gemm_fp32_methods[] =
 {
 // GEMV cases - starting with 'gemv_batched' wrapper to turn batched GEMV into GEMM.
 {
@@ -95,27 +95,27 @@ static const GemmImplementation<float, float, float> gemm_fp32_methods[] =
 #ifdef __aarch64__
 #ifdef ARM_COMPUTE_ENABLE_BF16
 // "fast mode" (BF16) kernels
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_bf16fp32_mmla_8x12",
     [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float>(args); }
 ),
 
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_fp32bf16fp32_mmla_6x16",
     [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_fp32bf16fp32_mmla_4x24",
     [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float>(args); }
 ),
 #endif // ARM_COMPUTE_ENABLE_BF16
 #ifdef ARM_COMPUTE_ENABLE_SVE
@@ -189,26 +189,26 @@ GemmImplementation<float, float, float>::with_estimate(
 },
 #endif // ARM_COMPUTE_ENABLE_SME2
 #ifdef ARM_COMPUTE_ENABLE_BF16
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_bf16fp32_mmla_8x3VL",
     [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_fp32bf16fp32_mmla_6x4VL",
     [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_fp32bf16fp32_mmla_4x6VL",
     [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float>(args); }
 ),
 #endif // ARM_COMPUTE_ENABLE_BF16
 #ifdef ARM_COMPUTE_ENABLE_SVEF32MM
@@ -218,8 +218,8 @@ GemmImplementation<float, float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_fp32_mmla_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_svef32mm() && (args._Ksize>4); },
-    [](const GemmArgs &args) { return !(args._fast_mode && args._ci->has_svebf16()); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mmla_8x3VL, float, float, float>(args); }
+    [](const GemmArgs &args) { return !(args._fast_mode && args._ci->has_bf16()); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mmla_8x3VL, float, float>(args); }
 },
 #endif // ARM_COMPUTE_ENABLE_SVEF32MM
 // SVE kernels
@@ -228,25 +228,25 @@ GemmImplementation<float, float, float>::with_estimate(
     "sve_hybrid_fp32_mla_8x1VL",
     [](const GemmArgs &args) { return args._ci->has_sve(); },
     [](const GemmArgs &args) { return (args._Nsize < 12); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_8x1VL, float, float, float>(args); }
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_8x1VL, float, float>(args); }
 },
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_fp32_mla_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_sve(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_fp32_mla_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_sve(); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float>(args); }
 ),
  #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
 #ifdef ARM_COMPUTE_ENABLE_BF16
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_ffinterleaved_bf16fp32_mmla_8x3VL",
     KernelWeightFormat::VL2VL_BL64_BF16,
@@ -254,7 +254,7 @@ GemmImplementation<float, float, float>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_ffhybrid_fp32bf16fp32_mmla_4x6VL",
     KernelWeightFormat::VL2VL_BL64_BF16,
@@ -263,7 +263,7 @@ GemmImplementation<float, float, float>::with_estimate(
     [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_sve_ffhybrid_fp32bf16fp32_mmla_4x6VL, float, float>(args); }
 ),
 #endif
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_ffinterleaved_fp32_mla_8x3VL",
     KernelWeightFormat::VL1VL_BL32,
@@ -271,7 +271,7 @@ GemmImplementation<float, float, float>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_fp32_mla_8x3VL, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_fp32_mla_8x3VL, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_ffhybrid_fp32_mla_6x4VL",
     KernelWeightFormat::VL1VL_BL32,
@@ -287,7 +287,7 @@ GemmImplementation<float, float, float>::with_estimate(
     "a64_sgemm_8x6",
     nullptr,
     [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A35; },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x6, float, float, float>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x6, float, float>(args); }
 },
 // Arm® Neon™ hybrid methods
 {
@@ -309,33 +309,33 @@ GemmImplementation<float, float, float>::with_estimate(
     "a64_hybrid_fp32_mla_8x4",
     nullptr,
     [](const GemmArgs &args) { return (args._Nsize < 12); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_8x4, float, float, float>(args); }
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_8x4, float, float>(args); }
 },
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_fp32_mla_4x24",
     nullptr,
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_fp32_mla_6x16",
     nullptr,
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_sgemm_8x12",
     nullptr,
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, float, float, float>::estimate_cycles<float>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, float, float, float>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, float, float>(args); }
 ),
 #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
 #ifdef ARM_COMPUTE_ENABLE_BF16
 // "fast mode" (BF16) kernels
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_ffinterleaved_bf16fp32_mmla_8x12",
     KernelWeightFormat::VL256_BL64_BF16,
@@ -343,7 +343,7 @@ GemmImplementation<float, float, float>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_ffhybrid_fp32bf16fp32_mmla_4x24",
     KernelWeightFormat::VL256_BL64_BF16,
@@ -351,7 +351,7 @@ GemmImplementation<float, float, float>::with_estimate(
     [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_4x24, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_4x24, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_ffhybrid_fp32bf16fp32_mmla_6x16",
     KernelWeightFormat::VL256_BL64_BF16,
@@ -359,9 +359,8 @@ GemmImplementation<float, float, float>::with_estimate(
     [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_6x16, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_6x16, float, float>(args); }
 ),
-
 #endif // BF16
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_ffinterleaved_fp32_mla_8x12",
     KernelWeightFormat::VL128_BL32,
@@ -369,7 +368,7 @@ GemmImplementation<float, float, float>::with_estimate(
     [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_fp32_mla_8x12, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_fp32_mla_8x12, float, float>(args); }
 ),
-GemmImplementation<float, float, float>::with_estimate(
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_ffhybrid_fp32_mla_6x16",
     KernelWeightFormat::VL128_BL32,
@@ -386,7 +385,7 @@ GemmImplementation<float, float, float>::with_estimate(
     "sgemm_8x6",
     nullptr,
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, float, float, float>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, float, float>(args); }
 },
 #endif // __arm__
 {
@@ -400,14 +399,14 @@ GemmImplementation<float, float, float>::with_estimate(
 
 /* Templated function to return this list. */
 template<>
-const GemmImplementation<float, float, float> *gemm_implementation_list<float, float, float>() {
+const GemmImplementation<float, float> *gemm_implementation_list<float, float>() {
     return gemm_fp32_methods;
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<float, float, float> gemm<float, float, float, Nothing>(const GemmArgs &args, const Nothing &);
-template bool has_opt_gemm<float, float, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
-template KernelDescription get_gemm_method<float, float, float, Nothing>(const GemmArgs &args, const Nothing &);
-template std::vector<KernelDescription> get_compatible_kernels<float, float, float, Nothing> (const GemmArgs &args, const Nothing &);
+template UniqueGemmCommon<float, float> gemm<float, float, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<float, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<float, float, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<float, float, Nothing> (const GemmArgs &args, const Nothing &);
 
 } // namespace arm_gemm
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_hybrid.hpp b/src/core/NEON/kernels/arm_gemm/gemm_hybrid.hpp
index 834751f1fe..a6c9677305 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_hybrid.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_hybrid.hpp
@@ -41,7 +41,7 @@ namespace arm_gemm {
 
 // Implementation of the GemmCommon abstract class.
 template<typename strategy, typename To, typename Tr>
-class GemmHybrid : public GemmCommon<To, To, Tr> {
+class GemmHybrid : public GemmCommon<To, Tr> {
     typedef typename strategy::operand_type Toi;
     typedef typename strategy::result_type Tri;
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp b/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp
index 8bbb877c1b..0cc4d4f3d9 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_hybrid_indirect.hpp
@@ -260,8 +260,8 @@ struct kernel_weight_format<strategy, false> {
 } // anonymous namespace
 
 // Implementation of the GemmCommon abstract class.
-template<typename strategy, typename To, typename Tw, typename Tr, typename OutputStage=Nothing, bool SeparateQuantize=false, bool FixedFormat=false>
-class GemmHybridIndirect : public GemmCommon<To, Tw, Tr> {
+template<typename strategy, typename To, typename Tr, typename OutputStage=Nothing, bool SeparateQuantize=false, bool FixedFormat=false>
+class GemmHybridIndirect : public GemmCommon<To, Tr> {
     typedef typename strategy::lhs_operand_type Tloi;
     typedef typename strategy::rhs_operand_type Troi;
     typedef typename strategy::result_type Tri;
@@ -618,7 +618,7 @@ public:
         return _args._nmulti * iceildiv(_args._Nsize, strategy::out_width());
     }
 
-    void requantize_bias(void *in_buffer, const Tw *B, const int ldb, const int B_multi_stride) override {
+    void requantize_bias(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override {
         if (std::is_same<OutputStage, Requantize32>::value) {
             _col_bias = reinterpret_cast<int32_t *>(in_buffer);
 
@@ -636,11 +636,11 @@ public:
         return strat.transforms.PrepareB_supports_transpose();
     }
 
-    void pretranspose_B_array(void *in_buffer, const Tw *B, const int ldb, const int B_multi_stride, bool transposed) override {
+    void pretranspose_B_array(void *in_buffer, const To *B, const int ldb, const int B_multi_stride, bool transposed) override {
         pretranspose_B_array_part(in_buffer, B, ldb, B_multi_stride, transposed, 0, get_B_pretranspose_window_size());
     }
 
-    void pretranspose_B_array_part(void *in_buffer, const Tw *B, const int ldb, const int B_multi_stride, bool transposed, size_t start, size_t end) override {
+    void pretranspose_B_array_part(void *in_buffer, const To *B, const int ldb, const int B_multi_stride, bool transposed, size_t start, size_t end) override {
         if (end >= get_B_pretranspose_window_size()) {
             requantize_bias(in_buffer, B, ldb, B_multi_stride);
         }
@@ -835,7 +835,7 @@ public:
 };
 
 template<typename strategy, typename To, typename Tr, typename OutputStage=Nothing>
-using GemmHybridIndirectFixedFormat = GemmHybridIndirect<strategy, To, To, Tr, OutputStage, false, true>;
+using GemmHybridIndirectFixedFormat = GemmHybridIndirect<strategy, To, Tr, OutputStage, false, true>;
 
 } // namespace arm_gemm
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_hybrid_quantized.hpp b/src/core/NEON/kernels/arm_gemm/gemm_hybrid_quantized.hpp
index 62184bcbd1..f12efe4282 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_hybrid_quantized.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_hybrid_quantized.hpp
@@ -42,7 +42,7 @@ namespace arm_gemm {
 
 // Implementation of the GemmCommon abstract class.
 template<typename strategy, typename To, typename Tr>
-class GemmHybridQuantized : public GemmCommon<To, To, Tr> {
+class GemmHybridQuantized : public GemmCommon<To, Tr> {
     typedef typename strategy::operand_type Toi;
     typedef typename strategy::result_type Tri;
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_implementation.hpp b/src/core/NEON/kernels/arm_gemm/gemm_implementation.hpp
index 19d5e3e23d..db5155f500 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_implementation.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_implementation.hpp
@@ -35,17 +35,17 @@ namespace arm_gemm {
  * of types, a static list of these structures is built up to describe the
  * implementations available.
  */
-template<typename Tlop, typename Trop, typename Tret, class OutputStage = Nothing>
+template<typename Top, typename Tret, class OutputStage = Nothing>
 struct GemmImplementation {
     const GemmMethod                                                               method;
     const char *                                                                   name;
     const KernelWeightFormat                                                       kernel_weight_format = KernelWeightFormat::NON_FIXED;
     std::function<bool(const GemmArgs &, const OutputStage &)>                     is_supported = {};
     std::function<uint64_t(const GemmArgs &, const OutputStage &)>                 cycle_estimate = {};
-    std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &, const OutputStage &)>  instantiate = {};
+    std::function<GemmCommon<Top, Tret> *(const GemmArgs &, const OutputStage &)>  instantiate = {};
 
     bool do_is_supported(const GemmArgs &args, const OutputStage &os) const {
-        // Check supplied is_supported() function first.
+	// Check supplied is_supported() function first.
         if (is_supported != nullptr && !is_supported(args, os)) {
             return false;
         }
@@ -68,7 +68,7 @@ struct GemmImplementation {
             // If we get here it means there is a config and it specifies a format.  Check it matches this kernel.
             // NOTE: this will execute SVE instructions if it's an SVE kernel, so it's important that is_supported()
             // was called above first.
-            return (args._cfg->weight_format == get_weight_format(kernel_weight_format, sizeof(Tlop)));
+            return (args._cfg->weight_format == get_weight_format(kernel_weight_format, sizeof(Top)));
         }
     }
 
@@ -80,13 +80,13 @@ struct GemmImplementation {
         }
     }
 
-    GemmCommon<Tlop, Trop, Tret> *do_instantiate(const GemmArgs &args, const OutputStage &os) const {
+    GemmCommon<Top, Tret> *do_instantiate(const GemmArgs &args, const OutputStage &os) const {
         return instantiate(args, os);
     }
 
     static GemmImplementation with_estimate(GemmMethod m, const char *n,
                        std::function<bool(const GemmArgs &, const OutputStage &)> is_supported, std::function<uint64_t(const GemmArgs &, const OutputStage &)> cycle_estimate,
-                       std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &, const OutputStage &)> instantiate) {
+                       std::function<GemmCommon<Top, Tret> *(const GemmArgs &, const OutputStage &)> instantiate) {
         GemmImplementation impl(m,n);
 
         impl.is_supported=is_supported;
@@ -103,14 +103,14 @@ struct GemmImplementation {
 
     GemmImplementation(GemmMethod m, const char *n,
                        std::function<bool(const GemmArgs &, const OutputStage &)> is_supported, std::function<bool(const GemmArgs &, const OutputStage &)> is_recommended,
-                       std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &, const OutputStage &)> instantiate) :
+                       std::function<GemmCommon<Top, Tret> *(const GemmArgs &, const OutputStage &)> instantiate) :
                        method(m), name(n), is_supported(is_supported),
                        cycle_estimate( [is_recommended](const GemmArgs &args, const OutputStage &os) { return (is_recommended == nullptr) ? 0 : (is_recommended(args, os) ? 0 : UINT64_MAX); } ),
                        instantiate(instantiate) {   }
 
     GemmImplementation(GemmMethod m, const char *n, KernelWeightFormat kwf,
                        std::function<bool(const GemmArgs &, const OutputStage &)> is_supported, std::function<bool(const GemmArgs &, const OutputStage &)> is_recommended,
-                       std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &, const OutputStage &)> instantiate) :
+                       std::function<GemmCommon<Top, Tret> *(const GemmArgs &, const OutputStage &)> instantiate) :
                        method(m), name(n), kernel_weight_format(kwf), is_supported(is_supported),
                        cycle_estimate( [is_recommended](const GemmArgs &args, const OutputStage &os) { return (is_recommended == nullptr) ? 0 : (is_recommended(args, os) ? 0 : UINT64_MAX); } ),
                        instantiate(instantiate) {   }
@@ -119,17 +119,17 @@ struct GemmImplementation {
 /* Slightly different version of above for straightforward GEMMs with no
  * output stage, so the std::functions there don't have to deal with the
  * unnecessary second argument.  */
-template<typename Tlop, typename Trop, typename Tret>
-struct GemmImplementation<Tlop, Trop, Tret, Nothing> {
+template<typename Top, typename Tret>
+struct GemmImplementation<Top, Tret, Nothing> {
     const GemmMethod                                          method;
     const char *                                              name;
     const KernelWeightFormat                                  kernel_weight_format = KernelWeightFormat::NON_FIXED;
     std::function<bool(const GemmArgs &)>                     is_supported = {};
     std::function<uint64_t(const GemmArgs &)>                 cycle_estimate = {};
-    std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &)>  instantiate = {};
+    std::function<GemmCommon<Top, Tret> *(const GemmArgs &)>  instantiate = {};
 
     bool do_is_supported(const GemmArgs &args, const Nothing &) const {
-        // Check supplied is_supported() function first.
+	// Check supplied is_supported() function first.
         if (is_supported != nullptr && !is_supported(args)) {
             return false;
         }
@@ -152,7 +152,7 @@ struct GemmImplementation<Tlop, Trop, Tret, Nothing> {
             // If we get here it means there is a config and it specifies a format.  Check it matches this kernel.
             // NOTE: this will execute SVE instructions if it's an SVE kernel, so it's important that is_supported()
             // was called above first.
-            return (args._cfg->weight_format == get_weight_format(kernel_weight_format, sizeof(Tlop)));
+            return (args._cfg->weight_format == get_weight_format(kernel_weight_format, sizeof(Top)));
         }
     }
 
@@ -164,13 +164,13 @@ struct GemmImplementation<Tlop, Trop, Tret, Nothing> {
         }
     }
 
-    GemmCommon<Tlop, Trop, Tret> *do_instantiate(const GemmArgs &args, const Nothing &) const {
+    GemmCommon<Top, Tret> *do_instantiate(const GemmArgs &args, const Nothing &) const {
         return instantiate(args);
     }
 
     static GemmImplementation with_estimate(GemmMethod m, const char *n,
                        std::function<bool(const GemmArgs &)> is_supported, std::function<uint64_t(const GemmArgs &)> cycle_estimate,
-                       std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &)> instantiate) {
+                       std::function<GemmCommon<Top, Tret> *(const GemmArgs &)> instantiate) {
         GemmImplementation impl(m,n);
 
         impl.is_supported=is_supported;
@@ -182,7 +182,7 @@ struct GemmImplementation<Tlop, Trop, Tret, Nothing> {
 
     static GemmImplementation with_estimate(GemmMethod m, const char *n, KernelWeightFormat f,
                        std::function<bool(const GemmArgs &)> is_supported, std::function<uint64_t(const GemmArgs &)> cycle_estimate,
-                       std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &)> instantiate) {
+                       std::function<GemmCommon<Top, Tret> *(const GemmArgs &)> instantiate) {
         GemmImplementation impl(m,n,f);
 
         impl.is_supported=is_supported;
@@ -199,14 +199,14 @@ struct GemmImplementation<Tlop, Trop, Tret, Nothing> {
 
     GemmImplementation(GemmMethod m, const char *n,
                        std::function<bool(const GemmArgs &)> is_supported, std::function<bool(const GemmArgs &)> is_recommended,
-                       std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &)> instantiate) :
+                       std::function<GemmCommon<Top, Tret> *(const GemmArgs &)> instantiate) :
                        method(m), name(n), is_supported(is_supported),
                        cycle_estimate( [is_recommended](const GemmArgs &args) -> uint64_t { return (is_recommended == nullptr) ? 0 : (is_recommended(args) ? 0 : UINT64_MAX); } ),
                        instantiate(instantiate) {   }
 
     GemmImplementation(GemmMethod m, const char *n, KernelWeightFormat kwf,
                        std::function<bool(const GemmArgs &)> is_supported, std::function<bool(const GemmArgs &)> is_recommended,
-                       std::function<GemmCommon<Tlop, Trop, Tret> *(const GemmArgs &)> instantiate) :
+                       std::function<GemmCommon<Top, Tret> *(const GemmArgs &)> instantiate) :
                        method(m), name(n), kernel_weight_format(kwf), is_supported(is_supported),
                        cycle_estimate( [is_recommended](const GemmArgs &args) -> uint64_t { return (is_recommended == nullptr) ? 0 : (is_recommended(args) ? 0 : UINT64_MAX); } ),
                        instantiate(instantiate) {   }
@@ -218,8 +218,8 @@ struct GemmImplementation<Tlop, Trop, Tret, Nothing> {
  * A specialised version is provided for each supported combination of types.
  * The end of the list is indicated by a sentinel descriptor with
  * method==GemmMethod::DEFAULT.  */
-template<typename Tlop, typename Trop, typename Tret, class OutputStage = Nothing>
-const GemmImplementation<Tlop, Trop, Tret, OutputStage> *gemm_implementation_list();
+template<typename Top, typename Tret, class OutputStage = Nothing>
+const GemmImplementation<Top, Tret, OutputStage> *gemm_implementation_list();
 
 /*
  * Select a GEMM implementation for the given arguments.
@@ -234,15 +234,15 @@ const GemmImplementation<Tlop, Trop, Tret, OutputStage> *gemm_implementation_lis
  * this function returns false and doesn't touch the provided pointer
  * reference.
  */
-template<typename Tlop, typename Trop, typename Tret, class OutputStage>
-bool find_implementation(const GemmArgs &args, const OutputStage &os, const GemmImplementation<Tlop, Trop, Tret, OutputStage> * &impl) {
-    auto gemms = gemm_implementation_list<Tlop, Trop, Tret, OutputStage>();
+template<typename Top, typename Tret, class OutputStage>
+bool find_implementation(const GemmArgs &args, const OutputStage &os, const GemmImplementation<Top, Tret, OutputStage> * &impl) {
+    auto gemms = gemm_implementation_list<Top, Tret, OutputStage>();
     const GemmConfig *cfg = args._cfg;
 
-    const GemmImplementation<Tlop, Trop, Tret, OutputStage> *saved_impl = nullptr;
+    const GemmImplementation<Top, Tret, OutputStage> *saved_impl = nullptr;
     uint64_t best_estimate = 0;
 
-    for (const GemmImplementation<Tlop, Trop, Tret, OutputStage> *i = gemms; i->method != GemmMethod::DEFAULT; i++) {
+    for (const GemmImplementation<Top, Tret, OutputStage> *i = gemms; i->method != GemmMethod::DEFAULT; i++) {
         /* Skip if this implementation doesn't support these args. */
         if (!i->do_is_supported(args, os)) {
             continue;
@@ -284,17 +284,17 @@ bool find_implementation(const GemmArgs &args, const OutputStage &os, const Gemm
     return false;
 }
 
-template<typename Tlop, typename Trop, typename Tret, class OutputStage>
+template<typename Top, typename Tret, class OutputStage>
 std::vector<KernelDescription> get_compatible_kernels(const GemmArgs &args, const OutputStage &os) {
     std::vector<KernelDescription> res;
 
     /* Find out what the default implementation in so we can set the flag accordingly later. */
-    const GemmImplementation<Tlop, Trop, Tret, OutputStage> *default_impl;
+    const GemmImplementation<Top, Tret, OutputStage> *default_impl;
     find_implementation(args, os, default_impl);
 
-    auto gemms = gemm_implementation_list<Tlop, Trop, Tret, OutputStage>();
+    auto gemms = gemm_implementation_list<Top, Tret, OutputStage>();
 
-    for (const GemmImplementation<Tlop, Trop, Tret, OutputStage> *i = gemms; i->method != GemmMethod::DEFAULT; i++) {
+    for (const GemmImplementation<Top, Tret, OutputStage> *i = gemms; i->method != GemmMethod::DEFAULT; i++) {
         /* Check that this implementation supports the presented problem. */
 
         if (!i->do_is_supported(args, os)) {
@@ -307,31 +307,31 @@ std::vector<KernelDescription> get_compatible_kernels(const GemmArgs &args, cons
     return res;
 }
 
-template<typename Tlop, typename Trop, typename Tret, class OutputStage>
+template<typename Top, typename Tret, class OutputStage>
 bool has_opt_gemm(WeightFormat &wf, const GemmArgs &args, const OutputStage &os) {
-    const GemmImplementation<Tlop, Trop, Tret, OutputStage> *impl;
-    const bool success =  find_implementation<Tlop, Trop, Tret, OutputStage>(args, os, impl);
+    const GemmImplementation<Top, Tret, OutputStage> *impl;
+    const bool success =  find_implementation<Top, Tret, OutputStage>(args, os, impl);
     if (success)
-      wf = UniqueGemmCommon<Tlop, Trop, Tret>(impl->do_instantiate(args, os))->get_config().weight_format;
+      wf = UniqueGemmCommon<Top, Tret>(impl->do_instantiate(args, os))->get_config().weight_format;
     return success;
 }
 
-template<typename Tlop, typename Trop, typename Tret, class OutputStage>
-UniqueGemmCommon<Tlop, Trop, Tret> gemm(const GemmArgs &args, const OutputStage &os) {
-    const GemmImplementation<Tlop, Trop, Tret, OutputStage> *impl;
+template<typename Top, typename Tret, class OutputStage>
+UniqueGemmCommon<Top, Tret> gemm(const GemmArgs &args, const OutputStage &os) {
+    const GemmImplementation<Top, Tret, OutputStage> *impl;
 
-    if (find_implementation<Tlop, Trop, Tret, OutputStage>(args, os, impl)) {
-        return UniqueGemmCommon<Tlop, Trop, Tret>(impl->do_instantiate(args, os));
+    if (find_implementation<Top, Tret, OutputStage>(args, os, impl)) {
+        return UniqueGemmCommon<Top, Tret>(impl->do_instantiate(args, os));
     }
 
-    return UniqueGemmCommon<Tlop, Trop, Tret>(nullptr);
+    return UniqueGemmCommon<Top, Tret>(nullptr);
 }
 
-template<typename Tlop, typename Trop, typename Tret, class OutputStage>
+template<typename Top, typename Tret, class OutputStage>
 KernelDescription get_gemm_method(const GemmArgs &args, const OutputStage &os) {
-    const GemmImplementation<Tlop, Trop, Tret, OutputStage> *impl;
+    const GemmImplementation<Top, Tret, OutputStage> *impl;
 
-    if (find_implementation<Tlop, Trop, Tret>(args, os, impl)) {
+    if (find_implementation<Top, Tret>(args, os, impl)) {
         return KernelDescription(impl->method, impl->name);
     }
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int16.cpp
index c44e7be4a3..befc1a58a3 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_int16.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_int16.cpp
@@ -32,13 +32,13 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<int16_t, int16_t, int32_t> gemm_s16_methods[] = {
+static const GemmImplementation<int16_t, int32_t> gemm_s16_methods[] = {
 {
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_s16_8x12",
     nullptr,
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s16_8x12, int16_t, int16_t, int32_t>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s16_8x12, int16_t, int32_t>(args); }
 },
 {
     GemmMethod::DEFAULT,
@@ -50,15 +50,15 @@ static const GemmImplementation<int16_t, int16_t, int32_t> gemm_s16_methods[] =
 };
 
 template<>
-const GemmImplementation<int16_t, int16_t, int32_t> *gemm_implementation_list<int16_t, int16_t, int32_t>() {
+const GemmImplementation<int16_t, int32_t> *gemm_implementation_list<int16_t, int32_t>() {
     return gemm_s16_methods;
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<int16_t, int16_t, int32_t> gemm<int16_t, int16_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
-template bool has_opt_gemm<int16_t, int16_t, int32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
-template KernelDescription get_gemm_method<int16_t, int16_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
-template std::vector<KernelDescription> get_compatible_kernels<int16_t, int16_t, int32_t, Nothing> (const GemmArgs &args, const Nothing &);
+template UniqueGemmCommon<int16_t, int32_t> gemm<int16_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<int16_t, int32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<int16_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<int16_t, int32_t, Nothing> (const GemmArgs &args, const Nothing &);
 
 } // namespace arm_gemm
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
index 464e2c6059..fedda3a47a 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
@@ -54,7 +54,7 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<int8_t, int8_t, int32_t> gemm_s8_methods[] = {
+static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = {
 #ifdef ARM_COMPUTE_ENABLE_SVE
 #ifdef ARM_COMPUTE_ENABLE_SME2
 // SME kernels
@@ -82,48 +82,48 @@ static const GemmImplementation<int8_t, int8_t, int32_t> gemm_s8_methods[] = {
     [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL, int8_t, int32_t>(args); }
 },
 #endif // ARM_COMPUTE_ENABLE_SME2
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_s8s32_mmla_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_svei8mm(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int32_t>(args); }
 ),
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_s8s32_mmla_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_svei8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int32_t>(args); }
 ),
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_s8s32_dot_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_sve() && args._Ksize>=16; },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int32_t>(args); }
 ),
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_s8s32_dot_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_sve() && (args._Ksize>4); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int32_t>(args); }
 ),
 #endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_s8s32_mmla_8x12",
     [](const GemmArgs &args) { return args._ci->has_i8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int32_t>(args); }
 ),
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_s8s32_mmla_6x16",
     [](const GemmArgs &args) { return args._ci->has_i8mm(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int32_t>(args); }
 ),
 {
     GemmMethod::GEMM_HYBRID,
@@ -144,29 +144,29 @@ GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
     "a64_gemm_s16_8x12",
     nullptr,
     [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && ((args._Msize > 28) || ((args._Msize % 8) > 4)); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s16_8x12, int8_t, int8_t, int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s16_8x12, int8_t, int32_t>(args); },
 },
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
 
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_s8s32_dot_6x16",
     [](const GemmArgs &args) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int32_t>(args); }
 ),
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_s8_8x12",
     [](const GemmArgs &args) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_s8_8x12, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s8_8x12, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_s8_8x12, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s8_8x12, int8_t, int32_t>(args); }
 ),
-GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
+GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_s8_4x4",
     nullptr,
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_s8_4x4, int8_t, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s8_4x4, int8_t, int8_t, int32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_s8_4x4, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s8_4x4, int8_t, int32_t>(args); }
 ),
 
 {
@@ -179,15 +179,15 @@ GemmImplementation<int8_t, int8_t, int32_t>::with_estimate(
 };
 
 template<>
-const GemmImplementation<int8_t, int8_t, int32_t> *gemm_implementation_list<int8_t, int8_t, int32_t>() {
+const GemmImplementation<int8_t, int32_t> *gemm_implementation_list<int8_t, int32_t>() {
     return gemm_s8_methods;
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<int8_t, int8_t, int32_t> gemm<int8_t, int8_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
-template bool has_opt_gemm<int8_t, int8_t, int32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
-template KernelDescription get_gemm_method<int8_t, int8_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
-template std::vector<KernelDescription> get_compatible_kernels<int8_t, int8_t, int32_t, Nothing> (const GemmArgs &args, const Nothing &);
+template UniqueGemmCommon<int8_t, int32_t> gemm<int8_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<int8_t, int32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<int8_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<int8_t, int32_t, Nothing> (const GemmArgs &args, const Nothing &);
 
 } // namespace arm_gemm
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
index 5214a71cce..897ec9d05f 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
@@ -62,12 +62,12 @@ namespace {
 template<bool MergeStep, bool FixedFormat, typename OutputStage>
 class kernel_and_merge {
 public:
-    template<typename strategy, typename Tlo, typename Tro, typename Tr, typename Tri, typename Tab>
+    template<typename strategy, typename To, typename Tr, typename Tri, typename Tab>
     static void run (
 #ifdef CYCLE_PROFILING
         profiler &prof,
 #endif
-        strategy &strat, const Tlo *a_ptr, const Tro *b_panel, size_t b_stride, Tri *c_panel,
+        strategy &strat, const To *a_ptr, const To *b_panel, size_t b_stride, Tri *c_panel,
         Tr *c_ptr, int ldc, int kern_k, unsigned int m_0,
         unsigned int m_max, unsigned int n_0, unsigned int n_max, const Tr *biasptr,
         const Activation &act, bool accumulate, const OutputStage &os, const int32_t *col_bias,
@@ -76,12 +76,12 @@ public:
 
 // Run a kernel and call the separate merge step
 template<>
-template<typename strategy, typename Tlo, typename Tro, typename Tr, typename Tri, typename Tab>
+template<typename strategy, typename To, typename Tr, typename Tri, typename Tab>
 void kernel_and_merge<true, false, Nothing>::run(
 #ifdef CYCLE_PROFILING
         profiler &prof,
 #endif
-        strategy &strat, const Tlo *a_ptr, const Tro *b_panel, size_t, Tri *c_panel,
+        strategy &strat, const To *a_ptr, const To *b_panel, size_t, Tri *c_panel,
         Tr *c_ptr, int ldc, int kern_k, unsigned int m_0,
         unsigned int m_max, unsigned int n_0, unsigned int n_max, const Tr *biasptr,
         const Activation &act, bool accumulate, const Nothing &, const int32_t *, Tab *)
@@ -106,12 +106,12 @@ void kernel_and_merge<true, false, Nothing>::run(
 
 // Run a fixed-format kernel and call the separate merge step
 template<>
-template<typename strategy, typename Tlo, typename Tro, typename Tr, typename Tri, typename Tab>
+template<typename strategy, typename To, typename Tr, typename Tri, typename Tab>
 void kernel_and_merge<true, true, Nothing>::run(
 #ifdef CYCLE_PROFILING
         profiler &prof,
 #endif
-        strategy &strat, const Tlo *a_ptr, const Tro *b_panel, size_t b_stride, Tri *c_panel,
+        strategy &strat, const To *a_ptr, const To *b_panel, size_t b_stride, Tri *c_panel,
         Tr *c_ptr, int ldc, int kern_k, unsigned int m_0,
         unsigned int m_max, unsigned int n_0, unsigned int n_max, const Tr *biasptr,
         const Activation &act, bool accumulate, const Nothing &, const int32_t *, Tab *)
@@ -136,12 +136,12 @@ void kernel_and_merge<true, true, Nothing>::run(
 
 // Run a kernel with integrated merge
 template<>
-template<typename strategy, typename Tlo, typename Tro, typename Tr, typename Tri, typename Tab>
+template<typename strategy, typename To, typename Tr, typename Tri, typename Tab>
 void kernel_and_merge<false, false, Nothing>::run(
 #ifdef CYCLE_PROFILING
         profiler &prof,
 #endif
-        strategy &strat, const Tlo *a_ptr, const Tro *b_panel, size_t, Tri *,
+        strategy &strat, const To *a_ptr, const To *b_panel, size_t, Tri *,
         Tr *c_ptr, int ldc, int kern_k, unsigned int m_0, unsigned int m_max,
         unsigned int n_0, unsigned int n_max, const Tr *biasptr,
         const Activation &act, bool accumulate, const Nothing &, const int32_t *,
@@ -175,12 +175,12 @@ void kernel_and_merge<false, false, Nothing>::run(
 
 // Run a kernel with integrated merge, quantizing
 template<>
-template<typename strategy, typename Tlo, typename Tro, typename Tr, typename Tri, typename Tab>
+template<typename strategy, typename To, typename Tr, typename Tri, typename Tab>
 void kernel_and_merge<false, false, Requantize32>::run(
 #ifdef CYCLE_PROFILING
         profiler &prof,
 #endif
-        strategy &strat, const Tlo *a_ptr, const Tro *b_panel, size_t, Tri *,
+        strategy &strat, const To *a_ptr, const To *b_panel, size_t, Tri *,
         Tr *c_ptr, int ldc, int kern_k, unsigned int m_0, unsigned int m_max,
         unsigned int n_0, unsigned int n_max, const Tr *,
         const Activation &, bool accumulate, const Requantize32 &qp, const int32_t *col_bias,
@@ -211,12 +211,12 @@ void kernel_and_merge<false, false, Requantize32>::run(
 
 // Run a kernel and call the separate quantize step
 template<>
-template<typename strategy, typename Tlo, typename Tro, typename Tr, typename Tri, typename Tab>
+template<typename strategy, typename To, typename Tr, typename Tri, typename Tab>
 void kernel_and_merge<true, false, Requantize32>::run(
 #ifdef CYCLE_PROFILING
         profiler &prof,
 #endif
-        strategy &strat, const Tlo *a_ptr, const Tro *b_panel, size_t, Tri *c_panel,
+        strategy &strat, const To *a_ptr, const To *b_panel, size_t, Tri *c_panel,
         Tr *c_ptr, int ldc, int kern_k, unsigned int m_0,
         unsigned int m_max, unsigned int n_0, unsigned int n_max, const Tr *,
         const Activation &, bool, const Requantize32 &qp, const int32_t *col_bias,
@@ -257,12 +257,12 @@ void kernel_and_merge<true, false, Requantize32>::run(
 
 // Run a kernel with integrated merge, dequantizing to FP32
 template<>
-template<typename strategy, typename Tlo, typename Tro, typename Tr, typename Tri, typename Tab>
+template<typename strategy, typename To, typename Tr, typename Tri, typename Tab>
 void kernel_and_merge<false, false, DequantizeFloat>::run(
 #ifdef CYCLE_PROFILING
         profiler &prof,
 #endif
-        strategy &strat, const Tlo *a_ptr, const Tro *b_panel, size_t, Tri *,
+        strategy &strat, const To *a_ptr, const To *b_panel, size_t, Tri *,
         Tr *c_ptr, int ldc, int kern_k, unsigned int m_0, unsigned int m_max,
         unsigned int n_0, unsigned int n_max, const Tr *bias,
         const Activation &act, bool accumulate, const DequantizeFloat &dq, const int32_t *col_bias,
@@ -294,12 +294,12 @@ void kernel_and_merge<false, false, DequantizeFloat>::run(
 }
 
 template<>
-template<typename strategy, typename Tlo, typename Tro, typename Tr, typename Tri, typename Tab>
+template<typename strategy, typename To, typename Tr, typename Tri, typename Tab>
 void kernel_and_merge<true, false, DequantizeFloat>::run(
 #ifdef CYCLE_PROFILING
         profiler &prof,
 #endif
-        strategy &strat, const Tlo *a_ptr, const Tro *b_panel, size_t, Tri *c_panel,
+        strategy &strat, const To *a_ptr, const To *b_panel, size_t, Tri *c_panel,
         Tr *c_ptr, int ldc, int kern_k, unsigned int m_0,
         unsigned int m_max, unsigned int n_0, unsigned int n_max, const Tr *bias,
         const Activation &act, bool accumulate, const DequantizeFloat &qp, const int32_t *,
@@ -394,21 +394,21 @@ struct get_stripe_width<strategy, true> {
 };
 
 // KernelWeightFormat is a similar story.
-template<typename strategy, bool FixedFormat, typename Tro>
+template<typename strategy, bool FixedFormat, typename To>
 struct get_kernel_weight_format {
     static KernelWeightFormat get() {
         return KernelWeightFormat::NON_FIXED;
     }
 };
 
-template<typename strategy, typename Tro>
-struct get_kernel_weight_format<strategy, true, Tro> {
+template<typename strategy, typename To>
+struct get_kernel_weight_format<strategy, true, To> {
     static KernelWeightFormat get() {
         KernelWeightFormat kwf = strategy::kernel_weight_format();
 
         // If we are using a BF16 kernel to do an FP32 problem (fast mode) then we need to set the BF16 flag on the
         // weight format.
-        if (std::is_same<Tro, float>::value && std::is_same<typename strategy::rhs_operand_type, bfloat16>::value) {
+        if (std::is_same<To, float>::value && std::is_same<typename strategy::operand_type, bfloat16>::value) {
             uint32_t kwf_i = static_cast<uint32_t>(kwf);
             kwf_i |= 0x10;
             kwf = static_cast<KernelWeightFormat>(kwf_i);
@@ -420,10 +420,9 @@ struct get_kernel_weight_format<strategy, true, Tro> {
 
 } // anonymous namespace
 
-template<typename strategy, typename Tlo, typename Tro, typename Tr, typename OutputStage=Nothing, bool MergeStep=true, bool FixedFormat=false, bool ForceThreadColumns=false, bool ForceFloatAccumulate=false>
-class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
-    typedef typename strategy::lhs_operand_type Tloi;
-    typedef typename strategy::rhs_operand_type Troi;
+template<typename strategy, typename To, typename Tr, typename OutputStage=Nothing, bool MergeStep=true, bool FixedFormat=false, bool ForceThreadColumns=false, bool ForceFloatAccumulate=false>
+class GemmInterleaved : public GemmCommon<To, Tr> {
+    typedef typename strategy::operand_type Toi;
     typedef typename strategy::result_type Tri;
     typedef typename accumulate_buffer_type<strategy, OutputStage, ForceFloatAccumulate>::type Tab;
 
@@ -454,7 +453,7 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
     unsigned int _Mround=0;
 
     /* Working space, pretransposed buffer, buffer manager */
-    const Troi *_B_transposed=nullptr;
+    const Toi *_B_transposed=nullptr;
     void *_working_space=nullptr;
 
     Tab *_accumulation_buffer=nullptr;
@@ -466,10 +465,10 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
     int32_t *col_bias = nullptr;
 
     /* Indirect parameters.  _indirect_buf doubles as a flag to indicate that "indirect" transform should be used. */
-    const Tlo * const * const * _indirect_buf = nullptr;
+    const To * const * const * _indirect_buf = nullptr;
 
     /* Convolver - only set up for convolution problems, so also doubles as a flag. */
-    std::unique_ptr<convolver<Tlo>>  _convolver = nullptr;
+    std::unique_ptr<convolver<To>>  _convolver = nullptr;
 
     unsigned int get_col_sum_size() const {
         if (std::is_same<OutputStage, Requantize32>::value) {
@@ -484,7 +483,7 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
     class blockwalker {
     private:
         /* Size loops, etc. based on our parent's configuration */
-        const GemmInterleaved<strategy, Tlo, Tro, Tr, OutputStage, MergeStep, FixedFormat, ForceThreadColumns, ForceFloatAccumulate> &_parent;
+        const GemmInterleaved<strategy, To, Tr, OutputStage, MergeStep, FixedFormat, ForceThreadColumns, ForceFloatAccumulate> &_parent;
 
         /* K, X and multi parameters for current iteration. */
         unsigned int _k0=0, _x0=0, _multi=0;
@@ -499,9 +498,9 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
         bool _newmulti=true;
 
     public:
-        blockwalker(const GemmInterleaved<strategy, Tlo, Tro, Tr, OutputStage, MergeStep, FixedFormat, ForceThreadColumns, ForceFloatAccumulate> &parent) : _parent(parent) { }
+        blockwalker(const GemmInterleaved<strategy, To, Tr, OutputStage, MergeStep, FixedFormat, ForceThreadColumns, ForceFloatAccumulate> &parent) : _parent(parent) { }
 
-        blockwalker(const GemmInterleaved<strategy, Tlo, Tro, Tr, OutputStage, MergeStep, FixedFormat, ForceThreadColumns, ForceFloatAccumulate> &parent,
+        blockwalker(const GemmInterleaved<strategy, To, Tr, OutputStage, MergeStep, FixedFormat, ForceThreadColumns, ForceFloatAccumulate> &parent,
                     unsigned int x_start, unsigned int x_end) : _parent(parent), _x0 (_x_start), _x_start(x_start), _x_end(x_end) { }
 
         unsigned int xmax() {
@@ -555,7 +554,7 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
         unsigned int k_depth = _k_block;
 
         if (std::is_same<OutputStage, Requantize32>::value) {
-            k_depth += sizeof(int32_t) / sizeof(Tloi);
+            k_depth += sizeof(int32_t) / sizeof(Toi);
         }
 
         return k_depth;
@@ -565,10 +564,10 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
     size_t get_a_working_size() const {
         if (_thread_columns) {
             // For 2D threading: allocate a buffer of one block of rows per thread
-            return ROUND_UP(sizeof(Tloi) * get_total_k_depth() * strategy::out_height() * _maxthreads);
+            return ROUND_UP(sizeof(Toi) * get_total_k_depth() * strategy::out_height() * _maxthreads);
         } else {
             // For 1D threaded: one of these needed, regardless of thread count.  Divided according to window.
-            return ROUND_UP(sizeof(Tloi) * get_total_k_depth() * _Mround * _nbatches);
+            return ROUND_UP(sizeof(Toi) * get_total_k_depth() * _Mround * _nbatches);
         }
     }
 
@@ -693,7 +692,7 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
             }
 
             // Don't bother to block below this size threshold (1.25X target size)
-            unsigned int scaling_threshold = ((target_bytes_per_block * 5) / 4) / sizeof(Tloi);
+            unsigned int scaling_threshold = ((target_bytes_per_block * 5) / 4) / sizeof(Toi);
 
             if (get_ktotal(args) <= scaling_threshold) {
                 return get_ktotal(args);
@@ -701,7 +700,7 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
 
             // Once we are blocking, this (lower) threshold determines when we should use more blocks
             // NOTE: Could be that some factor-based solution would work better here.
-            unsigned int max_block_size = target_bytes_per_block / sizeof(Tloi);
+            unsigned int max_block_size = target_bytes_per_block / sizeof(Toi);
 
             unsigned int num_k_blocks = iceildiv(get_ktotal(args), max_block_size);
 
@@ -714,7 +713,7 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
 
         // k_block: Find out how much of the larger array can be loaded into half the cache.
         // This should account for associative caches.
-        k_block = (L1_size / 2) / (sizeof(Tloi) * (std::max(strategy::out_width(), strategy::out_height())));
+        k_block = (L1_size / 2) / (sizeof(Toi) * (std::max(strategy::out_width(), strategy::out_height())));
 
         // Needs to be (at least a single) multiple of the K unroll level.
         k_block /= strategy::k_unroll();
@@ -762,14 +761,14 @@ class GemmInterleaved : public GemmCommon<Tlo, Tro, Tr> {
         // x_block: Work out how many rows (of length k_block) will fit in the L2
         // Don't allocate more than 90% of the L2 to allow for overheads, and subtract off the L1 contents.
         const unsigned int scaled_l2_size = (L2_size * 9) / 10;
-        const unsigned int k_block_area = k_block * sizeof(Tloi) * (strategy::out_width() + strategy::out_height());
+        const unsigned int k_block_area = k_block * sizeof(Toi) * (strategy::out_width() + strategy::out_height());
 
         // .. if the L1 contents is bigger than the L2, just return a minimal size block.
         if (k_block_area > scaled_l2_size) {
             return strategy::out_width();
         }
 
-        x_block = (scaled_l2_size - k_block_area) / (sizeof(Tloi) * k_block);
+        x_block = (scaled_l2_size - k_block_area) / (sizeof(Toi) * k_block);
 
         // Needs to be (at least a single) multiple of the kernel output width.
         x_block /= strategy::out_width();
@@ -867,8 +866,8 @@ public:
             const auto end_x = std::min(work_range.get_position_end(1) * strategy::out_width(), _Nsize);
 
             Tri * const c_panel = reinterpret_cast<Tri *>(working_space_bytes + (threadid * get_c_working_size()));
-            Tloi * const a_panel = reinterpret_cast<Tloi *>(working_space_bytes + (_maxthreads * get_c_working_size()) +
-                                       (threadid * sizeof(Tloi) * get_total_k_depth() * strategy::out_height()));
+            Toi * const a_panel = reinterpret_cast<Toi *>(working_space_bytes + (_maxthreads * get_c_working_size()) +
+                                       (threadid * sizeof(Toi) * get_total_k_depth() * strategy::out_height()));
 
             for (unsigned int multi=0; multi<_nmulti; multi++) {
                 for (unsigned int k0=0; k0<_Ktotal; k0+=_k_block) {
@@ -885,8 +884,8 @@ public:
                     // Figure out how many "K" the kernel will actually process.
                     unsigned int kern_k = roundup(kmax - k0, strategy::k_unroll());
 
-                    const Troi *b_ptr = FixedFormat ?
-                        reinterpret_cast<const Troi *>(this->_Bptr) + (multi * this->_B_multi_stride) +
+                    const Toi *b_ptr = FixedFormat ?
+                        reinterpret_cast<const Toi *>(this->_Bptr) + (multi * this->_B_multi_stride) +
                                                      ((start_x / get_stripe_width<strategy, FixedFormat>::get()) * this->_ldb) +
                                                      (k0 * get_stripe_width<strategy, FixedFormat>::get()) :
                         _B_transposed + (rounded_width * _Ktotal * multi) + (k0 * rounded_width) + (start_x * kern_k);
@@ -900,7 +899,7 @@ public:
                         // Set up transposed 'A' block
                         {
 #ifdef CYCLE_PROFILING
-                            auto p=prof.ScopedProfiler(PROFILE_PREPA, strategy::out_height() * (kmax-k0) * sizeof(Tloi));
+                            auto p=prof.ScopedProfiler(PROFILE_PREPA, strategy::out_height() * (kmax-k0) * sizeof(Toi));
 #endif
                             // See comment above on transform_type<> class: this extracts either 'transforms' or
                             // 'transforms_quantized' as appropriate.
@@ -968,10 +967,10 @@ public:
             // (one per thread) first, followed by the (window-divided) A
             // buffer.
             // Set a_panel to the base of the A buffers - compute offsets into it based on M/batches later.
-            Tloi * const a_panel = reinterpret_cast<Tloi *>(working_space_bytes + (_maxthreads * get_c_working_size()));
+            Toi * const a_panel = reinterpret_cast<Toi *>(working_space_bytes + (_maxthreads * get_c_working_size()));
             Tri * const c_panel = reinterpret_cast<Tri *>(working_space_bytes + (threadid * get_c_working_size()));
 
-            const Troi *b_panel;
+            const Toi *b_panel;
             b_panel = _B_transposed;
 
             // newkblock() is always true on the first iteration, so these will be set properly on the first loop.
@@ -990,7 +989,7 @@ public:
             for (;!current.done();current.advance()) {
                 if (current.newkblock()) {
 #ifdef CYCLE_PROFILING
-                    auto p=prof.ScopedProfiler(PROFILE_PREPA, (end - start) * strategy::out_height() * (current.kmax()-current.k0()) * sizeof(Tloi));
+                    auto p=prof.ScopedProfiler(PROFILE_PREPA, (end - start) * strategy::out_height() * (current.kmax()-current.k0()) * sizeof(Toi));
 #endif
                     // See comment above on transform_type<> class: this extracts either 'transforms' or
                     // 'transforms_quantized' as appropriate.
@@ -1026,7 +1025,7 @@ public:
                     // larger than the (rounded) K value.
 
                     if(std::is_same<OutputStage, Requantize32>::value) {
-                        a_panel_stride = kern_k + (sizeof(int32_t) / sizeof(Tloi));
+                        a_panel_stride = kern_k + (sizeof(int32_t) / sizeof(Toi));
                     } else {
                         a_panel_stride = kern_k;
                     }
@@ -1034,7 +1033,7 @@ public:
 
                 // For FixedFormat cases, figure out the B pointer.  The loop below moves through batches and vertically through the output so this will be the same throughout.
                 if (FixedFormat) {
-                    b_panel = reinterpret_cast<const Troi *>(this->_Bptr) + (current.multi() * this->_B_multi_stride) +
+                    b_panel = reinterpret_cast<const Toi *>(this->_Bptr) + (current.multi() * this->_B_multi_stride) +
                                                                            ((current.x0() / get_stripe_width<strategy, FixedFormat>::get()) * this->_ldb) +
                                                                            (current.k0() * get_stripe_width<strategy, FixedFormat>::get());
                 }
@@ -1044,7 +1043,7 @@ public:
                     unsigned int first_m = (batch == batch_0)   ? m_0   : 0;
                     unsigned int last_m  = (batch == batch_end) ? m_max : _Msize;
 
-                    const Tloi *a_ptr = a_panel + (batch * _Mround + first_m) * get_total_k_depth();
+                    const Toi *a_ptr = a_panel + (batch * _Mround + first_m) * get_total_k_depth();
 
                     if (first_m >= last_m)
                         continue;
@@ -1166,7 +1165,7 @@ public:
 
         unsigned int x_size = roundup(_Nsize, strategy::out_width());
 
-        return (x_size * _Ktotal * _nmulti * sizeof(Troi)) + get_col_sum_size();
+        return (x_size * _Ktotal * _nmulti * sizeof(Toi)) + get_col_sum_size();
     }
 
     size_t get_B_pretranspose_window_size() const override {
@@ -1176,7 +1175,7 @@ public:
         return n_blocks * k_blocks * _nmulti;
     }
 
-    void requantize_bias(void *in_buffer, const Tro *B, const int ldb, const int B_multi_stride) override {
+    void requantize_bias(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override {
         if (std::is_same<OutputStage, Requantize32>::value) {
             col_bias = reinterpret_cast<int32_t *>(in_buffer);
 
@@ -1196,11 +1195,11 @@ public:
         return transforms.PrepareB_supports_transpose();
     }
 
-    void pretranspose_B_array(void *in_buffer, const Tro *B, const int ldb, const int B_multi_stride, const bool transposed) override {
+    void pretranspose_B_array(void *in_buffer, const To *B, const int ldb, const int B_multi_stride, const bool transposed) override {
         pretranspose_B_array_part(in_buffer, B, ldb, B_multi_stride, transposed, 0, get_B_pretranspose_window_size());
     }
 
-    void pretranspose_B_array_part(void *in_buffer, const Tro *B, const int ldb, const int B_multi_stride, const bool transposed, size_t start, size_t end) override {
+    void pretranspose_B_array_part(void *in_buffer, const To *B, const int ldb, const int B_multi_stride, const bool transposed, size_t start, size_t end) override {
         // Perform column sums etc as part of the last block.
         if (end >= get_B_pretranspose_window_size()) {
             requantize_bias(in_buffer, B, ldb, B_multi_stride);
@@ -1208,7 +1207,7 @@ public:
 
         // Put the transposed data after the column sums - in non-quantized cases get_col_sum_size() == 0
         uintptr_t buffer_int = reinterpret_cast<uintptr_t>(in_buffer);
-        Troi *buffer = reinterpret_cast<Troi *>(buffer_int + get_col_sum_size());
+        Toi *buffer = reinterpret_cast<Toi *>(buffer_int + get_col_sum_size());
         _B_transposed = buffer;
 
         blockwalker current(*this);
@@ -1293,7 +1292,7 @@ public:
     void set_pretransposed_B_data(void *in_buffer) override {
         // Put the transposed data after the column sums - in non-quantized cases get_col_sum_size() == 0
         uintptr_t buffer_int = reinterpret_cast<uintptr_t>(in_buffer);
-        _B_transposed = reinterpret_cast<Troi *>(buffer_int + get_col_sum_size());
+        _B_transposed = reinterpret_cast<Toi *>(buffer_int + get_col_sum_size());
         col_bias = reinterpret_cast<int32_t *>(in_buffer);
     }
 
@@ -1313,14 +1312,14 @@ public:
         }
     }
 
-    void set_indirect_parameters(size_t string_len, const Tlo * const * const *ptr) override {
+    void set_indirect_parameters(size_t string_len, const To * const * const *ptr) override {
         assert(string_len == _Ksize);
         _indirect_buf = ptr;
     }
 
     void set_convolution_parameters(ConvolutionParameters parms) override {
         assert(parms.input_channels == _Ksize);
-        _convolver = std::unique_ptr<convolver<Tlo>>(new convolver<Tlo>(parms));
+        _convolver = std::unique_ptr<convolver<To>>(new convolver<To>(parms));
     }
 
     // Estimate cycles for given problem given provided parameters
@@ -1331,7 +1330,7 @@ public:
         const PerformanceParameters &params = strategy::template get_performance_parameters<perf_type>(args._ci);
 
         uint64_t total_macs    = static_cast<uint64_t>(args._nbatches) * args._nmulti * roundup(args._Msize, strategy::out_height()) * roundup(args._Nsize, strategy::out_width()) * get_ktotal(args);
-        uint64_t prepare_bytes = static_cast<uint64_t>(args._nbatches) * args._nmulti * roundup(args._Msize, strategy::out_height()) * get_ktotal(args) * sizeof(Tloi);
+        uint64_t prepare_bytes = static_cast<uint64_t>(args._nbatches) * args._nmulti * roundup(args._Msize, strategy::out_height()) * get_ktotal(args) * sizeof(Toi);
         uint64_t merge_bytes   = static_cast<uint64_t>(args._nbatches) * args._nmulti * k_blocks * args._Msize * roundup(args._Nsize, strategy::out_width()) * sizeof(Tr);
 
         float mac_cycles     = static_cast<float>(total_macs) / params.kernel_macs_cycle;
@@ -1358,7 +1357,7 @@ public:
         c.inner_block_size = _k_block;
         c.outer_block_size = _x_block;
         c.filter = get_type_name<strategy>();
-        c.weight_format = get_weight_format(get_kernel_weight_format<strategy, FixedFormat, Tro>::get(), sizeof(Tro));
+        c.weight_format = get_weight_format(get_kernel_weight_format<strategy, FixedFormat, To>::get(), sizeof(To));
 
         return c;
     }
@@ -1366,21 +1365,21 @@ public:
 
 // Aliases for the variations
 template<typename strategy, typename To, typename Tr, typename OutputStage=Nothing>
-using GemmInterleavedNoMerge = GemmInterleaved<strategy, To, To, Tr, OutputStage, false>;
+using GemmInterleavedNoMerge = GemmInterleaved<strategy, To, Tr, OutputStage, false>;
 
 template<typename strategy, typename To, typename Tr, typename OutputStage=Nothing>
-using GemmInterleavedFixedFormat = GemmInterleaved<strategy, To, To, Tr, OutputStage, true, true>;
+using GemmInterleavedFixedFormat = GemmInterleaved<strategy, To, Tr, OutputStage, true, true>;
 
 template<typename strategy, typename To, typename Tr>
-using GemmInterleavedPretransposedNoMergeQuantizedInline = GemmInterleaved<strategy, To, To, Tr, Requantize32, false>;
+using GemmInterleavedPretransposedNoMergeQuantizedInline = GemmInterleaved<strategy, To, Tr, Requantize32, false>;
 
-template<typename strategy, typename Tlo, typename Tro, typename Tr>
-using GemmInterleavedQuantized = GemmInterleaved<strategy, Tlo, Tro, Tr, Requantize32>;
+template<typename strategy, typename To, typename Tr>
+using GemmInterleavedQuantized = GemmInterleaved<strategy, To, Tr, Requantize32>;
 
 template<typename strategy, typename To, typename Tr>
-using GemmInterleavedNoMergeDequantized = GemmInterleaved<strategy, To, To, Tr, DequantizeFloat, false>;
+using GemmInterleavedNoMergeDequantized = GemmInterleaved<strategy, To, Tr, DequantizeFloat, false>;
 
-template<typename strategy, typename Tlo, typename Tro, typename Tr>
-using GemmInterleavedDequantized = GemmInterleaved<strategy, Tlo, Tro, Tr, DequantizeFloat>;
+template<typename strategy, typename To, typename Tr>
+using GemmInterleavedDequantized = GemmInterleaved<strategy, To, Tr, DequantizeFloat>;
 
 } // namespace arm_gemm
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_q8_mixed.cpp b/src/core/NEON/kernels/arm_gemm/gemm_q8_mixed.cpp
deleted file mode 100644
index a48244cb3c..0000000000
--- a/src/core/NEON/kernels/arm_gemm/gemm_q8_mixed.cpp
+++ /dev/null
@@ -1,138 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef __aarch64__
-
-#include "arm_gemm.hpp"
-
-#include "kernels/a64_hybrid_u8s8qa_dot_4x16.hpp"
-#include "kernels/a64_hybrid_u8s8qa_mmla_4x16.hpp"
-#include "kernels/a64_hybrid_u8s8s32_dot_6x16.hpp"
-#include "kernels/a64_hybrid_u8s8s32_mmla_6x16.hpp"
-#include "kernels/a64_interleaved_u8s8s32_mmla_8x12.hpp"
-
-#ifdef ARM_COMPUTE_ENABLE_SVE
-#include "kernels/sve_hybrid_u8s8qa_dot_4x4VL.hpp"
-#include "kernels/sve_interleaved_u8s8s32_mmla_8x3VL.hpp"
-#include "kernels/sve_hybrid_u8s8s32_mmla_6x4VL.hpp"
-#include "kernels/sve_hybrid_u8s8qa_mmla_4x4VL.hpp"
-#endif // ARM_COMPUTE_ENABLE_SVE
-
-#include "gemm_hybrid_indirect.hpp"
-#include "gemm_hybrid_quantized.hpp"
-#include "gemm_interleaved.hpp"
-#include "gemv_pretransposed.hpp"
-#include "quantize_wrapper.hpp"
-#include "utils.hpp"
-
-namespace arm_gemm {
-
-static const GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32> gemm_q8_mixed_methods[] =
-{
-#ifdef ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_HYBRID,
-    "sve_hybrid_u8s8qa_mmla_4x4VL",
-    [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_asymmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8s8qa_mmla_4x4VL, uint8_t, int8_t, uint8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8s8qa_mmla_4x4VL, uint8_t, int8_t, uint8_t, Requantize32>(args, qp); }
-),
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_INTERLEAVED,
-    "sve_interleaved_u8s8s32_mmla_8x3VL",
-    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8s8s32_mmla_8x3VL, uint8_t, int8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8s8s32_mmla_8x3VL, uint8_t, int8_t, uint8_t>(args, qp); }
-),
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_INTERLEAVED,
-    "sve_hybrid_u8s8s32_mmla_6x4VL",
-    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8s8s32_mmla_6x4VL, uint8_t, int8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8s8s32_mmla_6x4VL, uint8_t, int8_t, uint8_t, Requantize32, true>(args, qp); }
-),
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_HYBRID,
-    "sve_hybrid_u8s8qa_dot_4x4VL",
-    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8s8qa_dot_4x4VL, uint8_t, int8_t, uint8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8s8qa_dot_4x4VL, uint8_t, int8_t, uint8_t, Requantize32>(args, qp); }
-),
-#endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_HYBRID,
-    "a64_hybrid_u8s8qa_mmla_4x16",
-    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8s8qa_mmla_4x16, uint8_t, int8_t, uint8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8s8qa_mmla_4x16, uint8_t, int8_t, uint8_t, Requantize32>(args, qp); }
-),
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_INTERLEAVED,
-    "a64_interleaved_u8s8s32_mmla_8x12",
-    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_u8s8s32_mmla_8x12, uint8_t, int8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_u8s8s32_mmla_8x12, uint8_t, int8_t, uint8_t>(args, qp); }
-),
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_INTERLEAVED,
-    "a64_hybrid_u8s8s32_mmla_6x16",
-    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8s8s32_mmla_6x16, uint8_t, int8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8s8s32_mmla_6x16, uint8_t, int8_t, uint8_t, Requantize32, true>(args, qp); }
-),
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_HYBRID,
-    "a64_hybrid_u8s8qa_dot_4x16",
-    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_svei8mm() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8s8qa_dot_4x16, uint8_t, int8_t, uint8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8s8qa_dot_4x16, uint8_t, int8_t, uint8_t, Requantize32>(args, qp); }
-),
-GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32>::with_estimate(
-    GemmMethod::GEMM_HYBRID,
-    "a64_hybrid_u8s8s32_dot_6x16",
-    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && args._ci->has_i8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8s8s32_dot_6x16, uint8_t, int8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8s8s32_dot_6x16, uint8_t, int8_t, uint8_t, Requantize32, true>(args, qp); }
-),
-{
-    GemmMethod::DEFAULT,
-    "",
-    nullptr,
-    nullptr,
-    nullptr
-}
-};
-
-template<>
-const GemmImplementation<uint8_t, int8_t, uint8_t, Requantize32> *gemm_implementation_list<uint8_t, int8_t, uint8_t, Requantize32>() {
-    return gemm_q8_mixed_methods;
-}
-
-template UniqueGemmCommon<uint8_t, int8_t, uint8_t> gemm<uint8_t, int8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-template bool has_opt_gemm<uint8_t, int8_t, uint8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &);
-template KernelDescription get_gemm_method<uint8_t, int8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-template std::vector<KernelDescription> get_compatible_kernels<uint8_t, int8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-
-} // namespace arm_gemm
-
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
index 18008e713e..321c97262f 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
@@ -58,6 +58,7 @@
 
 #include "gemm_hybrid_indirect.hpp"
 #include "gemm_hybrid_quantized.hpp"
+#include "gemm_hybrid_quantized_inline.hpp"
 #include "gemm_interleaved.hpp"
 #include "gemv_pretransposed.hpp"
 #include "quantize_wrapper.hpp"
@@ -65,7 +66,7 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<int8_t, int8_t, int8_t, Requantize32> gemm_qint8_methods[] =
+static const GemmImplementation<int8_t, int8_t, Requantize32> gemm_qint8_methods[] =
 {
 #ifdef ARM_COMPUTE_ENABLE_SVE
 #ifdef ARM_COMPUTE_ENABLE_SME2
@@ -100,90 +101,90 @@ static const GemmImplementation<int8_t, int8_t, int8_t, Requantize32> gemm_qint8
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_2VLx2VL, int8_t, int8_t>(args, qp); }
 },
 #endif // ARM_COMPUTE_ENABLE_SME2
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_s8qa_mmla_4x4VL",
     [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_asymmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, int8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_s8qs_mmla_6x4VL",
     [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_symmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, int8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_s8s32_mmla_8x3VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t, int8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_hybrid_s8s32_mmla_6x4VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, int8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_s8qs_dot_6x4VL",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_symmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, int8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_s8qa_dot_4x4VL",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, int8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_s8s32_dot_6x4VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, int8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_s8s32_dot_8x3VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve() && (args._Ksize>4); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t, int8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t>(args, qp); }
 ),
 #endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_s8qa_mmla_4x16",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, int8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_s8qs_mmla_6x16",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_symmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, int8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_s8s32_mmla_8x12",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t, int8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_hybrid_s8s32_mmla_6x16",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, int8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, Requantize32, true>(args, qp); }
 ),
 {
     GemmMethod::GEMM_HYBRID_QUANTIZED,
@@ -204,42 +205,42 @@ GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
     "a64_gemm_s16_8x12",
     nullptr,
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->get_cpu_model() == CPUModel::A53 && ((args._Msize > 28) || ((args._Msize % 8) > 4)); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s16_8x12, int8_t, int8_t, int8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s16_8x12, int8_t, int8_t>(args, qp); }
 },
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_s8qs_dot_6x16",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_dotprod() && quant_hybrid_symmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, int8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_s8qa_dot_4x16",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_dotprod() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, int8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_s8s32_dot_6x16",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, int8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_s8_8x12",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t, int8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
+GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_s8_4x4",
     nullptr,
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t, int8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t>(args, qp); }
 ),
 {
     GemmMethod::QUANTIZE_WRAPPER,
@@ -258,14 +259,14 @@ GemmImplementation<int8_t, int8_t, int8_t, Requantize32>::with_estimate(
 };
 
 template<>
-const GemmImplementation<int8_t, int8_t, int8_t, Requantize32> *gemm_implementation_list<int8_t, int8_t, int8_t, Requantize32>() {
+const GemmImplementation<int8_t, int8_t, Requantize32> *gemm_implementation_list<int8_t, int8_t, Requantize32>() {
     return gemm_qint8_methods;
 }
 
-template UniqueGemmCommon<int8_t, int8_t, int8_t> gemm<int8_t, int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-template bool has_opt_gemm<int8_t, int8_t, int8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &os);
-template KernelDescription get_gemm_method<int8_t, int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-template std::vector<KernelDescription> get_compatible_kernels<int8_t, int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template UniqueGemmCommon<int8_t, int8_t> gemm<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template bool has_opt_gemm<int8_t, int8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &os);
+template KernelDescription get_gemm_method<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template std::vector<KernelDescription> get_compatible_kernels<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
 
 } // namespace arm_gemm
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
index 7c182b6777..93eecf991e 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
@@ -54,13 +54,14 @@
 
 #include "gemm_hybrid_indirect.hpp"
 #include "gemm_hybrid_quantized.hpp"
+#include "gemm_hybrid_quantized_inline.hpp"
 #include "gemm_interleaved.hpp"
 #include "gemv_pretransposed.hpp"
 #include "quantize_wrapper.hpp"
 
 namespace arm_gemm {
 
-static const GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32> gemm_quint8_methods[] =
+static const GemmImplementation<uint8_t, uint8_t, Requantize32> gemm_quint8_methods[] =
 {
 #ifdef ARM_COMPUTE_ENABLE_SVE
 #ifdef ARM_COMPUTE_ENABLE_SME2
@@ -96,69 +97,69 @@ static const GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32> gemm_qu
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_u8q_mopa_2VLx2VL, uint8_t, uint8_t>(args, qp); }
 },
 #endif // ARM_COMPUTE_ENABLE_SME2
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_u8qa_mmla_4x4VL",
     [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_asymmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8qa_mmla_4x4VL, uint8_t, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8qa_mmla_4x4VL, uint8_t, uint8_t, uint8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8qa_mmla_4x4VL, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8qa_mmla_4x4VL, uint8_t, uint8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_u8u32_mmla_8x3VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_hybrid_u8u32_mmla_6x4VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, uint8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_u8qa_dot_4x4VL",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8qa_dot_4x4VL, uint8_t, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8qa_dot_4x4VL, uint8_t, uint8_t, uint8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8qa_dot_4x4VL, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8qa_dot_4x4VL, uint8_t, uint8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_u8u32_dot_6x4VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, uint8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_u8u32_dot_8x3VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve() && (args._Ksize>4); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t>(args, qp); }
 ),
 #endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_u8qa_mmla_4x16",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_mmla_4x16, uint8_t, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8qa_mmla_4x16, uint8_t, uint8_t, uint8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_mmla_4x16, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8qa_mmla_4x16, uint8_t, uint8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_u8u32_mmla_8x12",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_hybrid_u8u32_mmla_6x16",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, uint8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, Requantize32, true>(args, qp); }
 ),
 {
     GemmMethod::GEMM_HYBRID_QUANTIZED,
@@ -179,35 +180,35 @@ GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
     "a64_gemm_u16_8x12",
     nullptr,
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4; },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u16_8x12, uint8_t, uint8_t, uint8_t>(args, qp); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u16_8x12, uint8_t, uint8_t>(args, qp); },
 },
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_u8qa_dot_4x16",
     [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_dotprod() && quant_hybrid_asymmetric(qp); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, uint8_t, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, uint8_t, uint8_t, uint8_t, Requantize32>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, uint8_t, uint8_t, Requantize32>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8qa_dot_4x16, uint8_t, uint8_t, Requantize32>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_u8u32_dot_6x16",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, uint8_t, Requantize32, true>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, Requantize32, true>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, Requantize32, true>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u8_8x12",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_8x12, uint8_t, uint8_t>(args, qp); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
+GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u8_4x4",
     nullptr,
-    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t, uint8_t>(args, qp); }
+    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t>::estimate_cycles<uint8_t>(args); },
+    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u8_4x4, uint8_t, uint8_t>(args, qp); }
 ),
 {
     GemmMethod::QUANTIZE_WRAPPER,
@@ -226,14 +227,14 @@ GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32>::with_estimate(
 };
 
 template<>
-const GemmImplementation<uint8_t, uint8_t, uint8_t, Requantize32> *gemm_implementation_list<uint8_t, uint8_t, uint8_t, Requantize32>() {
+const GemmImplementation<uint8_t, uint8_t, Requantize32> *gemm_implementation_list<uint8_t, uint8_t, Requantize32>() {
     return gemm_quint8_methods;
 }
 
-template UniqueGemmCommon<uint8_t, uint8_t, uint8_t> gemm<uint8_t, uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-template bool has_opt_gemm<uint8_t, uint8_t, uint8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &os);
-template KernelDescription get_gemm_method<uint8_t, uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
-template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template UniqueGemmCommon<uint8_t, uint8_t> gemm<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template bool has_opt_gemm<uint8_t, uint8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &os);
+template KernelDescription get_gemm_method<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
 
 } // namespace arm_gemm
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_s8fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_s8fp32.cpp
index 1d995a87b5..38d9b763f6 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_s8fp32.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_s8fp32.cpp
@@ -48,7 +48,7 @@
 #include <vector>
 namespace arm_gemm {
 
-static const GemmImplementation<int8_t, int8_t, float, DequantizeFloat> gemm_s8fp32_methods[] =
+static const GemmImplementation<int8_t, float, DequantizeFloat> gemm_s8fp32_methods[] =
 {
 #ifdef ARM_COMPUTE_ENABLE_SVE
 #ifdef ARM_COMPUTE_ENABLE_SME2
@@ -76,48 +76,48 @@ static const GemmImplementation<int8_t, int8_t, float, DequantizeFloat> gemm_s8f
     [](const GemmArgs &args, const DequantizeFloat &dq) { return new GemmInterleavedNoMergeDequantized<cls_sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL, int8_t, float>(args, dq); }
 },
 #endif // ARM_COMPUTE_ENABLE_SME2
-GemmImplementation<int8_t, int8_t, float, DequantizeFloat>::with_estimate(
+GemmImplementation<int8_t, float, DequantizeFloat>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_s8s32_mmla_8x3VL",
     [](const GemmArgs &args, const DequantizeFloat &) { return args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t, float>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t, float>(args, qp); }
+    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, float>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, float>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, float, DequantizeFloat>::with_estimate(
+GemmImplementation<int8_t, float, DequantizeFloat>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_s8s32_dot_8x3VL",
     [](const GemmArgs &args, const DequantizeFloat &) { return args._ci->has_sve(); },
-    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t, float>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t, float>(args, qp); }
+    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, float>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, float>(args, qp); }
 ),
 #endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<int8_t, int8_t, float, DequantizeFloat>::with_estimate(
+GemmImplementation<int8_t, float, DequantizeFloat>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_s8s32_mmla_8x12",
     [](const GemmArgs &args, const DequantizeFloat &) { return args._ci->has_i8mm(); },
-    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t, float>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t, float>(args, qp); }
+    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, float>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, float>(args, qp); }
 ),
 {
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_s16_8x12",
     nullptr,
     [](const GemmArgs &args, const DequantizeFloat &) { return args._ci->get_cpu_model() == CPUModel::A53 && ((args._Msize > 28) || ((args._Msize % 8) > 4)); },
-    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_gemm_s16_8x12, int8_t, int8_t, float>(args, qp); }
+    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_gemm_s16_8x12, int8_t, float>(args, qp); }
 },
-GemmImplementation<int8_t, int8_t, float, DequantizeFloat>::with_estimate(
+GemmImplementation<int8_t, float, DequantizeFloat>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_s8_8x12",
     [](const GemmArgs &args, const DequantizeFloat &) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_a64_gemm_s8_8x12, int8_t, int8_t, float>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_gemm_s8_8x12, int8_t, int8_t, float>(args, qp); }
+    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_a64_gemm_s8_8x12, int8_t, float>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_gemm_s8_8x12, int8_t, float>(args, qp); }
 ),
-GemmImplementation<int8_t, int8_t, float, DequantizeFloat>::with_estimate(
+GemmImplementation<int8_t, float, DequantizeFloat>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_s8_4x4",
     nullptr,
-    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_a64_gemm_s8_4x4, int8_t, int8_t, float>::estimate_cycles<int8_t>(args); },
-    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_gemm_s8_4x4, int8_t, int8_t, float>(args, qp); }
+    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_a64_gemm_s8_4x4, int8_t, float>::estimate_cycles<int8_t>(args); },
+    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_gemm_s8_4x4, int8_t, float>(args, qp); }
 ),
 {
     GemmMethod::DEFAULT,
@@ -129,13 +129,13 @@ GemmImplementation<int8_t, int8_t, float, DequantizeFloat>::with_estimate(
 };
 
 template<>
-const GemmImplementation<int8_t, int8_t, float, DequantizeFloat> *gemm_implementation_list<int8_t, int8_t, float, DequantizeFloat>() {
+const GemmImplementation<int8_t, float, DequantizeFloat> *gemm_implementation_list<int8_t, float, DequantizeFloat>() {
     return gemm_s8fp32_methods;
 }
 
-template UniqueGemmCommon<int8_t, int8_t, float> gemm<int8_t, int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
-template KernelDescription get_gemm_method<int8_t, int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
-template std::vector<KernelDescription> get_compatible_kernels<int8_t, int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
+template UniqueGemmCommon<int8_t, float> gemm<int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
+template KernelDescription get_gemm_method<int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
+template std::vector<KernelDescription> get_compatible_kernels<int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
 
 } // namespace arm_gemm
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_u8s8fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_u8s8fp32.cpp
deleted file mode 100644
index 606b422b0b..0000000000
--- a/src/core/NEON/kernels/arm_gemm/gemm_u8s8fp32.cpp
+++ /dev/null
@@ -1,89 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef __aarch64__
-
-#include "arm_gemm.hpp"
-
-#include "kernels/a64_hybrid_u8s8qa_dot_4x16.hpp"
-#include "kernels/a64_hybrid_u8s8qa_mmla_4x16.hpp"
-#include "kernels/a64_hybrid_u8s8s32_dot_6x16.hpp"
-#include "kernels/a64_hybrid_u8s8s32_mmla_6x16.hpp"
-#include "kernels/a64_interleaved_u8s8s32_mmla_8x12.hpp"
-
-#ifdef ARM_COMPUTE_ENABLE_SVE
-#include "kernels/sve_hybrid_u8s8qa_dot_4x4VL.hpp"
-#include "kernels/sve_interleaved_u8s8s32_mmla_8x3VL.hpp"
-#include "kernels/sve_hybrid_u8s8s32_mmla_6x4VL.hpp"
-#include "kernels/sve_hybrid_u8s8qa_mmla_4x4VL.hpp"
-#endif // ARM_COMPUTE_ENABLE_SVE
-
-#include "gemm_hybrid_indirect.hpp"
-#include "gemm_hybrid_quantized.hpp"
-#include "gemm_interleaved.hpp"
-#include "gemv_pretransposed.hpp"
-#include "quantize_wrapper.hpp"
-#include "utils.hpp"
-
-namespace arm_gemm {
-
-static const GemmImplementation<uint8_t, int8_t, float, DequantizeFloat> gemm_u8s8fp32_methods[] =
-{
-#ifdef ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<uint8_t, int8_t, float, DequantizeFloat>::with_estimate(
-    GemmMethod::GEMM_INTERLEAVED,
-    "sve_interleaved_u8s8s32_mmla_8x3VL",
-    [](const GemmArgs &args, const DequantizeFloat &) { return args._ci->has_svei8mm(); },
-    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_sve_interleaved_u8s8s32_mmla_8x3VL, uint8_t, int8_t, float>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_sve_interleaved_u8s8s32_mmla_8x3VL, uint8_t, int8_t, float>(args, qp); }
-),
-#endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<uint8_t, int8_t, float, DequantizeFloat>::with_estimate(
-    GemmMethod::GEMM_INTERLEAVED,
-    "a64_interleaved_u8s8s32_mmla_8x12",
-    [](const GemmArgs &args, const DequantizeFloat &) { return args._ci->has_i8mm(); },
-    [](const GemmArgs &args, const DequantizeFloat &) { return GemmInterleavedDequantized<cls_a64_interleaved_u8s8s32_mmla_8x12, uint8_t, int8_t, float>::estimate_cycles<uint8_t>(args); },
-    [](const GemmArgs &args, const DequantizeFloat &qp) { return new GemmInterleavedDequantized<cls_a64_interleaved_u8s8s32_mmla_8x12, uint8_t, int8_t, float>(args, qp); }
-),
-{
-    GemmMethod::DEFAULT,
-    "",
-    nullptr,
-    nullptr,
-    nullptr
-}
-};
-
-template<>
-const GemmImplementation<uint8_t, int8_t, float, DequantizeFloat> *gemm_implementation_list<uint8_t, int8_t, float, DequantizeFloat>() {
-    return gemm_u8s8fp32_methods;
-}
-
-template UniqueGemmCommon<uint8_t, int8_t, float> gemm<uint8_t, int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
-template bool has_opt_gemm<uint8_t, int8_t, float, DequantizeFloat>(WeightFormat &weight_format, const GemmArgs &args, const DequantizeFloat &os);
-template KernelDescription get_gemm_method<uint8_t, int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
-template std::vector<KernelDescription> get_compatible_kernels<uint8_t, int8_t, float, DequantizeFloat>(const GemmArgs &args, const DequantizeFloat &os);
-
-} // namespace arm_gemm
-
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp
index 308452e304..44f085c183 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp
@@ -32,13 +32,13 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<uint16_t, uint16_t, uint32_t> gemm_u16_methods[] = {
+static const GemmImplementation<uint16_t, uint32_t> gemm_u16_methods[] = {
 {
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u16_8x12",
     nullptr,
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u16_8x12, uint16_t, uint16_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u16_8x12, uint16_t, uint32_t>(args); }
 },
 {
     GemmMethod::DEFAULT,
@@ -50,15 +50,15 @@ static const GemmImplementation<uint16_t, uint16_t, uint32_t> gemm_u16_methods[]
 };
 
 template<>
-const GemmImplementation<uint16_t, uint16_t, uint32_t> *gemm_implementation_list<uint16_t, uint16_t, uint32_t>() {
+const GemmImplementation<uint16_t, uint32_t> *gemm_implementation_list<uint16_t, uint32_t>() {
     return gemm_u16_methods;
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<uint16_t, uint16_t, uint32_t> gemm<uint16_t, uint16_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
-template bool has_opt_gemm<uint16_t, uint16_t, uint32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
-template KernelDescription get_gemm_method<uint16_t, uint16_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
-template std::vector<KernelDescription> get_compatible_kernels<uint16_t, uint16_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template UniqueGemmCommon<uint16_t, uint32_t> gemm<uint16_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<uint16_t, uint32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<uint16_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<uint16_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
 
 } // namespace arm_gemm
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
index fb442419b7..dfacb687a8 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
@@ -46,50 +46,50 @@
 
 namespace arm_gemm {
 
-static const GemmImplementation<uint8_t, uint8_t, uint32_t> gemm_u8_methods[] = {
+static const GemmImplementation<uint8_t, uint32_t> gemm_u8_methods[] = {
 #ifdef ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_u8u32_mmla_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_svei8mm(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint32_t>(args); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_u8u32_mmla_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_svei8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint32_t>(args); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_u8u32_dot_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_sve(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint32_t>(args); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_u8u32_dot_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_sve() && (args._Ksize>4); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint32_t>(args); }
 ),
 #endif // ARM_COMPUTE_ENABLE_SVE
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_interleaved_u8u32_mmla_8x12",
     [](const GemmArgs &args) { return args._ci->has_i8mm() && (args._Ksize>8); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint32_t>(args); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_u8u32_mmla_6x16",
     [](const GemmArgs &args) { return args._ci->has_i8mm(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint32_t>(args); }
 ),
 {
     GemmMethod::GEMM_HYBRID,
@@ -110,28 +110,28 @@ GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
     "a64_gemm_u16_8x12",
     nullptr,
     [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4; },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u16_8x12, uint8_t, uint8_t, uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u16_8x12, uint8_t, uint32_t>(args); },
 },
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_u8u32_dot_6x16",
     [](const GemmArgs &args) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint32_t>(args); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u8_8x12",
     [](const GemmArgs &args) { return args._ci->has_dotprod(); },
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_u8_8x12, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u8_8x12, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_u8_8x12, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u8_8x12, uint8_t, uint32_t>(args); }
 ),
-GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_gemm_u8_4x4",
     nullptr,
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_u8_4x4, uint8_t, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
-    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u8_4x4, uint8_t, uint8_t, uint32_t>(args); }
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_u8_4x4, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u8_4x4, uint8_t, uint32_t>(args); }
 ),
 {
     GemmMethod::DEFAULT,
@@ -143,15 +143,15 @@ GemmImplementation<uint8_t, uint8_t, uint32_t>::with_estimate(
 };
 
 template<>
-const GemmImplementation<uint8_t, uint8_t, uint32_t> *gemm_implementation_list<uint8_t, uint8_t, uint32_t>() {
+const GemmImplementation<uint8_t, uint32_t> *gemm_implementation_list<uint8_t, uint32_t>() {
     return gemm_u8_methods;
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<uint8_t, uint8_t, uint32_t> gemm<uint8_t, uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
-template bool has_opt_gemm<uint8_t, uint8_t, uint32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
-template KernelDescription get_gemm_method<uint8_t, uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
-template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint8_t, uint32_t, Nothing> (const GemmArgs &args, const Nothing &);
+template UniqueGemmCommon<uint8_t, uint32_t> gemm<uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<uint8_t, uint32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
+template KernelDescription get_gemm_method<uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint32_t, Nothing> (const GemmArgs &args, const Nothing &);
 
 } // namespace arm_gemm
 
diff --git a/src/core/NEON/kernels/arm_gemm/gemv_batched.hpp b/src/core/NEON/kernels/arm_gemm/gemv_batched.hpp
index aa03fb6aa1..ad504f2664 100644
--- a/src/core/NEON/kernels/arm_gemm/gemv_batched.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemv_batched.hpp
@@ -31,9 +31,9 @@ namespace arm_gemm {
  * efficiently as a GEMM (with M'=nbatches and nbatches'=1).  This wrapper
  * implements this.  */
 template<typename To, typename Tr>
-class GemvBatched : public GemmCommon<To, To, Tr> {
+class GemvBatched : public GemmCommon<To, Tr> {
 private:
-    UniqueGemmCommon<To, To, Tr> _subgemm = nullptr;
+    UniqueGemmCommon<To, Tr> _subgemm = nullptr;
 
 public:
     GemvBatched(const GemmArgs &args) {
@@ -42,7 +42,7 @@ public:
         newargs._Msize = args._nbatches;
         newargs._nbatches = 1;
         newargs._cfg = nullptr;
-        _subgemm = gemm<To,To,Tr>(newargs);
+        _subgemm = gemm<To,Tr>(newargs);
     }
 
     void set_arrays(const To *A, const int, const int A_batch_stride, const int A_multi_stride,
diff --git a/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp b/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
index 1a7c51c7a4..dbada36052 100644
--- a/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemv_pretransposed.hpp
@@ -85,7 +85,7 @@ void run_gemv_kernel<Requantize32>::run(
 //
 // batches are not supported as a batched GEMV makes no sense (can be converted to a GEMM).
 template<typename strategy, typename To, typename Tr, typename OutputStage=Nothing>
-class GemvPretransposed : public GemmCommon<To, To, Tr> {
+class GemvPretransposed : public GemmCommon<To, Tr> {
     typedef typename strategy::operand_type Toi;
     typedef typename strategy::result_type Tri;
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a32_interleave6_block1_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a32_interleave6_block1_fp32_fp32.hpp
index 3ad32d8a50..ffd9d4b22a 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a32_interleave6_block1_fp32_fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a32_interleave6_block1_fp32_fp32.hpp
@@ -75,61 +75,61 @@ void interleave_block<6, 1, VLType::None, false>(
     for (;width>7;width-=8) {
         __asm __volatile (
             // Load up 8 elements (2 vectors) from each of 8 sources.
-            "VLD1.32    {d0-d3}, [%[inptr0]]!\n"   // q0=A0A1A2A3
-            "VLD1.32    {d4-d7}, [%[inptr1]]!\n"   // q2=B0B1B2B3
-            "VLD1.32    {d8-d11}, [%[inptr2]]!\n"  // q4=C0C1C2C3
-            "VZIP.32    q0, q4\n"     // q0=A0C0A1C1, q4 = A2C2A3C3
-            "VLD1.32    {d12-d15}, [%[inptr3]]!\n" // q6=D0D1D2D3
-            "VZIP.32    q2, q6\n"     // q2=B0D0B1D1, q6 = B2D2B3D3
-            "VLD1.32    {d16-d19}, [%[inptr4]]!\n"
-            "VLD1.32    {d20-d23}, [%[inptr5]]!\n"
-            "VZIP.32    q8, q10\n"    // q8=E0F0E1F1, q10 = E2F2E3F3
+            "VLD1.32	{d0-d3}, [%[inptr0]]!\n"   // q0=A0A1A2A3
+            "VLD1.32	{d4-d7}, [%[inptr1]]!\n"   // q2=B0B1B2B3
+            "VLD1.32	{d8-d11}, [%[inptr2]]!\n"  // q4=C0C1C2C3
+            "VZIP.32	q0, q4\n"     // q0=A0C0A1C1, q4 = A2C2A3C3
+            "VLD1.32	{d12-d15}, [%[inptr3]]!\n" // q6=D0D1D2D3
+            "VZIP.32	q2, q6\n"     // q2=B0D0B1D1, q6 = B2D2B3D3
+            "VLD1.32	{d16-d19}, [%[inptr4]]!\n"
+            "VLD1.32	{d20-d23}, [%[inptr5]]!\n"
+            "VZIP.32	q8, q10\n"    // q8=E0F0E1F1, q10 = E2F2E3F3
             ASM_PREFETCH("[%[inptr0], #128]")
-            "VZIP.32    q0, q2\n"    // q0 = A0B0C0D0, q2 = A1B1C1D1
+            "VZIP.32	q0, q2\n"    // q0 = A0B0C0D0, q2 = A1B1C1D1
 
             // Store first elements
-            "VST1.32    {d0-d1}, [%[outptr]]!\n"
-            "VST1.32    {d16}, [%[outptr]]!\n"
+            "VST1.32	{d0-d1}, [%[outptr]]!\n"
+            "VST1.32	{d16}, [%[outptr]]!\n"
 
-            "VZIP.32    q4, q6\n"    // q4 = A2B2C2D2, q6 = A3B3C3D3
+            "VZIP.32	q4, q6\n"    // q4 = A2B2C2D2, q6 = A3B3C3D3
 
             // Store second elements
-            "VST1.32    {d4-d5}, [%[outptr]]!\n"
-            "VZIP.32    q1, q5\n"
+            "VST1.32	{d4-d5}, [%[outptr]]!\n"
+            "VZIP.32	q1, q5\n"
             ASM_PREFETCH("[%[inptr1], #128]")
-            "VST1.32    {d17}, [%[outptr]]!\n"
-            "VZIP.32    q3, q7\n"
+            "VST1.32	{d17}, [%[outptr]]!\n"
+            "VZIP.32	q3, q7\n"
 
             // Store third elements
-            "VZIP.32    q9, q11\n"
-            "VST1.32    {d8-d9}, [%[outptr]]!\n"
-            "VZIP.32    q1, q3\n"
+            "VZIP.32	q9, q11\n"
+            "VST1.32	{d8-d9}, [%[outptr]]!\n"
+            "VZIP.32	q1, q3\n"
             ASM_PREFETCH("[%[inptr2], #128]")
-            "VST1.32    {d20}, [%[outptr]]!\n"
+            "VST1.32	{d20}, [%[outptr]]!\n"
 
             // Store fourth elements
-            "VZIP.32    q5, q7\n"
-            "VST1.32    {d12-d13}, [%[outptr]]!\n"
+            "VZIP.32	q5, q7\n"
+            "VST1.32	{d12-d13}, [%[outptr]]!\n"
             ASM_PREFETCH("[%[inptr3], #128]")
-            "VST1.32    {d21}, [%[outptr]]!\n"
+            "VST1.32	{d21}, [%[outptr]]!\n"
 
             // Fifth
-            "VST1.32    {d2-d3}, [%[outptr]]!\n"
+            "VST1.32	{d2-d3}, [%[outptr]]!\n"
             ASM_PREFETCH("[%[inptr4], #128]")
-            "VST1.32    {d18}, [%[outptr]]!\n"
+            "VST1.32	{d18}, [%[outptr]]!\n"
 
             // Sixth
-            "VST1.32    {d6-d7}, [%[outptr]]!\n"
+            "VST1.32	{d6-d7}, [%[outptr]]!\n"
             ASM_PREFETCH("[%[inptr5], #128]")
-            "VST1.32    {d19}, [%[outptr]]!\n"
+            "VST1.32	{d19}, [%[outptr]]!\n"
 
             // Seventh
-            "VST1.32    {d10-d11}, [%[outptr]]!\n"
-            "VST1.32    {d22}, [%[outptr]]!\n"
+            "VST1.32	{d10-d11}, [%[outptr]]!\n"
+            "VST1.32	{d22}, [%[outptr]]!\n"
 
             // Eighth
-            "VST1.32    {d14-d15}, [%[outptr]]!\n"
-            "VST1.32    {d23}, [%[outptr]]!\n"
+            "VST1.32	{d14-d15}, [%[outptr]]!\n"
+            "VST1.32	{d23}, [%[outptr]]!\n"
 
             : [inptr0] "+r" (inptr0), [inptr1] "+r" (inptr1), [inptr2] "+r" (inptr2), [inptr3] "+r" (inptr3),
               [inptr4] "+r" (inptr4), [inptr5] "+r" (inptr5), [outptr] "+r" (outptr)
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave4_block16_s8_s8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave4_block16_s8_s8.hpp
index 4d065300ae..d5a41a332d 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave4_block16_s8_s8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave4_block16_s8_s8.hpp
@@ -34,9 +34,9 @@ void interleave_block<4, 16, VLType::None, false>(
       "ldr x23, [%x[in], #0x0]\n"
       "ldr x22, [%x[in], #0x8]\n"
       "cmp %x[height], #0x4\n"
+      "add x23, x23, %x[row_offset]\n"
       "ldr x21, [%x[in], #0x10]\n"
       "ldr x20, [%x[in], #0x18]\n"
-      "add x23, x23, %x[row_offset]\n"
       "add x22, x22, %x[row_offset]\n"
       "add x21, x21, %x[row_offset]\n"
       "add x20, x20, %x[row_offset]\n"
@@ -60,12 +60,12 @@ void interleave_block<4, 16, VLType::None, false>(
       "ldr q19, [x23], #0x10\n"
       "ldr q18, [x22], #0x10\n"
       "subs %x[width], %x[width], #0x10\n"
+      "cmp %x[width], #0x10\n"
       "ldr q17, [x21], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "cmp %x[width], #0x10\n"
+      "str q19, [%x[out_ptr], #0x0]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
       "prfm pldl1keep, [x22, #0x70]\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
       "str q18, [%x[out_ptr], #0x10]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
       "prfm pldl1keep, [x20, #0x70]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp16_fp16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp16_fp16.hpp
index 1cd6523c76..35de179ed4 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp16_fp16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp16_fp16.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 1, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset], LSL #1\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset], LSL #1\n"
       "add x27, x27, %x[row_offset], LSL #1\n"
       "add x26, x26, %x[row_offset], LSL #1\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset], LSL #1\n"
       "add x24, x24, %x[row_offset], LSL #1\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset], LSL #1\n"
       "add x22, x22, %x[row_offset], LSL #1\n"
       "add x21, x21, %x[row_offset], LSL #1\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,52 +79,52 @@ void interleave_block<8, 1, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q23, [x28], #0x10\n"
-      "ldr q25, [x27], #0x10\n"
+      "ldr q25, [x28], #0x10\n"
+      "ldr q27, [x27], #0x10\n"
       "subs %x[width], %x[width], #0x8\n"
-      "ldr q22, [x26], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
       "cmp %x[width], #0x8\n"
-      "ldr q19, [x24], #0x10\n"
-      "ldr q18, [x23], #0x10\n"
+      "ldr q26, [x26], #0x10\n"
+      "ldr q24, [x25], #0x10\n"
+      "ldr q21, [x24], #0x10\n"
+      "ldr q20, [x23], #0x10\n"
+      "zip1 v23.8h, v25.8h, v21.8h\n"
+      "zip1 v22.8h, v27.8h, v20.8h\n"
       "ldr q17, [x22], #0x10\n"
       "ldr q16, [x21], #0x10\n"
+      "zip1 v19.8h, v26.8h, v17.8h\n"
+      "zip1 v18.8h, v24.8h, v16.8h\n"
+      "zip2 v25.8h, v25.8h, v21.8h\n"
+      "zip2 v21.8h, v26.8h, v17.8h\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
-      "zip1 v20.8h, v23.8h, v19.8h\n"
-      "zip1 v24.8h, v25.8h, v18.8h\n"
-      "zip2 v23.8h, v23.8h, v19.8h\n"
-      "zip2 v25.8h, v25.8h, v18.8h\n"
+      "zip2 v20.8h, v27.8h, v20.8h\n"
+      "zip2 v16.8h, v24.8h, v16.8h\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "zip1 v19.8h, v22.8h, v17.8h\n"
-      "zip1 v18.8h, v21.8h, v16.8h\n"
+      "zip1 v24.8h, v23.8h, v19.8h\n"
+      "zip1 v17.8h, v22.8h, v18.8h\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
-      "zip2 v22.8h, v22.8h, v17.8h\n"
-      "zip2 v17.8h, v21.8h, v16.8h\n"
+      "zip2 v23.8h, v23.8h, v19.8h\n"
+      "zip2 v19.8h, v22.8h, v18.8h\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip1 v16.8h, v24.8h, v18.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip2 v19.8h, v24.8h, v18.8h\n"
-      "zip1 v24.8h, v23.8h, v22.8h\n"
-      "zip1 v18.8h, v25.8h, v17.8h\n"
-      "zip2 v23.8h, v23.8h, v22.8h\n"
-      "zip2 v22.8h, v25.8h, v17.8h\n"
-      "zip1 v17.8h, v21.8h, v16.8h\n"
-      "zip2 v16.8h, v21.8h, v16.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip1 v19.8h, v24.8h, v18.8h\n"
-      "zip2 v18.8h, v24.8h, v18.8h\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip1 v22.8h, v25.8h, v21.8h\n"
+      "zip1 v18.8h, v20.8h, v16.8h\n"
+      "zip2 v21.8h, v25.8h, v21.8h\n"
+      "zip2 v20.8h, v20.8h, v16.8h\n"
+      "zip1 v16.8h, v24.8h, v17.8h\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.8h, v24.8h, v17.8h\n"
       "str q16, [%x[out_ptr], #0x10]\n"
-      "zip1 v17.8h, v23.8h, v22.8h\n"
-      "zip2 v16.8h, v23.8h, v22.8h\n"
-      "str q21, [%x[out_ptr], #0x20]\n"
-      "str q20, [%x[out_ptr], #0x30]\n"
+      "zip1 v17.8h, v23.8h, v19.8h\n"
+      "zip2 v16.8h, v23.8h, v19.8h\n"
+      "str q17, [%x[out_ptr], #0x20]\n"
+      "zip1 v19.8h, v22.8h, v18.8h\n"
+      "zip2 v18.8h, v22.8h, v18.8h\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip1 v17.8h, v21.8h, v20.8h\n"
+      "zip2 v16.8h, v21.8h, v20.8h\n"
       "str q19, [%x[out_ptr], #0x40]\n"
       "str q18, [%x[out_ptr], #0x50]\n"
       "str q17, [%x[out_ptr], #0x60]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp16_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp16_fp32.hpp
index 81c758f498..59981e9979 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp16_fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp16_fp32.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 1, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset], LSL #1\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset], LSL #1\n"
       "add x27, x27, %x[row_offset], LSL #1\n"
       "add x26, x26, %x[row_offset], LSL #1\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset], LSL #1\n"
       "add x24, x24, %x[row_offset], LSL #1\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset], LSL #1\n"
       "add x22, x22, %x[row_offset], LSL #1\n"
       "add x21, x21, %x[row_offset], LSL #1\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,53 +79,53 @@ void interleave_block<8, 1, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr d25, [x28], #0x8\n"
-      "ldr d24, [x27], #0x8\n"
+      "ldr d27, [x28], #0x8\n"
+      "ldr d26, [x27], #0x8\n"
+      "fcvtl v27.4s, v27.4h\n"
+      "fcvtl v26.4s, v26.4h\n"
+      "ldr d22, [x26], #0x8\n"
+      "ldr d21, [x25], #0x8\n"
+      "fcvtl v22.4s, v22.4h\n"
+      "fcvtl v21.4s, v21.4h\n"
+      "ldr d20, [x24], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "fcvtl v20.4s, v20.4h\n"
+      "fcvtl v25.4s, v25.4h\n"
+      "ldr d19, [x22], #0x8\n"
+      "ldr d16, [x21], #0x8\n"
+      "fcvtl v19.4s, v19.4h\n"
+      "fcvtl v16.4s, v16.4h\n"
+      "zip1 v24.4s, v27.4s, v22.4s\n"
+      "zip1 v23.4s, v26.4s, v21.4s\n"
       "subs %x[width], %x[width], #0x4\n"
-      "ldr d18, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
       "cmp %x[width], #0x4\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d16, [x21], #0x8\n"
-      "fcvtl v25.4s, v25.4h\n"
-      "fcvtl v24.4s, v24.4h\n"
-      "fcvtl v18.4s, v18.4h\n"
-      "fcvtl v17.4s, v17.4h\n"
+      "zip1 v18.4s, v20.4s, v19.4s\n"
+      "zip1 v17.4s, v25.4s, v16.4s\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
-      "fcvtl v23.4s, v23.4h\n"
-      "fcvtl v22.4s, v22.4h\n"
+      "zip2 v22.4s, v27.4s, v22.4s\n"
+      "zip2 v21.4s, v26.4s, v21.4s\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "fcvtl v21.4s, v21.4h\n"
-      "fcvtl v16.4s, v16.4h\n"
+      "zip2 v20.4s, v20.4s, v19.4s\n"
+      "zip2 v19.4s, v25.4s, v16.4s\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
-      "zip1 v20.4s, v25.4s, v18.4s\n"
-      "zip1 v19.4s, v24.4s, v17.4s\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "zip2 v26.4s, v25.4s, v18.4s\n"
-      "zip2 v25.4s, v24.4s, v17.4s\n"
-      "zip1 v18.4s, v23.4s, v21.4s\n"
-      "zip1 v17.4s, v22.4s, v16.4s\n"
-      "zip2 v24.4s, v23.4s, v21.4s\n"
-      "zip2 v23.4s, v22.4s, v16.4s\n"
-      "zip1 v16.4s, v20.4s, v19.4s\n"
-      "zip2 v22.4s, v20.4s, v19.4s\n"
-      "zip1 v21.4s, v18.4s, v17.4s\n"
-      "zip2 v20.4s, v18.4s, v17.4s\n"
-      "zip1 v19.4s, v26.4s, v25.4s\n"
-      "zip1 v18.4s, v24.4s, v23.4s\n"
+      "zip1 v16.4s, v24.4s, v23.4s\n"
       "str q16, [%x[out_ptr], #0x0]\n"
-      "zip2 v17.4s, v26.4s, v25.4s\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
       "zip2 v16.4s, v24.4s, v23.4s\n"
-      "str q21, [%x[out_ptr], #0x10]\n"
-      "str q22, [%x[out_ptr], #0x20]\n"
-      "str q20, [%x[out_ptr], #0x30]\n"
-      "str q19, [%x[out_ptr], #0x40]\n"
+      "str q16, [%x[out_ptr], #0x20]\n"
+      "zip2 v17.4s, v18.4s, v17.4s\n"
+      "zip1 v16.4s, v22.4s, v21.4s\n"
+      "str q17, [%x[out_ptr], #0x30]\n"
+      "zip1 v18.4s, v20.4s, v19.4s\n"
+      "zip2 v17.4s, v22.4s, v21.4s\n"
+      "str q16, [%x[out_ptr], #0x40]\n"
+      "zip2 v16.4s, v20.4s, v19.4s\n"
       "str q18, [%x[out_ptr], #0x50]\n"
       "str q17, [%x[out_ptr], #0x60]\n"
       "str q16, [%x[out_ptr], #0x70]\n"
@@ -134,76 +134,76 @@ void interleave_block<8, 1, VLType::None, false>(
       "3:"  // Main loop skip
       "cbz %x[width], 6f\n"
       "tbz %x[width], #1, 4f\n"
-      "ldr s29, [x28], #0x4\n"
-      "ldr s28, [x27], #0x4\n"
+      "ldr s28, [x28], #0x4\n"
+      "ldr s27, [x27], #0x4\n"
       "mov x20, #0x2\n"
-      "ldr s27, [x26], #0x4\n"
-      "ldr s26, [x25], #0x4\n"
-      "ldr s25, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s23, [x22], #0x4\n"
-      "ldr s22, [x21], #0x4\n"
+      "ldr s26, [x26], #0x4\n"
+      "ldr s25, [x25], #0x4\n"
+      "ldr s24, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s22, [x22], #0x4\n"
+      "ldr s21, [x21], #0x4\n"
       "tbz %x[width], #0, 5f\n"
-      "ld1 { v29.h }[2], [x28]\n"
-      "ld1 { v28.h }[2], [x27]\n"
+      "ld1 { v28.h }[2], [x28]\n"
+      "ld1 { v27.h }[2], [x27]\n"
       "mov x20, #0x3\n"
-      "ld1 { v27.h }[2], [x26]\n"
-      "ld1 { v26.h }[2], [x25]\n"
-      "ld1 { v25.h }[2], [x24]\n"
-      "ld1 { v24.h }[2], [x23]\n"
-      "ld1 { v23.h }[2], [x22]\n"
-      "ld1 { v22.h }[2], [x21]\n"
+      "ld1 { v26.h }[2], [x26]\n"
+      "ld1 { v25.h }[2], [x25]\n"
+      "ld1 { v24.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
+      "ld1 { v22.h }[2], [x22]\n"
+      "ld1 { v21.h }[2], [x21]\n"
       "b 5f\n"
       "4:"  // odd_loads_1_0
-      "ldr h29, [x28, #0x0]\n"
-      "ldr h28, [x27, #0x0]\n"
+      "ldr h28, [x28, #0x0]\n"
+      "ldr h27, [x27, #0x0]\n"
       "mov x20, #0x1\n"
-      "ldr h27, [x26, #0x0]\n"
-      "ldr h26, [x25, #0x0]\n"
-      "ldr h25, [x24, #0x0]\n"
-      "ldr h24, [x23, #0x0]\n"
-      "ldr h23, [x22, #0x0]\n"
-      "ldr h22, [x21, #0x0]\n"
+      "ldr h26, [x26, #0x0]\n"
+      "ldr h25, [x25, #0x0]\n"
+      "ldr h24, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
+      "ldr h22, [x22, #0x0]\n"
+      "ldr h21, [x21, #0x0]\n"
       "5:"  // Odd load end
-      "fcvtl v29.4s, v29.4h\n"
       "fcvtl v28.4s, v28.4h\n"
-      "subs x20, x20, #0x1\n"
       "fcvtl v27.4s, v27.4h\n"
+      "subs x20, x20, #0x1\n"
       "fcvtl v26.4s, v26.4h\n"
       "fcvtl v25.4s, v25.4h\n"
       "fcvtl v24.4s, v24.4h\n"
       "fcvtl v23.4s, v23.4h\n"
       "fcvtl v22.4s, v22.4h\n"
-      "zip1 v21.4s, v29.4s, v27.4s\n"
+      "fcvtl v21.4s, v21.4h\n"
       "zip1 v20.4s, v28.4s, v26.4s\n"
-      "zip1 v19.4s, v25.4s, v23.4s\n"
+      "zip1 v19.4s, v27.4s, v25.4s\n"
       "zip1 v18.4s, v24.4s, v22.4s\n"
-      "zip1 v17.4s, v21.4s, v20.4s\n"
-      "zip1 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip1 v17.4s, v23.4s, v21.4s\n"
+      "zip1 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 6f\n"
       "subs x20, x20, #0x1\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      "zip2 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 6f\n"
-      "zip2 v19.4s, v29.4s, v27.4s\n"
-      "zip2 v17.4s, v28.4s, v26.4s\n"
-      "zip2 v18.4s, v25.4s, v23.4s\n"
-      "zip2 v16.4s, v24.4s, v22.4s\n"
-      "zip1 v17.4s, v19.4s, v17.4s\n"
-      "zip1 v16.4s, v18.4s, v16.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v19.4s, v28.4s, v26.4s\n"
+      "zip2 v16.4s, v27.4s, v25.4s\n"
+      "zip2 v18.4s, v24.4s, v22.4s\n"
+      "zip2 v17.4s, v23.4s, v21.4s\n"
+      "zip1 v16.4s, v19.4s, v16.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "6:"  // Odds skip
       : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
-      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp
index 521b3bf8f5..9eeabfa9eb 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 1, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset], LSL #2\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset], LSL #2\n"
       "add x27, x27, %x[row_offset], LSL #2\n"
       "add x26, x26, %x[row_offset], LSL #2\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset], LSL #2\n"
       "add x24, x24, %x[row_offset], LSL #2\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset], LSL #2\n"
       "add x22, x22, %x[row_offset], LSL #2\n"
       "add x21, x21, %x[row_offset], LSL #2\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,46 +79,46 @@ void interleave_block<8, 1, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q25, [x28], #0x10\n"
-      "ldr q24, [x27], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "ldr q18, [x27], #0x10\n"
       "subs %x[width], %x[width], #0x4\n"
-      "ldr q18, [x26], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
       "cmp %x[width], #0x4\n"
-      "ldr q23, [x24], #0x10\n"
-      "ldr q22, [x23], #0x10\n"
-      "ldr q21, [x22], #0x10\n"
+      "ldr q17, [x26], #0x10\n"
+      "ldr q16, [x25], #0x10\n"
+      "zip1 v25.4s, v20.4s, v17.4s\n"
+      "zip1 v24.4s, v18.4s, v16.4s\n"
+      "ldr q19, [x24], #0x10\n"
+      "ldr q23, [x23], #0x10\n"
+      "zip2 v22.4s, v20.4s, v17.4s\n"
+      "zip2 v21.4s, v18.4s, v16.4s\n"
+      "ldr q18, [x22], #0x10\n"
       "ldr q16, [x21], #0x10\n"
-      "zip1 v20.4s, v25.4s, v18.4s\n"
-      "zip1 v19.4s, v24.4s, v17.4s\n"
-      "zip2 v26.4s, v25.4s, v18.4s\n"
-      "zip2 v25.4s, v24.4s, v17.4s\n"
+      "zip1 v20.4s, v19.4s, v18.4s\n"
+      "zip1 v17.4s, v23.4s, v16.4s\n"
+      "zip2 v19.4s, v19.4s, v18.4s\n"
+      "zip2 v18.4s, v23.4s, v16.4s\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "zip1 v18.4s, v23.4s, v21.4s\n"
-      "zip1 v17.4s, v22.4s, v16.4s\n"
-      "zip2 v24.4s, v23.4s, v21.4s\n"
-      "zip2 v23.4s, v22.4s, v16.4s\n"
+      "zip1 v16.4s, v25.4s, v24.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
+      "zip1 v16.4s, v20.4s, v17.4s\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "zip1 v16.4s, v20.4s, v19.4s\n"
-      "zip2 v22.4s, v20.4s, v19.4s\n"
-      "zip1 v21.4s, v18.4s, v17.4s\n"
-      "zip2 v20.4s, v18.4s, v17.4s\n"
-      "zip1 v19.4s, v26.4s, v25.4s\n"
-      "zip1 v18.4s, v24.4s, v23.4s\n"
-      "str q16, [%x[out_ptr], #0x0]\n"
-      "zip2 v17.4s, v26.4s, v25.4s\n"
-      "zip2 v16.4s, v24.4s, v23.4s\n"
-      "str q21, [%x[out_ptr], #0x10]\n"
-      "str q22, [%x[out_ptr], #0x20]\n"
-      "str q20, [%x[out_ptr], #0x30]\n"
-      "str q19, [%x[out_ptr], #0x40]\n"
-      "str q18, [%x[out_ptr], #0x50]\n"
+      "zip2 v16.4s, v25.4s, v24.4s\n"
+      "str q16, [%x[out_ptr], #0x20]\n"
+      "zip2 v16.4s, v20.4s, v17.4s\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip1 v16.4s, v22.4s, v21.4s\n"
+      "str q16, [%x[out_ptr], #0x40]\n"
+      "zip1 v16.4s, v19.4s, v18.4s\n"
+      "zip2 v17.4s, v22.4s, v21.4s\n"
+      "str q16, [%x[out_ptr], #0x50]\n"
+      "zip2 v16.4s, v19.4s, v18.4s\n"
       "str q17, [%x[out_ptr], #0x60]\n"
       "str q16, [%x[out_ptr], #0x70]\n"
       "add %x[out_ptr], %x[out_ptr], #0x80\n"
@@ -126,68 +126,68 @@ void interleave_block<8, 1, VLType::None, false>(
       "3:"  // Main loop skip
       "cbz %x[width], 6f\n"
       "tbz %x[width], #1, 4f\n"
-      "ldr d29, [x28], #0x8\n"
-      "ldr d28, [x27], #0x8\n"
+      "ldr d28, [x28], #0x8\n"
+      "ldr d27, [x27], #0x8\n"
       "mov x20, #0x2\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d26, [x25], #0x8\n"
-      "ldr d25, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "ldr d22, [x21], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
+      "ldr d21, [x21], #0x8\n"
       "tbz %x[width], #0, 5f\n"
-      "ld1 { v29.s }[2], [x28]\n"
-      "ld1 { v28.s }[2], [x27]\n"
+      "ld1 { v28.s }[2], [x28]\n"
+      "ld1 { v27.s }[2], [x27]\n"
       "mov x20, #0x3\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v26.s }[2], [x25]\n"
-      "ld1 { v25.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v23.s }[2], [x22]\n"
-      "ld1 { v22.s }[2], [x21]\n"
+      "ld1 { v26.s }[2], [x26]\n"
+      "ld1 { v25.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
+      "ld1 { v21.s }[2], [x21]\n"
       "b 5f\n"
       "4:"  // odd_loads_1_0
-      "ldr s29, [x28, #0x0]\n"
-      "ldr s28, [x27, #0x0]\n"
+      "ldr s28, [x28, #0x0]\n"
+      "ldr s27, [x27, #0x0]\n"
       "mov x20, #0x1\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s26, [x25, #0x0]\n"
-      "ldr s25, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s23, [x22, #0x0]\n"
-      "ldr s22, [x21, #0x0]\n"
+      "ldr s26, [x26, #0x0]\n"
+      "ldr s25, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
+      "ldr s21, [x21, #0x0]\n"
       "5:"  // Odd load end
-      "zip1 v21.4s, v29.4s, v27.4s\n"
       "zip1 v20.4s, v28.4s, v26.4s\n"
+      "zip1 v19.4s, v27.4s, v25.4s\n"
       "subs x20, x20, #0x1\n"
-      "zip1 v19.4s, v25.4s, v23.4s\n"
       "zip1 v18.4s, v24.4s, v22.4s\n"
-      "zip1 v17.4s, v21.4s, v20.4s\n"
-      "zip1 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip1 v17.4s, v23.4s, v21.4s\n"
+      "zip1 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 6f\n"
       "subs x20, x20, #0x1\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      "zip2 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 6f\n"
-      "zip2 v19.4s, v29.4s, v27.4s\n"
-      "zip2 v17.4s, v28.4s, v26.4s\n"
-      "zip2 v18.4s, v25.4s, v23.4s\n"
-      "zip2 v16.4s, v24.4s, v22.4s\n"
-      "zip1 v17.4s, v19.4s, v17.4s\n"
-      "zip1 v16.4s, v18.4s, v16.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v19.4s, v28.4s, v26.4s\n"
+      "zip2 v16.4s, v27.4s, v25.4s\n"
+      "zip2 v18.4s, v24.4s, v22.4s\n"
+      "zip2 v17.4s, v23.4s, v21.4s\n"
+      "zip1 v16.4s, v19.4s, v16.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "6:"  // Odds skip
       : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
-      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_s16_s16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_s16_s16.hpp
index 8f67a21d05..27b3335694 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_s16_s16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_s16_s16.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 1, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset], LSL #1\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset], LSL #1\n"
       "add x27, x27, %x[row_offset], LSL #1\n"
       "add x26, x26, %x[row_offset], LSL #1\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset], LSL #1\n"
       "add x24, x24, %x[row_offset], LSL #1\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset], LSL #1\n"
       "add x22, x22, %x[row_offset], LSL #1\n"
       "add x21, x21, %x[row_offset], LSL #1\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,52 +79,52 @@ void interleave_block<8, 1, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q23, [x28], #0x10\n"
-      "ldr q25, [x27], #0x10\n"
+      "ldr q25, [x28], #0x10\n"
+      "ldr q27, [x27], #0x10\n"
       "subs %x[width], %x[width], #0x8\n"
-      "ldr q22, [x26], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
       "cmp %x[width], #0x8\n"
-      "ldr q19, [x24], #0x10\n"
-      "ldr q18, [x23], #0x10\n"
+      "ldr q26, [x26], #0x10\n"
+      "ldr q24, [x25], #0x10\n"
+      "ldr q21, [x24], #0x10\n"
+      "ldr q20, [x23], #0x10\n"
+      "zip1 v23.8h, v25.8h, v21.8h\n"
+      "zip1 v22.8h, v27.8h, v20.8h\n"
       "ldr q17, [x22], #0x10\n"
       "ldr q16, [x21], #0x10\n"
+      "zip1 v19.8h, v26.8h, v17.8h\n"
+      "zip1 v18.8h, v24.8h, v16.8h\n"
+      "zip2 v25.8h, v25.8h, v21.8h\n"
+      "zip2 v21.8h, v26.8h, v17.8h\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
-      "zip1 v20.8h, v23.8h, v19.8h\n"
-      "zip1 v24.8h, v25.8h, v18.8h\n"
-      "zip2 v23.8h, v23.8h, v19.8h\n"
-      "zip2 v25.8h, v25.8h, v18.8h\n"
+      "zip2 v20.8h, v27.8h, v20.8h\n"
+      "zip2 v16.8h, v24.8h, v16.8h\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "zip1 v19.8h, v22.8h, v17.8h\n"
-      "zip1 v18.8h, v21.8h, v16.8h\n"
+      "zip1 v24.8h, v23.8h, v19.8h\n"
+      "zip1 v17.8h, v22.8h, v18.8h\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
-      "zip2 v22.8h, v22.8h, v17.8h\n"
-      "zip2 v17.8h, v21.8h, v16.8h\n"
+      "zip2 v23.8h, v23.8h, v19.8h\n"
+      "zip2 v19.8h, v22.8h, v18.8h\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip1 v16.8h, v24.8h, v18.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip2 v19.8h, v24.8h, v18.8h\n"
-      "zip1 v24.8h, v23.8h, v22.8h\n"
-      "zip1 v18.8h, v25.8h, v17.8h\n"
-      "zip2 v23.8h, v23.8h, v22.8h\n"
-      "zip2 v22.8h, v25.8h, v17.8h\n"
-      "zip1 v17.8h, v21.8h, v16.8h\n"
-      "zip2 v16.8h, v21.8h, v16.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip1 v19.8h, v24.8h, v18.8h\n"
-      "zip2 v18.8h, v24.8h, v18.8h\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip1 v22.8h, v25.8h, v21.8h\n"
+      "zip1 v18.8h, v20.8h, v16.8h\n"
+      "zip2 v21.8h, v25.8h, v21.8h\n"
+      "zip2 v20.8h, v20.8h, v16.8h\n"
+      "zip1 v16.8h, v24.8h, v17.8h\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.8h, v24.8h, v17.8h\n"
       "str q16, [%x[out_ptr], #0x10]\n"
-      "zip1 v17.8h, v23.8h, v22.8h\n"
-      "zip2 v16.8h, v23.8h, v22.8h\n"
-      "str q21, [%x[out_ptr], #0x20]\n"
-      "str q20, [%x[out_ptr], #0x30]\n"
+      "zip1 v17.8h, v23.8h, v19.8h\n"
+      "zip2 v16.8h, v23.8h, v19.8h\n"
+      "str q17, [%x[out_ptr], #0x20]\n"
+      "zip1 v19.8h, v22.8h, v18.8h\n"
+      "zip2 v18.8h, v22.8h, v18.8h\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip1 v17.8h, v21.8h, v20.8h\n"
+      "zip2 v16.8h, v21.8h, v20.8h\n"
       "str q19, [%x[out_ptr], #0x40]\n"
       "str q18, [%x[out_ptr], #0x50]\n"
       "str q17, [%x[out_ptr], #0x60]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_s8_s16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_s8_s16.hpp
index 0a76fa812e..4c90691edc 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_s8_s16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_s8_s16.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 1, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset]\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset]\n"
       "add x27, x27, %x[row_offset]\n"
       "add x26, x26, %x[row_offset]\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset]\n"
       "add x24, x24, %x[row_offset]\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset]\n"
       "add x22, x22, %x[row_offset]\n"
       "add x21, x21, %x[row_offset]\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,60 +79,60 @@ void interleave_block<8, 1, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr d23, [x28], #0x8\n"
-      "ldr d25, [x27], #0x8\n"
-      "subs %x[width], %x[width], #0x8\n"
-      "ldr d22, [x26], #0x8\n"
-      "ldr d21, [x25], #0x8\n"
-      "cmp %x[width], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d17, [x22], #0x8\n"
-      "ldr d16, [x21], #0x8\n"
-      "sshll v23.8h, v23.8b, #0x0\n"
+      "ldr d25, [x28], #0x8\n"
+      "ldr d27, [x27], #0x8\n"
       "sshll v25.8h, v25.8b, #0x0\n"
-      "sshll v22.8h, v22.8b, #0x0\n"
+      "sshll v27.8h, v27.8b, #0x0\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d24, [x25], #0x8\n"
+      "sshll v26.8h, v26.8b, #0x0\n"
+      "sshll v24.8h, v24.8b, #0x0\n"
+      "ldr d21, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "sshll v21.8h, v21.8b, #0x0\n"
+      "sshll v20.8h, v20.8b, #0x0\n"
+      "ldr d17, [x22], #0x8\n"
+      "ldr d16, [x21], #0x8\n"
+      "sshll v17.8h, v17.8b, #0x0\n"
+      "sshll v16.8h, v16.8b, #0x0\n"
+      "zip1 v23.8h, v25.8h, v21.8h\n"
+      "zip1 v22.8h, v26.8h, v17.8h\n"
+      "subs %x[width], %x[width], #0x8\n"
+      "cmp %x[width], #0x8\n"
+      "zip1 v19.8h, v27.8h, v20.8h\n"
+      "zip1 v18.8h, v24.8h, v16.8h\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
-      "sshll v19.8h, v19.8b, #0x0\n"
-      "sshll v18.8h, v18.8b, #0x0\n"
+      "zip2 v25.8h, v25.8h, v21.8h\n"
+      "zip2 v21.8h, v26.8h, v17.8h\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "sshll v17.8h, v17.8b, #0x0\n"
-      "sshll v16.8h, v16.8b, #0x0\n"
+      "zip2 v20.8h, v27.8h, v20.8h\n"
+      "zip2 v16.8h, v24.8h, v16.8h\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
+      "zip1 v24.8h, v23.8h, v22.8h\n"
+      "zip1 v17.8h, v19.8h, v18.8h\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "zip1 v20.8h, v23.8h, v19.8h\n"
-      "zip1 v24.8h, v25.8h, v18.8h\n"
-      "zip2 v23.8h, v23.8h, v19.8h\n"
-      "zip2 v25.8h, v25.8h, v18.8h\n"
-      "zip1 v19.8h, v22.8h, v17.8h\n"
-      "zip1 v18.8h, v21.8h, v16.8h\n"
-      "zip2 v22.8h, v22.8h, v17.8h\n"
-      "zip2 v17.8h, v21.8h, v16.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip1 v16.8h, v24.8h, v18.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip2 v19.8h, v24.8h, v18.8h\n"
-      "zip1 v24.8h, v23.8h, v22.8h\n"
-      "zip1 v18.8h, v25.8h, v17.8h\n"
       "zip2 v23.8h, v23.8h, v22.8h\n"
-      "zip2 v22.8h, v25.8h, v17.8h\n"
-      "zip1 v17.8h, v21.8h, v16.8h\n"
-      "zip2 v16.8h, v21.8h, v16.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip1 v19.8h, v24.8h, v18.8h\n"
-      "zip2 v18.8h, v24.8h, v18.8h\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v19.8h, v19.8h, v18.8h\n"
+      "zip1 v22.8h, v25.8h, v21.8h\n"
+      "zip1 v18.8h, v20.8h, v16.8h\n"
+      "zip2 v21.8h, v25.8h, v21.8h\n"
+      "zip2 v20.8h, v20.8h, v16.8h\n"
+      "zip1 v16.8h, v24.8h, v17.8h\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.8h, v24.8h, v17.8h\n"
       "str q16, [%x[out_ptr], #0x10]\n"
-      "zip1 v17.8h, v23.8h, v22.8h\n"
-      "zip2 v16.8h, v23.8h, v22.8h\n"
-      "str q21, [%x[out_ptr], #0x20]\n"
-      "str q20, [%x[out_ptr], #0x30]\n"
+      "zip1 v17.8h, v23.8h, v19.8h\n"
+      "zip2 v16.8h, v23.8h, v19.8h\n"
+      "str q17, [%x[out_ptr], #0x20]\n"
+      "zip1 v19.8h, v22.8h, v18.8h\n"
+      "zip2 v18.8h, v22.8h, v18.8h\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip1 v17.8h, v21.8h, v20.8h\n"
+      "zip2 v16.8h, v21.8h, v20.8h\n"
       "str q19, [%x[out_ptr], #0x40]\n"
       "str q18, [%x[out_ptr], #0x50]\n"
       "str q17, [%x[out_ptr], #0x60]\n"
@@ -227,11 +227,11 @@ void interleave_block<8, 1, VLType::None, false>(
       "sshll v24.8h, v24.8b, #0x0\n"
       "sshll v23.8h, v23.8b, #0x0\n"
       "zip1 v22.8h, v30.8h, v26.8h\n"
-      "zip1 v21.8h, v29.8h, v25.8h\n"
-      "zip1 v20.8h, v28.8h, v24.8h\n"
+      "zip1 v21.8h, v28.8h, v24.8h\n"
+      "zip1 v20.8h, v29.8h, v25.8h\n"
       "zip1 v19.8h, v27.8h, v23.8h\n"
-      "zip1 v18.8h, v22.8h, v20.8h\n"
-      "zip1 v17.8h, v21.8h, v19.8h\n"
+      "zip1 v18.8h, v22.8h, v21.8h\n"
+      "zip1 v17.8h, v20.8h, v19.8h\n"
       "zip1 v16.8h, v18.8h, v17.8h\n"
       "str q16, [%x[out_ptr], #0x0]\n"
       "add %x[out_ptr], %x[out_ptr], #0x10\n"
@@ -241,8 +241,8 @@ void interleave_block<8, 1, VLType::None, false>(
       "str q16, [%x[out_ptr], #0x0]\n"
       "add %x[out_ptr], %x[out_ptr], #0x10\n"
       "beq 8f\n"
-      "zip2 v18.8h, v22.8h, v20.8h\n"
-      "zip2 v17.8h, v21.8h, v19.8h\n"
+      "zip2 v18.8h, v22.8h, v21.8h\n"
+      "zip2 v17.8h, v20.8h, v19.8h\n"
       "subs x20, x20, #0x1\n"
       "zip1 v16.8h, v18.8h, v17.8h\n"
       "str q16, [%x[out_ptr], #0x0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_u8_u16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_u8_u16.hpp
index be6e8980f6..8901908140 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_u8_u16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_u8_u16.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 1, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset]\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset]\n"
       "add x27, x27, %x[row_offset]\n"
       "add x26, x26, %x[row_offset]\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset]\n"
       "add x24, x24, %x[row_offset]\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset]\n"
       "add x22, x22, %x[row_offset]\n"
       "add x21, x21, %x[row_offset]\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,60 +79,60 @@ void interleave_block<8, 1, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr d23, [x28], #0x8\n"
-      "ldr d25, [x27], #0x8\n"
-      "subs %x[width], %x[width], #0x8\n"
-      "ldr d22, [x26], #0x8\n"
-      "ldr d21, [x25], #0x8\n"
-      "cmp %x[width], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d17, [x22], #0x8\n"
-      "ldr d16, [x21], #0x8\n"
-      "ushll v23.8h, v23.8b, #0x0\n"
+      "ldr d25, [x28], #0x8\n"
+      "ldr d27, [x27], #0x8\n"
       "ushll v25.8h, v25.8b, #0x0\n"
-      "ushll v22.8h, v22.8b, #0x0\n"
+      "ushll v27.8h, v27.8b, #0x0\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d24, [x25], #0x8\n"
+      "ushll v26.8h, v26.8b, #0x0\n"
+      "ushll v24.8h, v24.8b, #0x0\n"
+      "ldr d21, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "ushll v21.8h, v21.8b, #0x0\n"
+      "ushll v20.8h, v20.8b, #0x0\n"
+      "ldr d17, [x22], #0x8\n"
+      "ldr d16, [x21], #0x8\n"
+      "ushll v17.8h, v17.8b, #0x0\n"
+      "ushll v16.8h, v16.8b, #0x0\n"
+      "zip1 v23.8h, v25.8h, v21.8h\n"
+      "zip1 v22.8h, v26.8h, v17.8h\n"
+      "subs %x[width], %x[width], #0x8\n"
+      "cmp %x[width], #0x8\n"
+      "zip1 v19.8h, v27.8h, v20.8h\n"
+      "zip1 v18.8h, v24.8h, v16.8h\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
-      "ushll v19.8h, v19.8b, #0x0\n"
-      "ushll v18.8h, v18.8b, #0x0\n"
+      "zip2 v25.8h, v25.8h, v21.8h\n"
+      "zip2 v21.8h, v26.8h, v17.8h\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "ushll v17.8h, v17.8b, #0x0\n"
-      "ushll v16.8h, v16.8b, #0x0\n"
+      "zip2 v20.8h, v27.8h, v20.8h\n"
+      "zip2 v16.8h, v24.8h, v16.8h\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
+      "zip1 v24.8h, v23.8h, v22.8h\n"
+      "zip1 v17.8h, v19.8h, v18.8h\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "zip1 v20.8h, v23.8h, v19.8h\n"
-      "zip1 v24.8h, v25.8h, v18.8h\n"
-      "zip2 v23.8h, v23.8h, v19.8h\n"
-      "zip2 v25.8h, v25.8h, v18.8h\n"
-      "zip1 v19.8h, v22.8h, v17.8h\n"
-      "zip1 v18.8h, v21.8h, v16.8h\n"
-      "zip2 v22.8h, v22.8h, v17.8h\n"
-      "zip2 v17.8h, v21.8h, v16.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip1 v16.8h, v24.8h, v18.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip2 v19.8h, v24.8h, v18.8h\n"
-      "zip1 v24.8h, v23.8h, v22.8h\n"
-      "zip1 v18.8h, v25.8h, v17.8h\n"
       "zip2 v23.8h, v23.8h, v22.8h\n"
-      "zip2 v22.8h, v25.8h, v17.8h\n"
-      "zip1 v17.8h, v21.8h, v16.8h\n"
-      "zip2 v16.8h, v21.8h, v16.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip1 v19.8h, v24.8h, v18.8h\n"
-      "zip2 v18.8h, v24.8h, v18.8h\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v19.8h, v19.8h, v18.8h\n"
+      "zip1 v22.8h, v25.8h, v21.8h\n"
+      "zip1 v18.8h, v20.8h, v16.8h\n"
+      "zip2 v21.8h, v25.8h, v21.8h\n"
+      "zip2 v20.8h, v20.8h, v16.8h\n"
+      "zip1 v16.8h, v24.8h, v17.8h\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.8h, v24.8h, v17.8h\n"
       "str q16, [%x[out_ptr], #0x10]\n"
-      "zip1 v17.8h, v23.8h, v22.8h\n"
-      "zip2 v16.8h, v23.8h, v22.8h\n"
-      "str q21, [%x[out_ptr], #0x20]\n"
-      "str q20, [%x[out_ptr], #0x30]\n"
+      "zip1 v17.8h, v23.8h, v19.8h\n"
+      "zip2 v16.8h, v23.8h, v19.8h\n"
+      "str q17, [%x[out_ptr], #0x20]\n"
+      "zip1 v19.8h, v22.8h, v18.8h\n"
+      "zip2 v18.8h, v22.8h, v18.8h\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip1 v17.8h, v21.8h, v20.8h\n"
+      "zip2 v16.8h, v21.8h, v20.8h\n"
       "str q19, [%x[out_ptr], #0x40]\n"
       "str q18, [%x[out_ptr], #0x50]\n"
       "str q17, [%x[out_ptr], #0x60]\n"
@@ -227,11 +227,11 @@ void interleave_block<8, 1, VLType::None, false>(
       "ushll v24.8h, v24.8b, #0x0\n"
       "ushll v23.8h, v23.8b, #0x0\n"
       "zip1 v22.8h, v30.8h, v26.8h\n"
-      "zip1 v21.8h, v29.8h, v25.8h\n"
-      "zip1 v20.8h, v28.8h, v24.8h\n"
+      "zip1 v21.8h, v28.8h, v24.8h\n"
+      "zip1 v20.8h, v29.8h, v25.8h\n"
       "zip1 v19.8h, v27.8h, v23.8h\n"
-      "zip1 v18.8h, v22.8h, v20.8h\n"
-      "zip1 v17.8h, v21.8h, v19.8h\n"
+      "zip1 v18.8h, v22.8h, v21.8h\n"
+      "zip1 v17.8h, v20.8h, v19.8h\n"
       "zip1 v16.8h, v18.8h, v17.8h\n"
       "str q16, [%x[out_ptr], #0x0]\n"
       "add %x[out_ptr], %x[out_ptr], #0x10\n"
@@ -241,8 +241,8 @@ void interleave_block<8, 1, VLType::None, false>(
       "str q16, [%x[out_ptr], #0x0]\n"
       "add %x[out_ptr], %x[out_ptr], #0x10\n"
       "beq 8f\n"
-      "zip2 v18.8h, v22.8h, v20.8h\n"
-      "zip2 v17.8h, v21.8h, v19.8h\n"
+      "zip2 v18.8h, v22.8h, v21.8h\n"
+      "zip2 v17.8h, v20.8h, v19.8h\n"
       "subs x20, x20, #0x1\n"
       "zip1 v16.8h, v18.8h, v17.8h\n"
       "str q16, [%x[out_ptr], #0x0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_bf16_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_bf16_bf16.hpp
index f034b2b45c..db610427a8 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_bf16_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_bf16_bf16.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 2, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset], LSL #1\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset], LSL #1\n"
       "add x27, x27, %x[row_offset], LSL #1\n"
       "add x26, x26, %x[row_offset], LSL #1\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset], LSL #1\n"
       "add x24, x24, %x[row_offset], LSL #1\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset], LSL #1\n"
       "add x22, x22, %x[row_offset], LSL #1\n"
       "add x21, x21, %x[row_offset], LSL #1\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,46 +79,46 @@ void interleave_block<8, 2, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q25, [x28], #0x10\n"
-      "ldr q24, [x27], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "ldr q18, [x27], #0x10\n"
       "subs %x[width], %x[width], #0x8\n"
-      "ldr q18, [x26], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
       "cmp %x[width], #0x8\n"
-      "ldr q23, [x24], #0x10\n"
-      "ldr q22, [x23], #0x10\n"
-      "ldr q21, [x22], #0x10\n"
+      "ldr q17, [x26], #0x10\n"
+      "ldr q16, [x25], #0x10\n"
+      "zip1 v25.4s, v20.4s, v17.4s\n"
+      "zip1 v24.4s, v18.4s, v16.4s\n"
+      "ldr q19, [x24], #0x10\n"
+      "ldr q23, [x23], #0x10\n"
+      "zip2 v22.4s, v20.4s, v17.4s\n"
+      "zip2 v21.4s, v18.4s, v16.4s\n"
+      "ldr q18, [x22], #0x10\n"
       "ldr q16, [x21], #0x10\n"
-      "zip1 v20.4s, v25.4s, v18.4s\n"
-      "zip1 v19.4s, v24.4s, v17.4s\n"
-      "zip2 v26.4s, v25.4s, v18.4s\n"
-      "zip2 v25.4s, v24.4s, v17.4s\n"
+      "zip1 v20.4s, v19.4s, v18.4s\n"
+      "zip1 v17.4s, v23.4s, v16.4s\n"
+      "zip2 v19.4s, v19.4s, v18.4s\n"
+      "zip2 v18.4s, v23.4s, v16.4s\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "zip1 v18.4s, v23.4s, v21.4s\n"
-      "zip1 v17.4s, v22.4s, v16.4s\n"
-      "zip2 v24.4s, v23.4s, v21.4s\n"
-      "zip2 v23.4s, v22.4s, v16.4s\n"
+      "zip1 v16.4s, v25.4s, v24.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
+      "zip1 v16.4s, v20.4s, v17.4s\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "zip1 v16.4s, v20.4s, v19.4s\n"
-      "zip2 v22.4s, v20.4s, v19.4s\n"
-      "zip1 v21.4s, v18.4s, v17.4s\n"
-      "zip2 v20.4s, v18.4s, v17.4s\n"
-      "zip1 v19.4s, v26.4s, v25.4s\n"
-      "zip1 v18.4s, v24.4s, v23.4s\n"
-      "str q16, [%x[out_ptr], #0x0]\n"
-      "zip2 v17.4s, v26.4s, v25.4s\n"
-      "zip2 v16.4s, v24.4s, v23.4s\n"
-      "str q21, [%x[out_ptr], #0x10]\n"
-      "str q22, [%x[out_ptr], #0x20]\n"
-      "str q20, [%x[out_ptr], #0x30]\n"
-      "str q19, [%x[out_ptr], #0x40]\n"
-      "str q18, [%x[out_ptr], #0x50]\n"
+      "zip2 v16.4s, v25.4s, v24.4s\n"
+      "str q16, [%x[out_ptr], #0x20]\n"
+      "zip2 v16.4s, v20.4s, v17.4s\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip1 v16.4s, v22.4s, v21.4s\n"
+      "str q16, [%x[out_ptr], #0x40]\n"
+      "zip1 v16.4s, v19.4s, v18.4s\n"
+      "zip2 v17.4s, v22.4s, v21.4s\n"
+      "str q16, [%x[out_ptr], #0x50]\n"
+      "zip2 v16.4s, v19.4s, v18.4s\n"
       "str q17, [%x[out_ptr], #0x60]\n"
       "str q16, [%x[out_ptr], #0x70]\n"
       "add %x[out_ptr], %x[out_ptr], #0x80\n"
@@ -126,119 +126,119 @@ void interleave_block<8, 2, VLType::None, false>(
       "3:"  // Main loop skip
       "cbz %x[width], 8f\n"
       "tbz %x[width], #2, 5f\n"
-      "ldr d29, [x28], #0x8\n"
-      "ldr d28, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d26, [x25], #0x8\n"
-      "ldr d25, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "ldr d22, [x21], #0x8\n"
+      "ldr d28, [x28], #0x8\n"
+      "ldr d27, [x27], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
+      "ldr d21, [x21], #0x8\n"
       "tbz %x[width], #1, 4f\n"
-      "ld1 { v29.s }[2], [x28], #0x4\n"
-      "ld1 { v28.s }[2], [x27], #0x4\n"
+      "ld1 { v28.s }[2], [x28], #0x4\n"
+      "ld1 { v27.s }[2], [x27], #0x4\n"
       "mov x20, #0x3\n"
-      "ld1 { v27.s }[2], [x26], #0x4\n"
-      "ld1 { v26.s }[2], [x25], #0x4\n"
-      "ld1 { v25.s }[2], [x24], #0x4\n"
-      "ld1 { v24.s }[2], [x23], #0x4\n"
-      "ld1 { v23.s }[2], [x22], #0x4\n"
-      "ld1 { v22.s }[2], [x21], #0x4\n"
+      "ld1 { v26.s }[2], [x26], #0x4\n"
+      "ld1 { v25.s }[2], [x25], #0x4\n"
+      "ld1 { v24.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
+      "ld1 { v22.s }[2], [x22], #0x4\n"
+      "ld1 { v21.s }[2], [x21], #0x4\n"
       "tbz %x[width], #0, 7f\n"
-      "ld1 { v29.h }[6], [x28]\n"
-      "ld1 { v28.h }[6], [x27]\n"
+      "ld1 { v28.h }[6], [x28]\n"
+      "ld1 { v27.h }[6], [x27]\n"
       "mov x20, #0x4\n"
-      "ld1 { v27.h }[6], [x26]\n"
-      "ld1 { v26.h }[6], [x25]\n"
-      "ld1 { v25.h }[6], [x24]\n"
-      "ld1 { v24.h }[6], [x23]\n"
-      "ld1 { v23.h }[6], [x22]\n"
-      "ld1 { v22.h }[6], [x21]\n"
+      "ld1 { v26.h }[6], [x26]\n"
+      "ld1 { v25.h }[6], [x25]\n"
+      "ld1 { v24.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
+      "ld1 { v22.h }[6], [x22]\n"
+      "ld1 { v21.h }[6], [x21]\n"
       "b 7f\n"
       "4:"  // odd_loads_1_4
       "mov x20, #0x2\n"
       "tbz %x[width], #0, 7f\n"
-      "ld1 { v29.h }[4], [x28]\n"
-      "ld1 { v28.h }[4], [x27]\n"
+      "ld1 { v28.h }[4], [x28]\n"
+      "ld1 { v27.h }[4], [x27]\n"
       "mov x20, #0x3\n"
-      "ld1 { v27.h }[4], [x26]\n"
-      "ld1 { v26.h }[4], [x25]\n"
-      "ld1 { v25.h }[4], [x24]\n"
-      "ld1 { v24.h }[4], [x23]\n"
-      "ld1 { v23.h }[4], [x22]\n"
-      "ld1 { v22.h }[4], [x21]\n"
+      "ld1 { v26.h }[4], [x26]\n"
+      "ld1 { v25.h }[4], [x25]\n"
+      "ld1 { v24.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
+      "ld1 { v22.h }[4], [x22]\n"
+      "ld1 { v21.h }[4], [x21]\n"
       "b 7f\n"
       "5:"  // odd_loads_2_0
       "tbz %x[width], #1, 6f\n"
-      "ldr s29, [x28], #0x4\n"
-      "ldr s28, [x27], #0x4\n"
+      "ldr s28, [x28], #0x4\n"
+      "ldr s27, [x27], #0x4\n"
       "mov x20, #0x1\n"
-      "ldr s27, [x26], #0x4\n"
-      "ldr s26, [x25], #0x4\n"
-      "ldr s25, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s23, [x22], #0x4\n"
-      "ldr s22, [x21], #0x4\n"
+      "ldr s26, [x26], #0x4\n"
+      "ldr s25, [x25], #0x4\n"
+      "ldr s24, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s22, [x22], #0x4\n"
+      "ldr s21, [x21], #0x4\n"
       "tbz %x[width], #0, 7f\n"
-      "ld1 { v29.h }[2], [x28]\n"
-      "ld1 { v28.h }[2], [x27]\n"
+      "ld1 { v28.h }[2], [x28]\n"
+      "ld1 { v27.h }[2], [x27]\n"
       "mov x20, #0x2\n"
-      "ld1 { v27.h }[2], [x26]\n"
-      "ld1 { v26.h }[2], [x25]\n"
-      "ld1 { v25.h }[2], [x24]\n"
-      "ld1 { v24.h }[2], [x23]\n"
-      "ld1 { v23.h }[2], [x22]\n"
-      "ld1 { v22.h }[2], [x21]\n"
+      "ld1 { v26.h }[2], [x26]\n"
+      "ld1 { v25.h }[2], [x25]\n"
+      "ld1 { v24.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
+      "ld1 { v22.h }[2], [x22]\n"
+      "ld1 { v21.h }[2], [x21]\n"
       "b 7f\n"
       "6:"  // odd_loads_1_0
-      "ldr h29, [x28, #0x0]\n"
-      "ldr h28, [x27, #0x0]\n"
+      "ldr h28, [x28, #0x0]\n"
+      "ldr h27, [x27, #0x0]\n"
       "mov x20, #0x1\n"
-      "ldr h27, [x26, #0x0]\n"
-      "ldr h26, [x25, #0x0]\n"
-      "ldr h25, [x24, #0x0]\n"
-      "ldr h24, [x23, #0x0]\n"
-      "ldr h23, [x22, #0x0]\n"
-      "ldr h22, [x21, #0x0]\n"
+      "ldr h26, [x26, #0x0]\n"
+      "ldr h25, [x25, #0x0]\n"
+      "ldr h24, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
+      "ldr h22, [x22, #0x0]\n"
+      "ldr h21, [x21, #0x0]\n"
       "7:"  // Odd load end
-      "zip1 v21.4s, v29.4s, v27.4s\n"
       "zip1 v20.4s, v28.4s, v26.4s\n"
+      "zip1 v19.4s, v27.4s, v25.4s\n"
       "subs x20, x20, #0x1\n"
-      "zip1 v19.4s, v25.4s, v23.4s\n"
       "zip1 v18.4s, v24.4s, v22.4s\n"
-      "zip1 v17.4s, v21.4s, v20.4s\n"
-      "zip1 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip1 v17.4s, v23.4s, v21.4s\n"
+      "zip1 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 8f\n"
       "subs x20, x20, #0x1\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      "zip2 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 8f\n"
-      "zip2 v21.4s, v29.4s, v27.4s\n"
       "zip2 v20.4s, v28.4s, v26.4s\n"
+      "zip2 v19.4s, v27.4s, v25.4s\n"
       "subs x20, x20, #0x1\n"
-      "zip2 v19.4s, v25.4s, v23.4s\n"
       "zip2 v18.4s, v24.4s, v22.4s\n"
-      "zip1 v17.4s, v21.4s, v20.4s\n"
-      "zip1 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v17.4s, v23.4s, v21.4s\n"
+      "zip1 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 8f\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      "zip2 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "8:"  // Odds skip
       : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
-      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp
index a57810ce20..b3a52451e8 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 2, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset], LSL #2\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset], LSL #2\n"
       "add x27, x27, %x[row_offset], LSL #2\n"
       "add x26, x26, %x[row_offset], LSL #2\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset], LSL #2\n"
       "add x24, x24, %x[row_offset], LSL #2\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset], LSL #2\n"
       "add x22, x22, %x[row_offset], LSL #2\n"
       "add x21, x21, %x[row_offset], LSL #2\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,100 +79,100 @@ void interleave_block<8, 2, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q21, [x28], #0x10\n"
-      "ldr q16, [x27], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "ldr q19, [x27], #0x10\n"
       "subs %x[width], %x[width], #0x4\n"
-      "ldr q20, [x26], #0x10\n"
-      "ldr q19, [x25], #0x10\n"
       "cmp %x[width], #0x4\n"
-      "ldr q24, [x24], #0x10\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q23, [x22], #0x10\n"
-      "ldr q22, [x21], #0x10\n"
-      "zip1 v17.2d, v21.2d, v16.2d\n"
-      "zip2 v21.2d, v21.2d, v16.2d\n"
+      "ldr q25, [x26], #0x10\n"
+      "ldr q24, [x25], #0x10\n"
+      "zip1 v16.2d, v20.2d, v19.2d\n"
+      "zip1 v18.2d, v25.2d, v24.2d\n"
+      "ldr q23, [x24], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "zip1 v17.2d, v23.2d, v22.2d\n"
+      "zip2 v21.2d, v20.2d, v19.2d\n"
+      "ldr q20, [x22], #0x10\n"
+      "ldr q19, [x21], #0x10\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.2d, v20.2d, v19.2d\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
-      "zip1 v16.2d, v20.2d, v19.2d\n"
-      "zip2 v20.2d, v20.2d, v19.2d\n"
+      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip2 v18.2d, v25.2d, v24.2d\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "zip1 v19.2d, v24.2d, v18.2d\n"
-      "zip2 v18.2d, v24.2d, v18.2d\n"
+      "str q17, [%x[out_ptr], #0x20]\n"
+      "zip2 v17.2d, v23.2d, v22.2d\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
-      "zip1 v17.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip2 v16.2d, v20.2d, v19.2d\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "str q16, [%x[out_ptr], #0x10]\n"
-      "zip2 v16.2d, v23.2d, v22.2d\n"
-      "str q19, [%x[out_ptr], #0x20]\n"
-      "str q17, [%x[out_ptr], #0x30]\n"
       "str q21, [%x[out_ptr], #0x40]\n"
-      "str q20, [%x[out_ptr], #0x50]\n"
-      "str q18, [%x[out_ptr], #0x60]\n"
+      "str q18, [%x[out_ptr], #0x50]\n"
+      "str q17, [%x[out_ptr], #0x60]\n"
       "str q16, [%x[out_ptr], #0x70]\n"
       "add %x[out_ptr], %x[out_ptr], #0x80\n"
       "bge 2b\n"
       "3:"  // Main loop skip
       "cbz %x[width], 6f\n"
       "tbz %x[width], #1, 4f\n"
-      "ldr d27, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
+      "ldr d25, [x28], #0x8\n"
+      "ldr d24, [x27], #0x8\n"
       "mov x20, #0x1\n"
-      "ldr d25, [x26], #0x8\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d20, [x21], #0x8\n"
+      "ldr d23, [x26], #0x8\n"
+      "ldr d22, [x25], #0x8\n"
+      "ldr d21, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d19, [x22], #0x8\n"
+      "ldr d18, [x21], #0x8\n"
       "tbz %x[width], #0, 5f\n"
-      "ld1 { v27.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
+      "ld1 { v25.s }[2], [x28]\n"
+      "ld1 { v24.s }[2], [x27]\n"
       "mov x20, #0x2\n"
-      "ld1 { v25.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
-      "ld1 { v21.s }[2], [x22]\n"
-      "ld1 { v20.s }[2], [x21]\n"
+      "ld1 { v23.s }[2], [x26]\n"
+      "ld1 { v22.s }[2], [x25]\n"
+      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v19.s }[2], [x22]\n"
+      "ld1 { v18.s }[2], [x21]\n"
       "b 5f\n"
       "4:"  // odd_loads_1_0
-      "ldr s27, [x28, #0x0]\n"
-      "ldr s26, [x27, #0x0]\n"
+      "ldr s25, [x28, #0x0]\n"
+      "ldr s24, [x27, #0x0]\n"
       "mov x20, #0x1\n"
-      "ldr s25, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
-      "ldr s21, [x22, #0x0]\n"
-      "ldr s20, [x21, #0x0]\n"
+      "ldr s23, [x26, #0x0]\n"
+      "ldr s22, [x25, #0x0]\n"
+      "ldr s21, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s19, [x22, #0x0]\n"
+      "ldr s18, [x21, #0x0]\n"
       "5:"  // Odd load end
       "subs x20, x20, #0x1\n"
-      "zip1 v19.2d, v27.2d, v26.2d\n"
-      "zip1 v18.2d, v25.2d, v24.2d\n"
-      "zip1 v17.2d, v23.2d, v22.2d\n"
-      "zip1 v16.2d, v21.2d, v20.2d\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
-      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip1 v16.2d, v25.2d, v24.2d\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
+      "zip1 v17.2d, v21.2d, v20.2d\n"
+      "zip1 v16.2d, v19.2d, v18.2d\n"
       "str q17, [%x[out_ptr], #0x20]\n"
       "str q16, [%x[out_ptr], #0x30]\n"
       "add %x[out_ptr], %x[out_ptr], #0x40\n"
       "beq 6f\n"
-      "zip2 v19.2d, v27.2d, v26.2d\n"
-      "zip2 v18.2d, v25.2d, v24.2d\n"
-      "zip2 v17.2d, v23.2d, v22.2d\n"
-      "zip2 v16.2d, v21.2d, v20.2d\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
-      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip2 v16.2d, v25.2d, v24.2d\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
+      "zip2 v17.2d, v21.2d, v20.2d\n"
+      "zip2 v16.2d, v19.2d, v18.2d\n"
       "str q17, [%x[out_ptr], #0x20]\n"
       "str q16, [%x[out_ptr], #0x30]\n"
       "add %x[out_ptr], %x[out_ptr], #0x40\n"
       "6:"  // Odds skip
       : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
-      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_bf16_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_bf16_bf16.hpp
index edc1375b02..33639c201e 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_bf16_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_bf16_bf16.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 4, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset], LSL #1\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset], LSL #1\n"
       "add x27, x27, %x[row_offset], LSL #1\n"
       "add x26, x26, %x[row_offset], LSL #1\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset], LSL #1\n"
       "add x24, x24, %x[row_offset], LSL #1\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset], LSL #1\n"
       "add x22, x22, %x[row_offset], LSL #1\n"
       "add x21, x21, %x[row_offset], LSL #1\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,142 +79,142 @@ void interleave_block<8, 4, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q21, [x28], #0x10\n"
-      "ldr q16, [x27], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "ldr q19, [x27], #0x10\n"
       "subs %x[width], %x[width], #0x8\n"
-      "ldr q20, [x26], #0x10\n"
-      "ldr q19, [x25], #0x10\n"
       "cmp %x[width], #0x8\n"
-      "ldr q24, [x24], #0x10\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q23, [x22], #0x10\n"
-      "ldr q22, [x21], #0x10\n"
-      "zip1 v17.2d, v21.2d, v16.2d\n"
-      "zip2 v21.2d, v21.2d, v16.2d\n"
+      "ldr q25, [x26], #0x10\n"
+      "ldr q24, [x25], #0x10\n"
+      "zip1 v16.2d, v20.2d, v19.2d\n"
+      "zip1 v18.2d, v25.2d, v24.2d\n"
+      "ldr q23, [x24], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "zip1 v17.2d, v23.2d, v22.2d\n"
+      "zip2 v21.2d, v20.2d, v19.2d\n"
+      "ldr q20, [x22], #0x10\n"
+      "ldr q19, [x21], #0x10\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.2d, v20.2d, v19.2d\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
-      "zip1 v16.2d, v20.2d, v19.2d\n"
-      "zip2 v20.2d, v20.2d, v19.2d\n"
+      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip2 v18.2d, v25.2d, v24.2d\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "zip1 v19.2d, v24.2d, v18.2d\n"
-      "zip2 v18.2d, v24.2d, v18.2d\n"
+      "str q17, [%x[out_ptr], #0x20]\n"
+      "zip2 v17.2d, v23.2d, v22.2d\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
-      "zip1 v17.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip2 v16.2d, v20.2d, v19.2d\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "str q16, [%x[out_ptr], #0x10]\n"
-      "zip2 v16.2d, v23.2d, v22.2d\n"
-      "str q19, [%x[out_ptr], #0x20]\n"
-      "str q17, [%x[out_ptr], #0x30]\n"
       "str q21, [%x[out_ptr], #0x40]\n"
-      "str q20, [%x[out_ptr], #0x50]\n"
-      "str q18, [%x[out_ptr], #0x60]\n"
+      "str q18, [%x[out_ptr], #0x50]\n"
+      "str q17, [%x[out_ptr], #0x60]\n"
       "str q16, [%x[out_ptr], #0x70]\n"
       "add %x[out_ptr], %x[out_ptr], #0x80\n"
       "bge 2b\n"
       "3:"  // Main loop skip
       "cbz %x[width], 8f\n"
       "tbz %x[width], #2, 5f\n"
-      "ldr d27, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d25, [x26], #0x8\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d20, [x21], #0x8\n"
+      "ldr d25, [x28], #0x8\n"
+      "ldr d24, [x27], #0x8\n"
+      "ldr d23, [x26], #0x8\n"
+      "ldr d22, [x25], #0x8\n"
+      "ldr d21, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d19, [x22], #0x8\n"
+      "ldr d18, [x21], #0x8\n"
       "tbz %x[width], #1, 4f\n"
-      "ld1 { v27.s }[2], [x28], #0x4\n"
-      "ld1 { v26.s }[2], [x27], #0x4\n"
+      "ld1 { v25.s }[2], [x28], #0x4\n"
+      "ld1 { v24.s }[2], [x27], #0x4\n"
       "mov x20, #0x2\n"
-      "ld1 { v25.s }[2], [x26], #0x4\n"
-      "ld1 { v24.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
-      "ld1 { v22.s }[2], [x23], #0x4\n"
-      "ld1 { v21.s }[2], [x22], #0x4\n"
-      "ld1 { v20.s }[2], [x21], #0x4\n"
+      "ld1 { v23.s }[2], [x26], #0x4\n"
+      "ld1 { v22.s }[2], [x25], #0x4\n"
+      "ld1 { v21.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
+      "ld1 { v19.s }[2], [x22], #0x4\n"
+      "ld1 { v18.s }[2], [x21], #0x4\n"
       "tbz %x[width], #0, 7f\n"
-      "ld1 { v27.h }[6], [x28]\n"
-      "ld1 { v26.h }[6], [x27]\n"
-      "ld1 { v25.h }[6], [x26]\n"
-      "ld1 { v24.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
-      "ld1 { v22.h }[6], [x23]\n"
-      "ld1 { v21.h }[6], [x22]\n"
-      "ld1 { v20.h }[6], [x21]\n"
+      "ld1 { v25.h }[6], [x28]\n"
+      "ld1 { v24.h }[6], [x27]\n"
+      "ld1 { v23.h }[6], [x26]\n"
+      "ld1 { v22.h }[6], [x25]\n"
+      "ld1 { v21.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
+      "ld1 { v19.h }[6], [x22]\n"
+      "ld1 { v18.h }[6], [x21]\n"
       "b 7f\n"
       "4:"  // odd_loads_1_4
       "mov x20, #0x1\n"
       "tbz %x[width], #0, 7f\n"
-      "ld1 { v27.h }[4], [x28]\n"
-      "ld1 { v26.h }[4], [x27]\n"
+      "ld1 { v25.h }[4], [x28]\n"
+      "ld1 { v24.h }[4], [x27]\n"
       "mov x20, #0x2\n"
-      "ld1 { v25.h }[4], [x26]\n"
-      "ld1 { v24.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
-      "ld1 { v22.h }[4], [x23]\n"
-      "ld1 { v21.h }[4], [x22]\n"
-      "ld1 { v20.h }[4], [x21]\n"
+      "ld1 { v23.h }[4], [x26]\n"
+      "ld1 { v22.h }[4], [x25]\n"
+      "ld1 { v21.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
+      "ld1 { v19.h }[4], [x22]\n"
+      "ld1 { v18.h }[4], [x21]\n"
       "b 7f\n"
       "5:"  // odd_loads_2_0
       "tbz %x[width], #1, 6f\n"
-      "ldr s27, [x28], #0x4\n"
-      "ldr s26, [x27], #0x4\n"
+      "ldr s25, [x28], #0x4\n"
+      "ldr s24, [x27], #0x4\n"
       "mov x20, #0x1\n"
-      "ldr s25, [x26], #0x4\n"
-      "ldr s24, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
-      "ldr s22, [x23], #0x4\n"
-      "ldr s21, [x22], #0x4\n"
-      "ldr s20, [x21], #0x4\n"
+      "ldr s23, [x26], #0x4\n"
+      "ldr s22, [x25], #0x4\n"
+      "ldr s21, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
+      "ldr s19, [x22], #0x4\n"
+      "ldr s18, [x21], #0x4\n"
       "tbz %x[width], #0, 7f\n"
-      "ld1 { v27.h }[2], [x28]\n"
-      "ld1 { v26.h }[2], [x27]\n"
-      "ld1 { v25.h }[2], [x26]\n"
-      "ld1 { v24.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
-      "ld1 { v22.h }[2], [x23]\n"
-      "ld1 { v21.h }[2], [x22]\n"
-      "ld1 { v20.h }[2], [x21]\n"
+      "ld1 { v25.h }[2], [x28]\n"
+      "ld1 { v24.h }[2], [x27]\n"
+      "ld1 { v23.h }[2], [x26]\n"
+      "ld1 { v22.h }[2], [x25]\n"
+      "ld1 { v21.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
+      "ld1 { v19.h }[2], [x22]\n"
+      "ld1 { v18.h }[2], [x21]\n"
       "b 7f\n"
       "6:"  // odd_loads_1_0
-      "ldr h27, [x28, #0x0]\n"
-      "ldr h26, [x27, #0x0]\n"
+      "ldr h25, [x28, #0x0]\n"
+      "ldr h24, [x27, #0x0]\n"
       "mov x20, #0x1\n"
-      "ldr h25, [x26, #0x0]\n"
-      "ldr h24, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
-      "ldr h22, [x23, #0x0]\n"
-      "ldr h21, [x22, #0x0]\n"
-      "ldr h20, [x21, #0x0]\n"
+      "ldr h23, [x26, #0x0]\n"
+      "ldr h22, [x25, #0x0]\n"
+      "ldr h21, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
+      "ldr h19, [x22, #0x0]\n"
+      "ldr h18, [x21, #0x0]\n"
       "7:"  // Odd load end
       "subs x20, x20, #0x1\n"
-      "zip1 v19.2d, v27.2d, v26.2d\n"
-      "zip1 v18.2d, v25.2d, v24.2d\n"
-      "zip1 v17.2d, v23.2d, v22.2d\n"
-      "zip1 v16.2d, v21.2d, v20.2d\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
-      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip1 v16.2d, v25.2d, v24.2d\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
+      "zip1 v17.2d, v21.2d, v20.2d\n"
+      "zip1 v16.2d, v19.2d, v18.2d\n"
       "str q17, [%x[out_ptr], #0x20]\n"
       "str q16, [%x[out_ptr], #0x30]\n"
       "add %x[out_ptr], %x[out_ptr], #0x40\n"
       "beq 8f\n"
-      "zip2 v19.2d, v27.2d, v26.2d\n"
-      "zip2 v18.2d, v25.2d, v24.2d\n"
-      "zip2 v17.2d, v23.2d, v22.2d\n"
-      "zip2 v16.2d, v21.2d, v20.2d\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
-      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip2 v16.2d, v25.2d, v24.2d\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
+      "zip2 v17.2d, v21.2d, v20.2d\n"
+      "zip2 v16.2d, v19.2d, v18.2d\n"
       "str q17, [%x[out_ptr], #0x20]\n"
       "str q16, [%x[out_ptr], #0x30]\n"
       "add %x[out_ptr], %x[out_ptr], #0x40\n"
       "8:"  // Odds skip
       : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
-      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp
index ef1493b605..3044cfde48 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 4, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset], LSL #2\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset], LSL #2\n"
       "add x27, x27, %x[row_offset], LSL #2\n"
       "add x26, x26, %x[row_offset], LSL #2\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset], LSL #2\n"
       "add x24, x24, %x[row_offset], LSL #2\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset], LSL #2\n"
       "add x22, x22, %x[row_offset], LSL #2\n"
       "add x21, x21, %x[row_offset], LSL #2\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,36 +79,36 @@ void interleave_block<8, 4, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q19, [x28], #0x10\n"
-      "ldr q18, [x26], #0x10\n"
-      "subs %x[width], %x[width], #0x4\n"
+      "ldr q17, [x28], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      ".inst 0x0ea16a37  // bfcvtn v23.4h, v17.4s\n"
+      ".inst 0x0ea16a16  // bfcvtn v22.4h, v16.4s\n"
       "ldr q17, [x24], #0x10\n"
       "ldr q16, [x22], #0x10\n"
+      ".inst 0x0ea16a35  // bfcvtn v21.4h, v17.4s\n"
+      ".inst 0x0ea16a14  // bfcvtn v20.4h, v16.4s\n"
+      "ldr q19, [x27], #0x10\n"
+      "ldr q18, [x25], #0x10\n"
+      "subs %x[width], %x[width], #0x4\n"
       "cmp %x[width], #0x4\n"
-      "ldr q23, [x27], #0x10\n"
-      "ldr q22, [x25], #0x10\n"
-      "ldr q21, [x23], #0x10\n"
-      "ldr q20, [x21], #0x10\n"
-      ".inst 0x0ea16a73  // bfcvtn v19.4h, v19.4s\n"
-      ".inst 0x0ea16a52  // bfcvtn v18.4h, v18.4s\n"
-      ".inst 0x0ea16a31  // bfcvtn v17.4h, v17.4s\n"
-      ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
+      "ldr q17, [x23], #0x10\n"
+      "ldr q16, [x21], #0x10\n"
+      ".inst 0x4ea16a77  // bfcvtn2 v23.8h, v19.4s\n"
+      ".inst 0x4ea16a56  // bfcvtn2 v22.8h, v18.4s\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
+      ".inst 0x4ea16a35  // bfcvtn2 v21.8h, v17.4s\n"
+      ".inst 0x4ea16a14  // bfcvtn2 v20.8h, v16.4s\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
+      "str q23, [%x[out_ptr], #0x0]\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
-      ".inst 0x4ea16af3  // bfcvtn2 v19.8h, v23.4s\n"
-      ".inst 0x4ea16ad2  // bfcvtn2 v18.8h, v22.4s\n"
+      "str q22, [%x[out_ptr], #0x10]\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      ".inst 0x4ea16ab1  // bfcvtn2 v17.8h, v21.4s\n"
-      ".inst 0x4ea16a90  // bfcvtn2 v16.8h, v20.4s\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
-      "str q18, [%x[out_ptr], #0x10]\n"
-      "str q17, [%x[out_ptr], #0x20]\n"
-      "str q16, [%x[out_ptr], #0x30]\n"
+      "str q21, [%x[out_ptr], #0x20]\n"
+      "str q20, [%x[out_ptr], #0x30]\n"
       "add %x[out_ptr], %x[out_ptr], #0x40\n"
       "bge 2b\n"
       "3:"  // Main loop skip
@@ -150,9 +150,9 @@ void interleave_block<8, 4, VLType::None, false>(
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
       ".inst 0x4ea16af3  // bfcvtn2 v19.8h, v23.4s\n"
       ".inst 0x4ea16ad2  // bfcvtn2 v18.8h, v22.4s\n"
+      "str q19, [%x[out_ptr], #0x0]\n"
       ".inst 0x4ea16ab1  // bfcvtn2 v17.8h, v21.4s\n"
       ".inst 0x4ea16a90  // bfcvtn2 v16.8h, v20.4s\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
       "str q18, [%x[out_ptr], #0x10]\n"
       "str q17, [%x[out_ptr], #0x20]\n"
       "str q16, [%x[out_ptr], #0x30]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_s8_s8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_s8_s8.hpp
index ad213db3e5..8bb9898861 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_s8_s8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_s8_s8.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 4, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset]\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset]\n"
       "add x27, x27, %x[row_offset]\n"
       "add x26, x26, %x[row_offset]\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset]\n"
       "add x24, x24, %x[row_offset]\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset]\n"
       "add x22, x22, %x[row_offset]\n"
       "add x21, x21, %x[row_offset]\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,46 +79,46 @@ void interleave_block<8, 4, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q25, [x28], #0x10\n"
-      "ldr q24, [x27], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "ldr q18, [x27], #0x10\n"
       "subs %x[width], %x[width], #0x10\n"
-      "ldr q18, [x26], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
       "cmp %x[width], #0x10\n"
-      "ldr q23, [x24], #0x10\n"
-      "ldr q22, [x23], #0x10\n"
-      "ldr q21, [x22], #0x10\n"
+      "ldr q17, [x26], #0x10\n"
+      "ldr q16, [x25], #0x10\n"
+      "zip1 v25.4s, v20.4s, v17.4s\n"
+      "zip1 v24.4s, v18.4s, v16.4s\n"
+      "ldr q19, [x24], #0x10\n"
+      "ldr q23, [x23], #0x10\n"
+      "zip2 v22.4s, v20.4s, v17.4s\n"
+      "zip2 v21.4s, v18.4s, v16.4s\n"
+      "ldr q18, [x22], #0x10\n"
       "ldr q16, [x21], #0x10\n"
-      "zip1 v20.4s, v25.4s, v18.4s\n"
-      "zip1 v19.4s, v24.4s, v17.4s\n"
-      "zip2 v26.4s, v25.4s, v18.4s\n"
-      "zip2 v25.4s, v24.4s, v17.4s\n"
+      "zip1 v20.4s, v19.4s, v18.4s\n"
+      "zip1 v17.4s, v23.4s, v16.4s\n"
+      "zip2 v19.4s, v19.4s, v18.4s\n"
+      "zip2 v18.4s, v23.4s, v16.4s\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "zip1 v18.4s, v23.4s, v21.4s\n"
-      "zip1 v17.4s, v22.4s, v16.4s\n"
-      "zip2 v24.4s, v23.4s, v21.4s\n"
-      "zip2 v23.4s, v22.4s, v16.4s\n"
+      "zip1 v16.4s, v25.4s, v24.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
+      "zip1 v16.4s, v20.4s, v17.4s\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "zip1 v16.4s, v20.4s, v19.4s\n"
-      "zip2 v22.4s, v20.4s, v19.4s\n"
-      "zip1 v21.4s, v18.4s, v17.4s\n"
-      "zip2 v20.4s, v18.4s, v17.4s\n"
-      "zip1 v19.4s, v26.4s, v25.4s\n"
-      "zip1 v18.4s, v24.4s, v23.4s\n"
-      "str q16, [%x[out_ptr], #0x0]\n"
-      "zip2 v17.4s, v26.4s, v25.4s\n"
-      "zip2 v16.4s, v24.4s, v23.4s\n"
-      "str q21, [%x[out_ptr], #0x10]\n"
-      "str q22, [%x[out_ptr], #0x20]\n"
-      "str q20, [%x[out_ptr], #0x30]\n"
-      "str q19, [%x[out_ptr], #0x40]\n"
-      "str q18, [%x[out_ptr], #0x50]\n"
+      "zip2 v16.4s, v25.4s, v24.4s\n"
+      "str q16, [%x[out_ptr], #0x20]\n"
+      "zip2 v16.4s, v20.4s, v17.4s\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip1 v16.4s, v22.4s, v21.4s\n"
+      "str q16, [%x[out_ptr], #0x40]\n"
+      "zip1 v16.4s, v19.4s, v18.4s\n"
+      "zip2 v17.4s, v22.4s, v21.4s\n"
+      "str q16, [%x[out_ptr], #0x50]\n"
+      "zip2 v16.4s, v19.4s, v18.4s\n"
       "str q17, [%x[out_ptr], #0x60]\n"
       "str q16, [%x[out_ptr], #0x70]\n"
       "add %x[out_ptr], %x[out_ptr], #0x80\n"
@@ -126,203 +126,203 @@ void interleave_block<8, 4, VLType::None, false>(
       "3:"  // Main loop skip
       "cbz %x[width], 12f\n"
       "tbz %x[width], #3, 7f\n"
-      "ldr d29, [x28], #0x8\n"
-      "ldr d28, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d26, [x25], #0x8\n"
-      "ldr d25, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "ldr d22, [x21], #0x8\n"
+      "ldr d28, [x28], #0x8\n"
+      "ldr d27, [x27], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
+      "ldr d21, [x21], #0x8\n"
       "tbz %x[width], #2, 5f\n"
-      "ld1 { v29.s }[2], [x28], #0x4\n"
-      "ld1 { v28.s }[2], [x27], #0x4\n"
-      "ld1 { v27.s }[2], [x26], #0x4\n"
-      "ld1 { v26.s }[2], [x25], #0x4\n"
-      "ld1 { v25.s }[2], [x24], #0x4\n"
-      "ld1 { v24.s }[2], [x23], #0x4\n"
-      "ld1 { v23.s }[2], [x22], #0x4\n"
-      "ld1 { v22.s }[2], [x21], #0x4\n"
+      "ld1 { v28.s }[2], [x28], #0x4\n"
+      "ld1 { v27.s }[2], [x27], #0x4\n"
+      "ld1 { v26.s }[2], [x26], #0x4\n"
+      "ld1 { v25.s }[2], [x25], #0x4\n"
+      "ld1 { v24.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
+      "ld1 { v22.s }[2], [x22], #0x4\n"
+      "ld1 { v21.s }[2], [x21], #0x4\n"
       "tbz %x[width], #1, 4f\n"
-      "ld1 { v29.h }[6], [x28], #0x2\n"
-      "ld1 { v28.h }[6], [x27], #0x2\n"
+      "ld1 { v28.h }[6], [x28], #0x2\n"
+      "ld1 { v27.h }[6], [x27], #0x2\n"
       "mov x20, #0x4\n"
-      "ld1 { v27.h }[6], [x26], #0x2\n"
-      "ld1 { v26.h }[6], [x25], #0x2\n"
-      "ld1 { v25.h }[6], [x24], #0x2\n"
-      "ld1 { v24.h }[6], [x23], #0x2\n"
-      "ld1 { v23.h }[6], [x22], #0x2\n"
-      "ld1 { v22.h }[6], [x21], #0x2\n"
+      "ld1 { v26.h }[6], [x26], #0x2\n"
+      "ld1 { v25.h }[6], [x25], #0x2\n"
+      "ld1 { v24.h }[6], [x24], #0x2\n"
+      "ld1 { v23.h }[6], [x23], #0x2\n"
+      "ld1 { v22.h }[6], [x22], #0x2\n"
+      "ld1 { v21.h }[6], [x21], #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v29.b }[14], [x28]\n"
-      "ld1 { v28.b }[14], [x27]\n"
-      "ld1 { v27.b }[14], [x26]\n"
-      "ld1 { v26.b }[14], [x25]\n"
-      "ld1 { v25.b }[14], [x24]\n"
-      "ld1 { v24.b }[14], [x23]\n"
-      "ld1 { v23.b }[14], [x22]\n"
-      "ld1 { v22.b }[14], [x21]\n"
+      "ld1 { v28.b }[14], [x28]\n"
+      "ld1 { v27.b }[14], [x27]\n"
+      "ld1 { v26.b }[14], [x26]\n"
+      "ld1 { v25.b }[14], [x25]\n"
+      "ld1 { v24.b }[14], [x24]\n"
+      "ld1 { v23.b }[14], [x23]\n"
+      "ld1 { v22.b }[14], [x22]\n"
+      "ld1 { v21.b }[14], [x21]\n"
       "b 11f\n"
       "4:"  // odd_loads_1_12
       "mov x20, #0x3\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v29.b }[12], [x28]\n"
-      "ld1 { v28.b }[12], [x27]\n"
+      "ld1 { v28.b }[12], [x28]\n"
+      "ld1 { v27.b }[12], [x27]\n"
       "mov x20, #0x4\n"
-      "ld1 { v27.b }[12], [x26]\n"
-      "ld1 { v26.b }[12], [x25]\n"
-      "ld1 { v25.b }[12], [x24]\n"
-      "ld1 { v24.b }[12], [x23]\n"
-      "ld1 { v23.b }[12], [x22]\n"
-      "ld1 { v22.b }[12], [x21]\n"
+      "ld1 { v26.b }[12], [x26]\n"
+      "ld1 { v25.b }[12], [x25]\n"
+      "ld1 { v24.b }[12], [x24]\n"
+      "ld1 { v23.b }[12], [x23]\n"
+      "ld1 { v22.b }[12], [x22]\n"
+      "ld1 { v21.b }[12], [x21]\n"
       "b 11f\n"
       "5:"  // odd_loads_2_8
       "tbz %x[width], #1, 6f\n"
-      "ld1 { v29.h }[4], [x28], #0x2\n"
-      "ld1 { v28.h }[4], [x27], #0x2\n"
+      "ld1 { v28.h }[4], [x28], #0x2\n"
+      "ld1 { v27.h }[4], [x27], #0x2\n"
       "mov x20, #0x3\n"
-      "ld1 { v27.h }[4], [x26], #0x2\n"
-      "ld1 { v26.h }[4], [x25], #0x2\n"
-      "ld1 { v25.h }[4], [x24], #0x2\n"
-      "ld1 { v24.h }[4], [x23], #0x2\n"
-      "ld1 { v23.h }[4], [x22], #0x2\n"
-      "ld1 { v22.h }[4], [x21], #0x2\n"
+      "ld1 { v26.h }[4], [x26], #0x2\n"
+      "ld1 { v25.h }[4], [x25], #0x2\n"
+      "ld1 { v24.h }[4], [x24], #0x2\n"
+      "ld1 { v23.h }[4], [x23], #0x2\n"
+      "ld1 { v22.h }[4], [x22], #0x2\n"
+      "ld1 { v21.h }[4], [x21], #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v29.b }[10], [x28]\n"
-      "ld1 { v28.b }[10], [x27]\n"
-      "ld1 { v27.b }[10], [x26]\n"
-      "ld1 { v26.b }[10], [x25]\n"
-      "ld1 { v25.b }[10], [x24]\n"
-      "ld1 { v24.b }[10], [x23]\n"
-      "ld1 { v23.b }[10], [x22]\n"
-      "ld1 { v22.b }[10], [x21]\n"
+      "ld1 { v28.b }[10], [x28]\n"
+      "ld1 { v27.b }[10], [x27]\n"
+      "ld1 { v26.b }[10], [x26]\n"
+      "ld1 { v25.b }[10], [x25]\n"
+      "ld1 { v24.b }[10], [x24]\n"
+      "ld1 { v23.b }[10], [x23]\n"
+      "ld1 { v22.b }[10], [x22]\n"
+      "ld1 { v21.b }[10], [x21]\n"
       "b 11f\n"
       "6:"  // odd_loads_1_8
       "mov x20, #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v29.b }[8], [x28]\n"
-      "ld1 { v28.b }[8], [x27]\n"
+      "ld1 { v28.b }[8], [x28]\n"
+      "ld1 { v27.b }[8], [x27]\n"
       "mov x20, #0x3\n"
-      "ld1 { v27.b }[8], [x26]\n"
-      "ld1 { v26.b }[8], [x25]\n"
-      "ld1 { v25.b }[8], [x24]\n"
-      "ld1 { v24.b }[8], [x23]\n"
-      "ld1 { v23.b }[8], [x22]\n"
-      "ld1 { v22.b }[8], [x21]\n"
+      "ld1 { v26.b }[8], [x26]\n"
+      "ld1 { v25.b }[8], [x25]\n"
+      "ld1 { v24.b }[8], [x24]\n"
+      "ld1 { v23.b }[8], [x23]\n"
+      "ld1 { v22.b }[8], [x22]\n"
+      "ld1 { v21.b }[8], [x21]\n"
       "b 11f\n"
       "7:"  // odd_loads_4_0
       "tbz %x[width], #2, 9f\n"
-      "ldr s29, [x28], #0x4\n"
-      "ldr s28, [x27], #0x4\n"
-      "ldr s27, [x26], #0x4\n"
-      "ldr s26, [x25], #0x4\n"
-      "ldr s25, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s23, [x22], #0x4\n"
-      "ldr s22, [x21], #0x4\n"
+      "ldr s28, [x28], #0x4\n"
+      "ldr s27, [x27], #0x4\n"
+      "ldr s26, [x26], #0x4\n"
+      "ldr s25, [x25], #0x4\n"
+      "ldr s24, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s22, [x22], #0x4\n"
+      "ldr s21, [x21], #0x4\n"
       "tbz %x[width], #1, 8f\n"
-      "ld1 { v29.h }[2], [x28], #0x2\n"
-      "ld1 { v28.h }[2], [x27], #0x2\n"
+      "ld1 { v28.h }[2], [x28], #0x2\n"
+      "ld1 { v27.h }[2], [x27], #0x2\n"
       "mov x20, #0x2\n"
-      "ld1 { v27.h }[2], [x26], #0x2\n"
-      "ld1 { v26.h }[2], [x25], #0x2\n"
-      "ld1 { v25.h }[2], [x24], #0x2\n"
-      "ld1 { v24.h }[2], [x23], #0x2\n"
-      "ld1 { v23.h }[2], [x22], #0x2\n"
-      "ld1 { v22.h }[2], [x21], #0x2\n"
+      "ld1 { v26.h }[2], [x26], #0x2\n"
+      "ld1 { v25.h }[2], [x25], #0x2\n"
+      "ld1 { v24.h }[2], [x24], #0x2\n"
+      "ld1 { v23.h }[2], [x23], #0x2\n"
+      "ld1 { v22.h }[2], [x22], #0x2\n"
+      "ld1 { v21.h }[2], [x21], #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v29.b }[6], [x28]\n"
-      "ld1 { v28.b }[6], [x27]\n"
-      "ld1 { v27.b }[6], [x26]\n"
-      "ld1 { v26.b }[6], [x25]\n"
-      "ld1 { v25.b }[6], [x24]\n"
-      "ld1 { v24.b }[6], [x23]\n"
-      "ld1 { v23.b }[6], [x22]\n"
-      "ld1 { v22.b }[6], [x21]\n"
+      "ld1 { v28.b }[6], [x28]\n"
+      "ld1 { v27.b }[6], [x27]\n"
+      "ld1 { v26.b }[6], [x26]\n"
+      "ld1 { v25.b }[6], [x25]\n"
+      "ld1 { v24.b }[6], [x24]\n"
+      "ld1 { v23.b }[6], [x23]\n"
+      "ld1 { v22.b }[6], [x22]\n"
+      "ld1 { v21.b }[6], [x21]\n"
       "b 11f\n"
       "8:"  // odd_loads_1_4
       "mov x20, #0x1\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v29.b }[4], [x28]\n"
-      "ld1 { v28.b }[4], [x27]\n"
+      "ld1 { v28.b }[4], [x28]\n"
+      "ld1 { v27.b }[4], [x27]\n"
       "mov x20, #0x2\n"
-      "ld1 { v27.b }[4], [x26]\n"
-      "ld1 { v26.b }[4], [x25]\n"
-      "ld1 { v25.b }[4], [x24]\n"
-      "ld1 { v24.b }[4], [x23]\n"
-      "ld1 { v23.b }[4], [x22]\n"
-      "ld1 { v22.b }[4], [x21]\n"
+      "ld1 { v26.b }[4], [x26]\n"
+      "ld1 { v25.b }[4], [x25]\n"
+      "ld1 { v24.b }[4], [x24]\n"
+      "ld1 { v23.b }[4], [x23]\n"
+      "ld1 { v22.b }[4], [x22]\n"
+      "ld1 { v21.b }[4], [x21]\n"
       "b 11f\n"
       "9:"  // odd_loads_2_0
       "tbz %x[width], #1, 10f\n"
-      "ldr h29, [x28], #0x2\n"
-      "ldr h28, [x27], #0x2\n"
+      "ldr h28, [x28], #0x2\n"
+      "ldr h27, [x27], #0x2\n"
       "mov x20, #0x1\n"
-      "ldr h27, [x26], #0x2\n"
-      "ldr h26, [x25], #0x2\n"
-      "ldr h25, [x24], #0x2\n"
-      "ldr h24, [x23], #0x2\n"
-      "ldr h23, [x22], #0x2\n"
-      "ldr h22, [x21], #0x2\n"
+      "ldr h26, [x26], #0x2\n"
+      "ldr h25, [x25], #0x2\n"
+      "ldr h24, [x24], #0x2\n"
+      "ldr h23, [x23], #0x2\n"
+      "ldr h22, [x22], #0x2\n"
+      "ldr h21, [x21], #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v29.b }[2], [x28]\n"
-      "ld1 { v28.b }[2], [x27]\n"
-      "ld1 { v27.b }[2], [x26]\n"
-      "ld1 { v26.b }[2], [x25]\n"
-      "ld1 { v25.b }[2], [x24]\n"
-      "ld1 { v24.b }[2], [x23]\n"
-      "ld1 { v23.b }[2], [x22]\n"
-      "ld1 { v22.b }[2], [x21]\n"
+      "ld1 { v28.b }[2], [x28]\n"
+      "ld1 { v27.b }[2], [x27]\n"
+      "ld1 { v26.b }[2], [x26]\n"
+      "ld1 { v25.b }[2], [x25]\n"
+      "ld1 { v24.b }[2], [x24]\n"
+      "ld1 { v23.b }[2], [x23]\n"
+      "ld1 { v22.b }[2], [x22]\n"
+      "ld1 { v21.b }[2], [x21]\n"
       "b 11f\n"
       "10:"  // odd_loads_1_0
-      "ldr b29, [x28, #0x0]\n"
-      "ldr b28, [x27, #0x0]\n"
+      "ldr b28, [x28, #0x0]\n"
+      "ldr b27, [x27, #0x0]\n"
       "mov x20, #0x1\n"
-      "ldr b27, [x26, #0x0]\n"
-      "ldr b26, [x25, #0x0]\n"
-      "ldr b25, [x24, #0x0]\n"
-      "ldr b24, [x23, #0x0]\n"
-      "ldr b23, [x22, #0x0]\n"
-      "ldr b22, [x21, #0x0]\n"
+      "ldr b26, [x26, #0x0]\n"
+      "ldr b25, [x25, #0x0]\n"
+      "ldr b24, [x24, #0x0]\n"
+      "ldr b23, [x23, #0x0]\n"
+      "ldr b22, [x22, #0x0]\n"
+      "ldr b21, [x21, #0x0]\n"
       "11:"  // Odd load end
-      "zip1 v21.4s, v29.4s, v27.4s\n"
       "zip1 v20.4s, v28.4s, v26.4s\n"
+      "zip1 v19.4s, v27.4s, v25.4s\n"
       "subs x20, x20, #0x1\n"
-      "zip1 v19.4s, v25.4s, v23.4s\n"
       "zip1 v18.4s, v24.4s, v22.4s\n"
-      "zip1 v17.4s, v21.4s, v20.4s\n"
-      "zip1 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip1 v17.4s, v23.4s, v21.4s\n"
+      "zip1 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 12f\n"
       "subs x20, x20, #0x1\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      "zip2 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 12f\n"
-      "zip2 v21.4s, v29.4s, v27.4s\n"
       "zip2 v20.4s, v28.4s, v26.4s\n"
+      "zip2 v19.4s, v27.4s, v25.4s\n"
       "subs x20, x20, #0x1\n"
-      "zip2 v19.4s, v25.4s, v23.4s\n"
       "zip2 v18.4s, v24.4s, v22.4s\n"
-      "zip1 v17.4s, v21.4s, v20.4s\n"
-      "zip1 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v17.4s, v23.4s, v21.4s\n"
+      "zip1 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "beq 12f\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      "zip2 v16.4s, v19.4s, v18.4s\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v20.4s, v19.4s\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.4s, v18.4s, v17.4s\n"
       "str q16, [%x[out_ptr], #0x10]\n"
       "add %x[out_ptr], %x[out_ptr], #0x20\n"
       "12:"  // Odds skip
       : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
-      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp
index 28d2196ade..6e1efa3814 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp
@@ -34,29 +34,29 @@ void interleave_block<8, 8, VLType::None, false>(
       "ldr x28, [%x[in], #0x0]\n"
       "ldr x27, [%x[in], #0x8]\n"
       "cmp %x[height], #0x8\n"
+      "add x28, x28, %x[row_offset]\n"
       "ldr x26, [%x[in], #0x10]\n"
       "ldr x25, [%x[in], #0x18]\n"
-      "ldr x24, [%x[in], #0x20]\n"
-      "ldr x23, [%x[in], #0x28]\n"
-      "ldr x22, [%x[in], #0x30]\n"
-      "ldr x21, [%x[in], #0x38]\n"
-      "add x28, x28, %x[row_offset]\n"
       "add x27, x27, %x[row_offset]\n"
       "add x26, x26, %x[row_offset]\n"
+      "ldr x24, [%x[in], #0x20]\n"
+      "ldr x23, [%x[in], #0x28]\n"
       "add x25, x25, %x[row_offset]\n"
       "add x24, x24, %x[row_offset]\n"
+      "ldr x22, [%x[in], #0x30]\n"
+      "ldr x21, [%x[in], #0x38]\n"
       "add x23, x23, %x[row_offset]\n"
       "add x22, x22, %x[row_offset]\n"
       "add x21, x21, %x[row_offset]\n"
       "beq 1f\n"
       "cmp %x[height], #0x2\n"
-      "mov x21, x28\n"
       "csel x27, x27, x28, GE\n"
       "csel x26, x26, x28, GT\n"
       "cmp %x[height], #0x4\n"
       "csel x25, x25, x28, GE\n"
       "csel x24, x24, x28, GT\n"
       "cmp %x[height], #0x6\n"
+      "mov x21, x28\n"
       "csel x23, x23, x28, GE\n"
       "csel x22, x22, x28, GT\n"
       "1:"  // no_pointer_adj
@@ -79,226 +79,226 @@ void interleave_block<8, 8, VLType::None, false>(
       "prfm pldl1keep, [x21, #0x40]\n"
       "blt 3f\n"
       "2:"  // Main loop head
-      "ldr q21, [x28], #0x10\n"
-      "ldr q16, [x27], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "ldr q19, [x27], #0x10\n"
       "subs %x[width], %x[width], #0x10\n"
-      "ldr q20, [x26], #0x10\n"
-      "ldr q19, [x25], #0x10\n"
       "cmp %x[width], #0x10\n"
-      "ldr q24, [x24], #0x10\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q23, [x22], #0x10\n"
-      "ldr q22, [x21], #0x10\n"
-      "zip1 v17.2d, v21.2d, v16.2d\n"
-      "zip2 v21.2d, v21.2d, v16.2d\n"
+      "ldr q25, [x26], #0x10\n"
+      "ldr q24, [x25], #0x10\n"
+      "zip1 v16.2d, v20.2d, v19.2d\n"
+      "zip1 v18.2d, v25.2d, v24.2d\n"
+      "ldr q23, [x24], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "zip1 v17.2d, v23.2d, v22.2d\n"
+      "zip2 v21.2d, v20.2d, v19.2d\n"
+      "ldr q20, [x22], #0x10\n"
+      "ldr q19, [x21], #0x10\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.2d, v20.2d, v19.2d\n"
       "prfm pldl1keep, [x28, #0x70]\n"
       "prfm pldl1keep, [x27, #0x70]\n"
-      "zip1 v16.2d, v20.2d, v19.2d\n"
-      "zip2 v20.2d, v20.2d, v19.2d\n"
+      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip2 v18.2d, v25.2d, v24.2d\n"
       "prfm pldl1keep, [x26, #0x70]\n"
       "prfm pldl1keep, [x25, #0x70]\n"
-      "zip1 v19.2d, v24.2d, v18.2d\n"
-      "zip2 v18.2d, v24.2d, v18.2d\n"
+      "str q17, [%x[out_ptr], #0x20]\n"
+      "zip2 v17.2d, v23.2d, v22.2d\n"
       "prfm pldl1keep, [x24, #0x70]\n"
       "prfm pldl1keep, [x23, #0x70]\n"
-      "str q17, [%x[out_ptr], #0x0]\n"
-      "zip1 v17.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x30]\n"
+      "zip2 v16.2d, v20.2d, v19.2d\n"
       "prfm pldl1keep, [x22, #0x70]\n"
       "prfm pldl1keep, [x21, #0x70]\n"
-      "str q16, [%x[out_ptr], #0x10]\n"
-      "zip2 v16.2d, v23.2d, v22.2d\n"
-      "str q19, [%x[out_ptr], #0x20]\n"
-      "str q17, [%x[out_ptr], #0x30]\n"
       "str q21, [%x[out_ptr], #0x40]\n"
-      "str q20, [%x[out_ptr], #0x50]\n"
-      "str q18, [%x[out_ptr], #0x60]\n"
+      "str q18, [%x[out_ptr], #0x50]\n"
+      "str q17, [%x[out_ptr], #0x60]\n"
       "str q16, [%x[out_ptr], #0x70]\n"
       "add %x[out_ptr], %x[out_ptr], #0x80\n"
       "bge 2b\n"
       "3:"  // Main loop skip
       "cbz %x[width], 12f\n"
       "tbz %x[width], #3, 7f\n"
-      "ldr d27, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d25, [x26], #0x8\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d20, [x21], #0x8\n"
+      "ldr d25, [x28], #0x8\n"
+      "ldr d24, [x27], #0x8\n"
+      "ldr d23, [x26], #0x8\n"
+      "ldr d22, [x25], #0x8\n"
+      "ldr d21, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d19, [x22], #0x8\n"
+      "ldr d18, [x21], #0x8\n"
       "tbz %x[width], #2, 5f\n"
-      "ld1 { v27.s }[2], [x28], #0x4\n"
-      "ld1 { v26.s }[2], [x27], #0x4\n"
-      "ld1 { v25.s }[2], [x26], #0x4\n"
-      "ld1 { v24.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
-      "ld1 { v22.s }[2], [x23], #0x4\n"
-      "ld1 { v21.s }[2], [x22], #0x4\n"
-      "ld1 { v20.s }[2], [x21], #0x4\n"
+      "ld1 { v25.s }[2], [x28], #0x4\n"
+      "ld1 { v24.s }[2], [x27], #0x4\n"
+      "ld1 { v23.s }[2], [x26], #0x4\n"
+      "ld1 { v22.s }[2], [x25], #0x4\n"
+      "ld1 { v21.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
+      "ld1 { v19.s }[2], [x22], #0x4\n"
+      "ld1 { v18.s }[2], [x21], #0x4\n"
       "tbz %x[width], #1, 4f\n"
-      "ld1 { v27.h }[6], [x28], #0x2\n"
-      "ld1 { v26.h }[6], [x27], #0x2\n"
+      "ld1 { v25.h }[6], [x28], #0x2\n"
+      "ld1 { v24.h }[6], [x27], #0x2\n"
       "mov x20, #0x2\n"
-      "ld1 { v25.h }[6], [x26], #0x2\n"
-      "ld1 { v24.h }[6], [x25], #0x2\n"
-      "ld1 { v23.h }[6], [x24], #0x2\n"
-      "ld1 { v22.h }[6], [x23], #0x2\n"
-      "ld1 { v21.h }[6], [x22], #0x2\n"
-      "ld1 { v20.h }[6], [x21], #0x2\n"
+      "ld1 { v23.h }[6], [x26], #0x2\n"
+      "ld1 { v22.h }[6], [x25], #0x2\n"
+      "ld1 { v21.h }[6], [x24], #0x2\n"
+      "ld1 { v20.h }[6], [x23], #0x2\n"
+      "ld1 { v19.h }[6], [x22], #0x2\n"
+      "ld1 { v18.h }[6], [x21], #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v27.b }[14], [x28]\n"
-      "ld1 { v26.b }[14], [x27]\n"
-      "ld1 { v25.b }[14], [x26]\n"
-      "ld1 { v24.b }[14], [x25]\n"
-      "ld1 { v23.b }[14], [x24]\n"
-      "ld1 { v22.b }[14], [x23]\n"
-      "ld1 { v21.b }[14], [x22]\n"
-      "ld1 { v20.b }[14], [x21]\n"
+      "ld1 { v25.b }[14], [x28]\n"
+      "ld1 { v24.b }[14], [x27]\n"
+      "ld1 { v23.b }[14], [x26]\n"
+      "ld1 { v22.b }[14], [x25]\n"
+      "ld1 { v21.b }[14], [x24]\n"
+      "ld1 { v20.b }[14], [x23]\n"
+      "ld1 { v19.b }[14], [x22]\n"
+      "ld1 { v18.b }[14], [x21]\n"
       "b 11f\n"
       "4:"  // odd_loads_1_12
       "mov x20, #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v27.b }[12], [x28]\n"
-      "ld1 { v26.b }[12], [x27]\n"
-      "ld1 { v25.b }[12], [x26]\n"
-      "ld1 { v24.b }[12], [x25]\n"
-      "ld1 { v23.b }[12], [x24]\n"
-      "ld1 { v22.b }[12], [x23]\n"
-      "ld1 { v21.b }[12], [x22]\n"
-      "ld1 { v20.b }[12], [x21]\n"
+      "ld1 { v25.b }[12], [x28]\n"
+      "ld1 { v24.b }[12], [x27]\n"
+      "ld1 { v23.b }[12], [x26]\n"
+      "ld1 { v22.b }[12], [x25]\n"
+      "ld1 { v21.b }[12], [x24]\n"
+      "ld1 { v20.b }[12], [x23]\n"
+      "ld1 { v19.b }[12], [x22]\n"
+      "ld1 { v18.b }[12], [x21]\n"
       "b 11f\n"
       "5:"  // odd_loads_2_8
       "tbz %x[width], #1, 6f\n"
-      "ld1 { v27.h }[4], [x28], #0x2\n"
-      "ld1 { v26.h }[4], [x27], #0x2\n"
+      "ld1 { v25.h }[4], [x28], #0x2\n"
+      "ld1 { v24.h }[4], [x27], #0x2\n"
       "mov x20, #0x2\n"
-      "ld1 { v25.h }[4], [x26], #0x2\n"
-      "ld1 { v24.h }[4], [x25], #0x2\n"
-      "ld1 { v23.h }[4], [x24], #0x2\n"
-      "ld1 { v22.h }[4], [x23], #0x2\n"
-      "ld1 { v21.h }[4], [x22], #0x2\n"
-      "ld1 { v20.h }[4], [x21], #0x2\n"
+      "ld1 { v23.h }[4], [x26], #0x2\n"
+      "ld1 { v22.h }[4], [x25], #0x2\n"
+      "ld1 { v21.h }[4], [x24], #0x2\n"
+      "ld1 { v20.h }[4], [x23], #0x2\n"
+      "ld1 { v19.h }[4], [x22], #0x2\n"
+      "ld1 { v18.h }[4], [x21], #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v27.b }[10], [x28]\n"
-      "ld1 { v26.b }[10], [x27]\n"
-      "ld1 { v25.b }[10], [x26]\n"
-      "ld1 { v24.b }[10], [x25]\n"
-      "ld1 { v23.b }[10], [x24]\n"
-      "ld1 { v22.b }[10], [x23]\n"
-      "ld1 { v21.b }[10], [x22]\n"
-      "ld1 { v20.b }[10], [x21]\n"
+      "ld1 { v25.b }[10], [x28]\n"
+      "ld1 { v24.b }[10], [x27]\n"
+      "ld1 { v23.b }[10], [x26]\n"
+      "ld1 { v22.b }[10], [x25]\n"
+      "ld1 { v21.b }[10], [x24]\n"
+      "ld1 { v20.b }[10], [x23]\n"
+      "ld1 { v19.b }[10], [x22]\n"
+      "ld1 { v18.b }[10], [x21]\n"
       "b 11f\n"
       "6:"  // odd_loads_1_8
       "mov x20, #0x1\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v27.b }[8], [x28]\n"
-      "ld1 { v26.b }[8], [x27]\n"
+      "ld1 { v25.b }[8], [x28]\n"
+      "ld1 { v24.b }[8], [x27]\n"
       "mov x20, #0x2\n"
-      "ld1 { v25.b }[8], [x26]\n"
-      "ld1 { v24.b }[8], [x25]\n"
-      "ld1 { v23.b }[8], [x24]\n"
-      "ld1 { v22.b }[8], [x23]\n"
-      "ld1 { v21.b }[8], [x22]\n"
-      "ld1 { v20.b }[8], [x21]\n"
+      "ld1 { v23.b }[8], [x26]\n"
+      "ld1 { v22.b }[8], [x25]\n"
+      "ld1 { v21.b }[8], [x24]\n"
+      "ld1 { v20.b }[8], [x23]\n"
+      "ld1 { v19.b }[8], [x22]\n"
+      "ld1 { v18.b }[8], [x21]\n"
       "b 11f\n"
       "7:"  // odd_loads_4_0
       "tbz %x[width], #2, 9f\n"
-      "ldr s27, [x28], #0x4\n"
-      "ldr s26, [x27], #0x4\n"
-      "ldr s25, [x26], #0x4\n"
-      "ldr s24, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
-      "ldr s22, [x23], #0x4\n"
-      "ldr s21, [x22], #0x4\n"
-      "ldr s20, [x21], #0x4\n"
+      "ldr s25, [x28], #0x4\n"
+      "ldr s24, [x27], #0x4\n"
+      "ldr s23, [x26], #0x4\n"
+      "ldr s22, [x25], #0x4\n"
+      "ldr s21, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
+      "ldr s19, [x22], #0x4\n"
+      "ldr s18, [x21], #0x4\n"
       "tbz %x[width], #1, 8f\n"
-      "ld1 { v27.h }[2], [x28], #0x2\n"
-      "ld1 { v26.h }[2], [x27], #0x2\n"
+      "ld1 { v25.h }[2], [x28], #0x2\n"
+      "ld1 { v24.h }[2], [x27], #0x2\n"
       "mov x20, #0x1\n"
-      "ld1 { v25.h }[2], [x26], #0x2\n"
-      "ld1 { v24.h }[2], [x25], #0x2\n"
-      "ld1 { v23.h }[2], [x24], #0x2\n"
-      "ld1 { v22.h }[2], [x23], #0x2\n"
-      "ld1 { v21.h }[2], [x22], #0x2\n"
-      "ld1 { v20.h }[2], [x21], #0x2\n"
+      "ld1 { v23.h }[2], [x26], #0x2\n"
+      "ld1 { v22.h }[2], [x25], #0x2\n"
+      "ld1 { v21.h }[2], [x24], #0x2\n"
+      "ld1 { v20.h }[2], [x23], #0x2\n"
+      "ld1 { v19.h }[2], [x22], #0x2\n"
+      "ld1 { v18.h }[2], [x21], #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v27.b }[6], [x28]\n"
-      "ld1 { v26.b }[6], [x27]\n"
-      "ld1 { v25.b }[6], [x26]\n"
-      "ld1 { v24.b }[6], [x25]\n"
-      "ld1 { v23.b }[6], [x24]\n"
-      "ld1 { v22.b }[6], [x23]\n"
-      "ld1 { v21.b }[6], [x22]\n"
-      "ld1 { v20.b }[6], [x21]\n"
+      "ld1 { v25.b }[6], [x28]\n"
+      "ld1 { v24.b }[6], [x27]\n"
+      "ld1 { v23.b }[6], [x26]\n"
+      "ld1 { v22.b }[6], [x25]\n"
+      "ld1 { v21.b }[6], [x24]\n"
+      "ld1 { v20.b }[6], [x23]\n"
+      "ld1 { v19.b }[6], [x22]\n"
+      "ld1 { v18.b }[6], [x21]\n"
       "b 11f\n"
       "8:"  // odd_loads_1_4
       "mov x20, #0x1\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v27.b }[4], [x28]\n"
-      "ld1 { v26.b }[4], [x27]\n"
-      "ld1 { v25.b }[4], [x26]\n"
-      "ld1 { v24.b }[4], [x25]\n"
-      "ld1 { v23.b }[4], [x24]\n"
-      "ld1 { v22.b }[4], [x23]\n"
-      "ld1 { v21.b }[4], [x22]\n"
-      "ld1 { v20.b }[4], [x21]\n"
+      "ld1 { v25.b }[4], [x28]\n"
+      "ld1 { v24.b }[4], [x27]\n"
+      "ld1 { v23.b }[4], [x26]\n"
+      "ld1 { v22.b }[4], [x25]\n"
+      "ld1 { v21.b }[4], [x24]\n"
+      "ld1 { v20.b }[4], [x23]\n"
+      "ld1 { v19.b }[4], [x22]\n"
+      "ld1 { v18.b }[4], [x21]\n"
       "b 11f\n"
       "9:"  // odd_loads_2_0
       "tbz %x[width], #1, 10f\n"
-      "ldr h27, [x28], #0x2\n"
-      "ldr h26, [x27], #0x2\n"
+      "ldr h25, [x28], #0x2\n"
+      "ldr h24, [x27], #0x2\n"
       "mov x20, #0x1\n"
-      "ldr h25, [x26], #0x2\n"
-      "ldr h24, [x25], #0x2\n"
-      "ldr h23, [x24], #0x2\n"
-      "ldr h22, [x23], #0x2\n"
-      "ldr h21, [x22], #0x2\n"
-      "ldr h20, [x21], #0x2\n"
+      "ldr h23, [x26], #0x2\n"
+      "ldr h22, [x25], #0x2\n"
+      "ldr h21, [x24], #0x2\n"
+      "ldr h20, [x23], #0x2\n"
+      "ldr h19, [x22], #0x2\n"
+      "ldr h18, [x21], #0x2\n"
       "tbz %x[width], #0, 11f\n"
-      "ld1 { v27.b }[2], [x28]\n"
-      "ld1 { v26.b }[2], [x27]\n"
-      "ld1 { v25.b }[2], [x26]\n"
-      "ld1 { v24.b }[2], [x25]\n"
-      "ld1 { v23.b }[2], [x24]\n"
-      "ld1 { v22.b }[2], [x23]\n"
-      "ld1 { v21.b }[2], [x22]\n"
-      "ld1 { v20.b }[2], [x21]\n"
+      "ld1 { v25.b }[2], [x28]\n"
+      "ld1 { v24.b }[2], [x27]\n"
+      "ld1 { v23.b }[2], [x26]\n"
+      "ld1 { v22.b }[2], [x25]\n"
+      "ld1 { v21.b }[2], [x24]\n"
+      "ld1 { v20.b }[2], [x23]\n"
+      "ld1 { v19.b }[2], [x22]\n"
+      "ld1 { v18.b }[2], [x21]\n"
       "b 11f\n"
       "10:"  // odd_loads_1_0
-      "ldr b27, [x28, #0x0]\n"
-      "ldr b26, [x27, #0x0]\n"
+      "ldr b25, [x28, #0x0]\n"
+      "ldr b24, [x27, #0x0]\n"
       "mov x20, #0x1\n"
-      "ldr b25, [x26, #0x0]\n"
-      "ldr b24, [x25, #0x0]\n"
-      "ldr b23, [x24, #0x0]\n"
-      "ldr b22, [x23, #0x0]\n"
-      "ldr b21, [x22, #0x0]\n"
-      "ldr b20, [x21, #0x0]\n"
+      "ldr b23, [x26, #0x0]\n"
+      "ldr b22, [x25, #0x0]\n"
+      "ldr b21, [x24, #0x0]\n"
+      "ldr b20, [x23, #0x0]\n"
+      "ldr b19, [x22, #0x0]\n"
+      "ldr b18, [x21, #0x0]\n"
       "11:"  // Odd load end
       "subs x20, x20, #0x1\n"
-      "zip1 v19.2d, v27.2d, v26.2d\n"
-      "zip1 v18.2d, v25.2d, v24.2d\n"
-      "zip1 v17.2d, v23.2d, v22.2d\n"
-      "zip1 v16.2d, v21.2d, v20.2d\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
-      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip1 v16.2d, v25.2d, v24.2d\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip1 v16.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
+      "zip1 v17.2d, v21.2d, v20.2d\n"
+      "zip1 v16.2d, v19.2d, v18.2d\n"
       "str q17, [%x[out_ptr], #0x20]\n"
       "str q16, [%x[out_ptr], #0x30]\n"
       "add %x[out_ptr], %x[out_ptr], #0x40\n"
       "beq 12f\n"
-      "zip2 v19.2d, v27.2d, v26.2d\n"
-      "zip2 v18.2d, v25.2d, v24.2d\n"
-      "zip2 v17.2d, v23.2d, v22.2d\n"
-      "zip2 v16.2d, v21.2d, v20.2d\n"
-      "str q19, [%x[out_ptr], #0x0]\n"
-      "str q18, [%x[out_ptr], #0x10]\n"
+      "zip2 v16.2d, v25.2d, v24.2d\n"
+      "str q16, [%x[out_ptr], #0x0]\n"
+      "zip2 v16.2d, v23.2d, v22.2d\n"
+      "str q16, [%x[out_ptr], #0x10]\n"
+      "zip2 v17.2d, v21.2d, v20.2d\n"
+      "zip2 v16.2d, v19.2d, v18.2d\n"
       "str q17, [%x[out_ptr], #0x20]\n"
       "str q16, [%x[out_ptr], #0x30]\n"
       "add %x[out_ptr], %x[out_ptr], #0x40\n"
       "12:"  // Odds skip
       : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
-      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp
index ff171984e7..8ed20a1e48 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp
@@ -32,103 +32,103 @@ void interleave_block<1, 2, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "sub x28, %x[width], #0x1\n"
-      "mov x27, #0x0\n"
       "cntw x22, ALL, MUL #2\n"
+      "sub x28, %x[width], #0x1\n"
       "cntw x21, ALL, MUL #2\n"
       "sub x20, x22, #0x1\n"
       "whilelt p10.s, XZR, %x[height]\n"
       "add x28, x28, x21\n"
-      "ands x26, %x[width], x20\n"
+      "ands x27, %x[width], x20\n"
       "udiv x28, x28, x21\n"
-      "csel x26, x26, x22, NE\n"
+      "csel x27, x27, x22, NE\n"
+      "mov x26, #0x0\n"
       "and x25, x28, #0x1\n"
       "sub x28, x28, #0x1\n"
-      "add x26, x26, #0x1\n"
+      "add x27, x27, #0x1\n"
       "mov x20, %x[width]\n"
       "ptrue p0.b\n"
       "mov x24, %x[outptr_raw]\n"
       "mov x23, %x[row_offset]\n"
       "cntw x22\n"
       "lsr x28, x28, #0x1\n"
-      "lsr x26, x26, #0x1\n"
+      "lsr x27, x27, #0x1\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44771  // whilelt pn9.s, x27, x20, VLx2\n"
+      ".inst 0x25b44751  // whilelt pn9.s, x26, x20, VLx2\n"
       "mov x21, %x[in]\n"
       "1:"  // Width loop: Preamble: Loop
       "ldr x20, [x21], #0x8\n"
       ".inst 0x25306548  // psel p8.s, p9.s/Z, p10.s[w12]\n"
-      ".inst 0xa0174294  // ld1w { z20.s-z21.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
-      ".inst 0xc160e294  // bfcvt z20.h, { z20.s-z21.s }\n"
-      ".inst 0xc0800280  // mova za0h.s[x12], p0/M, z20.s\n"
+      ".inst 0xa0174286  // ld1w { z6.s-z7.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
+      ".inst 0xc160e0c6  // bfcvt z6.h, { z6.s-z7.s }\n"
+      ".inst 0xc08000c0  // mova za0h.s[x12], p0/M, z6.s\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x22\n"
       "blt 1b\n"
       "incw x23, ALL, MUL #2\n"
-      "incw x27, ALL, MUL #2\n"
+      "incw x26, ALL, MUL #2\n"
       "cbz x28, 5f\n"
       "2:"  // Width loop
       "mov x20, %x[width]\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44771  // whilelt pn9.s, x27, x20, VLx2\n"
+      ".inst 0x25b44751  // whilelt pn9.s, x26, x20, VLx2\n"
       "mov x21, %x[in]\n"
       "3:"  // Width loop: Odd: Loop
       "ldr x20, [x21], #0x8\n"
       ".inst 0x25306548  // psel p8.s, p9.s/Z, p10.s[w12]\n"
-      ".inst 0xc0828018  // mova z24.s, p0/M, za0v.s[x12]\n"
-      ".inst 0xa0174294  // ld1w { z20.s-z21.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
-      "st1w { z24.s }, p0, [x24]\n"
-      "addvl x24, x24, #1\n"
-      ".inst 0xc160e294  // bfcvt z20.h, { z20.s-z21.s }\n"
-      ".inst 0xc0800288  // mova za2h.s[x12], p0/M, z20.s\n"
+      ".inst 0xa017429e  // ld1w { z30.s-z31.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
+      ".inst 0xc160e3de  // bfcvt z30.h, { z30.s-z31.s }\n"
+      ".inst 0xc08003c8  // mova za2h.s[x12], p0/M, z30.s\n"
+      ".inst 0xc082800f  // mova z15.s, p0/M, za0v.s[x12]\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x22\n"
+      "st1w { z15.s }, p0, [x24]\n"
+      "addvl x24, x24, #1\n"
       "blt 3b\n"
-      "incw x27, ALL, MUL #2\n"
+      "incw x26, ALL, MUL #2\n"
       "mov x20, %x[width]\n"
       "incw x23, ALL, MUL #2\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44771  // whilelt pn9.s, x27, x20, VLx2\n"
+      ".inst 0x25b44751  // whilelt pn9.s, x26, x20, VLx2\n"
       "mov x21, %x[in]\n"
       "4:"  // Width loop: Even: Loop
       "ldr x20, [x21], #0x8\n"
       ".inst 0x25306548  // psel p8.s, p9.s/Z, p10.s[w12]\n"
-      ".inst 0xc082810c  // mova z12.s, p0/M, za2v.s[x12]\n"
-      ".inst 0xa0174284  // ld1w { z4.s-z5.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
-      "st1w { z12.s }, p0, [x24]\n"
-      "addvl x24, x24, #1\n"
-      ".inst 0xc160e084  // bfcvt z4.h, { z4.s-z5.s }\n"
-      ".inst 0xc0800080  // mova za0h.s[x12], p0/M, z4.s\n"
+      ".inst 0xa0174298  // ld1w { z24.s-z25.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
+      ".inst 0xc160e318  // bfcvt z24.h, { z24.s-z25.s }\n"
+      ".inst 0xc0800300  // mova za0h.s[x12], p0/M, z24.s\n"
+      ".inst 0xc0828110  // mova z16.s, p0/M, za2v.s[x12]\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x22\n"
+      "st1w { z16.s }, p0, [x24]\n"
+      "addvl x24, x24, #1\n"
       "blt 4b\n"
       "subs x28, x28, #0x1\n"
       "incw x23, ALL, MUL #2\n"
-      "incw x27, ALL, MUL #2\n"
+      "incw x26, ALL, MUL #2\n"
       "bgt 2b\n"
       "5:"  // Width loop: Tails
       "cbnz x25, 8f\n"
       "mov x20, %x[width]\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44771  // whilelt pn9.s, x27, x20, VLx2\n"
+      ".inst 0x25b44751  // whilelt pn9.s, x26, x20, VLx2\n"
       "mov x21, %x[in]\n"
       "6:"  // Width loop: Tails: Even: Odd: Loop
       "ldr x20, [x21], #0x8\n"
       ".inst 0x25306548  // psel p8.s, p9.s/Z, p10.s[w12]\n"
-      ".inst 0xc0828010  // mova z16.s, p0/M, za0v.s[x12]\n"
       ".inst 0xa017428e  // ld1w { z14.s-z15.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
-      "st1w { z16.s }, p0, [x24]\n"
-      "addvl x24, x24, #1\n"
       ".inst 0xc160e1ce  // bfcvt z14.h, { z14.s-z15.s }\n"
       ".inst 0xc08001c8  // mova za2h.s[x12], p0/M, z14.s\n"
+      ".inst 0xc0828010  // mova z16.s, p0/M, za0v.s[x12]\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x22\n"
+      "st1w { z16.s }, p0, [x24]\n"
+      "addvl x24, x24, #1\n"
       "blt 6b\n"
       "mov x12, #0x0\n"
       "7:"  // Width loop: Tails: Even: Even: Loop
       ".inst 0xc0828110  // mova z16.s, p0/M, za2v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x26\n"
+      "cmp x12, x27\n"
       "st1w { z16.s }, p0, [x24]\n"
       "addvl x24, x24, #1\n"
       "blt 7b\n"
@@ -138,7 +138,7 @@ void interleave_block<1, 2, VLType::SME, false>(
       "9:"  // Width loop: Tails: Odd: Loop
       ".inst 0xc0828010  // mova z16.s, p0/M, za0v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x26\n"
+      "cmp x12, x27\n"
       "st1w { z16.s }, p0, [x24]\n"
       "addvl x24, x24, #1\n"
       "blt 9b\n"
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave2VL_block2_fp32_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave2VL_block2_fp32_bf16.hpp
index 6d3601d165..aaa01039cf 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave2VL_block2_fp32_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave2VL_block2_fp32_bf16.hpp
@@ -32,126 +32,126 @@ void interleave_block<2, 2, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "sub x10, %x[width], #0x1\n"
-      "mov x9, #0x0\n"
       "cntw x22, ALL, MUL #2\n"
-      "cntw x28\n"
+      "cntw x9\n"
+      "sub x28, %x[width], #0x1\n"
       "cntw x21, ALL, MUL #2\n"
       "sub x20, x22, #0x1\n"
       ".inst 0x25207815  // ptrue pn13.b\n"
       "whilelt p12.s, XZR, %x[height]\n"
-      "whilelt p11.s, x28, %x[height]\n"
-      "add x10, x10, x21\n"
+      "whilelt p11.s, x9, %x[height]\n"
+      "add x28, x28, x21\n"
       "ands x27, %x[width], x20\n"
-      "udiv x10, x10, x21\n"
+      "udiv x28, x28, x21\n"
       "csel x27, x27, x22, NE\n"
-      "and x26, x10, #0x1\n"
-      "sub x10, x10, #0x1\n"
+      "mov x26, #0x0\n"
+      "and x25, x28, #0x1\n"
+      "sub x28, x28, #0x1\n"
       "add x27, x27, #0x1\n"
       "mov x20, %x[width]\n"
-      "mov x25, %x[in]\n"
+      "mov x24, %x[in]\n"
       "ptrue p0.b\n"
-      "mov x24, %x[outptr_raw]\n"
-      "mov x23, %x[row_offset]\n"
-      "lsr x10, x10, #0x1\n"
+      "mov x23, %x[outptr_raw]\n"
+      "mov x22, %x[row_offset]\n"
+      "lsr x28, x28, #0x1\n"
       "lsr x27, x27, #0x1\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44532  // whilelt pn10.s, x9, x20, VLx2\n"
-      "add x22, x25, x28, LSL #3\n"
+      ".inst 0x25b44752  // whilelt pn10.s, x26, x20, VLx2\n"
+      "add x21, x24, x9, LSL #3\n"
       "1:"  // Width loop: Preamble: Loop
-      "ldr x21, [x25], #0x8\n"
+      "ldr x20, [x24], #0x8\n"
       ".inst 0x25306989  // psel p9.s, p10.s/Z, p12.s[w12]\n"
       ".inst 0x25306968  // psel p8.s, p10.s/Z, p11.s[w12]\n"
-      "ldr x20, [x22], #0x8\n"
-      ".inst 0xa01746b4  // ld1w { z20.s-z21.s }, pn9.s/Z, [x21, x23, LSL #2]\n"
-      ".inst 0xa017428c  // ld1w { z12.s-z13.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
-      ".inst 0xc160e294  // bfcvt z20.h, { z20.s-z21.s }\n"
-      ".inst 0xc160e18c  // bfcvt z12.h, { z12.s-z13.s }\n"
-      ".inst 0xc0800280  // mova za0h.s[x12], p0/M, z20.s\n"
-      ".inst 0xc0800184  // mova za1h.s[x12], p0/M, z12.s\n"
+      ".inst 0xa0164698  // ld1w { z24.s-z25.s }, pn9.s/Z, [x20, x22, LSL #2]\n"
+      "ldr x20, [x21], #0x8\n"
+      ".inst 0xa0164296  // ld1w { z22.s-z23.s }, pn8.s/Z, [x20, x22, LSL #2]\n"
+      ".inst 0xc160e318  // bfcvt z24.h, { z24.s-z25.s }\n"
+      ".inst 0xc160e2d6  // bfcvt z22.h, { z22.s-z23.s }\n"
+      ".inst 0xc0800300  // mova za0h.s[x12], p0/M, z24.s\n"
+      ".inst 0xc08002c4  // mova za1h.s[x12], p0/M, z22.s\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x28\n"
+      "cmp x12, x9\n"
       "blt 1b\n"
-      "incw x23, ALL, MUL #2\n"
-      "incw x9, ALL, MUL #2\n"
-      "cbz x10, 5f\n"
+      "incw x22, ALL, MUL #2\n"
+      "incw x26, ALL, MUL #2\n"
+      "cbz x28, 5f\n"
       "2:"  // Width loop
       "mov x20, %x[width]\n"
-      "mov x25, %x[in]\n"
+      "mov x24, %x[in]\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44532  // whilelt pn10.s, x9, x20, VLx2\n"
-      "add x22, x25, x28, LSL #3\n"
+      ".inst 0x25b44752  // whilelt pn10.s, x26, x20, VLx2\n"
+      "add x21, x24, x9, LSL #3\n"
       "3:"  // Width loop: Odd: Loop
-      "ldr x21, [x25], #0x8\n"
+      "ldr x20, [x24], #0x8\n"
       ".inst 0x25306989  // psel p9.s, p10.s/Z, p12.s[w12]\n"
       ".inst 0x25306968  // psel p8.s, p10.s/Z, p11.s[w12]\n"
-      ".inst 0xc0828007  // mova z7.s, p0/M, za0v.s[x12]\n"
-      "ldr x20, [x22], #0x8\n"
-      ".inst 0xc082808f  // mova z15.s, p0/M, za1v.s[x12]\n"
-      ".inst 0xa01746b6  // ld1w { z22.s-z23.s }, pn9.s/Z, [x21, x23, LSL #2]\n"
-      ".inst 0xa017429a  // ld1w { z26.s-z27.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
-      ".inst 0xa1605707  // st1w { z7.s, z15.s }, pn13.b, [x24]\n"
-      "addvl x24, x24, #2\n"
+      ".inst 0xa0164696  // ld1w { z22.s-z23.s }, pn9.s/Z, [x20, x22, LSL #2]\n"
+      "ldr x20, [x21], #0x8\n"
+      ".inst 0xa016428a  // ld1w { z10.s-z11.s }, pn8.s/Z, [x20, x22, LSL #2]\n"
       ".inst 0xc160e2d6  // bfcvt z22.h, { z22.s-z23.s }\n"
-      ".inst 0xc160e35a  // bfcvt z26.h, { z26.s-z27.s }\n"
+      ".inst 0xc160e14a  // bfcvt z10.h, { z10.s-z11.s }\n"
       ".inst 0xc08002c8  // mova za2h.s[x12], p0/M, z22.s\n"
-      ".inst 0xc080034c  // mova za3h.s[x12], p0/M, z26.s\n"
+      ".inst 0xc080014c  // mova za3h.s[x12], p0/M, z10.s\n"
+      ".inst 0xc0828008  // mova z8.s, p0/M, za0v.s[x12]\n"
+      ".inst 0xc0828089  // mova z9.s, p0/M, za1v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x28\n"
+      "cmp x12, x9\n"
+      ".inst 0xa06056e8  // st1w { z8.s-z9.s }, pn13.b, [x23]\n"
+      "addvl x23, x23, #2\n"
       "blt 3b\n"
-      "incw x9, ALL, MUL #2\n"
+      "incw x26, ALL, MUL #2\n"
       "mov x20, %x[width]\n"
-      "mov x25, %x[in]\n"
-      "incw x23, ALL, MUL #2\n"
+      "mov x24, %x[in]\n"
+      "incw x22, ALL, MUL #2\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44532  // whilelt pn10.s, x9, x20, VLx2\n"
-      "add x22, x25, x28, LSL #3\n"
+      ".inst 0x25b44752  // whilelt pn10.s, x26, x20, VLx2\n"
+      "add x21, x24, x9, LSL #3\n"
       "4:"  // Width loop: Even: Loop
-      "ldr x21, [x25], #0x8\n"
+      "ldr x20, [x24], #0x8\n"
       ".inst 0x25306989  // psel p9.s, p10.s/Z, p12.s[w12]\n"
       ".inst 0x25306968  // psel p8.s, p10.s/Z, p11.s[w12]\n"
-      ".inst 0xc0828108  // mova z8.s, p0/M, za2v.s[x12]\n"
-      "ldr x20, [x22], #0x8\n"
-      ".inst 0xc0828189  // mova z9.s, p0/M, za3v.s[x12]\n"
-      ".inst 0xa01746ae  // ld1w { z14.s-z15.s }, pn9.s/Z, [x21, x23, LSL #2]\n"
-      ".inst 0xa017428c  // ld1w { z12.s-z13.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
-      ".inst 0xa0605708  // st1w { z8.s-z9.s }, pn13.b, [x24]\n"
-      "addvl x24, x24, #2\n"
-      ".inst 0xc160e1ce  // bfcvt z14.h, { z14.s-z15.s }\n"
-      ".inst 0xc160e18c  // bfcvt z12.h, { z12.s-z13.s }\n"
-      ".inst 0xc08001c0  // mova za0h.s[x12], p0/M, z14.s\n"
-      ".inst 0xc0800184  // mova za1h.s[x12], p0/M, z12.s\n"
+      ".inst 0xa016469a  // ld1w { z26.s-z27.s }, pn9.s/Z, [x20, x22, LSL #2]\n"
+      "ldr x20, [x21], #0x8\n"
+      ".inst 0xa016429e  // ld1w { z30.s-z31.s }, pn8.s/Z, [x20, x22, LSL #2]\n"
+      ".inst 0xc160e35a  // bfcvt z26.h, { z26.s-z27.s }\n"
+      ".inst 0xc160e3de  // bfcvt z30.h, { z30.s-z31.s }\n"
+      ".inst 0xc0800340  // mova za0h.s[x12], p0/M, z26.s\n"
+      ".inst 0xc08003c4  // mova za1h.s[x12], p0/M, z30.s\n"
+      ".inst 0xc0828106  // mova z6.s, p0/M, za2v.s[x12]\n"
+      ".inst 0xc082818e  // mova z14.s, p0/M, za3v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x28\n"
+      "cmp x12, x9\n"
+      ".inst 0xa16056e6  // st1w { z6.s, z14.s }, pn13.b, [x23]\n"
+      "addvl x23, x23, #2\n"
       "blt 4b\n"
-      "subs x10, x10, #0x1\n"
-      "incw x23, ALL, MUL #2\n"
-      "incw x9, ALL, MUL #2\n"
+      "subs x28, x28, #0x1\n"
+      "incw x22, ALL, MUL #2\n"
+      "incw x26, ALL, MUL #2\n"
       "bgt 2b\n"
       "5:"  // Width loop: Tails
-      "cbnz x26, 8f\n"
+      "cbnz x25, 8f\n"
       "mov x20, %x[width]\n"
-      "mov x25, %x[in]\n"
+      "mov x24, %x[in]\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44532  // whilelt pn10.s, x9, x20, VLx2\n"
-      "add x22, x25, x28, LSL #3\n"
+      ".inst 0x25b44752  // whilelt pn10.s, x26, x20, VLx2\n"
+      "add x21, x24, x9, LSL #3\n"
       "6:"  // Width loop: Tails: Even: Odd: Loop
-      "ldr x21, [x25], #0x8\n"
+      "ldr x20, [x24], #0x8\n"
       ".inst 0x25306989  // psel p9.s, p10.s/Z, p12.s[w12]\n"
       ".inst 0x25306968  // psel p8.s, p10.s/Z, p11.s[w12]\n"
-      ".inst 0xc0828003  // mova z3.s, p0/M, za0v.s[x12]\n"
-      "ldr x20, [x22], #0x8\n"
-      ".inst 0xc082808b  // mova z11.s, p0/M, za1v.s[x12]\n"
-      ".inst 0xa01746ac  // ld1w { z12.s-z13.s }, pn9.s/Z, [x21, x23, LSL #2]\n"
-      ".inst 0xa017428e  // ld1w { z14.s-z15.s }, pn8.s/Z, [x20, x23, LSL #2]\n"
-      ".inst 0xa1605703  // st1w { z3.s, z11.s }, pn13.b, [x24]\n"
-      "addvl x24, x24, #2\n"
+      ".inst 0xa016468c  // ld1w { z12.s-z13.s }, pn9.s/Z, [x20, x22, LSL #2]\n"
+      "ldr x20, [x21], #0x8\n"
+      ".inst 0xa016428e  // ld1w { z14.s-z15.s }, pn8.s/Z, [x20, x22, LSL #2]\n"
       ".inst 0xc160e18c  // bfcvt z12.h, { z12.s-z13.s }\n"
       ".inst 0xc160e1ce  // bfcvt z14.h, { z14.s-z15.s }\n"
       ".inst 0xc0800188  // mova za2h.s[x12], p0/M, z12.s\n"
       ".inst 0xc08001cc  // mova za3h.s[x12], p0/M, z14.s\n"
+      ".inst 0xc0828007  // mova z7.s, p0/M, za0v.s[x12]\n"
+      ".inst 0xc082808f  // mova z15.s, p0/M, za1v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x28\n"
+      "cmp x12, x9\n"
+      ".inst 0xa16056e7  // st1w { z7.s, z15.s }, pn13.b, [x23]\n"
+      "addvl x23, x23, #2\n"
       "blt 6b\n"
       "mov x12, #0x0\n"
       "7:"  // Width loop: Tails: Even: Even: Loop
@@ -159,8 +159,8 @@ void interleave_block<2, 2, VLType::SME, false>(
       ".inst 0xc082818f  // mova z15.s, p0/M, za3v.s[x12]\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x27\n"
-      ".inst 0xa060570e  // st1w { z14.s-z15.s }, pn13.b, [x24]\n"
-      "addvl x24, x24, #2\n"
+      ".inst 0xa06056ee  // st1w { z14.s-z15.s }, pn13.b, [x23]\n"
+      "addvl x23, x23, #2\n"
       "blt 7b\n"
       "b 10f\n"
       "8:"  // Width loop: Tails: Odd
@@ -170,15 +170,15 @@ void interleave_block<2, 2, VLType::SME, false>(
       ".inst 0xc0828095  // mova z21.s, p0/M, za1v.s[x12]\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x27\n"
-      ".inst 0xa0605714  // st1w { z20.s-z21.s }, pn13.b, [x24]\n"
-      "addvl x24, x24, #2\n"
+      ".inst 0xa06056f4  // st1w { z20.s-z21.s }, pn13.b, [x23]\n"
+      "addvl x23, x23, #2\n"
       "blt 9b\n"
       "10:"  // End
-      "mov %x[outptr_raw], x24\n"
+      "mov %x[outptr_raw], x23\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [outptr_raw] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp
index a8187f78e8..83a7f62693 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp
@@ -32,63 +32,63 @@ void interleave_block<4, 2, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "sub x14, %x[width], #0x1\n"
-      "mov x13, %x[in]\n"
       "cntw x23, ALL, MUL #2\n"
-      "cntw x11\n"
+      "cntw x10\n"
       "cntw x22, ALL, MUL #2\n"
       "cntw x20, ALL, MUL #3\n"
       "sub x21, x23, #0x1\n"
       ".inst 0x25207817  // ptrue pn15.b\n"
-      "whilelt p2.s, XZR, %x[height]\n"
-      "whilelt p1.s, x11, %x[height]\n"
-      "whilelt p14.s, x22, %x[height]\n"
-      "whilelt p13.s, x20, %x[height]\n"
+      "whilelt p1.s, XZR, %x[height]\n"
+      "whilelt p14.s, x10, %x[height]\n"
+      "whilelt p13.s, x22, %x[height]\n"
+      "whilelt p12.s, x20, %x[height]\n"
+      "sub x9, %x[width], #0x1\n"
       "cntw x20, ALL, MUL #2\n"
-      "ands x10, %x[width], x21\n"
-      "add x14, x14, x20\n"
-      "csel x10, x10, x23, NE\n"
-      "add x9, x13, x11, LSL #3\n"
-      "mov x28, #0x0\n"
-      "udiv x14, x14, x20\n"
-      "add x10, x10, #0x1\n"
+      "ands x28, %x[width], x21\n"
+      "mov x27, %x[in]\n"
+      "add x9, x9, x20\n"
+      "csel x28, x28, x23, NE\n"
+      "add x26, x27, x10, LSL #3\n"
+      "mov x25, #0x0\n"
+      "udiv x9, x9, x20\n"
+      "add x28, x28, #0x1\n"
       "mov x20, %x[width]\n"
-      "add x27, x9, x11, LSL #3\n"
+      "add x24, x26, x10, LSL #3\n"
       "ptrue p0.b\n"
-      "mov x26, %x[outptr_raw]\n"
-      "mov x25, %x[row_offset]\n"
-      "sub x14, x14, #0x1\n"
-      "lsr x10, x10, #0x1\n"
+      "mov x23, %x[outptr_raw]\n"
+      "mov x22, %x[row_offset]\n"
+      "sub x9, x9, #0x1\n"
+      "lsr x28, x28, #0x1\n"
       "mov x12, #0x0\n"
-      ".inst 0x25b44794  // whilelt pn12.s, x28, x20, VLx2\n"
-      "add x24, x27, x11, LSL #3\n"
+      ".inst 0x25b44733  // whilelt pn11.s, x25, x20, VLx2\n"
+      "add x21, x24, x10, LSL #3\n"
       "1:"  // Width loop: Preamble: Loop
-      "ldr x23, [x13], #0x8\n"
-      ".inst 0x2530704b  // psel p11.s, p12.s/Z, p2.s[w12]\n"
-      ".inst 0x2530702a  // psel p10.s, p12.s/Z, p1.s[w12]\n"
-      "ldr x22, [x9], #0x8\n"
-      ".inst 0x253071c9  // psel p9.s, p12.s/Z, p14.s[w12]\n"
-      ".inst 0x253071a8  // psel p8.s, p12.s/Z, p13.s[w12]\n"
-      "ldr x21, [x27], #0x8\n"
+      "ldr x20, [x27], #0x8\n"
+      ".inst 0x25306c28  // psel p8.s, p11.s/Z, p1.s[w12]\n"
+      ".inst 0x25306dca  // psel p10.s, p11.s/Z, p14.s[w12]\n"
+      ".inst 0xa0164298  // ld1w { z24.s-z25.s }, pn8.s/Z, [x20, x22, LSL #2]\n"
+      "ldr x20, [x26], #0x8\n"
+      ".inst 0x25306da9  // psel p9.s, p11.s/Z, p13.s[w12]\n"
+      ".inst 0x25306d88  // psel p8.s, p11.s/Z, p12.s[w12]\n"
+      ".inst 0xa0164a82  // ld1w { z2.s-z3.s }, pn10.s/Z, [x20, x22, LSL #2]\n"
       "ldr x20, [x24], #0x8\n"
-      ".inst 0xa0194eea  // ld1w { z10.s-z11.s }, pn11.s/Z, [x23, x25, LSL #2]\n"
-      ".inst 0xa0194ada  // ld1w { z26.s-z27.s }, pn10.s/Z, [x22, x25, LSL #2]\n"
-      ".inst 0xa01946be  // ld1w { z30.s-z31.s }, pn9.s/Z, [x21, x25, LSL #2]\n"
-      ".inst 0xa019428c  // ld1w { z12.s-z13.s }, pn8.s/Z, [x20, x25, LSL #2]\n"
+      ".inst 0xa016468a  // ld1w { z10.s-z11.s }, pn9.s/Z, [x20, x22, LSL #2]\n"
+      ".inst 0xc160e318  // bfcvt z24.h, { z24.s-z25.s }\n"
+      ".inst 0xc160e042  // bfcvt z2.h, { z2.s-z3.s }\n"
+      "ldr x20, [x21], #0x8\n"
+      ".inst 0xa016428c  // ld1w { z12.s-z13.s }, pn8.s/Z, [x20, x22, LSL #2]\n"
       ".inst 0xc160e14a  // bfcvt z10.h, { z10.s-z11.s }\n"
-      ".inst 0xc160e35a  // bfcvt z26.h, { z26.s-z27.s }\n"
-      ".inst 0xc0800140  // mova za0h.s[x12], p0/M, z10.s\n"
-      ".inst 0xc160e3de  // bfcvt z30.h, { z30.s-z31.s }\n"
-      ".inst 0xc0800344  // mova za1h.s[x12], p0/M, z26.s\n"
       ".inst 0xc160e18c  // bfcvt z12.h, { z12.s-z13.s }\n"
-      ".inst 0xc08003c8  // mova za2h.s[x12], p0/M, z30.s\n"
+      ".inst 0xc0800300  // mova za0h.s[x12], p0/M, z24.s\n"
+      ".inst 0xc0800044  // mova za1h.s[x12], p0/M, z2.s\n"
+      ".inst 0xc0800148  // mova za2h.s[x12], p0/M, z10.s\n"
       ".inst 0xc080018c  // mova za3h.s[x12], p0/M, z12.s\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x11\n"
+      "cmp x12, x10\n"
       "blt 1b\n"
+      "incw x22, ALL, MUL #2\n"
       "incw x25, ALL, MUL #2\n"
-      "incw x28, ALL, MUL #2\n"
-      "cbz x14, 5f\n"
+      "cbz x9, 5f\n"
       "2:"  // Width loop
       "mov x12, #0x0\n"
       "3:"  // Width loop: Store: Loop
@@ -97,44 +97,44 @@ void interleave_block<4, 2, VLType::SME, false>(
       ".inst 0xc0828119  // mova z25.s, p0/M, za2v.s[x12]\n"
       ".inst 0xc082819d  // mova z29.s, p0/M, za3v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x11\n"
-      ".inst 0xa160df51  // st1w { z17.s, z21.s, z25.s, z29.s }, pn15.b, [x26]\n"
-      "addvl x26, x26, #4\n"
+      "cmp x12, x10\n"
+      ".inst 0xa160def1  // st1w { z17.s, z21.s, z25.s, z29.s }, pn15.b, [x23]\n"
+      "addvl x23, x23, #4\n"
       "blt 3b\n"
-      "mov x13, %x[in]\n"
+      "mov x27, %x[in]\n"
+      "add x26, x27, x10, LSL #3\n"
       "mov x20, %x[width]\n"
-      "add x9, x13, x11, LSL #3\n"
+      "add x24, x26, x10, LSL #3\n"
       "mov x12, #0x0\n"
-      "add x27, x9, x11, LSL #3\n"
-      ".inst 0x25b44794  // whilelt pn12.s, x28, x20, VLx2\n"
-      "add x24, x27, x11, LSL #3\n"
+      ".inst 0x25b44733  // whilelt pn11.s, x25, x20, VLx2\n"
+      "add x21, x24, x10, LSL #3\n"
       "4:"  // Width loop: Load: Loop
-      "ldr x23, [x13], #0x8\n"
-      ".inst 0x2530704b  // psel p11.s, p12.s/Z, p2.s[w12]\n"
-      ".inst 0x2530702a  // psel p10.s, p12.s/Z, p1.s[w12]\n"
-      "ldr x22, [x9], #0x8\n"
-      ".inst 0x253071c9  // psel p9.s, p12.s/Z, p14.s[w12]\n"
-      ".inst 0x253071a8  // psel p8.s, p12.s/Z, p13.s[w12]\n"
-      "ldr x21, [x27], #0x8\n"
+      "ldr x20, [x27], #0x8\n"
+      ".inst 0x25306c28  // psel p8.s, p11.s/Z, p1.s[w12]\n"
+      ".inst 0x25306dca  // psel p10.s, p11.s/Z, p14.s[w12]\n"
+      ".inst 0xa016428c  // ld1w { z12.s-z13.s }, pn8.s/Z, [x20, x22, LSL #2]\n"
+      "ldr x20, [x26], #0x8\n"
+      ".inst 0x25306da9  // psel p9.s, p11.s/Z, p13.s[w12]\n"
+      ".inst 0x25306d88  // psel p8.s, p11.s/Z, p12.s[w12]\n"
+      ".inst 0xa0164a8e  // ld1w { z14.s-z15.s }, pn10.s/Z, [x20, x22, LSL #2]\n"
       "ldr x20, [x24], #0x8\n"
-      ".inst 0xa0194eec  // ld1w { z12.s-z13.s }, pn11.s/Z, [x23, x25, LSL #2]\n"
-      ".inst 0xa0194ace  // ld1w { z14.s-z15.s }, pn10.s/Z, [x22, x25, LSL #2]\n"
-      ".inst 0xa01946b2  // ld1w { z18.s-z19.s }, pn9.s/Z, [x21, x25, LSL #2]\n"
-      ".inst 0xa019429e  // ld1w { z30.s-z31.s }, pn8.s/Z, [x20, x25, LSL #2]\n"
+      ".inst 0xa0164692  // ld1w { z18.s-z19.s }, pn9.s/Z, [x20, x22, LSL #2]\n"
       ".inst 0xc160e18c  // bfcvt z12.h, { z12.s-z13.s }\n"
       ".inst 0xc160e1ce  // bfcvt z14.h, { z14.s-z15.s }\n"
-      ".inst 0xc0800180  // mova za0h.s[x12], p0/M, z12.s\n"
+      "ldr x20, [x21], #0x8\n"
+      ".inst 0xa016429e  // ld1w { z30.s-z31.s }, pn8.s/Z, [x20, x22, LSL #2]\n"
       ".inst 0xc160e252  // bfcvt z18.h, { z18.s-z19.s }\n"
-      ".inst 0xc08001c4  // mova za1h.s[x12], p0/M, z14.s\n"
       ".inst 0xc160e3de  // bfcvt z30.h, { z30.s-z31.s }\n"
+      ".inst 0xc0800180  // mova za0h.s[x12], p0/M, z12.s\n"
+      ".inst 0xc08001c4  // mova za1h.s[x12], p0/M, z14.s\n"
       ".inst 0xc0800248  // mova za2h.s[x12], p0/M, z18.s\n"
       ".inst 0xc08003cc  // mova za3h.s[x12], p0/M, z30.s\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x11\n"
+      "cmp x12, x10\n"
       "blt 4b\n"
-      "subs x14, x14, #0x1\n"
+      "subs x9, x9, #0x1\n"
+      "incw x22, ALL, MUL #2\n"
       "incw x25, ALL, MUL #2\n"
-      "incw x28, ALL, MUL #2\n"
       "bgt 2b\n"
       "5:"  // Width loop: Tails
       "mov x12, #0x0\n"
@@ -144,16 +144,16 @@ void interleave_block<4, 2, VLType::SME, false>(
       ".inst 0xc0828119  // mova z25.s, p0/M, za2v.s[x12]\n"
       ".inst 0xc082819d  // mova z29.s, p0/M, za3v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x10\n"
-      ".inst 0xa160df51  // st1w { z17.s, z21.s, z25.s, z29.s }, pn15.b, [x26]\n"
-      "addvl x26, x26, #4\n"
+      "cmp x12, x28\n"
+      ".inst 0xa160def1  // st1w { z17.s, z21.s, z25.s, z29.s }, pn15.b, [x23]\n"
+      "addvl x23, x23, #4\n"
       "blt 6b\n"
       "7:"  // End
-      "mov %x[outptr_raw], x26\n"
+      "mov %x[outptr_raw], x23\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [outptr_raw] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_bf16_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_bf16_bf16.hpp
index bc9f68ed72..45f660fec1 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_bf16_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_bf16_bf16.hpp
@@ -33,25 +33,25 @@ void interleave_block<1, 1, VLType::SME, false>(
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
       "mov x21, %x[width]\n"
-      "mov x20, %x[width]\n"
       "inch x21\n"
       "cnth x11\n"
       "sub x21, x21, #0x1\n"
-      "sub x10, x11, #0x1\n"
       "udiv x21, x21, x11\n"  // n_passes = ceildiv(width, VL<T>)
-      "ands x10, x20, x10\n"
+      "mov x20, %x[width]\n"
+      "sub x10, x11, #0x1\n"
       "sub x9, x21, #0x1\n"
+      "ands x10, x20, x10\n"
       "sub x28, x11, #0x2\n"
       "lsl x20, %x[height], #0x1\n"  // height * 2
       "mov x27, #0x0\n"
       "mov x26, %x[in]\n"
       "lsr x9, x9, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x25, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
+      "ldr x25, [x26, #0x0]\n"
+      "and x24, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "csel x10, x10, x11, NE\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x23, [x26, #0x8]\n"
       "ptrue p11.h\n"
       "whilelt p10.h, XZR, x20\n"
-      "ldr x23, [x26, #0x8]\n"
       "mov x22, %x[row_offset]\n"
       "mov x21, %x[out]\n"
       "whilelt p9.h, x27, %x[width]\n"
@@ -60,119 +60,119 @@ void interleave_block<1, 1, VLType::SME, false>(
       "mov x12, #0x0\n"
       "cbz x28, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25286141  // psel p1.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       ".inst 0x25386140  // psel p0.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0xe0560700  // ld1h { za0h.h[x12] }, p1/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x25, [x26, #0x0]\n"
       ".inst 0xe05602e1  // ld1h { za0h.h[x12, #1] }, p0/Z, [x23, x22, LSL #1]\n"
       "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "cmp x12, x28\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25286141  // psel p1.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       ".inst 0x25386140  // psel p0.h, p8.h/Z, p10.h[w12, #1]\n"
       "mov x26, %x[in]\n"
-      "inch x27\n"
-      ".inst 0xe0560700  // ld1h { za0h.h[x12] }, p1/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       ".inst 0xe05602e1  // ld1h { za0h.h[x12, #1] }, p0/Z, [x23, x22, LSL #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      "inch x22\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "inch x22\n"
+      "inch x27\n"
       "cbz x9, 8f\n"
       "mov x20, x9\n"
       "3:"  // K loop: Main loop
       "whilelt p8.h, x27, %x[width]\n"
-      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
       "cbz x28, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25296143  // psel p3.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25396142  // psel p2.h, p8.h/Z, p10.h[w13, #1]\n"
-      ".inst 0x25296d21  // psel p1.h, p11.h/Z, p9.h[w13]\n"
-      ".inst 0x25396d20  // psel p0.h, p11.h/Z, p9.h[w13, #1]\n"
-      ".inst 0xe0562f08  // ld1h { za1h.h[x13] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0562ae9  // ld1h { za1h.h[x13, #1] }, p2/Z, [x23, x22, LSL #1]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560328  // ld1h { za1h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
+      ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe05606e9  // ld1h { za1h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
       "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe07fa6a0  // st1h { za0v.h[x13] }, p1/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0xe07f82a0  // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
+      ".inst 0xe06b82a1  // st1h { za0v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "add x26, x26, #0x10\n"
-      ".inst 0xe06ba2a1  // st1h { za0v.h[x13, #1] }, p0/Z, [x21, x11, LSL #1]\n"
-      "add x13, x13, #0x2\n"
       "addvl x21, x21, #2\n"
-      "cmp x13, x28\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25296143  // psel p3.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25396142  // psel p2.h, p8.h/Z, p10.h[w13, #1]\n"
-      ".inst 0x25296d21  // psel p1.h, p11.h/Z, p9.h[w13]\n"
-      ".inst 0x25396d20  // psel p0.h, p11.h/Z, p9.h[w13, #1]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560328  // ld1h { za1h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe05606e9  // ld1h { za1h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe07f82a0  // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
       "whilelt p9.h, x27, %x[width]\n"
-      ".inst 0xe0562f08  // ld1h { za1h.h[x13] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       "inch x27\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0562ae9  // ld1h { za1h.h[x13, #1] }, p2/Z, [x23, x22, LSL #1]\n"
-      "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe07fa6a0  // st1h { za0v.h[x13] }, p1/Z, [x21, XZR, LSL #1]\n"
       "add x26, x26, #0x10\n"
-      ".inst 0xe06ba2a1  // st1h { za0v.h[x13, #1] }, p0/Z, [x21, x11, LSL #1]\n"
+      ".inst 0xe06b82a1  // st1h { za0v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
       "addvl x21, x21, #2\n"
       "inch x22\n"
       "whilelt p8.h, x27, %x[width]\n"
+      "mov x12, #0x0\n"
       "cbz x28, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25286143  // psel p3.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25386142  // psel p2.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0x25286d21  // psel p1.h, p11.h/Z, p9.h[w12]\n"
-      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
-      ".inst 0xe0560f00  // ld1h { za0h.h[x12] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0560ae1  // ld1h { za0h.h[x12, #1] }, p2/Z, [x23, x22, LSL #1]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
+      ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe05606e1  // ld1h { za0h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
       "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe07f86a8  // st1h { za1v.h[x12] }, p1/Z, [x21, XZR, LSL #1]\n"
-      "add x26, x26, #0x10\n"
+      ".inst 0xe07f82a8  // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
       ".inst 0xe06b82a9  // st1h { za1v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
       "add x12, x12, #0x2\n"
-      "addvl x21, x21, #2\n"
       "cmp x12, x28\n"
+      "add x26, x26, #0x10\n"
+      "addvl x21, x21, #2\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25286143  // psel p3.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25386142  // psel p2.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0x25286d21  // psel p1.h, p11.h/Z, p9.h[w12]\n"
-      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe05606e1  // ld1h { za0h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe07f82a8  // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
       "whilelt p9.h, x27, %x[width]\n"
-      ".inst 0xe0560f00  // ld1h { za0h.h[x12] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       "subs x20, x20, #0x1\n"
-      "inch x27\n"
-      ".inst 0xe0560ae1  // ld1h { za0h.h[x12, #1] }, p2/Z, [x23, x22, LSL #1]\n"
-      "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe07f86a8  // st1h { za1v.h[x12] }, p1/Z, [x21, XZR, LSL #1]\n"
       "add x26, x26, #0x10\n"
       ".inst 0xe06b82a9  // st1h { za1v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
       "addvl x21, x21, #2\n"
+      "inch x27\n"
       "inch x22\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x25, 11f\n"
+      "cbnz x24, 11f\n"
       "mov x26, %x[in]\n"
       "whilelt p8.h, x27, %x[width]\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25286d21  // psel p1.h, p11.h/Z, p9.h[w12]\n"
-      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0xe07f86a0  // st1h { za0v.h[x12] }, p1/Z, [x21, XZR, LSL #1]\n"
-      "addvl x21, x21, #1\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe07f82a0  // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
       "ldr x20, [x26, #0x0]\n"
-      "add x26, x26, #0x8\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
       ".inst 0xe0560288  // ld1h { za1h.h[x12] }, p0/Z, [x20, x22, LSL #1]\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x11\n"
+      "add x26, x26, #0x8\n"
+      "addvl x21, x21, #1\n"
       "blt 9b\n"
       "whilelt p9.h, x27, %x[width]\n"
       "whilelt p8.h, x27, %x[width]\n"
@@ -181,8 +181,8 @@ void interleave_block<1, 1, VLType::SME, false>(
       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
       ".inst 0xe07f82a8  // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
       "blt 10b\n"
       "whilelt p8.h, x27, %x[width]\n"
       "b 13f\n"
@@ -192,15 +192,15 @@ void interleave_block<1, 1, VLType::SME, false>(
       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
       ".inst 0xe07f82a0  // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block2_bf16_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block2_bf16_bf16.hpp
index 793bc80524..ce7192afe6 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block2_bf16_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block2_bf16_bf16.hpp
@@ -32,28 +32,28 @@ void interleave_block<1, 2, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x22, %x[width]\n"
+      "cnth x22\n"
       "mov x21, %x[width]\n"
-      "cnth x20\n"
-      "inch x22\n"
-      "sub x11, x20, #0x1\n"
-      "sub x22, x22, #0x1\n"
-      "ands x11, x21, x11\n"
+      "inch x21\n"
+      "mov x20, %x[width]\n"
+      "sub x11, x22, #0x1\n"
+      "sub x21, x21, #0x1\n"
+      "ands x11, x20, x11\n"
       "cntw x10\n"
-      "udiv x22, x22, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x11, x11, x20, NE\n"
-      "sub x9, x22, #0x1\n"
+      "udiv x21, x21, x22\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x11, x11, x22, NE\n"
+      "sub x9, x21, #0x1\n"
       "add x11, x11, #0x1\n"
       "sub x28, x10, #0x2\n"
       "lsl x20, %x[height], #0x1\n"  // height * 2
       "mov x27, #0x0\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
       "lsr x9, x9, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x25, x22, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "ldr x24, [x26, #0x0]\n"
+      "and x24, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
+      "ldr x23, [x26, #0x8]\n"
       "lsr x11, x11, #0x1\n"
       "ptrue p11.s\n"
-      "ldr x23, [x26, #0x8]\n"
       "whilelt p10.h, XZR, x20\n"
       "mov x22, %x[row_offset]\n"
       "mov x21, %x[out]\n"
@@ -63,124 +63,124 @@ void interleave_block<1, 2, VLType::SME, false>(
       "mov x12, #0x0\n"
       "cbz x28, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25286141  // psel p1.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       ".inst 0x25686140  // psel p0.h, p8.h/Z, p10.h[w12, #2]\n"
-      ".inst 0xe0560700  // ld1h { za0h.h[x12] }, p1/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x25, [x26, #0x0]\n"
       ".inst 0xe05602e2  // ld1h { za0h.h[x12, #2] }, p0/Z, [x23, x22, LSL #1]\n"
       "add x12, x12, #0x4\n"
+      "cmp x12, x28, LSL #1\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "cmp x12, x28, LSL #1\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25286141  // psel p1.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       ".inst 0x25686140  // psel p0.h, p8.h/Z, p10.h[w12, #2]\n"
       "mov x26, %x[in]\n"
-      "inch x27\n"
-      ".inst 0xe0560700  // ld1h { za0h.h[x12] }, p1/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       ".inst 0xe05602e2  // ld1h { za0h.h[x12, #2] }, p0/Z, [x23, x22, LSL #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      "inch x22\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "inch x22\n"
+      "inch x27\n"
       "cbz x9, 8f\n"
       "mov x20, x9\n"
       "3:"  // K loop: Main loop
       "whilelt p8.h, x27, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      "mov x14, #0x0\n"
       "cbz x28, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25386143  // psel p3.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0x25786142  // psel p2.h, p8.h/Z, p10.h[w12, #3]\n"
-      ".inst 0x252a6d21  // psel p1.h, p11.h/Z, p9.h[w14]\n"
-      ".inst 0x253a6d20  // psel p0.h, p11.h/Z, p9.h[w14, #1]\n"
-      ".inst 0xe0560f01  // ld1h { za0h.h[x12, #1] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0560ae3  // ld1h { za0h.h[x12, #3] }, p2/Z, [x23, x22, LSL #1]\n"
+      ".inst 0x25396140  // psel p0.h, p8.h/Z, p10.h[w13, #1]\n"
+      ".inst 0xe0562321  // ld1h { za0h.h[x13, #1] }, p0/Z, [x25, x22, LSL #1]\n"
+      ".inst 0x25796141  // psel p1.h, p8.h/Z, p10.h[w13, #3]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe05626e3  // ld1h { za0h.h[x13, #3] }, p1/Z, [x23, x22, LSL #1]\n"
       "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0aa82a1  // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "add x26, x26, #0x10\n"
-      "add x12, x12, #0x4\n"
-      ".inst 0xe0bfc6a0  // st1w { za0v.s[x14] }, p1/Z, [x21, XZR, LSL #2]\n"
-      ".inst 0xe0aac2a1  // st1w { za0v.s[x14, #1] }, p0/Z, [x21, x10, LSL #2]\n"
-      "add x14, x14, #0x2\n"
       "addvl x21, x21, #2\n"
-      "cmp x14, x28\n"
+      "add x13, x13, #0x4\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25386143  // psel p3.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0x25786142  // psel p2.h, p8.h/Z, p10.h[w12, #3]\n"
-      ".inst 0x252a6d21  // psel p1.h, p11.h/Z, p9.h[w14]\n"
-      ".inst 0x253a6d20  // psel p0.h, p11.h/Z, p9.h[w14, #1]\n"
+      ".inst 0x25396140  // psel p0.h, p8.h/Z, p10.h[w13, #1]\n"
+      ".inst 0xe0562321  // ld1h { za0h.h[x13, #1] }, p0/Z, [x25, x22, LSL #1]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25796141  // psel p1.h, p8.h/Z, p10.h[w13, #3]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe05626e3  // ld1h { za0h.h[x13, #3] }, p1/Z, [x23, x22, LSL #1]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "whilelt p9.h, x27, %x[width]\n"
-      ".inst 0xe0560f01  // ld1h { za0h.h[x12, #1] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       "inch x27\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0560ae3  // ld1h { za0h.h[x12, #3] }, p2/Z, [x23, x22, LSL #1]\n"
-      "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
+      ".inst 0xe0aa82a1  // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n"
+      "addvl x21, x21, #2\n"
       "inch x22\n"
-      ".inst 0xe0bfc6a0  // st1w { za0v.s[x14] }, p1/Z, [x21, XZR, LSL #2]\n"
       "whilelt p8.h, x27, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      ".inst 0xe0aac2a1  // st1w { za0v.s[x14, #1] }, p0/Z, [x21, x10, LSL #2]\n"
-      "addvl x21, x21, #2\n"
       "cbz x28, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25296143  // psel p3.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25696142  // psel p2.h, p8.h/Z, p10.h[w13, #2]\n"
-      ".inst 0x25286d21  // psel p1.h, p11.h/Z, p9.h[w12]\n"
-      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
-      ".inst 0xe0562f00  // ld1h { za0h.h[x13] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0562ae2  // ld1h { za0h.h[x13, #2] }, p2/Z, [x23, x22, LSL #1]\n"
+      ".inst 0x25296140  // psel p0.h, p8.h/Z, p10.h[w13]\n"
+      ".inst 0xe0562320  // ld1h { za0h.h[x13] }, p0/Z, [x25, x22, LSL #1]\n"
+      ".inst 0x25696141  // psel p1.h, p8.h/Z, p10.h[w13, #2]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe05626e2  // ld1h { za0h.h[x13, #2] }, p1/Z, [x23, x22, LSL #1]\n"
       "ldr x23, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      "add x13, x13, #0x4\n"
-      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       ".inst 0xe0aa82a9  // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n"
       "add x12, x12, #0x2\n"
-      "addvl x21, x21, #2\n"
       "cmp x12, x28\n"
+      "add x26, x26, #0x10\n"
+      "addvl x21, x21, #2\n"
+      "add x13, x13, #0x4\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25296143  // psel p3.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25696142  // psel p2.h, p8.h/Z, p10.h[w13, #2]\n"
-      ".inst 0x25286d21  // psel p1.h, p11.h/Z, p9.h[w12]\n"
-      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
+      ".inst 0x25296140  // psel p0.h, p8.h/Z, p10.h[w13]\n"
+      ".inst 0xe0562320  // ld1h { za0h.h[x13] }, p0/Z, [x25, x22, LSL #1]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25696141  // psel p1.h, p8.h/Z, p10.h[w13, #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe05626e2  // ld1h { za0h.h[x13, #2] }, p1/Z, [x23, x22, LSL #1]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "whilelt p9.h, x27, %x[width]\n"
-      ".inst 0xe0562f00  // ld1h { za0h.h[x13] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       "subs x20, x20, #0x1\n"
-      "inch x27\n"
-      ".inst 0xe0562ae2  // ld1h { za0h.h[x13, #2] }, p2/Z, [x23, x22, LSL #1]\n"
-      "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "inch x22\n"
-      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
       ".inst 0xe0aa82a9  // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n"
       "addvl x21, x21, #2\n"
+      "inch x27\n"
+      "inch x22\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x25, 11f\n"
+      "cbnz x24, 11f\n"
       "mov x26, %x[in]\n"
       "whilelt p8.h, x27, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25396140  // psel p0.h, p8.h/Z, p10.h[w13, #1]\n"
-      ".inst 0xe0bf86a0  // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "ldr x20, [x26, #0x0]\n"
+      "add x12, x12, #0x1\n"
+      ".inst 0x25396140  // psel p0.h, p8.h/Z, p10.h[w13, #1]\n"
       "cmp x12, x10\n"
-      "add x26, x26, #0x8\n"
       ".inst 0xe0562281  // ld1h { za0h.h[x13, #1] }, p0/Z, [x20, x22, LSL #1]\n"
+      "add x26, x26, #0x8\n"
+      "addvl x21, x21, #1\n"
       "add x13, x13, #0x2\n"
       "blt 9b\n"
       "whilelt p9.h, x27, %x[width]\n"
@@ -189,11 +189,11 @@ void interleave_block<1, 2, VLType::SME, false>(
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: Second
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      "add x20, x20, #0x2\n"
       ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x11\n"
+      "addvl x21, x21, #1\n"
+      "add x20, x20, #0x2\n"
       "blt 10b\n"
       "whilelt p8.h, x27, %x[width]\n"
       "b 13f\n"
@@ -203,15 +203,15 @@ void interleave_block<1, 2, VLType::SME, false>(
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x11\n"
+      "addvl x21, x21, #1\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block2_fp16_fp16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block2_fp16_fp16.hpp
index 0e0e4e462c..982cfa6d40 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block2_fp16_fp16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block2_fp16_fp16.hpp
@@ -22,7 +22,7 @@
  * SOFTWARE.
  */
 
-#if defined(ARM_COMPUTE_ENABLE_SME)
+#if defined(__ARM_FEATURE_SVE)
 
 template <>
 void interleave_block<1, 2, VLType::SME, false>(
@@ -63,11 +63,11 @@ void interleave_block<1, 2, VLType::SME, false>(
       "mov x12, #0x0\n"
       "cbz x28, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25286141  // psel p1.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25686140  // psel p0.h, p8.h/Z, p10.h[w12, #2]\n"
-      ".inst 0xe0560700  // ld1h { za0h.h[x12] }, p1/Z, [x24, x22, LSL #1]\n"
+      ".inst 0x25286143  // psel p3.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0x25686142  // psel p2.h, p8.h/Z, p10.h[w12, #2]\n"
+      ".inst 0xe0560f00  // ld1h { za0h.h[x12] }, p3/Z, [x24, x22, LSL #1]\n"
       "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe05602e2  // ld1h { za0h.h[x12, #2] }, p0/Z, [x23, x22, LSL #1]\n"
+      ".inst 0xe0560ae2  // ld1h { za0h.h[x12, #2] }, p2/Z, [x23, x22, LSL #1]\n"
       "add x12, x12, #0x4\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
@@ -88,20 +88,20 @@ void interleave_block<1, 2, VLType::SME, false>(
       "mov x20, x9\n"
       "3:"  // K loop: Main loop
       "whilelt p8.h, x27, %x[width]\n"
-      "mov x12, #0x0\n"
+      "mov x15, #0x0\n"
       "mov x14, #0x0\n"
       "cbz x28, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25386143  // psel p3.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0x25786142  // psel p2.h, p8.h/Z, p10.h[w12, #3]\n"
+      ".inst 0x253b6143  // psel p3.h, p8.h/Z, p10.h[w15, #1]\n"
+      ".inst 0x257b6142  // psel p2.h, p8.h/Z, p10.h[w15, #3]\n"
       ".inst 0x252a6d21  // psel p1.h, p11.h/Z, p9.h[w14]\n"
       ".inst 0x253a6d20  // psel p0.h, p11.h/Z, p9.h[w14, #1]\n"
-      ".inst 0xe0560f01  // ld1h { za0h.h[x12, #1] }, p3/Z, [x24, x22, LSL #1]\n"
+      ".inst 0xe0566f01  // ld1h { za0h.h[x15, #1] }, p3/Z, [x24, x22, LSL #1]\n"
       "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0560ae3  // ld1h { za0h.h[x12, #3] }, p2/Z, [x23, x22, LSL #1]\n"
+      ".inst 0xe0566ae3  // ld1h { za0h.h[x15, #3] }, p2/Z, [x23, x22, LSL #1]\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "add x12, x12, #0x4\n"
+      "add x15, x15, #0x4\n"
       ".inst 0xe0bfc6a0  // st1w { za0v.s[x14] }, p1/Z, [x21, XZR, LSL #2]\n"
       ".inst 0xe0aac2a1  // st1w { za0v.s[x14, #1] }, p0/Z, [x21, x10, LSL #2]\n"
       "add x14, x14, #0x2\n"
@@ -109,23 +109,23 @@ void interleave_block<1, 2, VLType::SME, false>(
       "cmp x14, x28\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25386143  // psel p3.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0x25786142  // psel p2.h, p8.h/Z, p10.h[w12, #3]\n"
+      ".inst 0x253b6143  // psel p3.h, p8.h/Z, p10.h[w15, #1]\n"
+      ".inst 0x257b6142  // psel p2.h, p8.h/Z, p10.h[w15, #3]\n"
       ".inst 0x252a6d21  // psel p1.h, p11.h/Z, p9.h[w14]\n"
       ".inst 0x253a6d20  // psel p0.h, p11.h/Z, p9.h[w14, #1]\n"
       "mov x26, %x[in]\n"
       "whilelt p9.h, x27, %x[width]\n"
-      ".inst 0xe0560f01  // ld1h { za0h.h[x12, #1] }, p3/Z, [x24, x22, LSL #1]\n"
+      ".inst 0xe0566f01  // ld1h { za0h.h[x15, #1] }, p3/Z, [x24, x22, LSL #1]\n"
       "ldr x24, [x26, #0x0]\n"
       "inch x27\n"
       "mov x13, #0x0\n"
-      ".inst 0xe0560ae3  // ld1h { za0h.h[x12, #3] }, p2/Z, [x23, x22, LSL #1]\n"
+      "whilelt p8.h, x27, %x[width]\n"
+      "mov x12, #0x0\n"
+      ".inst 0xe0566ae3  // ld1h { za0h.h[x15, #3] }, p2/Z, [x23, x22, LSL #1]\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
       "inch x22\n"
       ".inst 0xe0bfc6a0  // st1w { za0v.s[x14] }, p1/Z, [x21, XZR, LSL #2]\n"
-      "whilelt p8.h, x27, %x[width]\n"
-      "mov x12, #0x0\n"
       ".inst 0xe0aac2a1  // st1w { za0v.s[x14, #1] }, p0/Z, [x21, x10, LSL #2]\n"
       "addvl x21, x21, #2\n"
       "cbz x28, 7f\n"
@@ -172,15 +172,15 @@ void interleave_block<1, 2, VLType::SME, false>(
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25396140  // psel p0.h, p8.h/Z, p10.h[w13, #1]\n"
-      ".inst 0xe0bf86a0  // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25396143  // psel p3.h, p8.h/Z, p10.h[w13, #1]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      "ldr x24, [x26, #0x0]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
-      "ldr x20, [x26, #0x0]\n"
-      "cmp x12, x10\n"
       "add x26, x26, #0x8\n"
-      ".inst 0xe0562281  // ld1h { za0h.h[x13, #1] }, p0/Z, [x20, x22, LSL #1]\n"
+      "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
+      ".inst 0xe0562f01  // ld1h { za0h.h[x13, #1] }, p3/Z, [x24, x22, LSL #1]\n"
       "add x13, x13, #0x2\n"
       "blt 9b\n"
       "whilelt p9.h, x27, %x[width]\n"
@@ -195,7 +195,7 @@ void interleave_block<1, 2, VLType::SME, false>(
       "addvl x21, x21, #1\n"
       "cmp x12, x11\n"
       "blt 10b\n"
-      "whilelt p8.h, x27, %x[width]\n"
+      "whilelt p9.h, x27, %x[width]\n"
       "b 13f\n"
       "11:"  // K loop: Tails: Odd
       "mov x12, #0x0\n"
@@ -203,16 +203,16 @@ void interleave_block<1, 2, VLType::SME, false>(
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x11\n"
+      "addvl x21, x21, #1\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
-#endif  // defined(ARM_COMPUTE_ENABLE_SME)
+#endif  // defined(__ARM_FEATURE_SVE)
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8.hpp
index 558a5d7637..79cd668a84 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8.hpp
@@ -32,16 +32,16 @@ void interleave_block<1, 4, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "cntb x21\n"
       "mov x23, %x[width]\n"
-      "mov x21, %x[width]\n"
-      "cntb x20\n"
       "incb x23\n"
-      "sub x10, x20, #0x1\n"
+      "mov x20, %x[width]\n"
+      "sub x10, x21, #0x1\n"
       "cntw x9\n"
       "sub x23, x23, #0x1\n"
-      "ands x10, x21, x10\n"
-      "udiv x23, x23, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x10, x10, x20, NE\n"
+      "ands x10, x20, x10\n"
+      "udiv x23, x23, x21\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x10, x10, x21, NE\n"
       "lsl x22, %x[height], #0x1\n"  // height * 2
       "lsl x21, x9, #0x1\n"
       "sub x20, x23, #0x1\n"
@@ -52,12 +52,12 @@ void interleave_block<1, 4, VLType::SME, false>(
       "mov x27, #0x0\n"
       "mov x26, %x[in]\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x25, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
+      "ldr x25, [x26, #0x0]\n"
+      "and x24, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "lsr x10, x10, #0x2\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x23, [x26, #0x8]\n"
       "ptrue p11.s\n"
       "zip1 p10.b, p9.b, p8.b\n"
-      "ldr x23, [x26, #0x8]\n"
       "mov x22, %x[row_offset]\n"
       "mov x21, %x[out]\n"
       "whilelt p9.b, x27, %x[width]\n"
@@ -66,124 +66,124 @@ void interleave_block<1, 4, VLType::SME, false>(
       "mov x12, #0x0\n"
       "cbz x28, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25246141  // psel p1.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0160320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x22]\n"
       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
-      ".inst 0xe0160700  // ld1b { za0h.b[x12] }, p1/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x25, [x26, #0x0]\n"
       ".inst 0xe01602e4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x23, x22]\n"
       "add x12, x12, #0x8\n"
+      "cmp x12, x28, LSL #2\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "cmp x12, x28, LSL #2\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25246141  // psel p1.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0160320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x22]\n"
       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
       "mov x26, %x[in]\n"
-      "incb x27\n"
-      ".inst 0xe0160700  // ld1b { za0h.b[x12] }, p1/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
       ".inst 0xe01602e4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x23, x22]\n"
+      "ldr x25, [x26, #0x0]\n"
+      "incb x22\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "incb x22\n"
+      "incb x27\n"
       "cbz x20, 8f\n"
       "mov x20, x20\n"
       "3:"  // K loop: Main loop
       "whilelt p8.b, x27, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      "mov x14, #0x0\n"
       "cbz x28, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25346143  // psel p3.b, p8.b/Z, p10.b[w12, #2]\n"
-      ".inst 0x25746142  // psel p2.b, p8.b/Z, p10.b[w12, #6]\n"
-      ".inst 0x25266d21  // psel p1.b, p11.b/Z, p9.b[w14]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      ".inst 0xe0160f02  // ld1b { za0h.b[x12, #2] }, p3/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0160ae6  // ld1b { za0h.b[x12, #6] }, p2/Z, [x23, x22]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162322  // ld1b { za0h.b[x13, #2] }, p0/Z, [x25, x22]\n"
+      ".inst 0x25756141  // psel p1.b, p8.b/Z, p10.b[w13, #6]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e6  // ld1b { za0h.b[x13, #6] }, p1/Z, [x23, x22]\n"
       "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0a982a1  // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x9, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "add x26, x26, #0x10\n"
-      "add x12, x12, #0x8\n"
-      ".inst 0xe0bfc6a0  // st1w { za0v.s[x14] }, p1/Z, [x21, XZR, LSL #2]\n"
-      ".inst 0xe0a9c2a1  // st1w { za0v.s[x14, #1] }, p0/Z, [x21, x9, LSL #2]\n"
-      "add x14, x14, #0x2\n"
       "addvl x21, x21, #2\n"
-      "cmp x14, x28\n"
+      "add x13, x13, #0x8\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25346143  // psel p3.b, p8.b/Z, p10.b[w12, #2]\n"
-      ".inst 0x25746142  // psel p2.b, p8.b/Z, p10.b[w12, #6]\n"
-      ".inst 0x25266d21  // psel p1.b, p11.b/Z, p9.b[w14]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162322  // ld1b { za0h.b[x13, #2] }, p0/Z, [x25, x22]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25756141  // psel p1.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e6  // ld1b { za0h.b[x13, #6] }, p1/Z, [x23, x22]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "whilelt p9.b, x27, %x[width]\n"
-      ".inst 0xe0160f02  // ld1b { za0h.b[x12, #2] }, p3/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
       "incb x27\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0160ae6  // ld1b { za0h.b[x12, #6] }, p2/Z, [x23, x22]\n"
-      "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
+      ".inst 0xe0a982a1  // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x9, LSL #2]\n"
+      "addvl x21, x21, #2\n"
       "incb x22\n"
-      ".inst 0xe0bfc6a0  // st1w { za0v.s[x14] }, p1/Z, [x21, XZR, LSL #2]\n"
       "whilelt p8.b, x27, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      ".inst 0xe0a9c2a1  // st1w { za0v.s[x14, #1] }, p0/Z, [x21, x9, LSL #2]\n"
-      "addvl x21, x21, #2\n"
       "cbz x28, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x25656142  // psel p2.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x25246d21  // psel p1.b, p11.b/Z, p9.b[w12]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      ".inst 0xe0162f00  // ld1b { za0h.b[x13] }, p3/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0162ae4  // ld1b { za0h.b[x13, #4] }, p2/Z, [x23, x22]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe0162320  // ld1b { za0h.b[x13] }, p0/Z, [x25, x22]\n"
+      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x23, x22]\n"
       "ldr x23, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      "add x13, x13, #0x8\n"
-      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       ".inst 0xe0a982a9  // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x9, LSL #2]\n"
       "add x12, x12, #0x2\n"
-      "addvl x21, x21, #2\n"
       "cmp x12, x28\n"
+      "add x26, x26, #0x10\n"
+      "addvl x21, x21, #2\n"
+      "add x13, x13, #0x8\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x25656142  // psel p2.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x25246d21  // psel p1.b, p11.b/Z, p9.b[w12]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe0162320  // ld1b { za0h.b[x13] }, p0/Z, [x25, x22]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x23, x22]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "whilelt p9.b, x27, %x[width]\n"
-      ".inst 0xe0162f00  // ld1b { za0h.b[x13] }, p3/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
       "subs x20, x20, #0x1\n"
-      "incb x27\n"
-      ".inst 0xe0162ae4  // ld1b { za0h.b[x13, #4] }, p2/Z, [x23, x22]\n"
-      "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "incb x22\n"
-      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
       ".inst 0xe0a982a9  // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x9, LSL #2]\n"
       "addvl x21, x21, #2\n"
+      "incb x27\n"
+      "incb x22\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x25, 11f\n"
+      "cbnz x24, 11f\n"
       "mov x26, %x[in]\n"
       "whilelt p8.b, x27, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0xe0bf86a0  // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "ldr x20, [x26, #0x0]\n"
+      "add x12, x12, #0x1\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
       "cmp x12, x9\n"
-      "add x26, x26, #0x8\n"
       ".inst 0xe0162282  // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x22]\n"
+      "add x26, x26, #0x8\n"
+      "addvl x21, x21, #1\n"
       "add x13, x13, #0x4\n"
       "blt 9b\n"
       "whilelt p9.b, x27, %x[width]\n"
@@ -192,11 +192,11 @@ void interleave_block<1, 4, VLType::SME, false>(
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: Second
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      "add x20, x20, #0x4\n"
       ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
+      "add x20, x20, #0x4\n"
       "blt 10b\n"
       "whilelt p8.b, x27, %x[width]\n"
       "b 13f\n"
@@ -206,15 +206,15 @@ void interleave_block<1, 4, VLType::SME, false>(
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp
index ba8be81ade..fe98bd86b5 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp
@@ -32,18 +32,18 @@ void interleave_block<1, 4, VLType::SME, true>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "cntb x21\n"
       "mov x23, %x[width]\n"
-      "mov x21, %x[width]\n"
-      "cntb x20\n"
-      "incb x23\n"
       "mov z18.b, #0x1\n"
+      "incb x23\n"
+      "mov x20, %x[width]\n"
       "mov z17.s, #0x0\n"
-      "sub x10, x20, #0x1\n"
+      "sub x10, x21, #0x1\n"
       "cntw x9\n"
       "sub x23, x23, #0x1\n"
-      "ands x10, x21, x10\n"
-      "udiv x23, x23, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x10, x10, x20, NE\n"
+      "ands x10, x20, x10\n"
+      "udiv x23, x23, x21\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x10, x10, x21, NE\n"
       "lsl x22, %x[height], #0x1\n"  // height * 2
       "lsl x21, x9, #0x1\n"
       "sub x20, x23, #0x1\n"
@@ -51,7 +51,7 @@ void interleave_block<1, 4, VLType::SME, true>(
       "whilelt p9.b, XZR, x22\n"
       "whilelt p8.b, x21, x22\n"
       "mov x28, #0x0\n"
-      "ptrue p4.b\n"
+      "ptrue p2.b\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
       "and x27, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "lsr x10, x10, #0x2\n"
@@ -64,132 +64,132 @@ void interleave_block<1, 4, VLType::SME, true>(
       "whilelt p8.b, x28, %x[width]\n"
       "cbnz %x[first], 1f\n"
       "addvl x24, x24, #-1\n"
-      "ld1w { z17.s }, p4/Z, [x24]\n"
+      "ld1w { z17.s }, p2/Z, [x24]\n"
       "1:"  // K loop: Load row sums: End
       "mov x23, %x[in]\n"
-      "mov x12, #0x0\n"
       "ldr x22, [x23, #0x0]\n"
+      "mov x12, #0x0\n"
       "ldr x21, [x23, #0x8]\n"
       "add x23, x23, #0x10\n"
       "cbz x26, 3f\n"
       "2:"  // K loop: Charge: Loop
-      ".inst 0x25246141  // psel p1.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe01902c0  // ld1b { za0h.b[x12] }, p0/Z, [x22, x25]\n"
       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
-      ".inst 0xe01906c0  // ld1b { za0h.b[x12] }, p1/Z, [x22, x25]\n"
       "ldr x22, [x23, #0x0]\n"
       ".inst 0xe01902a4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n"
       "add x12, x12, #0x8\n"
+      "cmp x12, x26, LSL #2\n"
       "ldr x21, [x23, #0x8]\n"
       "add x23, x23, #0x10\n"
-      "cmp x12, x26, LSL #2\n"
       "blt 2b\n"
       "3:"  // K loop: Charge: End
-      ".inst 0x25246141  // psel p1.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe01902c0  // ld1b { za0h.b[x12] }, p0/Z, [x22, x25]\n"
       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
       "mov x23, %x[in]\n"
-      "incb x28\n"
-      ".inst 0xe01906c0  // ld1b { za0h.b[x12] }, p1/Z, [x22, x25]\n"
-      "ldr x22, [x23, #0x0]\n"
       ".inst 0xe01902a4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n"
+      "ldr x22, [x23, #0x0]\n"
+      "incb x25\n"
       "ldr x21, [x23, #0x8]\n"
       "add x23, x23, #0x10\n"
-      "incb x25\n"
+      "incb x28\n"
       "cbz x20, 9f\n"
       "mov x20, x20\n"
       "4:"  // K loop: Main loop
       "whilelt p8.b, x28, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      "mov x14, #0x0\n"
       "cbz x26, 6f\n"
       "5:"  // K loop: Main loop: First: Loop
-      ".inst 0x25346143  // psel p3.b, p8.b/Z, p10.b[w12, #2]\n"
-      ".inst 0x25746142  // psel p2.b, p8.b/Z, p10.b[w12, #6]\n"
-      ".inst 0x25266d21  // psel p1.b, p11.b/Z, p9.b[w14]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      ".inst 0xe0190ec2  // ld1b { za0h.b[x12, #2] }, p3/Z, [x22, x25]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe01922c2  // ld1b { za0h.b[x13, #2] }, p0/Z, [x22, x25]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
       "ldr x22, [x23, #0x0]\n"
-      ".inst 0xe0190aa6  // ld1b { za0h.b[x12, #6] }, p2/Z, [x21, x25]\n"
+      ".inst 0xe01922a6  // ld1b { za0h.b[x13, #6] }, p0/Z, [x21, x25]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      "sdot z17.s, z16.b, z18.b\n"
       "ldr x21, [x23, #0x8]\n"
-      "add x23, x23, #0x10\n"
-      "add x12, x12, #0x8\n"
-      ".inst 0xc082d010  // mova z16.s, p4/M, za0v.s[x14]\n"
-      ".inst 0xe0bfc700  // st1w { za0v.s[x14] }, p1/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0828830  // mova z16.s, p2/M, za0v.s[x12, #1]\n"
+      ".inst 0xe0a98301  // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x26\n"
       "sdot z17.s, z16.b, z18.b\n"
-      ".inst 0xe0a9c301  // st1w { za0v.s[x14, #1] }, p0/Z, [x24, x9, LSL #2]\n"
-      ".inst 0xc082d030  // mova z16.s, p4/M, za0v.s[x14, #1]\n"
-      "add x14, x14, #0x2\n"
-      "cmp x14, x26\n"
+      "add x23, x23, #0x10\n"
       "addvl x24, x24, #2\n"
-      "sdot z17.s, z16.b, z18.b\n"
+      "add x13, x13, #0x8\n"
       "blt 5b\n"
       "6:"  // K loop: Main loop: First: Tail
-      ".inst 0x25346143  // psel p3.b, p8.b/Z, p10.b[w12, #2]\n"
-      ".inst 0x25746142  // psel p2.b, p8.b/Z, p10.b[w12, #6]\n"
-      ".inst 0x25266d21  // psel p1.b, p11.b/Z, p9.b[w14]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe01922c2  // ld1b { za0h.b[x13, #2] }, p0/Z, [x22, x25]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0xe01922a6  // ld1b { za0h.b[x13, #6] }, p0/Z, [x21, x25]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
+      "sdot z17.s, z16.b, z18.b\n"
       "mov x23, %x[in]\n"
-      "whilelt p9.b, x28, %x[width]\n"
-      ".inst 0xe0190ec2  // ld1b { za0h.b[x12, #2] }, p3/Z, [x22, x25]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x23, #0x0]\n"
-      "incb x28\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0190aa6  // ld1b { za0h.b[x12, #6] }, p2/Z, [x21, x25]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0828830  // mova z16.s, p2/M, za0v.s[x12, #1]\n"
       "ldr x21, [x23, #0x8]\n"
+      ".inst 0xe0bf8700  // st1w { za0v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
+      "whilelt p9.b, x28, %x[width]\n"
+      "incb x28\n"
       "add x23, x23, #0x10\n"
+      ".inst 0xe0a98301  // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
+      "sdot z17.s, z16.b, z18.b\n"
+      "addvl x24, x24, #2\n"
       "incb x25\n"
-      ".inst 0xc082d010  // mova z16.s, p4/M, za0v.s[x14]\n"
-      ".inst 0xe0bfc700  // st1w { za0v.s[x14] }, p1/Z, [x24, XZR, LSL #2]\n"
       "whilelt p8.b, x28, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      "sdot z17.s, z16.b, z18.b\n"
-      ".inst 0xc082d030  // mova z16.s, p4/M, za0v.s[x14, #1]\n"
-      ".inst 0xe0a9c301  // st1w { za0v.s[x14, #1] }, p0/Z, [x24, x9, LSL #2]\n"
-      "addvl x24, x24, #2\n"
-      "sdot z17.s, z16.b, z18.b\n"
       "cbz x26, 8f\n"
       "7:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x25656142  // psel p2.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x25246d21  // psel p1.b, p11.b/Z, p9.b[w12]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      ".inst 0xe0192ec0  // ld1b { za0h.b[x13] }, p3/Z, [x22, x25]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe01922c0  // ld1b { za0h.b[x13] }, p0/Z, [x22, x25]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
       "ldr x22, [x23, #0x0]\n"
-      ".inst 0xe0192aa4  // ld1b { za0h.b[x13, #4] }, p2/Z, [x21, x25]\n"
-      "ldr x21, [x23, #0x8]\n"
-      "add x23, x23, #0x10\n"
-      "add x13, x13, #0x8\n"
-      ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
-      ".inst 0xe0bf8708  // st1w { za2v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0xe01922a4  // ld1b { za0h.b[x13, #4] }, p0/Z, [x21, x25]\n"
+      ".inst 0xc0828910  // mova z16.s, p2/M, za2v.s[x12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       "sdot z17.s, z16.b, z18.b\n"
+      "ldr x21, [x23, #0x8]\n"
+      ".inst 0xe0bf8308  // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0828930  // mova z16.s, p2/M, za2v.s[x12, #1]\n"
       ".inst 0xe0a98309  // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
-      ".inst 0xc0829130  // mova z16.s, p4/M, za2v.s[x12, #1]\n"
       "add x12, x12, #0x2\n"
       "cmp x12, x26\n"
-      "addvl x24, x24, #2\n"
       "sdot z17.s, z16.b, z18.b\n"
+      "add x23, x23, #0x10\n"
+      "addvl x24, x24, #2\n"
+      "add x13, x13, #0x8\n"
       "blt 7b\n"
       "8:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x25656142  // psel p2.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x25246d21  // psel p1.b, p11.b/Z, p9.b[w12]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe01922c0  // ld1b { za0h.b[x13] }, p0/Z, [x22, x25]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
+      ".inst 0xe01922a4  // ld1b { za0h.b[x13, #4] }, p0/Z, [x21, x25]\n"
+      ".inst 0xc0828910  // mova z16.s, p2/M, za2v.s[x12]\n"
+      "sdot z17.s, z16.b, z18.b\n"
       "mov x23, %x[in]\n"
-      "whilelt p9.b, x28, %x[width]\n"
-      ".inst 0xe0192ec0  // ld1b { za0h.b[x13] }, p3/Z, [x22, x25]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x23, #0x0]\n"
-      "subs x20, x20, #0x1\n"
-      "incb x28\n"
-      ".inst 0xe0192aa4  // ld1b { za0h.b[x13, #4] }, p2/Z, [x21, x25]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0828930  // mova z16.s, p2/M, za2v.s[x12, #1]\n"
       "ldr x21, [x23, #0x8]\n"
-      "add x23, x23, #0x10\n"
-      "incb x25\n"
-      ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
       ".inst 0xe0bf8708  // st1w { za2v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
-      "sdot z17.s, z16.b, z18.b\n"
-      ".inst 0xc0829130  // mova z16.s, p4/M, za2v.s[x12, #1]\n"
+      "whilelt p9.b, x28, %x[width]\n"
+      "subs x20, x20, #0x1\n"
+      "add x23, x23, #0x10\n"
       ".inst 0xe0a98309  // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
-      "addvl x24, x24, #2\n"
       "sdot z17.s, z16.b, z18.b\n"
+      "addvl x24, x24, #2\n"
+      "incb x28\n"
+      "incb x25\n"
       "bgt 4b\n"
       "9:"  // K loop: Tails
       "cbnz x27, 12f\n"
@@ -198,17 +198,17 @@ void interleave_block<1, 4, VLType::SME, true>(
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: First
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      "ldr x20, [x23, #0x0]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
+      "add x12, x12, #0x1\n"
       ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0xc0829010  // mova z16.s, p4/M, za0v.s[x12]\n"
       "sdot z17.s, z16.b, z18.b\n"
-      ".inst 0xe0bf8700  // st1w { za0v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x24, x24, #1\n"
-      "ldr x20, [x23, #0x0]\n"
+      ".inst 0xe0192282  // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x25]\n"
       "cmp x12, x9\n"
       "add x23, x23, #0x8\n"
-      ".inst 0xe0192282  // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x25]\n"
+      "addvl x24, x24, #1\n"
       "add x13, x13, #0x4\n"
       "blt 10b\n"
       "whilelt p9.b, x28, %x[width]\n"
@@ -217,13 +217,13 @@ void interleave_block<1, 4, VLType::SME, true>(
       "mov x12, #0x0\n"
       "11:"  // K loop: Tails: Even: Second
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
-      "add x20, x20, #0x4\n"
-      "sdot z17.s, z16.b, z18.b\n"
       ".inst 0xe0bf8308  // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0xc0828910  // mova z16.s, p2/M, za2v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "addvl x24, x24, #1\n"
       "cmp x12, x10\n"
+      "sdot z17.s, z16.b, z18.b\n"
+      "addvl x24, x24, #1\n"
+      "add x20, x20, #0x4\n"
       "blt 11b\n"
       "whilelt p8.b, x28, %x[width]\n"
       "b 14f\n"
@@ -231,21 +231,21 @@ void interleave_block<1, 4, VLType::SME, true>(
       "mov x12, #0x0\n"
       "13:"  // K loop: Tails: Odd: Loop
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0829010  // mova z16.s, p4/M, za0v.s[x12]\n"
-      "sdot z17.s, z16.b, z18.b\n"
       ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "addvl x24, x24, #1\n"
       "cmp x12, x10\n"
+      "sdot z17.s, z16.b, z18.b\n"
+      "addvl x24, x24, #1\n"
       "blt 13b\n"
       "14:"  // K loop: End
-      "st1w { z17.s }, p4, [x24]\n"
+      "st1w { z17.s }, p2, [x24]\n"
       "addvl x24, x24, #1\n"
-      ".inst 0xd503467f  // SMSTOP\n"
       "mov %x[out], x24\n"
+      ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [first] "r" (first), [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp
index 6d1c1a207f..bc7d013798 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8.hpp
@@ -32,16 +32,16 @@ void interleave_block<1, 4, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "cntb x21\n"
       "mov x23, %x[width]\n"
-      "mov x21, %x[width]\n"
-      "cntb x20\n"
       "incb x23\n"
-      "sub x10, x20, #0x1\n"
+      "mov x20, %x[width]\n"
+      "sub x10, x21, #0x1\n"
       "cntw x9\n"
       "sub x23, x23, #0x1\n"
-      "ands x10, x21, x10\n"
-      "udiv x23, x23, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x10, x10, x20, NE\n"
+      "ands x10, x20, x10\n"
+      "udiv x23, x23, x21\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x10, x10, x21, NE\n"
       "lsl x22, %x[height], #0x1\n"  // height * 2
       "lsl x21, x9, #0x1\n"
       "sub x20, x23, #0x1\n"
@@ -52,12 +52,12 @@ void interleave_block<1, 4, VLType::SME, false>(
       "mov x27, #0x0\n"
       "mov x26, %x[in]\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x25, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
+      "ldr x25, [x26, #0x0]\n"
+      "and x24, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "lsr x10, x10, #0x2\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x23, [x26, #0x8]\n"
       "ptrue p11.s\n"
       "zip1 p10.b, p9.b, p8.b\n"
-      "ldr x23, [x26, #0x8]\n"
       "mov x22, %x[row_offset]\n"
       "mov x21, %x[out]\n"
       "whilelt p9.b, x27, %x[width]\n"
@@ -66,124 +66,124 @@ void interleave_block<1, 4, VLType::SME, false>(
       "mov x12, #0x0\n"
       "cbz x28, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25246141  // psel p1.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0160320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x22]\n"
       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
-      ".inst 0xe0160700  // ld1b { za0h.b[x12] }, p1/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x25, [x26, #0x0]\n"
       ".inst 0xe01602e4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x23, x22]\n"
       "add x12, x12, #0x8\n"
+      "cmp x12, x28, LSL #2\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "cmp x12, x28, LSL #2\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25246141  // psel p1.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0160320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x22]\n"
       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
       "mov x26, %x[in]\n"
-      "incb x27\n"
-      ".inst 0xe0160700  // ld1b { za0h.b[x12] }, p1/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
       ".inst 0xe01602e4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x23, x22]\n"
+      "ldr x25, [x26, #0x0]\n"
+      "incb x22\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "incb x22\n"
+      "incb x27\n"
       "cbz x20, 8f\n"
       "mov x20, x20\n"
       "3:"  // K loop: Main loop
       "whilelt p8.b, x27, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      "mov x14, #0x0\n"
       "cbz x28, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25346143  // psel p3.b, p8.b/Z, p10.b[w12, #2]\n"
-      ".inst 0x25746142  // psel p2.b, p8.b/Z, p10.b[w12, #6]\n"
-      ".inst 0x25266d21  // psel p1.b, p11.b/Z, p9.b[w14]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      ".inst 0xe0160f02  // ld1b { za0h.b[x12, #2] }, p3/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0160ae6  // ld1b { za0h.b[x12, #6] }, p2/Z, [x23, x22]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162322  // ld1b { za0h.b[x13, #2] }, p0/Z, [x25, x22]\n"
+      ".inst 0x25756141  // psel p1.b, p8.b/Z, p10.b[w13, #6]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e6  // ld1b { za0h.b[x13, #6] }, p1/Z, [x23, x22]\n"
       "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0a982a1  // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x9, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "add x26, x26, #0x10\n"
-      "add x12, x12, #0x8\n"
-      ".inst 0xe0bfc6a0  // st1w { za0v.s[x14] }, p1/Z, [x21, XZR, LSL #2]\n"
-      ".inst 0xe0a9c2a1  // st1w { za0v.s[x14, #1] }, p0/Z, [x21, x9, LSL #2]\n"
-      "add x14, x14, #0x2\n"
       "addvl x21, x21, #2\n"
-      "cmp x14, x28\n"
+      "add x13, x13, #0x8\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25346143  // psel p3.b, p8.b/Z, p10.b[w12, #2]\n"
-      ".inst 0x25746142  // psel p2.b, p8.b/Z, p10.b[w12, #6]\n"
-      ".inst 0x25266d21  // psel p1.b, p11.b/Z, p9.b[w14]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162322  // ld1b { za0h.b[x13, #2] }, p0/Z, [x25, x22]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25756141  // psel p1.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e6  // ld1b { za0h.b[x13, #6] }, p1/Z, [x23, x22]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "whilelt p9.b, x27, %x[width]\n"
-      ".inst 0xe0160f02  // ld1b { za0h.b[x12, #2] }, p3/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
       "incb x27\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0160ae6  // ld1b { za0h.b[x12, #6] }, p2/Z, [x23, x22]\n"
-      "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
+      ".inst 0xe0a982a1  // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x9, LSL #2]\n"
+      "addvl x21, x21, #2\n"
       "incb x22\n"
-      ".inst 0xe0bfc6a0  // st1w { za0v.s[x14] }, p1/Z, [x21, XZR, LSL #2]\n"
       "whilelt p8.b, x27, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      ".inst 0xe0a9c2a1  // st1w { za0v.s[x14, #1] }, p0/Z, [x21, x9, LSL #2]\n"
-      "addvl x21, x21, #2\n"
       "cbz x28, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x25656142  // psel p2.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x25246d21  // psel p1.b, p11.b/Z, p9.b[w12]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      ".inst 0xe0162f00  // ld1b { za0h.b[x13] }, p3/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0162ae4  // ld1b { za0h.b[x13, #4] }, p2/Z, [x23, x22]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe0162320  // ld1b { za0h.b[x13] }, p0/Z, [x25, x22]\n"
+      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x23, x22]\n"
       "ldr x23, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      "add x13, x13, #0x8\n"
-      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       ".inst 0xe0a982a9  // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x9, LSL #2]\n"
       "add x12, x12, #0x2\n"
-      "addvl x21, x21, #2\n"
       "cmp x12, x28\n"
+      "add x26, x26, #0x10\n"
+      "addvl x21, x21, #2\n"
+      "add x13, x13, #0x8\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x25656142  // psel p2.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x25246d21  // psel p1.b, p11.b/Z, p9.b[w12]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe0162320  // ld1b { za0h.b[x13] }, p0/Z, [x25, x22]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x23, x22]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "whilelt p9.b, x27, %x[width]\n"
-      ".inst 0xe0162f00  // ld1b { za0h.b[x13] }, p3/Z, [x24, x22]\n"
-      "ldr x24, [x26, #0x0]\n"
       "subs x20, x20, #0x1\n"
-      "incb x27\n"
-      ".inst 0xe0162ae4  // ld1b { za0h.b[x13, #4] }, p2/Z, [x23, x22]\n"
-      "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "incb x22\n"
-      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
       ".inst 0xe0a982a9  // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x9, LSL #2]\n"
       "addvl x21, x21, #2\n"
+      "incb x27\n"
+      "incb x22\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x25, 11f\n"
+      "cbnz x24, 11f\n"
       "mov x26, %x[in]\n"
       "whilelt p8.b, x27, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0xe0bf86a0  // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "ldr x20, [x26, #0x0]\n"
+      "add x12, x12, #0x1\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
       "cmp x12, x9\n"
-      "add x26, x26, #0x8\n"
       ".inst 0xe0162282  // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x22]\n"
+      "add x26, x26, #0x8\n"
+      "addvl x21, x21, #1\n"
       "add x13, x13, #0x4\n"
       "blt 9b\n"
       "whilelt p9.b, x27, %x[width]\n"
@@ -192,11 +192,11 @@ void interleave_block<1, 4, VLType::SME, false>(
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: Second
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      "add x20, x20, #0x4\n"
       ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
+      "add x20, x20, #0x4\n"
       "blt 10b\n"
       "whilelt p8.b, x27, %x[width]\n"
       "b 13f\n"
@@ -206,15 +206,15 @@ void interleave_block<1, 4, VLType::SME, false>(
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8_summing.hpp
index dbcd18678b..66fcd800d4 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8_summing.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_u8_u8_summing.hpp
@@ -32,18 +32,18 @@ void interleave_block<1, 4, VLType::SME, true>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "cntb x21\n"
       "mov x23, %x[width]\n"
-      "mov x21, %x[width]\n"
-      "cntb x20\n"
-      "incb x23\n"
       "mov z18.b, #0x1\n"
+      "incb x23\n"
+      "mov x20, %x[width]\n"
       "mov z17.s, #0x0\n"
-      "sub x10, x20, #0x1\n"
+      "sub x10, x21, #0x1\n"
       "cntw x9\n"
       "sub x23, x23, #0x1\n"
-      "ands x10, x21, x10\n"
-      "udiv x23, x23, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x10, x10, x20, NE\n"
+      "ands x10, x20, x10\n"
+      "udiv x23, x23, x21\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x10, x10, x21, NE\n"
       "lsl x22, %x[height], #0x1\n"  // height * 2
       "lsl x21, x9, #0x1\n"
       "sub x20, x23, #0x1\n"
@@ -51,7 +51,7 @@ void interleave_block<1, 4, VLType::SME, true>(
       "whilelt p9.b, XZR, x22\n"
       "whilelt p8.b, x21, x22\n"
       "mov x28, #0x0\n"
-      "ptrue p4.b\n"
+      "ptrue p2.b\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
       "and x27, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "lsr x10, x10, #0x2\n"
@@ -64,132 +64,132 @@ void interleave_block<1, 4, VLType::SME, true>(
       "whilelt p8.b, x28, %x[width]\n"
       "cbnz %x[first], 1f\n"
       "addvl x24, x24, #-1\n"
-      "ld1w { z17.s }, p4/Z, [x24]\n"
+      "ld1w { z17.s }, p2/Z, [x24]\n"
       "1:"  // K loop: Load row sums: End
       "mov x23, %x[in]\n"
-      "mov x12, #0x0\n"
       "ldr x22, [x23, #0x0]\n"
+      "mov x12, #0x0\n"
       "ldr x21, [x23, #0x8]\n"
       "add x23, x23, #0x10\n"
       "cbz x26, 3f\n"
       "2:"  // K loop: Charge: Loop
-      ".inst 0x25246141  // psel p1.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe01902c0  // ld1b { za0h.b[x12] }, p0/Z, [x22, x25]\n"
       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
-      ".inst 0xe01906c0  // ld1b { za0h.b[x12] }, p1/Z, [x22, x25]\n"
       "ldr x22, [x23, #0x0]\n"
       ".inst 0xe01902a4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n"
       "add x12, x12, #0x8\n"
+      "cmp x12, x26, LSL #2\n"
       "ldr x21, [x23, #0x8]\n"
       "add x23, x23, #0x10\n"
-      "cmp x12, x26, LSL #2\n"
       "blt 2b\n"
       "3:"  // K loop: Charge: End
-      ".inst 0x25246141  // psel p1.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe01902c0  // ld1b { za0h.b[x12] }, p0/Z, [x22, x25]\n"
       ".inst 0x25646140  // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
       "mov x23, %x[in]\n"
-      "incb x28\n"
-      ".inst 0xe01906c0  // ld1b { za0h.b[x12] }, p1/Z, [x22, x25]\n"
-      "ldr x22, [x23, #0x0]\n"
       ".inst 0xe01902a4  // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n"
+      "ldr x22, [x23, #0x0]\n"
+      "incb x25\n"
       "ldr x21, [x23, #0x8]\n"
       "add x23, x23, #0x10\n"
-      "incb x25\n"
+      "incb x28\n"
       "cbz x20, 9f\n"
       "mov x20, x20\n"
       "4:"  // K loop: Main loop
       "whilelt p8.b, x28, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      "mov x14, #0x0\n"
       "cbz x26, 6f\n"
       "5:"  // K loop: Main loop: First: Loop
-      ".inst 0x25346143  // psel p3.b, p8.b/Z, p10.b[w12, #2]\n"
-      ".inst 0x25746142  // psel p2.b, p8.b/Z, p10.b[w12, #6]\n"
-      ".inst 0x25266d21  // psel p1.b, p11.b/Z, p9.b[w14]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      ".inst 0xe0190ec2  // ld1b { za0h.b[x12, #2] }, p3/Z, [x22, x25]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe01922c2  // ld1b { za0h.b[x13, #2] }, p0/Z, [x22, x25]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
       "ldr x22, [x23, #0x0]\n"
-      ".inst 0xe0190aa6  // ld1b { za0h.b[x12, #6] }, p2/Z, [x21, x25]\n"
+      ".inst 0xe01922a6  // ld1b { za0h.b[x13, #6] }, p0/Z, [x21, x25]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      "udot z17.s, z16.b, z18.b\n"
       "ldr x21, [x23, #0x8]\n"
-      "add x23, x23, #0x10\n"
-      "add x12, x12, #0x8\n"
-      ".inst 0xc082d010  // mova z16.s, p4/M, za0v.s[x14]\n"
-      ".inst 0xe0bfc700  // st1w { za0v.s[x14] }, p1/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0828830  // mova z16.s, p2/M, za0v.s[x12, #1]\n"
+      ".inst 0xe0a98301  // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x26\n"
       "udot z17.s, z16.b, z18.b\n"
-      ".inst 0xe0a9c301  // st1w { za0v.s[x14, #1] }, p0/Z, [x24, x9, LSL #2]\n"
-      ".inst 0xc082d030  // mova z16.s, p4/M, za0v.s[x14, #1]\n"
-      "add x14, x14, #0x2\n"
-      "cmp x14, x26\n"
+      "add x23, x23, #0x10\n"
       "addvl x24, x24, #2\n"
-      "udot z17.s, z16.b, z18.b\n"
+      "add x13, x13, #0x8\n"
       "blt 5b\n"
       "6:"  // K loop: Main loop: First: Tail
-      ".inst 0x25346143  // psel p3.b, p8.b/Z, p10.b[w12, #2]\n"
-      ".inst 0x25746142  // psel p2.b, p8.b/Z, p10.b[w12, #6]\n"
-      ".inst 0x25266d21  // psel p1.b, p11.b/Z, p9.b[w14]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe01922c2  // ld1b { za0h.b[x13, #2] }, p0/Z, [x22, x25]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0xe01922a6  // ld1b { za0h.b[x13, #6] }, p0/Z, [x21, x25]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
+      "udot z17.s, z16.b, z18.b\n"
       "mov x23, %x[in]\n"
-      "whilelt p9.b, x28, %x[width]\n"
-      ".inst 0xe0190ec2  // ld1b { za0h.b[x12, #2] }, p3/Z, [x22, x25]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x23, #0x0]\n"
-      "incb x28\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0190aa6  // ld1b { za0h.b[x12, #6] }, p2/Z, [x21, x25]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0828830  // mova z16.s, p2/M, za0v.s[x12, #1]\n"
       "ldr x21, [x23, #0x8]\n"
+      ".inst 0xe0bf8700  // st1w { za0v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
+      "whilelt p9.b, x28, %x[width]\n"
+      "incb x28\n"
       "add x23, x23, #0x10\n"
+      ".inst 0xe0a98301  // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
+      "udot z17.s, z16.b, z18.b\n"
+      "addvl x24, x24, #2\n"
       "incb x25\n"
-      ".inst 0xc082d010  // mova z16.s, p4/M, za0v.s[x14]\n"
-      ".inst 0xe0bfc700  // st1w { za0v.s[x14] }, p1/Z, [x24, XZR, LSL #2]\n"
       "whilelt p8.b, x28, %x[width]\n"
+      "mov x13, #0x0\n"
       "mov x12, #0x0\n"
-      "udot z17.s, z16.b, z18.b\n"
-      ".inst 0xc082d030  // mova z16.s, p4/M, za0v.s[x14, #1]\n"
-      ".inst 0xe0a9c301  // st1w { za0v.s[x14, #1] }, p0/Z, [x24, x9, LSL #2]\n"
-      "addvl x24, x24, #2\n"
-      "udot z17.s, z16.b, z18.b\n"
       "cbz x26, 8f\n"
       "7:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x25656142  // psel p2.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x25246d21  // psel p1.b, p11.b/Z, p9.b[w12]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      ".inst 0xe0192ec0  // ld1b { za0h.b[x13] }, p3/Z, [x22, x25]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe01922c0  // ld1b { za0h.b[x13] }, p0/Z, [x22, x25]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
       "ldr x22, [x23, #0x0]\n"
-      ".inst 0xe0192aa4  // ld1b { za0h.b[x13, #4] }, p2/Z, [x21, x25]\n"
-      "ldr x21, [x23, #0x8]\n"
-      "add x23, x23, #0x10\n"
-      "add x13, x13, #0x8\n"
-      ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
-      ".inst 0xe0bf8708  // st1w { za2v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0xe01922a4  // ld1b { za0h.b[x13, #4] }, p0/Z, [x21, x25]\n"
+      ".inst 0xc0828910  // mova z16.s, p2/M, za2v.s[x12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       "udot z17.s, z16.b, z18.b\n"
+      "ldr x21, [x23, #0x8]\n"
+      ".inst 0xe0bf8308  // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0828930  // mova z16.s, p2/M, za2v.s[x12, #1]\n"
       ".inst 0xe0a98309  // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
-      ".inst 0xc0829130  // mova z16.s, p4/M, za2v.s[x12, #1]\n"
       "add x12, x12, #0x2\n"
       "cmp x12, x26\n"
-      "addvl x24, x24, #2\n"
       "udot z17.s, z16.b, z18.b\n"
+      "add x23, x23, #0x10\n"
+      "addvl x24, x24, #2\n"
+      "add x13, x13, #0x8\n"
       "blt 7b\n"
       "8:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x25656142  // psel p2.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x25246d21  // psel p1.b, p11.b/Z, p9.b[w12]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe01922c0  // ld1b { za0h.b[x13] }, p0/Z, [x22, x25]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
+      ".inst 0xe01922a4  // ld1b { za0h.b[x13, #4] }, p0/Z, [x21, x25]\n"
+      ".inst 0xc0828910  // mova z16.s, p2/M, za2v.s[x12]\n"
+      "udot z17.s, z16.b, z18.b\n"
       "mov x23, %x[in]\n"
-      "whilelt p9.b, x28, %x[width]\n"
-      ".inst 0xe0192ec0  // ld1b { za0h.b[x13] }, p3/Z, [x22, x25]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x23, #0x0]\n"
-      "subs x20, x20, #0x1\n"
-      "incb x28\n"
-      ".inst 0xe0192aa4  // ld1b { za0h.b[x13, #4] }, p2/Z, [x21, x25]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0828930  // mova z16.s, p2/M, za2v.s[x12, #1]\n"
       "ldr x21, [x23, #0x8]\n"
-      "add x23, x23, #0x10\n"
-      "incb x25\n"
-      ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
       ".inst 0xe0bf8708  // st1w { za2v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
-      "udot z17.s, z16.b, z18.b\n"
-      ".inst 0xc0829130  // mova z16.s, p4/M, za2v.s[x12, #1]\n"
+      "whilelt p9.b, x28, %x[width]\n"
+      "subs x20, x20, #0x1\n"
+      "add x23, x23, #0x10\n"
       ".inst 0xe0a98309  // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
-      "addvl x24, x24, #2\n"
       "udot z17.s, z16.b, z18.b\n"
+      "addvl x24, x24, #2\n"
+      "incb x28\n"
+      "incb x25\n"
       "bgt 4b\n"
       "9:"  // K loop: Tails
       "cbnz x27, 12f\n"
@@ -198,17 +198,17 @@ void interleave_block<1, 4, VLType::SME, true>(
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: First
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      "ldr x20, [x23, #0x0]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
+      "add x12, x12, #0x1\n"
       ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0xc0829010  // mova z16.s, p4/M, za0v.s[x12]\n"
       "udot z17.s, z16.b, z18.b\n"
-      ".inst 0xe0bf8700  // st1w { za0v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x24, x24, #1\n"
-      "ldr x20, [x23, #0x0]\n"
+      ".inst 0xe0192282  // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x25]\n"
       "cmp x12, x9\n"
       "add x23, x23, #0x8\n"
-      ".inst 0xe0192282  // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x25]\n"
+      "addvl x24, x24, #1\n"
       "add x13, x13, #0x4\n"
       "blt 10b\n"
       "whilelt p9.b, x28, %x[width]\n"
@@ -217,13 +217,13 @@ void interleave_block<1, 4, VLType::SME, true>(
       "mov x12, #0x0\n"
       "11:"  // K loop: Tails: Even: Second
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
-      "add x20, x20, #0x4\n"
-      "udot z17.s, z16.b, z18.b\n"
       ".inst 0xe0bf8308  // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0xc0828910  // mova z16.s, p2/M, za2v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "addvl x24, x24, #1\n"
       "cmp x12, x10\n"
+      "udot z17.s, z16.b, z18.b\n"
+      "addvl x24, x24, #1\n"
+      "add x20, x20, #0x4\n"
       "blt 11b\n"
       "whilelt p8.b, x28, %x[width]\n"
       "b 14f\n"
@@ -231,21 +231,21 @@ void interleave_block<1, 4, VLType::SME, true>(
       "mov x12, #0x0\n"
       "13:"  // K loop: Tails: Odd: Loop
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0829010  // mova z16.s, p4/M, za0v.s[x12]\n"
-      "udot z17.s, z16.b, z18.b\n"
       ".inst 0xe0bf8300  // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "addvl x24, x24, #1\n"
       "cmp x12, x10\n"
+      "udot z17.s, z16.b, z18.b\n"
+      "addvl x24, x24, #1\n"
       "blt 13b\n"
       "14:"  // K loop: End
-      "st1w { z17.s }, p4, [x24]\n"
+      "st1w { z17.s }, p2, [x24]\n"
       "addvl x24, x24, #1\n"
-      ".inst 0xd503467f  // SMSTOP\n"
       "mov %x[out], x24\n"
+      ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [first] "r" (first), [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp16_fp16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp16_fp16.hpp
index 591c08dcb2..fb0a74b3bd 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp16_fp16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp16_fp16.hpp
@@ -33,25 +33,25 @@ void interleave_block<1, 1, VLType::SME, false>(
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
       "mov x21, %x[width]\n"
-      "mov x20, %x[width]\n"
       "inch x21\n"
       "cnth x11\n"
       "sub x21, x21, #0x1\n"
-      "sub x10, x11, #0x1\n"
       "udiv x21, x21, x11\n"  // n_passes = ceildiv(width, VL<T>)
-      "ands x10, x20, x10\n"
+      "mov x20, %x[width]\n"
+      "sub x10, x11, #0x1\n"
       "sub x9, x21, #0x1\n"
+      "ands x10, x20, x10\n"
       "sub x28, x11, #0x2\n"
       "lsl x20, %x[height], #0x1\n"  // height * 2
       "mov x27, #0x0\n"
       "mov x26, %x[in]\n"
       "lsr x9, x9, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x25, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
+      "ldr x25, [x26, #0x0]\n"
+      "and x24, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "csel x10, x10, x11, NE\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x23, [x26, #0x8]\n"
       "ptrue p11.h\n"
       "whilelt p10.h, XZR, x20\n"
-      "ldr x23, [x26, #0x8]\n"
       "mov x22, %x[row_offset]\n"
       "mov x21, %x[out]\n"
       "whilelt p9.h, x27, %x[width]\n"
@@ -60,119 +60,119 @@ void interleave_block<1, 1, VLType::SME, false>(
       "mov x12, #0x0\n"
       "cbz x28, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25286141  // psel p1.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       ".inst 0x25386140  // psel p0.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0xe0560700  // ld1h { za0h.h[x12] }, p1/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x25, [x26, #0x0]\n"
       ".inst 0xe05602e1  // ld1h { za0h.h[x12, #1] }, p0/Z, [x23, x22, LSL #1]\n"
       "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "cmp x12, x28\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25286141  // psel p1.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       ".inst 0x25386140  // psel p0.h, p8.h/Z, p10.h[w12, #1]\n"
       "mov x26, %x[in]\n"
-      "inch x27\n"
-      ".inst 0xe0560700  // ld1h { za0h.h[x12] }, p1/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       ".inst 0xe05602e1  // ld1h { za0h.h[x12, #1] }, p0/Z, [x23, x22, LSL #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      "inch x22\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "inch x22\n"
+      "inch x27\n"
       "cbz x9, 8f\n"
       "mov x20, x9\n"
       "3:"  // K loop: Main loop
       "whilelt p8.h, x27, %x[width]\n"
-      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
       "cbz x28, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25296143  // psel p3.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25396142  // psel p2.h, p8.h/Z, p10.h[w13, #1]\n"
-      ".inst 0x25296d21  // psel p1.h, p11.h/Z, p9.h[w13]\n"
-      ".inst 0x25396d20  // psel p0.h, p11.h/Z, p9.h[w13, #1]\n"
-      ".inst 0xe0562f08  // ld1h { za1h.h[x13] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0562ae9  // ld1h { za1h.h[x13, #1] }, p2/Z, [x23, x22, LSL #1]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560328  // ld1h { za1h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
+      ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe05606e9  // ld1h { za1h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
       "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe07fa6a0  // st1h { za0v.h[x13] }, p1/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0xe07f82a0  // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
+      ".inst 0xe06b82a1  // st1h { za0v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "add x26, x26, #0x10\n"
-      ".inst 0xe06ba2a1  // st1h { za0v.h[x13, #1] }, p0/Z, [x21, x11, LSL #1]\n"
-      "add x13, x13, #0x2\n"
       "addvl x21, x21, #2\n"
-      "cmp x13, x28\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25296143  // psel p3.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25396142  // psel p2.h, p8.h/Z, p10.h[w13, #1]\n"
-      ".inst 0x25296d21  // psel p1.h, p11.h/Z, p9.h[w13]\n"
-      ".inst 0x25396d20  // psel p0.h, p11.h/Z, p9.h[w13, #1]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560328  // ld1h { za1h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe05606e9  // ld1h { za1h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe07f82a0  // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
       "whilelt p9.h, x27, %x[width]\n"
-      ".inst 0xe0562f08  // ld1h { za1h.h[x13] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       "inch x27\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0562ae9  // ld1h { za1h.h[x13, #1] }, p2/Z, [x23, x22, LSL #1]\n"
-      "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe07fa6a0  // st1h { za0v.h[x13] }, p1/Z, [x21, XZR, LSL #1]\n"
       "add x26, x26, #0x10\n"
-      ".inst 0xe06ba2a1  // st1h { za0v.h[x13, #1] }, p0/Z, [x21, x11, LSL #1]\n"
+      ".inst 0xe06b82a1  // st1h { za0v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
       "addvl x21, x21, #2\n"
       "inch x22\n"
       "whilelt p8.h, x27, %x[width]\n"
+      "mov x12, #0x0\n"
       "cbz x28, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25286143  // psel p3.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25386142  // psel p2.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0x25286d21  // psel p1.h, p11.h/Z, p9.h[w12]\n"
-      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
-      ".inst 0xe0560f00  // ld1h { za0h.h[x12] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0560ae1  // ld1h { za0h.h[x12, #1] }, p2/Z, [x23, x22, LSL #1]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
+      ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe05606e1  // ld1h { za0h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
       "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe07f86a8  // st1h { za1v.h[x12] }, p1/Z, [x21, XZR, LSL #1]\n"
-      "add x26, x26, #0x10\n"
+      ".inst 0xe07f82a8  // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
       ".inst 0xe06b82a9  // st1h { za1v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
       "add x12, x12, #0x2\n"
-      "addvl x21, x21, #2\n"
       "cmp x12, x28\n"
+      "add x26, x26, #0x10\n"
+      "addvl x21, x21, #2\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25286143  // psel p3.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25386142  // psel p2.h, p8.h/Z, p10.h[w12, #1]\n"
-      ".inst 0x25286d21  // psel p1.h, p11.h/Z, p9.h[w12]\n"
-      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
+      ".inst 0xe0560320  // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25386141  // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe05606e1  // ld1h { za0h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe07f82a8  // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
+      ".inst 0x25386d20  // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
       "whilelt p9.h, x27, %x[width]\n"
-      ".inst 0xe0560f00  // ld1h { za0h.h[x12] }, p3/Z, [x24, x22, LSL #1]\n"
-      "ldr x24, [x26, #0x0]\n"
       "subs x20, x20, #0x1\n"
-      "inch x27\n"
-      ".inst 0xe0560ae1  // ld1h { za0h.h[x12, #1] }, p2/Z, [x23, x22, LSL #1]\n"
-      "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe07f86a8  // st1h { za1v.h[x12] }, p1/Z, [x21, XZR, LSL #1]\n"
       "add x26, x26, #0x10\n"
       ".inst 0xe06b82a9  // st1h { za1v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
       "addvl x21, x21, #2\n"
+      "inch x27\n"
       "inch x22\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x25, 11f\n"
+      "cbnz x24, 11f\n"
       "mov x26, %x[in]\n"
       "whilelt p8.h, x27, %x[width]\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25286d21  // psel p1.h, p11.h/Z, p9.h[w12]\n"
-      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0xe07f86a0  // st1h { za0v.h[x12] }, p1/Z, [x21, XZR, LSL #1]\n"
-      "addvl x21, x21, #1\n"
+      ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
+      ".inst 0xe07f82a0  // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
       "ldr x20, [x26, #0x0]\n"
-      "add x26, x26, #0x8\n"
+      ".inst 0x25286140  // psel p0.h, p8.h/Z, p10.h[w12]\n"
       ".inst 0xe0560288  // ld1h { za1h.h[x12] }, p0/Z, [x20, x22, LSL #1]\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x11\n"
+      "add x26, x26, #0x8\n"
+      "addvl x21, x21, #1\n"
       "blt 9b\n"
       "whilelt p9.h, x27, %x[width]\n"
       "whilelt p8.h, x27, %x[width]\n"
@@ -181,8 +181,8 @@ void interleave_block<1, 1, VLType::SME, false>(
       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
       ".inst 0xe07f82a8  // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
       "blt 10b\n"
       "whilelt p8.h, x27, %x[width]\n"
       "b 13f\n"
@@ -192,15 +192,15 @@ void interleave_block<1, 1, VLType::SME, false>(
       ".inst 0x25286d20  // psel p0.h, p11.h/Z, p9.h[w12]\n"
       ".inst 0xe07f82a0  // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x10\n"
+      "addvl x21, x21, #1\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp
index b76ec57d22..3fe3885068 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp
@@ -32,24 +32,24 @@ void interleave_block<1, 1, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x21, %x[width]\n"
-      "mov x20, %x[width]\n"
-      "incw x21\n"
+      "mov x22, %x[width]\n"
+      "incw x22\n"
       "cntw x10\n"
-      "sub x21, x21, #0x1\n"
+      "sub x22, x22, #0x1\n"
+      "udiv x22, x22, x10\n"  // n_passes = ceildiv(width, VL<T>)
+      "mov x21, %x[width]\n"
       "sub x9, x10, #0x1\n"
-      "udiv x21, x21, x10\n"  // n_passes = ceildiv(width, VL<T>)
-      "ands x9, x20, x9\n"
-      "sub x20, x21, #0x1\n"
+      "sub x20, x22, #0x1\n"
+      "ands x9, x21, x9\n"
       "sub x28, x10, #0x2\n"
       "mov x27, #0x0\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x25, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "ldr x24, [x26, #0x0]\n"
+      "and x24, x22, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
+      "ldr x23, [x26, #0x8]\n"
       "csel x9, x9, x10, NE\n"
       "ptrue p11.s\n"
-      "ldr x23, [x26, #0x8]\n"
       "whilelt p10.s, XZR, %x[height]\n"
       "mov x22, %x[row_offset]\n"
       "mov x21, %x[out]\n"
@@ -59,119 +59,119 @@ void interleave_block<1, 1, VLType::SME, false>(
       "mov x12, #0x0\n"
       "cbz x28, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25306141  // psel p1.s, p8.s/Z, p10.s[w12]\n"
+      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
+      ".inst 0xe0960320  // ld1w { za0h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n"
       ".inst 0x25706140  // psel p0.s, p8.s/Z, p10.s[w12, #1]\n"
-      ".inst 0xe0960700  // ld1w { za0h.s[x12] }, p1/Z, [x24, x22, LSL #2]\n"
-      "ldr x24, [x26, #0x0]\n"
+      "ldr x25, [x26, #0x0]\n"
       ".inst 0xe09602e1  // ld1w { za0h.s[x12, #1] }, p0/Z, [x23, x22, LSL #2]\n"
       "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "cmp x12, x28\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25306141  // psel p1.s, p8.s/Z, p10.s[w12]\n"
+      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
+      ".inst 0xe0960320  // ld1w { za0h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n"
       ".inst 0x25706140  // psel p0.s, p8.s/Z, p10.s[w12, #1]\n"
       "mov x26, %x[in]\n"
-      "incw x27\n"
-      ".inst 0xe0960700  // ld1w { za0h.s[x12] }, p1/Z, [x24, x22, LSL #2]\n"
-      "ldr x24, [x26, #0x0]\n"
       ".inst 0xe09602e1  // ld1w { za0h.s[x12, #1] }, p0/Z, [x23, x22, LSL #2]\n"
+      "ldr x25, [x26, #0x0]\n"
+      "incw x22\n"
       "ldr x23, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      "incw x22\n"
+      "incw x27\n"
       "cbz x20, 8f\n"
       "mov x20, x20\n"
       "3:"  // K loop: Main loop
       "whilelt p8.s, x27, %x[width]\n"
-      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
       "cbz x28, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25316143  // psel p3.s, p8.s/Z, p10.s[w13]\n"
-      ".inst 0x25716142  // psel p2.s, p8.s/Z, p10.s[w13, #1]\n"
-      ".inst 0x25316d21  // psel p1.s, p11.s/Z, p9.s[w13]\n"
-      ".inst 0x25716d20  // psel p0.s, p11.s/Z, p9.s[w13, #1]\n"
-      ".inst 0xe0962f08  // ld1w { za2h.s[x13] }, p3/Z, [x24, x22, LSL #2]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0962ae9  // ld1w { za2h.s[x13, #1] }, p2/Z, [x23, x22, LSL #2]\n"
+      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
+      ".inst 0xe0960328  // ld1w { za2h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n"
+      ".inst 0x25706141  // psel p1.s, p8.s/Z, p10.s[w12, #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe09606e9  // ld1w { za2h.s[x12, #1] }, p1/Z, [x23, x22, LSL #2]\n"
       "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe0bfa6a0  // st1w { za0v.s[x13] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0aa82a1  // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x28\n"
       "add x26, x26, #0x10\n"
-      ".inst 0xe0aaa2a1  // st1w { za0v.s[x13, #1] }, p0/Z, [x21, x10, LSL #2]\n"
-      "add x13, x13, #0x2\n"
       "addvl x21, x21, #2\n"
-      "cmp x13, x28\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25316143  // psel p3.s, p8.s/Z, p10.s[w13]\n"
-      ".inst 0x25716142  // psel p2.s, p8.s/Z, p10.s[w13, #1]\n"
-      ".inst 0x25316d21  // psel p1.s, p11.s/Z, p9.s[w13]\n"
-      ".inst 0x25716d20  // psel p0.s, p11.s/Z, p9.s[w13, #1]\n"
+      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
+      ".inst 0xe0960328  // ld1w { za2h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25706141  // psel p1.s, p8.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe09606e9  // ld1w { za2h.s[x12, #1] }, p1/Z, [x23, x22, LSL #2]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "whilelt p9.s, x27, %x[width]\n"
-      ".inst 0xe0962f08  // ld1w { za2h.s[x13] }, p3/Z, [x24, x22, LSL #2]\n"
-      "ldr x24, [x26, #0x0]\n"
       "incw x27\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0962ae9  // ld1w { za2h.s[x13, #1] }, p2/Z, [x23, x22, LSL #2]\n"
-      "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe0bfa6a0  // st1w { za0v.s[x13] }, p1/Z, [x21, XZR, LSL #2]\n"
       "add x26, x26, #0x10\n"
-      ".inst 0xe0aaa2a1  // st1w { za0v.s[x13, #1] }, p0/Z, [x21, x10, LSL #2]\n"
+      ".inst 0xe0aa82a1  // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n"
       "addvl x21, x21, #2\n"
       "incw x22\n"
       "whilelt p8.s, x27, %x[width]\n"
+      "mov x12, #0x0\n"
       "cbz x28, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25306143  // psel p3.s, p8.s/Z, p10.s[w12]\n"
-      ".inst 0x25706142  // psel p2.s, p8.s/Z, p10.s[w12, #1]\n"
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
-      ".inst 0xe0960f00  // ld1w { za0h.s[x12] }, p3/Z, [x24, x22, LSL #2]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe0960ae1  // ld1w { za0h.s[x12, #1] }, p2/Z, [x23, x22, LSL #2]\n"
+      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
+      ".inst 0xe0960320  // ld1w { za0h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n"
+      ".inst 0x25706141  // psel p1.s, p8.s/Z, p10.s[w12, #1]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe09606e1  // ld1w { za0h.s[x12, #1] }, p1/Z, [x23, x22, LSL #2]\n"
       "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
-      "add x26, x26, #0x10\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       ".inst 0xe0aa82a9  // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n"
       "add x12, x12, #0x2\n"
-      "addvl x21, x21, #2\n"
       "cmp x12, x28\n"
+      "add x26, x26, #0x10\n"
+      "addvl x21, x21, #2\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25306143  // psel p3.s, p8.s/Z, p10.s[w12]\n"
-      ".inst 0x25706142  // psel p2.s, p8.s/Z, p10.s[w12, #1]\n"
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
+      ".inst 0xe0960320  // ld1w { za0h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n"
       "mov x26, %x[in]\n"
+      "ldr x25, [x26, #0x0]\n"
+      ".inst 0x25706141  // psel p1.s, p8.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe09606e1  // ld1w { za0h.s[x12, #1] }, p1/Z, [x23, x22, LSL #2]\n"
+      "ldr x23, [x26, #0x8]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "whilelt p9.s, x27, %x[width]\n"
-      ".inst 0xe0960f00  // ld1w { za0h.s[x12] }, p3/Z, [x24, x22, LSL #2]\n"
-      "ldr x24, [x26, #0x0]\n"
       "subs x20, x20, #0x1\n"
-      "incw x27\n"
-      ".inst 0xe0960ae1  // ld1w { za0h.s[x12, #1] }, p2/Z, [x23, x22, LSL #2]\n"
-      "ldr x23, [x26, #0x8]\n"
-      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
       "add x26, x26, #0x10\n"
       ".inst 0xe0aa82a9  // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n"
       "addvl x21, x21, #2\n"
+      "incw x27\n"
       "incw x22\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x25, 11f\n"
+      "cbnz x24, 11f\n"
       "mov x26, %x[in]\n"
       "whilelt p8.s, x27, %x[width]\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
-      ".inst 0xe0bf86a0  // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
-      "addvl x21, x21, #1\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "ldr x20, [x26, #0x0]\n"
-      "add x26, x26, #0x8\n"
+      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
       ".inst 0xe0960288  // ld1w { za2h.s[x12] }, p0/Z, [x20, x22, LSL #2]\n"
       "add x12, x12, #0x1\n"
       "cmp x12, x10\n"
+      "add x26, x26, #0x8\n"
+      "addvl x21, x21, #1\n"
       "blt 9b\n"
       "whilelt p9.s, x27, %x[width]\n"
       "whilelt p8.s, x27, %x[width]\n"
@@ -180,8 +180,8 @@ void interleave_block<1, 1, VLType::SME, false>(
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x9\n"
+      "addvl x21, x21, #1\n"
       "blt 10b\n"
       "whilelt p8.s, x27, %x[width]\n"
       "b 13f\n"
@@ -191,15 +191,15 @@ void interleave_block<1, 1, VLType::SME, false>(
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x21, x21, #1\n"
       "cmp x12, x9\n"
+      "addvl x21, x21, #1\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_bf16_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_bf16_bf16.hpp
index a07831e7bd..1ed835b21b 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_bf16_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_bf16_bf16.hpp
@@ -32,66 +32,66 @@ void interleave_block<2, 1, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x28, #0x0\n"
-      "mov x27, %x[row_offset]\n"
-      "cnth x26\n"
-      "cnth x25\n"
-      "cmp %x[height], x26\n"
+      "cnth x28\n"
+      "cmp %x[height], x28\n"
+      "cnth x27\n"
+      "csel x28, %x[height], x28, LT\n"
+      "mov x26, #0x0\n"
       "ptrue p13.s\n"
-      "csel x26, %x[height], x26, LT\n"
+      "sub x28, x28, #0x1\n"
       "whilelt p12.h, XZR, %x[height]\n"
-      "sub x26, x26, #0x1\n"
-      "whilelt p11.h, x25, %x[height]\n"
+      "whilelt p11.h, x27, %x[height]\n"
+      "mov x25, %x[row_offset]\n"
       "mov x24, %x[out]\n"
-      "whilelt p10.h, x28, %x[width]\n"
-      "whilelt p9.h, x28, %x[width]\n"
-      "whilelt p8.h, x28, %x[width]\n"
+      "whilelt p10.h, x26, %x[width]\n"
+      "whilelt p9.h, x26, %x[width]\n"
+      "whilelt p8.h, x26, %x[width]\n"
       "1:"  // Width loop
       "add x23, %x[in], XZR, LSL #3\n"
-      "add x20, %x[in], x25, LSL #3\n"
-      "mov x13, #0x0\n"
+      "add x20, %x[in], x27, LSL #3\n"
       "ldr x22, [x23], #0x8\n"
+      "mov x12, #0x0\n"
       "ldr x21, [x20], #0x8\n"
-      "cbz x26, 3f\n"
+      "cbz x28, 3f\n"
       "2:"  // Loads: Loop
-      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
-      ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
-      ".inst 0xe05b26c0  // ld1h { za0h.h[x13] }, p1/Z, [x22, x27, LSL #1]\n"
+      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      ".inst 0xe05906c0  // ld1h { za0h.h[x12] }, p1/Z, [x22, x25, LSL #1]\n"
       "ldr x22, [x23], #0x8\n"
-      ".inst 0xe05b22a8  // ld1h { za1h.h[x13] }, p0/Z, [x21, x27, LSL #1]\n"
-      "add x13, x13, #0x2\n"
+      ".inst 0xe05902a8  // ld1h { za1h.h[x12] }, p0/Z, [x21, x25, LSL #1]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x28, LSL #1\n"
       "ldr x21, [x20], #0x8\n"
-      "cmp x13, x26, LSL #1\n"
       "blt 2b\n"
       "3:"  // Loads: Tail
-      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
-      ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
-      "sub x20, %x[width], x28\n"
+      "sub x20, %x[width], x26\n"
+      ".inst 0x25286580  // psel p0.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0xe05902c0  // ld1h { za0h.h[x12] }, p0/Z, [x22, x25, LSL #1]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      "cmp x20, x27\n"
+      ".inst 0xe05902a8  // ld1h { za1h.h[x12] }, p0/Z, [x21, x25, LSL #1]\n"
       "mov x12, #0x0\n"
-      "cmp x20, x25\n"
-      ".inst 0xe05b26c0  // ld1h { za0h.h[x13] }, p1/Z, [x22, x27, LSL #1]\n"
-      "csel x20, x20, x25, LT\n"
-      ".inst 0xe05b22a8  // ld1h { za1h.h[x13] }, p0/Z, [x21, x27, LSL #1]\n"
+      "csel x20, x20, x27, LT\n"
       "4:"  // Stores: Loop
-      ".inst 0x25287541  // psel p1.h, p13.h/Z, p10.h[w12]\n"
       ".inst 0x25287540  // psel p0.h, p13.h/Z, p10.h[w12]\n"
-      ".inst 0xe07f8700  // st1h { za0v.h[x12] }, p1/Z, [x24, XZR, LSL #1]\n"
-      ".inst 0xe0798308  // st1h { za1v.h[x12] }, p0/Z, [x24, x25, LSL #1]\n"
+      ".inst 0xe07f8300  // st1h { za0v.h[x12] }, p0/Z, [x24, XZR, LSL #1]\n"
+      ".inst 0x25287540  // psel p0.h, p13.h/Z, p10.h[w12]\n"
+      ".inst 0xe07b8308  // st1h { za1v.h[x12] }, p0/Z, [x24, x27, LSL #1]\n"
       "add x12, x12, #0x1\n"
-      "addvl x24, x24, #4\n"
       "cmp x12, x20\n"
+      "addvl x24, x24, #4\n"
       "blt 4b\n"
-      "inch x28\n"
-      "inch x27\n"
-      "whilelt p10.h, x28, %x[width]\n"
-      "whilelt p9.h, x28, %x[width]\n"
-      "whilelt p8.h, x28, %x[width]\n"
+      "inch x26\n"
+      "whilelt p10.h, x26, %x[width]\n"
+      "whilelt p9.h, x26, %x[width]\n"
+      "whilelt p8.h, x26, %x[width]\n"
+      "inch x25\n"
       "b.any 1b\n"
       "mov %x[out], x24\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_bf16_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_bf16_bf16.hpp
index 01dfecc4ef..715810ddea 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_bf16_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_bf16_bf16.hpp
@@ -32,264 +32,269 @@ void interleave_block<2, 2, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x22, %x[width]\n"
+      "cnth x22\n"
       "mov x21, %x[width]\n"
-      "cnth x20\n"
-      "inch x22\n"
-      "sub x7, x20, #0x1\n"
-      "sub x22, x22, #0x1\n"
-      "ands x7, x21, x7\n"
-      "cntw x8\n"
-      "udiv x22, x22, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x7, x7, x20, NE\n"
-      "sub x13, x22, #0x1\n"
-      "add x7, x7, #0x1\n"
-      "sub x17, x8, #0x2\n"
-      "lsl x21, %x[height], #0x1\n"  // height * 2
-      "lsl x20, x8, #0x1\n"
-      "mov x16, #0x0\n"
+      "inch x21\n"
+      "mov x20, %x[width]\n"
+      "sub x17, x22, #0x1\n"
+      "sub x21, x21, #0x1\n"
+      "ands x17, x20, x17\n"
+      "cntw x16\n"
+      "udiv x21, x21, x22\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x17, x17, x22, NE\n"
+      "sub x13, x21, #0x1\n"
+      "add x17, x17, #0x1\n"
+      "sub x15, x16, #0x2\n"
+      "lsl x22, %x[height], #0x1\n"  // height * 2
+      "lsl x20, x16, #0x1\n"
+      "mov x14, #0x0\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      "cntw x9, ALL, MUL #2\n"
-      "cntw x28, ALL, MUL #3\n"
-      "ldr x27, [x11, #0x0]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      "cntw x28, ALL, MUL #2\n"
+      "cntw x27, ALL, MUL #3\n"
+      "ldr x26, [x10, #0x0]\n"
       "lsr x13, x13, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x26, x22, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "ldr x25, [x10, #0x0]\n"
-      "lsr x7, x7, #0x1\n"
-      "ptrue p12.s\n"
+      "and x25, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "ldr x24, [x11, #0x8]\n"
-      "whilelt p11.h, XZR, x21\n"
-      "whilelt p10.h, x20, x21\n"
+      "lsr x17, x17, #0x1\n"
+      "ptrue p13.s\n"
       "ldr x21, [x10, #0x8]\n"
+      "whilelt p12.h, XZR, x22\n"
+      "whilelt p11.h, x20, x22\n"
       "mov x23, %x[row_offset]\n"
       "mov x22, %x[out]\n"
-      "whilelt p9.h, x16, %x[width]\n"
-      "whilelt p8.h, x16, %x[width]\n"
+      "whilelt p10.h, x14, %x[width]\n"
+      "whilelt p9.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
       "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
       "mov x12, #0x0\n"
-      "cbz x17, 2f\n"
+      "cbz x15, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25286163  // psel p3.h, p8.h/Z, p11.h[w12]\n"
-      ".inst 0x25286142  // psel p2.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25686161  // psel p1.h, p8.h/Z, p11.h[w12, #2]\n"
-      ".inst 0x25686140  // psel p0.h, p8.h/Z, p10.h[w12, #2]\n"
-      ".inst 0xe0570f60  // ld1h { za0h.h[x12] }, p3/Z, [x27, x23, LSL #1]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0570b28  // ld1h { za1h.h[x12] }, p2/Z, [x25, x23, LSL #1]\n"
-      "ldr x25, [x10, #0x0]\n"
+      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      ".inst 0xe0570520  // ld1h { za0h.h[x12] }, p1/Z, [x9, x23, LSL #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0570348  // ld1h { za1h.h[x12] }, p0/Z, [x26, x23, LSL #1]\n"
+      ".inst 0x25686581  // psel p1.h, p9.h/Z, p12.h[w12, #2]\n"
+      ".inst 0x25686160  // psel p0.h, p8.h/Z, p11.h[w12, #2]\n"
+      "ldr x26, [x10, #0x0]\n"
       ".inst 0xe0570702  // ld1h { za0h.h[x12, #2] }, p1/Z, [x24, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
       ".inst 0xe05702aa  // ld1h { za1h.h[x12, #2] }, p0/Z, [x21, x23, LSL #1]\n"
       "add x12, x12, #0x4\n"
+      "cmp x12, x15, LSL #1\n"
       "ldr x21, [x10, #0x8]\n"
       "add x10, x10, #0x10\n"
-      "cmp x12, x17, LSL #1\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25286163  // psel p3.h, p8.h/Z, p11.h[w12]\n"
-      ".inst 0x25286142  // psel p2.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25686161  // psel p1.h, p8.h/Z, p11.h[w12, #2]\n"
-      ".inst 0x25686140  // psel p0.h, p8.h/Z, p10.h[w12, #2]\n"
+      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      ".inst 0xe0570520  // ld1h { za0h.h[x12] }, p1/Z, [x9, x23, LSL #1]\n"
+      ".inst 0xe0570348  // ld1h { za1h.h[x12] }, p0/Z, [x26, x23, LSL #1]\n"
+      ".inst 0x25686581  // psel p1.h, p9.h/Z, p12.h[w12, #2]\n"
+      ".inst 0x25686160  // psel p0.h, p8.h/Z, p11.h[w12, #2]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0570f60  // ld1h { za0h.h[x12] }, p3/Z, [x27, x23, LSL #1]\n"
-      "ldr x27, [x11, #0x0]\n"
-      "inch x16\n"
-      ".inst 0xe0570b28  // ld1h { za1h.h[x12] }, p2/Z, [x25, x23, LSL #1]\n"
-      "ldr x25, [x10, #0x0]\n"
+      "add x10, %x[in], x16, LSL #3\n"
       ".inst 0xe0570702  // ld1h { za0h.h[x12, #2] }, p1/Z, [x24, x23, LSL #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe05702aa  // ld1h { za1h.h[x12, #2] }, p0/Z, [x21, x23, LSL #1]\n"
+      "ldr x26, [x10, #0x0]\n"
+      "inch x23\n"
+      "inch x14\n"
       "ldr x24, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe05702aa  // ld1h { za1h.h[x12, #2] }, p0/Z, [x21, x23, LSL #1]\n"
       "ldr x21, [x10, #0x8]\n"
       "add x10, x10, #0x10\n"
-      "inch x23\n"
       "cbz x13, 8f\n"
       "mov x20, x13\n"
       "3:"  // K loop: Main loop
-      "whilelt p8.h, x16, %x[width]\n"
-      "mov x15, #0x0\n"
-      "mov x14, #0x0\n"
-      "cbz x17, 5f\n"
+      "whilelt p9.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
+      "cbz x15, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x253b6160  // psel p0.h, p8.h/Z, p11.h[w15, #1]\n"
-      ".inst 0x253b6142  // psel p2.h, p8.h/Z, p10.h[w15, #1]\n"
-      ".inst 0x257b6161  // psel p1.h, p8.h/Z, p11.h[w15, #3]\n"
-      ".inst 0x257b6143  // psel p3.h, p8.h/Z, p10.h[w15, #3]\n"
-      ".inst 0xe0576361  // ld1h { za0h.h[x15, #1] }, p0/Z, [x27, x23, LSL #1]\n"
-      ".inst 0x252a7120  // psel p0.h, p12.h/Z, p9.h[w14]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0576b29  // ld1h { za1h.h[x15, #1] }, p2/Z, [x25, x23, LSL #1]\n"
-      ".inst 0x252a7122  // psel p2.h, p12.h/Z, p9.h[w14]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0576703  // ld1h { za0h.h[x15, #3] }, p1/Z, [x24, x23, LSL #1]\n"
-      ".inst 0x253a7121  // psel p1.h, p12.h/Z, p9.h[w14, #1]\n"
+      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
+      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
+      ".inst 0xe0572521  // ld1h { za0h.h[x13, #1] }, p1/Z, [x9, x23, LSL #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0572349  // ld1h { za1h.h[x13, #1] }, p0/Z, [x26, x23, LSL #1]\n"
+      ".inst 0x25796580  // psel p0.h, p9.h/Z, p12.h[w13, #3]\n"
+      ".inst 0x25796162  // psel p2.h, p8.h/Z, p11.h[w13, #3]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0572303  // ld1h { za0h.h[x13, #3] }, p0/Z, [x24, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0576eab  // ld1h { za1h.h[x15, #3] }, p3/Z, [x21, x23, LSL #1]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0572aab  // ld1h { za1h.h[x13, #3] }, p2/Z, [x21, x23, LSL #1]\n"
       "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bfc2c0  // st1w { za0v.s[x14] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x253a7120  // psel p0.h, p12.h/Z, p9.h[w14, #1]\n"
-      ".inst 0xe0a8cac4  // st1w { za1v.s[x14] }, p2/Z, [x22, x8, LSL #2]\n"
+      ".inst 0xe0bf86c0  // st1w { za0v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      "add x11, x11, #0x10\n"
+      ".inst 0xe0bc86c1  // st1w { za0v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
       "add x10, x10, #0x10\n"
-      "add x15, x15, #0x4\n"
-      ".inst 0xe0a9c6c1  // st1w { za0v.s[x14, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bcc2c5  // st1w { za1v.s[x14, #1] }, p0/Z, [x22, x28, LSL #2]\n"
-      "add x14, x14, #0x2\n"
+      "add x13, x13, #0x4\n"
+      ".inst 0xe0bb82c5  // st1w { za1v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x15\n"
       "addvl x22, x22, #4\n"
-      "cmp x14, x17\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x253b6160  // psel p0.h, p8.h/Z, p11.h[w15, #1]\n"
-      ".inst 0x253b6142  // psel p2.h, p8.h/Z, p10.h[w15, #1]\n"
-      ".inst 0x257b6161  // psel p1.h, p8.h/Z, p11.h[w15, #3]\n"
-      ".inst 0x257b6143  // psel p3.h, p8.h/Z, p10.h[w15, #3]\n"
+      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
+      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
+      ".inst 0xe0572521  // ld1h { za0h.h[x13, #1] }, p1/Z, [x9, x23, LSL #1]\n"
+      ".inst 0xe0572349  // ld1h { za1h.h[x13, #1] }, p0/Z, [x26, x23, LSL #1]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0576361  // ld1h { za0h.h[x15, #1] }, p0/Z, [x27, x23, LSL #1]\n"
-      ".inst 0x252a7120  // psel p0.h, p12.h/Z, p9.h[w14]\n"
-      "ldr x27, [x11, #0x0]\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0576b29  // ld1h { za1h.h[x15, #1] }, p2/Z, [x25, x23, LSL #1]\n"
-      ".inst 0x252a7122  // psel p2.h, p12.h/Z, p9.h[w14]\n"
-      "ldr x25, [x10, #0x0]\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0576703  // ld1h { za0h.h[x15, #3] }, p1/Z, [x24, x23, LSL #1]\n"
-      ".inst 0x253a7121  // psel p1.h, p12.h/Z, p9.h[w14, #1]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0x25796580  // psel p0.h, p9.h/Z, p12.h[w13, #3]\n"
+      ".inst 0x25796161  // psel p1.h, p8.h/Z, p11.h[w13, #3]\n"
+      ".inst 0xe0572303  // ld1h { za0h.h[x13, #3] }, p0/Z, [x24, x23, LSL #1]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe05726ab  // ld1h { za1h.h[x13, #3] }, p1/Z, [x21, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0576eab  // ld1h { za1h.h[x15, #3] }, p3/Z, [x21, x23, LSL #1]\n"
+      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
       "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bfc2c0  // st1w { za0v.s[x14] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x253a7120  // psel p0.h, p12.h/Z, p9.h[w14, #1]\n"
-      ".inst 0xe0a8cac4  // st1w { za1v.s[x14] }, p2/Z, [x22, x8, LSL #2]\n"
-      "whilelt p9.h, x16, %x[width]\n"
-      "inch x16\n"
-      ".inst 0xe0a9c6c1  // st1w { za0v.s[x14, #1] }, p1/Z, [x22, x9, LSL #2]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b08ac4  // st1w { za1v.s[x12] }, p2/Z, [x22, x16, LSL #2]\n"
+      "whilelt p10.h, x14, %x[width]\n"
+      "inch x14\n"
+      ".inst 0xe0bc86c1  // st1w { za0v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
+      "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
-      "inch x23\n"
-      ".inst 0xe0bcc2c5  // st1w { za1v.s[x14, #1] }, p0/Z, [x22, x28, LSL #2]\n"
+      ".inst 0xe0bb82c5  // st1w { za1v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
       "addvl x22, x22, #4\n"
-      "whilelt p8.h, x16, %x[width]\n"
-      "cbz x17, 7f\n"
+      "inch x23\n"
+      "whilelt p9.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
+      "cbz x15, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
+      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
       ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
-      ".inst 0x25296142  // psel p2.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25696161  // psel p1.h, p8.h/Z, p11.h[w13, #2]\n"
-      ".inst 0x25696143  // psel p3.h, p8.h/Z, p10.h[w13, #2]\n"
-      ".inst 0xe0572360  // ld1h { za0h.h[x13] }, p0/Z, [x27, x23, LSL #1]\n"
-      ".inst 0x25287120  // psel p0.h, p12.h/Z, p9.h[w12]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0572b28  // ld1h { za1h.h[x13] }, p2/Z, [x25, x23, LSL #1]\n"
-      ".inst 0x25287122  // psel p2.h, p12.h/Z, p9.h[w12]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0572702  // ld1h { za0h.h[x13, #2] }, p1/Z, [x24, x23, LSL #1]\n"
-      ".inst 0x25387121  // psel p1.h, p12.h/Z, p9.h[w12, #1]\n"
+      ".inst 0xe0572520  // ld1h { za0h.h[x13] }, p1/Z, [x9, x23, LSL #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0572348  // ld1h { za1h.h[x13] }, p0/Z, [x26, x23, LSL #1]\n"
+      ".inst 0x25696580  // psel p0.h, p9.h/Z, p12.h[w13, #2]\n"
+      ".inst 0x25696162  // psel p2.h, p8.h/Z, p11.h[w13, #2]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0572302  // ld1h { za0h.h[x13, #2] }, p0/Z, [x24, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0572eaa  // ld1h { za1h.h[x13, #2] }, p3/Z, [x21, x23, LSL #1]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0572aaa  // ld1h { za1h.h[x13, #2] }, p2/Z, [x21, x23, LSL #1]\n"
       "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x25387120  // psel p0.h, p12.h/Z, p9.h[w12, #1]\n"
-      ".inst 0xe0a88acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
+      ".inst 0xe0bf86c8  // st1w { za2v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b082cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      "add x11, x11, #0x10\n"
+      ".inst 0xe0bc86c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
       "add x10, x10, #0x10\n"
       "add x13, x13, #0x4\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bc82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x28, LSL #2]\n"
+      ".inst 0xe0bb82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
       "add x12, x12, #0x2\n"
+      "cmp x12, x15\n"
       "addvl x22, x22, #4\n"
-      "cmp x12, x17\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
+      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
       ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
-      ".inst 0x25296142  // psel p2.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25696161  // psel p1.h, p8.h/Z, p11.h[w13, #2]\n"
-      ".inst 0x25696143  // psel p3.h, p8.h/Z, p10.h[w13, #2]\n"
+      ".inst 0xe0572520  // ld1h { za0h.h[x13] }, p1/Z, [x9, x23, LSL #1]\n"
+      ".inst 0xe0572348  // ld1h { za1h.h[x13] }, p0/Z, [x26, x23, LSL #1]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0572360  // ld1h { za0h.h[x13] }, p0/Z, [x27, x23, LSL #1]\n"
-      ".inst 0x25287120  // psel p0.h, p12.h/Z, p9.h[w12]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0572b28  // ld1h { za1h.h[x13] }, p2/Z, [x25, x23, LSL #1]\n"
-      ".inst 0x25287122  // psel p2.h, p12.h/Z, p9.h[w12]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0572702  // ld1h { za0h.h[x13, #2] }, p1/Z, [x24, x23, LSL #1]\n"
-      ".inst 0x25387121  // psel p1.h, p12.h/Z, p9.h[w12, #1]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0x25696580  // psel p0.h, p9.h/Z, p12.h[w13, #2]\n"
+      ".inst 0x25696161  // psel p1.h, p8.h/Z, p11.h[w13, #2]\n"
+      ".inst 0xe0572302  // ld1h { za0h.h[x13, #2] }, p0/Z, [x24, x23, LSL #1]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe05726aa  // ld1h { za1h.h[x13, #2] }, p1/Z, [x21, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0572eaa  // ld1h { za1h.h[x13, #2] }, p3/Z, [x21, x23, LSL #1]\n"
+      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
       "ldr x21, [x10, #0x8]\n"
       ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x25387120  // psel p0.h, p12.h/Z, p9.h[w12, #1]\n"
-      ".inst 0xe0a88acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
-      "whilelt p9.h, x16, %x[width]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b08acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x16, LSL #2]\n"
+      "whilelt p10.h, x14, %x[width]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
+      ".inst 0xe0bc86c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
+      "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
-      "inch x16\n"
-      ".inst 0xe0bc82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x28, LSL #2]\n"
+      ".inst 0xe0bb82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
       "addvl x22, x22, #4\n"
+      "inch x14\n"
       "inch x23\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x26, 11f\n"
+      "cbnz x25, 11f\n"
       "mov x11, %x[in]\n"
-      "whilelt p8.h, x16, %x[width]\n"
+      "whilelt p9.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25307123  // psel p3.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307122  // psel p2.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25396161  // psel p1.h, p8.h/Z, p11.h[w13, #1]\n"
-      ".inst 0x25396140  // psel p0.h, p8.h/Z, p10.h[w13, #1]\n"
-      ".inst 0xe0bf8ec0  // st1w { za0v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a88ac4  // st1w { za1v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x22, x22, #2\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "ldr x21, [x11, #0x0]\n"
-      "cmp x12, x8\n"
-      "ldr x20, [x11, x8, LSL #0x3]\n"
-      "add x11, x11, #0x8\n"
+      "add x12, x12, #0x1\n"
+      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
+      "ldr x20, [x11, x16, LSL #0x3]\n"
+      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
+      "cmp x12, x16\n"
       ".inst 0xe05726a1  // ld1h { za0h.h[x13, #1] }, p1/Z, [x21, x23, LSL #1]\n"
       ".inst 0xe0572289  // ld1h { za1h.h[x13, #1] }, p0/Z, [x20, x23, LSL #1]\n"
+      "add x11, x11, #0x8\n"
+      "addvl x22, x22, #2\n"
       "add x13, x13, #0x2\n"
       "blt 9b\n"
-      "whilelt p9.h, x16, %x[width]\n"
-      "whilelt p8.h, x16, %x[width]\n"
+      "whilelt p10.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
       "mov x20, #0x0\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: Second
-      ".inst 0x25307121  // psel p1.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307120  // psel p0.s, p12.s/Z, p9.s[w12]\n"
-      "add x20, x20, #0x2\n"
-      ".inst 0xe0bf86c8  // st1w { za2v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a882cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x8, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
+      "cmp x12, x17\n"
       "addvl x22, x22, #2\n"
-      "cmp x12, x7\n"
+      "add x20, x20, #0x2\n"
       "blt 10b\n"
-      "whilelt p8.h, x16, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
       "b 13f\n"
       "11:"  // K loop: Tails: Odd
       "mov x12, #0x0\n"
       "12:"  // K loop: Tails: Odd: Loop
-      ".inst 0x25307121  // psel p1.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307120  // psel p0.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0xe0bf86c0  // st1w { za0v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a882c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x8, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
+      "cmp x12, x17\n"
       "addvl x22, x22, #2\n"
-      "cmp x12, x7\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x22\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_fp16_fp16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_fp16_fp16.hpp
index b1e9226773..849f6c3228 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_fp16_fp16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block2_fp16_fp16.hpp
@@ -32,264 +32,269 @@ void interleave_block<2, 2, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x22, %x[width]\n"
+      "cnth x22\n"
       "mov x21, %x[width]\n"
-      "cnth x20\n"
-      "inch x22\n"
-      "sub x7, x20, #0x1\n"
-      "sub x22, x22, #0x1\n"
-      "ands x7, x21, x7\n"
-      "cntw x8\n"
-      "udiv x22, x22, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x7, x7, x20, NE\n"
-      "sub x13, x22, #0x1\n"
-      "add x7, x7, #0x1\n"
-      "sub x17, x8, #0x2\n"
-      "lsl x21, %x[height], #0x1\n"  // height * 2
-      "lsl x20, x8, #0x1\n"
-      "mov x16, #0x0\n"
+      "inch x21\n"
+      "mov x20, %x[width]\n"
+      "sub x17, x22, #0x1\n"
+      "sub x21, x21, #0x1\n"
+      "ands x17, x20, x17\n"
+      "cntw x16\n"
+      "udiv x21, x21, x22\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x17, x17, x22, NE\n"
+      "sub x13, x21, #0x1\n"
+      "add x17, x17, #0x1\n"
+      "sub x15, x16, #0x2\n"
+      "lsl x22, %x[height], #0x1\n"  // height * 2
+      "lsl x20, x16, #0x1\n"
+      "mov x14, #0x0\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      "cntw x9, ALL, MUL #2\n"
-      "cntw x28, ALL, MUL #3\n"
-      "ldr x27, [x11, #0x0]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      "cntw x28, ALL, MUL #2\n"
+      "cntw x27, ALL, MUL #3\n"
+      "ldr x26, [x10, #0x0]\n"
       "lsr x13, x13, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x26, x22, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "ldr x25, [x10, #0x0]\n"
-      "lsr x7, x7, #0x1\n"
-      "ptrue p12.s\n"
+      "and x25, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "ldr x24, [x11, #0x8]\n"
-      "whilelt p11.h, XZR, x21\n"
-      "whilelt p10.h, x20, x21\n"
+      "lsr x17, x17, #0x1\n"
+      "ptrue p13.s\n"
       "ldr x21, [x10, #0x8]\n"
+      "whilelt p12.h, XZR, x22\n"
+      "whilelt p11.h, x20, x22\n"
       "mov x23, %x[row_offset]\n"
       "mov x22, %x[out]\n"
-      "whilelt p9.h, x16, %x[width]\n"
-      "whilelt p8.h, x16, %x[width]\n"
+      "whilelt p10.h, x14, %x[width]\n"
+      "whilelt p9.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
       "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
       "mov x12, #0x0\n"
-      "cbz x17, 2f\n"
+      "cbz x15, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25286163  // psel p3.h, p8.h/Z, p11.h[w12]\n"
-      ".inst 0x25286142  // psel p2.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25686161  // psel p1.h, p8.h/Z, p11.h[w12, #2]\n"
-      ".inst 0x25686140  // psel p0.h, p8.h/Z, p10.h[w12, #2]\n"
-      ".inst 0xe0570f60  // ld1h { za0h.h[x12] }, p3/Z, [x27, x23, LSL #1]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0570b28  // ld1h { za1h.h[x12] }, p2/Z, [x25, x23, LSL #1]\n"
-      "ldr x25, [x10, #0x0]\n"
+      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      ".inst 0xe0570520  // ld1h { za0h.h[x12] }, p1/Z, [x9, x23, LSL #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0570348  // ld1h { za1h.h[x12] }, p0/Z, [x26, x23, LSL #1]\n"
+      ".inst 0x25686581  // psel p1.h, p9.h/Z, p12.h[w12, #2]\n"
+      ".inst 0x25686160  // psel p0.h, p8.h/Z, p11.h[w12, #2]\n"
+      "ldr x26, [x10, #0x0]\n"
       ".inst 0xe0570702  // ld1h { za0h.h[x12, #2] }, p1/Z, [x24, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
       ".inst 0xe05702aa  // ld1h { za1h.h[x12, #2] }, p0/Z, [x21, x23, LSL #1]\n"
       "add x12, x12, #0x4\n"
+      "cmp x12, x15, LSL #1\n"
       "ldr x21, [x10, #0x8]\n"
       "add x10, x10, #0x10\n"
-      "cmp x12, x17, LSL #1\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25286163  // psel p3.h, p8.h/Z, p11.h[w12]\n"
-      ".inst 0x25286142  // psel p2.h, p8.h/Z, p10.h[w12]\n"
-      ".inst 0x25686161  // psel p1.h, p8.h/Z, p11.h[w12, #2]\n"
-      ".inst 0x25686140  // psel p0.h, p8.h/Z, p10.h[w12, #2]\n"
+      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      ".inst 0xe0570520  // ld1h { za0h.h[x12] }, p1/Z, [x9, x23, LSL #1]\n"
+      ".inst 0xe0570348  // ld1h { za1h.h[x12] }, p0/Z, [x26, x23, LSL #1]\n"
+      ".inst 0x25686581  // psel p1.h, p9.h/Z, p12.h[w12, #2]\n"
+      ".inst 0x25686160  // psel p0.h, p8.h/Z, p11.h[w12, #2]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0570f60  // ld1h { za0h.h[x12] }, p3/Z, [x27, x23, LSL #1]\n"
-      "ldr x27, [x11, #0x0]\n"
-      "inch x16\n"
-      ".inst 0xe0570b28  // ld1h { za1h.h[x12] }, p2/Z, [x25, x23, LSL #1]\n"
-      "ldr x25, [x10, #0x0]\n"
+      "add x10, %x[in], x16, LSL #3\n"
       ".inst 0xe0570702  // ld1h { za0h.h[x12, #2] }, p1/Z, [x24, x23, LSL #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe05702aa  // ld1h { za1h.h[x12, #2] }, p0/Z, [x21, x23, LSL #1]\n"
+      "ldr x26, [x10, #0x0]\n"
+      "inch x23\n"
+      "inch x14\n"
       "ldr x24, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe05702aa  // ld1h { za1h.h[x12, #2] }, p0/Z, [x21, x23, LSL #1]\n"
       "ldr x21, [x10, #0x8]\n"
       "add x10, x10, #0x10\n"
-      "inch x23\n"
       "cbz x13, 8f\n"
       "mov x20, x13\n"
       "3:"  // K loop: Main loop
-      "whilelt p8.h, x16, %x[width]\n"
-      "mov x15, #0x0\n"
-      "mov x14, #0x0\n"
-      "cbz x17, 5f\n"
+      "whilelt p9.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
+      "cbz x15, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x253b6160  // psel p0.h, p8.h/Z, p11.h[w15, #1]\n"
-      ".inst 0x253b6142  // psel p2.h, p8.h/Z, p10.h[w15, #1]\n"
-      ".inst 0x257b6161  // psel p1.h, p8.h/Z, p11.h[w15, #3]\n"
-      ".inst 0x257b6143  // psel p3.h, p8.h/Z, p10.h[w15, #3]\n"
-      ".inst 0xe0576361  // ld1h { za0h.h[x15, #1] }, p0/Z, [x27, x23, LSL #1]\n"
-      ".inst 0x252a7120  // psel p0.h, p12.h/Z, p9.h[w14]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0576b29  // ld1h { za1h.h[x15, #1] }, p2/Z, [x25, x23, LSL #1]\n"
-      ".inst 0x252a7122  // psel p2.h, p12.h/Z, p9.h[w14]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0576703  // ld1h { za0h.h[x15, #3] }, p1/Z, [x24, x23, LSL #1]\n"
-      ".inst 0x253a7121  // psel p1.h, p12.h/Z, p9.h[w14, #1]\n"
+      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
+      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
+      ".inst 0xe0572521  // ld1h { za0h.h[x13, #1] }, p1/Z, [x9, x23, LSL #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0572349  // ld1h { za1h.h[x13, #1] }, p0/Z, [x26, x23, LSL #1]\n"
+      ".inst 0x25796580  // psel p0.h, p9.h/Z, p12.h[w13, #3]\n"
+      ".inst 0x25796162  // psel p2.h, p8.h/Z, p11.h[w13, #3]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0572303  // ld1h { za0h.h[x13, #3] }, p0/Z, [x24, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0576eab  // ld1h { za1h.h[x15, #3] }, p3/Z, [x21, x23, LSL #1]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0572aab  // ld1h { za1h.h[x13, #3] }, p2/Z, [x21, x23, LSL #1]\n"
       "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bfc2c0  // st1w { za0v.s[x14] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x253a7120  // psel p0.h, p12.h/Z, p9.h[w14, #1]\n"
-      ".inst 0xe0a8cac4  // st1w { za1v.s[x14] }, p2/Z, [x22, x8, LSL #2]\n"
+      ".inst 0xe0bf86c0  // st1w { za0v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      "add x11, x11, #0x10\n"
+      ".inst 0xe0bc86c1  // st1w { za0v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
       "add x10, x10, #0x10\n"
-      "add x15, x15, #0x4\n"
-      ".inst 0xe0a9c6c1  // st1w { za0v.s[x14, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bcc2c5  // st1w { za1v.s[x14, #1] }, p0/Z, [x22, x28, LSL #2]\n"
-      "add x14, x14, #0x2\n"
+      "add x13, x13, #0x4\n"
+      ".inst 0xe0bb82c5  // st1w { za1v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x15\n"
       "addvl x22, x22, #4\n"
-      "cmp x14, x17\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x253b6160  // psel p0.h, p8.h/Z, p11.h[w15, #1]\n"
-      ".inst 0x253b6142  // psel p2.h, p8.h/Z, p10.h[w15, #1]\n"
-      ".inst 0x257b6161  // psel p1.h, p8.h/Z, p11.h[w15, #3]\n"
-      ".inst 0x257b6143  // psel p3.h, p8.h/Z, p10.h[w15, #3]\n"
+      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
+      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
+      ".inst 0xe0572521  // ld1h { za0h.h[x13, #1] }, p1/Z, [x9, x23, LSL #1]\n"
+      ".inst 0xe0572349  // ld1h { za1h.h[x13, #1] }, p0/Z, [x26, x23, LSL #1]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0576361  // ld1h { za0h.h[x15, #1] }, p0/Z, [x27, x23, LSL #1]\n"
-      ".inst 0x252a7120  // psel p0.h, p12.h/Z, p9.h[w14]\n"
-      "ldr x27, [x11, #0x0]\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0576b29  // ld1h { za1h.h[x15, #1] }, p2/Z, [x25, x23, LSL #1]\n"
-      ".inst 0x252a7122  // psel p2.h, p12.h/Z, p9.h[w14]\n"
-      "ldr x25, [x10, #0x0]\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0576703  // ld1h { za0h.h[x15, #3] }, p1/Z, [x24, x23, LSL #1]\n"
-      ".inst 0x253a7121  // psel p1.h, p12.h/Z, p9.h[w14, #1]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0x25796580  // psel p0.h, p9.h/Z, p12.h[w13, #3]\n"
+      ".inst 0x25796161  // psel p1.h, p8.h/Z, p11.h[w13, #3]\n"
+      ".inst 0xe0572303  // ld1h { za0h.h[x13, #3] }, p0/Z, [x24, x23, LSL #1]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe05726ab  // ld1h { za1h.h[x13, #3] }, p1/Z, [x21, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0576eab  // ld1h { za1h.h[x15, #3] }, p3/Z, [x21, x23, LSL #1]\n"
+      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
       "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bfc2c0  // st1w { za0v.s[x14] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x253a7120  // psel p0.h, p12.h/Z, p9.h[w14, #1]\n"
-      ".inst 0xe0a8cac4  // st1w { za1v.s[x14] }, p2/Z, [x22, x8, LSL #2]\n"
-      "whilelt p9.h, x16, %x[width]\n"
-      "inch x16\n"
-      ".inst 0xe0a9c6c1  // st1w { za0v.s[x14, #1] }, p1/Z, [x22, x9, LSL #2]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b08ac4  // st1w { za1v.s[x12] }, p2/Z, [x22, x16, LSL #2]\n"
+      "whilelt p10.h, x14, %x[width]\n"
+      "inch x14\n"
+      ".inst 0xe0bc86c1  // st1w { za0v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
+      "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
-      "inch x23\n"
-      ".inst 0xe0bcc2c5  // st1w { za1v.s[x14, #1] }, p0/Z, [x22, x28, LSL #2]\n"
+      ".inst 0xe0bb82c5  // st1w { za1v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
       "addvl x22, x22, #4\n"
-      "whilelt p8.h, x16, %x[width]\n"
-      "cbz x17, 7f\n"
+      "inch x23\n"
+      "whilelt p9.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
+      "cbz x15, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
+      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
       ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
-      ".inst 0x25296142  // psel p2.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25696161  // psel p1.h, p8.h/Z, p11.h[w13, #2]\n"
-      ".inst 0x25696143  // psel p3.h, p8.h/Z, p10.h[w13, #2]\n"
-      ".inst 0xe0572360  // ld1h { za0h.h[x13] }, p0/Z, [x27, x23, LSL #1]\n"
-      ".inst 0x25287120  // psel p0.h, p12.h/Z, p9.h[w12]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0572b28  // ld1h { za1h.h[x13] }, p2/Z, [x25, x23, LSL #1]\n"
-      ".inst 0x25287122  // psel p2.h, p12.h/Z, p9.h[w12]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0572702  // ld1h { za0h.h[x13, #2] }, p1/Z, [x24, x23, LSL #1]\n"
-      ".inst 0x25387121  // psel p1.h, p12.h/Z, p9.h[w12, #1]\n"
+      ".inst 0xe0572520  // ld1h { za0h.h[x13] }, p1/Z, [x9, x23, LSL #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0572348  // ld1h { za1h.h[x13] }, p0/Z, [x26, x23, LSL #1]\n"
+      ".inst 0x25696580  // psel p0.h, p9.h/Z, p12.h[w13, #2]\n"
+      ".inst 0x25696162  // psel p2.h, p8.h/Z, p11.h[w13, #2]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0572302  // ld1h { za0h.h[x13, #2] }, p0/Z, [x24, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0572eaa  // ld1h { za1h.h[x13, #2] }, p3/Z, [x21, x23, LSL #1]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0572aaa  // ld1h { za1h.h[x13, #2] }, p2/Z, [x21, x23, LSL #1]\n"
       "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x25387120  // psel p0.h, p12.h/Z, p9.h[w12, #1]\n"
-      ".inst 0xe0a88acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
+      ".inst 0xe0bf86c8  // st1w { za2v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b082cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      "add x11, x11, #0x10\n"
+      ".inst 0xe0bc86c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
       "add x10, x10, #0x10\n"
       "add x13, x13, #0x4\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bc82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x28, LSL #2]\n"
+      ".inst 0xe0bb82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
       "add x12, x12, #0x2\n"
+      "cmp x12, x15\n"
       "addvl x22, x22, #4\n"
-      "cmp x12, x17\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
+      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
       ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
-      ".inst 0x25296142  // psel p2.h, p8.h/Z, p10.h[w13]\n"
-      ".inst 0x25696161  // psel p1.h, p8.h/Z, p11.h[w13, #2]\n"
-      ".inst 0x25696143  // psel p3.h, p8.h/Z, p10.h[w13, #2]\n"
+      ".inst 0xe0572520  // ld1h { za0h.h[x13] }, p1/Z, [x9, x23, LSL #1]\n"
+      ".inst 0xe0572348  // ld1h { za1h.h[x13] }, p0/Z, [x26, x23, LSL #1]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0572360  // ld1h { za0h.h[x13] }, p0/Z, [x27, x23, LSL #1]\n"
-      ".inst 0x25287120  // psel p0.h, p12.h/Z, p9.h[w12]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0572b28  // ld1h { za1h.h[x13] }, p2/Z, [x25, x23, LSL #1]\n"
-      ".inst 0x25287122  // psel p2.h, p12.h/Z, p9.h[w12]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0572702  // ld1h { za0h.h[x13, #2] }, p1/Z, [x24, x23, LSL #1]\n"
-      ".inst 0x25387121  // psel p1.h, p12.h/Z, p9.h[w12, #1]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0x25696580  // psel p0.h, p9.h/Z, p12.h[w13, #2]\n"
+      ".inst 0x25696161  // psel p1.h, p8.h/Z, p11.h[w13, #2]\n"
+      ".inst 0xe0572302  // ld1h { za0h.h[x13, #2] }, p0/Z, [x24, x23, LSL #1]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe05726aa  // ld1h { za1h.h[x13, #2] }, p1/Z, [x21, x23, LSL #1]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0572eaa  // ld1h { za1h.h[x13, #2] }, p3/Z, [x21, x23, LSL #1]\n"
+      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
       "ldr x21, [x10, #0x8]\n"
       ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x25387120  // psel p0.h, p12.h/Z, p9.h[w12, #1]\n"
-      ".inst 0xe0a88acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
-      "whilelt p9.h, x16, %x[width]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b08acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x16, LSL #2]\n"
+      "whilelt p10.h, x14, %x[width]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
+      ".inst 0xe0bc86c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
+      "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
-      "inch x16\n"
-      ".inst 0xe0bc82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x28, LSL #2]\n"
+      ".inst 0xe0bb82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
       "addvl x22, x22, #4\n"
+      "inch x14\n"
       "inch x23\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x26, 11f\n"
+      "cbnz x25, 11f\n"
       "mov x11, %x[in]\n"
-      "whilelt p8.h, x16, %x[width]\n"
+      "whilelt p9.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25307123  // psel p3.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307122  // psel p2.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25396161  // psel p1.h, p8.h/Z, p11.h[w13, #1]\n"
-      ".inst 0x25396140  // psel p0.h, p8.h/Z, p10.h[w13, #1]\n"
-      ".inst 0xe0bf8ec0  // st1w { za0v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a88ac4  // st1w { za1v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x22, x22, #2\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "ldr x21, [x11, #0x0]\n"
-      "cmp x12, x8\n"
-      "ldr x20, [x11, x8, LSL #0x3]\n"
-      "add x11, x11, #0x8\n"
+      "add x12, x12, #0x1\n"
+      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
+      "ldr x20, [x11, x16, LSL #0x3]\n"
+      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
+      "cmp x12, x16\n"
       ".inst 0xe05726a1  // ld1h { za0h.h[x13, #1] }, p1/Z, [x21, x23, LSL #1]\n"
       ".inst 0xe0572289  // ld1h { za1h.h[x13, #1] }, p0/Z, [x20, x23, LSL #1]\n"
+      "add x11, x11, #0x8\n"
+      "addvl x22, x22, #2\n"
       "add x13, x13, #0x2\n"
       "blt 9b\n"
-      "whilelt p9.h, x16, %x[width]\n"
-      "whilelt p8.h, x16, %x[width]\n"
+      "whilelt p10.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
       "mov x20, #0x0\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: Second
-      ".inst 0x25307121  // psel p1.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307120  // psel p0.s, p12.s/Z, p9.s[w12]\n"
-      "add x20, x20, #0x2\n"
-      ".inst 0xe0bf86c8  // st1w { za2v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a882cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x8, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
+      "cmp x12, x17\n"
       "addvl x22, x22, #2\n"
-      "cmp x12, x7\n"
+      "add x20, x20, #0x2\n"
       "blt 10b\n"
-      "whilelt p8.h, x16, %x[width]\n"
+      "whilelt p8.h, x14, %x[width]\n"
       "b 13f\n"
       "11:"  // K loop: Tails: Odd
       "mov x12, #0x0\n"
       "12:"  // K loop: Tails: Odd: Loop
-      ".inst 0x25307121  // psel p1.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307120  // psel p0.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0xe0bf86c0  // st1w { za0v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a882c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x8, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
+      "cmp x12, x17\n"
       "addvl x22, x22, #2\n"
-      "cmp x12, x7\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x22\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp
index 7b42b6fb93..2d6e1ce6c7 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8.hpp
@@ -32,265 +32,265 @@ void interleave_block<2, 4, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "cntb x21\n"
       "mov x23, %x[width]\n"
-      "mov x21, %x[width]\n"
-      "cntb x20\n"
       "incb x23\n"
-      "sub x7, x20, #0x1\n"
-      "cntw x8\n"
+      "mov x20, %x[width]\n"
+      "sub x17, x21, #0x1\n"
+      "cntw x16\n"
       "sub x23, x23, #0x1\n"
-      "ands x7, x21, x7\n"
-      "udiv x23, x23, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x7, x7, x20, NE\n"
+      "ands x17, x20, x17\n"
+      "udiv x23, x23, x21\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x17, x17, x21, NE\n"
       "lsl x22, %x[height], #0x1\n"  // height * 2
-      "lsl x21, x8, #0x1\n"
+      "lsl x21, x16, #0x1\n"
       "sub x20, x23, #0x1\n"
-      "add x7, x7, #0x3\n"
-      "sub x17, x8, #0x2\n"
+      "add x17, x17, #0x3\n"
+      "sub x15, x16, #0x2\n"
       "whilelt p9.b, XZR, x22\n"
       "whilelt p8.b, x21, x22\n"
-      "mov x16, #0x0\n"
+      "mov x14, #0x0\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      "cntw x9, ALL, MUL #2\n"
-      "cntw x28, ALL, MUL #3\n"
-      "ldr x27, [x11, #0x0]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      "cntw x28, ALL, MUL #2\n"
+      "cntw x27, ALL, MUL #3\n"
+      "ldr x26, [x10, #0x0]\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x26, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "ldr x25, [x10, #0x0]\n"
-      "lsr x7, x7, #0x2\n"
-      "ptrue p11.s\n"
+      "and x25, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "ldr x24, [x11, #0x8]\n"
+      "lsr x17, x17, #0x2\n"
+      "ptrue p11.s\n"
+      "ldr x23, [x10, #0x8]\n"
       "zip1 p10.b, p9.b, p8.b\n"
-      "mov x23, %x[row_offset]\n"
-      "ldr x21, [x10, #0x8]\n"
-      "mov x22, %x[out]\n"
-      "whilelt p9.b, x16, %x[width]\n"
-      "whilelt p8.b, x16, %x[width]\n"
+      "mov x22, %x[row_offset]\n"
+      "mov x21, %x[out]\n"
+      "whilelt p9.b, x14, %x[width]\n"
+      "whilelt p8.b, x14, %x[width]\n"
       "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
       "mov x12, #0x0\n"
-      "cbz x17, 2f\n"
+      "cbz x15, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25246143  // psel p3.b, p8.b/Z, p10.b[w12]\n"
-      ".inst 0x252c6142  // psel p2.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0160120  // ld1b { za0h.b[x12] }, p0/Z, [x9, x22]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0160341  // ld1b { za0h.b[x12, #1] }, p0/Z, [x26, x22]\n"
       ".inst 0x25646141  // psel p1.b, p8.b/Z, p10.b[w12, #4]\n"
       ".inst 0x256c6140  // psel p0.b, p8.b/Z, p10.b[w12, #5]\n"
-      ".inst 0xe0170f60  // ld1b { za0h.b[x12] }, p3/Z, [x27, x23]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0170b21  // ld1b { za0h.b[x12, #1] }, p2/Z, [x25, x23]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0170704  // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x23]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0xe0160704  // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x22]\n"
       "ldr x24, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01702a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x23]\n"
+      ".inst 0xe01602e5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x23, x22]\n"
       "add x12, x12, #0x8\n"
-      "ldr x21, [x10, #0x8]\n"
+      "cmp x12, x15, LSL #2\n"
+      "ldr x23, [x10, #0x8]\n"
       "add x10, x10, #0x10\n"
-      "cmp x12, x17, LSL #2\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25246143  // psel p3.b, p8.b/Z, p10.b[w12]\n"
-      ".inst 0x252c6142  // psel p2.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0160120  // ld1b { za0h.b[x12] }, p0/Z, [x9, x22]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0xe0160341  // ld1b { za0h.b[x12, #1] }, p0/Z, [x26, x22]\n"
       ".inst 0x25646141  // psel p1.b, p8.b/Z, p10.b[w12, #4]\n"
       ".inst 0x256c6140  // psel p0.b, p8.b/Z, p10.b[w12, #5]\n"
+      ".inst 0xe0160704  // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x22]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0170f60  // ld1b { za0h.b[x12] }, p3/Z, [x27, x23]\n"
-      "ldr x27, [x11, #0x0]\n"
-      "incb x16\n"
-      ".inst 0xe0170b21  // ld1b { za0h.b[x12, #1] }, p2/Z, [x25, x23]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0170704  // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x23]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe01602e5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x23, x22]\n"
+      "ldr x26, [x10, #0x0]\n"
+      "incb x22\n"
+      "incb x14\n"
       "ldr x24, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01702a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
+      "ldr x23, [x10, #0x8]\n"
       "add x10, x10, #0x10\n"
-      "incb x23\n"
       "cbz x20, 8f\n"
       "mov x20, x20\n"
       "3:"  // K loop: Main loop
-      "whilelt p8.b, x16, %x[width]\n"
-      "mov x15, #0x0\n"
-      "mov x14, #0x0\n"
-      "cbz x17, 5f\n"
+      "whilelt p8.b, x14, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
+      "cbz x15, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25376143  // psel p3.b, p8.b/Z, p10.b[w15, #2]\n"
-      ".inst 0x253f6142  // psel p2.b, p8.b/Z, p10.b[w15, #3]\n"
-      ".inst 0x25776141  // psel p1.b, p8.b/Z, p10.b[w15, #6]\n"
-      ".inst 0x257f6140  // psel p0.b, p8.b/Z, p10.b[w15, #7]\n"
-      ".inst 0xe0176f62  // ld1b { za0h.b[x15, #2] }, p3/Z, [x27, x23]\n"
-      ".inst 0x25266d23  // psel p3.b, p11.b/Z, p9.b[w14]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0176b23  // ld1b { za0h.b[x15, #3] }, p2/Z, [x25, x23]\n"
-      ".inst 0x25266d22  // psel p2.b, p11.b/Z, p9.b[w14]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0176706  // ld1b { za0h.b[x15, #6] }, p1/Z, [x24, x23]\n"
-      ".inst 0x252e6d21  // psel p1.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162122  // ld1b { za0h.b[x13, #2] }, p0/Z, [x9, x22]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0162343  // ld1b { za0h.b[x13, #3] }, p0/Z, [x26, x22]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0x257d6142  // psel p2.b, p8.b/Z, p10.b[w13, #7]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0xe0162306  // ld1b { za0h.b[x13, #6] }, p0/Z, [x24, x22]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x24, [x11, #0x8]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0162ae7  // ld1b { za0h.b[x13, #7] }, p2/Z, [x23, x22]\n"
+      "ldr x23, [x10, #0x8]\n"
+      ".inst 0xe0bf86a0  // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01762a7  // ld1b { za0h.b[x15, #7] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0xe0bc86a1  // st1w { za0v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
       "add x10, x10, #0x10\n"
-      ".inst 0xe0bfcec0  // st1w { za0v.s[x14] }, p3/Z, [x22, XZR, LSL #2]\n"
-      "add x15, x15, #0x8\n"
-      ".inst 0xe0a8cac4  // st1w { za1v.s[x14] }, p2/Z, [x22, x8, LSL #2]\n"
-      ".inst 0xe0a9c6c1  // st1w { za0v.s[x14, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bcc2c5  // st1w { za1v.s[x14, #1] }, p0/Z, [x22, x28, LSL #2]\n"
-      "add x14, x14, #0x2\n"
-      "addvl x22, x22, #4\n"
-      "cmp x14, x17\n"
+      "add x13, x13, #0x8\n"
+      ".inst 0xe0bb82a5  // st1w { za1v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x15\n"
+      "addvl x21, x21, #4\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25376143  // psel p3.b, p8.b/Z, p10.b[w15, #2]\n"
-      ".inst 0x253f6142  // psel p2.b, p8.b/Z, p10.b[w15, #3]\n"
-      ".inst 0x25776141  // psel p1.b, p8.b/Z, p10.b[w15, #6]\n"
-      ".inst 0x257f6140  // psel p0.b, p8.b/Z, p10.b[w15, #7]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162122  // ld1b { za0h.b[x13, #2] }, p0/Z, [x9, x22]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      ".inst 0xe0162343  // ld1b { za0h.b[x13, #3] }, p0/Z, [x26, x22]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0176f62  // ld1b { za0h.b[x15, #2] }, p3/Z, [x27, x23]\n"
-      ".inst 0x25266d23  // psel p3.b, p11.b/Z, p9.b[w14]\n"
-      "ldr x27, [x11, #0x0]\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0176b23  // ld1b { za0h.b[x15, #3] }, p2/Z, [x25, x23]\n"
-      ".inst 0x25266d22  // psel p2.b, p11.b/Z, p9.b[w14]\n"
-      "ldr x25, [x10, #0x0]\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0176706  // ld1b { za0h.b[x15, #6] }, p1/Z, [x24, x23]\n"
-      ".inst 0x252e6d21  // psel p1.b, p11.b/Z, p9.b[w14, #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0162306  // ld1b { za0h.b[x13, #6] }, p0/Z, [x24, x22]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      ".inst 0x257d6141  // psel p1.b, p8.b/Z, p10.b[w13, #7]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e7  // ld1b { za0h.b[x13, #7] }, p1/Z, [x23, x22]\n"
       "ldr x24, [x11, #0x8]\n"
+      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
+      "ldr x23, [x10, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0b08aa4  // st1w { za1v.s[x12] }, p2/Z, [x21, x16, LSL #2]\n"
+      "whilelt p9.b, x14, %x[width]\n"
+      "incb x14\n"
+      ".inst 0xe0bc86a1  // st1w { za0v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01762a7  // ld1b { za0h.b[x15, #7] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      "whilelt p9.b, x16, %x[width]\n"
-      ".inst 0xe0bfcec0  // st1w { za0v.s[x14] }, p3/Z, [x22, XZR, LSL #2]\n"
-      "incb x16\n"
       "add x10, x10, #0x10\n"
-      ".inst 0xe0a8cac4  // st1w { za1v.s[x14] }, p2/Z, [x22, x8, LSL #2]\n"
-      "incb x23\n"
-      "whilelt p8.b, x16, %x[width]\n"
-      ".inst 0xe0a9c6c1  // st1w { za0v.s[x14, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bcc2c5  // st1w { za1v.s[x14, #1] }, p0/Z, [x22, x28, LSL #2]\n"
-      "addvl x22, x22, #4\n"
-      "cbz x17, 7f\n"
+      ".inst 0xe0bb82a5  // st1w { za1v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
+      "addvl x21, x21, #4\n"
+      "incb x22\n"
+      "whilelt p8.b, x14, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
+      "cbz x15, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x256d6140  // psel p0.b, p8.b/Z, p10.b[w13, #5]\n"
-      ".inst 0xe0172f60  // ld1b { za0h.b[x13] }, p3/Z, [x27, x23]\n"
-      ".inst 0x25246d23  // psel p3.b, p11.b/Z, p9.b[w12]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0172b21  // ld1b { za0h.b[x13, #1] }, p2/Z, [x25, x23]\n"
-      ".inst 0x25246d22  // psel p2.b, p11.b/Z, p9.b[w12]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0172704  // ld1b { za0h.b[x13, #4] }, p1/Z, [x24, x23]\n"
-      ".inst 0x252c6d21  // psel p1.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe0162120  // ld1b { za0h.b[x13] }, p0/Z, [x9, x22]\n"
+      ".inst 0x252d6140  // psel p0.b, p8.b/Z, p10.b[w13, #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0162341  // ld1b { za0h.b[x13, #1] }, p0/Z, [x26, x22]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
+      ".inst 0x256d6142  // psel p2.b, p8.b/Z, p10.b[w13, #5]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0xe0162304  // ld1b { za0h.b[x13, #4] }, p0/Z, [x24, x22]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x24, [x11, #0x8]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0162ae5  // ld1b { za0h.b[x13, #5] }, p2/Z, [x23, x22]\n"
+      "ldr x23, [x10, #0x8]\n"
+      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0b082ac  // st1w { za3v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01722a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0xe0bc86a9  // st1w { za2v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
       "add x10, x10, #0x10\n"
-      ".inst 0xe0bf8ec8  // st1w { za2v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
       "add x13, x13, #0x8\n"
-      ".inst 0xe0a88acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bc82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x28, LSL #2]\n"
+      ".inst 0xe0bb82ad  // st1w { za3v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
       "add x12, x12, #0x2\n"
-      "addvl x22, x22, #4\n"
-      "cmp x12, x17\n"
+      "cmp x12, x15\n"
+      "addvl x21, x21, #4\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x256d6140  // psel p0.b, p8.b/Z, p10.b[w13, #5]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe0162120  // ld1b { za0h.b[x13] }, p0/Z, [x9, x22]\n"
+      ".inst 0x252d6140  // psel p0.b, p8.b/Z, p10.b[w13, #1]\n"
+      ".inst 0xe0162341  // ld1b { za0h.b[x13, #1] }, p0/Z, [x26, x22]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0172f60  // ld1b { za0h.b[x13] }, p3/Z, [x27, x23]\n"
-      ".inst 0x25246d23  // psel p3.b, p11.b/Z, p9.b[w12]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0172b21  // ld1b { za0h.b[x13, #1] }, p2/Z, [x25, x23]\n"
-      ".inst 0x25246d22  // psel p2.b, p11.b/Z, p9.b[w12]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0172704  // ld1b { za0h.b[x13, #4] }, p1/Z, [x24, x23]\n"
-      ".inst 0x252c6d21  // psel p1.b, p11.b/Z, p9.b[w12, #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0162304  // ld1b { za0h.b[x13, #4] }, p0/Z, [x24, x22]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      ".inst 0x256d6141  // psel p1.b, p8.b/Z, p10.b[w13, #5]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e5  // ld1b { za0h.b[x13, #5] }, p1/Z, [x23, x22]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe01722a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      "whilelt p9.b, x16, %x[width]\n"
-      ".inst 0xe0bf8ec8  // st1w { za2v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
+      "ldr x23, [x10, #0x8]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0b08aac  // st1w { za3v.s[x12] }, p2/Z, [x21, x16, LSL #2]\n"
+      "whilelt p9.b, x14, %x[width]\n"
       "subs x20, x20, #0x1\n"
+      ".inst 0xe0bc86a9  // st1w { za2v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
+      "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
-      ".inst 0xe0a88acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
-      "incb x16\n"
-      "incb x23\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bc82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x28, LSL #2]\n"
-      "addvl x22, x22, #4\n"
+      ".inst 0xe0bb82ad  // st1w { za3v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
+      "addvl x21, x21, #4\n"
+      "incb x14\n"
+      "incb x22\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x26, 11f\n"
+      "cbnz x25, 11f\n"
       "mov x11, %x[in]\n"
-      "whilelt p8.b, x16, %x[width]\n"
+      "whilelt p8.b, x14, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25306d23  // psel p3.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      ".inst 0xe0bf8ec0  // st1w { za0v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a88ac4  // st1w { za1v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
+      "ldr x20, [x11, #0x0]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162282  // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x22]\n"
+      "ldr x20, [x11, x16, LSL #0x3]\n"
       "add x12, x12, #0x1\n"
-      "addvl x22, x22, #2\n"
-      "ldr x21, [x11, #0x0]\n"
-      "cmp x12, x8\n"
-      "ldr x20, [x11, x8, LSL #0x3]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      ".inst 0xe0162283  // ld1b { za0h.b[x13, #3] }, p0/Z, [x20, x22]\n"
+      "cmp x12, x16\n"
       "add x11, x11, #0x8\n"
-      ".inst 0xe01726a2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x21, x23]\n"
-      ".inst 0xe0172283  // ld1b { za0h.b[x13, #3] }, p0/Z, [x20, x23]\n"
+      "addvl x21, x21, #2\n"
       "add x13, x13, #0x4\n"
       "blt 9b\n"
-      "whilelt p9.b, x16, %x[width]\n"
-      "whilelt p8.b, x16, %x[width]\n"
+      "whilelt p9.b, x14, %x[width]\n"
+      "whilelt p8.b, x14, %x[width]\n"
       "mov x20, #0x0\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: Second
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      "add x20, x20, #0x4\n"
-      ".inst 0xe0bf86c8  // st1w { za2v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a882cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x8, LSL #2]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0b082ac  // st1w { za3v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x22, x22, #2\n"
-      "cmp x12, x7\n"
+      "cmp x12, x17\n"
+      "addvl x21, x21, #2\n"
+      "add x20, x20, #0x4\n"
       "blt 10b\n"
-      "whilelt p8.b, x16, %x[width]\n"
+      "whilelt p8.b, x14, %x[width]\n"
       "b 13f\n"
       "11:"  // K loop: Tails: Odd
       "mov x12, #0x0\n"
       "12:"  // K loop: Tails: Odd: Loop
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xe0bf86c0  // st1w { za0v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a882c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x8, LSL #2]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x22, x22, #2\n"
-      "cmp x12, x7\n"
+      "cmp x12, x17\n"
+      "addvl x21, x21, #2\n"
       "blt 12b\n"
       "13:"  // K loop: End
-      "mov %x[out], x22\n"
+      "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8_summing.hpp
index 6930bf4056..27b9bc3806 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8_summing.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_s8_s8_summing.hpp
@@ -32,321 +32,321 @@ void interleave_block<2, 4, VLType::SME, true>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "cntb x21\n"
       "mov x23, %x[width]\n"
-      "mov x21, %x[width]\n"
-      "cntb x20\n"
-      "incb x23\n"
       "mov z20.b, #0x1\n"
+      "incb x23\n"
+      "mov x20, %x[width]\n"
       "mov z19.s, #0x0\n"
-      "sub x7, x20, #0x1\n"
-      "cntw x8\n"
       "mov z18.s, #0x0\n"
+      "sub x17, x21, #0x1\n"
+      "cntw x16\n"
       "sub x23, x23, #0x1\n"
-      "ands x7, x21, x7\n"
-      "udiv x23, x23, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x7, x7, x20, NE\n"
+      "ands x17, x20, x17\n"
+      "udiv x23, x23, x21\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x17, x17, x21, NE\n"
       "lsl x22, %x[height], #0x1\n"  // height * 2
-      "lsl x21, x8, #0x1\n"
+      "lsl x21, x16, #0x1\n"
       "sub x20, x23, #0x1\n"
-      "add x7, x7, #0x3\n"
+      "add x17, x17, #0x3\n"
       "whilelt p9.b, XZR, x22\n"
       "whilelt p8.b, x21, x22\n"
-      "mov x17, #0x0\n"
-      "cntw x16, ALL, MUL #2\n"
+      "mov x15, #0x0\n"
+      "cntw x14, ALL, MUL #2\n"
       "cntw x11, ALL, MUL #3\n"
       "ptrue p4.b\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
       "and x10, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "lsr x7, x7, #0x2\n"
-      "sub x9, x8, #0x2\n"
+      "lsr x17, x17, #0x2\n"
+      "sub x9, x16, #0x2\n"
       "ptrue p11.s\n"
       "zip1 p10.b, p9.b, p8.b\n"
       "mov x28, %x[row_offset]\n"
       "mov x27, %x[out]\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p9.b, x15, %x[width]\n"
+      "whilelt p8.b, x15, %x[width]\n"
       "cbnz %x[first], 1f\n"
       "addvl x27, x27, #-2\n"
       "ld1w { z19.s }, p4/Z, [x27]\n"
       "ld1w { z18.s }, p4/Z, [x27, #1, MUL VL]\n"
       "1:"  // K loop: Load row sums: End
       "mov x26, %x[in]\n"
-      "add x25, %x[in], x8, LSL #3\n"
+      "add x25, %x[in], x16, LSL #3\n"
       "ldr x24, [x26, #0x0]\n"
-      "mov x12, #0x0\n"
       "ldr x23, [x25, #0x0]\n"
+      "mov x12, #0x0\n"
       "ldr x22, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
       "ldr x21, [x25, #0x8]\n"
       "add x25, x25, #0x10\n"
       "cbz x9, 3f\n"
       "2:"  // K loop: Charge: Loop
-      ".inst 0x25246143  // psel p3.b, p8.b/Z, p10.b[w12]\n"
-      ".inst 0x252c6142  // psel p2.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe01c0300  // ld1b { za0h.b[x12] }, p0/Z, [x24, x28]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      "ldr x24, [x26, #0x0]\n"
+      ".inst 0xe01c02e1  // ld1b { za0h.b[x12, #1] }, p0/Z, [x23, x28]\n"
       ".inst 0x25646141  // psel p1.b, p8.b/Z, p10.b[w12, #4]\n"
       ".inst 0x256c6140  // psel p0.b, p8.b/Z, p10.b[w12, #5]\n"
-      ".inst 0xe01c0f00  // ld1b { za0h.b[x12] }, p3/Z, [x24, x28]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe01c0ae1  // ld1b { za0h.b[x12, #1] }, p2/Z, [x23, x28]\n"
       "ldr x23, [x25, #0x0]\n"
       ".inst 0xe01c06c4  // ld1b { za0h.b[x12, #4] }, p1/Z, [x22, x28]\n"
       "ldr x22, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
       ".inst 0xe01c02a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x28]\n"
       "add x12, x12, #0x8\n"
+      "cmp x12, x9, LSL #2\n"
       "ldr x21, [x25, #0x8]\n"
       "add x25, x25, #0x10\n"
-      "cmp x12, x9, LSL #2\n"
       "blt 2b\n"
       "3:"  // K loop: Charge: End
-      ".inst 0x25246143  // psel p3.b, p8.b/Z, p10.b[w12]\n"
-      ".inst 0x252c6142  // psel p2.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe01c0300  // ld1b { za0h.b[x12] }, p0/Z, [x24, x28]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0xe01c02e1  // ld1b { za0h.b[x12, #1] }, p0/Z, [x23, x28]\n"
       ".inst 0x25646141  // psel p1.b, p8.b/Z, p10.b[w12, #4]\n"
       ".inst 0x256c6140  // psel p0.b, p8.b/Z, p10.b[w12, #5]\n"
+      ".inst 0xe01c06c4  // ld1b { za0h.b[x12, #4] }, p1/Z, [x22, x28]\n"
       "mov x26, %x[in]\n"
-      "add x25, %x[in], x8, LSL #3\n"
-      ".inst 0xe01c0f00  // ld1b { za0h.b[x12] }, p3/Z, [x24, x28]\n"
+      "add x25, %x[in], x16, LSL #3\n"
       "ldr x24, [x26, #0x0]\n"
-      "incb x17\n"
-      ".inst 0xe01c0ae1  // ld1b { za0h.b[x12, #1] }, p2/Z, [x23, x28]\n"
+      ".inst 0xe01c02a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x28]\n"
       "ldr x23, [x25, #0x0]\n"
-      ".inst 0xe01c06c4  // ld1b { za0h.b[x12, #4] }, p1/Z, [x22, x28]\n"
+      "incb x28\n"
+      "incb x15\n"
       "ldr x22, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      ".inst 0xe01c02a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x28]\n"
       "ldr x21, [x25, #0x8]\n"
       "add x25, x25, #0x10\n"
-      "incb x28\n"
       "cbz x20, 9f\n"
       "mov x20, x20\n"
       "4:"  // K loop: Main loop
-      "whilelt p8.b, x17, %x[width]\n"
-      "mov x15, #0x0\n"
-      "mov x14, #0x0\n"
+      "whilelt p8.b, x15, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
       "cbz x9, 6f\n"
       "5:"  // K loop: Main loop: First: Loop
-      ".inst 0x25376143  // psel p3.b, p8.b/Z, p10.b[w15, #2]\n"
-      ".inst 0x253f6142  // psel p2.b, p8.b/Z, p10.b[w15, #3]\n"
-      ".inst 0x25776141  // psel p1.b, p8.b/Z, p10.b[w15, #6]\n"
-      ".inst 0x257f6140  // psel p0.b, p8.b/Z, p10.b[w15, #7]\n"
-      ".inst 0xe01c6f02  // ld1b { za0h.b[x15, #2] }, p3/Z, [x24, x28]\n"
-      ".inst 0x25266d23  // psel p3.b, p11.b/Z, p9.b[w14]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe01c2302  // ld1b { za0h.b[x13, #2] }, p0/Z, [x24, x28]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
       "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe01c6ae3  // ld1b { za0h.b[x15, #3] }, p2/Z, [x23, x28]\n"
-      ".inst 0x25266d22  // psel p2.b, p11.b/Z, p9.b[w14]\n"
+      ".inst 0xe01c22e3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x23, x28]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0x257d6142  // psel p2.b, p8.b/Z, p10.b[w13, #7]\n"
       "ldr x23, [x25, #0x0]\n"
-      ".inst 0xe01c66c6  // ld1b { za0h.b[x15, #6] }, p1/Z, [x22, x28]\n"
-      ".inst 0x252e6d21  // psel p1.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0xe01c22c6  // ld1b { za0h.b[x13, #6] }, p0/Z, [x22, x28]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0xe01c62a7  // ld1b { za0h.b[x15, #7] }, p0/Z, [x21, x28]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01c2aa7  // ld1b { za0h.b[x13, #7] }, p2/Z, [x21, x28]\n"
       "ldr x21, [x25, #0x8]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0xe0bfcf60  // st1w { za0v.s[x14] }, p3/Z, [x27, XZR, LSL #2]\n"
-      ".inst 0xc082d011  // mova z17.s, p4/M, za0v.s[x14]\n"
-      "add x15, x15, #0x8\n"
-      ".inst 0xc082d090  // mova z16.s, p4/M, za1v.s[x14]\n"
+      ".inst 0xe0bf8760  // st1w { za0v.s[x12] }, p1/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0xc0829011  // mova z17.s, p4/M, za0v.s[x12]\n"
+      ".inst 0xc0829090  // mova z16.s, p4/M, za1v.s[x12]\n"
       "sdot z19.s, z17.b, z20.b\n"
-      ".inst 0xe0a8cb64  // st1w { za1v.s[x14] }, p2/Z, [x27, x8, LSL #2]\n"
+      ".inst 0xe0b08364  // st1w { za1v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "sdot z18.s, z16.b, z20.b\n"
-      ".inst 0xe0b0c761  // st1w { za0v.s[x14, #1] }, p1/Z, [x27, x16, LSL #2]\n"
-      ".inst 0xc082d031  // mova z17.s, p4/M, za0v.s[x14, #1]\n"
-      ".inst 0xe0abc365  // st1w { za1v.s[x14, #1] }, p0/Z, [x27, x11, LSL #2]\n"
-      ".inst 0xc082d0b0  // mova z16.s, p4/M, za1v.s[x14, #1]\n"
-      "add x14, x14, #0x2\n"
-      "addvl x27, x27, #4\n"
-      "cmp x14, x9\n"
+      ".inst 0xe0ae8361  // st1w { za0v.s[x12, #1] }, p0/Z, [x27, x14, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0829031  // mova z17.s, p4/M, za0v.s[x12, #1]\n"
+      ".inst 0xe0ab8365  // st1w { za1v.s[x12, #1] }, p0/Z, [x27, x11, LSL #2]\n"
+      ".inst 0xc08290b0  // mova z16.s, p4/M, za1v.s[x12, #1]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x9\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "sdot z19.s, z17.b, z20.b\n"
       "sdot z18.s, z16.b, z20.b\n"
+      "addvl x27, x27, #4\n"
+      "add x13, x13, #0x8\n"
       "blt 5b\n"
       "6:"  // K loop: Main loop: First: Tail
-      ".inst 0x25376143  // psel p3.b, p8.b/Z, p10.b[w15, #2]\n"
-      ".inst 0x253f6142  // psel p2.b, p8.b/Z, p10.b[w15, #3]\n"
-      ".inst 0x25776141  // psel p1.b, p8.b/Z, p10.b[w15, #6]\n"
-      ".inst 0x257f6140  // psel p0.b, p8.b/Z, p10.b[w15, #7]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe01c2302  // ld1b { za0h.b[x13, #2] }, p0/Z, [x24, x28]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      ".inst 0xe01c22e3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x23, x28]\n"
+      ".inst 0x25756141  // psel p1.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0x257d6140  // psel p0.b, p8.b/Z, p10.b[w13, #7]\n"
+      ".inst 0xe01c26c6  // ld1b { za0h.b[x13, #6] }, p1/Z, [x22, x28]\n"
       "mov x26, %x[in]\n"
-      "add x25, %x[in], x8, LSL #3\n"
-      ".inst 0xe01c6f02  // ld1b { za0h.b[x15, #2] }, p3/Z, [x24, x28]\n"
-      ".inst 0x25266d23  // psel p3.b, p11.b/Z, p9.b[w14]\n"
+      "add x25, %x[in], x16, LSL #3\n"
       "ldr x24, [x26, #0x0]\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe01c6ae3  // ld1b { za0h.b[x15, #3] }, p2/Z, [x23, x28]\n"
-      ".inst 0x25266d22  // psel p2.b, p11.b/Z, p9.b[w14]\n"
+      ".inst 0xe01c22a7  // ld1b { za0h.b[x13, #7] }, p0/Z, [x21, x28]\n"
+      ".inst 0xc0829011  // mova z17.s, p4/M, za0v.s[x12]\n"
+      ".inst 0x25306d23  // psel p3.s, p11.s/Z, p9.s[w12]\n"
+      "sdot z19.s, z17.b, z20.b\n"
+      ".inst 0xc0829090  // mova z16.s, p4/M, za1v.s[x12]\n"
+      "sdot z18.s, z16.b, z20.b\n"
       "ldr x23, [x25, #0x0]\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe01c66c6  // ld1b { za0h.b[x15, #6] }, p1/Z, [x22, x28]\n"
-      ".inst 0x252e6d21  // psel p1.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0xe01c62a7  // ld1b { za0h.b[x15, #7] }, p0/Z, [x21, x28]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0829031  // mova z17.s, p4/M, za0v.s[x12, #1]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "ldr x21, [x25, #0x8]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      ".inst 0xc082d011  // mova z17.s, p4/M, za0v.s[x14]\n"
-      ".inst 0xe0bfcf60  // st1w { za0v.s[x14] }, p3/Z, [x27, XZR, LSL #2]\n"
-      "incb x17\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0xc082d090  // mova z16.s, p4/M, za1v.s[x14]\n"
+      ".inst 0xe0bf8f60  // st1w { za0v.s[x12] }, p3/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0xc08290b0  // mova z16.s, p4/M, za1v.s[x12, #1]\n"
+      "whilelt p9.b, x15, %x[width]\n"
+      ".inst 0xe0b08b64  // st1w { za1v.s[x12] }, p2/Z, [x27, x16, LSL #2]\n"
+      "incb x15\n"
+      "add x26, x26, #0x10\n"
       "sdot z19.s, z17.b, z20.b\n"
-      ".inst 0xe0a8cb64  // st1w { za1v.s[x14] }, p2/Z, [x27, x8, LSL #2]\n"
-      "incb x28\n"
+      ".inst 0xe0ae8761  // st1w { za0v.s[x12, #1] }, p1/Z, [x27, x14, LSL #2]\n"
+      "add x25, x25, #0x10\n"
       "sdot z18.s, z16.b, z20.b\n"
-      ".inst 0xe0b0c761  // st1w { za0v.s[x14, #1] }, p1/Z, [x27, x16, LSL #2]\n"
-      ".inst 0xc082d031  // mova z17.s, p4/M, za0v.s[x14, #1]\n"
-      "whilelt p8.b, x17, %x[width]\n"
-      ".inst 0xc082d0b0  // mova z16.s, p4/M, za1v.s[x14, #1]\n"
-      ".inst 0xe0abc365  // st1w { za1v.s[x14, #1] }, p0/Z, [x27, x11, LSL #2]\n"
+      "incb x28\n"
+      ".inst 0xe0ab8365  // st1w { za1v.s[x12, #1] }, p0/Z, [x27, x11, LSL #2]\n"
       "addvl x27, x27, #4\n"
-      "sdot z19.s, z17.b, z20.b\n"
-      "sdot z18.s, z16.b, z20.b\n"
+      "whilelt p8.b, x15, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
       "cbz x9, 8f\n"
       "7:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x256d6140  // psel p0.b, p8.b/Z, p10.b[w13, #5]\n"
-      ".inst 0xe01c2f00  // ld1b { za0h.b[x13] }, p3/Z, [x24, x28]\n"
-      ".inst 0x25246d23  // psel p3.b, p11.b/Z, p9.b[w12]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe01c2300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x28]\n"
+      ".inst 0x252d6140  // psel p0.b, p8.b/Z, p10.b[w13, #1]\n"
       "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe01c2ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x28]\n"
-      ".inst 0x25246d22  // psel p2.b, p11.b/Z, p9.b[w12]\n"
+      ".inst 0xe01c22e1  // ld1b { za0h.b[x13, #1] }, p0/Z, [x23, x28]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
+      ".inst 0x256d6142  // psel p2.b, p8.b/Z, p10.b[w13, #5]\n"
       "ldr x23, [x25, #0x0]\n"
-      ".inst 0xe01c26c4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x22, x28]\n"
-      ".inst 0x252c6d21  // psel p1.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0xe01c22c4  // ld1b { za0h.b[x13, #4] }, p0/Z, [x22, x28]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0xe01c22a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x28]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01c2aa5  // ld1b { za0h.b[x13, #5] }, p2/Z, [x21, x28]\n"
       "ldr x21, [x25, #0x8]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0xe0bf8f68  // st1w { za2v.s[x12] }, p3/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0xe0bf8768  // st1w { za2v.s[x12] }, p1/Z, [x27, XZR, LSL #2]\n"
       ".inst 0xc0829111  // mova z17.s, p4/M, za2v.s[x12]\n"
-      "add x13, x13, #0x8\n"
       ".inst 0xc0829190  // mova z16.s, p4/M, za3v.s[x12]\n"
       "sdot z19.s, z17.b, z20.b\n"
-      ".inst 0xe0a88b6c  // st1w { za3v.s[x12] }, p2/Z, [x27, x8, LSL #2]\n"
+      ".inst 0xe0b0836c  // st1w { za3v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "sdot z18.s, z16.b, z20.b\n"
-      ".inst 0xe0b08769  // st1w { za2v.s[x12, #1] }, p1/Z, [x27, x16, LSL #2]\n"
+      ".inst 0xe0ae8369  // st1w { za2v.s[x12, #1] }, p0/Z, [x27, x14, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       ".inst 0xc0829131  // mova z17.s, p4/M, za2v.s[x12, #1]\n"
       ".inst 0xe0ab836d  // st1w { za3v.s[x12, #1] }, p0/Z, [x27, x11, LSL #2]\n"
       ".inst 0xc08291b0  // mova z16.s, p4/M, za3v.s[x12, #1]\n"
       "add x12, x12, #0x2\n"
-      "addvl x27, x27, #4\n"
       "cmp x12, x9\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "sdot z19.s, z17.b, z20.b\n"
       "sdot z18.s, z16.b, z20.b\n"
+      "addvl x27, x27, #4\n"
+      "add x13, x13, #0x8\n"
       "blt 7b\n"
       "8:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe01c2300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x28]\n"
+      ".inst 0x252d6140  // psel p0.b, p8.b/Z, p10.b[w13, #1]\n"
+      ".inst 0xe01c22e1  // ld1b { za0h.b[x13, #1] }, p0/Z, [x23, x28]\n"
       ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
       ".inst 0x256d6140  // psel p0.b, p8.b/Z, p10.b[w13, #5]\n"
+      ".inst 0xe01c26c4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x22, x28]\n"
       "mov x26, %x[in]\n"
-      "add x25, %x[in], x8, LSL #3\n"
-      ".inst 0xe01c2f00  // ld1b { za0h.b[x13] }, p3/Z, [x24, x28]\n"
-      ".inst 0x25246d23  // psel p3.b, p11.b/Z, p9.b[w12]\n"
+      "add x25, %x[in], x16, LSL #3\n"
       "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe01c2ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x28]\n"
-      ".inst 0x25246d22  // psel p2.b, p11.b/Z, p9.b[w12]\n"
+      ".inst 0xe01c22a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x28]\n"
+      ".inst 0xc0829111  // mova z17.s, p4/M, za2v.s[x12]\n"
+      ".inst 0x25306d23  // psel p3.s, p11.s/Z, p9.s[w12]\n"
+      "sdot z19.s, z17.b, z20.b\n"
+      ".inst 0xc0829190  // mova z16.s, p4/M, za3v.s[x12]\n"
+      "sdot z18.s, z16.b, z20.b\n"
       "ldr x23, [x25, #0x0]\n"
-      ".inst 0xe01c26c4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x22, x28]\n"
-      ".inst 0x252c6d21  // psel p1.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0xe01c22a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x28]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0829131  // mova z17.s, p4/M, za2v.s[x12, #1]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "ldr x21, [x25, #0x8]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      ".inst 0xc0829111  // mova z17.s, p4/M, za2v.s[x12]\n"
       ".inst 0xe0bf8f68  // st1w { za2v.s[x12] }, p3/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0xc08291b0  // mova z16.s, p4/M, za3v.s[x12, #1]\n"
+      "whilelt p9.b, x15, %x[width]\n"
+      ".inst 0xe0b08b6c  // st1w { za3v.s[x12] }, p2/Z, [x27, x16, LSL #2]\n"
       "subs x20, x20, #0x1\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0xc0829190  // mova z16.s, p4/M, za3v.s[x12]\n"
+      "add x26, x26, #0x10\n"
       "sdot z19.s, z17.b, z20.b\n"
-      ".inst 0xe0a88b6c  // st1w { za3v.s[x12] }, p2/Z, [x27, x8, LSL #2]\n"
-      "incb x17\n"
+      ".inst 0xe0ae8769  // st1w { za2v.s[x12, #1] }, p1/Z, [x27, x14, LSL #2]\n"
+      "add x25, x25, #0x10\n"
       "sdot z18.s, z16.b, z20.b\n"
-      ".inst 0xe0b08769  // st1w { za2v.s[x12, #1] }, p1/Z, [x27, x16, LSL #2]\n"
-      ".inst 0xc0829131  // mova z17.s, p4/M, za2v.s[x12, #1]\n"
-      "incb x28\n"
-      ".inst 0xc08291b0  // mova z16.s, p4/M, za3v.s[x12, #1]\n"
+      "incb x15\n"
       ".inst 0xe0ab836d  // st1w { za3v.s[x12, #1] }, p0/Z, [x27, x11, LSL #2]\n"
       "addvl x27, x27, #4\n"
-      "sdot z19.s, z17.b, z20.b\n"
-      "sdot z18.s, z16.b, z20.b\n"
+      "incb x28\n"
       "bgt 4b\n"
       "9:"  // K loop: Tails
       "cbnz x10, 12f\n"
       "mov x26, %x[in]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p8.b, x15, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: First
-      ".inst 0x25306d23  // psel p3.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf8360  // st1w { za0v.s[x12] }, p0/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xc0829011  // mova z17.s, p4/M, za0v.s[x12]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      ".inst 0xe0b08364  // st1w { za1v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
+      "ldr x21, [x26, #0x0]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
       ".inst 0xc0829090  // mova z16.s, p4/M, za1v.s[x12]\n"
+      "ldr x20, [x26, x16, LSL #0x3]\n"
+      ".inst 0xe01c22a2  // ld1b { za0h.b[x13, #2] }, p0/Z, [x21, x28]\n"
+      "add x12, x12, #0x1\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      "cmp x12, x16\n"
       "sdot z19.s, z17.b, z20.b\n"
       "sdot z18.s, z16.b, z20.b\n"
-      ".inst 0xe0bf8f60  // st1w { za0v.s[x12] }, p3/Z, [x27, XZR, LSL #2]\n"
-      ".inst 0xe0a88b64  // st1w { za1v.s[x12] }, p2/Z, [x27, x8, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x27, x27, #2\n"
-      "ldr x21, [x26, #0x0]\n"
-      "cmp x12, x8\n"
-      "ldr x20, [x26, x8, LSL #0x3]\n"
-      "add x26, x26, #0x8\n"
-      ".inst 0xe01c26a2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x21, x28]\n"
       ".inst 0xe01c2283  // ld1b { za0h.b[x13, #3] }, p0/Z, [x20, x28]\n"
+      "add x26, x26, #0x8\n"
+      "addvl x27, x27, #2\n"
       "add x13, x13, #0x4\n"
       "blt 10b\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p9.b, x15, %x[width]\n"
+      "whilelt p8.b, x15, %x[width]\n"
       "mov x20, #0x0\n"
       "mov x12, #0x0\n"
       "11:"  // K loop: Tails: Even: Second
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf8368  // st1w { za2v.s[x12] }, p0/Z, [x27, XZR, LSL #2]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xc0829111  // mova z17.s, p4/M, za2v.s[x12]\n"
+      ".inst 0xe0b0836c  // st1w { za3v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
       ".inst 0xc0829190  // mova z16.s, p4/M, za3v.s[x12]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x17\n"
       "sdot z19.s, z17.b, z20.b\n"
-      "add x20, x20, #0x4\n"
       "sdot z18.s, z16.b, z20.b\n"
-      ".inst 0xe0bf8768  // st1w { za2v.s[x12] }, p1/Z, [x27, XZR, LSL #2]\n"
-      ".inst 0xe0a8836c  // st1w { za3v.s[x12] }, p0/Z, [x27, x8, LSL #2]\n"
-      "add x12, x12, #0x1\n"
       "addvl x27, x27, #2\n"
-      "cmp x12, x7\n"
+      "add x20, x20, #0x4\n"
       "blt 11b\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p8.b, x15, %x[width]\n"
       "b 14f\n"
       "12:"  // K loop: Tails: Odd
       "mov x12, #0x0\n"
       "13:"  // K loop: Tails: Odd: Loop
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf8360  // st1w { za0v.s[x12] }, p0/Z, [x27, XZR, LSL #2]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xc0829011  // mova z17.s, p4/M, za0v.s[x12]\n"
+      ".inst 0xe0b08364  // st1w { za1v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
       ".inst 0xc0829090  // mova z16.s, p4/M, za1v.s[x12]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x17\n"
       "sdot z19.s, z17.b, z20.b\n"
       "sdot z18.s, z16.b, z20.b\n"
-      ".inst 0xe0bf8760  // st1w { za0v.s[x12] }, p1/Z, [x27, XZR, LSL #2]\n"
-      ".inst 0xe0a88364  // st1w { za1v.s[x12] }, p0/Z, [x27, x8, LSL #2]\n"
-      "add x12, x12, #0x1\n"
       "addvl x27, x27, #2\n"
-      "cmp x12, x7\n"
       "blt 13b\n"
       "14:"  // K loop: End
       "st1w { z19.s }, p4, [x27]\n"
       "st1w { z18.s }, p4, [x27, #1, MUL VL]\n"
       "addvl x27, x27, #2\n"
-      ".inst 0xd503467f  // SMSTOP\n"
       "mov %x[out], x27\n"
+      ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [first] "r" (first), [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_u8_u8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_u8_u8.hpp
index 9ce93ed95c..3f3863720a 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_u8_u8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_u8_u8.hpp
@@ -32,265 +32,265 @@ void interleave_block<2, 4, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "cntb x21\n"
       "mov x23, %x[width]\n"
-      "mov x21, %x[width]\n"
-      "cntb x20\n"
       "incb x23\n"
-      "sub x7, x20, #0x1\n"
-      "cntw x8\n"
+      "mov x20, %x[width]\n"
+      "sub x17, x21, #0x1\n"
+      "cntw x16\n"
       "sub x23, x23, #0x1\n"
-      "ands x7, x21, x7\n"
-      "udiv x23, x23, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x7, x7, x20, NE\n"
+      "ands x17, x20, x17\n"
+      "udiv x23, x23, x21\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x17, x17, x21, NE\n"
       "lsl x22, %x[height], #0x1\n"  // height * 2
-      "lsl x21, x8, #0x1\n"
+      "lsl x21, x16, #0x1\n"
       "sub x20, x23, #0x1\n"
-      "add x7, x7, #0x3\n"
-      "sub x17, x8, #0x2\n"
+      "add x17, x17, #0x3\n"
+      "sub x15, x16, #0x2\n"
       "whilelt p9.b, XZR, x22\n"
       "whilelt p8.b, x21, x22\n"
-      "mov x16, #0x0\n"
+      "mov x14, #0x0\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      "cntw x9, ALL, MUL #2\n"
-      "cntw x28, ALL, MUL #3\n"
-      "ldr x27, [x11, #0x0]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      "cntw x28, ALL, MUL #2\n"
+      "cntw x27, ALL, MUL #3\n"
+      "ldr x26, [x10, #0x0]\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "and x26, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "ldr x25, [x10, #0x0]\n"
-      "lsr x7, x7, #0x2\n"
-      "ptrue p11.s\n"
+      "and x25, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
       "ldr x24, [x11, #0x8]\n"
+      "lsr x17, x17, #0x2\n"
+      "ptrue p11.s\n"
+      "ldr x23, [x10, #0x8]\n"
       "zip1 p10.b, p9.b, p8.b\n"
-      "mov x23, %x[row_offset]\n"
-      "ldr x21, [x10, #0x8]\n"
-      "mov x22, %x[out]\n"
-      "whilelt p9.b, x16, %x[width]\n"
-      "whilelt p8.b, x16, %x[width]\n"
+      "mov x22, %x[row_offset]\n"
+      "mov x21, %x[out]\n"
+      "whilelt p9.b, x14, %x[width]\n"
+      "whilelt p8.b, x14, %x[width]\n"
       "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
       "mov x12, #0x0\n"
-      "cbz x17, 2f\n"
+      "cbz x15, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25246143  // psel p3.b, p8.b/Z, p10.b[w12]\n"
-      ".inst 0x252c6142  // psel p2.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0160120  // ld1b { za0h.b[x12] }, p0/Z, [x9, x22]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0160341  // ld1b { za0h.b[x12, #1] }, p0/Z, [x26, x22]\n"
       ".inst 0x25646141  // psel p1.b, p8.b/Z, p10.b[w12, #4]\n"
       ".inst 0x256c6140  // psel p0.b, p8.b/Z, p10.b[w12, #5]\n"
-      ".inst 0xe0170f60  // ld1b { za0h.b[x12] }, p3/Z, [x27, x23]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0170b21  // ld1b { za0h.b[x12, #1] }, p2/Z, [x25, x23]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0170704  // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x23]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0xe0160704  // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x22]\n"
       "ldr x24, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01702a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x23]\n"
+      ".inst 0xe01602e5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x23, x22]\n"
       "add x12, x12, #0x8\n"
-      "ldr x21, [x10, #0x8]\n"
+      "cmp x12, x15, LSL #2\n"
+      "ldr x23, [x10, #0x8]\n"
       "add x10, x10, #0x10\n"
-      "cmp x12, x17, LSL #2\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25246143  // psel p3.b, p8.b/Z, p10.b[w12]\n"
-      ".inst 0x252c6142  // psel p2.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0160120  // ld1b { za0h.b[x12] }, p0/Z, [x9, x22]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0xe0160341  // ld1b { za0h.b[x12, #1] }, p0/Z, [x26, x22]\n"
       ".inst 0x25646141  // psel p1.b, p8.b/Z, p10.b[w12, #4]\n"
       ".inst 0x256c6140  // psel p0.b, p8.b/Z, p10.b[w12, #5]\n"
+      ".inst 0xe0160704  // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x22]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0170f60  // ld1b { za0h.b[x12] }, p3/Z, [x27, x23]\n"
-      "ldr x27, [x11, #0x0]\n"
-      "incb x16\n"
-      ".inst 0xe0170b21  // ld1b { za0h.b[x12, #1] }, p2/Z, [x25, x23]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0170704  // ld1b { za0h.b[x12, #4] }, p1/Z, [x24, x23]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe01602e5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x23, x22]\n"
+      "ldr x26, [x10, #0x0]\n"
+      "incb x22\n"
+      "incb x14\n"
       "ldr x24, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01702a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
+      "ldr x23, [x10, #0x8]\n"
       "add x10, x10, #0x10\n"
-      "incb x23\n"
       "cbz x20, 8f\n"
       "mov x20, x20\n"
       "3:"  // K loop: Main loop
-      "whilelt p8.b, x16, %x[width]\n"
-      "mov x15, #0x0\n"
-      "mov x14, #0x0\n"
-      "cbz x17, 5f\n"
+      "whilelt p8.b, x14, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
+      "cbz x15, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25376143  // psel p3.b, p8.b/Z, p10.b[w15, #2]\n"
-      ".inst 0x253f6142  // psel p2.b, p8.b/Z, p10.b[w15, #3]\n"
-      ".inst 0x25776141  // psel p1.b, p8.b/Z, p10.b[w15, #6]\n"
-      ".inst 0x257f6140  // psel p0.b, p8.b/Z, p10.b[w15, #7]\n"
-      ".inst 0xe0176f62  // ld1b { za0h.b[x15, #2] }, p3/Z, [x27, x23]\n"
-      ".inst 0x25266d23  // psel p3.b, p11.b/Z, p9.b[w14]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0176b23  // ld1b { za0h.b[x15, #3] }, p2/Z, [x25, x23]\n"
-      ".inst 0x25266d22  // psel p2.b, p11.b/Z, p9.b[w14]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0176706  // ld1b { za0h.b[x15, #6] }, p1/Z, [x24, x23]\n"
-      ".inst 0x252e6d21  // psel p1.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162122  // ld1b { za0h.b[x13, #2] }, p0/Z, [x9, x22]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0162343  // ld1b { za0h.b[x13, #3] }, p0/Z, [x26, x22]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0x257d6142  // psel p2.b, p8.b/Z, p10.b[w13, #7]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0xe0162306  // ld1b { za0h.b[x13, #6] }, p0/Z, [x24, x22]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x24, [x11, #0x8]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0162ae7  // ld1b { za0h.b[x13, #7] }, p2/Z, [x23, x22]\n"
+      "ldr x23, [x10, #0x8]\n"
+      ".inst 0xe0bf86a0  // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01762a7  // ld1b { za0h.b[x15, #7] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0xe0bc86a1  // st1w { za0v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
       "add x10, x10, #0x10\n"
-      ".inst 0xe0bfcec0  // st1w { za0v.s[x14] }, p3/Z, [x22, XZR, LSL #2]\n"
-      "add x15, x15, #0x8\n"
-      ".inst 0xe0a8cac4  // st1w { za1v.s[x14] }, p2/Z, [x22, x8, LSL #2]\n"
-      ".inst 0xe0a9c6c1  // st1w { za0v.s[x14, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bcc2c5  // st1w { za1v.s[x14, #1] }, p0/Z, [x22, x28, LSL #2]\n"
-      "add x14, x14, #0x2\n"
-      "addvl x22, x22, #4\n"
-      "cmp x14, x17\n"
+      "add x13, x13, #0x8\n"
+      ".inst 0xe0bb82a5  // st1w { za1v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x15\n"
+      "addvl x21, x21, #4\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25376143  // psel p3.b, p8.b/Z, p10.b[w15, #2]\n"
-      ".inst 0x253f6142  // psel p2.b, p8.b/Z, p10.b[w15, #3]\n"
-      ".inst 0x25776141  // psel p1.b, p8.b/Z, p10.b[w15, #6]\n"
-      ".inst 0x257f6140  // psel p0.b, p8.b/Z, p10.b[w15, #7]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162122  // ld1b { za0h.b[x13, #2] }, p0/Z, [x9, x22]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      ".inst 0xe0162343  // ld1b { za0h.b[x13, #3] }, p0/Z, [x26, x22]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0176f62  // ld1b { za0h.b[x15, #2] }, p3/Z, [x27, x23]\n"
-      ".inst 0x25266d23  // psel p3.b, p11.b/Z, p9.b[w14]\n"
-      "ldr x27, [x11, #0x0]\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe0176b23  // ld1b { za0h.b[x15, #3] }, p2/Z, [x25, x23]\n"
-      ".inst 0x25266d22  // psel p2.b, p11.b/Z, p9.b[w14]\n"
-      "ldr x25, [x10, #0x0]\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0176706  // ld1b { za0h.b[x15, #6] }, p1/Z, [x24, x23]\n"
-      ".inst 0x252e6d21  // psel p1.b, p11.b/Z, p9.b[w14, #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0162306  // ld1b { za0h.b[x13, #6] }, p0/Z, [x24, x22]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      ".inst 0x257d6141  // psel p1.b, p8.b/Z, p10.b[w13, #7]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e7  // ld1b { za0h.b[x13, #7] }, p1/Z, [x23, x22]\n"
       "ldr x24, [x11, #0x8]\n"
+      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
+      "ldr x23, [x10, #0x8]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0b08aa4  // st1w { za1v.s[x12] }, p2/Z, [x21, x16, LSL #2]\n"
+      "whilelt p9.b, x14, %x[width]\n"
+      "incb x14\n"
+      ".inst 0xe0bc86a1  // st1w { za0v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01762a7  // ld1b { za0h.b[x15, #7] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      "whilelt p9.b, x16, %x[width]\n"
-      ".inst 0xe0bfcec0  // st1w { za0v.s[x14] }, p3/Z, [x22, XZR, LSL #2]\n"
-      "incb x16\n"
       "add x10, x10, #0x10\n"
-      ".inst 0xe0a8cac4  // st1w { za1v.s[x14] }, p2/Z, [x22, x8, LSL #2]\n"
-      "incb x23\n"
-      "whilelt p8.b, x16, %x[width]\n"
-      ".inst 0xe0a9c6c1  // st1w { za0v.s[x14, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bcc2c5  // st1w { za1v.s[x14, #1] }, p0/Z, [x22, x28, LSL #2]\n"
-      "addvl x22, x22, #4\n"
-      "cbz x17, 7f\n"
+      ".inst 0xe0bb82a5  // st1w { za1v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
+      "addvl x21, x21, #4\n"
+      "incb x22\n"
+      "whilelt p8.b, x14, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
+      "cbz x15, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x256d6140  // psel p0.b, p8.b/Z, p10.b[w13, #5]\n"
-      ".inst 0xe0172f60  // ld1b { za0h.b[x13] }, p3/Z, [x27, x23]\n"
-      ".inst 0x25246d23  // psel p3.b, p11.b/Z, p9.b[w12]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0172b21  // ld1b { za0h.b[x13, #1] }, p2/Z, [x25, x23]\n"
-      ".inst 0x25246d22  // psel p2.b, p11.b/Z, p9.b[w12]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0172704  // ld1b { za0h.b[x13, #4] }, p1/Z, [x24, x23]\n"
-      ".inst 0x252c6d21  // psel p1.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe0162120  // ld1b { za0h.b[x13] }, p0/Z, [x9, x22]\n"
+      ".inst 0x252d6140  // psel p0.b, p8.b/Z, p10.b[w13, #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0162341  // ld1b { za0h.b[x13, #1] }, p0/Z, [x26, x22]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
+      ".inst 0x256d6142  // psel p2.b, p8.b/Z, p10.b[w13, #5]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0xe0162304  // ld1b { za0h.b[x13, #4] }, p0/Z, [x24, x22]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x24, [x11, #0x8]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0162ae5  // ld1b { za0h.b[x13, #5] }, p2/Z, [x23, x22]\n"
+      "ldr x23, [x10, #0x8]\n"
+      ".inst 0xe0bf86a8  // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0b082ac  // st1w { za3v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe01722a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0xe0bc86a9  // st1w { za2v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
       "add x10, x10, #0x10\n"
-      ".inst 0xe0bf8ec8  // st1w { za2v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
       "add x13, x13, #0x8\n"
-      ".inst 0xe0a88acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bc82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x28, LSL #2]\n"
+      ".inst 0xe0bb82ad  // st1w { za3v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
       "add x12, x12, #0x2\n"
-      "addvl x22, x22, #4\n"
-      "cmp x12, x17\n"
+      "cmp x12, x15\n"
+      "addvl x21, x21, #4\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x256d6140  // psel p0.b, p8.b/Z, p10.b[w13, #5]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe0162120  // ld1b { za0h.b[x13] }, p0/Z, [x9, x22]\n"
+      ".inst 0x252d6140  // psel p0.b, p8.b/Z, p10.b[w13, #1]\n"
+      ".inst 0xe0162341  // ld1b { za0h.b[x13, #1] }, p0/Z, [x26, x22]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x8, LSL #3\n"
-      ".inst 0xe0172f60  // ld1b { za0h.b[x13] }, p3/Z, [x27, x23]\n"
-      ".inst 0x25246d23  // psel p3.b, p11.b/Z, p9.b[w12]\n"
-      "ldr x27, [x11, #0x0]\n"
-      ".inst 0xe0172b21  // ld1b { za0h.b[x13, #1] }, p2/Z, [x25, x23]\n"
-      ".inst 0x25246d22  // psel p2.b, p11.b/Z, p9.b[w12]\n"
-      "ldr x25, [x10, #0x0]\n"
-      ".inst 0xe0172704  // ld1b { za0h.b[x13, #4] }, p1/Z, [x24, x23]\n"
-      ".inst 0x252c6d21  // psel p1.b, p11.b/Z, p9.b[w12, #1]\n"
+      "ldr x9, [x11, #0x0]\n"
+      ".inst 0xe0162304  // ld1b { za0h.b[x13, #4] }, p0/Z, [x24, x22]\n"
+      "add x10, %x[in], x16, LSL #3\n"
+      ".inst 0x256d6141  // psel p1.b, p8.b/Z, p10.b[w13, #5]\n"
+      "ldr x26, [x10, #0x0]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01626e5  // ld1b { za0h.b[x13, #5] }, p1/Z, [x23, x22]\n"
       "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe01722a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x23]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      "whilelt p9.b, x16, %x[width]\n"
-      ".inst 0xe0bf8ec8  // st1w { za2v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
+      "ldr x23, [x10, #0x8]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xe0b08aac  // st1w { za3v.s[x12] }, p2/Z, [x21, x16, LSL #2]\n"
+      "whilelt p9.b, x14, %x[width]\n"
       "subs x20, x20, #0x1\n"
+      ".inst 0xe0bc86a9  // st1w { za2v.s[x12, #1] }, p1/Z, [x21, x28, LSL #2]\n"
+      "add x11, x11, #0x10\n"
       "add x10, x10, #0x10\n"
-      ".inst 0xe0a88acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
-      "incb x16\n"
-      "incb x23\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bc82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x28, LSL #2]\n"
-      "addvl x22, x22, #4\n"
+      ".inst 0xe0bb82ad  // st1w { za3v.s[x12, #1] }, p0/Z, [x21, x27, LSL #2]\n"
+      "addvl x21, x21, #4\n"
+      "incb x14\n"
+      "incb x22\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x26, 11f\n"
+      "cbnz x25, 11f\n"
       "mov x11, %x[in]\n"
-      "whilelt p8.b, x16, %x[width]\n"
+      "whilelt p8.b, x14, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25306d23  // psel p3.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      ".inst 0xe0bf8ec0  // st1w { za0v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a88ac4  // st1w { za1v.s[x12] }, p2/Z, [x22, x8, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
+      "ldr x20, [x11, #0x0]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe0162282  // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x22]\n"
+      "ldr x20, [x11, x16, LSL #0x3]\n"
       "add x12, x12, #0x1\n"
-      "addvl x22, x22, #2\n"
-      "ldr x21, [x11, #0x0]\n"
-      "cmp x12, x8\n"
-      "ldr x20, [x11, x8, LSL #0x3]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      ".inst 0xe0162283  // ld1b { za0h.b[x13, #3] }, p0/Z, [x20, x22]\n"
+      "cmp x12, x16\n"
       "add x11, x11, #0x8\n"
-      ".inst 0xe01726a2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x21, x23]\n"
-      ".inst 0xe0172283  // ld1b { za0h.b[x13, #3] }, p0/Z, [x20, x23]\n"
+      "addvl x21, x21, #2\n"
       "add x13, x13, #0x4\n"
       "blt 9b\n"
-      "whilelt p9.b, x16, %x[width]\n"
-      "whilelt p8.b, x16, %x[width]\n"
+      "whilelt p9.b, x14, %x[width]\n"
+      "whilelt p8.b, x14, %x[width]\n"
       "mov x20, #0x0\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: Second
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      "add x20, x20, #0x4\n"
-      ".inst 0xe0bf86c8  // st1w { za2v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a882cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x8, LSL #2]\n"
+      ".inst 0xe0bf82a8  // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0b082ac  // st1w { za3v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x22, x22, #2\n"
-      "cmp x12, x7\n"
+      "cmp x12, x17\n"
+      "addvl x21, x21, #2\n"
+      "add x20, x20, #0x4\n"
       "blt 10b\n"
-      "whilelt p8.b, x16, %x[width]\n"
+      "whilelt p8.b, x14, %x[width]\n"
       "b 13f\n"
       "11:"  // K loop: Tails: Odd
       "mov x12, #0x0\n"
       "12:"  // K loop: Tails: Odd: Loop
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xe0bf86c0  // st1w { za0v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0a882c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x8, LSL #2]\n"
+      ".inst 0xe0bf82a0  // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0b082a4  // st1w { za1v.s[x12] }, p0/Z, [x21, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x22, x22, #2\n"
-      "cmp x12, x7\n"
+      "cmp x12, x17\n"
+      "addvl x21, x21, #2\n"
       "blt 12b\n"
       "13:"  // K loop: End
-      "mov %x[out], x22\n"
+      "mov %x[out], x21\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_u8_u8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_u8_u8_summing.hpp
index 915381334e..c8657fad1c 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_u8_u8_summing.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_block4_u8_u8_summing.hpp
@@ -32,321 +32,321 @@ void interleave_block<2, 4, VLType::SME, true>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "cntb x21\n"
       "mov x23, %x[width]\n"
-      "mov x21, %x[width]\n"
-      "cntb x20\n"
-      "incb x23\n"
       "mov z20.b, #0x1\n"
+      "incb x23\n"
+      "mov x20, %x[width]\n"
       "mov z19.s, #0x0\n"
-      "sub x7, x20, #0x1\n"
-      "cntw x8\n"
       "mov z18.s, #0x0\n"
+      "sub x17, x21, #0x1\n"
+      "cntw x16\n"
       "sub x23, x23, #0x1\n"
-      "ands x7, x21, x7\n"
-      "udiv x23, x23, x20\n"  // n_passes = ceildiv(width, VL<T>)
-      "csel x7, x7, x20, NE\n"
+      "ands x17, x20, x17\n"
+      "udiv x23, x23, x21\n"  // n_passes = ceildiv(width, VL<T>)
+      "csel x17, x17, x21, NE\n"
       "lsl x22, %x[height], #0x1\n"  // height * 2
-      "lsl x21, x8, #0x1\n"
+      "lsl x21, x16, #0x1\n"
       "sub x20, x23, #0x1\n"
-      "add x7, x7, #0x3\n"
+      "add x17, x17, #0x3\n"
       "whilelt p9.b, XZR, x22\n"
       "whilelt p8.b, x21, x22\n"
-      "mov x17, #0x0\n"
-      "cntw x16, ALL, MUL #2\n"
+      "mov x15, #0x0\n"
+      "cntw x14, ALL, MUL #2\n"
       "cntw x11, ALL, MUL #3\n"
       "ptrue p4.b\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
       "and x10, x23, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "lsr x7, x7, #0x2\n"
-      "sub x9, x8, #0x2\n"
+      "lsr x17, x17, #0x2\n"
+      "sub x9, x16, #0x2\n"
       "ptrue p11.s\n"
       "zip1 p10.b, p9.b, p8.b\n"
       "mov x28, %x[row_offset]\n"
       "mov x27, %x[out]\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p9.b, x15, %x[width]\n"
+      "whilelt p8.b, x15, %x[width]\n"
       "cbnz %x[first], 1f\n"
       "addvl x27, x27, #-2\n"
       "ld1w { z19.s }, p4/Z, [x27]\n"
       "ld1w { z18.s }, p4/Z, [x27, #1, MUL VL]\n"
       "1:"  // K loop: Load row sums: End
       "mov x26, %x[in]\n"
-      "add x25, %x[in], x8, LSL #3\n"
+      "add x25, %x[in], x16, LSL #3\n"
       "ldr x24, [x26, #0x0]\n"
-      "mov x12, #0x0\n"
       "ldr x23, [x25, #0x0]\n"
+      "mov x12, #0x0\n"
       "ldr x22, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
       "ldr x21, [x25, #0x8]\n"
       "add x25, x25, #0x10\n"
       "cbz x9, 3f\n"
       "2:"  // K loop: Charge: Loop
-      ".inst 0x25246143  // psel p3.b, p8.b/Z, p10.b[w12]\n"
-      ".inst 0x252c6142  // psel p2.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe01c0300  // ld1b { za0h.b[x12] }, p0/Z, [x24, x28]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      "ldr x24, [x26, #0x0]\n"
+      ".inst 0xe01c02e1  // ld1b { za0h.b[x12, #1] }, p0/Z, [x23, x28]\n"
       ".inst 0x25646141  // psel p1.b, p8.b/Z, p10.b[w12, #4]\n"
       ".inst 0x256c6140  // psel p0.b, p8.b/Z, p10.b[w12, #5]\n"
-      ".inst 0xe01c0f00  // ld1b { za0h.b[x12] }, p3/Z, [x24, x28]\n"
-      "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe01c0ae1  // ld1b { za0h.b[x12, #1] }, p2/Z, [x23, x28]\n"
       "ldr x23, [x25, #0x0]\n"
       ".inst 0xe01c06c4  // ld1b { za0h.b[x12, #4] }, p1/Z, [x22, x28]\n"
       "ldr x22, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
       ".inst 0xe01c02a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x28]\n"
       "add x12, x12, #0x8\n"
+      "cmp x12, x9, LSL #2\n"
       "ldr x21, [x25, #0x8]\n"
       "add x25, x25, #0x10\n"
-      "cmp x12, x9, LSL #2\n"
       "blt 2b\n"
       "3:"  // K loop: Charge: End
-      ".inst 0x25246143  // psel p3.b, p8.b/Z, p10.b[w12]\n"
-      ".inst 0x252c6142  // psel p2.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe01c0300  // ld1b { za0h.b[x12] }, p0/Z, [x24, x28]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0xe01c02e1  // ld1b { za0h.b[x12, #1] }, p0/Z, [x23, x28]\n"
       ".inst 0x25646141  // psel p1.b, p8.b/Z, p10.b[w12, #4]\n"
       ".inst 0x256c6140  // psel p0.b, p8.b/Z, p10.b[w12, #5]\n"
+      ".inst 0xe01c06c4  // ld1b { za0h.b[x12, #4] }, p1/Z, [x22, x28]\n"
       "mov x26, %x[in]\n"
-      "add x25, %x[in], x8, LSL #3\n"
-      ".inst 0xe01c0f00  // ld1b { za0h.b[x12] }, p3/Z, [x24, x28]\n"
+      "add x25, %x[in], x16, LSL #3\n"
       "ldr x24, [x26, #0x0]\n"
-      "incb x17\n"
-      ".inst 0xe01c0ae1  // ld1b { za0h.b[x12, #1] }, p2/Z, [x23, x28]\n"
+      ".inst 0xe01c02a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x28]\n"
       "ldr x23, [x25, #0x0]\n"
-      ".inst 0xe01c06c4  // ld1b { za0h.b[x12, #4] }, p1/Z, [x22, x28]\n"
+      "incb x28\n"
+      "incb x15\n"
       "ldr x22, [x26, #0x8]\n"
       "add x26, x26, #0x10\n"
-      ".inst 0xe01c02a5  // ld1b { za0h.b[x12, #5] }, p0/Z, [x21, x28]\n"
       "ldr x21, [x25, #0x8]\n"
       "add x25, x25, #0x10\n"
-      "incb x28\n"
       "cbz x20, 9f\n"
       "mov x20, x20\n"
       "4:"  // K loop: Main loop
-      "whilelt p8.b, x17, %x[width]\n"
-      "mov x15, #0x0\n"
-      "mov x14, #0x0\n"
+      "whilelt p8.b, x15, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
       "cbz x9, 6f\n"
       "5:"  // K loop: Main loop: First: Loop
-      ".inst 0x25376143  // psel p3.b, p8.b/Z, p10.b[w15, #2]\n"
-      ".inst 0x253f6142  // psel p2.b, p8.b/Z, p10.b[w15, #3]\n"
-      ".inst 0x25776141  // psel p1.b, p8.b/Z, p10.b[w15, #6]\n"
-      ".inst 0x257f6140  // psel p0.b, p8.b/Z, p10.b[w15, #7]\n"
-      ".inst 0xe01c6f02  // ld1b { za0h.b[x15, #2] }, p3/Z, [x24, x28]\n"
-      ".inst 0x25266d23  // psel p3.b, p11.b/Z, p9.b[w14]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe01c2302  // ld1b { za0h.b[x13, #2] }, p0/Z, [x24, x28]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
       "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe01c6ae3  // ld1b { za0h.b[x15, #3] }, p2/Z, [x23, x28]\n"
-      ".inst 0x25266d22  // psel p2.b, p11.b/Z, p9.b[w14]\n"
+      ".inst 0xe01c22e3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x23, x28]\n"
+      ".inst 0x25756140  // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0x257d6142  // psel p2.b, p8.b/Z, p10.b[w13, #7]\n"
       "ldr x23, [x25, #0x0]\n"
-      ".inst 0xe01c66c6  // ld1b { za0h.b[x15, #6] }, p1/Z, [x22, x28]\n"
-      ".inst 0x252e6d21  // psel p1.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0xe01c22c6  // ld1b { za0h.b[x13, #6] }, p0/Z, [x22, x28]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0xe01c62a7  // ld1b { za0h.b[x15, #7] }, p0/Z, [x21, x28]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01c2aa7  // ld1b { za0h.b[x13, #7] }, p2/Z, [x21, x28]\n"
       "ldr x21, [x25, #0x8]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0xe0bfcf60  // st1w { za0v.s[x14] }, p3/Z, [x27, XZR, LSL #2]\n"
-      ".inst 0xc082d010  // mova z16.s, p4/M, za0v.s[x14]\n"
-      "add x15, x15, #0x8\n"
-      ".inst 0xc082d091  // mova z17.s, p4/M, za1v.s[x14]\n"
+      ".inst 0xe0bf8760  // st1w { za0v.s[x12] }, p1/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0xc0829010  // mova z16.s, p4/M, za0v.s[x12]\n"
+      ".inst 0xc0829091  // mova z17.s, p4/M, za1v.s[x12]\n"
       "udot z19.s, z16.b, z20.b\n"
-      ".inst 0xe0a8cb64  // st1w { za1v.s[x14] }, p2/Z, [x27, x8, LSL #2]\n"
+      ".inst 0xe0b08364  // st1w { za1v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "udot z18.s, z17.b, z20.b\n"
-      ".inst 0xe0b0c761  // st1w { za0v.s[x14, #1] }, p1/Z, [x27, x16, LSL #2]\n"
-      ".inst 0xc082d030  // mova z16.s, p4/M, za0v.s[x14, #1]\n"
-      ".inst 0xe0abc365  // st1w { za1v.s[x14, #1] }, p0/Z, [x27, x11, LSL #2]\n"
-      ".inst 0xc082d0b1  // mova z17.s, p4/M, za1v.s[x14, #1]\n"
-      "add x14, x14, #0x2\n"
-      "addvl x27, x27, #4\n"
-      "cmp x14, x9\n"
+      ".inst 0xe0ae8361  // st1w { za0v.s[x12, #1] }, p0/Z, [x27, x14, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0829030  // mova z16.s, p4/M, za0v.s[x12, #1]\n"
+      ".inst 0xe0ab8365  // st1w { za1v.s[x12, #1] }, p0/Z, [x27, x11, LSL #2]\n"
+      ".inst 0xc08290b1  // mova z17.s, p4/M, za1v.s[x12, #1]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x9\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "udot z19.s, z16.b, z20.b\n"
       "udot z18.s, z17.b, z20.b\n"
+      "addvl x27, x27, #4\n"
+      "add x13, x13, #0x8\n"
       "blt 5b\n"
       "6:"  // K loop: Main loop: First: Tail
-      ".inst 0x25376143  // psel p3.b, p8.b/Z, p10.b[w15, #2]\n"
-      ".inst 0x253f6142  // psel p2.b, p8.b/Z, p10.b[w15, #3]\n"
-      ".inst 0x25776141  // psel p1.b, p8.b/Z, p10.b[w15, #6]\n"
-      ".inst 0x257f6140  // psel p0.b, p8.b/Z, p10.b[w15, #7]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0xe01c2302  // ld1b { za0h.b[x13, #2] }, p0/Z, [x24, x28]\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      ".inst 0xe01c22e3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x23, x28]\n"
+      ".inst 0x25756141  // psel p1.b, p8.b/Z, p10.b[w13, #6]\n"
+      ".inst 0x257d6140  // psel p0.b, p8.b/Z, p10.b[w13, #7]\n"
+      ".inst 0xe01c26c6  // ld1b { za0h.b[x13, #6] }, p1/Z, [x22, x28]\n"
       "mov x26, %x[in]\n"
-      "add x25, %x[in], x8, LSL #3\n"
-      ".inst 0xe01c6f02  // ld1b { za0h.b[x15, #2] }, p3/Z, [x24, x28]\n"
-      ".inst 0x25266d23  // psel p3.b, p11.b/Z, p9.b[w14]\n"
+      "add x25, %x[in], x16, LSL #3\n"
       "ldr x24, [x26, #0x0]\n"
-      "mov x13, #0x0\n"
-      ".inst 0xe01c6ae3  // ld1b { za0h.b[x15, #3] }, p2/Z, [x23, x28]\n"
-      ".inst 0x25266d22  // psel p2.b, p11.b/Z, p9.b[w14]\n"
+      ".inst 0xe01c22a7  // ld1b { za0h.b[x13, #7] }, p0/Z, [x21, x28]\n"
+      ".inst 0xc0829010  // mova z16.s, p4/M, za0v.s[x12]\n"
+      ".inst 0x25306d23  // psel p3.s, p11.s/Z, p9.s[w12]\n"
+      "udot z19.s, z16.b, z20.b\n"
+      ".inst 0xc0829091  // mova z17.s, p4/M, za1v.s[x12]\n"
+      "udot z18.s, z17.b, z20.b\n"
       "ldr x23, [x25, #0x0]\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe01c66c6  // ld1b { za0h.b[x15, #6] }, p1/Z, [x22, x28]\n"
-      ".inst 0x252e6d21  // psel p1.b, p11.b/Z, p9.b[w14, #1]\n"
+      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0xe01c62a7  // ld1b { za0h.b[x15, #7] }, p0/Z, [x21, x28]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0829030  // mova z16.s, p4/M, za0v.s[x12, #1]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "ldr x21, [x25, #0x8]\n"
-      ".inst 0x252e6d20  // psel p0.b, p11.b/Z, p9.b[w14, #1]\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      ".inst 0xc082d010  // mova z16.s, p4/M, za0v.s[x14]\n"
-      ".inst 0xe0bfcf60  // st1w { za0v.s[x14] }, p3/Z, [x27, XZR, LSL #2]\n"
-      "incb x17\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0xc082d091  // mova z17.s, p4/M, za1v.s[x14]\n"
+      ".inst 0xe0bf8f60  // st1w { za0v.s[x12] }, p3/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0xc08290b1  // mova z17.s, p4/M, za1v.s[x12, #1]\n"
+      "whilelt p9.b, x15, %x[width]\n"
+      ".inst 0xe0b08b64  // st1w { za1v.s[x12] }, p2/Z, [x27, x16, LSL #2]\n"
+      "incb x15\n"
+      "add x26, x26, #0x10\n"
       "udot z19.s, z16.b, z20.b\n"
-      ".inst 0xe0a8cb64  // st1w { za1v.s[x14] }, p2/Z, [x27, x8, LSL #2]\n"
-      "incb x28\n"
+      ".inst 0xe0ae8761  // st1w { za0v.s[x12, #1] }, p1/Z, [x27, x14, LSL #2]\n"
+      "add x25, x25, #0x10\n"
       "udot z18.s, z17.b, z20.b\n"
-      ".inst 0xe0b0c761  // st1w { za0v.s[x14, #1] }, p1/Z, [x27, x16, LSL #2]\n"
-      ".inst 0xc082d030  // mova z16.s, p4/M, za0v.s[x14, #1]\n"
-      "whilelt p8.b, x17, %x[width]\n"
-      ".inst 0xc082d0b1  // mova z17.s, p4/M, za1v.s[x14, #1]\n"
-      ".inst 0xe0abc365  // st1w { za1v.s[x14, #1] }, p0/Z, [x27, x11, LSL #2]\n"
+      "incb x28\n"
+      ".inst 0xe0ab8365  // st1w { za1v.s[x12, #1] }, p0/Z, [x27, x11, LSL #2]\n"
       "addvl x27, x27, #4\n"
-      "udot z19.s, z16.b, z20.b\n"
-      "udot z18.s, z17.b, z20.b\n"
+      "whilelt p8.b, x15, %x[width]\n"
+      "mov x13, #0x0\n"
+      "mov x12, #0x0\n"
       "cbz x9, 8f\n"
       "7:"  // K loop: Main loop: Second: Loop
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
-      ".inst 0x256d6140  // psel p0.b, p8.b/Z, p10.b[w13, #5]\n"
-      ".inst 0xe01c2f00  // ld1b { za0h.b[x13] }, p3/Z, [x24, x28]\n"
-      ".inst 0x25246d23  // psel p3.b, p11.b/Z, p9.b[w12]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe01c2300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x28]\n"
+      ".inst 0x252d6140  // psel p0.b, p8.b/Z, p10.b[w13, #1]\n"
       "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe01c2ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x28]\n"
-      ".inst 0x25246d22  // psel p2.b, p11.b/Z, p9.b[w12]\n"
+      ".inst 0xe01c22e1  // ld1b { za0h.b[x13, #1] }, p0/Z, [x23, x28]\n"
+      ".inst 0x25656140  // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
+      ".inst 0x256d6142  // psel p2.b, p8.b/Z, p10.b[w13, #5]\n"
       "ldr x23, [x25, #0x0]\n"
-      ".inst 0xe01c26c4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x22, x28]\n"
-      ".inst 0x252c6d21  // psel p1.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0xe01c22c4  // ld1b { za0h.b[x13, #4] }, p0/Z, [x22, x28]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0xe01c22a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x28]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe01c2aa5  // ld1b { za0h.b[x13, #5] }, p2/Z, [x21, x28]\n"
       "ldr x21, [x25, #0x8]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0xe0bf8f68  // st1w { za2v.s[x12] }, p3/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0xe0bf8768  // st1w { za2v.s[x12] }, p1/Z, [x27, XZR, LSL #2]\n"
       ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
-      "add x13, x13, #0x8\n"
       ".inst 0xc0829191  // mova z17.s, p4/M, za3v.s[x12]\n"
       "udot z19.s, z16.b, z20.b\n"
-      ".inst 0xe0a88b6c  // st1w { za3v.s[x12] }, p2/Z, [x27, x8, LSL #2]\n"
+      ".inst 0xe0b0836c  // st1w { za3v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "udot z18.s, z17.b, z20.b\n"
-      ".inst 0xe0b08769  // st1w { za2v.s[x12, #1] }, p1/Z, [x27, x16, LSL #2]\n"
+      ".inst 0xe0ae8369  // st1w { za2v.s[x12, #1] }, p0/Z, [x27, x14, LSL #2]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       ".inst 0xc0829130  // mova z16.s, p4/M, za2v.s[x12, #1]\n"
       ".inst 0xe0ab836d  // st1w { za3v.s[x12, #1] }, p0/Z, [x27, x11, LSL #2]\n"
       ".inst 0xc08291b1  // mova z17.s, p4/M, za3v.s[x12, #1]\n"
       "add x12, x12, #0x2\n"
-      "addvl x27, x27, #4\n"
       "cmp x12, x9\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "udot z19.s, z16.b, z20.b\n"
       "udot z18.s, z17.b, z20.b\n"
+      "addvl x27, x27, #4\n"
+      "add x13, x13, #0x8\n"
       "blt 7b\n"
       "8:"  // K loop: Main loop: Second: Tail
-      ".inst 0x25256143  // psel p3.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
+      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
+      ".inst 0xe01c2300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x28]\n"
+      ".inst 0x252d6140  // psel p0.b, p8.b/Z, p10.b[w13, #1]\n"
+      ".inst 0xe01c22e1  // ld1b { za0h.b[x13, #1] }, p0/Z, [x23, x28]\n"
       ".inst 0x25656141  // psel p1.b, p8.b/Z, p10.b[w13, #4]\n"
       ".inst 0x256d6140  // psel p0.b, p8.b/Z, p10.b[w13, #5]\n"
+      ".inst 0xe01c26c4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x22, x28]\n"
       "mov x26, %x[in]\n"
-      "add x25, %x[in], x8, LSL #3\n"
-      ".inst 0xe01c2f00  // ld1b { za0h.b[x13] }, p3/Z, [x24, x28]\n"
-      ".inst 0x25246d23  // psel p3.b, p11.b/Z, p9.b[w12]\n"
+      "add x25, %x[in], x16, LSL #3\n"
       "ldr x24, [x26, #0x0]\n"
-      ".inst 0xe01c2ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x28]\n"
-      ".inst 0x25246d22  // psel p2.b, p11.b/Z, p9.b[w12]\n"
+      ".inst 0xe01c22a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x28]\n"
+      ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
+      ".inst 0x25306d23  // psel p3.s, p11.s/Z, p9.s[w12]\n"
+      "udot z19.s, z16.b, z20.b\n"
+      ".inst 0xc0829191  // mova z17.s, p4/M, za3v.s[x12]\n"
+      "udot z18.s, z17.b, z20.b\n"
       "ldr x23, [x25, #0x0]\n"
-      ".inst 0xe01c26c4  // ld1b { za0h.b[x13, #4] }, p1/Z, [x22, x28]\n"
-      ".inst 0x252c6d21  // psel p1.b, p11.b/Z, p9.b[w12, #1]\n"
+      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
       "ldr x22, [x26, #0x8]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0xe01c22a5  // ld1b { za0h.b[x13, #5] }, p0/Z, [x21, x28]\n"
+      ".inst 0x25706d21  // psel p1.s, p11.s/Z, p9.s[w12, #1]\n"
+      ".inst 0xc0829130  // mova z16.s, p4/M, za2v.s[x12, #1]\n"
+      ".inst 0x25706d20  // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
       "ldr x21, [x25, #0x8]\n"
-      ".inst 0x252c6d20  // psel p0.b, p11.b/Z, p9.b[w12, #1]\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
       ".inst 0xe0bf8f68  // st1w { za2v.s[x12] }, p3/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0xc08291b1  // mova z17.s, p4/M, za3v.s[x12, #1]\n"
+      "whilelt p9.b, x15, %x[width]\n"
+      ".inst 0xe0b08b6c  // st1w { za3v.s[x12] }, p2/Z, [x27, x16, LSL #2]\n"
       "subs x20, x20, #0x1\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0xc0829191  // mova z17.s, p4/M, za3v.s[x12]\n"
+      "add x26, x26, #0x10\n"
       "udot z19.s, z16.b, z20.b\n"
-      ".inst 0xe0a88b6c  // st1w { za3v.s[x12] }, p2/Z, [x27, x8, LSL #2]\n"
-      "incb x17\n"
+      ".inst 0xe0ae8769  // st1w { za2v.s[x12, #1] }, p1/Z, [x27, x14, LSL #2]\n"
+      "add x25, x25, #0x10\n"
       "udot z18.s, z17.b, z20.b\n"
-      ".inst 0xe0b08769  // st1w { za2v.s[x12, #1] }, p1/Z, [x27, x16, LSL #2]\n"
-      ".inst 0xc0829130  // mova z16.s, p4/M, za2v.s[x12, #1]\n"
-      "incb x28\n"
-      ".inst 0xc08291b1  // mova z17.s, p4/M, za3v.s[x12, #1]\n"
+      "incb x15\n"
       ".inst 0xe0ab836d  // st1w { za3v.s[x12, #1] }, p0/Z, [x27, x11, LSL #2]\n"
       "addvl x27, x27, #4\n"
-      "udot z19.s, z16.b, z20.b\n"
-      "udot z18.s, z17.b, z20.b\n"
+      "incb x28\n"
       "bgt 4b\n"
       "9:"  // K loop: Tails
       "cbnz x10, 12f\n"
       "mov x26, %x[in]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p8.b, x15, %x[width]\n"
       "mov x13, #0x0\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: First
-      ".inst 0x25306d23  // psel p3.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf8360  // st1w { za0v.s[x12] }, p0/Z, [x27, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xc0829010  // mova z16.s, p4/M, za0v.s[x12]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      ".inst 0xe0b08364  // st1w { za1v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
+      "ldr x21, [x26, #0x0]\n"
+      ".inst 0x25356140  // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
       ".inst 0xc0829091  // mova z17.s, p4/M, za1v.s[x12]\n"
+      "ldr x20, [x26, x16, LSL #0x3]\n"
+      ".inst 0xe01c22a2  // ld1b { za0h.b[x13, #2] }, p0/Z, [x21, x28]\n"
+      "add x12, x12, #0x1\n"
+      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
+      "cmp x12, x16\n"
       "udot z19.s, z16.b, z20.b\n"
       "udot z18.s, z17.b, z20.b\n"
-      ".inst 0xe0bf8f60  // st1w { za0v.s[x12] }, p3/Z, [x27, XZR, LSL #2]\n"
-      ".inst 0xe0a88b64  // st1w { za1v.s[x12] }, p2/Z, [x27, x8, LSL #2]\n"
-      "add x12, x12, #0x1\n"
-      "addvl x27, x27, #2\n"
-      "ldr x21, [x26, #0x0]\n"
-      "cmp x12, x8\n"
-      "ldr x20, [x26, x8, LSL #0x3]\n"
-      "add x26, x26, #0x8\n"
-      ".inst 0xe01c26a2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x21, x28]\n"
       ".inst 0xe01c2283  // ld1b { za0h.b[x13, #3] }, p0/Z, [x20, x28]\n"
+      "add x26, x26, #0x8\n"
+      "addvl x27, x27, #2\n"
       "add x13, x13, #0x4\n"
       "blt 10b\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p9.b, x15, %x[width]\n"
+      "whilelt p8.b, x15, %x[width]\n"
       "mov x20, #0x0\n"
       "mov x12, #0x0\n"
       "11:"  // K loop: Tails: Even: Second
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf8368  // st1w { za2v.s[x12] }, p0/Z, [x27, XZR, LSL #2]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xc0829110  // mova z16.s, p4/M, za2v.s[x12]\n"
+      ".inst 0xe0b0836c  // st1w { za3v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
       ".inst 0xc0829191  // mova z17.s, p4/M, za3v.s[x12]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x17\n"
       "udot z19.s, z16.b, z20.b\n"
-      "add x20, x20, #0x4\n"
       "udot z18.s, z17.b, z20.b\n"
-      ".inst 0xe0bf8768  // st1w { za2v.s[x12] }, p1/Z, [x27, XZR, LSL #2]\n"
-      ".inst 0xe0a8836c  // st1w { za3v.s[x12] }, p0/Z, [x27, x8, LSL #2]\n"
-      "add x12, x12, #0x1\n"
       "addvl x27, x27, #2\n"
-      "cmp x12, x7\n"
+      "add x20, x20, #0x4\n"
       "blt 11b\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p8.b, x15, %x[width]\n"
       "b 14f\n"
       "12:"  // K loop: Tails: Odd
       "mov x12, #0x0\n"
       "13:"  // K loop: Tails: Odd: Loop
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0bf8360  // st1w { za0v.s[x12] }, p0/Z, [x27, XZR, LSL #2]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xc0829010  // mova z16.s, p4/M, za0v.s[x12]\n"
+      ".inst 0xe0b08364  // st1w { za1v.s[x12] }, p0/Z, [x27, x16, LSL #2]\n"
       ".inst 0xc0829091  // mova z17.s, p4/M, za1v.s[x12]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x17\n"
       "udot z19.s, z16.b, z20.b\n"
       "udot z18.s, z17.b, z20.b\n"
-      ".inst 0xe0bf8760  // st1w { za0v.s[x12] }, p1/Z, [x27, XZR, LSL #2]\n"
-      ".inst 0xe0a88364  // st1w { za1v.s[x12] }, p0/Z, [x27, x8, LSL #2]\n"
-      "add x12, x12, #0x1\n"
       "addvl x27, x27, #2\n"
-      "cmp x12, x7\n"
       "blt 13b\n"
       "14:"  // K loop: End
       "st1w { z19.s }, p4, [x27]\n"
       "st1w { z18.s }, p4, [x27, #1, MUL VL]\n"
       "addvl x27, x27, #2\n"
-      ".inst 0xd503467f  // SMSTOP\n"
       "mov %x[out], x27\n"
+      ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [first] "r" (first), [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_fp16_fp16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_fp16_fp16.hpp
index 19d87039fe..9403efc7c6 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_fp16_fp16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_fp16_fp16.hpp
@@ -32,66 +32,66 @@ void interleave_block<2, 1, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x28, #0x0\n"
-      "mov x27, %x[row_offset]\n"
-      "cnth x26\n"
-      "cnth x25\n"
-      "cmp %x[height], x26\n"
+      "cnth x28\n"
+      "cmp %x[height], x28\n"
+      "cnth x27\n"
+      "csel x28, %x[height], x28, LT\n"
+      "mov x26, #0x0\n"
       "ptrue p13.s\n"
-      "csel x26, %x[height], x26, LT\n"
+      "sub x28, x28, #0x1\n"
       "whilelt p12.h, XZR, %x[height]\n"
-      "sub x26, x26, #0x1\n"
-      "whilelt p11.h, x25, %x[height]\n"
+      "whilelt p11.h, x27, %x[height]\n"
+      "mov x25, %x[row_offset]\n"
       "mov x24, %x[out]\n"
-      "whilelt p10.h, x28, %x[width]\n"
-      "whilelt p9.h, x28, %x[width]\n"
-      "whilelt p8.h, x28, %x[width]\n"
+      "whilelt p10.h, x26, %x[width]\n"
+      "whilelt p9.h, x26, %x[width]\n"
+      "whilelt p8.h, x26, %x[width]\n"
       "1:"  // Width loop
       "add x23, %x[in], XZR, LSL #3\n"
-      "add x20, %x[in], x25, LSL #3\n"
-      "mov x13, #0x0\n"
+      "add x20, %x[in], x27, LSL #3\n"
       "ldr x22, [x23], #0x8\n"
+      "mov x12, #0x0\n"
       "ldr x21, [x20], #0x8\n"
-      "cbz x26, 3f\n"
+      "cbz x28, 3f\n"
       "2:"  // Loads: Loop
-      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
-      ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
-      ".inst 0xe05b26c0  // ld1h { za0h.h[x13] }, p1/Z, [x22, x27, LSL #1]\n"
+      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      ".inst 0xe05906c0  // ld1h { za0h.h[x12] }, p1/Z, [x22, x25, LSL #1]\n"
       "ldr x22, [x23], #0x8\n"
-      ".inst 0xe05b22a8  // ld1h { za1h.h[x13] }, p0/Z, [x21, x27, LSL #1]\n"
-      "add x13, x13, #0x2\n"
+      ".inst 0xe05902a8  // ld1h { za1h.h[x12] }, p0/Z, [x21, x25, LSL #1]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x28, LSL #1\n"
       "ldr x21, [x20], #0x8\n"
-      "cmp x13, x26, LSL #1\n"
       "blt 2b\n"
       "3:"  // Loads: Tail
-      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
-      ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
-      "sub x20, %x[width], x28\n"
+      "sub x20, %x[width], x26\n"
+      ".inst 0x25286580  // psel p0.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0xe05902c0  // ld1h { za0h.h[x12] }, p0/Z, [x22, x25, LSL #1]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      "cmp x20, x27\n"
+      ".inst 0xe05902a8  // ld1h { za1h.h[x12] }, p0/Z, [x21, x25, LSL #1]\n"
       "mov x12, #0x0\n"
-      "cmp x20, x25\n"
-      ".inst 0xe05b26c0  // ld1h { za0h.h[x13] }, p1/Z, [x22, x27, LSL #1]\n"
-      "csel x20, x20, x25, LT\n"
-      ".inst 0xe05b22a8  // ld1h { za1h.h[x13] }, p0/Z, [x21, x27, LSL #1]\n"
+      "csel x20, x20, x27, LT\n"
       "4:"  // Stores: Loop
-      ".inst 0x25287541  // psel p1.h, p13.h/Z, p10.h[w12]\n"
       ".inst 0x25287540  // psel p0.h, p13.h/Z, p10.h[w12]\n"
-      ".inst 0xe07f8700  // st1h { za0v.h[x12] }, p1/Z, [x24, XZR, LSL #1]\n"
-      ".inst 0xe0798308  // st1h { za1v.h[x12] }, p0/Z, [x24, x25, LSL #1]\n"
+      ".inst 0xe07f8300  // st1h { za0v.h[x12] }, p0/Z, [x24, XZR, LSL #1]\n"
+      ".inst 0x25287540  // psel p0.h, p13.h/Z, p10.h[w12]\n"
+      ".inst 0xe07b8308  // st1h { za1v.h[x12] }, p0/Z, [x24, x27, LSL #1]\n"
       "add x12, x12, #0x1\n"
-      "addvl x24, x24, #4\n"
       "cmp x12, x20\n"
+      "addvl x24, x24, #4\n"
       "blt 4b\n"
-      "inch x28\n"
-      "inch x27\n"
-      "whilelt p10.h, x28, %x[width]\n"
-      "whilelt p9.h, x28, %x[width]\n"
-      "whilelt p8.h, x28, %x[width]\n"
+      "inch x26\n"
+      "whilelt p10.h, x26, %x[width]\n"
+      "whilelt p9.h, x26, %x[width]\n"
+      "whilelt p8.h, x26, %x[width]\n"
+      "inch x25\n"
       "b.any 1b\n"
       "mov %x[out], x24\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_fp32_fp32.hpp
index 68fabe3523..b310651525 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_fp32_fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave2VL_fp32_fp32.hpp
@@ -32,251 +32,256 @@ void interleave_block<2, 1, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
+      "mov x22, %x[width]\n"
+      "incw x22\n"
+      "cntw x16\n"
+      "sub x22, x22, #0x1\n"
+      "udiv x22, x22, x16\n"  // n_passes = ceildiv(width, VL<T>)
       "mov x21, %x[width]\n"
-      "mov x20, %x[width]\n"
-      "incw x21\n"
-      "cntw x17\n"
-      "sub x21, x21, #0x1\n"
-      "sub x16, x17, #0x1\n"
-      "udiv x21, x21, x17\n"  // n_passes = ceildiv(width, VL<T>)
-      "ands x16, x20, x16\n"
-      "sub x20, x21, #0x1\n"
-      "sub x15, x17, #0x2\n"
-      "mov x14, #0x0\n"
+      "sub x15, x16, #0x1\n"
+      "sub x20, x22, #0x1\n"
+      "ands x15, x21, x15\n"
+      "sub x14, x16, #0x2\n"
+      "mov x13, #0x0\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x17, LSL #3\n"
-      "cntw x9, ALL, MUL #2\n"
-      "ldr x28, [x11, #0x0]\n"
-      "cntw x27, ALL, MUL #3\n"
+      "ldr x10, [x11, #0x0]\n"
+      "add x9, %x[in], x16, LSL #3\n"
+      "cntw x28, ALL, MUL #2\n"
+      "ldr x27, [x9, #0x0]\n"
+      "cntw x26, ALL, MUL #3\n"
       "lsr x20, x20, #0x1\n"  // n_loops = (n_passes - 1) / 2
-      "ldr x26, [x10, #0x0]\n"
-      "and x25, x21, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
-      "csel x16, x16, x17, NE\n"
-      "ldr x24, [x11, #0x8]\n"
-      "ptrue p12.s\n"
-      "whilelt p11.s, XZR, %x[height]\n"
-      "ldr x21, [x10, #0x8]\n"
-      "whilelt p10.s, x17, %x[height]\n"
+      "ldr x25, [x11, #0x8]\n"
+      "and x24, x22, #0x1\n"  // odd_tail = bool(n_passes & 0x1)
+      "csel x15, x15, x16, NE\n"
+      "ldr x21, [x9, #0x8]\n"
+      "ptrue p13.s\n"
+      "whilelt p12.s, XZR, %x[height]\n"
+      "whilelt p11.s, x16, %x[height]\n"
       "mov x23, %x[row_offset]\n"
       "mov x22, %x[out]\n"
-      "whilelt p9.s, x14, %x[width]\n"
-      "whilelt p8.s, x14, %x[width]\n"
+      "whilelt p10.s, x13, %x[width]\n"
+      "whilelt p9.s, x13, %x[width]\n"
+      "whilelt p8.s, x13, %x[width]\n"
       "add x11, x11, #0x10\n"
-      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       "mov x12, #0x0\n"
-      "cbz x15, 2f\n"
+      "cbz x14, 2f\n"
       "1:"  // K loop: Charge: Loop
-      ".inst 0x25306163  // psel p3.s, p8.s/Z, p11.s[w12]\n"
-      ".inst 0x25306142  // psel p2.s, p8.s/Z, p10.s[w12]\n"
-      ".inst 0x25706161  // psel p1.s, p8.s/Z, p11.s[w12, #1]\n"
-      ".inst 0x25706140  // psel p0.s, p8.s/Z, p10.s[w12, #1]\n"
-      ".inst 0xe0970f80  // ld1w { za0h.s[x12] }, p3/Z, [x28, x23, LSL #2]\n"
-      "ldr x28, [x11, #0x0]\n"
-      ".inst 0xe0970b44  // ld1w { za1h.s[x12] }, p2/Z, [x26, x23, LSL #2]\n"
-      "ldr x26, [x10, #0x0]\n"
-      ".inst 0xe0970701  // ld1w { za0h.s[x12, #1] }, p1/Z, [x24, x23, LSL #2]\n"
-      "ldr x24, [x11, #0x8]\n"
+      ".inst 0x25306581  // psel p1.s, p9.s/Z, p12.s[w12]\n"
+      ".inst 0x25306160  // psel p0.s, p8.s/Z, p11.s[w12]\n"
+      ".inst 0xe0970540  // ld1w { za0h.s[x12] }, p1/Z, [x10, x23, LSL #2]\n"
+      "ldr x10, [x11, #0x0]\n"
+      ".inst 0xe0970364  // ld1w { za1h.s[x12] }, p0/Z, [x27, x23, LSL #2]\n"
+      ".inst 0x25706581  // psel p1.s, p9.s/Z, p12.s[w12, #1]\n"
+      ".inst 0x25706160  // psel p0.s, p8.s/Z, p11.s[w12, #1]\n"
+      "ldr x27, [x9, #0x0]\n"
+      ".inst 0xe0970721  // ld1w { za0h.s[x12, #1] }, p1/Z, [x25, x23, LSL #2]\n"
+      "ldr x25, [x11, #0x8]\n"
       "add x11, x11, #0x10\n"
       ".inst 0xe09702a5  // ld1w { za1h.s[x12, #1] }, p0/Z, [x21, x23, LSL #2]\n"
       "add x12, x12, #0x2\n"
-      "ldr x21, [x10, #0x8]\n"
-      "add x10, x10, #0x10\n"
-      "cmp x12, x15\n"
+      "cmp x12, x14\n"
+      "ldr x21, [x9, #0x8]\n"
+      "add x9, x9, #0x10\n"
       "blt 1b\n"
       "2:"  // K loop: Charge: End
-      ".inst 0x25306163  // psel p3.s, p8.s/Z, p11.s[w12]\n"
-      ".inst 0x25306142  // psel p2.s, p8.s/Z, p10.s[w12]\n"
-      ".inst 0x25706161  // psel p1.s, p8.s/Z, p11.s[w12, #1]\n"
-      ".inst 0x25706140  // psel p0.s, p8.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25306581  // psel p1.s, p9.s/Z, p12.s[w12]\n"
+      ".inst 0x25306160  // psel p0.s, p8.s/Z, p11.s[w12]\n"
+      ".inst 0xe0970540  // ld1w { za0h.s[x12] }, p1/Z, [x10, x23, LSL #2]\n"
+      ".inst 0xe0970364  // ld1w { za1h.s[x12] }, p0/Z, [x27, x23, LSL #2]\n"
+      ".inst 0x25706581  // psel p1.s, p9.s/Z, p12.s[w12, #1]\n"
+      ".inst 0x25706160  // psel p0.s, p8.s/Z, p11.s[w12, #1]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x17, LSL #3\n"
-      ".inst 0xe0970f80  // ld1w { za0h.s[x12] }, p3/Z, [x28, x23, LSL #2]\n"
-      "ldr x28, [x11, #0x0]\n"
-      "incw x14\n"
-      ".inst 0xe0970b44  // ld1w { za1h.s[x12] }, p2/Z, [x26, x23, LSL #2]\n"
-      "ldr x26, [x10, #0x0]\n"
-      ".inst 0xe0970701  // ld1w { za0h.s[x12, #1] }, p1/Z, [x24, x23, LSL #2]\n"
-      "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
+      "add x9, %x[in], x16, LSL #3\n"
+      ".inst 0xe0970721  // ld1w { za0h.s[x12, #1] }, p1/Z, [x25, x23, LSL #2]\n"
+      "ldr x10, [x11, #0x0]\n"
       ".inst 0xe09702a5  // ld1w { za1h.s[x12, #1] }, p0/Z, [x21, x23, LSL #2]\n"
-      "ldr x21, [x10, #0x8]\n"
-      "add x10, x10, #0x10\n"
+      "ldr x27, [x9, #0x0]\n"
       "incw x23\n"
+      "incw x13\n"
+      "ldr x25, [x11, #0x8]\n"
+      "add x11, x11, #0x10\n"
+      "ldr x21, [x9, #0x8]\n"
+      "add x9, x9, #0x10\n"
       "cbz x20, 8f\n"
       "mov x20, x20\n"
       "3:"  // K loop: Main loop
-      "whilelt p8.s, x14, %x[width]\n"
-      "mov x13, #0x0\n"
-      "cbz x15, 5f\n"
+      "whilelt p9.s, x13, %x[width]\n"
+      "whilelt p8.s, x13, %x[width]\n"
+      "mov x12, #0x0\n"
+      "cbz x14, 5f\n"
       "4:"  // K loop: Main loop: First: Loop
-      ".inst 0x25316160  // psel p0.s, p8.s/Z, p11.s[w13]\n"
-      ".inst 0x25316142  // psel p2.s, p8.s/Z, p10.s[w13]\n"
-      ".inst 0x25716161  // psel p1.s, p8.s/Z, p11.s[w13, #1]\n"
-      ".inst 0x25716143  // psel p3.s, p8.s/Z, p10.s[w13, #1]\n"
-      ".inst 0xe0972388  // ld1w { za2h.s[x13] }, p0/Z, [x28, x23, LSL #2]\n"
-      ".inst 0x25317120  // psel p0.s, p12.s/Z, p9.s[w13]\n"
-      "ldr x28, [x11, #0x0]\n"
-      ".inst 0xe0972b4c  // ld1w { za3h.s[x13] }, p2/Z, [x26, x23, LSL #2]\n"
-      ".inst 0x25317122  // psel p2.s, p12.s/Z, p9.s[w13]\n"
-      "ldr x26, [x10, #0x0]\n"
-      ".inst 0xe0972709  // ld1w { za2h.s[x13, #1] }, p1/Z, [x24, x23, LSL #2]\n"
-      ".inst 0x25717121  // psel p1.s, p12.s/Z, p9.s[w13, #1]\n"
-      "ldr x24, [x11, #0x8]\n"
+      ".inst 0x25306581  // psel p1.s, p9.s/Z, p12.s[w12]\n"
+      ".inst 0x25306160  // psel p0.s, p8.s/Z, p11.s[w12]\n"
+      ".inst 0xe0970548  // ld1w { za2h.s[x12] }, p1/Z, [x10, x23, LSL #2]\n"
+      "ldr x10, [x11, #0x0]\n"
+      ".inst 0xe097036c  // ld1w { za3h.s[x12] }, p0/Z, [x27, x23, LSL #2]\n"
+      ".inst 0x25706580  // psel p0.s, p9.s/Z, p12.s[w12, #1]\n"
+      ".inst 0x25706162  // psel p2.s, p8.s/Z, p11.s[w12, #1]\n"
+      "ldr x27, [x9, #0x0]\n"
+      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0970329  // ld1w { za2h.s[x12, #1] }, p0/Z, [x25, x23, LSL #2]\n"
+      "ldr x25, [x11, #0x8]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0970aad  // ld1w { za3h.s[x12, #1] }, p2/Z, [x21, x23, LSL #2]\n"
+      "ldr x21, [x9, #0x8]\n"
+      ".inst 0xe0bf86c0  // st1w { za0v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe0972ead  // ld1w { za3h.s[x13, #1] }, p3/Z, [x21, x23, LSL #2]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bfa2c0  // st1w { za0v.s[x13] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x25717120  // psel p0.s, p12.s/Z, p9.s[w13, #1]\n"
-      ".inst 0xe0b1aac4  // st1w { za1v.s[x13] }, p2/Z, [x22, x17, LSL #2]\n"
-      "add x10, x10, #0x10\n"
-      ".inst 0xe0a9a6c1  // st1w { za0v.s[x13, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bba2c5  // st1w { za1v.s[x13, #1] }, p0/Z, [x22, x27, LSL #2]\n"
-      "add x13, x13, #0x2\n"
+      ".inst 0xe0bc86c1  // st1w { za0v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
+      "add x9, x9, #0x10\n"
+      ".inst 0xe0ba82c5  // st1w { za1v.s[x12, #1] }, p0/Z, [x22, x26, LSL #2]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x14\n"
       "addvl x22, x22, #4\n"
-      "cmp x13, x15\n"
       "blt 4b\n"
       "5:"  // K loop: Main loop: First: Tail
-      ".inst 0x25316160  // psel p0.s, p8.s/Z, p11.s[w13]\n"
-      ".inst 0x25316142  // psel p2.s, p8.s/Z, p10.s[w13]\n"
-      ".inst 0x25716161  // psel p1.s, p8.s/Z, p11.s[w13, #1]\n"
-      ".inst 0x25716143  // psel p3.s, p8.s/Z, p10.s[w13, #1]\n"
+      ".inst 0x25306581  // psel p1.s, p9.s/Z, p12.s[w12]\n"
+      ".inst 0x25306160  // psel p0.s, p8.s/Z, p11.s[w12]\n"
+      ".inst 0xe0970548  // ld1w { za2h.s[x12] }, p1/Z, [x10, x23, LSL #2]\n"
+      ".inst 0xe097036c  // ld1w { za3h.s[x12] }, p0/Z, [x27, x23, LSL #2]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x17, LSL #3\n"
-      ".inst 0xe0972388  // ld1w { za2h.s[x13] }, p0/Z, [x28, x23, LSL #2]\n"
-      ".inst 0x25317120  // psel p0.s, p12.s/Z, p9.s[w13]\n"
-      "ldr x28, [x11, #0x0]\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0972b4c  // ld1w { za3h.s[x13] }, p2/Z, [x26, x23, LSL #2]\n"
-      ".inst 0x25317122  // psel p2.s, p12.s/Z, p9.s[w13]\n"
-      "ldr x26, [x10, #0x0]\n"
-      ".inst 0xe0972709  // ld1w { za2h.s[x13, #1] }, p1/Z, [x24, x23, LSL #2]\n"
-      ".inst 0x25717121  // psel p1.s, p12.s/Z, p9.s[w13, #1]\n"
-      "ldr x24, [x11, #0x8]\n"
+      "add x9, %x[in], x16, LSL #3\n"
+      "ldr x10, [x11, #0x0]\n"
+      ".inst 0x25706580  // psel p0.s, p9.s/Z, p12.s[w12, #1]\n"
+      ".inst 0x25706161  // psel p1.s, p8.s/Z, p11.s[w12, #1]\n"
+      ".inst 0xe0970329  // ld1w { za2h.s[x12, #1] }, p0/Z, [x25, x23, LSL #2]\n"
+      "ldr x27, [x9, #0x0]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe09706ad  // ld1w { za3h.s[x12, #1] }, p1/Z, [x21, x23, LSL #2]\n"
+      "ldr x25, [x11, #0x8]\n"
+      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
+      "ldr x21, [x9, #0x8]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b08ac4  // st1w { za1v.s[x12] }, p2/Z, [x22, x16, LSL #2]\n"
+      "whilelt p10.s, x13, %x[width]\n"
+      "incw x13\n"
+      ".inst 0xe0bc86c1  // st1w { za0v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe0972ead  // ld1w { za3h.s[x13, #1] }, p3/Z, [x21, x23, LSL #2]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bfa2c0  // st1w { za0v.s[x13] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x25717120  // psel p0.s, p12.s/Z, p9.s[w13, #1]\n"
-      ".inst 0xe0b1aac4  // st1w { za1v.s[x13] }, p2/Z, [x22, x17, LSL #2]\n"
-      "whilelt p9.s, x14, %x[width]\n"
-      "incw x14\n"
-      ".inst 0xe0a9a6c1  // st1w { za0v.s[x13, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      "add x10, x10, #0x10\n"
-      "incw x23\n"
-      ".inst 0xe0bba2c5  // st1w { za1v.s[x13, #1] }, p0/Z, [x22, x27, LSL #2]\n"
+      "add x9, x9, #0x10\n"
+      ".inst 0xe0ba82c5  // st1w { za1v.s[x12, #1] }, p0/Z, [x22, x26, LSL #2]\n"
       "addvl x22, x22, #4\n"
-      "whilelt p8.s, x14, %x[width]\n"
-      "cbz x15, 7f\n"
+      "incw x23\n"
+      "whilelt p9.s, x13, %x[width]\n"
+      "whilelt p8.s, x13, %x[width]\n"
+      "mov x12, #0x0\n"
+      "cbz x14, 7f\n"
       "6:"  // K loop: Main loop: Second: Loop
+      ".inst 0x25306581  // psel p1.s, p9.s/Z, p12.s[w12]\n"
       ".inst 0x25306160  // psel p0.s, p8.s/Z, p11.s[w12]\n"
-      ".inst 0x25306142  // psel p2.s, p8.s/Z, p10.s[w12]\n"
-      ".inst 0x25706161  // psel p1.s, p8.s/Z, p11.s[w12, #1]\n"
-      ".inst 0x25706143  // psel p3.s, p8.s/Z, p10.s[w12, #1]\n"
-      ".inst 0xe0970380  // ld1w { za0h.s[x12] }, p0/Z, [x28, x23, LSL #2]\n"
-      ".inst 0x25307120  // psel p0.s, p12.s/Z, p9.s[w12]\n"
-      "ldr x28, [x11, #0x0]\n"
-      ".inst 0xe0970b44  // ld1w { za1h.s[x12] }, p2/Z, [x26, x23, LSL #2]\n"
-      ".inst 0x25307122  // psel p2.s, p12.s/Z, p9.s[w12]\n"
-      "ldr x26, [x10, #0x0]\n"
-      ".inst 0xe0970701  // ld1w { za0h.s[x12, #1] }, p1/Z, [x24, x23, LSL #2]\n"
-      ".inst 0x25707121  // psel p1.s, p12.s/Z, p9.s[w12, #1]\n"
-      "ldr x24, [x11, #0x8]\n"
+      ".inst 0xe0970540  // ld1w { za0h.s[x12] }, p1/Z, [x10, x23, LSL #2]\n"
+      "ldr x10, [x11, #0x0]\n"
+      ".inst 0xe0970364  // ld1w { za1h.s[x12] }, p0/Z, [x27, x23, LSL #2]\n"
+      ".inst 0x25706580  // psel p0.s, p9.s/Z, p12.s[w12, #1]\n"
+      ".inst 0x25706162  // psel p2.s, p8.s/Z, p11.s[w12, #1]\n"
+      "ldr x27, [x9, #0x0]\n"
+      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0970321  // ld1w { za0h.s[x12, #1] }, p0/Z, [x25, x23, LSL #2]\n"
+      "ldr x25, [x11, #0x8]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0970aa5  // ld1w { za1h.s[x12, #1] }, p2/Z, [x21, x23, LSL #2]\n"
+      "ldr x21, [x9, #0x8]\n"
+      ".inst 0xe0bf86c8  // st1w { za2v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b082cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
       "add x11, x11, #0x10\n"
-      ".inst 0xe0970ea5  // ld1w { za1h.s[x12, #1] }, p3/Z, [x21, x23, LSL #2]\n"
-      "ldr x21, [x10, #0x8]\n"
-      ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x25707120  // psel p0.s, p12.s/Z, p9.s[w12, #1]\n"
-      ".inst 0xe0b18acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x17, LSL #2]\n"
-      "add x10, x10, #0x10\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      ".inst 0xe0bb82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
+      ".inst 0xe0bc86c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
+      "add x9, x9, #0x10\n"
+      ".inst 0xe0ba82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x26, LSL #2]\n"
       "add x12, x12, #0x2\n"
+      "cmp x12, x14\n"
       "addvl x22, x22, #4\n"
-      "cmp x12, x15\n"
       "blt 6b\n"
       "7:"  // K loop: Main loop: Second: Tail
+      ".inst 0x25306581  // psel p1.s, p9.s/Z, p12.s[w12]\n"
       ".inst 0x25306160  // psel p0.s, p8.s/Z, p11.s[w12]\n"
-      ".inst 0x25306142  // psel p2.s, p8.s/Z, p10.s[w12]\n"
-      ".inst 0x25706161  // psel p1.s, p8.s/Z, p11.s[w12, #1]\n"
-      ".inst 0x25706143  // psel p3.s, p8.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0970540  // ld1w { za0h.s[x12] }, p1/Z, [x10, x23, LSL #2]\n"
+      ".inst 0xe0970364  // ld1w { za1h.s[x12] }, p0/Z, [x27, x23, LSL #2]\n"
       "mov x11, %x[in]\n"
-      "add x10, %x[in], x17, LSL #3\n"
-      ".inst 0xe0970380  // ld1w { za0h.s[x12] }, p0/Z, [x28, x23, LSL #2]\n"
-      ".inst 0x25307120  // psel p0.s, p12.s/Z, p9.s[w12]\n"
-      "ldr x28, [x11, #0x0]\n"
-      ".inst 0xe0970b44  // ld1w { za1h.s[x12] }, p2/Z, [x26, x23, LSL #2]\n"
-      ".inst 0x25307122  // psel p2.s, p12.s/Z, p9.s[w12]\n"
-      "ldr x26, [x10, #0x0]\n"
-      ".inst 0xe0970701  // ld1w { za0h.s[x12, #1] }, p1/Z, [x24, x23, LSL #2]\n"
-      ".inst 0x25707121  // psel p1.s, p12.s/Z, p9.s[w12, #1]\n"
-      "ldr x24, [x11, #0x8]\n"
-      "add x11, x11, #0x10\n"
-      ".inst 0xe0970ea5  // ld1w { za1h.s[x12, #1] }, p3/Z, [x21, x23, LSL #2]\n"
-      "ldr x21, [x10, #0x8]\n"
+      "add x9, %x[in], x16, LSL #3\n"
+      "ldr x10, [x11, #0x0]\n"
+      ".inst 0x25706580  // psel p0.s, p9.s/Z, p12.s[w12, #1]\n"
+      ".inst 0x25706161  // psel p1.s, p8.s/Z, p11.s[w12, #1]\n"
+      ".inst 0xe0970321  // ld1w { za0h.s[x12, #1] }, p0/Z, [x25, x23, LSL #2]\n"
+      "ldr x27, [x9, #0x0]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe09706a5  // ld1w { za1h.s[x12, #1] }, p1/Z, [x21, x23, LSL #2]\n"
+      "ldr x25, [x11, #0x8]\n"
+      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
+      "ldr x21, [x9, #0x8]\n"
       ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0x25707120  // psel p0.s, p12.s/Z, p9.s[w12, #1]\n"
-      ".inst 0xe0b18acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x17, LSL #2]\n"
-      "whilelt p9.s, x14, %x[width]\n"
+      ".inst 0x25707541  // psel p1.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0x25707540  // psel p0.s, p13.s/Z, p10.s[w12, #1]\n"
+      ".inst 0xe0b08acc  // st1w { za3v.s[x12] }, p2/Z, [x22, x16, LSL #2]\n"
+      "whilelt p10.s, x13, %x[width]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xe0a986c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x9, LSL #2]\n"
-      "add x10, x10, #0x10\n"
-      "incw x14\n"
-      ".inst 0xe0bb82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x27, LSL #2]\n"
+      ".inst 0xe0bc86c9  // st1w { za2v.s[x12, #1] }, p1/Z, [x22, x28, LSL #2]\n"
+      "add x11, x11, #0x10\n"
+      "add x9, x9, #0x10\n"
+      ".inst 0xe0ba82cd  // st1w { za3v.s[x12, #1] }, p0/Z, [x22, x26, LSL #2]\n"
       "addvl x22, x22, #4\n"
+      "incw x13\n"
       "incw x23\n"
       "bgt 3b\n"
       "8:"  // K loop: Tails
-      "cbnz x25, 11f\n"
+      "cbnz x24, 11f\n"
       "mov x11, %x[in]\n"
-      "whilelt p8.s, x14, %x[width]\n"
+      "whilelt p9.s, x13, %x[width]\n"
+      "whilelt p8.s, x13, %x[width]\n"
       "mov x12, #0x0\n"
       "9:"  // K loop: Tails: Even: First
-      ".inst 0x25307123  // psel p3.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307122  // psel p2.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25306161  // psel p1.s, p8.s/Z, p11.s[w12]\n"
-      ".inst 0x25306140  // psel p0.s, p8.s/Z, p10.s[w12]\n"
-      ".inst 0xe0bf8ec0  // st1w { za0v.s[x12] }, p3/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0b18ac4  // st1w { za1v.s[x12] }, p2/Z, [x22, x17, LSL #2]\n"
-      "addvl x22, x22, #2\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "ldr x21, [x11, #0x0]\n"
-      "ldr x20, [x11, x17, LSL #0x3]\n"
-      "add x11, x11, #0x8\n"
+      ".inst 0x25306581  // psel p1.s, p9.s/Z, p12.s[w12]\n"
+      ".inst 0x25306160  // psel p0.s, p8.s/Z, p11.s[w12]\n"
+      "ldr x20, [x11, x16, LSL #0x3]\n"
       ".inst 0xe09706a8  // ld1w { za2h.s[x12] }, p1/Z, [x21, x23, LSL #2]\n"
+      "add x11, x11, #0x8\n"
+      "addvl x22, x22, #2\n"
       ".inst 0xe097028c  // ld1w { za3h.s[x12] }, p0/Z, [x20, x23, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "cmp x12, x17\n"
+      "cmp x12, x16\n"
       "blt 9b\n"
-      "whilelt p9.s, x14, %x[width]\n"
-      "whilelt p8.s, x14, %x[width]\n"
+      "whilelt p10.s, x13, %x[width]\n"
+      "whilelt p8.s, x13, %x[width]\n"
+      "whilelt p8.s, x13, %x[width]\n"
       "mov x12, #0x0\n"
       "10:"  // K loop: Tails: Even: Second
-      ".inst 0x25307121  // psel p1.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307120  // psel p0.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0xe0bf86c8  // st1w { za2v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0b182cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x17, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c8  // st1w { za2v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082cc  // st1w { za3v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
+      "cmp x12, x15\n"
       "addvl x22, x22, #2\n"
-      "cmp x12, x16\n"
       "blt 10b\n"
-      "whilelt p8.s, x14, %x[width]\n"
+      "whilelt p8.s, x13, %x[width]\n"
       "b 13f\n"
       "11:"  // K loop: Tails: Odd
       "mov x12, #0x0\n"
       "12:"  // K loop: Tails: Odd: Loop
-      ".inst 0x25307121  // psel p1.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0x25307120  // psel p0.s, p12.s/Z, p9.s[w12]\n"
-      ".inst 0xe0bf86c0  // st1w { za0v.s[x12] }, p1/Z, [x22, XZR, LSL #2]\n"
-      ".inst 0xe0b182c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x17, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0bf82c0  // st1w { za0v.s[x12] }, p0/Z, [x22, XZR, LSL #2]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0b082c4  // st1w { za1v.s[x12] }, p0/Z, [x22, x16, LSL #2]\n"
       "add x12, x12, #0x1\n"
+      "cmp x12, x15\n"
       "addvl x22, x22, #2\n"
-      "cmp x12, x16\n"
       "blt 12b\n"
       "13:"  // K loop: End
       "mov %x[out], x22\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block2_bf16_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block2_bf16_bf16.hpp
index f5c756eba6..6903945536 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block2_bf16_bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block2_bf16_bf16.hpp
@@ -32,93 +32,93 @@ void interleave_block<4, 2, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x17, #0x0\n"
-      "mov x16, %x[row_offset]\n"
+      "cntw x16\n"
       "cntw x15\n"
-      "cntw x14\n"
-      "cntw x11, ALL, MUL #2\n"
-      "cntw x10, ALL, MUL #3\n"
-      "cmp %x[height], x15\n"
-      "cnth x9\n"
-      "csel x15, %x[height], x15, LT\n"
+      "cntw x14, ALL, MUL #2\n"
+      "cntw x13, ALL, MUL #3\n"
+      "cmp %x[height], x16\n"
+      "csel x16, %x[height], x16, LT\n"
       "whilelt p11.h, XZR, %x[height]\n"
-      "whilelt p10.h, x14, %x[height]\n"
-      "whilelt p9.h, x11, %x[height]\n"
-      "whilelt p8.h, x10, %x[height]\n"
+      "whilelt p10.h, x15, %x[height]\n"
+      "whilelt p9.h, x14, %x[height]\n"
+      "whilelt p8.h, x13, %x[height]\n"
+      "mov x11, #0x0\n"
+      "cnth x10\n"
       "ptrue p13.s\n"
-      "sub x15, x15, #0x1\n"
+      "sub x16, x16, #0x1\n"
       "zip1 p12.h, p11.h, p9.h\n"
       "zip1 p11.h, p10.h, p8.h\n"
+      "mov x9, %x[row_offset]\n"
       "mov x28, %x[out]\n"
-      "whilelt p10.h, x17, %x[width]\n"
-      "whilelt p9.h, x17, %x[width]\n"
-      "whilelt p8.h, x17, %x[width]\n"
+      "whilelt p10.h, x11, %x[width]\n"
+      "whilelt p9.h, x11, %x[width]\n"
+      "whilelt p8.h, x11, %x[width]\n"
       "1:"  // Width loop
       "add x27, %x[in], XZR, LSL #3\n"
-      "add x26, %x[in], x14, LSL #3\n"
-      "add x25, %x[in], x11, LSL #3\n"
-      "add x20, %x[in], x10, LSL #3\n"
-      "ldr x24, [x27], #0x8\n"
-      "mov x13, #0x0\n"
+      "add x26, %x[in], x15, LSL #3\n"
+      "ldr x25, [x27], #0x8\n"
+      "add x24, %x[in], x14, LSL #3\n"
+      "add x20, %x[in], x13, LSL #3\n"
       "ldr x23, [x26], #0x8\n"
-      "ldr x22, [x25], #0x8\n"
+      "mov x12, #0x0\n"
+      "ldr x22, [x24], #0x8\n"
       "ldr x21, [x20], #0x8\n"
-      "cbz x15, 3f\n"
+      "cbz x16, 3f\n"
       "2:"  // Loads: Loop
-      ".inst 0x25296580  // psel p0.h, p9.h/Z, p12.h[w13]\n"
-      ".inst 0x25296162  // psel p2.h, p8.h/Z, p11.h[w13]\n"
-      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
-      ".inst 0xe0502300  // ld1h { za0h.h[x13] }, p0/Z, [x24, x16, LSL #1]\n"
-      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
-      "ldr x24, [x27], #0x8\n"
-      ".inst 0xe0502ae8  // ld1h { za1h.h[x13] }, p2/Z, [x23, x16, LSL #1]\n"
+      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      ".inst 0xe0490720  // ld1h { za0h.h[x12] }, p1/Z, [x25, x9, LSL #1]\n"
+      "ldr x25, [x27], #0x8\n"
+      ".inst 0xe04902e8  // ld1h { za1h.h[x12] }, p0/Z, [x23, x9, LSL #1]\n"
+      ".inst 0x25386581  // psel p1.h, p9.h/Z, p12.h[w12, #1]\n"
+      ".inst 0x25386160  // psel p0.h, p8.h/Z, p11.h[w12, #1]\n"
       "ldr x23, [x26], #0x8\n"
-      ".inst 0xe05026c1  // ld1h { za0h.h[x13, #1] }, p1/Z, [x22, x16, LSL #1]\n"
-      "ldr x22, [x25], #0x8\n"
-      ".inst 0xe05022a9  // ld1h { za1h.h[x13, #1] }, p0/Z, [x21, x16, LSL #1]\n"
-      "add x13, x13, #0x2\n"
+      ".inst 0xe04906c1  // ld1h { za0h.h[x12, #1] }, p1/Z, [x22, x9, LSL #1]\n"
+      "ldr x22, [x24], #0x8\n"
+      ".inst 0xe04902a9  // ld1h { za1h.h[x12, #1] }, p0/Z, [x21, x9, LSL #1]\n"
+      "add x12, x12, #0x2\n"
+      "cmp x12, x16, LSL #1\n"
       "ldr x21, [x20], #0x8\n"
-      "cmp x13, x15, LSL #1\n"
       "blt 2b\n"
       "3:"  // Loads: Tail
-      ".inst 0x25296580  // psel p0.h, p9.h/Z, p12.h[w13]\n"
-      ".inst 0x25296162  // psel p2.h, p8.h/Z, p11.h[w13]\n"
-      "sub x20, %x[width], x17\n"
-      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
-      "cmp x20, x9\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0502300  // ld1h { za0h.h[x13] }, p0/Z, [x24, x16, LSL #1]\n"
-      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
-      "csel x20, x20, x9, LT\n"
-      ".inst 0xe0502ae8  // ld1h { za1h.h[x13] }, p2/Z, [x23, x16, LSL #1]\n"
+      ".inst 0x25286581  // psel p1.h, p9.h/Z, p12.h[w12]\n"
+      ".inst 0x25286160  // psel p0.h, p8.h/Z, p11.h[w12]\n"
+      ".inst 0xe0490720  // ld1h { za0h.h[x12] }, p1/Z, [x25, x9, LSL #1]\n"
+      "sub x20, %x[width], x11\n"
+      ".inst 0xe04902e8  // ld1h { za1h.h[x12] }, p0/Z, [x23, x9, LSL #1]\n"
+      "cmp x20, x10\n"
+      "csel x20, x20, x10, LT\n"
+      ".inst 0x25386580  // psel p0.h, p9.h/Z, p12.h[w12, #1]\n"
+      ".inst 0xe04902c1  // ld1h { za0h.h[x12, #1] }, p0/Z, [x22, x9, LSL #1]\n"
+      ".inst 0x25386160  // psel p0.h, p8.h/Z, p11.h[w12, #1]\n"
       "add x20, x20, #0x1\n"
-      ".inst 0xe05026c1  // ld1h { za0h.h[x13, #1] }, p1/Z, [x22, x16, LSL #1]\n"
+      ".inst 0xe04902a9  // ld1h { za1h.h[x12, #1] }, p0/Z, [x21, x9, LSL #1]\n"
+      "mov x12, #0x0\n"
       "lsr x20, x20, #0x1\n"
-      ".inst 0xe05022a9  // ld1h { za1h.h[x13, #1] }, p0/Z, [x21, x16, LSL #1]\n"
       "4:"  // Stores: Loop
       ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
-      ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
-      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
       ".inst 0xe0bf8380  // st1w { za0v.s[x12] }, p0/Z, [x28, XZR, LSL #2]\n"
       ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
-      ".inst 0xe0ae8b84  // st1w { za1v.s[x12] }, p2/Z, [x28, x14, LSL #2]\n"
-      ".inst 0xe0ab8788  // st1w { za2v.s[x12] }, p1/Z, [x28, x11, LSL #2]\n"
-      ".inst 0xe0aa838c  // st1w { za3v.s[x12] }, p0/Z, [x28, x10, LSL #2]\n"
+      ".inst 0xe0af8384  // st1w { za1v.s[x12] }, p0/Z, [x28, x15, LSL #2]\n"
+      ".inst 0x25307541  // psel p1.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
+      ".inst 0xe0ae8788  // st1w { za2v.s[x12] }, p1/Z, [x28, x14, LSL #2]\n"
+      ".inst 0xe0ad838c  // st1w { za3v.s[x12] }, p0/Z, [x28, x13, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x28, x28, #4\n"
       "cmp x12, x20\n"
+      "addvl x28, x28, #4\n"
       "blt 4b\n"
-      "inch x17\n"
-      "inch x16\n"
-      "whilelt p10.h, x17, %x[width]\n"
-      "whilelt p9.h, x17, %x[width]\n"
-      "whilelt p8.h, x17, %x[width]\n"
+      "inch x11\n"
+      "whilelt p10.h, x11, %x[width]\n"
+      "whilelt p9.h, x11, %x[width]\n"
+      "whilelt p8.h, x11, %x[width]\n"
+      "inch x9\n"
       "b.any 1b\n"
       "mov %x[out], x28\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block2_fp16_fp16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block2_fp16_fp16.hpp
index 9e0ab463be..b1ba9a4fe7 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block2_fp16_fp16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block2_fp16_fp16.hpp
@@ -22,7 +22,7 @@
  * SOFTWARE.
  */
 
-#if defined(ARM_COMPUTE_ENABLE_SME)
+#if defined(__ARM_FEATURE_SVE)
 
 template <>
 void interleave_block<4, 2, VLType::SME, false>(
@@ -65,36 +65,36 @@ void interleave_block<4, 2, VLType::SME, false>(
       "ldr x21, [x20], #0x8\n"
       "cbz x15, 3f\n"
       "2:"  // Loads: Loop
-      ".inst 0x25296580  // psel p0.h, p9.h/Z, p12.h[w13]\n"
-      ".inst 0x25296162  // psel p2.h, p8.h/Z, p11.h[w13]\n"
-      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
-      ".inst 0xe0502300  // ld1h { za0h.h[x13] }, p0/Z, [x24, x16, LSL #1]\n"
-      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
+      ".inst 0x25296582  // psel p2.h, p9.h/Z, p12.h[w13]\n"
+      ".inst 0x25296161  // psel p1.h, p8.h/Z, p11.h[w13]\n"
+      ".inst 0x25396580  // psel p0.h, p9.h/Z, p12.h[w13, #1]\n"
+      ".inst 0xe0502b00  // ld1h { za0h.h[x13] }, p2/Z, [x24, x16, LSL #1]\n"
+      ".inst 0x25396162  // psel p2.h, p8.h/Z, p11.h[w13, #1]\n"
       "ldr x24, [x27], #0x8\n"
-      ".inst 0xe0502ae8  // ld1h { za1h.h[x13] }, p2/Z, [x23, x16, LSL #1]\n"
+      ".inst 0xe05026e8  // ld1h { za1h.h[x13] }, p1/Z, [x23, x16, LSL #1]\n"
       "ldr x23, [x26], #0x8\n"
-      ".inst 0xe05026c1  // ld1h { za0h.h[x13, #1] }, p1/Z, [x22, x16, LSL #1]\n"
+      ".inst 0xe05022c1  // ld1h { za0h.h[x13, #1] }, p0/Z, [x22, x16, LSL #1]\n"
       "ldr x22, [x25], #0x8\n"
-      ".inst 0xe05022a9  // ld1h { za1h.h[x13, #1] }, p0/Z, [x21, x16, LSL #1]\n"
+      ".inst 0xe0502aa9  // ld1h { za1h.h[x13, #1] }, p2/Z, [x21, x16, LSL #1]\n"
       "add x13, x13, #0x2\n"
       "ldr x21, [x20], #0x8\n"
       "cmp x13, x15, LSL #1\n"
       "blt 2b\n"
       "3:"  // Loads: Tail
-      ".inst 0x25296580  // psel p0.h, p9.h/Z, p12.h[w13]\n"
-      ".inst 0x25296162  // psel p2.h, p8.h/Z, p11.h[w13]\n"
+      ".inst 0x25296581  // psel p1.h, p9.h/Z, p12.h[w13]\n"
+      ".inst 0x25296160  // psel p0.h, p8.h/Z, p11.h[w13]\n"
       "sub x20, %x[width], x17\n"
-      ".inst 0x25396581  // psel p1.h, p9.h/Z, p12.h[w13, #1]\n"
+      ".inst 0x25396582  // psel p2.h, p9.h/Z, p12.h[w13, #1]\n"
       "cmp x20, x9\n"
       "mov x12, #0x0\n"
-      ".inst 0xe0502300  // ld1h { za0h.h[x13] }, p0/Z, [x24, x16, LSL #1]\n"
-      ".inst 0x25396160  // psel p0.h, p8.h/Z, p11.h[w13, #1]\n"
+      ".inst 0xe0502700  // ld1h { za0h.h[x13] }, p1/Z, [x24, x16, LSL #1]\n"
+      ".inst 0xe05022e8  // ld1h { za1h.h[x13] }, p0/Z, [x23, x16, LSL #1]\n"
+      ".inst 0x25396161  // psel p1.h, p8.h/Z, p11.h[w13, #1]\n"
       "csel x20, x20, x9, LT\n"
-      ".inst 0xe0502ae8  // ld1h { za1h.h[x13] }, p2/Z, [x23, x16, LSL #1]\n"
       "add x20, x20, #0x1\n"
-      ".inst 0xe05026c1  // ld1h { za0h.h[x13, #1] }, p1/Z, [x22, x16, LSL #1]\n"
+      ".inst 0xe0502ac1  // ld1h { za0h.h[x13, #1] }, p2/Z, [x22, x16, LSL #1]\n"
       "lsr x20, x20, #0x1\n"
-      ".inst 0xe05022a9  // ld1h { za1h.h[x13, #1] }, p0/Z, [x21, x16, LSL #1]\n"
+      ".inst 0xe05026a9  // ld1h { za1h.h[x13, #1] }, p1/Z, [x21, x16, LSL #1]\n"
       "4:"  // Stores: Loop
       ".inst 0x25307540  // psel p0.s, p13.s/Z, p10.s[w12]\n"
       ".inst 0x25307542  // psel p2.s, p13.s/Z, p10.s[w12]\n"
@@ -122,4 +122,4 @@ void interleave_block<4, 2, VLType::SME, false>(
     );
 }
 
-#endif  // defined(ARM_COMPUTE_ENABLE_SME)
+#endif  // defined(__ARM_FEATURE_SVE)
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_s8_s8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_s8_s8.hpp
index b0b3aa85c1..9a5521aa8a 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_s8_s8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_s8_s8.hpp
@@ -32,92 +32,92 @@ void interleave_block<4, 4, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x17, #0x0\n"
-      "mov x16, %x[row_offset]\n"
+      "cntw x16\n"
       "cntw x15\n"
-      "cntw x14\n"
-      "cntw x11, ALL, MUL #2\n"
-      "cntw x10, ALL, MUL #3\n"
-      "cmp %x[height], x15\n"
-      "cntb x9\n"
-      "csel x15, %x[height], x15, LT\n"
+      "cntw x14, ALL, MUL #2\n"
+      "cntw x13, ALL, MUL #3\n"
+      "cmp %x[height], x16\n"
+      "csel x16, %x[height], x16, LT\n"
       "whilelt p12.b, XZR, %x[height]\n"
-      "whilelt p10.b, x14, %x[height]\n"
-      "whilelt p9.b, x11, %x[height]\n"
-      "whilelt p8.b, x10, %x[height]\n"
+      "whilelt p10.b, x15, %x[height]\n"
+      "whilelt p9.b, x14, %x[height]\n"
+      "whilelt p8.b, x13, %x[height]\n"
       "zip1 p12.b, p12.b, p9.b\n"
       "zip1 p10.b, p10.b, p8.b\n"
+      "mov x11, #0x0\n"
+      "cntb x10\n"
       "ptrue p11.s\n"
-      "sub x15, x15, #0x1\n"
+      "sub x16, x16, #0x1\n"
       "zip1 p10.b, p12.b, p10.b\n"
+      "mov x9, %x[row_offset]\n"
       "mov x28, %x[out]\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p9.b, x11, %x[width]\n"
+      "whilelt p8.b, x11, %x[width]\n"
       "1:"  // Width loop
       "add x27, %x[in], XZR, LSL #3\n"
-      "add x26, %x[in], x14, LSL #3\n"
-      "add x25, %x[in], x11, LSL #3\n"
-      "add x20, %x[in], x10, LSL #3\n"
-      "ldr x24, [x27], #0x8\n"
-      "mov x13, #0x0\n"
-      "ldr x23, [x26], #0x8\n"
-      "ldr x22, [x25], #0x8\n"
-      "ldr x21, [x20], #0x8\n"
-      "cbz x15, 3f\n"
+      "add x26, %x[in], x15, LSL #3\n"
+      "ldr x25, [x27], #0x8\n"
+      "add x24, %x[in], x14, LSL #3\n"
+      "add x23, %x[in], x13, LSL #3\n"
+      "ldr x20, [x26], #0x8\n"
+      "mov x12, #0x0\n"
+      "ldr x22, [x24], #0x8\n"
+      "ldr x21, [x23], #0x8\n"
+      "cbz x16, 3f\n"
       "2:"  // Loads: Loop
-      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0xe0102300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x16]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      "ldr x24, [x27], #0x8\n"
-      ".inst 0xe0102ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x16]\n"
-      "ldr x23, [x26], #0x8\n"
-      ".inst 0xe01026c2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x22, x16]\n"
-      "ldr x22, [x25], #0x8\n"
-      ".inst 0xe01022a3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x21, x16]\n"
-      "add x13, x13, #0x4\n"
-      "ldr x21, [x20], #0x8\n"
-      "cmp x13, x15, LSL #2\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0090320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x9]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      "ldr x25, [x27], #0x8\n"
+      ".inst 0xe0090281  // ld1b { za0h.b[x12, #1] }, p0/Z, [x20, x9]\n"
+      ".inst 0x25346141  // psel p1.b, p8.b/Z, p10.b[w12, #2]\n"
+      ".inst 0x253c6140  // psel p0.b, p8.b/Z, p10.b[w12, #3]\n"
+      "ldr x20, [x26], #0x8\n"
+      ".inst 0xe00906c2  // ld1b { za0h.b[x12, #2] }, p1/Z, [x22, x9]\n"
+      "ldr x22, [x24], #0x8\n"
+      ".inst 0xe00902a3  // ld1b { za0h.b[x12, #3] }, p0/Z, [x21, x9]\n"
+      "add x12, x12, #0x4\n"
+      "cmp x12, x16, LSL #2\n"
+      "ldr x21, [x23], #0x8\n"
       "blt 2b\n"
       "3:"  // Loads: Tail
-      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      "sub x20, %x[width], x17\n"
-      "cmp x20, x9\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0102300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x16]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      "csel x20, x20, x9, LT\n"
-      ".inst 0xe0102ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x16]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0090320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x9]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0xe0090281  // ld1b { za0h.b[x12, #1] }, p0/Z, [x20, x9]\n"
+      ".inst 0x25346140  // psel p0.b, p8.b/Z, p10.b[w12, #2]\n"
+      "sub x20, %x[width], x11\n"
+      ".inst 0xe00902c2  // ld1b { za0h.b[x12, #2] }, p0/Z, [x22, x9]\n"
+      "cmp x20, x10\n"
+      "csel x20, x20, x10, LT\n"
+      ".inst 0x253c6140  // psel p0.b, p8.b/Z, p10.b[w12, #3]\n"
       "add x20, x20, #0x3\n"
-      ".inst 0xe01026c2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x22, x16]\n"
+      ".inst 0xe00902a3  // ld1b { za0h.b[x12, #3] }, p0/Z, [x21, x9]\n"
+      "mov x12, #0x0\n"
       "lsr x20, x20, #0x2\n"
-      ".inst 0xe01022a3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x21, x16]\n"
       "4:"  // Stores: Loop
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xe0bf8380  // st1w { za0v.s[x12] }, p0/Z, [x28, XZR, LSL #2]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xe0ae8b84  // st1w { za1v.s[x12] }, p2/Z, [x28, x14, LSL #2]\n"
-      ".inst 0xe0ab8788  // st1w { za2v.s[x12] }, p1/Z, [x28, x11, LSL #2]\n"
-      ".inst 0xe0aa838c  // st1w { za3v.s[x12] }, p0/Z, [x28, x10, LSL #2]\n"
+      ".inst 0xe0af8384  // st1w { za1v.s[x12] }, p0/Z, [x28, x15, LSL #2]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0ae8788  // st1w { za2v.s[x12] }, p1/Z, [x28, x14, LSL #2]\n"
+      ".inst 0xe0ad838c  // st1w { za3v.s[x12] }, p0/Z, [x28, x13, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x28, x28, #4\n"
       "cmp x12, x20\n"
+      "addvl x28, x28, #4\n"
       "blt 4b\n"
-      "incb x17\n"
-      "incb x16\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "incb x11\n"
+      "whilelt p9.b, x11, %x[width]\n"
+      "whilelt p8.b, x11, %x[width]\n"
+      "incb x9\n"
       "b.any 1b\n"
       "mov %x[out], x28\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_s8_s8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_s8_s8_summing.hpp
index a4696816e9..22ec1011cc 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_s8_s8_summing.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_s8_s8_summing.hpp
@@ -32,118 +32,118 @@ void interleave_block<4, 4, VLType::SME, true>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x17, %x[row_offset]\n"
-      "mov x16, %x[out]\n"
+      "cntw x16\n"
       "cntw x15\n"
-      "cntw x14\n"
       "mov z24.b, #0x1\n"
+      "cntw x14, ALL, MUL #2\n"
+      "cntw x13, ALL, MUL #3\n"
       "mov z23.s, #0x0\n"
-      "cntw x11, ALL, MUL #2\n"
-      "cntw x10, ALL, MUL #3\n"
       "mov z22.s, #0x0\n"
+      "cmp %x[height], x16\n"
+      "csel x16, %x[height], x16, LT\n"
       "mov z21.s, #0x0\n"
-      "cmp %x[height], x15\n"
-      "ptrue p3.b\n"
       "mov z20.s, #0x0\n"
-      "csel x15, %x[height], x15, LT\n"
       "whilelt p12.b, XZR, %x[height]\n"
-      "whilelt p10.b, x14, %x[height]\n"
-      "whilelt p9.b, x11, %x[height]\n"
-      "whilelt p8.b, x10, %x[height]\n"
+      "whilelt p10.b, x15, %x[height]\n"
+      "whilelt p9.b, x14, %x[height]\n"
+      "whilelt p8.b, x13, %x[height]\n"
       "zip1 p12.b, p12.b, p9.b\n"
       "zip1 p10.b, p10.b, p8.b\n"
-      "cntb x9\n"
+      "ptrue p2.b\n"
+      "cntb x11\n"
       "ptrue p11.s\n"
-      "sub x15, x15, #0x1\n"
+      "sub x16, x16, #0x1\n"
       "zip1 p10.b, p12.b, p10.b\n"
+      "mov x10, %x[row_offset]\n"
+      "mov x9, %x[out]\n"
       "cbnz %x[first], 1f\n"
-      "addvl x16, x16, #-4\n"
-      "ld1w { z23.s }, p3/Z, [x16]\n"
-      "ld1w { z22.s }, p3/Z, [x16, #1, MUL VL]\n"
-      "ld1w { z21.s }, p3/Z, [x16, #2, MUL VL]\n"
-      "ld1w { z20.s }, p3/Z, [x16, #3, MUL VL]\n"
+      "addvl x9, x9, #-4\n"
+      "ld1w { z23.s }, p2/Z, [x9]\n"
+      "ld1w { z22.s }, p2/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z21.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x9, #3, MUL VL]\n"
       "1:"  // Initialise row sums: End
       "mov x28, #0x0\n"
       "whilelt p9.b, x28, %x[width]\n"
       "whilelt p8.b, x28, %x[width]\n"
       "2:"  // Width loop
       "add x27, %x[in], XZR, LSL #3\n"
-      "add x26, %x[in], x14, LSL #3\n"
-      "add x25, %x[in], x11, LSL #3\n"
-      "add x20, %x[in], x10, LSL #3\n"
-      "ldr x24, [x27], #0x8\n"
-      "mov x13, #0x0\n"
-      "ldr x23, [x26], #0x8\n"
-      "ldr x22, [x25], #0x8\n"
-      "ldr x21, [x20], #0x8\n"
-      "cbz x15, 4f\n"
+      "add x26, %x[in], x15, LSL #3\n"
+      "ldr x25, [x27], #0x8\n"
+      "add x24, %x[in], x14, LSL #3\n"
+      "add x23, %x[in], x13, LSL #3\n"
+      "ldr x20, [x26], #0x8\n"
+      "mov x12, #0x0\n"
+      "ldr x22, [x24], #0x8\n"
+      "ldr x21, [x23], #0x8\n"
+      "cbz x16, 4f\n"
       "3:"  // Loads: Loop
-      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0xe0112300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x17]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      "ldr x24, [x27], #0x8\n"
-      ".inst 0xe0112ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x17]\n"
-      "ldr x23, [x26], #0x8\n"
-      ".inst 0xe01126c2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x22, x17]\n"
-      "ldr x22, [x25], #0x8\n"
-      ".inst 0xe01122a3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x21, x17]\n"
-      "add x13, x13, #0x4\n"
-      "ldr x21, [x20], #0x8\n"
-      "cmp x13, x15, LSL #2\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe00a0320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x10]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      "ldr x25, [x27], #0x8\n"
+      ".inst 0xe00a0281  // ld1b { za0h.b[x12, #1] }, p0/Z, [x20, x10]\n"
+      ".inst 0x25346141  // psel p1.b, p8.b/Z, p10.b[w12, #2]\n"
+      ".inst 0x253c6140  // psel p0.b, p8.b/Z, p10.b[w12, #3]\n"
+      "ldr x20, [x26], #0x8\n"
+      ".inst 0xe00a06c2  // ld1b { za0h.b[x12, #2] }, p1/Z, [x22, x10]\n"
+      "ldr x22, [x24], #0x8\n"
+      ".inst 0xe00a02a3  // ld1b { za0h.b[x12, #3] }, p0/Z, [x21, x10]\n"
+      "add x12, x12, #0x4\n"
+      "cmp x12, x16, LSL #2\n"
+      "ldr x21, [x23], #0x8\n"
       "blt 3b\n"
       "4:"  // Loads: Tail
-      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe00a0320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x10]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0xe00a0281  // ld1b { za0h.b[x12, #1] }, p0/Z, [x20, x10]\n"
+      ".inst 0x25346140  // psel p0.b, p8.b/Z, p10.b[w12, #2]\n"
       "sub x20, %x[width], x28\n"
-      "cmp x20, x9\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0112300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x17]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      "csel x20, x20, x9, LT\n"
-      ".inst 0xe0112ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x17]\n"
+      ".inst 0xe00a02c2  // ld1b { za0h.b[x12, #2] }, p0/Z, [x22, x10]\n"
+      "cmp x20, x11\n"
+      "csel x20, x20, x11, LT\n"
+      ".inst 0x253c6140  // psel p0.b, p8.b/Z, p10.b[w12, #3]\n"
       "add x20, x20, #0x3\n"
-      ".inst 0xe01126c2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x22, x17]\n"
+      ".inst 0xe00a02a3  // ld1b { za0h.b[x12, #3] }, p0/Z, [x21, x10]\n"
+      "mov x12, #0x0\n"
       "lsr x20, x20, #0x2\n"
-      ".inst 0xe01122a3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x21, x17]\n"
       "5:"  // Stores: Loop
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0828c11  // mova z17.s, p3/M, za0v.s[x12]\n"
+      ".inst 0xe0bf8120  // st1w { za0v.s[x12] }, p0/Z, [x9, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xc0828811  // mova z17.s, p2/M, za0v.s[x12]\n"
+      ".inst 0xe0af8124  // st1w { za1v.s[x12] }, p0/Z, [x9, x15, LSL #2]\n"
       ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0828c90  // mova z16.s, p3/M, za1v.s[x12]\n"
-      "sdot z23.s, z17.b, z24.b\n"
-      ".inst 0xc0828d13  // mova z19.s, p3/M, za2v.s[x12]\n"
-      "sdot z22.s, z16.b, z24.b\n"
-      ".inst 0xe0bf8200  // st1w { za0v.s[x12] }, p0/Z, [x16, XZR, LSL #2]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0828d92  // mova z18.s, p3/M, za3v.s[x12]\n"
-      "sdot z21.s, z19.b, z24.b\n"
-      ".inst 0xe0ae8a04  // st1w { za1v.s[x12] }, p2/Z, [x16, x14, LSL #2]\n"
-      "sdot z20.s, z18.b, z24.b\n"
-      ".inst 0xe0ab8608  // st1w { za2v.s[x12] }, p1/Z, [x16, x11, LSL #2]\n"
-      ".inst 0xe0aa820c  // st1w { za3v.s[x12] }, p0/Z, [x16, x10, LSL #2]\n"
+      ".inst 0xc0828893  // mova z19.s, p2/M, za1v.s[x12]\n"
+      ".inst 0xe0ae8528  // st1w { za2v.s[x12] }, p1/Z, [x9, x14, LSL #2]\n"
+      ".inst 0xc0828910  // mova z16.s, p2/M, za2v.s[x12]\n"
+      "sdot z23.s, z17.b, z24.b\n"
+      ".inst 0xe0ad812c  // st1w { za3v.s[x12] }, p0/Z, [x9, x13, LSL #2]\n"
+      ".inst 0xc0828992  // mova z18.s, p2/M, za3v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "addvl x16, x16, #4\n"
       "cmp x12, x20\n"
+      "sdot z22.s, z19.b, z24.b\n"
+      "sdot z21.s, z16.b, z24.b\n"
+      "addvl x9, x9, #4\n"
+      "sdot z20.s, z18.b, z24.b\n"
       "blt 5b\n"
       "incb x28\n"
-      "incb x17\n"
       "whilelt p9.b, x28, %x[width]\n"
       "whilelt p8.b, x28, %x[width]\n"
+      "incb x10\n"
       "b.any 2b\n"
-      "st1w { z23.s }, p3, [x16]\n"
-      "st1w { z22.s }, p3, [x16, #1, MUL VL]\n"
-      "st1w { z21.s }, p3, [x16, #2, MUL VL]\n"
-      "st1w { z20.s }, p3, [x16, #3, MUL VL]\n"
-      "addvl x16, x16, #4\n"
+      "st1w { z23.s }, p2, [x9]\n"
+      "st1w { z22.s }, p2, [x9, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x9, #2, MUL VL]\n"
+      "st1w { z20.s }, p2, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
+      "mov %x[out], x9\n"
       ".inst 0xd503467f  // SMSTOP\n"
-      "mov %x[out], x16\n"
       : [out] "+&r" (out)
       : [first] "r" (first), [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_u8_u8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_u8_u8.hpp
index df77398acc..a99c2ea91b 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_u8_u8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_u8_u8.hpp
@@ -32,92 +32,92 @@ void interleave_block<4, 4, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x17, #0x0\n"
-      "mov x16, %x[row_offset]\n"
+      "cntw x16\n"
       "cntw x15\n"
-      "cntw x14\n"
-      "cntw x11, ALL, MUL #2\n"
-      "cntw x10, ALL, MUL #3\n"
-      "cmp %x[height], x15\n"
-      "cntb x9\n"
-      "csel x15, %x[height], x15, LT\n"
+      "cntw x14, ALL, MUL #2\n"
+      "cntw x13, ALL, MUL #3\n"
+      "cmp %x[height], x16\n"
+      "csel x16, %x[height], x16, LT\n"
       "whilelt p12.b, XZR, %x[height]\n"
-      "whilelt p10.b, x14, %x[height]\n"
-      "whilelt p9.b, x11, %x[height]\n"
-      "whilelt p8.b, x10, %x[height]\n"
+      "whilelt p10.b, x15, %x[height]\n"
+      "whilelt p9.b, x14, %x[height]\n"
+      "whilelt p8.b, x13, %x[height]\n"
       "zip1 p12.b, p12.b, p9.b\n"
       "zip1 p10.b, p10.b, p8.b\n"
+      "mov x11, #0x0\n"
+      "cntb x10\n"
       "ptrue p11.s\n"
-      "sub x15, x15, #0x1\n"
+      "sub x16, x16, #0x1\n"
       "zip1 p10.b, p12.b, p10.b\n"
+      "mov x9, %x[row_offset]\n"
       "mov x28, %x[out]\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "whilelt p9.b, x11, %x[width]\n"
+      "whilelt p8.b, x11, %x[width]\n"
       "1:"  // Width loop
       "add x27, %x[in], XZR, LSL #3\n"
-      "add x26, %x[in], x14, LSL #3\n"
-      "add x25, %x[in], x11, LSL #3\n"
-      "add x20, %x[in], x10, LSL #3\n"
-      "ldr x24, [x27], #0x8\n"
-      "mov x13, #0x0\n"
-      "ldr x23, [x26], #0x8\n"
-      "ldr x22, [x25], #0x8\n"
-      "ldr x21, [x20], #0x8\n"
-      "cbz x15, 3f\n"
+      "add x26, %x[in], x15, LSL #3\n"
+      "ldr x25, [x27], #0x8\n"
+      "add x24, %x[in], x14, LSL #3\n"
+      "add x23, %x[in], x13, LSL #3\n"
+      "ldr x20, [x26], #0x8\n"
+      "mov x12, #0x0\n"
+      "ldr x22, [x24], #0x8\n"
+      "ldr x21, [x23], #0x8\n"
+      "cbz x16, 3f\n"
       "2:"  // Loads: Loop
-      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0xe0102300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x16]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      "ldr x24, [x27], #0x8\n"
-      ".inst 0xe0102ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x16]\n"
-      "ldr x23, [x26], #0x8\n"
-      ".inst 0xe01026c2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x22, x16]\n"
-      "ldr x22, [x25], #0x8\n"
-      ".inst 0xe01022a3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x21, x16]\n"
-      "add x13, x13, #0x4\n"
-      "ldr x21, [x20], #0x8\n"
-      "cmp x13, x15, LSL #2\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0090320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x9]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      "ldr x25, [x27], #0x8\n"
+      ".inst 0xe0090281  // ld1b { za0h.b[x12, #1] }, p0/Z, [x20, x9]\n"
+      ".inst 0x25346141  // psel p1.b, p8.b/Z, p10.b[w12, #2]\n"
+      ".inst 0x253c6140  // psel p0.b, p8.b/Z, p10.b[w12, #3]\n"
+      "ldr x20, [x26], #0x8\n"
+      ".inst 0xe00906c2  // ld1b { za0h.b[x12, #2] }, p1/Z, [x22, x9]\n"
+      "ldr x22, [x24], #0x8\n"
+      ".inst 0xe00902a3  // ld1b { za0h.b[x12, #3] }, p0/Z, [x21, x9]\n"
+      "add x12, x12, #0x4\n"
+      "cmp x12, x16, LSL #2\n"
+      "ldr x21, [x23], #0x8\n"
       "blt 2b\n"
       "3:"  // Loads: Tail
-      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      "sub x20, %x[width], x17\n"
-      "cmp x20, x9\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0102300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x16]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      "csel x20, x20, x9, LT\n"
-      ".inst 0xe0102ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x16]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe0090320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x9]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0xe0090281  // ld1b { za0h.b[x12, #1] }, p0/Z, [x20, x9]\n"
+      ".inst 0x25346140  // psel p0.b, p8.b/Z, p10.b[w12, #2]\n"
+      "sub x20, %x[width], x11\n"
+      ".inst 0xe00902c2  // ld1b { za0h.b[x12, #2] }, p0/Z, [x22, x9]\n"
+      "cmp x20, x10\n"
+      "csel x20, x20, x10, LT\n"
+      ".inst 0x253c6140  // psel p0.b, p8.b/Z, p10.b[w12, #3]\n"
       "add x20, x20, #0x3\n"
-      ".inst 0xe01026c2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x22, x16]\n"
+      ".inst 0xe00902a3  // ld1b { za0h.b[x12, #3] }, p0/Z, [x21, x9]\n"
+      "mov x12, #0x0\n"
       "lsr x20, x20, #0x2\n"
-      ".inst 0xe01022a3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x21, x16]\n"
       "4:"  // Stores: Loop
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
       ".inst 0xe0bf8380  // st1w { za0v.s[x12] }, p0/Z, [x28, XZR, LSL #2]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xe0ae8b84  // st1w { za1v.s[x12] }, p2/Z, [x28, x14, LSL #2]\n"
-      ".inst 0xe0ab8788  // st1w { za2v.s[x12] }, p1/Z, [x28, x11, LSL #2]\n"
-      ".inst 0xe0aa838c  // st1w { za3v.s[x12] }, p0/Z, [x28, x10, LSL #2]\n"
+      ".inst 0xe0af8384  // st1w { za1v.s[x12] }, p0/Z, [x28, x15, LSL #2]\n"
+      ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xe0ae8788  // st1w { za2v.s[x12] }, p1/Z, [x28, x14, LSL #2]\n"
+      ".inst 0xe0ad838c  // st1w { za3v.s[x12] }, p0/Z, [x28, x13, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x28, x28, #4\n"
       "cmp x12, x20\n"
+      "addvl x28, x28, #4\n"
       "blt 4b\n"
-      "incb x17\n"
-      "incb x16\n"
-      "whilelt p9.b, x17, %x[width]\n"
-      "whilelt p8.b, x17, %x[width]\n"
+      "incb x11\n"
+      "whilelt p9.b, x11, %x[width]\n"
+      "whilelt p8.b, x11, %x[width]\n"
+      "incb x9\n"
       "b.any 1b\n"
       "mov %x[out], x28\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_u8_u8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_u8_u8_summing.hpp
index 14ab3f476b..0a826d38f2 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_u8_u8_summing.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_block4_u8_u8_summing.hpp
@@ -32,118 +32,118 @@ void interleave_block<4, 4, VLType::SME, true>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x17, %x[row_offset]\n"
-      "mov x16, %x[out]\n"
+      "cntw x16\n"
       "cntw x15\n"
-      "cntw x14\n"
       "mov z24.b, #0x1\n"
+      "cntw x14, ALL, MUL #2\n"
+      "cntw x13, ALL, MUL #3\n"
       "mov z23.s, #0x0\n"
-      "cntw x11, ALL, MUL #2\n"
-      "cntw x10, ALL, MUL #3\n"
       "mov z22.s, #0x0\n"
+      "cmp %x[height], x16\n"
+      "csel x16, %x[height], x16, LT\n"
       "mov z21.s, #0x0\n"
-      "cmp %x[height], x15\n"
-      "ptrue p3.b\n"
       "mov z20.s, #0x0\n"
-      "csel x15, %x[height], x15, LT\n"
       "whilelt p12.b, XZR, %x[height]\n"
-      "whilelt p10.b, x14, %x[height]\n"
-      "whilelt p9.b, x11, %x[height]\n"
-      "whilelt p8.b, x10, %x[height]\n"
+      "whilelt p10.b, x15, %x[height]\n"
+      "whilelt p9.b, x14, %x[height]\n"
+      "whilelt p8.b, x13, %x[height]\n"
       "zip1 p12.b, p12.b, p9.b\n"
       "zip1 p10.b, p10.b, p8.b\n"
-      "cntb x9\n"
+      "ptrue p2.b\n"
+      "cntb x11\n"
       "ptrue p11.s\n"
-      "sub x15, x15, #0x1\n"
+      "sub x16, x16, #0x1\n"
       "zip1 p10.b, p12.b, p10.b\n"
+      "mov x10, %x[row_offset]\n"
+      "mov x9, %x[out]\n"
       "cbnz %x[first], 1f\n"
-      "addvl x16, x16, #-4\n"
-      "ld1w { z23.s }, p3/Z, [x16]\n"
-      "ld1w { z22.s }, p3/Z, [x16, #1, MUL VL]\n"
-      "ld1w { z21.s }, p3/Z, [x16, #2, MUL VL]\n"
-      "ld1w { z20.s }, p3/Z, [x16, #3, MUL VL]\n"
+      "addvl x9, x9, #-4\n"
+      "ld1w { z23.s }, p2/Z, [x9]\n"
+      "ld1w { z22.s }, p2/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z21.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x9, #3, MUL VL]\n"
       "1:"  // Initialise row sums: End
       "mov x28, #0x0\n"
       "whilelt p9.b, x28, %x[width]\n"
       "whilelt p8.b, x28, %x[width]\n"
       "2:"  // Width loop
       "add x27, %x[in], XZR, LSL #3\n"
-      "add x26, %x[in], x14, LSL #3\n"
-      "add x25, %x[in], x11, LSL #3\n"
-      "add x20, %x[in], x10, LSL #3\n"
-      "ldr x24, [x27], #0x8\n"
-      "mov x13, #0x0\n"
-      "ldr x23, [x26], #0x8\n"
-      "ldr x22, [x25], #0x8\n"
-      "ldr x21, [x20], #0x8\n"
-      "cbz x15, 4f\n"
+      "add x26, %x[in], x15, LSL #3\n"
+      "ldr x25, [x27], #0x8\n"
+      "add x24, %x[in], x14, LSL #3\n"
+      "add x23, %x[in], x13, LSL #3\n"
+      "ldr x20, [x26], #0x8\n"
+      "mov x12, #0x0\n"
+      "ldr x22, [x24], #0x8\n"
+      "ldr x21, [x23], #0x8\n"
+      "cbz x16, 4f\n"
       "3:"  // Loads: Loop
-      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
-      ".inst 0xe0112300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x17]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      "ldr x24, [x27], #0x8\n"
-      ".inst 0xe0112ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x17]\n"
-      "ldr x23, [x26], #0x8\n"
-      ".inst 0xe01126c2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x22, x17]\n"
-      "ldr x22, [x25], #0x8\n"
-      ".inst 0xe01122a3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x21, x17]\n"
-      "add x13, x13, #0x4\n"
-      "ldr x21, [x20], #0x8\n"
-      "cmp x13, x15, LSL #2\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe00a0320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x10]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      "ldr x25, [x27], #0x8\n"
+      ".inst 0xe00a0281  // ld1b { za0h.b[x12, #1] }, p0/Z, [x20, x10]\n"
+      ".inst 0x25346141  // psel p1.b, p8.b/Z, p10.b[w12, #2]\n"
+      ".inst 0x253c6140  // psel p0.b, p8.b/Z, p10.b[w12, #3]\n"
+      "ldr x20, [x26], #0x8\n"
+      ".inst 0xe00a06c2  // ld1b { za0h.b[x12, #2] }, p1/Z, [x22, x10]\n"
+      "ldr x22, [x24], #0x8\n"
+      ".inst 0xe00a02a3  // ld1b { za0h.b[x12, #3] }, p0/Z, [x21, x10]\n"
+      "add x12, x12, #0x4\n"
+      "cmp x12, x16, LSL #2\n"
+      "ldr x21, [x23], #0x8\n"
       "blt 3b\n"
       "4:"  // Loads: Tail
-      ".inst 0x25256140  // psel p0.b, p8.b/Z, p10.b[w13]\n"
-      ".inst 0x252d6142  // psel p2.b, p8.b/Z, p10.b[w13, #1]\n"
-      ".inst 0x25356141  // psel p1.b, p8.b/Z, p10.b[w13, #2]\n"
+      ".inst 0x25246140  // psel p0.b, p8.b/Z, p10.b[w12]\n"
+      ".inst 0xe00a0320  // ld1b { za0h.b[x12] }, p0/Z, [x25, x10]\n"
+      ".inst 0x252c6140  // psel p0.b, p8.b/Z, p10.b[w12, #1]\n"
+      ".inst 0xe00a0281  // ld1b { za0h.b[x12, #1] }, p0/Z, [x20, x10]\n"
+      ".inst 0x25346140  // psel p0.b, p8.b/Z, p10.b[w12, #2]\n"
       "sub x20, %x[width], x28\n"
-      "cmp x20, x9\n"
-      "mov x12, #0x0\n"
-      ".inst 0xe0112300  // ld1b { za0h.b[x13] }, p0/Z, [x24, x17]\n"
-      ".inst 0x253d6140  // psel p0.b, p8.b/Z, p10.b[w13, #3]\n"
-      "csel x20, x20, x9, LT\n"
-      ".inst 0xe0112ae1  // ld1b { za0h.b[x13, #1] }, p2/Z, [x23, x17]\n"
+      ".inst 0xe00a02c2  // ld1b { za0h.b[x12, #2] }, p0/Z, [x22, x10]\n"
+      "cmp x20, x11\n"
+      "csel x20, x20, x11, LT\n"
+      ".inst 0x253c6140  // psel p0.b, p8.b/Z, p10.b[w12, #3]\n"
       "add x20, x20, #0x3\n"
-      ".inst 0xe01126c2  // ld1b { za0h.b[x13, #2] }, p1/Z, [x22, x17]\n"
+      ".inst 0xe00a02a3  // ld1b { za0h.b[x12, #3] }, p0/Z, [x21, x10]\n"
+      "mov x12, #0x0\n"
       "lsr x20, x20, #0x2\n"
-      ".inst 0xe01122a3  // ld1b { za0h.b[x13, #3] }, p0/Z, [x21, x17]\n"
       "5:"  // Stores: Loop
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0x25306d22  // psel p2.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0828c13  // mova z19.s, p3/M, za0v.s[x12]\n"
+      ".inst 0xe0bf8120  // st1w { za0v.s[x12] }, p0/Z, [x9, XZR, LSL #2]\n"
+      ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
+      ".inst 0xc0828810  // mova z16.s, p2/M, za0v.s[x12]\n"
+      ".inst 0xe0af8124  // st1w { za1v.s[x12] }, p0/Z, [x9, x15, LSL #2]\n"
       ".inst 0x25306d21  // psel p1.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0828c92  // mova z18.s, p3/M, za1v.s[x12]\n"
-      "udot z23.s, z19.b, z24.b\n"
-      ".inst 0xc0828d11  // mova z17.s, p3/M, za2v.s[x12]\n"
-      "udot z22.s, z18.b, z24.b\n"
-      ".inst 0xe0bf8200  // st1w { za0v.s[x12] }, p0/Z, [x16, XZR, LSL #2]\n"
       ".inst 0x25306d20  // psel p0.s, p11.s/Z, p9.s[w12]\n"
-      ".inst 0xc0828d90  // mova z16.s, p3/M, za3v.s[x12]\n"
-      "udot z21.s, z17.b, z24.b\n"
-      ".inst 0xe0ae8a04  // st1w { za1v.s[x12] }, p2/Z, [x16, x14, LSL #2]\n"
-      "udot z20.s, z16.b, z24.b\n"
-      ".inst 0xe0ab8608  // st1w { za2v.s[x12] }, p1/Z, [x16, x11, LSL #2]\n"
-      ".inst 0xe0aa820c  // st1w { za3v.s[x12] }, p0/Z, [x16, x10, LSL #2]\n"
+      ".inst 0xc0828891  // mova z17.s, p2/M, za1v.s[x12]\n"
+      ".inst 0xe0ae8528  // st1w { za2v.s[x12] }, p1/Z, [x9, x14, LSL #2]\n"
+      ".inst 0xc0828913  // mova z19.s, p2/M, za2v.s[x12]\n"
+      "udot z23.s, z16.b, z24.b\n"
+      ".inst 0xe0ad812c  // st1w { za3v.s[x12] }, p0/Z, [x9, x13, LSL #2]\n"
+      ".inst 0xc0828992  // mova z18.s, p2/M, za3v.s[x12]\n"
       "add x12, x12, #0x1\n"
-      "addvl x16, x16, #4\n"
       "cmp x12, x20\n"
+      "udot z22.s, z17.b, z24.b\n"
+      "udot z21.s, z19.b, z24.b\n"
+      "addvl x9, x9, #4\n"
+      "udot z20.s, z18.b, z24.b\n"
       "blt 5b\n"
       "incb x28\n"
-      "incb x17\n"
       "whilelt p9.b, x28, %x[width]\n"
       "whilelt p8.b, x28, %x[width]\n"
+      "incb x10\n"
       "b.any 2b\n"
-      "st1w { z23.s }, p3, [x16]\n"
-      "st1w { z22.s }, p3, [x16, #1, MUL VL]\n"
-      "st1w { z21.s }, p3, [x16, #2, MUL VL]\n"
-      "st1w { z20.s }, p3, [x16, #3, MUL VL]\n"
-      "addvl x16, x16, #4\n"
+      "st1w { z23.s }, p2, [x9]\n"
+      "st1w { z22.s }, p2, [x9, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x9, #2, MUL VL]\n"
+      "st1w { z20.s }, p2, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
+      "mov %x[out], x9\n"
       ".inst 0xd503467f  // SMSTOP\n"
-      "mov %x[out], x16\n"
       : [out] "+&r" (out)
       : [first] "r" (first), [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_fp32_fp32.hpp
index 09c961f1cd..f6326100b7 100644
--- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_fp32_fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave4VL_fp32_fp32.hpp
@@ -32,92 +32,92 @@ void interleave_block<4, 1, VLType::SME, false>(
 {
   __asm__ __volatile__(
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x16, #0x0\n"
-      "mov x15, %x[row_offset]\n"
+      "cntw x15\n"
+      "cmp %x[height], x15\n"
       "cntw x14\n"
-      "cntw x11\n"
-      "cmp %x[height], x14\n"
-      "cntw x10, ALL, MUL #2\n"
-      "cntw x9, ALL, MUL #3\n"
-      "csel x14, %x[height], x14, LT\n"
+      "cntw x13, ALL, MUL #2\n"
+      "cntw x11, ALL, MUL #3\n"
+      "csel x15, %x[height], x15, LT\n"
+      "mov x10, #0x0\n"
       "ptrue p4.s\n"
-      "sub x14, x14, #0x1\n"
+      "sub x15, x15, #0x1\n"
       "whilelt p3.s, XZR, %x[height]\n"
-      "whilelt p15.s, x11, %x[height]\n"
-      "whilelt p14.s, x10, %x[height]\n"
-      "whilelt p13.s, x9, %x[height]\n"
+      "whilelt p15.s, x14, %x[height]\n"
+      "whilelt p14.s, x13, %x[height]\n"
+      "whilelt p13.s, x11, %x[height]\n"
+      "mov x9, %x[row_offset]\n"
       "mov x28, %x[out]\n"
-      "whilelt p12.s, x16, %x[width]\n"
-      "whilelt p11.s, x16, %x[width]\n"
-      "whilelt p10.s, x16, %x[width]\n"
-      "whilelt p9.s, x16, %x[width]\n"
-      "whilelt p8.s, x16, %x[width]\n"
+      "whilelt p12.s, x10, %x[width]\n"
+      "whilelt p11.s, x10, %x[width]\n"
+      "whilelt p10.s, x10, %x[width]\n"
+      "whilelt p9.s, x10, %x[width]\n"
+      "whilelt p8.s, x10, %x[width]\n"
       "1:"  // Width loop
       "add x27, %x[in], XZR, LSL #3\n"
-      "add x26, %x[in], x11, LSL #3\n"
-      "add x25, %x[in], x10, LSL #3\n"
-      "add x20, %x[in], x9, LSL #3\n"
-      "ldr x24, [x27], #0x8\n"
-      "mov x13, #0x0\n"
+      "add x26, %x[in], x14, LSL #3\n"
+      "ldr x25, [x27], #0x8\n"
+      "add x24, %x[in], x13, LSL #3\n"
+      "add x20, %x[in], x11, LSL #3\n"
       "ldr x23, [x26], #0x8\n"
-      "ldr x22, [x25], #0x8\n"
+      "mov x12, #0x0\n"
+      "ldr x22, [x24], #0x8\n"
       "ldr x21, [x20], #0x8\n"
-      "cbz x14, 3f\n"
+      "cbz x15, 3f\n"
       "2:"  // Loads: Loop
-      ".inst 0x25316c60  // psel p0.s, p11.s/Z, p3.s[w13]\n"
-      ".inst 0x253169e2  // psel p2.s, p10.s/Z, p15.s[w13]\n"
-      ".inst 0x253165c1  // psel p1.s, p9.s/Z, p14.s[w13]\n"
-      ".inst 0xe08f2300  // ld1w { za0h.s[x13] }, p0/Z, [x24, x15, LSL #2]\n"
-      ".inst 0x253161a0  // psel p0.s, p8.s/Z, p13.s[w13]\n"
-      "ldr x24, [x27], #0x8\n"
-      ".inst 0xe08f2ae4  // ld1w { za1h.s[x13] }, p2/Z, [x23, x15, LSL #2]\n"
+      ".inst 0x25306c60  // psel p0.s, p11.s/Z, p3.s[w12]\n"
+      ".inst 0x253069e2  // psel p2.s, p10.s/Z, p15.s[w12]\n"
+      ".inst 0xe0890320  // ld1w { za0h.s[x12] }, p0/Z, [x25, x9, LSL #2]\n"
+      "ldr x25, [x27], #0x8\n"
+      ".inst 0x253065c1  // psel p1.s, p9.s/Z, p14.s[w12]\n"
+      ".inst 0x253061a0  // psel p0.s, p8.s/Z, p13.s[w12]\n"
+      ".inst 0xe0890ae4  // ld1w { za1h.s[x12] }, p2/Z, [x23, x9, LSL #2]\n"
       "ldr x23, [x26], #0x8\n"
-      ".inst 0xe08f26c8  // ld1w { za2h.s[x13] }, p1/Z, [x22, x15, LSL #2]\n"
-      "ldr x22, [x25], #0x8\n"
-      ".inst 0xe08f22ac  // ld1w { za3h.s[x13] }, p0/Z, [x21, x15, LSL #2]\n"
-      "add x13, x13, #0x1\n"
+      ".inst 0xe08906c8  // ld1w { za2h.s[x12] }, p1/Z, [x22, x9, LSL #2]\n"
+      "ldr x22, [x24], #0x8\n"
+      ".inst 0xe08902ac  // ld1w { za3h.s[x12] }, p0/Z, [x21, x9, LSL #2]\n"
+      "add x12, x12, #0x1\n"
+      "cmp x12, x15\n"
       "ldr x21, [x20], #0x8\n"
-      "cmp x13, x14\n"
       "blt 2b\n"
       "3:"  // Loads: Tail
-      ".inst 0x25316c60  // psel p0.s, p11.s/Z, p3.s[w13]\n"
-      ".inst 0x253169e2  // psel p2.s, p10.s/Z, p15.s[w13]\n"
-      ".inst 0x253165c1  // psel p1.s, p9.s/Z, p14.s[w13]\n"
-      "sub x20, %x[width], x16\n"
-      "cmp x20, x11\n"
+      "sub x20, %x[width], x10\n"
+      ".inst 0x25306c60  // psel p0.s, p11.s/Z, p3.s[w12]\n"
+      ".inst 0xe0890320  // ld1w { za0h.s[x12] }, p0/Z, [x25, x9, LSL #2]\n"
+      ".inst 0x253069e0  // psel p0.s, p10.s/Z, p15.s[w12]\n"
+      ".inst 0x253065c1  // psel p1.s, p9.s/Z, p14.s[w12]\n"
+      ".inst 0xe08902e4  // ld1w { za1h.s[x12] }, p0/Z, [x23, x9, LSL #2]\n"
+      ".inst 0x253061a0  // psel p0.s, p8.s/Z, p13.s[w12]\n"
+      "cmp x20, x14\n"
+      ".inst 0xe08906c8  // ld1w { za2h.s[x12] }, p1/Z, [x22, x9, LSL #2]\n"
+      ".inst 0xe08902ac  // ld1w { za3h.s[x12] }, p0/Z, [x21, x9, LSL #2]\n"
       "mov x12, #0x0\n"
-      ".inst 0xe08f2300  // ld1w { za0h.s[x13] }, p0/Z, [x24, x15, LSL #2]\n"
-      ".inst 0x253161a0  // psel p0.s, p8.s/Z, p13.s[w13]\n"
-      "csel x20, x20, x11, LT\n"
-      ".inst 0xe08f2ae4  // ld1w { za1h.s[x13] }, p2/Z, [x23, x15, LSL #2]\n"
-      ".inst 0xe08f26c8  // ld1w { za2h.s[x13] }, p1/Z, [x22, x15, LSL #2]\n"
-      ".inst 0xe08f22ac  // ld1w { za3h.s[x13] }, p0/Z, [x21, x15, LSL #2]\n"
+      "csel x20, x20, x14, LT\n"
       "4:"  // Stores: Loop
       ".inst 0x25305180  // psel p0.s, p4.s/Z, p12.s[w12]\n"
-      ".inst 0x25305182  // psel p2.s, p4.s/Z, p12.s[w12]\n"
-      ".inst 0x25305181  // psel p1.s, p4.s/Z, p12.s[w12]\n"
       ".inst 0xe0bf8380  // st1w { za0v.s[x12] }, p0/Z, [x28, XZR, LSL #2]\n"
       ".inst 0x25305180  // psel p0.s, p4.s/Z, p12.s[w12]\n"
-      ".inst 0xe0ab8b84  // st1w { za1v.s[x12] }, p2/Z, [x28, x11, LSL #2]\n"
-      ".inst 0xe0aa8788  // st1w { za2v.s[x12] }, p1/Z, [x28, x10, LSL #2]\n"
-      ".inst 0xe0a9838c  // st1w { za3v.s[x12] }, p0/Z, [x28, x9, LSL #2]\n"
+      ".inst 0xe0ae8384  // st1w { za1v.s[x12] }, p0/Z, [x28, x14, LSL #2]\n"
+      ".inst 0x25305181  // psel p1.s, p4.s/Z, p12.s[w12]\n"
+      ".inst 0x25305180  // psel p0.s, p4.s/Z, p12.s[w12]\n"
+      ".inst 0xe0ad8788  // st1w { za2v.s[x12] }, p1/Z, [x28, x13, LSL #2]\n"
+      ".inst 0xe0ab838c  // st1w { za3v.s[x12] }, p0/Z, [x28, x11, LSL #2]\n"
       "add x12, x12, #0x1\n"
-      "addvl x28, x28, #4\n"
       "cmp x12, x20\n"
+      "addvl x28, x28, #4\n"
       "blt 4b\n"
-      "incw x16\n"
-      "incw x15\n"
-      "whilelt p12.s, x16, %x[width]\n"
-      "whilelt p11.s, x16, %x[width]\n"
-      "whilelt p10.s, x16, %x[width]\n"
-      "whilelt p9.s, x16, %x[width]\n"
-      "whilelt p8.s, x16, %x[width]\n"
+      "incw x10\n"
+      "whilelt p12.s, x10, %x[width]\n"
+      "whilelt p11.s, x10, %x[width]\n"
+      "whilelt p10.s, x10, %x[width]\n"
+      "whilelt p9.s, x10, %x[width]\n"
+      "whilelt p8.s, x10, %x[width]\n"
+      "incw x9\n"
       "b.any 1b\n"
       "mov %x[out], x28\n"
       ".inst 0xd503467f  // SMSTOP\n"
       : [out] "+&r" (out)
       : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp b/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp
index da27f31428..7c09608e3e 100644
--- a/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp
+++ b/src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp
@@ -47,7 +47,245 @@
 
 namespace arm_gemm {
 
-#include "interleave_indirect_impl.hpp"
+/*
+ * Core function that does heavy lifting - interleave 'int_by' rows of width 'width' together.
+ *
+ * 'height' indicates the actual number of rows to interleave, so if it's less than int_by then the remaining
+ * entries are padded (note that this is "GEMM" padding rather than convolution padding, so there is no need to pad
+ * with a particular value.
+ *
+ * Note that it is not expected for this templated version to ever be used - all cases that matter should be
+ * explicitly specialized with an optimized implementation.
+ */
+template<unsigned int height_vectors, unsigned int block, VLType vlt, bool integrate_sums, typename TIn, typename TOut>
+void interleave_block( TOut * &out, const TIn * const *in, size_t width, size_t height, size_t row_offset, bool first) {
+    const unsigned int int_by = height_vectors * (vlt == VLType::SVE ? get_vector_length<TOut>() / block : 1);
+
+    std::vector<int32_t> the_sums;
+
+    if (integrate_sums) {
+        the_sums = std::vector<int32_t>(int_by, 0);
+
+        if (!first) {
+            // In 'integrate sums' mode, we dump the sums at the end on each pass.
+
+            // On the last pass this is correct, but on other passes it is not -
+            // so on the subsequent pass we need to take the output written by
+            // the previous pass as starting point for the sums, and then
+            // overwrite them with new interleaved data.
+            int32_t *out_int32 = reinterpret_cast<int32_t *>(out);
+
+            // Rewind pointer to where we wrote out the sums last time.
+            out_int32 -= int_by;
+
+            // Restore the running sums.
+            memcpy(the_sums.data(), out_int32, int_by * sizeof(int32_t));
+
+            // Update the "real" pointer so that the next output will clobber the old sums.
+            out = reinterpret_cast<TOut *>(out_int32);
+        }
+    }
+
+    for (unsigned int pos=0; pos<width; pos+=block) {
+        for (unsigned int row=0; row<int_by; row++) {
+            // Row out of range - pad 'block' entries.
+            if (row >= height) {
+                for (unsigned int col=0; col<block; col++) {
+                    *out++ = 0;
+                }
+                continue;
+            }
+
+            for (unsigned int col=0; col<block; col++) {
+                // Column out of range - pad a single entry
+                if (pos + col >= width) {
+                    *out++ = 0;
+                    continue;
+                }
+
+                if (integrate_sums) {
+                    the_sums[row] += in[row][row_offset + pos + col];
+                }
+
+                *out++ = in[row][row_offset + pos + col];
+            }
+        }
+    }
+
+    if (integrate_sums) {
+        int32_t *out_int32 = reinterpret_cast<int32_t *>(out);
+
+        memcpy(out_int32, the_sums.data(), int_by * sizeof(int32_t));
+
+        out = reinterpret_cast<TOut *>(out_int32 + int_by);
+    }
+}
+
+template<unsigned int height_vectors, unsigned int block, VLType vlt, typename TOut>
+inline void FixupRowSums(TOut * &out, const int32_t row_sum_multiplier) {
+    const unsigned int height = height_vectors * (vlt == VLType::SVE ? get_vector_length<TOut>() / block : 1);
+
+    // If we are integrating row sums, we need to do some fix up, depending on whether the multiplier is non-zero or not.
+    if (row_sum_multiplier) {
+        // Non-zero: interleave_block<>() will have done the sums, so 'out' will point to the start of the
+        // next block (post sums).
+        // We need to go back and apply the multiplier to the computed sums.  We don't need to change 'out'.
+        int32_t *out_int32 = reinterpret_cast<int32_t *>(out);
+
+        out_int32 -= height;
+        for (unsigned int i=0; i<height; i++) {
+            out_int32[i] *= row_sum_multiplier;
+        }
+    } else {
+        // Zero: interleave_block<>() will *not* have done the sums, so 'out' will point to the start of the
+        // sum block.  We need to insert the (zero) sums, and advance 'out'.
+        int32_t *out_int32 = reinterpret_cast<int32_t *>(out);
+
+        for (unsigned int i=0; i<height; i++) {
+            out_int32[i] = 0;
+        }
+
+        out_int32 += height;
+
+        out = reinterpret_cast<TOut *>(out_int32);
+    }
+}
+
+template<unsigned int height_vectors, unsigned int block, VLType vlt, typename TIn, typename TOut>
+void IndirectInterleave(TOut *out, const TIn * const * const *ptr, unsigned int stringlen,
+                        unsigned int rounded_stringlen, const unsigned int y0, const unsigned int ymax,
+                        const unsigned int k0, const unsigned int kmax, bool integrate_sums,
+                        const int32_t row_sum_multiplier) {
+    const unsigned int height = height_vectors * (vlt == VLType::SVE ? get_vector_length<TOut>() / block : 1);
+
+    // 'interleave_block' implementations are entitled to read a pointer for each row they handle from the input
+    // pointer array, even for out of range rows (although they must not subsequently dereference those pointers for
+    // out of range rows).  This allows interleave_block to use techniques like row predication, or loading all
+    // pointers and conditionally overriding the out of range ones.
+
+    // This is problematic in the "pure" indirect case when we get to the last rows, where it can lead to out of
+    // range reads.  Avoid this with a local buffer to use in last-rows cases.  Use alloca as a std::vector can be
+    // expensive in highly threaded scenarios.
+    const TIn **row_ptrs = reinterpret_cast<const TIn **>(alloca(height * sizeof(const TIn *)));
+
+    // Figure out the starting position based on k0 (with rounded length)
+    unsigned int start_string      = k0 / rounded_stringlen;
+    unsigned int start_stringpos   = k0 % rounded_stringlen;
+
+    // Process blocks of 'height' height...
+    for (unsigned int ybase = y0; ybase < ymax; ybase+=height) {
+        // Height to process
+        unsigned int active_height = std::min(ymax - ybase, height);
+
+        // Track our progress through the various strings
+        unsigned int k_left    = (kmax - k0);
+        unsigned int string    = start_string;
+        unsigned int stringpos = start_stringpos;
+
+        bool first = true;
+
+        // Prepare to call 'interleave_block' above for each string encompassed by K range
+        while (k_left > 0) {
+            // Width to process - and the width we will generate (with padding)
+            unsigned int in_width   = std::min(k_left, stringlen - stringpos);
+            unsigned int out_width  = std::min(k_left, rounded_stringlen - stringpos);
+
+            const TIn * const *row_base = ptr[string] + ybase;
+
+            // If not all rows are valid, copy the ones that are into local array (see above comment).
+            if (active_height < height) {
+                for (unsigned int i=0; i<active_height; i++) {
+                    row_ptrs[i] = ptr[string][ybase + i];
+                }
+
+                row_base = row_ptrs;
+            }
+
+            // 'integrate_sums' is a function parameter rather than a template parameter to prevent duplicating too
+            // much code.  However, integrated sums make no sense for non-integral types and won't ever be
+            // requested.  So put a type trait check here to avoid generating pointless code.
+            if (std::is_integral<TOut>::value && integrate_sums && row_sum_multiplier) {
+                interleave_block<height_vectors, block, vlt, true>(out, row_base, in_width, active_height, stringpos, first);
+            } else {
+                interleave_block<height_vectors, block, vlt, false>(out, row_base, in_width, active_height, stringpos, first);
+            }
+
+            k_left -= out_width;
+            string++;
+            stringpos=0;
+            first=false;
+        }
+
+        if (std::is_integral<TOut>::value && integrate_sums) {
+            FixupRowSums<height_vectors, block, vlt>(out, row_sum_multiplier);
+        }
+    }
+}
+
+template<unsigned int height_vectors, unsigned int block, VLType vlt, typename TIn, typename TOut>
+void ConvolutionInterleave(TOut *out, const TIn *in, size_t in_stride, const convolver<TIn> &conv, const unsigned int rounded_stringlen,
+        const unsigned int y0, const unsigned int ymax, const unsigned int k0, const unsigned int kmax, bool integrate_sums, const int32_t row_sum_multiplier) {
+    const unsigned int height = height_vectors * (vlt == VLType::SVE ? get_vector_length<TOut>() / block : 1);
+
+    auto conv_cols = conv.process_columns(in, in_stride, k0, kmax, rounded_stringlen);
+
+    // Use alloca here as a std::vector can be expensive in highly threaded scenarios.
+    const TIn **row_ptrs = reinterpret_cast<const TIn **>(alloca(height * sizeof(const TIn *)));
+
+    for (unsigned int ybase = y0; ybase < ymax; ybase += height) {
+        // How many of the rows are active - the rest will get padded in interleave_block.
+        unsigned int active_height   = std::min(ymax - ybase, height);
+        bool first = true;
+
+        auto conv_rows = conv_cols.process_rows(ybase, active_height);
+
+        while (!conv_rows.finished()) {
+            unsigned int width, offset;
+
+            // Get next set of parameters
+            std::tie(width, offset) = conv_rows.next_block(row_ptrs);
+
+            // Perform the interleave
+            if (std::is_integral<TOut>::value && integrate_sums && row_sum_multiplier) {
+                interleave_block<height_vectors, block, vlt, true>(out, row_ptrs, width, active_height, offset, first);
+            } else {
+                interleave_block<height_vectors, block, vlt, false>(out, row_ptrs, width, active_height, offset, first);
+            }
+
+            first=false;
+        }
+
+        if (std::is_integral<TOut>::value && integrate_sums) {
+            FixupRowSums<height_vectors, block, vlt>(out, row_sum_multiplier);
+        }
+    }
+}
+
+template<unsigned int height_vectors, unsigned int block, VLType vlt, typename TIn, typename TOut>
+void Interleave(TOut *out, const TIn *in, size_t in_stride, const unsigned int y0, const unsigned int ymax, const unsigned int k0, const unsigned int kmax, bool integrate_sums, const int32_t row_sum_multiplier) {
+    const unsigned int height = height_vectors * (vlt == VLType::SVE ? get_vector_length<TOut>() / block : 1);
+
+    // Use alloca here as a std::vector can be expensive in highly threaded scenarios.
+    const TIn **row_ptrs = reinterpret_cast<const TIn **>(alloca(height * sizeof(const TIn *)));
+
+    const unsigned int width=kmax-k0;
+
+    for (unsigned int y=y0; y<ymax; y+=height) {
+        for (unsigned int r=0; r<height; r++) {
+            row_ptrs[r] = in + ((y + r) * in_stride);
+        }
+
+        if (std::is_integral<TOut>::value && integrate_sums && row_sum_multiplier) {
+            interleave_block<height_vectors, block, vlt, true>(out, row_ptrs, width, std::min(height, ymax-y), k0, true);
+        } else {
+            interleave_block<height_vectors, block, vlt, false>(out, row_ptrs, width, std::min(height, ymax-y), k0, true);
+        }
+
+        if (std::is_integral<TOut>::value && integrate_sums) {
+            FixupRowSums<height_vectors, block, vlt>(out, row_sum_multiplier);
+        }
+    }
+}
 
 #include "indirect-interleaves/list.hpp"
 
@@ -78,7 +316,6 @@ template void Interleave<6, 1, VLType::None>(float *, const bfloat16 *, size_t,
 
 /* AArch64 */
 #ifdef __aarch64__
-
 /* FP32 */
 /* Arm® Neon™/SVE implementation (height 8) */
 template void IndirectInterleave<8, 1, VLType::None>(float *, const float * const * const *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int, bool, int32_t);
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp
index 011459c157..935958b224 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp
@@ -44,8 +44,7 @@ void a32_sgemm_8x6_a55r1(const float *, const float *, float *, int, int, int);
 // structure.
 class sgemm_8x6 {
 public:
-    typedef float lhs_operand_type;
-    typedef float rhs_operand_type;
+    typedef float operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
@@ -64,7 +63,7 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 8> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 6, 8> transforms = {};
 
     kern_type kernel = a32_sgemm_8x6;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp
index 95c2682bf6..32c9515582 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp
@@ -56,347 +56,347 @@ void a32_sgemm_8x6_a53(const float *Apanel, const float *Bpanel, float *Cpanel,
             int k = ((K+3)/4) - 1;
 
             __asm __volatile (
-                "vmov.i32    q4, #0\n"
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]\n"
-                "vmov.i32    q5, #0\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]\n"
-                "vmov.i32    q6, #0\n"
-                "ldr        r0, [%[a_ptr], #0x10]\n"
-                "vmov.i32    q7, #0\n"
-                "ldr        r1, [%[a_ptr], #0x14]\n"
-                "vmov.i32    q8, #0\n"
+                "vmov.i32	q4, #0\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]\n"
+                "vmov.i32	q5, #0\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]\n"
+                "vmov.i32	q6, #0\n"
+                "ldr		r0, [%[a_ptr], #0x10]\n"
+                "vmov.i32	q7, #0\n"
+                "ldr		r1, [%[a_ptr], #0x14]\n"
+                "vmov.i32	q8, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #0x40]")
-                "vmov.i32    q9, #0\n"
+                "vmov.i32	q9, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #0x40]")
-                "vmov.i32    q10, #0\n"
+                "vmov.i32	q10, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #0x80]")
-                "vmov.i32    q11, #0\n"
+                "vmov.i32	q11, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #0x80]")
-                "vmov.i32    q12, #0\n"
-                "vmov.i32    q13, #0\n"
+                "vmov.i32	q12, #0\n"
+                "vmov.i32	q13, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #0xC0]")
-                "vmov.i32    q14, #0\n"
+                "vmov.i32	q14, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #0XC0]")
-                "vmov.i32    q15, #0\n"
-                "cmp        %[k], #0\n"
-                "beq        6f\n"
+                "vmov.i32	q15, #0\n"
+                "cmp		%[k], #0\n"
+                "beq		6f\n"
 
                 "1:\n"
                 // Unroll 0
-                "vldr        d6, [%[b_ptr], #0x10]\n"
-                "vmov        d2, r0, r1\n"
-                "vmla.f32    q4, q2, d0[0]\n"
-                "ldr        r0, [%[b_ptr], #0x18]\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "ldr        r1, [%[b_ptr], #0x1C]\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-
-                "vldr        d3, [%[a_ptr], #0x18]\n"
-                "vmov        d7, r0, r1\n"
-                "vmla.f32    q7, q2, d1[1]\n"
+                "vldr		d6, [%[b_ptr], #0x10]\n"
+                "vmov		d2, r0, r1\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "ldr		r0, [%[b_ptr], #0x18]\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "ldr		r1, [%[b_ptr], #0x1C]\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+
+                "vldr		d3, [%[a_ptr], #0x18]\n"
+                "vmov		d7, r0, r1\n"
+                "vmla.f32	q7, q2, d1[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #0x100]")
-                "vmla.f32    q8, q2, d2[0]\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-
-                "vldr        d4, [%[b_ptr], #0x20]\n"
-                "vmla.f32    q10, q3, d0[0]\n"
-                "ldr        r0, [%[b_ptr], #0x28]\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "ldr        r1, [%[b_ptr], #0x2C]\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-
-                "vldr        d0, [%[a_ptr], #0x20]\n"
-                "vmov        d5, r0, r1\n"
-                "vmla.f32    q13, q3, d1[1]\n"
-                "ldr        r0, [%[a_ptr], #0x28]\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-                "ldr        r1, [%[a_ptr], #0x2C]\n"
-                "vmla.f32    q15, q3, d2[1]\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+
+                "vldr		d4, [%[b_ptr], #0x20]\n"
+                "vmla.f32	q10, q3, d0[0]\n"
+                "ldr		r0, [%[b_ptr], #0x28]\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "ldr		r1, [%[b_ptr], #0x2C]\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+
+                "vldr		d0, [%[a_ptr], #0x20]\n"
+                "vmov		d5, r0, r1\n"
+                "vmla.f32	q13, q3, d1[1]\n"
+                "ldr		r0, [%[a_ptr], #0x28]\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+                "ldr		r1, [%[a_ptr], #0x2C]\n"
+                "vmla.f32	q15, q3, d2[1]\n"
 
                 // Unroll 1
-                "vldr        d6, [%[b_ptr], #0x30]\n"
-                "vmov        d1, r0, r1\n"
-                "vmla.f32    q4, q2, d3[0]\n"
-                "ldr        r0, [%[b_ptr], #0x38]\n"
-                "vmla.f32    q5, q2, d3[1]\n"
-                "ldr        r1, [%[b_ptr], #0x3C]\n"
-                "vmla.f32    q6, q2, d0[0]\n"
-
-                "vldr        d2, [%[a_ptr], #0x30]\n"
-                "vmov        d7, r0, r1\n"
-                "vmla.f32    q7, q2, d0[1]\n"
+                "vldr		d6, [%[b_ptr], #0x30]\n"
+                "vmov		d1, r0, r1\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "ldr		r0, [%[b_ptr], #0x38]\n"
+                "vmla.f32	q5, q2, d3[1]\n"
+                "ldr		r1, [%[b_ptr], #0x3C]\n"
+                "vmla.f32	q6, q2, d0[0]\n"
+
+                "vldr		d2, [%[a_ptr], #0x30]\n"
+                "vmov		d7, r0, r1\n"
+                "vmla.f32	q7, q2, d0[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #0x100]")
-                "vmla.f32    q8, q2, d1[0]\n"
-                "vmla.f32    q9, q2, d1[1]\n"
-
-                "vldr        d4, [%[b_ptr], #0x40]\n"
-                "vmla.f32    q10, q3, d3[0]\n"
-                "ldr        r0, [%[b_ptr], #0x48]\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "ldr        r1, [%[b_ptr], #0x4C]\n"
-                "vmla.f32    q12, q3, d0[0]\n"
-
-                "vldr        d3, [%[a_ptr], #0x38]\n"
-                "vmov        d5, r0, r1\n"
-                "vmla.f32    q13, q3, d0[1]\n"
-                "ldr        r0, [%[a_ptr], #0x40]\n"
-                "vmla.f32    q14, q3, d1[0]\n"
-                "ldr        r1, [%[a_ptr], #0x44]\n"
-                "vmla.f32    q15, q3, d1[1]\n"
+                "vmla.f32	q8, q2, d1[0]\n"
+                "vmla.f32	q9, q2, d1[1]\n"
+
+                "vldr		d4, [%[b_ptr], #0x40]\n"
+                "vmla.f32	q10, q3, d3[0]\n"
+                "ldr		r0, [%[b_ptr], #0x48]\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "ldr		r1, [%[b_ptr], #0x4C]\n"
+                "vmla.f32	q12, q3, d0[0]\n"
+
+                "vldr		d3, [%[a_ptr], #0x38]\n"
+                "vmov		d5, r0, r1\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "ldr		r0, [%[a_ptr], #0x40]\n"
+                "vmla.f32	q14, q3, d1[0]\n"
+                "ldr		r1, [%[a_ptr], #0x44]\n"
+                "vmla.f32	q15, q3, d1[1]\n"
 
                 // Unroll 2
-                "vldr        d6, [%[b_ptr], #0x50]\n"
-                "vmov        d0, r0, r1\n"
-                "vmla.f32    q4, q2, d2[0]\n"
-                "ldr        r0, [%[b_ptr], #0x58]\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "ldr        r1, [%[b_ptr], #0x5C]\n"
-                "vmla.f32    q6, q2, d3[0]\n"
-
-                "vldr        d1, [%[a_ptr], #0x48]\n"
-                "vmov        d7, r0, r1\n"
-                "vmla.f32    q7, q2, d3[1]\n"
+                "vldr		d6, [%[b_ptr], #0x50]\n"
+                "vmov		d0, r0, r1\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "ldr		r0, [%[b_ptr], #0x58]\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "ldr		r1, [%[b_ptr], #0x5C]\n"
+                "vmla.f32	q6, q2, d3[0]\n"
+
+                "vldr		d1, [%[a_ptr], #0x48]\n"
+                "vmov		d7, r0, r1\n"
+                "vmla.f32	q7, q2, d3[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #0x140]")
-                "vmla.f32    q8, q2, d0[0]\n"
-                "vmla.f32    q9, q2, d0[1]\n"
-
-                "vldr        d4, [%[b_ptr], #0x60]\n"
-                "vmla.f32    q10, q3, d2[0]\n"
-                "ldr        r0, [%[b_ptr], #0x68]\n"
-                "vmla.f32    q11, q3, d2[1]\n"
-                "ldr        r1, [%[b_ptr], #0x6C]\n"
-                "vmla.f32    q12, q3, d3[0]\n"
-
-                "vldr        d2, [%[a_ptr], #0x50]\n"
-                "vmov        d5, r0, r1\n"
-                "vmla.f32    q13, q3, d3[1]\n"
-                "ldr        r0, [%[a_ptr], #0x58]\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "ldr        r1, [%[a_ptr], #0x5C]\n"
-                "vmla.f32    q15, q3, d0[1]\n"
-                "add        %[a_ptr], %[a_ptr], #0x60\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+                "vmla.f32	q9, q2, d0[1]\n"
+
+                "vldr		d4, [%[b_ptr], #0x60]\n"
+                "vmla.f32	q10, q3, d2[0]\n"
+                "ldr		r0, [%[b_ptr], #0x68]\n"
+                "vmla.f32	q11, q3, d2[1]\n"
+                "ldr		r1, [%[b_ptr], #0x6C]\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+
+                "vldr		d2, [%[a_ptr], #0x50]\n"
+                "vmov		d5, r0, r1\n"
+                "vmla.f32	q13, q3, d3[1]\n"
+                "ldr		r0, [%[a_ptr], #0x58]\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "ldr		r1, [%[a_ptr], #0x5C]\n"
+                "vmla.f32	q15, q3, d0[1]\n"
+                "add		%[a_ptr], %[a_ptr], #0x60\n"
 
                 // Unroll 3
-                "vldr        d6, [%[b_ptr], #0x70]\n"
-                "vmov        d3, r0, r1\n"
-                "vmla.f32    q4, q2, d1[0]\n"
-                "ldr        r0, [%[b_ptr], #0x78]\n"
-                "vmla.f32    q5, q2, d1[1]\n"
-                "ldr        r1, [%[b_ptr], #0x7C]\n"
-                "vmla.f32    q6, q2, d2[0]\n"
-                "add        %[b_ptr], %[b_ptr], #0x80\n"
-
-                "vldr        d0, [%[a_ptr], #0x00]\n"
-                "vmov        d7, r0, r1\n"
-                "vmla.f32    q7, q2, d2[1]\n"
+                "vldr		d6, [%[b_ptr], #0x70]\n"
+                "vmov		d3, r0, r1\n"
+                "vmla.f32	q4, q2, d1[0]\n"
+                "ldr		r0, [%[b_ptr], #0x78]\n"
+                "vmla.f32	q5, q2, d1[1]\n"
+                "ldr		r1, [%[b_ptr], #0x7C]\n"
+                "vmla.f32	q6, q2, d2[0]\n"
+                "add		%[b_ptr], %[b_ptr], #0x80\n"
+
+                "vldr		d0, [%[a_ptr], #0x00]\n"
+                "vmov		d7, r0, r1\n"
+                "vmla.f32	q7, q2, d2[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #0xC0]")
-                "vmla.f32    q8, q2, d3[0]\n"
-                "vmla.f32    q9, q2, d3[1]\n"
-
-                "vldr        d4, [%[b_ptr], #0x00]\n"
-                "vmla.f32    q10, q3, d1[0]\n"
-                "ldr        r0, [%[b_ptr], #0x08]\n"
-                "vmla.f32    q11, q3, d1[1]\n"
-                "ldr        r1, [%[b_ptr], #0x0C]\n"
-                "vmla.f32    q12, q3, d2[0]\n"
-                "subs        %[k], %[k], #1\n"
-
-                "vldr        d1, [%[a_ptr], #0x08]\n"
-                "vmov        d5, r0, r1\n"
-                "vmla.f32    q13, q3, d2[1]\n"
-                "ldr        r0, [%[a_ptr], #0x10]\n"
-                "vmla.f32    q14, q3, d3[0]\n"
-                "ldr        r1, [%[a_ptr], #0x14]\n"
-                "vmla.f32    q15, q3, d3[1]\n"
-                "bne        1b\n"
+                "vmla.f32	q8, q2, d3[0]\n"
+                "vmla.f32	q9, q2, d3[1]\n"
+
+                "vldr		d4, [%[b_ptr], #0x00]\n"
+                "vmla.f32	q10, q3, d1[0]\n"
+                "ldr		r0, [%[b_ptr], #0x08]\n"
+                "vmla.f32	q11, q3, d1[1]\n"
+                "ldr		r1, [%[b_ptr], #0x0C]\n"
+                "vmla.f32	q12, q3, d2[0]\n"
+                "subs		%[k], %[k], #1\n"
+
+                "vldr		d1, [%[a_ptr], #0x08]\n"
+                "vmov		d5, r0, r1\n"
+                "vmla.f32	q13, q3, d2[1]\n"
+                "ldr		r0, [%[a_ptr], #0x10]\n"
+                "vmla.f32	q14, q3, d3[0]\n"
+                "ldr		r1, [%[a_ptr], #0x14]\n"
+                "vmla.f32	q15, q3, d3[1]\n"
+                "bne		1b\n"
 
                 // "Tails" shows how many multiply blocks are needed at the
                 // end, must be 1-4 inclusive.  Bail out to alternative tail
                 // immediately if it's 1.
                 "6:\n"
-                "subs        %[tails], %[tails], #1\n"
-                "beq        3f\n"
+                "subs		%[tails], %[tails], #1\n"
+                "beq		3f\n"
 
                 // Detached final iteration - for now adapt the generic
                 // tails rather than reimplementing for A53.
 
                 // Unroll 0
-                "vmov        d2, r0, r1\n"
-                "add        %[a_ptr], %[a_ptr], #0x18\n"
-                "vmla.f32    q4, q2, d0[0]\n"
-                "vld1.32    {d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "add        %[b_ptr], %[b_ptr], #0x10\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d1[1]\n"
-                "vmla.f32    q8, q2, d2[0]\n"
-                "subs        %[tails], %[tails], #1\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d0[0]\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-                "vmla.f32    q13, q3, d1[1]\n"
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-                "vmla.f32    q15, q3, d2[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "beq        4f\n"
+                "vmov		d2, r0, r1\n"
+                "add		%[a_ptr], %[a_ptr], #0x18\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "vld1.32	{d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "add		%[b_ptr], %[b_ptr], #0x10\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d1[1]\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "subs		%[tails], %[tails], #1\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d0[0]\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+                "vmla.f32	q13, q3, d1[1]\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+                "vmla.f32	q15, q3, d2[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "beq		4f\n"
 
                 // Unroll 1
-                "vmla.f32    q4, q2, d3[0]\n"
-                "vmla.f32    q5, q2, d3[1]\n"
-                "subs        %[tails], %[tails], #1\n"
-                "vmla.f32    q6, q2, d0[0]\n"
-                "vmla.f32    q7, q2, d0[1]\n"
-                "vmla.f32    q8, q2, d1[0]\n"
-                "vmla.f32    q9, q2, d1[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d3[0]\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "vld1.32    {d2-d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q12, q3, d0[0]\n"
-                "vmla.f32    q13, q3, d0[1]\n"
-                "vmla.f32    q14, q3, d1[0]\n"
-                "vmla.f32    q15, q3, d1[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "beq        5f\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "vmla.f32	q5, q2, d3[1]\n"
+                "subs		%[tails], %[tails], #1\n"
+                "vmla.f32	q6, q2, d0[0]\n"
+                "vmla.f32	q7, q2, d0[1]\n"
+                "vmla.f32	q8, q2, d1[0]\n"
+                "vmla.f32	q9, q2, d1[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d3[0]\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "vld1.32	{d2-d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q12, q3, d0[0]\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "vmla.f32	q14, q3, d1[0]\n"
+                "vmla.f32	q15, q3, d1[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "beq		5f\n"
 
                 // Unroll 2
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q4, q2, d2[0]\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "vmla.f32    q6, q2, d3[0]\n"
-                "vmla.f32    q7, q2, d3[1]\n"
-                "vmla.f32    q8, q2, d0[0]\n"
-                "vmla.f32    q9, q2, d0[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d2[0]\n"
-                "vmla.f32    q11, q3, d2[1]\n"
-                "vmla.f32    q12, q3, d3[0]\n"
-                "vmla.f32    q13, q3, d3[1]\n"
-                "vld1.32    {d2-d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "vmla.f32    q15, q3, d0[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "vmla.f32	q6, q2, d3[0]\n"
+                "vmla.f32	q7, q2, d3[1]\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+                "vmla.f32	q9, q2, d0[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d2[0]\n"
+                "vmla.f32	q11, q3, d2[1]\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+                "vmla.f32	q13, q3, d3[1]\n"
+                "vld1.32	{d2-d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "vmla.f32	q15, q3, d0[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
 
                 // Unroll 3
-                "vmla.f32    q4, q2, d1[0]\n"
-                "vmla.f32    q10, q3, d1[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q5, q2, d1[1]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d1[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d2[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d2[0]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d2[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d2[1]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d3[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d3[0]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d3[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d3[1]\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "b        2f\n"
+                "vmla.f32	q4, q2, d1[0]\n"
+                "vmla.f32	q10, q3, d1[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q5, q2, d1[1]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d1[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d2[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d2[0]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d2[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d2[1]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d3[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d3[0]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d3[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d3[1]\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "b		2f\n"
 
                 // tails==1 final tail
                 "3:\n"
-                "vmov        d2, r0, r1\n"
-                "add        %[b_ptr], %[b_ptr], #0x10\n"
-                "vmla.f32    q4, q2, d0[0]\n"
-                "add        %[a_ptr], %[a_ptr], #0x18\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q10, q3, d0[0]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d1[1]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d1[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d2[0]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d2[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "b        2f\n"
+                "vmov		d2, r0, r1\n"
+                "add		%[b_ptr], %[b_ptr], #0x10\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "add		%[a_ptr], %[a_ptr], #0x18\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q10, q3, d0[0]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d1[1]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d1[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d2[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "b		2f\n"
 
                 // tails==2 final tail
                 "4:\n"
-                "vmla.f32    q4, q2, d3[0]\n"
-                "vmla.f32    q10, q3, d3[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q5, q2, d3[1]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d0[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d0[0]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d0[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d0[1]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d1[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d1[0]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d1[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d1[1]\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "b        2f\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "vmla.f32	q10, q3, d3[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q5, q2, d3[1]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d0[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d0[0]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d0[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d1[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d1[0]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d1[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d1[1]\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "b		2f\n"
 
                 // tails==3 final tail
                 "5:\n"
-                "vmla.f32    q4, q2, d2[0]\n"
-                "vld1.32    {d0}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "vmla.f32    q6, q2, d3[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q10, q3, d2[0]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d2[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d3[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d3[1]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d3[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d0[0]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d0[1]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d0[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "vld1.32	{d0}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "vmla.f32	q6, q2, d3[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q10, q3, d2[0]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d2[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d3[1]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d3[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d0[1]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d0[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
 
                 "2:\n"
-                "vst1.32    {d30-d31}, [%[c_ptr] :128]!\n"
+                "vst1.32	{d30-d31}, [%[c_ptr] :128]!\n"
             : [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr), [k] "+r" (k), [tails] "+r" (tails)
             :
             : "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15",
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp
index 54e0a26843..972b66ed50 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp
@@ -62,345 +62,345 @@ void a32_sgemm_8x6_a55r1(const float *Apanel, const float *Bpanel, float *Cpanel
             a_ptr = a_ptr0;
 
             __asm __volatile (
-                "vldr        d0, [%[a_ptr]]\n"
-                "vmov.i32    q4, #0\n"
-                "vldr        d1, [%[a_ptr], #0x08]\n"
-                "vmov.i32    q5, #0\n"
-                "vldr        d4, [%[b_ptr]]\n"
-                "vmov.i32    q6, #0\n"
-                "vldr        d5, [%[b_ptr], #0x08]\n"
-                "vmov.i32    q7, #0\n"
-                "vldr        d2, [%[a_ptr], #0x10]\n"
-                "vmov.i32    q8, #0\n"
+                "vldr		d0, [%[a_ptr]]\n"
+                "vmov.i32	q4, #0\n"
+                "vldr		d1, [%[a_ptr], #0x08]\n"
+                "vmov.i32	q5, #0\n"
+                "vldr		d4, [%[b_ptr]]\n"
+                "vmov.i32	q6, #0\n"
+                "vldr		d5, [%[b_ptr], #0x08]\n"
+                "vmov.i32	q7, #0\n"
+                "vldr		d2, [%[a_ptr], #0x10]\n"
+                "vmov.i32	q8, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #0x40]")
-                "vmov.i32    q9, #0\n"
+                "vmov.i32	q9, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #0x40]")
-                "vmov.i32    q10, #0\n"
+                "vmov.i32	q10, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #0x80]")
-                "vmov.i32    q11, #0\n"
+                "vmov.i32	q11, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #0x80]")
-                "vmov.i32    q12, #0\n"
+                "vmov.i32	q12, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #0XC0]")
-                "vmov.i32    q13, #0\n"
+                "vmov.i32	q13, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #0xC0]")
-                "vmov.i32    q14, #0\n"
+                "vmov.i32	q14, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #0x100]")
-                "vmov.i32    q15, #0\n"
+                "vmov.i32	q15, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #0x100]")
-                "cmp        %[k], #0\n"
+                "cmp		%[k], #0\n"
                 ASM_PREFETCH("[%[b_ptr], #0x140]")
-                "beq        6f\n"
+                "beq		6f\n"
                 ASM_PREFETCH("[%[b_ptr], #0x180]")
 
                 "1:\n"
                 // Unroll 0
-                "vmla.f32    q4, q2, d0[0]\n"
-                "vldr        d6, [%[b_ptr], #0x10]\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "vldr        d7, [%[b_ptr], #0x18]\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-                "vldr        d3, [%[a_ptr], #0x18]\n"
-                "vmla.f32    q7, q2, d1[1]\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "vldr		d6, [%[b_ptr], #0x10]\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "vldr		d7, [%[b_ptr], #0x18]\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+                "vldr		d3, [%[a_ptr], #0x18]\n"
+                "vmla.f32	q7, q2, d1[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #0x140]")
-                "vmla.f32    q8, q2, d2[0]\n"
-                "subs        %[k], %[k], #1\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-                "vldr        d4, [%[b_ptr], #0x20]\n"
-                "vmla.f32    q10, q3, d0[0]\n"
-                "vldr        d5, [%[b_ptr], #0x28]\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "vldr        d0, [%[a_ptr], #0x20]\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-
-                "vmla.f32    q13, q3, d1[1]\n"
-                "vldr        d1, [%[a_ptr], #0x28]\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-
-                "vmla.f32    q15, q3, d2[1]\n"
-                "vldr        d6, [%[b_ptr], #0x30]\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "subs		%[k], %[k], #1\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+                "vldr		d4, [%[b_ptr], #0x20]\n"
+                "vmla.f32	q10, q3, d0[0]\n"
+                "vldr		d5, [%[b_ptr], #0x28]\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "vldr		d0, [%[a_ptr], #0x20]\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+
+                "vmla.f32	q13, q3, d1[1]\n"
+                "vldr		d1, [%[a_ptr], #0x28]\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+
+                "vmla.f32	q15, q3, d2[1]\n"
+                "vldr		d6, [%[b_ptr], #0x30]\n"
 
                 // Unroll 1
-                "vmla.f32    q4, q2, d3[0]\n"
-                "vldr        d7, [%[b_ptr], #0x38]\n"
-                "vmla.f32    q5, q2, d3[1]\n"
-                "vldr        d2, [%[a_ptr], #0x30]\n"
-                "vmla.f32    q6, q2, d0[0]\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "vldr		d7, [%[b_ptr], #0x38]\n"
+                "vmla.f32	q5, q2, d3[1]\n"
+                "vldr		d2, [%[a_ptr], #0x30]\n"
+                "vmla.f32	q6, q2, d0[0]\n"
 
-                "vmla.f32    q7, q2, d0[1]\n"
+                "vmla.f32	q7, q2, d0[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #0x1C0]")
-                "vmla.f32    q8, q2, d1[0]\n"
+                "vmla.f32	q8, q2, d1[0]\n"
 
-                "vmla.f32    q9, q2, d1[1]\n"
-                "vldr        d4, [%[b_ptr], #0x40]\n"
-                "vmla.f32    q10, q3, d3[0]\n"
-                "vldr        d5, [%[b_ptr], #0x48]\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "vldr        d3, [%[a_ptr], #0x38]\n"
-                "vmla.f32    q12, q3, d0[0]\n"
+                "vmla.f32	q9, q2, d1[1]\n"
+                "vldr		d4, [%[b_ptr], #0x40]\n"
+                "vmla.f32	q10, q3, d3[0]\n"
+                "vldr		d5, [%[b_ptr], #0x48]\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "vldr		d3, [%[a_ptr], #0x38]\n"
+                "vmla.f32	q12, q3, d0[0]\n"
 
-                "vmla.f32    q13, q3, d0[1]\n"
-                "vldr        d0, [%[a_ptr], #0x40]\n"
-                "vmla.f32    q14, q3, d1[0]\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "vldr		d0, [%[a_ptr], #0x40]\n"
+                "vmla.f32	q14, q3, d1[0]\n"
 
-                "vmla.f32    q15, q3, d1[1]\n"
-                "vldr        d6, [%[b_ptr], #0x50]\n"
+                "vmla.f32	q15, q3, d1[1]\n"
+                "vldr		d6, [%[b_ptr], #0x50]\n"
 
                 // Unroll 2
-                "vmla.f32    q4, q2, d2[0]\n"
-                "vldr        d7, [%[b_ptr], #0x58]\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "vldr        d1, [%[a_ptr], #0x48]\n"
-                "vmla.f32    q6, q2, d3[0]\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "vldr		d7, [%[b_ptr], #0x58]\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "vldr		d1, [%[a_ptr], #0x48]\n"
+                "vmla.f32	q6, q2, d3[0]\n"
 
-                "vmla.f32    q7, q2, d3[1]\n"
+                "vmla.f32	q7, q2, d3[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #0x180]")
-                "vmla.f32    q8, q2, d0[0]\n"
-
-                "vmla.f32    q9, q2, d0[1]\n"
-                "vldr        d4, [%[b_ptr], #0x60]\n"
-                "vmla.f32    q10, q3, d2[0]\n"
-                "vldr        d5, [%[b_ptr], #0x68]\n"
-                "vmla.f32    q11, q3, d2[1]\n"
-                "vldr        d2, [%[a_ptr], #0x50]\n"
-                "vmla.f32    q12, q3, d3[0]\n"
-
-                "vmla.f32    q13, q3, d3[1]\n"
-                "vldr        d3, [%[a_ptr], #0x58]\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "add        %[a_ptr], %[a_ptr], #0x60\n"
-                "vmla.f32    q15, q3, d0[1]\n"
-                "vldr        d6, [%[b_ptr], #0x70]\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+
+                "vmla.f32	q9, q2, d0[1]\n"
+                "vldr		d4, [%[b_ptr], #0x60]\n"
+                "vmla.f32	q10, q3, d2[0]\n"
+                "vldr		d5, [%[b_ptr], #0x68]\n"
+                "vmla.f32	q11, q3, d2[1]\n"
+                "vldr		d2, [%[a_ptr], #0x50]\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+
+                "vmla.f32	q13, q3, d3[1]\n"
+                "vldr		d3, [%[a_ptr], #0x58]\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "add		%[a_ptr], %[a_ptr], #0x60\n"
+                "vmla.f32	q15, q3, d0[1]\n"
+                "vldr		d6, [%[b_ptr], #0x70]\n"
 
                 // Unroll 3
-                "vmla.f32    q4, q2, d1[0]\n"
-                "vldr        d7, [%[b_ptr], #0x78]\n"
-                "vmla.f32    q5, q2, d1[1]\n"
-                "add        %[b_ptr], %[b_ptr], #0x80\n"
-                "vmla.f32    q6, q2, d2[0]\n"
-                "vldr        d0, [%[a_ptr], #0x00]\n"
-                "vmla.f32    q7, q2, d2[1]\n"
+                "vmla.f32	q4, q2, d1[0]\n"
+                "vldr		d7, [%[b_ptr], #0x78]\n"
+                "vmla.f32	q5, q2, d1[1]\n"
+                "add		%[b_ptr], %[b_ptr], #0x80\n"
+                "vmla.f32	q6, q2, d2[0]\n"
+                "vldr		d0, [%[a_ptr], #0x00]\n"
+                "vmla.f32	q7, q2, d2[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #0x180]")
-                "vmla.f32    q8, q2, d3[0]\n"
+                "vmla.f32	q8, q2, d3[0]\n"
 
-                "vmla.f32    q9, q2, d3[1]\n"
-                "vldr        d4, [%[b_ptr], #0x00]\n"
-                "vmla.f32    q10, q3, d1[0]\n"
-                "vldr        d5, [%[b_ptr], #0x08]\n"
-                "vmla.f32    q11, q3, d1[1]\n"
-                "vldr        d1, [%[a_ptr], #0x08]\n"
-                "vmla.f32    q12, q3, d2[0]\n"
+                "vmla.f32	q9, q2, d3[1]\n"
+                "vldr		d4, [%[b_ptr], #0x00]\n"
+                "vmla.f32	q10, q3, d1[0]\n"
+                "vldr		d5, [%[b_ptr], #0x08]\n"
+                "vmla.f32	q11, q3, d1[1]\n"
+                "vldr		d1, [%[a_ptr], #0x08]\n"
+                "vmla.f32	q12, q3, d2[0]\n"
 
-                "vmla.f32    q13, q3, d2[1]\n"
-                "vldr        d2, [%[a_ptr], #0x10]\n"
-                "vmla.f32    q14, q3, d3[0]\n"
+                "vmla.f32	q13, q3, d2[1]\n"
+                "vldr		d2, [%[a_ptr], #0x10]\n"
+                "vmla.f32	q14, q3, d3[0]\n"
 
-                "vmla.f32    q15, q3, d3[1]\n"
-                "bne        1b\n"
+                "vmla.f32	q15, q3, d3[1]\n"
+                "bne		1b\n"
 
                 // "Tails" shows how many multiply blocks are needed at the
                 // end, must be 1-4 inclusive.  Bail out to alternative tail
                 // immediately if it's 1.
                 "6:\n"
-                "subs        %[tails], %[tails], #1\n"
-                "beq        3f\n"
+                "subs		%[tails], %[tails], #1\n"
+                "beq		3f\n"
 
                 // Detached final iteration
 
                 // Unroll 0
-                "vmla.f32    q4, q2, d0[0]\n"
-                "vldr        d6, [%[b_ptr], #0x10]\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "vldr        d7, [%[b_ptr], #0x18]\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-                "vldr        d3, [%[a_ptr], #0x18]\n"
-                "vmla.f32    q7, q2, d1[1]\n"
-                "subs        %[tails], %[tails], #1\n"
-                "vmla.f32    q8, q2, d2[0]\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-                "vldr        d4, [%[b_ptr], #0x20]\n"
-
-                "vmla.f32    q10, q3, d0[0]\n"
-                "vldr        d5, [%[b_ptr], #0x28]\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "vldr        d0, [%[a_ptr], #0x20]\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-                "vmla.f32    q13, q3, d1[1]\n"
-                "vldr        d1, [%[a_ptr], #0x28]\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-                "vmla.f32    q15, q3, d2[1]\n"
-                "beq        4f\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "vldr		d6, [%[b_ptr], #0x10]\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "vldr		d7, [%[b_ptr], #0x18]\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+                "vldr		d3, [%[a_ptr], #0x18]\n"
+                "vmla.f32	q7, q2, d1[1]\n"
+                "subs		%[tails], %[tails], #1\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+                "vldr		d4, [%[b_ptr], #0x20]\n"
+
+                "vmla.f32	q10, q3, d0[0]\n"
+                "vldr		d5, [%[b_ptr], #0x28]\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "vldr		d0, [%[a_ptr], #0x20]\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+                "vmla.f32	q13, q3, d1[1]\n"
+                "vldr		d1, [%[a_ptr], #0x28]\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+                "vmla.f32	q15, q3, d2[1]\n"
+                "beq		4f\n"
 
                 // Unroll 1
-                "vmla.f32    q4, q2, d3[0]\n"
-                "vldr        d6, [%[b_ptr], #0x30]\n"
-                "vmla.f32    q5, q2, d3[1]\n"
-                "vldr        d7, [%[b_ptr], #0x38]\n"
-                "vmla.f32    q6, q2, d0[0]\n"
-                "vldr        d2, [%[a_ptr], #0x30]\n"
-                "vmla.f32    q7, q2, d0[1]\n"
-                "subs        %[tails], %[tails], #1\n"
-                "vmla.f32    q8, q2, d1[0]\n"
-
-                "vmla.f32    q9, q2, d1[1]\n"
-
-                "vmla.f32    q10, q3, d3[0]\n"
-                "vldr        d4, [%[b_ptr], #0x40]\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "vldr        d5, [%[b_ptr], #0x48]\n"
-                "vmla.f32    q12, q3, d0[0]\n"
-                "vldr        d3, [%[a_ptr], #0x38]\n"
-                "vmla.f32    q13, q3, d0[1]\n"
-                "vldr        d0, [%[a_ptr], #0x40]\n"
-                "vmla.f32    q14, q3, d1[0]\n"
-                "vmla.f32    q15, q3, d1[1]\n"
-                "beq        5f\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "vldr		d6, [%[b_ptr], #0x30]\n"
+                "vmla.f32	q5, q2, d3[1]\n"
+                "vldr		d7, [%[b_ptr], #0x38]\n"
+                "vmla.f32	q6, q2, d0[0]\n"
+                "vldr		d2, [%[a_ptr], #0x30]\n"
+                "vmla.f32	q7, q2, d0[1]\n"
+                "subs		%[tails], %[tails], #1\n"
+                "vmla.f32	q8, q2, d1[0]\n"
+
+                "vmla.f32	q9, q2, d1[1]\n"
+
+                "vmla.f32	q10, q3, d3[0]\n"
+                "vldr		d4, [%[b_ptr], #0x40]\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "vldr		d5, [%[b_ptr], #0x48]\n"
+                "vmla.f32	q12, q3, d0[0]\n"
+                "vldr		d3, [%[a_ptr], #0x38]\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "vldr		d0, [%[a_ptr], #0x40]\n"
+                "vmla.f32	q14, q3, d1[0]\n"
+                "vmla.f32	q15, q3, d1[1]\n"
+                "beq		5f\n"
 
                 // Unroll 2
-                "vmla.f32    q4, q2, d2[0]\n"
-                "vldr        d6, [%[b_ptr], #0x50]\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "vldr        d7, [%[b_ptr], #0x58]\n"
-                "vmla.f32    q6, q2, d3[0]\n"
-                "vldr        d1, [%[a_ptr], #0x48]\n"
-                "vmla.f32    q7, q2, d3[1]\n"
-                "vmla.f32    q8, q2, d0[0]\n"
-                "vmla.f32    q9, q2, d0[1]\n"
-
-                "vmla.f32    q10, q3, d2[0]\n"
-                "vldr        d4, [%[b_ptr], #0x60]\n"
-                "vmla.f32    q11, q3, d2[1]\n"
-                "vldr        d5, [%[b_ptr], #0x68]\n"
-                "vmla.f32    q12, q3, d3[0]\n"
-                "vldr        d2, [%[a_ptr], #0x50]\n"
-                "vmla.f32    q13, q3, d3[1]\n"
-                "vldr        d3, [%[a_ptr], #0x58]\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "vmla.f32    q15, q3, d0[1]\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "vldr		d6, [%[b_ptr], #0x50]\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "vldr		d7, [%[b_ptr], #0x58]\n"
+                "vmla.f32	q6, q2, d3[0]\n"
+                "vldr		d1, [%[a_ptr], #0x48]\n"
+                "vmla.f32	q7, q2, d3[1]\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+                "vmla.f32	q9, q2, d0[1]\n"
+
+                "vmla.f32	q10, q3, d2[0]\n"
+                "vldr		d4, [%[b_ptr], #0x60]\n"
+                "vmla.f32	q11, q3, d2[1]\n"
+                "vldr		d5, [%[b_ptr], #0x68]\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+                "vldr		d2, [%[a_ptr], #0x50]\n"
+                "vmla.f32	q13, q3, d3[1]\n"
+                "vldr		d3, [%[a_ptr], #0x58]\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "vmla.f32	q15, q3, d0[1]\n"
 
                 // Unroll 3
-                "vmla.f32    q4, q2, d1[0]\n"
-                "vldr        d6, [%[b_ptr], #0x70]\n"
-                "vmla.f32    q5, q2, d1[1]\n"
-                "vldr        d7, [%[b_ptr], #0x78]\n"
-                "vmla.f32    q10, q3, d1[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d1[1]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d2[0]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d2[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d2[1]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d2[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d3[0]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d3[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d3[1]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d3[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "add        %[a_ptr], %[a_ptr], #0x60\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "add        %[b_ptr], %[b_ptr], #0x80\n"
-                "b        2f\n"
+                "vmla.f32	q4, q2, d1[0]\n"
+                "vldr		d6, [%[b_ptr], #0x70]\n"
+                "vmla.f32	q5, q2, d1[1]\n"
+                "vldr		d7, [%[b_ptr], #0x78]\n"
+                "vmla.f32	q10, q3, d1[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d1[1]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d2[0]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d2[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d2[1]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d2[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d3[0]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d3[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d3[1]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d3[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "add		%[a_ptr], %[a_ptr], #0x60\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "add		%[b_ptr], %[b_ptr], #0x80\n"
+                "b		2f\n"
 
                 // tails==1 final tail
                 "3:\n"
-                "vmla.f32    q4, q2, d0[0]\n"
-                "vldr        d6, [%[b_ptr], #0x10]\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "vldr        d7, [%[b_ptr], #0x18]\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q10, q3, d0[0]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d1[1]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d1[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d2[0]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d2[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "add        %[a_ptr], %[a_ptr], #0x18\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "add        %[b_ptr], %[b_ptr], #0x20\n"
-                "b        2f\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "vldr		d6, [%[b_ptr], #0x10]\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "vldr		d7, [%[b_ptr], #0x18]\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q10, q3, d0[0]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d1[1]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d1[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d2[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "add		%[a_ptr], %[a_ptr], #0x18\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "add		%[b_ptr], %[b_ptr], #0x20\n"
+                "b		2f\n"
 
                 // tails==2 final tail
                 "4:\n"
-                "vmla.f32    q4, q2, d3[0]\n"
-                "vldr        d6, [%[b_ptr], #0x30]\n"
-                "vmla.f32    q5, q2, d3[1]\n"
-                "vldr        d7, [%[b_ptr], #0x38]\n"
-                "vmla.f32    q10, q3, d3[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d0[0]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d0[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d0[1]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d0[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d1[0]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d1[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d1[1]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d1[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "add        %[b_ptr], %[b_ptr], #0x40\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "add        %[a_ptr], %[a_ptr], #0x30\n"
-                "b        2f\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "vldr		d6, [%[b_ptr], #0x30]\n"
+                "vmla.f32	q5, q2, d3[1]\n"
+                "vldr		d7, [%[b_ptr], #0x38]\n"
+                "vmla.f32	q10, q3, d3[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d0[0]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d0[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d0[1]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d1[0]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d1[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d1[1]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d1[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "add		%[b_ptr], %[b_ptr], #0x40\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "add		%[a_ptr], %[a_ptr], #0x30\n"
+                "b		2f\n"
 
                 // tails==3 final tail
                 "5:\n"
-                "vmla.f32    q4, q2, d2[0]\n"
-                "vldr        d6, [%[b_ptr], #0x50]\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "vldr        d7, [%[b_ptr], #0x58]\n"
-                "vmla.f32    q6, q2, d3[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q10, q3, d2[0]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d2[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d3[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d3[1]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d3[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d0[0]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d0[1]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d0[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "add        %[a_ptr], %[a_ptr], #0x48\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "add        %[b_ptr], %[b_ptr], #0x60\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "vldr		d6, [%[b_ptr], #0x50]\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "vldr		d7, [%[b_ptr], #0x58]\n"
+                "vmla.f32	q6, q2, d3[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q10, q3, d2[0]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d2[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d3[1]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d3[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d0[1]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d0[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "add		%[a_ptr], %[a_ptr], #0x48\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "add		%[b_ptr], %[b_ptr], #0x60\n"
 
                 "2:\n"
-                "vst1.32    {d30-d31}, [%[c_ptr] :128]!\n"
+                "vst1.32	{d30-d31}, [%[c_ptr] :128]!\n"
             : [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr), [k] "+r" (k), [tails] "+r" (tails)
             :
             : "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15",
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp
index b230dc1fb7..877247e052 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp
@@ -56,110 +56,110 @@ void a32_sgemm_8x6(const float *Apanel, const float *Bpanel, float *Cpanel, int
             int k = ((K+3)/4) - 1;
 
             __asm __volatile (
-                "vmov.i32    q4, #0\n"
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]!\n"
-                "vmov.i32    q5, #0\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-                "vmov.i32    q6, #0\n"
+                "vmov.i32	q4, #0\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]!\n"
+                "vmov.i32	q5, #0\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+                "vmov.i32	q6, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #48]")
-                "vmov.i32    q7, #0\n"
+                "vmov.i32	q7, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #48]")
-                "vmov.i32    q8, #0\n"
+                "vmov.i32	q8, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #112]")
-                "vmov.i32    q9, #0\n"
+                "vmov.i32	q9, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #112]")
-                "vmov.i32    q10, #0\n"
-                "vmov.i32    q11, #0\n"
-                "vmov.i32    q12, #0\n"
-                "vmov.i32    q13, #0\n"
+                "vmov.i32	q10, #0\n"
+                "vmov.i32	q11, #0\n"
+                "vmov.i32	q12, #0\n"
+                "vmov.i32	q13, #0\n"
                 ASM_PREFETCH("[%[a_ptr], #176]")
-                "vmov.i32    q14, #0\n"
+                "vmov.i32	q14, #0\n"
                 ASM_PREFETCH("[%[b_ptr], #176]")
-                "vmov.i32    q15, #0\n"
+                "vmov.i32	q15, #0\n"
 
-                "cmp        %[k], #0\n"
-                "beq        6f\n"
+                "cmp		%[k], #0\n"
+                "beq		6f\n"
 
                 "1:\n"
                 // Unroll 0
-                "vmla.f32    q4, q2, d0[0]\n"
-                "vld1.32    {d2-d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d1[1]\n"
-                "vmla.f32    q8, q2, d2[0]\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d0[0]\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-                "vmla.f32    q13, q3, d1[1]\n"
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-                "vmla.f32    q15, q3, d2[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "vld1.32	{d2-d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d1[1]\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d0[0]\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+                "vmla.f32	q13, q3, d1[1]\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+                "vmla.f32	q15, q3, d2[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
 
                 // Unroll 1
-                "vmla.f32    q4, q2, d3[0]\n"
-                "subs        %[k], %[k], #1\n"
-                "vmla.f32    q5, q2, d3[1]\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "subs		%[k], %[k], #1\n"
+                "vmla.f32	q5, q2, d3[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #208]")
-                "vmla.f32    q6, q2, d0[0]\n"
-                "vmla.f32    q7, q2, d0[1]\n"
+                "vmla.f32	q6, q2, d0[0]\n"
+                "vmla.f32	q7, q2, d0[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "vmla.f32    q8, q2, d1[0]\n"
-                "vmla.f32    q9, q2, d1[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d3[0]\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "vld1.32    {d2-d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q12, q3, d0[0]\n"
-                "vmla.f32    q13, q3, d0[1]\n"
-                "vmla.f32    q14, q3, d1[0]\n"
-                "vmla.f32    q15, q3, d1[1]\n"
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q8, q2, d1[0]\n"
+                "vmla.f32	q9, q2, d1[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d3[0]\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "vld1.32	{d2-d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q12, q3, d0[0]\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "vmla.f32	q14, q3, d1[0]\n"
+                "vmla.f32	q15, q3, d1[1]\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]!\n"
 
                 // Unroll 2
-                "vmla.f32    q4, q2, d2[0]\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d3[0]\n"
-                "vmla.f32    q7, q2, d3[1]\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d3[0]\n"
+                "vmla.f32	q7, q2, d3[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #240]")
-                "vmla.f32    q8, q2, d0[0]\n"
-                "vmla.f32    q9, q2, d0[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+                "vmla.f32	q9, q2, d0[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
 
-                "vmla.f32    q10, q3, d2[0]\n"
-                "vmla.f32    q11, q3, d2[1]\n"
+                "vmla.f32	q10, q3, d2[0]\n"
+                "vmla.f32	q11, q3, d2[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #208]")
-                "vmla.f32    q12, q3, d3[0]\n"
-                "vmla.f32    q13, q3, d3[1]\n"
-                "vld1.32    {d2-d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "vmla.f32    q15, q3, d0[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+                "vmla.f32	q13, q3, d3[1]\n"
+                "vld1.32	{d2-d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "vmla.f32	q15, q3, d0[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
 
                 // Unroll 3
-                "vmla.f32    q4, q2, d1[0]\n"
-                "vmla.f32    q5, q2, d1[1]\n"
-                "vmla.f32    q6, q2, d2[0]\n"
-                "vmla.f32    q7, q2, d2[1]\n"
-                "vmla.f32    q8, q2, d3[0]\n"
-                "vmla.f32    q9, q2, d3[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d1[0]\n"
-                "vmla.f32    q11, q3, d1[1]\n"
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q12, q3, d2[0]\n"
-                "vmla.f32    q13, q3, d2[1]\n"
-                "vmla.f32    q14, q3, d3[0]\n"
-                "vmla.f32    q15, q3, d3[1]\n"
-                "bne        1b\n"
+                "vmla.f32	q4, q2, d1[0]\n"
+                "vmla.f32	q5, q2, d1[1]\n"
+                "vmla.f32	q6, q2, d2[0]\n"
+                "vmla.f32	q7, q2, d2[1]\n"
+                "vmla.f32	q8, q2, d3[0]\n"
+                "vmla.f32	q9, q2, d3[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d1[0]\n"
+                "vmla.f32	q11, q3, d1[1]\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q12, q3, d2[0]\n"
+                "vmla.f32	q13, q3, d2[1]\n"
+                "vmla.f32	q14, q3, d3[0]\n"
+                "vmla.f32	q15, q3, d3[1]\n"
+                "bne		1b\n"
 
                 // Branch here if we never execute main loop.
                 "6:\n"
@@ -167,182 +167,182 @@ void a32_sgemm_8x6(const float *Apanel, const float *Bpanel, float *Cpanel, int
                 // "Tails" shows how many multiply blocks are needed at the
                 // end, must be 1-4 inclusive.  Bail out to alternative tail
                 // immediately if it's 1.
-                "subs        %[tails], %[tails], #1\n"
-                "beq        3f\n"
+                "subs		%[tails], %[tails], #1\n"
+                "beq		3f\n"
 
                 // Detached final iteration
                 // Unroll 0
-                "vmla.f32    q4, q2, d0[0]\n"
-                "vld1.32    {d2-d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d1[1]\n"
-                "vmla.f32    q8, q2, d2[0]\n"
-                "subs        %[tails], %[tails], #1\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d0[0]\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-                "vmla.f32    q13, q3, d1[1]\n"
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-                "vmla.f32    q15, q3, d2[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "beq        4f\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "vld1.32	{d2-d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d1[1]\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "subs		%[tails], %[tails], #1\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d0[0]\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+                "vmla.f32	q13, q3, d1[1]\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+                "vmla.f32	q15, q3, d2[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "beq		4f\n"
 
                 // Unroll 1
-                "vmla.f32    q4, q2, d3[0]\n"
-                "vmla.f32    q5, q2, d3[1]\n"
-                "subs        %[tails], %[tails], #1\n"
-                "vmla.f32    q6, q2, d0[0]\n"
-                "vmla.f32    q7, q2, d0[1]\n"
-                "vmla.f32    q8, q2, d1[0]\n"
-                "vmla.f32    q9, q2, d1[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d3[0]\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "vld1.32    {d2-d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q12, q3, d0[0]\n"
-                "vmla.f32    q13, q3, d0[1]\n"
-                "vmla.f32    q14, q3, d1[0]\n"
-                "vmla.f32    q15, q3, d1[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "beq        5f\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "vmla.f32	q5, q2, d3[1]\n"
+                "subs		%[tails], %[tails], #1\n"
+                "vmla.f32	q6, q2, d0[0]\n"
+                "vmla.f32	q7, q2, d0[1]\n"
+                "vmla.f32	q8, q2, d1[0]\n"
+                "vmla.f32	q9, q2, d1[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d3[0]\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "vld1.32	{d2-d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q12, q3, d0[0]\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "vmla.f32	q14, q3, d1[0]\n"
+                "vmla.f32	q15, q3, d1[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "beq		5f\n"
 
                 // Unroll 2
-                "vld1.32    {d0-d1}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q4, q2, d2[0]\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "vmla.f32    q6, q2, d3[0]\n"
-                "vmla.f32    q7, q2, d3[1]\n"
-                "vmla.f32    q8, q2, d0[0]\n"
-                "vmla.f32    q9, q2, d0[1]\n"
-                "vld1.32    {d4-d5}, [%[b_ptr] :128]!\n"
-
-                "vmla.f32    q10, q3, d2[0]\n"
-                "vmla.f32    q11, q3, d2[1]\n"
-                "vmla.f32    q12, q3, d3[0]\n"
-                "vmla.f32    q13, q3, d3[1]\n"
-                "vld1.32    {d2-d3}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "vmla.f32    q15, q3, d0[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
+                "vld1.32	{d0-d1}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "vmla.f32	q6, q2, d3[0]\n"
+                "vmla.f32	q7, q2, d3[1]\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+                "vmla.f32	q9, q2, d0[1]\n"
+                "vld1.32	{d4-d5}, [%[b_ptr] :128]!\n"
+
+                "vmla.f32	q10, q3, d2[0]\n"
+                "vmla.f32	q11, q3, d2[1]\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+                "vmla.f32	q13, q3, d3[1]\n"
+                "vld1.32	{d2-d3}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "vmla.f32	q15, q3, d0[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
 
                 // Unroll 3
-                "vmla.f32    q4, q2, d1[0]\n"
-                "vmla.f32    q10, q3, d1[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q5, q2, d1[1]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d1[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d2[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d2[0]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d2[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d2[1]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d3[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d3[0]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d3[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d3[1]\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "b        2f\n"
+                "vmla.f32	q4, q2, d1[0]\n"
+                "vmla.f32	q10, q3, d1[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q5, q2, d1[1]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d1[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d2[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d2[0]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d2[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d2[1]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d3[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d3[0]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d3[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d3[1]\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "b		2f\n"
 
                 // tails==1 final tail
                 "3:\n"
-                "vmla.f32    q4, q2, d0[0]\n"
-                "vld1.32    {d2}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q5, q2, d0[1]\n"
-                "vld1.32    {d6-d7}, [%[b_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d1[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q10, q3, d0[0]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d0[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d1[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d1[1]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d1[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d2[0]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d2[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d2[1]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d2[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "b        2f\n"
+                "vmla.f32	q4, q2, d0[0]\n"
+                "vld1.32	{d2}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q5, q2, d0[1]\n"
+                "vld1.32	{d6-d7}, [%[b_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d1[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q10, q3, d0[0]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d0[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d1[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d1[1]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d1[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d2[0]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d2[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d2[1]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d2[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "b		2f\n"
 
                 // tails==2 final tail
                 "4:\n"
-                "vmla.f32    q4, q2, d3[0]\n"
-                "vmla.f32    q10, q3, d3[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q5, q2, d3[1]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d3[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q6, q2, d0[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d0[0]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d0[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d0[1]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d1[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d1[0]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d1[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d1[1]\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
-                "b        2f\n"
+                "vmla.f32	q4, q2, d3[0]\n"
+                "vmla.f32	q10, q3, d3[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q5, q2, d3[1]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d3[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q6, q2, d0[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d0[0]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d0[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d0[1]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d1[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d1[0]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d1[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d1[1]\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
+                "b		2f\n"
 
                 // tails==3 final tail
                 "5:\n"
-                "vmla.f32    q4, q2, d2[0]\n"
-                "vld1.32    {d0}, [%[a_ptr] :64]!\n"
-                "vmla.f32    q5, q2, d2[1]\n"
-                "vmla.f32    q6, q2, d3[0]\n"
-                "vst1.32    {d8-d9}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q10, q3, d2[0]\n"
-                "vst1.32    {d20-d21}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q11, q3, d2[1]\n"
-                "vst1.32    {d10-d11}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q12, q3, d3[0]\n"
-                "vst1.32    {d22-d23}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q7, q2, d3[1]\n"
-                "vst1.32    {d12-d13}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q13, q3, d3[1]\n"
-                "vst1.32    {d24-d25}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q8, q2, d0[0]\n"
-                "vst1.32    {d14-d15}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q14, q3, d0[0]\n"
-                "vst1.32    {d26-d27}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q9, q2, d0[1]\n"
-                "vst1.32    {d16-d17}, [%[c_ptr] :128]!\n"
-                "vmla.f32    q15, q3, d0[1]\n"
-                "vst1.32    {d28-d29}, [%[c_ptr] :128]!\n"
-                "vst1.32    {d18-d19}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q4, q2, d2[0]\n"
+                "vld1.32	{d0}, [%[a_ptr] :64]!\n"
+                "vmla.f32	q5, q2, d2[1]\n"
+                "vmla.f32	q6, q2, d3[0]\n"
+                "vst1.32	{d8-d9}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q10, q3, d2[0]\n"
+                "vst1.32	{d20-d21}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q11, q3, d2[1]\n"
+                "vst1.32	{d10-d11}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q12, q3, d3[0]\n"
+                "vst1.32	{d22-d23}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q7, q2, d3[1]\n"
+                "vst1.32	{d12-d13}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q13, q3, d3[1]\n"
+                "vst1.32	{d24-d25}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q8, q2, d0[0]\n"
+                "vst1.32	{d14-d15}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q14, q3, d0[0]\n"
+                "vst1.32	{d26-d27}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q9, q2, d0[1]\n"
+                "vst1.32	{d16-d17}, [%[c_ptr] :128]!\n"
+                "vmla.f32	q15, q3, d0[1]\n"
+                "vst1.32	{d28-d29}, [%[c_ptr] :128]!\n"
+                "vst1.32	{d18-d19}, [%[c_ptr] :128]!\n"
 
                 "2:\n"
-                "vst1.32    {d30-d31}, [%[c_ptr] :128]!\n"
+                "vst1.32	{d30-d31}, [%[c_ptr] :128]!\n"
             : [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr), [k] "+r" (k), [tails] "+r" (tails)
             :
             : "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15",
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16.hpp
index fe939b1084..8d4146ab3a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16.hpp
@@ -82,7 +82,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp
index 022a34fdcd..d713ca5d53 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp
@@ -50,19 +50,18 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -83,7 +82,6 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -108,19 +106,19 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "bgt 77f\n"
       "beq 39f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 3f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -133,15 +131,15 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cbz x15, 4f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x15, x15, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "b 16f\n"
@@ -232,8 +230,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "17:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 18f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -253,11 +251,7 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q6, [x12, #0x10]\n"
       "blt 21f\n"
       "20:"  // Height 1: Multiply loop: Main loop head
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x10\n"
-      "add x26, x26, #0x10\n"
       "trn1 v20.2d, v1.2d, v21.2d\n"
-      "trn2 v1.2d, v1.2d, v21.2d\n"
       ".inst 0x6e47ee88  // bfmmla v8.4s, v20.8h, v7.8h\n"
       "ldr q17, [x11, #0x0]\n"
       ".inst 0x6e46ee8c  // bfmmla v12.4s, v20.8h, v6.8h\n"
@@ -270,37 +264,38 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x9, #0x0]\n"
       ".inst 0x6e51ee8e  // bfmmla v14.4s, v20.8h, v17.8h\n"
       "ldr q17, [x9, #0x10]\n"
+      "trn2 v1.2d, v1.2d, v21.2d\n"
       ".inst 0x6e52ee8b  // bfmmla v11.4s, v20.8h, v18.8h\n"
       "ldr q18, [x12, #0x20]\n"
       ".inst 0x6e51ee8f  // bfmmla v15.4s, v20.8h, v17.8h\n"
       "ldr q17, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e52ec28  // bfmmla v8.4s, v1.8h, v18.8h\n"
       "ldr q18, [x11, #0x20]\n"
       ".inst 0x6e51ec2c  // bfmmla v12.4s, v1.8h, v17.8h\n"
       "ldr q17, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e52ec29  // bfmmla v9.4s, v1.8h, v18.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e51ec2d  // bfmmla v13.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e52ec2a  // bfmmla v10.4s, v1.8h, v18.8h\n"
       "ldr q18, [x9, #0x20]\n"
       ".inst 0x6e51ec2e  // bfmmla v14.4s, v1.8h, v17.8h\n"
       "ldr q17, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e52ec2b  // bfmmla v11.4s, v1.8h, v18.8h\n"
-      "ldr q7, [x12, #0x0]\n"
       ".inst 0x6e51ec2f  // bfmmla v15.4s, v1.8h, v17.8h\n"
       "ldr q1, [x26, #0x0]\n"
+      "add x12, x12, #0x40\n"
+      "ldr q7, [x12, #0x0]\n"
+      "add x11, x11, #0x40\n"
       "ldr q6, [x12, #0x10]\n"
+      "add x10, x10, #0x40\n"
+      "add x9, x9, #0x40\n"
       "bge 20b\n"
       "21:"  // Height 1: Multiply loop: Single iteration only
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "trn1 v19.2d, v1.2d, v17.2d\n"
-      "trn2 v1.2d, v1.2d, v17.2d\n"
+      "trn1 v19.2d, v1.2d, v20.2d\n"
       ".inst 0x6e47ee68  // bfmmla v8.4s, v19.8h, v7.8h\n"
       "ldr q17, [x11, #0x0]\n"
       ".inst 0x6e46ee6c  // bfmmla v12.4s, v19.8h, v6.8h\n"
@@ -312,58 +307,61 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e51ee6a  // bfmmla v10.4s, v19.8h, v17.8h\n"
       "ldr q17, [x9, #0x0]\n"
       ".inst 0x6e52ee6e  // bfmmla v14.4s, v19.8h, v18.8h\n"
-      "ldr q25, [x9, #0x10]\n"
+      "ldr q24, [x9, #0x10]\n"
+      "trn2 v1.2d, v1.2d, v20.2d\n"
       ".inst 0x6e51ee6b  // bfmmla v11.4s, v19.8h, v17.8h\n"
-      "ldr q17, [x12, #0x20]\n"
-      ".inst 0x6e59ee6f  // bfmmla v15.4s, v19.8h, v25.8h\n"
-      "ldr q3, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
-      ".inst 0x6e51ec28  // bfmmla v8.4s, v1.8h, v17.8h\n"
+      "ldr q18, [x12, #0x20]\n"
+      ".inst 0x6e58ee6f  // bfmmla v15.4s, v19.8h, v24.8h\n"
+      "ldr q17, [x12, #0x30]\n"
+      ".inst 0x6e52ec28  // bfmmla v8.4s, v1.8h, v18.8h\n"
       "ldr q19, [x11, #0x20]\n"
-      ".inst 0x6e43ec2c  // bfmmla v12.4s, v1.8h, v3.8h\n"
+      ".inst 0x6e51ec2c  // bfmmla v12.4s, v1.8h, v17.8h\n"
       "ldr q17, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e53ec29  // bfmmla v9.4s, v1.8h, v19.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e51ec2d  // bfmmla v13.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e52ec2a  // bfmmla v10.4s, v1.8h, v18.8h\n"
       "ldr q18, [x9, #0x20]\n"
       ".inst 0x6e51ec2e  // bfmmla v14.4s, v1.8h, v17.8h\n"
       "ldr q17, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e52ec2b  // bfmmla v11.4s, v1.8h, v18.8h\n"
       ".inst 0x6e51ec2f  // bfmmla v15.4s, v1.8h, v17.8h\n"
+      "add x26, x26, #0x10\n"
+      "add x12, x12, #0x40\n"
+      "add x11, x11, #0x40\n"
+      "add x10, x10, #0x40\n"
+      "add x9, x9, #0x40\n"
       "22:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 27f\n"
       "cmp x27, #0x4\n"
       "blt 24f\n"
       "23:"  // Height 1: Multiply loop: Odd block loop
       "ldr d19, [x26], #0x8\n"
-      "ldr q20, [x12, #0x0]\n"
-      "sub x27, x27, #0x4\n"
+      "ldr q18, [x12, #0x0]\n"
+      "trn1 v19.2d, v19.2d, v17.2d\n"
       "ldr q17, [x12, #0x10]\n"
-      "cmp x27, #0x4\n"
-      "add x12, x12, #0x20\n"
-      "trn1 v19.2d, v19.2d, v18.2d\n"
-      ".inst 0x6e54ee68  // bfmmla v8.4s, v19.8h, v20.8h\n"
+      ".inst 0x6e52ee68  // bfmmla v8.4s, v19.8h, v18.8h\n"
       "ldr q18, [x11, #0x0]\n"
       ".inst 0x6e51ee6c  // bfmmla v12.4s, v19.8h, v17.8h\n"
       "ldr q17, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e52ee69  // bfmmla v9.4s, v19.8h, v18.8h\n"
       "ldr q18, [x10, #0x0]\n"
       ".inst 0x6e51ee6d  // bfmmla v13.4s, v19.8h, v17.8h\n"
       "ldr q17, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e52ee6a  // bfmmla v10.4s, v19.8h, v18.8h\n"
       "ldr q18, [x9, #0x0]\n"
       ".inst 0x6e51ee6e  // bfmmla v14.4s, v19.8h, v17.8h\n"
       "ldr q17, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
+      "add x12, x12, #0x20\n"
+      "add x11, x11, #0x20\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
       "bge 23b\n"
       "24:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 27f\n"
@@ -377,25 +375,25 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "26:"  // Height 1: Multiply loop: Ragged operand read: Done
       "ldr q20, [x12, #0x0]\n"
       "ldr q18, [x12, #0x10]\n"
-      "add x12, x12, #0x20\n"
       "trn1 v19.2d, v1.2d, v17.2d\n"
       ".inst 0x6e54ee68  // bfmmla v8.4s, v19.8h, v20.8h\n"
       "ldr q17, [x11, #0x0]\n"
       ".inst 0x6e52ee6c  // bfmmla v12.4s, v19.8h, v18.8h\n"
       "ldr q18, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e51ee69  // bfmmla v9.4s, v19.8h, v17.8h\n"
       "ldr q17, [x10, #0x0]\n"
       ".inst 0x6e52ee6d  // bfmmla v13.4s, v19.8h, v18.8h\n"
-      "ldr q2, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
+      "ldr q6, [x10, #0x10]\n"
       ".inst 0x6e51ee6a  // bfmmla v10.4s, v19.8h, v17.8h\n"
       "ldr q18, [x9, #0x0]\n"
-      ".inst 0x6e42ee6e  // bfmmla v14.4s, v19.8h, v2.8h\n"
+      ".inst 0x6e46ee6e  // bfmmla v14.4s, v19.8h, v6.8h\n"
       "ldr q17, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
+      "add x12, x12, #0x20\n"
+      "add x11, x11, #0x20\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
       "27:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -406,9 +404,9 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "uzp1 v10.2d, v10.2d, v14.2d\n"
       "uzp1 v11.2d, v11.2d, v15.2d\n"
       "tbz %x[flags], #1, 28f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v18.4s }, [x21]\n"
       "ld1r { v17.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v18.4s\n"
       "fmin v9.4s, v9.4s, v18.4s\n"
@@ -479,19 +477,19 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "b 230f\n"
       "39:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "40:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 41f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -504,15 +502,15 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cbz x15, 42f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x15, x15, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "b 54f\n"
@@ -520,75 +518,75 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "tbz %x[flags], #0, 53f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "bge 51f\n"
       "tbz x14, #3, 46f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
       "tbz x14, #2, 44f\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
       "tbz x14, #1, 43f\n"
       "ldr d16, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x14, #0, 50f\n"
       "ld1 { v16.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v15.s }[2], [x25]\n"
       "b 50f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 50f\n"
       "ldr s16, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
       "b 50f\n"
       "44:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x14, #1, 45f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x14, #0, 50f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v14.s }[2], [x25]\n"
       "b 50f\n"
       "45:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 50f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
       "b 50f\n"
       "46:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x14, #2, 48f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "tbz x14, #1, 47f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x14, #0, 50f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v13.s }[2], [x25]\n"
       "b 50f\n"
       "47:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 50f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
       "b 50f\n"
       "48:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x14, #1, 49f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x14, #0, 50f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v12.s }[2], [x25]\n"
       "b 50f\n"
       "49:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s9, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
       "50:"  // Height 2: Partial accumulate: Done
       "sub x13, x13, x20\n"
@@ -598,10 +596,10 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x13, #0x10]\n"
       "ldr q11, [x13, #0x20]\n"
       "ldr q16, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "52:"  // Height 2: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -625,8 +623,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "55:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 56f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -651,12 +649,6 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "blt 59f\n"
       "58:"  // Height 2: Multiply loop: Main loop head
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "ldr q2, [x25, #0x0]\n"
       ".inst 0x6e47ee68  // bfmmla v8.4s, v19.8h, v7.8h\n"
       "ldr q18, [x11, #0x0]\n"
       ".inst 0x6e46ee6c  // bfmmla v12.4s, v19.8h, v6.8h\n"
@@ -669,38 +661,40 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x9, #0x0]\n"
       ".inst 0x6e51ee6e  // bfmmla v14.4s, v19.8h, v17.8h\n"
       "ldr q17, [x9, #0x10]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       "ldr q18, [x12, #0x20]\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
       "ldr q17, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e52ec28  // bfmmla v8.4s, v1.8h, v18.8h\n"
       "ldr q18, [x11, #0x20]\n"
       ".inst 0x6e51ec2c  // bfmmla v12.4s, v1.8h, v17.8h\n"
       "ldr q17, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e52ec29  // bfmmla v9.4s, v1.8h, v18.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e51ec2d  // bfmmla v13.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e52ec2a  // bfmmla v10.4s, v1.8h, v18.8h\n"
       "ldr q18, [x9, #0x20]\n"
       ".inst 0x6e51ec2e  // bfmmla v14.4s, v1.8h, v17.8h\n"
       "ldr q17, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "ldr q2, [x25, #0x0]\n"
       ".inst 0x6e52ec2b  // bfmmla v11.4s, v1.8h, v18.8h\n"
+      "add x12, x12, #0x40\n"
       "ldr q7, [x12, #0x0]\n"
       ".inst 0x6e51ec2f  // bfmmla v15.4s, v1.8h, v17.8h\n"
       "ldr q1, [x26, #0x0]\n"
       "ldr q6, [x12, #0x10]\n"
+      "add x11, x11, #0x40\n"
+      "add x10, x10, #0x40\n"
+      "add x9, x9, #0x40\n"
       "bge 58b\n"
       "59:"  // Height 2: Multiply loop: Single iteration only
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       ".inst 0x6e47ee68  // bfmmla v8.4s, v19.8h, v7.8h\n"
       "ldr q18, [x11, #0x0]\n"
       ".inst 0x6e46ee6c  // bfmmla v12.4s, v19.8h, v6.8h\n"
@@ -713,58 +707,62 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x9, #0x0]\n"
       ".inst 0x6e51ee6e  // bfmmla v14.4s, v19.8h, v17.8h\n"
       "ldr q17, [x9, #0x10]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       "ldr q18, [x12, #0x20]\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
       "ldr q17, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e52ec28  // bfmmla v8.4s, v1.8h, v18.8h\n"
       "ldr q18, [x11, #0x20]\n"
       ".inst 0x6e51ec2c  // bfmmla v12.4s, v1.8h, v17.8h\n"
       "ldr q17, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e52ec29  // bfmmla v9.4s, v1.8h, v18.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e51ec2d  // bfmmla v13.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e52ec2a  // bfmmla v10.4s, v1.8h, v18.8h\n"
       "ldr q18, [x9, #0x20]\n"
       ".inst 0x6e51ec2e  // bfmmla v14.4s, v1.8h, v17.8h\n"
       "ldr q17, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e52ec2b  // bfmmla v11.4s, v1.8h, v18.8h\n"
       ".inst 0x6e51ec2f  // bfmmla v15.4s, v1.8h, v17.8h\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x12, x12, #0x40\n"
+      "add x11, x11, #0x40\n"
+      "add x10, x10, #0x40\n"
+      "add x9, x9, #0x40\n"
       "60:"  // Height 2: Multiply loop: Main loop skip
       "cbz x27, 65f\n"
       "cmp x27, #0x4\n"
       "blt 62f\n"
       "61:"  // Height 2: Multiply loop: Odd block loop
-      "ldr d20, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d18, [x26], #0x8\n"
+      "ldr d17, [x25], #0x8\n"
+      "trn1 v19.2d, v18.2d, v17.2d\n"
       "sub x27, x27, #0x4\n"
       "ldr q18, [x12, #0x0]\n"
       "ldr q17, [x12, #0x10]\n"
-      "cmp x27, #0x4\n"
-      "add x12, x12, #0x20\n"
-      "trn1 v19.2d, v20.2d, v19.2d\n"
       ".inst 0x6e52ee68  // bfmmla v8.4s, v19.8h, v18.8h\n"
-      "ldr q26, [x11, #0x0]\n"
       ".inst 0x6e51ee6c  // bfmmla v12.4s, v19.8h, v17.8h\n"
+      "ldr q26, [x11, #0x0]\n"
       "ldr q6, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e5aee69  // bfmmla v9.4s, v19.8h, v26.8h\n"
-      "ldr q18, [x10, #0x0]\n"
       ".inst 0x6e46ee6d  // bfmmla v13.4s, v19.8h, v6.8h\n"
+      "ldr q18, [x10, #0x0]\n"
       "ldr q17, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e52ee6a  // bfmmla v10.4s, v19.8h, v18.8h\n"
-      "ldr q18, [x9, #0x0]\n"
       ".inst 0x6e51ee6e  // bfmmla v14.4s, v19.8h, v17.8h\n"
+      "ldr q18, [x9, #0x0]\n"
       "ldr q17, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
+      "add x12, x12, #0x20\n"
+      "add x11, x11, #0x20\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
       "bge 61b\n"
       "62:"  // Height 2: Multiply loop: Skip odd blocks
       "cbz x27, 65f\n"
@@ -782,24 +780,24 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x12, #0x0]\n"
       "ldr q17, [x12, #0x10]\n"
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "add x12, x12, #0x20\n"
       ".inst 0x6e52ee68  // bfmmla v8.4s, v19.8h, v18.8h\n"
       "ldr q18, [x11, #0x0]\n"
       ".inst 0x6e51ee6c  // bfmmla v12.4s, v19.8h, v17.8h\n"
       "ldr q17, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e52ee69  // bfmmla v9.4s, v19.8h, v18.8h\n"
-      "ldr q30, [x10, #0x0]\n"
+      "ldr q3, [x10, #0x0]\n"
       ".inst 0x6e51ee6d  // bfmmla v13.4s, v19.8h, v17.8h\n"
-      "ldr q31, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
-      ".inst 0x6e5eee6a  // bfmmla v10.4s, v19.8h, v30.8h\n"
+      "ldr q27, [x10, #0x10]\n"
+      ".inst 0x6e43ee6a  // bfmmla v10.4s, v19.8h, v3.8h\n"
       "ldr q18, [x9, #0x0]\n"
-      ".inst 0x6e5fee6e  // bfmmla v14.4s, v19.8h, v31.8h\n"
+      ".inst 0x6e5bee6e  // bfmmla v14.4s, v19.8h, v27.8h\n"
       "ldr q17, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
+      "add x12, x12, #0x20\n"
+      "add x11, x11, #0x20\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
       "65:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -808,17 +806,17 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
+      "add x25, x13, x20, LSL #2\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x26, x13, x20, LSL #2\n"
       "tbz %x[flags], #1, 66f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v18.4s }, [x21]\n"
       "ld1r { v17.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v18.4s\n"
       "fmin v12.4s, v12.4s, v18.4s\n"
@@ -842,63 +840,63 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "tbz x14, #3, 70f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
       "st1 { v12.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
       "tbz x14, #2, 68f\n"
       "st1 { v13.4s }, [x13], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
       "tbz x14, #1, 67f\n"
       "str d14, [x13], #0x8\n"
-      "str d11, [x26], #0x8\n"
+      "str d11, [x25], #0x8\n"
       "tbz x14, #0, 74f\n"
       "st1 { v14.s }[2], [x13]\n"
-      "st1 { v11.s }[2], [x26]\n"
+      "st1 { v11.s }[2], [x25]\n"
       "b 74f\n"
       "67:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x14, #0, 74f\n"
       "str s14, [x13, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
       "b 74f\n"
       "68:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x14, #1, 69f\n"
       "str d13, [x13], #0x8\n"
-      "str d10, [x26], #0x8\n"
+      "str d10, [x25], #0x8\n"
       "tbz x14, #0, 74f\n"
       "st1 { v13.s }[2], [x13]\n"
-      "st1 { v10.s }[2], [x26]\n"
+      "st1 { v10.s }[2], [x25]\n"
       "b 74f\n"
       "69:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x14, #0, 74f\n"
       "str s13, [x13, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
       "b 74f\n"
       "70:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x14, #2, 72f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
       "tbz x14, #1, 71f\n"
       "str d12, [x13], #0x8\n"
-      "str d9, [x26], #0x8\n"
+      "str d9, [x25], #0x8\n"
       "tbz x14, #0, 74f\n"
       "st1 { v12.s }[2], [x13]\n"
-      "st1 { v9.s }[2], [x26]\n"
+      "st1 { v9.s }[2], [x25]\n"
       "b 74f\n"
       "71:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x14, #0, 74f\n"
       "str s12, [x13, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
       "b 74f\n"
       "72:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x14, #1, 73f\n"
       "str d7, [x13], #0x8\n"
-      "str d8, [x26], #0x8\n"
+      "str d8, [x25], #0x8\n"
       "tbz x14, #0, 74f\n"
       "st1 { v7.s }[2], [x13]\n"
-      "st1 { v8.s }[2], [x26]\n"
+      "st1 { v8.s }[2], [x25]\n"
       "b 74f\n"
       "73:"  // Height 2: Partial direct writeback: partial_1_0
       "str s7, [x13, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
       "74:"  // Height 2: Partial direct writeback: Done
       "b 76f\n"
       "75:"  // Height 2: Full writeback
@@ -907,29 +905,29 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "str q13, [x13, #0x20]\n"
       "str q14, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
       "76:"  // Height 2: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 40b\n"
       "b 230f\n"
       "77:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "78:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 79f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -942,15 +940,15 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cbz x15, 80f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x15, x15, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -965,94 +963,94 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "80:"  // Height 3: no bias
       "tbz %x[flags], #0, 91f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
       "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "bge 89f\n"
       "tbz x14, #3, 84f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
       "tbz x14, #2, 82f\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
       "tbz x14, #1, 81f\n"
       "ldr d16, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
       "tbz x14, #0, 88f\n"
       "ld1 { v16.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
       "b 88f\n"
       "81:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 88f\n"
       "ldr s16, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
       "b 88f\n"
       "82:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x14, #1, 83f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x14, #0, 88f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
       "b 88f\n"
       "83:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 88f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
       "b 88f\n"
       "84:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x14, #2, 86f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "tbz x14, #1, 85f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x14, #0, 88f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
       "b 88f\n"
       "85:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 88f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
       "b 88f\n"
       "86:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x14, #1, 87f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x14, #0, 88f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
       "b 88f\n"
       "87:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s9, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
       "88:"  // Height 3: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 90f\n"
@@ -1061,14 +1059,14 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x13, #0x10]\n"
       "ldr q11, [x13, #0x20]\n"
       "ldr q16, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
       "90:"  // Height 3: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -1108,8 +1106,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "93:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 94f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1139,54 +1137,54 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "96:"  // Height 3: Multiply loop: Main loop head
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
-      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
+      "trn1 v27.2d, v3.2d, v29.2d\n"
       ".inst 0x6e47ef70  // bfmmla v16.4s, v27.8h, v7.8h\n"
       "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e46ef74  // bfmmla v20.4s, v27.8h, v6.8h\n"
       "ldr q25, [x11, #0x10]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "trn2 v3.2d, v3.2d, v29.2d\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x0]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x10]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x9, #0x0]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x9, #0x10]\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       "ldr q26, [x12, #0x20]\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
       "ldr q25, [x12, #0x30]\n"
       "ldr q2, [x25, #0x0]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e5aec28  // bfmmla v8.4s, v1.8h, v26.8h\n"
       ".inst 0x6e5aec70  // bfmmla v16.4s, v3.8h, v26.8h\n"
       "ldr q26, [x11, #0x20]\n"
       ".inst 0x6e59ec2c  // bfmmla v12.4s, v1.8h, v25.8h\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6e59ec74  // bfmmla v20.4s, v3.8h, v25.8h\n"
       "ldr q25, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e5aec29  // bfmmla v9.4s, v1.8h, v26.8h\n"
+      "add x11, x11, #0x40\n"
       ".inst 0x6e5aec71  // bfmmla v17.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e59ec2d  // bfmmla v13.4s, v1.8h, v25.8h\n"
       ".inst 0x6e59ec75  // bfmmla v21.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e5aec2a  // bfmmla v10.4s, v1.8h, v26.8h\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6e5aec72  // bfmmla v18.4s, v3.8h, v26.8h\n"
       "ldr q26, [x9, #0x20]\n"
       ".inst 0x6e59ec2e  // bfmmla v14.4s, v1.8h, v25.8h\n"
@@ -1205,52 +1203,52 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "97:"  // Height 3: Multiply loop: Single iteration only
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
-      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
+      "trn1 v27.2d, v3.2d, v29.2d\n"
       ".inst 0x6e47ef70  // bfmmla v16.4s, v27.8h, v7.8h\n"
       "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e46ef74  // bfmmla v20.4s, v27.8h, v6.8h\n"
       "ldr q25, [x11, #0x10]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "trn2 v3.2d, v3.2d, v29.2d\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x0]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x10]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x9, #0x0]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x9, #0x10]\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       "ldr q26, [x12, #0x20]\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
       "ldr q25, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e5aec28  // bfmmla v8.4s, v1.8h, v26.8h\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6e5aec70  // bfmmla v16.4s, v3.8h, v26.8h\n"
       "ldr q26, [x11, #0x20]\n"
       ".inst 0x6e59ec2c  // bfmmla v12.4s, v1.8h, v25.8h\n"
       ".inst 0x6e59ec74  // bfmmla v20.4s, v3.8h, v25.8h\n"
       "ldr q25, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e5aec29  // bfmmla v9.4s, v1.8h, v26.8h\n"
+      "add x11, x11, #0x40\n"
       ".inst 0x6e5aec71  // bfmmla v17.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e59ec2d  // bfmmla v13.4s, v1.8h, v25.8h\n"
       ".inst 0x6e59ec75  // bfmmla v21.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e5aec2a  // bfmmla v10.4s, v1.8h, v26.8h\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6e5aec72  // bfmmla v18.4s, v3.8h, v26.8h\n"
       "ldr q26, [x9, #0x20]\n"
       ".inst 0x6e59ec2e  // bfmmla v14.4s, v1.8h, v25.8h\n"
@@ -1266,40 +1264,40 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cmp x27, #0x4\n"
       "blt 100f\n"
       "99:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "sub x27, x27, #0x4\n"
-      "ldr d27, [x24], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "trn1 v28.2d, v26.2d, v25.2d\n"
+      "ldr d25, [x24], #0x8\n"
       "ldr q26, [x12, #0x0]\n"
-      "cmp x27, #0x4\n"
-      "ldr q25, [x12, #0x10]\n"
-      "add x12, x12, #0x20\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v27.2d, v29.2d\n"
+      "trn1 v27.2d, v25.2d, v27.2d\n"
       ".inst 0x6e5aef88  // bfmmla v8.4s, v28.8h, v26.8h\n"
+      "ldr q25, [x12, #0x10]\n"
       ".inst 0x6e5aef70  // bfmmla v16.4s, v27.8h, v26.8h\n"
       "ldr q26, [x11, #0x0]\n"
       ".inst 0x6e59ef8c  // bfmmla v12.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef74  // bfmmla v20.4s, v27.8h, v25.8h\n"
       "ldr q25, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "sub x27, x27, #0x4\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x0]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x9, #0x0]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "add x10, x10, #0x20\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
+      "add x9, x9, #0x20\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
       "bge 99b\n"
       "100:"  // Height 3: Multiply loop: Skip odd blocks
@@ -1321,18 +1319,18 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q26, [x12, #0x0]\n"
       "ldr q29, [x12, #0x10]\n"
       "trn1 v28.2d, v1.2d, v2.2d\n"
-      "add x12, x12, #0x20\n"
       "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e5aef88  // bfmmla v8.4s, v28.8h, v26.8h\n"
-      ".inst 0x6e5def8c  // bfmmla v12.4s, v28.8h, v29.8h\n"
       ".inst 0x6e5aef70  // bfmmla v16.4s, v27.8h, v26.8h\n"
       "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e5def8c  // bfmmla v12.4s, v28.8h, v29.8h\n"
       ".inst 0x6e5def74  // bfmmla v20.4s, v27.8h, v29.8h\n"
       "ldr q25, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x0]\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x10]\n"
@@ -1354,24 +1352,24 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 93b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x13, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v18.2d, v18.2d, v22.2d\n"
       "uzp1 v19.2d, v19.2d, v23.2d\n"
       "tbz %x[flags], #1, 104f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v26.4s }, [x21]\n"
       "ld1r { v25.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v26.4s\n"
       "fmin v12.4s, v12.4s, v26.4s\n"
@@ -1403,79 +1401,79 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "tbz x14, #3, 108f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
       "st1 { v12.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
       "tbz x14, #2, 106f\n"
       "st1 { v13.4s }, [x13], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
       "tbz x14, #1, 105f\n"
       "str d14, [x13], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x14, #0, 112f\n"
       "st1 { v14.s }[2], [x13]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
       "b 112f\n"
       "105:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x14, #0, 112f\n"
       "str s14, [x13, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
       "b 112f\n"
       "106:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x14, #1, 107f\n"
       "str d13, [x13], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x14, #0, 112f\n"
       "st1 { v13.s }[2], [x13]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
       "b 112f\n"
       "107:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x14, #0, 112f\n"
       "str s13, [x13, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
       "b 112f\n"
       "108:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x14, #2, 110f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
       "tbz x14, #1, 109f\n"
       "str d12, [x13], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x14, #0, 112f\n"
       "st1 { v12.s }[2], [x13]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
       "b 112f\n"
       "109:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x14, #0, 112f\n"
       "str s12, [x13, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
       "b 112f\n"
       "110:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x14, #1, 111f\n"
       "str d7, [x13], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x14, #0, 112f\n"
       "st1 { v7.s }[2], [x13]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
       "b 112f\n"
       "111:"  // Height 3: Partial direct writeback: partial_1_0
       "str s7, [x13, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
       "112:"  // Height 3: Partial direct writeback: Done
       "b 114f\n"
       "113:"  // Height 3: Full writeback
@@ -1484,33 +1482,33 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "str q13, [x13, #0x20]\n"
       "str q14, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "114:"  // Height 3: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 78b\n"
       "b 230f\n"
       "115:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "116:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 117f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -1523,15 +1521,15 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cbz x15, 118f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x15, x15, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -1546,111 +1544,111 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "118:"  // Height 4: no bias
       "tbz %x[flags], #0, 129f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "cmp x14, #0x10\n"
+      "add x23, x24, x20, LSL #2\n"
       "bge 127f\n"
       "tbz x14, #3, 122f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
       "tbz x14, #2, 120f\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
       "tbz x14, #1, 119f\n"
       "ldr d16, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x14, #0, 126f\n"
       "ld1 { v16.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
       "b 126f\n"
       "119:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 126f\n"
       "ldr s16, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
       "b 126f\n"
       "120:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x14, #1, 121f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x14, #0, 126f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
       "b 126f\n"
       "121:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 126f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
       "b 126f\n"
       "122:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x14, #2, 124f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "tbz x14, #1, 123f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x14, #0, 126f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
       "b 126f\n"
       "123:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 126f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
       "b 126f\n"
       "124:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x14, #1, 125f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x14, #0, 126f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
       "b 126f\n"
       "125:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s9, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
       "126:"  // Height 4: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 128f\n"
@@ -1659,18 +1657,18 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x13, #0x10]\n"
       "ldr q11, [x13, #0x20]\n"
       "ldr q16, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "128:"  // Height 4: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -1710,8 +1708,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "131:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 132f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1745,56 +1743,56 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "134:"  // Height 4: Multiply loop: Main loop head
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "cmp x27, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ldr q4, [x23, #0x0]\n"
-      ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
-      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e47ef70  // bfmmla v16.4s, v27.8h, v7.8h\n"
       "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e46ef74  // bfmmla v20.4s, v27.8h, v6.8h\n"
       "ldr q25, [x11, #0x10]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x0]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x10]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x9, #0x0]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x9, #0x10]\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       "ldr q26, [x12, #0x20]\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
+      "add x23, x23, #0x10\n"
+      "ldr q4, [x23, #0x0]\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
       "ldr q25, [x12, #0x30]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e5aec28  // bfmmla v8.4s, v1.8h, v26.8h\n"
+      "ldr q2, [x25, #0x0]\n"
       ".inst 0x6e5aec70  // bfmmla v16.4s, v3.8h, v26.8h\n"
       "ldr q26, [x11, #0x20]\n"
       ".inst 0x6e59ec2c  // bfmmla v12.4s, v1.8h, v25.8h\n"
       ".inst 0x6e59ec74  // bfmmla v20.4s, v3.8h, v25.8h\n"
       "ldr q25, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e5aec29  // bfmmla v9.4s, v1.8h, v26.8h\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6e5aec71  // bfmmla v17.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e59ec2d  // bfmmla v13.4s, v1.8h, v25.8h\n"
+      "add x11, x11, #0x40\n"
       ".inst 0x6e59ec75  // bfmmla v21.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e5aec2a  // bfmmla v10.4s, v1.8h, v26.8h\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6e5aec72  // bfmmla v18.4s, v3.8h, v26.8h\n"
       "ldr q26, [x9, #0x20]\n"
       ".inst 0x6e59ec2e  // bfmmla v14.4s, v1.8h, v25.8h\n"
@@ -1813,53 +1811,53 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "135:"  // Height 4: Multiply loop: Single iteration only
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
-      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e47ef70  // bfmmla v16.4s, v27.8h, v7.8h\n"
       "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e46ef74  // bfmmla v20.4s, v27.8h, v6.8h\n"
       "ldr q25, [x11, #0x10]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x0]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x10]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x9, #0x0]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x9, #0x10]\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       "ldr q26, [x12, #0x20]\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
       "ldr q25, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e5aec28  // bfmmla v8.4s, v1.8h, v26.8h\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6e5aec70  // bfmmla v16.4s, v3.8h, v26.8h\n"
       "ldr q26, [x11, #0x20]\n"
       ".inst 0x6e59ec2c  // bfmmla v12.4s, v1.8h, v25.8h\n"
       ".inst 0x6e59ec74  // bfmmla v20.4s, v3.8h, v25.8h\n"
       "ldr q25, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e5aec29  // bfmmla v9.4s, v1.8h, v26.8h\n"
+      "add x11, x11, #0x40\n"
       ".inst 0x6e5aec71  // bfmmla v17.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e59ec2d  // bfmmla v13.4s, v1.8h, v25.8h\n"
       ".inst 0x6e59ec75  // bfmmla v21.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e5aec2a  // bfmmla v10.4s, v1.8h, v26.8h\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6e5aec72  // bfmmla v18.4s, v3.8h, v26.8h\n"
       "ldr q26, [x9, #0x20]\n"
       ".inst 0x6e59ec2e  // bfmmla v14.4s, v1.8h, v25.8h\n"
@@ -1875,34 +1873,34 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cmp x27, #0x4\n"
       "blt 138f\n"
       "137:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "trn1 v28.2d, v26.2d, v25.2d\n"
       "sub x27, x27, #0x4\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d26, [x24], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "trn1 v27.2d, v26.2d, v25.2d\n"
       "cmp x27, #0x4\n"
       "ldr q26, [x12, #0x0]\n"
       "ldr q25, [x12, #0x10]\n"
-      "add x12, x12, #0x20\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v29.2d, v27.2d\n"
       ".inst 0x6e5aef88  // bfmmla v8.4s, v28.8h, v26.8h\n"
-      ".inst 0x6e59ef8c  // bfmmla v12.4s, v28.8h, v25.8h\n"
       ".inst 0x6e5aef70  // bfmmla v16.4s, v27.8h, v26.8h\n"
       "ldr q26, [x11, #0x0]\n"
+      ".inst 0x6e59ef8c  // bfmmla v12.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef74  // bfmmla v20.4s, v27.8h, v25.8h\n"
       "ldr q25, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x0]\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x9, #0x0]\n"
+      "add x10, x10, #0x20\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x9, #0x10]\n"
@@ -1935,10 +1933,10 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q25, [x12, #0x10]\n"
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "add x12, x12, #0x20\n"
       ".inst 0x6e5aef88  // bfmmla v8.4s, v28.8h, v26.8h\n"
       ".inst 0x6e5aef70  // bfmmla v16.4s, v27.8h, v26.8h\n"
       "ldr q26, [x11, #0x0]\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e59ef8c  // bfmmla v12.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef74  // bfmmla v20.4s, v27.8h, v25.8h\n"
       "ldr q25, [x11, #0x10]\n"
@@ -1967,17 +1965,17 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 131b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
@@ -1987,9 +1985,9 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
       "tbz %x[flags], #1, 142f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v26.4s }, [x21]\n"
       "ld1r { v25.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v26.4s\n"
       "fmin v12.4s, v12.4s, v26.4s\n"
@@ -2029,95 +2027,95 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "tbz x14, #3, 146f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
       "st1 { v12.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
       "tbz x14, #2, 144f\n"
       "st1 { v13.4s }, [x13], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
       "tbz x14, #1, 143f\n"
       "str d14, [x13], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
       "tbz x14, #0, 150f\n"
       "st1 { v14.s }[2], [x13]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
       "b 150f\n"
       "143:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x14, #0, 150f\n"
       "str s14, [x13, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
       "b 150f\n"
       "144:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x14, #1, 145f\n"
       "str d13, [x13], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
       "tbz x14, #0, 150f\n"
       "st1 { v13.s }[2], [x13]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
       "b 150f\n"
       "145:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x14, #0, 150f\n"
       "str s13, [x13, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
       "b 150f\n"
       "146:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x14, #2, 148f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
       "tbz x14, #1, 147f\n"
       "str d12, [x13], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
       "tbz x14, #0, 150f\n"
       "st1 { v12.s }[2], [x13]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
       "b 150f\n"
       "147:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x14, #0, 150f\n"
       "str s12, [x13, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
       "b 150f\n"
       "148:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x14, #1, 149f\n"
       "str d7, [x13], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
       "tbz x14, #0, 150f\n"
       "st1 { v7.s }[2], [x13]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
       "b 150f\n"
       "149:"  // Height 4: Partial direct writeback: partial_1_0
       "str s7, [x13, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
       "150:"  // Height 4: Partial direct writeback: Done
       "b 152f\n"
       "151:"  // Height 4: Full writeback
@@ -2126,37 +2124,37 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "str q13, [x13, #0x20]\n"
       "str q14, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
       "152:"  // Height 4: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 116b\n"
       "b 230f\n"
       "153:"  // Height 5
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "154:"  // Height 5: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 155f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -2169,15 +2167,15 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cbz x15, 156f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x15, x15, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -2200,128 +2198,128 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "156:"  // Height 5: no bias
       "tbz %x[flags], #0, 167f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x14, #0x10\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 165f\n"
       "tbz x14, #3, 160f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
       "tbz x14, #2, 158f\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v27.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v27.4s }, [x22], #0x10\n"
       "tbz x14, #1, 157f\n"
       "ldr d16, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d6, [x23], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d6, [x22], #0x8\n"
       "tbz x14, #0, 164f\n"
       "ld1 { v16.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v6.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v6.s }[2], [x22]\n"
       "b 164f\n"
       "157:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 164f\n"
       "ldr s16, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s6, [x23, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s6, [x22, #0x0]\n"
       "b 164f\n"
       "158:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x14, #1, 159f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x14, #0, 164f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
       "b 164f\n"
       "159:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 164f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
       "b 164f\n"
       "160:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x14, #2, 162f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x14, #1, 161f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x14, #0, 164f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
       "b 164f\n"
       "161:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 164f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
       "b 164f\n"
       "162:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x14, #1, 163f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x14, #0, 164f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 164f\n"
       "163:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s9, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "164:"  // Height 5: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 166f\n"
@@ -2330,22 +2328,22 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x13, #0x10]\n"
       "ldr q11, [x13, #0x20]\n"
       "ldr q16, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q25, [x23, #0x0]\n"
-      "ldr q26, [x23, #0x10]\n"
-      "ldr q27, [x23, #0x20]\n"
-      "ldr q6, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q25, [x22, #0x0]\n"
+      "ldr q26, [x22, #0x10]\n"
+      "ldr q27, [x22, #0x20]\n"
+      "ldr q6, [x22, #0x30]\n"
       "166:"  // Height 5: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -2401,8 +2399,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "169:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 170f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2439,35 +2437,35 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "172:"  // Height 5: Multiply loop: Main loop head
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
+      ".inst 0x6e47ecc8  // bfmmla v8.4s, v6.8h, v7.8h\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
+      "sub x27, x27, #0x8\n"
       "trn1 v4.2d, v5.2d, v0.2d\n"
       "trn2 v5.2d, v5.2d, v0.2d\n"
       "ldr q0, [x12, #0x10]\n"
-      ".inst 0x6e47ecc8  // bfmmla v8.4s, v6.8h, v7.8h\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q7, [x11, #0x0]\n"
       ".inst 0x6e40eccc  // bfmmla v12.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec54  // bfmmla v20.4s, v2.8h, v0.8h\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e40ec9c  // bfmmla v28.4s, v4.8h, v0.8h\n"
       "ldr q0, [x11, #0x10]\n"
       ".inst 0x6e47ecc9  // bfmmla v9.4s, v6.8h, v7.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e47ec51  // bfmmla v17.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec99  // bfmmla v25.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x0]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e40eccd  // bfmmla v13.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec55  // bfmmla v21.4s, v2.8h, v0.8h\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e40ec9d  // bfmmla v29.4s, v4.8h, v0.8h\n"
       "ldr q0, [x10, #0x10]\n"
       ".inst 0x6e47ecca  // bfmmla v10.4s, v6.8h, v7.8h\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e47ec52  // bfmmla v18.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9a  // bfmmla v26.4s, v4.8h, v7.8h\n"
       "ldr q7, [x9, #0x0]\n"
@@ -2485,17 +2483,17 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e40ec9f  // bfmmla v31.4s, v4.8h, v0.8h\n"
       "ldr q0, [x12, #0x30]\n"
       "ldr q4, [x23, #0x0]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e47ec28  // bfmmla v8.4s, v1.8h, v7.8h\n"
       ".inst 0x6e47ec70  // bfmmla v16.4s, v3.8h, v7.8h\n"
       ".inst 0x6e47ecb8  // bfmmla v24.4s, v5.8h, v7.8h\n"
       "ldr q6, [x11, #0x20]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6e40ec2c  // bfmmla v12.4s, v1.8h, v0.8h\n"
       ".inst 0x6e40ec74  // bfmmla v20.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ecbc  // bfmmla v28.4s, v5.8h, v0.8h\n"
       "ldr q0, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e46ec29  // bfmmla v9.4s, v1.8h, v6.8h\n"
+      "add x11, x11, #0x40\n"
       ".inst 0x6e46ec71  // bfmmla v17.4s, v3.8h, v6.8h\n"
       ".inst 0x6e46ecb9  // bfmmla v25.4s, v5.8h, v6.8h\n"
       "ldr q6, [x10, #0x20]\n"
@@ -2503,8 +2501,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e40ec75  // bfmmla v21.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ecbd  // bfmmla v29.4s, v5.8h, v0.8h\n"
       "ldr q0, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e46ec2a  // bfmmla v10.4s, v1.8h, v6.8h\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6e46ec72  // bfmmla v18.4s, v3.8h, v6.8h\n"
       ".inst 0x6e46ecba  // bfmmla v26.4s, v5.8h, v6.8h\n"
       "ldr q6, [x9, #0x20]\n"
@@ -2527,31 +2525,31 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "173:"  // Height 5: Multiply loop: Single iteration only
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
+      ".inst 0x6e47ecc8  // bfmmla v8.4s, v6.8h, v7.8h\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
+      "sub x27, x27, #0x8\n"
       "trn1 v4.2d, v5.2d, v0.2d\n"
       "trn2 v5.2d, v5.2d, v0.2d\n"
       "ldr q0, [x12, #0x10]\n"
-      ".inst 0x6e47ecc8  // bfmmla v8.4s, v6.8h, v7.8h\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q7, [x11, #0x0]\n"
       ".inst 0x6e40eccc  // bfmmla v12.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec54  // bfmmla v20.4s, v2.8h, v0.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e40ec9c  // bfmmla v28.4s, v4.8h, v0.8h\n"
       "ldr q0, [x11, #0x10]\n"
       ".inst 0x6e47ecc9  // bfmmla v9.4s, v6.8h, v7.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e47ec51  // bfmmla v17.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec99  // bfmmla v25.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x0]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e40eccd  // bfmmla v13.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec55  // bfmmla v21.4s, v2.8h, v0.8h\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e40ec9d  // bfmmla v29.4s, v4.8h, v0.8h\n"
       "ldr q0, [x10, #0x10]\n"
       ".inst 0x6e47ecca  // bfmmla v10.4s, v6.8h, v7.8h\n"
@@ -2570,8 +2568,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e40ec57  // bfmmla v23.4s, v2.8h, v0.8h\n"
       ".inst 0x6e40ec9f  // bfmmla v31.4s, v4.8h, v0.8h\n"
       "ldr q2, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e47ec28  // bfmmla v8.4s, v1.8h, v7.8h\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6e47ec70  // bfmmla v16.4s, v3.8h, v7.8h\n"
       ".inst 0x6e47ecb8  // bfmmla v24.4s, v5.8h, v7.8h\n"
       "ldr q0, [x11, #0x20]\n"
@@ -2579,8 +2577,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e42ec74  // bfmmla v20.4s, v3.8h, v2.8h\n"
       ".inst 0x6e42ecbc  // bfmmla v28.4s, v5.8h, v2.8h\n"
       "ldr q2, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e40ec29  // bfmmla v9.4s, v1.8h, v0.8h\n"
+      "add x11, x11, #0x40\n"
       ".inst 0x6e40ec71  // bfmmla v17.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ecb9  // bfmmla v25.4s, v5.8h, v0.8h\n"
       "ldr q0, [x10, #0x20]\n"
@@ -2588,8 +2586,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e42ec75  // bfmmla v21.4s, v3.8h, v2.8h\n"
       ".inst 0x6e42ecbd  // bfmmla v29.4s, v5.8h, v2.8h\n"
       "ldr q2, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e40ec2a  // bfmmla v10.4s, v1.8h, v0.8h\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6e40ec72  // bfmmla v18.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ecba  // bfmmla v26.4s, v5.8h, v0.8h\n"
       "ldr q0, [x9, #0x20]\n"
@@ -2609,29 +2607,29 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cmp x27, #0x4\n"
       "blt 176f\n"
       "175:"  // Height 5: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x4\n"
-      "ldr d3, [x24], #0x8\n"
-      "ldr d2, [x23], #0x8\n"
-      "cmp x27, #0x4\n"
       "ldr d0, [x22], #0x8\n"
       "ldr q1, [x12, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v3.2d, v2.2d\n"
-      "trn1 v2.2d, v0.2d, v5.2d\n"
-      "ldr q0, [x12, #0x10]\n"
-      "add x12, x12, #0x20\n"
+      "trn1 v2.2d, v0.2d, v2.2d\n"
       ".inst 0x6e41ec88  // bfmmla v8.4s, v4.8h, v1.8h\n"
+      "ldr q0, [x12, #0x10]\n"
       ".inst 0x6e41ec70  // bfmmla v16.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec58  // bfmmla v24.4s, v2.8h, v1.8h\n"
       "ldr q1, [x11, #0x0]\n"
       ".inst 0x6e40ec8c  // bfmmla v12.4s, v4.8h, v0.8h\n"
       ".inst 0x6e40ec74  // bfmmla v20.4s, v3.8h, v0.8h\n"
+      "cmp x27, #0x4\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e40ec5c  // bfmmla v28.4s, v2.8h, v0.8h\n"
       "ldr q0, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e41ec89  // bfmmla v9.4s, v4.8h, v1.8h\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e41ec71  // bfmmla v17.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec59  // bfmmla v25.4s, v2.8h, v1.8h\n"
       "ldr q1, [x10, #0x0]\n"
@@ -2639,8 +2637,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e40ec75  // bfmmla v21.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec5d  // bfmmla v29.4s, v2.8h, v0.8h\n"
       "ldr q0, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e41ec8a  // bfmmla v10.4s, v4.8h, v1.8h\n"
+      "add x10, x10, #0x20\n"
       ".inst 0x6e41ec72  // bfmmla v18.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec5a  // bfmmla v26.4s, v2.8h, v1.8h\n"
       "ldr q6, [x9, #0x0]\n"
@@ -2648,8 +2646,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e40ec76  // bfmmla v22.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec5e  // bfmmla v30.4s, v2.8h, v0.8h\n"
       "ldr q0, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e46ec8b  // bfmmla v11.4s, v4.8h, v6.8h\n"
+      "add x9, x9, #0x20\n"
       ".inst 0x6e46ec73  // bfmmla v19.4s, v3.8h, v6.8h\n"
       ".inst 0x6e46ec5b  // bfmmla v27.4s, v2.8h, v6.8h\n"
       ".inst 0x6e40ec8f  // bfmmla v15.4s, v4.8h, v0.8h\n"
@@ -2683,12 +2681,12 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "trn1 v3.2d, v3.2d, v4.2d\n"
       "trn1 v2.2d, v5.2d, v0.2d\n"
       "ldr q1, [x12, #0x10]\n"
-      "add x12, x12, #0x20\n"
       ".inst 0x6e46ece8  // bfmmla v8.4s, v7.8h, v6.8h\n"
       ".inst 0x6e46ec70  // bfmmla v16.4s, v3.8h, v6.8h\n"
       ".inst 0x6e46ec58  // bfmmla v24.4s, v2.8h, v6.8h\n"
       "ldr q0, [x11, #0x0]\n"
       ".inst 0x6e41ecec  // bfmmla v12.4s, v7.8h, v1.8h\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e41ec74  // bfmmla v20.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec5c  // bfmmla v28.4s, v2.8h, v1.8h\n"
       "ldr q1, [x11, #0x10]\n"
@@ -2723,20 +2721,20 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 169b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
@@ -2748,9 +2746,9 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "uzp1 v26.2d, v26.2d, v30.2d\n"
       "uzp1 v27.2d, v27.2d, v31.2d\n"
       "tbz %x[flags], #1, 180f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v1.4s\n"
       "fmin v12.4s, v12.4s, v1.4s\n"
@@ -2798,111 +2796,111 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "tbz x14, #3, 184f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
       "st1 { v12.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
       "tbz x14, #2, 182f\n"
       "st1 { v13.4s }, [x13], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
       "tbz x14, #1, 181f\n"
       "str d14, [x13], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x14, #0, 188f\n"
       "st1 { v14.s }[2], [x13]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
       "b 188f\n"
       "181:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x14, #0, 188f\n"
       "str s14, [x13, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
       "b 188f\n"
       "182:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x14, #1, 183f\n"
       "str d13, [x13], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x14, #0, 188f\n"
       "st1 { v13.s }[2], [x13]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
       "b 188f\n"
       "183:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x14, #0, 188f\n"
       "str s13, [x13, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
       "b 188f\n"
       "184:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x14, #2, 186f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x14, #1, 185f\n"
       "str d12, [x13], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x14, #0, 188f\n"
       "st1 { v12.s }[2], [x13]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 188f\n"
       "185:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x14, #0, 188f\n"
       "str s12, [x13, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 188f\n"
       "186:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x14, #1, 187f\n"
       "str d7, [x13], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x14, #0, 188f\n"
       "st1 { v7.s }[2], [x13]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 188f\n"
       "187:"  // Height 5: Partial direct writeback: partial_1_0
       "str s7, [x13, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "188:"  // Height 5: Partial direct writeback: Done
       "b 190f\n"
       "189:"  // Height 5: Full writeback
@@ -2911,45 +2909,44 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "str q13, [x13, #0x20]\n"
       "str q14, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "190:"  // Height 5: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 154b\n"
       "b 230f\n"
       "191:"  // Height 6
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0x18\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0x18\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "192:"  // Height 6: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 193f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -2962,15 +2959,15 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cbz x15, 194f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x15, x15, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -2993,145 +2990,145 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "194:"  // Height 6: no bias
       "tbz %x[flags], #0, 205f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x14, #0x10\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 203f\n"
       "tbz x14, #3, 198f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x14, #2, 196f\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v27.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x14, #1, 195f\n"
       "ldr d16, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d6, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d6, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x14, #0, 202f\n"
       "ld1 { v16.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v6.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v6.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 202f\n"
       "195:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 202f\n"
       "ldr s16, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s6, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s6, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 202f\n"
       "196:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x14, #1, 197f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x14, #0, 202f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 202f\n"
       "197:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 202f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 202f\n"
       "198:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x14, #2, 200f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x14, #1, 199f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x14, #0, 202f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 202f\n"
       "199:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 202f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 202f\n"
       "200:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x14, #1, 201f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x14, #0, 202f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 202f\n"
       "201:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s9, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "202:"  // Height 6: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 204f\n"
@@ -3140,26 +3137,26 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x13, #0x10]\n"
       "ldr q11, [x13, #0x20]\n"
       "ldr q16, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q25, [x23, #0x0]\n"
-      "ldr q26, [x23, #0x10]\n"
-      "ldr q27, [x23, #0x20]\n"
-      "ldr q6, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q25, [x22, #0x0]\n"
+      "ldr q26, [x22, #0x10]\n"
+      "ldr q27, [x22, #0x20]\n"
+      "ldr q6, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "204:"  // Height 6: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -3215,8 +3212,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "207:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 208f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -3257,36 +3254,36 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "210:"  // Height 6: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
       "cmp x27, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
       "ldr q6, [x12, #0x10]\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
-      "add x21, x21, #0x10\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q7, [x11, #0x0]\n"
       ".inst 0x6e46ec0c  // bfmmla v12.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec54  // bfmmla v20.4s, v2.8h, v6.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e46ec9c  // bfmmla v28.4s, v4.8h, v6.8h\n"
       "ldr q6, [x11, #0x10]\n"
       ".inst 0x6e47ec09  // bfmmla v9.4s, v0.8h, v7.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e47ec51  // bfmmla v17.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec99  // bfmmla v25.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x0]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e46ec0d  // bfmmla v13.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e46ec9d  // bfmmla v29.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x10]\n"
       ".inst 0x6e47ec0a  // bfmmla v10.4s, v0.8h, v7.8h\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x6e47ec52  // bfmmla v18.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9a  // bfmmla v26.4s, v4.8h, v7.8h\n"
       "ldr q7, [x9, #0x0]\n"
@@ -3304,17 +3301,17 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e46ec9f  // bfmmla v31.4s, v4.8h, v6.8h\n"
       "ldr q0, [x12, #0x30]\n"
       "ldr q4, [x23, #0x0]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e47ec28  // bfmmla v8.4s, v1.8h, v7.8h\n"
       ".inst 0x6e47ec70  // bfmmla v16.4s, v3.8h, v7.8h\n"
       ".inst 0x6e47ecb8  // bfmmla v24.4s, v5.8h, v7.8h\n"
       "ldr q6, [x11, #0x20]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6e40ec2c  // bfmmla v12.4s, v1.8h, v0.8h\n"
       ".inst 0x6e40ec74  // bfmmla v20.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ecbc  // bfmmla v28.4s, v5.8h, v0.8h\n"
       "ldr q0, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e46ec29  // bfmmla v9.4s, v1.8h, v6.8h\n"
+      "add x11, x11, #0x40\n"
       ".inst 0x6e46ec71  // bfmmla v17.4s, v3.8h, v6.8h\n"
       ".inst 0x6e46ecb9  // bfmmla v25.4s, v5.8h, v6.8h\n"
       "ldr q6, [x10, #0x20]\n"
@@ -3322,8 +3319,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e40ec75  // bfmmla v21.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ecbd  // bfmmla v29.4s, v5.8h, v0.8h\n"
       "ldr q0, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e46ec2a  // bfmmla v10.4s, v1.8h, v6.8h\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6e46ec72  // bfmmla v18.4s, v3.8h, v6.8h\n"
       ".inst 0x6e46ecba  // bfmmla v26.4s, v5.8h, v6.8h\n"
       "ldr q6, [x9, #0x20]\n"
@@ -3347,32 +3344,32 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "211:"  // Height 6: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
       "ldr q6, [x12, #0x10]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q7, [x11, #0x0]\n"
       ".inst 0x6e46ec0c  // bfmmla v12.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec54  // bfmmla v20.4s, v2.8h, v6.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e46ec9c  // bfmmla v28.4s, v4.8h, v6.8h\n"
       "ldr q6, [x11, #0x10]\n"
       ".inst 0x6e47ec09  // bfmmla v9.4s, v0.8h, v7.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e47ec51  // bfmmla v17.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec99  // bfmmla v25.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x0]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e46ec0d  // bfmmla v13.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
+      "add x22, x22, #0x10\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x6e46ec9d  // bfmmla v29.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x10]\n"
       ".inst 0x6e47ec0a  // bfmmla v10.4s, v0.8h, v7.8h\n"
@@ -3391,8 +3388,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e46ec57  // bfmmla v23.4s, v2.8h, v6.8h\n"
       ".inst 0x6e46ec9f  // bfmmla v31.4s, v4.8h, v6.8h\n"
       "ldr q2, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       ".inst 0x6e47ec28  // bfmmla v8.4s, v1.8h, v7.8h\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6e47ec70  // bfmmla v16.4s, v3.8h, v7.8h\n"
       ".inst 0x6e47ecb8  // bfmmla v24.4s, v5.8h, v7.8h\n"
       "ldr q0, [x11, #0x20]\n"
@@ -3400,8 +3397,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e42ec74  // bfmmla v20.4s, v3.8h, v2.8h\n"
       ".inst 0x6e42ecbc  // bfmmla v28.4s, v5.8h, v2.8h\n"
       "ldr q2, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       ".inst 0x6e40ec29  // bfmmla v9.4s, v1.8h, v0.8h\n"
+      "add x11, x11, #0x40\n"
       ".inst 0x6e40ec71  // bfmmla v17.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ecb9  // bfmmla v25.4s, v5.8h, v0.8h\n"
       "ldr q0, [x10, #0x20]\n"
@@ -3409,8 +3406,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e42ec75  // bfmmla v21.4s, v3.8h, v2.8h\n"
       ".inst 0x6e42ecbd  // bfmmla v29.4s, v5.8h, v2.8h\n"
       "ldr q2, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6e40ec2a  // bfmmla v10.4s, v1.8h, v0.8h\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6e40ec72  // bfmmla v18.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ecba  // bfmmla v26.4s, v5.8h, v0.8h\n"
       "ldr q0, [x9, #0x20]\n"
@@ -3430,25 +3427,25 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cmp x27, #0x4\n"
       "blt 214f\n"
       "213:"  // Height 6: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x4\n"
-      "ldr d5, [x24], #0x8\n"
-      "ldr d3, [x23], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "cmp x27, #0x4\n"
-      "ldr d2, [x22], #0x8\n"
+      "ldr d1, [x22], #0x8\n"
       "ldr d0, [x21], #0x8\n"
+      "trn1 v2.2d, v1.2d, v0.2d\n"
       "ldr q1, [x12, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v5.2d, v3.2d\n"
-      "trn1 v2.2d, v2.2d, v0.2d\n"
       "ldr q0, [x12, #0x10]\n"
-      "add x12, x12, #0x20\n"
       ".inst 0x6e41ec88  // bfmmla v8.4s, v4.8h, v1.8h\n"
       ".inst 0x6e41ec70  // bfmmla v16.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec58  // bfmmla v24.4s, v2.8h, v1.8h\n"
       "ldr q1, [x11, #0x0]\n"
       ".inst 0x6e40ec8c  // bfmmla v12.4s, v4.8h, v0.8h\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e40ec74  // bfmmla v20.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec5c  // bfmmla v28.4s, v2.8h, v0.8h\n"
       "ldr q0, [x11, #0x10]\n"
@@ -3506,19 +3503,19 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "ldr q0, [x12, #0x0]\n"
       "trn1 v7.2d, v1.2d, v2.2d\n"
       "trn1 v3.2d, v3.2d, v4.2d\n"
+      ".inst 0x6e40ece8  // bfmmla v8.4s, v7.8h, v0.8h\n"
       "trn1 v2.2d, v5.2d, v6.2d\n"
       "ldr q1, [x12, #0x10]\n"
-      "add x12, x12, #0x20\n"
-      ".inst 0x6e40ece8  // bfmmla v8.4s, v7.8h, v0.8h\n"
       ".inst 0x6e40ec70  // bfmmla v16.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec58  // bfmmla v24.4s, v2.8h, v0.8h\n"
       "ldr q0, [x11, #0x0]\n"
       ".inst 0x6e41ecec  // bfmmla v12.4s, v7.8h, v1.8h\n"
       ".inst 0x6e41ec74  // bfmmla v20.4s, v3.8h, v1.8h\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e41ec5c  // bfmmla v28.4s, v2.8h, v1.8h\n"
       "ldr q1, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e40ece9  // bfmmla v9.4s, v7.8h, v0.8h\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e40ec71  // bfmmla v17.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec59  // bfmmla v25.4s, v2.8h, v0.8h\n"
       "ldr q0, [x10, #0x0]\n"
@@ -3526,8 +3523,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e41ec75  // bfmmla v21.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec5d  // bfmmla v29.4s, v2.8h, v1.8h\n"
       "ldr q1, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e40ecea  // bfmmla v10.4s, v7.8h, v0.8h\n"
+      "add x10, x10, #0x20\n"
       ".inst 0x6e40ec72  // bfmmla v18.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec5a  // bfmmla v26.4s, v2.8h, v0.8h\n"
       "ldr q0, [x9, #0x0]\n"
@@ -3535,8 +3532,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e41ec76  // bfmmla v22.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec5e  // bfmmla v30.4s, v2.8h, v1.8h\n"
       "ldr q6, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e40eceb  // bfmmla v11.4s, v7.8h, v0.8h\n"
+      "add x9, x9, #0x20\n"
       ".inst 0x6e40ec73  // bfmmla v19.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec5b  // bfmmla v27.4s, v2.8h, v0.8h\n"
       ".inst 0x6e46ecef  // bfmmla v15.4s, v7.8h, v6.8h\n"
@@ -3548,21 +3545,21 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 207b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
@@ -3578,9 +3575,9 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "uzp1 v30.2d, v27.2d, v31.2d\n"
       "uzp2 v27.2d, v27.2d, v31.2d\n"
       "tbz %x[flags], #1, 218f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v1.4s\n"
       "fmin v12.4s, v12.4s, v1.4s\n"
@@ -3636,127 +3633,127 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "tbz x14, #3, 222f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
       "st1 { v12.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x23], #0x10\n"
-      "st1 { v24.4s }, [x22], #0x10\n"
-      "st1 { v25.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v24.4s }, [x21], #0x10\n"
+      "st1 { v25.4s }, [x21], #0x10\n"
       "tbz x14, #2, 220f\n"
       "st1 { v13.4s }, [x13], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v29.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
       "tbz x14, #1, 219f\n"
       "str d14, [x13], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
-      "str d27, [x22], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d30, [x22], #0x8\n"
+      "str d27, [x21], #0x8\n"
       "tbz x14, #0, 226f\n"
       "st1 { v14.s }[2], [x13]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
-      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v27.s }[2], [x21]\n"
       "b 226f\n"
       "219:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x14, #0, 226f\n"
       "str s14, [x13, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
-      "str s27, [x22, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s30, [x22, #0x0]\n"
+      "str s27, [x21, #0x0]\n"
       "b 226f\n"
       "220:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x14, #1, 221f\n"
       "str d13, [x13], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d29, [x23], #0x8\n"
-      "str d26, [x22], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d29, [x22], #0x8\n"
+      "str d26, [x21], #0x8\n"
       "tbz x14, #0, 226f\n"
       "st1 { v13.s }[2], [x13]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v29.s }[2], [x23]\n"
-      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v26.s }[2], [x21]\n"
       "b 226f\n"
       "221:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x14, #0, 226f\n"
       "str s13, [x13, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s29, [x23, #0x0]\n"
-      "str s26, [x22, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s29, [x22, #0x0]\n"
+      "str s26, [x21, #0x0]\n"
       "b 226f\n"
       "222:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x14, #2, 224f\n"
       "st1 { v7.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
-      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v24.4s }, [x21], #0x10\n"
       "tbz x14, #1, 223f\n"
       "str d12, [x13], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d28, [x23], #0x8\n"
-      "str d25, [x22], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d28, [x22], #0x8\n"
+      "str d25, [x21], #0x8\n"
       "tbz x14, #0, 226f\n"
       "st1 { v12.s }[2], [x13]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v28.s }[2], [x23]\n"
-      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v25.s }[2], [x21]\n"
       "b 226f\n"
       "223:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x14, #0, 226f\n"
       "str s12, [x13, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s28, [x23, #0x0]\n"
-      "str s25, [x22, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s28, [x22, #0x0]\n"
+      "str s25, [x21, #0x0]\n"
       "b 226f\n"
       "224:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x14, #1, 225f\n"
       "str d7, [x13], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
+      "str d24, [x21], #0x8\n"
       "tbz x14, #0, 226f\n"
       "st1 { v7.s }[2], [x13]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
-      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
+      "st1 { v24.s }[2], [x21]\n"
       "b 226f\n"
       "225:"  // Height 6: Partial direct writeback: partial_1_0
       "str s7, [x13, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
-      "str s24, [x22, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
+      "str s24, [x21, #0x0]\n"
       "226:"  // Height 6: Partial direct writeback: Done
       "b 228f\n"
       "227:"  // Height 6: Full writeback
@@ -3765,26 +3762,26 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "str q13, [x13, #0x20]\n"
       "str q14, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
-      "str q23, [x23, #0x0]\n"
-      "str q28, [x23, #0x10]\n"
-      "str q29, [x23, #0x20]\n"
-      "str q30, [x23, #0x30]\n"
-      "str q24, [x22, #0x0]\n"
-      "str q25, [x22, #0x10]\n"
-      "str q26, [x22, #0x20]\n"
-      "str q27, [x22, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
+      "str q23, [x22, #0x0]\n"
+      "str q28, [x22, #0x10]\n"
+      "str q29, [x22, #0x20]\n"
+      "str q30, [x22, #0x30]\n"
+      "str q24, [x21, #0x0]\n"
+      "str q25, [x21, #0x10]\n"
+      "str q26, [x21, #0x20]\n"
+      "str q27, [x21, #0x30]\n"
       "228:"  // Height 6: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 192b\n"
@@ -3800,8 +3797,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "230:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32.hpp
index 20138ffe7e..f6a7461740 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32.hpp
@@ -81,7 +81,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 32, 1> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 32, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp
index 93e5e051f8..2686b98092 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp16_mla_6x32/generic.cpp
@@ -49,19 +49,18 @@ void a64_ffhybrid_fp16_mla_6x32 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const __fp16 *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -82,7 +81,6 @@ void a64_ffhybrid_fp16_mla_6x32 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -107,19 +105,19 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "bgt 101f\n"
       "beq 51f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x18\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x18\n"
       "bgt 3f\n"
       "cmp x14, #0x10\n"
       "mov x9, x12\n"
@@ -266,8 +264,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "24:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 25f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -291,9 +289,6 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr q17, [x10, #0x0]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "ldr q16, [x9, #0x0]\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x10\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "ldr q17, [x12, #0x10]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
@@ -344,29 +339,30 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr q16, [x9, #0x60]\n"
       "fmla v10.8h, v17.8h, v0.h[6]\n"
       "ldr q17, [x12, #0x70]\n"
-      "add x12, x12, #0x80\n"
       "fmla v11.8h, v16.8h, v0.h[6]\n"
       "ldr q16, [x11, #0x70]\n"
-      "add x11, x11, #0x80\n"
       "fmla v8.8h, v17.8h, v0.h[7]\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       "fmla v9.8h, v16.8h, v0.h[7]\n"
       "ldr q16, [x9, #0x70]\n"
-      "add x9, x9, #0x80\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x10\n"
       "fmla v10.8h, v17.8h, v0.h[7]\n"
-      "ldr q6, [x12, #0x0]\n"
       "fmla v11.8h, v16.8h, v0.h[7]\n"
+      "add x26, x26, #0x10\n"
       "ldr q0, [x26, #0x0]\n"
+      "add x12, x12, #0x80\n"
+      "ldr q6, [x12, #0x0]\n"
+      "add x11, x11, #0x80\n"
       "ldr q7, [x11, #0x0]\n"
+      "add x10, x10, #0x80\n"
+      "add x9, x9, #0x80\n"
       "bge 27b\n"
       "28:"  // Height 1: Multiply loop: Single iteration only
       "fmla v8.8h, v6.8h, v0.h[0]\n"
       "ldr q17, [x10, #0x0]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "ldr q16, [x9, #0x0]\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "ldr q17, [x12, #0x10]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
@@ -417,35 +413,37 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr q16, [x9, #0x60]\n"
       "fmla v10.8h, v17.8h, v0.h[6]\n"
       "ldr q17, [x12, #0x70]\n"
-      "add x12, x12, #0x80\n"
       "fmla v11.8h, v16.8h, v0.h[6]\n"
       "ldr q16, [x11, #0x70]\n"
-      "add x11, x11, #0x80\n"
       "fmla v8.8h, v17.8h, v0.h[7]\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       "fmla v9.8h, v16.8h, v0.h[7]\n"
       "ldr q16, [x9, #0x70]\n"
-      "add x9, x9, #0x80\n"
+      "sub x27, x27, #0x8\n"
       "fmla v10.8h, v17.8h, v0.h[7]\n"
       "fmla v11.8h, v16.8h, v0.h[7]\n"
+      "add x26, x26, #0x10\n"
+      "add x12, x12, #0x80\n"
+      "add x11, x11, #0x80\n"
+      "add x10, x10, #0x80\n"
+      "add x9, x9, #0x80\n"
       "29:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 31f\n"
       "30:"  // Height 1: Multiply loop: Odd block loop
       "ldr h0, [x26], #0x2\n"
-      "ldr q17, [x12, #0x0]\n"
+      "ldr q16, [x12, #0x0]\n"
+      "fmla v8.8h, v16.8h, v0.h[0]\n"
       "sub x27, x27, #0x1\n"
+      "ldr q17, [x11, #0x0]\n"
+      "ldr q16, [x10, #0x0]\n"
+      "fmla v9.8h, v17.8h, v0.h[0]\n"
+      "fmla v10.8h, v16.8h, v0.h[0]\n"
+      "ldr q16, [x9, #0x0]\n"
+      "fmla v11.8h, v16.8h, v0.h[0]\n"
       "add x12, x12, #0x10\n"
-      "ldr q16, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "fmla v8.8h, v17.8h, v0.h[0]\n"
-      "ldr q17, [x10, #0x0]\n"
       "add x10, x10, #0x10\n"
-      "fmla v9.8h, v16.8h, v0.h[0]\n"
-      "ldr q16, [x9, #0x0]\n"
       "add x9, x9, #0x10\n"
-      "fmla v10.8h, v17.8h, v0.h[0]\n"
-      "fmla v11.8h, v16.8h, v0.h[0]\n"
       "cbnz x27, 30b\n"
       "31:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -453,9 +451,9 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 24b\n"
       "tbz %x[flags], #1, 32f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.8h }, [x21]\n"
       "ld1r { v16.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v17.8h\n"
       "fmin v9.8h, v9.8h, v17.8h\n"
@@ -574,19 +572,19 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "b 302f\n"
       "51:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "52:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x18\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x18\n"
       "bgt 53f\n"
       "cmp x14, #0x10\n"
       "mov x9, x12\n"
@@ -599,159 +597,159 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cbz x15, 54f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "add x15, x15, #0x40\n"
       "b 73f\n"
       "54:"  // Height 2: no bias
       "tbz %x[flags], #0, 72f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x14, #0x20\n"
-      "add x26, x13, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "bge 71f\n"
       "tbz x14, #4, 62f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
       "ld1 { v9.8h }, [x13], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
       "tbz x14, #3, 58f\n"
       "ld1 { v10.8h }, [x13], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
       "tbz x14, #2, 56f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "tbz x14, #1, 55f\n"
       "ld1 { v11.s }[2], [x13], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v11.h }[6], [x13]\n"
-      "ld1 { v15.h }[6], [x26]\n"
+      "ld1 { v15.h }[6], [x25]\n"
       "b 70f\n"
       "55:"  // Height 2: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v11.h }[4], [x13]\n"
-      "ld1 { v15.h }[4], [x26]\n"
+      "ld1 { v15.h }[4], [x25]\n"
       "b 70f\n"
       "56:"  // Height 2: Partial accumulate: partial_2_24
       "tbz x14, #1, 57f\n"
       "ldr s11, [x13], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v11.h }[2], [x13]\n"
-      "ld1 { v15.h }[2], [x26]\n"
+      "ld1 { v15.h }[2], [x25]\n"
       "b 70f\n"
       "57:"  // Height 2: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x14, #0, 70f\n"
       "ldr h11, [x13, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
       "b 70f\n"
       "58:"  // Height 2: Partial accumulate: partial_4_16
       "tbz x14, #2, 60f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "tbz x14, #1, 59f\n"
       "ld1 { v10.s }[2], [x13], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v10.h }[6], [x13]\n"
-      "ld1 { v14.h }[6], [x26]\n"
+      "ld1 { v14.h }[6], [x25]\n"
       "b 70f\n"
       "59:"  // Height 2: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v10.h }[4], [x13]\n"
-      "ld1 { v14.h }[4], [x26]\n"
+      "ld1 { v14.h }[4], [x25]\n"
       "b 70f\n"
       "60:"  // Height 2: Partial accumulate: partial_2_16
       "tbz x14, #1, 61f\n"
       "ldr s10, [x13], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v10.h }[2], [x13]\n"
-      "ld1 { v14.h }[2], [x26]\n"
+      "ld1 { v14.h }[2], [x25]\n"
       "b 70f\n"
       "61:"  // Height 2: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x14, #0, 70f\n"
       "ldr h10, [x13, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
       "b 70f\n"
       "62:"  // Height 2: Partial accumulate: partial_8_0
       "tbz x14, #3, 66f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
       "tbz x14, #2, 64f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "tbz x14, #1, 63f\n"
       "ld1 { v9.s }[2], [x13], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v9.h }[6], [x13]\n"
-      "ld1 { v13.h }[6], [x26]\n"
+      "ld1 { v13.h }[6], [x25]\n"
       "b 70f\n"
       "63:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v9.h }[4], [x13]\n"
-      "ld1 { v13.h }[4], [x26]\n"
+      "ld1 { v13.h }[4], [x25]\n"
       "b 70f\n"
       "64:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x14, #1, 65f\n"
       "ldr s9, [x13], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v9.h }[2], [x13]\n"
-      "ld1 { v13.h }[2], [x26]\n"
+      "ld1 { v13.h }[2], [x25]\n"
       "b 70f\n"
       "65:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x14, #0, 70f\n"
       "ldr h9, [x13, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
       "b 70f\n"
       "66:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x14, #2, 68f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "tbz x14, #1, 67f\n"
       "ld1 { v8.s }[2], [x13], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v8.h }[6], [x13]\n"
-      "ld1 { v12.h }[6], [x26]\n"
+      "ld1 { v12.h }[6], [x25]\n"
       "b 70f\n"
       "67:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v8.h }[4], [x13]\n"
-      "ld1 { v12.h }[4], [x26]\n"
+      "ld1 { v12.h }[4], [x25]\n"
       "b 70f\n"
       "68:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x14, #1, 69f\n"
       "ldr s8, [x13], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
       "tbz x14, #0, 70f\n"
       "ld1 { v8.h }[2], [x13]\n"
-      "ld1 { v12.h }[2], [x26]\n"
+      "ld1 { v12.h }[2], [x25]\n"
       "b 70f\n"
       "69:"  // Height 2: Partial accumulate: partial_1_0
       "ldr h8, [x13, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
       "70:"  // Height 2: Partial accumulate: Done
       "sub x13, x13, x20\n"
@@ -761,10 +759,10 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "b 73f\n"
       "72:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -779,8 +777,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "74:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 75f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -811,15 +809,15 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "ldr q16, [x9, #0x0]\n"
-      "add x26, x26, #0x10\n"
       "cmp x27, #0x10\n"
-      "add x25, x25, #0x10\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "fmla v14.8h, v17.8h, v1.h[0]\n"
       "ldr q17, [x12, #0x10]\n"
+      "add x26, x26, #0x10\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
       "fmla v15.8h, v16.8h, v1.h[0]\n"
       "ldr q16, [x11, #0x10]\n"
+      "add x25, x25, #0x10\n"
       "fmla v8.8h, v17.8h, v0.h[1]\n"
       "fmla v12.8h, v17.8h, v1.h[1]\n"
       "ldr q17, [x10, #0x10]\n"
@@ -920,10 +918,10 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "ldr q16, [x9, #0x0]\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "fmla v14.8h, v17.8h, v1.h[0]\n"
       "ldr q17, [x12, #0x10]\n"
+      "add x25, x25, #0x10\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
       "fmla v15.8h, v16.8h, v1.h[0]\n"
       "ldr q16, [x11, #0x10]\n"
@@ -1021,20 +1019,20 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "sub x27, x27, #0x1\n"
       "ldr q17, [x12, #0x0]\n"
       "ldr q16, [x11, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.8h, v17.8h, v1.h[0]\n"
       "fmla v12.8h, v17.8h, v0.h[0]\n"
       "ldr q17, [x10, #0x0]\n"
-      "add x10, x10, #0x10\n"
       "fmla v9.8h, v16.8h, v1.h[0]\n"
       "fmla v13.8h, v16.8h, v0.h[0]\n"
       "ldr q16, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
       "fmla v10.8h, v17.8h, v1.h[0]\n"
       "fmla v14.8h, v17.8h, v0.h[0]\n"
+      "add x12, x12, #0x10\n"
+      "add x11, x11, #0x10\n"
       "fmla v11.8h, v16.8h, v1.h[0]\n"
       "fmla v15.8h, v16.8h, v0.h[0]\n"
+      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       "cbnz x27, 80b\n"
       "81:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -1042,11 +1040,11 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 74b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "tbz %x[flags], #1, 82f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.8h }, [x21]\n"
       "ld1r { v16.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v17.8h\n"
       "fmin v9.8h, v9.8h, v17.8h\n"
@@ -1070,127 +1068,127 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "tbz x14, #4, 90f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
       "st1 { v9.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
       "tbz x14, #3, 86f\n"
       "st1 { v10.8h }, [x13], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
       "tbz x14, #2, 84f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
       "tbz x14, #1, 83f\n"
       "st1 { v11.s }[2], [x13], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
       "tbz x14, #0, 98f\n"
       "st1 { v11.h }[6], [x13]\n"
-      "st1 { v15.h }[6], [x26]\n"
+      "st1 { v15.h }[6], [x25]\n"
       "b 98f\n"
       "83:"  // Height 2: Partial direct writeback: partial_1_28
       "tbz x14, #0, 98f\n"
       "st1 { v11.h }[4], [x13]\n"
-      "st1 { v15.h }[4], [x26]\n"
+      "st1 { v15.h }[4], [x25]\n"
       "b 98f\n"
       "84:"  // Height 2: Partial direct writeback: partial_2_24
       "tbz x14, #1, 85f\n"
       "str s11, [x13], #0x4\n"
-      "str s15, [x26], #0x4\n"
+      "str s15, [x25], #0x4\n"
       "tbz x14, #0, 98f\n"
       "st1 { v11.h }[2], [x13]\n"
-      "st1 { v15.h }[2], [x26]\n"
+      "st1 { v15.h }[2], [x25]\n"
       "b 98f\n"
       "85:"  // Height 2: Partial direct writeback: partial_1_24
       "tbz x14, #0, 98f\n"
       "str h11, [x13, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
       "b 98f\n"
       "86:"  // Height 2: Partial direct writeback: partial_4_16
       "tbz x14, #2, 88f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
+      "str d14, [x25], #0x8\n"
       "tbz x14, #1, 87f\n"
       "st1 { v10.s }[2], [x13], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
       "tbz x14, #0, 98f\n"
       "st1 { v10.h }[6], [x13]\n"
-      "st1 { v14.h }[6], [x26]\n"
+      "st1 { v14.h }[6], [x25]\n"
       "b 98f\n"
       "87:"  // Height 2: Partial direct writeback: partial_1_20
       "tbz x14, #0, 98f\n"
       "st1 { v10.h }[4], [x13]\n"
-      "st1 { v14.h }[4], [x26]\n"
+      "st1 { v14.h }[4], [x25]\n"
       "b 98f\n"
       "88:"  // Height 2: Partial direct writeback: partial_2_16
       "tbz x14, #1, 89f\n"
       "str s10, [x13], #0x4\n"
-      "str s14, [x26], #0x4\n"
+      "str s14, [x25], #0x4\n"
       "tbz x14, #0, 98f\n"
       "st1 { v10.h }[2], [x13]\n"
-      "st1 { v14.h }[2], [x26]\n"
+      "st1 { v14.h }[2], [x25]\n"
       "b 98f\n"
       "89:"  // Height 2: Partial direct writeback: partial_1_16
       "tbz x14, #0, 98f\n"
       "str h10, [x13, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
       "b 98f\n"
       "90:"  // Height 2: Partial direct writeback: partial_8_0
       "tbz x14, #3, 94f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
       "tbz x14, #2, 92f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
+      "str d13, [x25], #0x8\n"
       "tbz x14, #1, 91f\n"
       "st1 { v9.s }[2], [x13], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
       "tbz x14, #0, 98f\n"
       "st1 { v9.h }[6], [x13]\n"
-      "st1 { v13.h }[6], [x26]\n"
+      "st1 { v13.h }[6], [x25]\n"
       "b 98f\n"
       "91:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x14, #0, 98f\n"
       "st1 { v9.h }[4], [x13]\n"
-      "st1 { v13.h }[4], [x26]\n"
+      "st1 { v13.h }[4], [x25]\n"
       "b 98f\n"
       "92:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x14, #1, 93f\n"
       "str s9, [x13], #0x4\n"
-      "str s13, [x26], #0x4\n"
+      "str s13, [x25], #0x4\n"
       "tbz x14, #0, 98f\n"
       "st1 { v9.h }[2], [x13]\n"
-      "st1 { v13.h }[2], [x26]\n"
+      "st1 { v13.h }[2], [x25]\n"
       "b 98f\n"
       "93:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x14, #0, 98f\n"
       "str h9, [x13, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
       "b 98f\n"
       "94:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x14, #2, 96f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
+      "str d12, [x25], #0x8\n"
       "tbz x14, #1, 95f\n"
       "st1 { v8.s }[2], [x13], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
       "tbz x14, #0, 98f\n"
       "st1 { v8.h }[6], [x13]\n"
-      "st1 { v12.h }[6], [x26]\n"
+      "st1 { v12.h }[6], [x25]\n"
       "b 98f\n"
       "95:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x14, #0, 98f\n"
       "st1 { v8.h }[4], [x13]\n"
-      "st1 { v12.h }[4], [x26]\n"
+      "st1 { v12.h }[4], [x25]\n"
       "b 98f\n"
       "96:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x14, #1, 97f\n"
       "str s8, [x13], #0x4\n"
-      "str s12, [x26], #0x4\n"
+      "str s12, [x25], #0x4\n"
       "tbz x14, #0, 98f\n"
       "st1 { v8.h }[2], [x13]\n"
-      "st1 { v12.h }[2], [x26]\n"
+      "st1 { v12.h }[2], [x25]\n"
       "b 98f\n"
       "97:"  // Height 2: Partial direct writeback: partial_1_0
       "str h8, [x13, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
       "98:"  // Height 2: Partial direct writeback: Done
       "b 100f\n"
       "99:"  // Height 2: Full writeback
@@ -1199,29 +1197,29 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
       "100:"  // Height 2: Writeback done
       "subs x14, x14, #0x20\n"
       "bgt 52b\n"
       "b 302f\n"
       "101:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "102:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x18\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x18\n"
       "bgt 103f\n"
       "cmp x14, #0x10\n"
       "mov x9, x12\n"
@@ -1234,197 +1232,197 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cbz x15, 104f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x15, x15, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
       "b 123f\n"
       "104:"  // Height 3: no bias
       "tbz %x[flags], #0, 122f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #1\n"
       "cmp x14, #0x20\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "bge 121f\n"
       "tbz x14, #4, 112f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
       "ld1 { v9.8h }, [x13], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
       "tbz x14, #3, 108f\n"
       "ld1 { v10.8h }, [x13], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
       "tbz x14, #2, 106f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x14, #1, 105f\n"
       "ld1 { v11.s }[2], [x13], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v11.h }[6], [x13]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
       "b 120f\n"
       "105:"  // Height 3: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v11.h }[4], [x13]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
       "b 120f\n"
       "106:"  // Height 3: Partial accumulate: partial_2_24
       "tbz x14, #1, 107f\n"
       "ldr s11, [x13], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s19, [x25], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v11.h }[2], [x13]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
       "b 120f\n"
       "107:"  // Height 3: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x14, #0, 120f\n"
       "ldr h11, [x13, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
       "b 120f\n"
       "108:"  // Height 3: Partial accumulate: partial_4_16
       "tbz x14, #2, 110f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x14, #1, 109f\n"
       "ld1 { v10.s }[2], [x13], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v10.h }[6], [x13]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
       "b 120f\n"
       "109:"  // Height 3: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v10.h }[4], [x13]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
       "b 120f\n"
       "110:"  // Height 3: Partial accumulate: partial_2_16
       "tbz x14, #1, 111f\n"
       "ldr s10, [x13], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s18, [x25], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v10.h }[2], [x13]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
       "b 120f\n"
       "111:"  // Height 3: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x14, #0, 120f\n"
       "ldr h10, [x13, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
       "b 120f\n"
       "112:"  // Height 3: Partial accumulate: partial_8_0
       "tbz x14, #3, 116f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
       "tbz x14, #2, 114f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x14, #1, 113f\n"
       "ld1 { v9.s }[2], [x13], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v9.h }[6], [x13]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
       "b 120f\n"
       "113:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v9.h }[4], [x13]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
       "b 120f\n"
       "114:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x14, #1, 115f\n"
       "ldr s9, [x13], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s17, [x25], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v9.h }[2], [x13]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
       "b 120f\n"
       "115:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x14, #0, 120f\n"
       "ldr h9, [x13, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
       "b 120f\n"
       "116:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x14, #2, 118f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
       "tbz x14, #1, 117f\n"
       "ld1 { v8.s }[2], [x13], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v8.h }[6], [x13]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
       "b 120f\n"
       "117:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v8.h }[4], [x13]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
       "b 120f\n"
       "118:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x14, #1, 119f\n"
       "ldr s8, [x13], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s16, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
       "tbz x14, #0, 120f\n"
       "ld1 { v8.h }[2], [x13]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
       "b 120f\n"
       "119:"  // Height 3: Partial accumulate: partial_1_0
       "ldr h8, [x13, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h16, [x25, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
       "120:"  // Height 3: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 123f\n"
@@ -1433,14 +1431,14 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
       "b 123f\n"
       "122:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1459,8 +1457,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "124:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 125f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1491,18 +1489,18 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v8.8h, v6.8h, v0.h[0]\n"
       "fmla v12.8h, v6.8h, v1.h[0]\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
+      "cmp x27, #0x10\n"
       "fmla v16.8h, v6.8h, v2.h[0]\n"
       "ldr q21, [x10, #0x0]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
-      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "fmla v17.8h, v7.8h, v2.h[0]\n"
       "ldr q20, [x9, #0x0]\n"
       "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
       "fmla v10.8h, v21.8h, v0.h[0]\n"
       "fmla v14.8h, v21.8h, v1.h[0]\n"
+      "add x24, x24, #0x10\n"
       "fmla v18.8h, v21.8h, v2.h[0]\n"
       "ldr q21, [x12, #0x10]\n"
       "fmla v11.8h, v20.8h, v0.h[0]\n"
@@ -1601,8 +1599,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v14.8h, v21.8h, v1.h[6]\n"
       "fmla v18.8h, v21.8h, v2.h[6]\n"
       "ldr q21, [x12, #0x70]\n"
-      "add x12, x12, #0x80\n"
       "fmla v11.8h, v20.8h, v0.h[6]\n"
+      "add x12, x12, #0x80\n"
       "fmla v15.8h, v20.8h, v1.h[6]\n"
       "fmla v19.8h, v20.8h, v2.h[6]\n"
       "ldr q20, [x11, #0x70]\n"
@@ -1611,8 +1609,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v12.8h, v21.8h, v1.h[7]\n"
       "fmla v16.8h, v21.8h, v2.h[7]\n"
       "ldr q21, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       "fmla v9.8h, v20.8h, v0.h[7]\n"
+      "add x10, x10, #0x80\n"
       "fmla v13.8h, v20.8h, v1.h[7]\n"
       "fmla v17.8h, v20.8h, v2.h[7]\n"
       "ldr q20, [x9, #0x70]\n"
@@ -1742,8 +1740,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v14.8h, v21.8h, v1.h[6]\n"
       "fmla v18.8h, v21.8h, v2.h[6]\n"
       "ldr q21, [x12, #0x70]\n"
-      "add x12, x12, #0x80\n"
       "fmla v11.8h, v20.8h, v0.h[6]\n"
+      "add x12, x12, #0x80\n"
       "fmla v15.8h, v20.8h, v1.h[6]\n"
       "fmla v19.8h, v20.8h, v2.h[6]\n"
       "ldr q20, [x11, #0x70]\n"
@@ -1752,8 +1750,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v12.8h, v21.8h, v1.h[7]\n"
       "fmla v16.8h, v21.8h, v2.h[7]\n"
       "ldr q21, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       "fmla v9.8h, v20.8h, v0.h[7]\n"
+      "add x10, x10, #0x80\n"
       "fmla v13.8h, v20.8h, v1.h[7]\n"
       "fmla v17.8h, v20.8h, v2.h[7]\n"
       "ldr q20, [x9, #0x70]\n"
@@ -1772,23 +1770,23 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "sub x27, x27, #0x1\n"
       "ldr h0, [x24], #0x2\n"
       "ldr q21, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "ldr q20, [x11, #0x0]\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.8h, v21.8h, v2.h[0]\n"
       "fmla v12.8h, v21.8h, v1.h[0]\n"
+      "ldr q20, [x11, #0x0]\n"
       "fmla v16.8h, v21.8h, v0.h[0]\n"
       "ldr q21, [x10, #0x0]\n"
-      "add x10, x10, #0x10\n"
       "fmla v9.8h, v20.8h, v2.h[0]\n"
       "fmla v13.8h, v20.8h, v1.h[0]\n"
       "fmla v17.8h, v20.8h, v0.h[0]\n"
       "ldr q20, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
+      "add x12, x12, #0x10\n"
       "fmla v10.8h, v21.8h, v2.h[0]\n"
       "fmla v14.8h, v21.8h, v1.h[0]\n"
+      "add x11, x11, #0x10\n"
+      "add x10, x10, #0x10\n"
       "fmla v18.8h, v21.8h, v0.h[0]\n"
       "fmla v11.8h, v20.8h, v2.h[0]\n"
+      "add x9, x9, #0x10\n"
       "fmla v15.8h, v20.8h, v1.h[0]\n"
       "fmla v19.8h, v20.8h, v0.h[0]\n"
       "cbnz x27, 130b\n"
@@ -1798,12 +1796,12 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 124b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "tbz %x[flags], #1, 132f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v21.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.8h }, [x21]\n"
       "ld1r { v20.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v21.8h\n"
       "fmin v9.8h, v9.8h, v21.8h\n"
@@ -1835,159 +1833,159 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "tbz x14, #4, 140f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
       "st1 { v9.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
       "tbz x14, #3, 136f\n"
       "st1 { v10.8h }, [x13], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
       "tbz x14, #2, 134f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x14, #1, 133f\n"
       "st1 { v11.s }[2], [x13], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
       "tbz x14, #0, 148f\n"
       "st1 { v11.h }[6], [x13]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
       "b 148f\n"
       "133:"  // Height 3: Partial direct writeback: partial_1_28
       "tbz x14, #0, 148f\n"
       "st1 { v11.h }[4], [x13]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
       "b 148f\n"
       "134:"  // Height 3: Partial direct writeback: partial_2_24
       "tbz x14, #1, 135f\n"
       "str s11, [x13], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
       "tbz x14, #0, 148f\n"
       "st1 { v11.h }[2], [x13]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
       "b 148f\n"
       "135:"  // Height 3: Partial direct writeback: partial_1_24
       "tbz x14, #0, 148f\n"
       "str h11, [x13, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
       "b 148f\n"
       "136:"  // Height 3: Partial direct writeback: partial_4_16
       "tbz x14, #2, 138f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x14, #1, 137f\n"
       "st1 { v10.s }[2], [x13], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
       "tbz x14, #0, 148f\n"
       "st1 { v10.h }[6], [x13]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
       "b 148f\n"
       "137:"  // Height 3: Partial direct writeback: partial_1_20
       "tbz x14, #0, 148f\n"
       "st1 { v10.h }[4], [x13]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
       "b 148f\n"
       "138:"  // Height 3: Partial direct writeback: partial_2_16
       "tbz x14, #1, 139f\n"
       "str s10, [x13], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
       "tbz x14, #0, 148f\n"
       "st1 { v10.h }[2], [x13]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
       "b 148f\n"
       "139:"  // Height 3: Partial direct writeback: partial_1_16
       "tbz x14, #0, 148f\n"
       "str h10, [x13, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
       "b 148f\n"
       "140:"  // Height 3: Partial direct writeback: partial_8_0
       "tbz x14, #3, 144f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
       "tbz x14, #2, 142f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x14, #1, 141f\n"
       "st1 { v9.s }[2], [x13], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
       "tbz x14, #0, 148f\n"
       "st1 { v9.h }[6], [x13]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
       "b 148f\n"
       "141:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x14, #0, 148f\n"
       "st1 { v9.h }[4], [x13]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
       "b 148f\n"
       "142:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x14, #1, 143f\n"
       "str s9, [x13], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
       "tbz x14, #0, 148f\n"
       "st1 { v9.h }[2], [x13]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
       "b 148f\n"
       "143:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x14, #0, 148f\n"
       "str h9, [x13, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
       "b 148f\n"
       "144:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x14, #2, 146f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x14, #1, 145f\n"
       "st1 { v8.s }[2], [x13], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
       "tbz x14, #0, 148f\n"
       "st1 { v8.h }[6], [x13]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
       "b 148f\n"
       "145:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x14, #0, 148f\n"
       "st1 { v8.h }[4], [x13]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
       "b 148f\n"
       "146:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x14, #1, 147f\n"
       "str s8, [x13], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
       "tbz x14, #0, 148f\n"
       "st1 { v8.h }[2], [x13]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
       "b 148f\n"
       "147:"  // Height 3: Partial direct writeback: partial_1_0
       "str h8, [x13, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
       "148:"  // Height 3: Partial direct writeback: Done
       "b 150f\n"
       "149:"  // Height 3: Full writeback
@@ -1996,33 +1994,33 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "150:"  // Height 3: Writeback done
       "subs x14, x14, #0x20\n"
       "bgt 102b\n"
       "b 302f\n"
       "151:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "152:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x18\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x18\n"
       "bgt 153f\n"
       "cmp x14, #0x10\n"
       "mov x9, x12\n"
@@ -2035,18 +2033,18 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cbz x15, 154f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x15, x15, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -2054,215 +2052,215 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "154:"  // Height 4: no bias
       "tbz %x[flags], #0, 172f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x20\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "cmp x14, #0x20\n"
+      "add x23, x24, x20, LSL #1\n"
       "bge 171f\n"
       "tbz x14, #4, 162f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
       "ld1 { v9.8h }, [x13], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
       "tbz x14, #3, 158f\n"
       "ld1 { v10.8h }, [x13], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
       "tbz x14, #2, 156f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x14, #1, 155f\n"
       "ld1 { v11.s }[2], [x13], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v11.h }[6], [x13]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
       "b 170f\n"
       "155:"  // Height 4: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v11.h }[4], [x13]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
       "b 170f\n"
       "156:"  // Height 4: Partial accumulate: partial_2_24
       "tbz x14, #1, 157f\n"
       "ldr s11, [x13], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v11.h }[2], [x13]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
       "b 170f\n"
       "157:"  // Height 4: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x14, #0, 170f\n"
       "ldr h11, [x13, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
       "b 170f\n"
       "158:"  // Height 4: Partial accumulate: partial_4_16
       "tbz x14, #2, 160f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x14, #1, 159f\n"
       "ld1 { v10.s }[2], [x13], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v10.h }[6], [x13]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
       "b 170f\n"
       "159:"  // Height 4: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v10.h }[4], [x13]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
       "b 170f\n"
       "160:"  // Height 4: Partial accumulate: partial_2_16
       "tbz x14, #1, 161f\n"
       "ldr s10, [x13], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v10.h }[2], [x13]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
       "b 170f\n"
       "161:"  // Height 4: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x14, #0, 170f\n"
       "ldr h10, [x13, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
       "b 170f\n"
       "162:"  // Height 4: Partial accumulate: partial_8_0
       "tbz x14, #3, 166f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
       "tbz x14, #2, 164f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x14, #1, 163f\n"
       "ld1 { v9.s }[2], [x13], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v9.h }[6], [x13]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
       "b 170f\n"
       "163:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v9.h }[4], [x13]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
       "b 170f\n"
       "164:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x14, #1, 165f\n"
       "ldr s9, [x13], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v9.h }[2], [x13]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
       "b 170f\n"
       "165:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x14, #0, 170f\n"
       "ldr h9, [x13, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
       "b 170f\n"
       "166:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x14, #2, 168f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x14, #1, 167f\n"
       "ld1 { v8.s }[2], [x13], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v8.h }[6], [x13]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
       "b 170f\n"
       "167:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v8.h }[4], [x13]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
       "b 170f\n"
       "168:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x14, #1, 169f\n"
       "ldr s8, [x13], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
       "tbz x14, #0, 170f\n"
       "ld1 { v8.h }[2], [x13]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
       "b 170f\n"
       "169:"  // Height 4: Partial accumulate: partial_1_0
       "ldr h8, [x13, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
       "170:"  // Height 4: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 173f\n"
@@ -2271,18 +2269,18 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "b 173f\n"
       "172:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2305,8 +2303,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "174:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 175f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2341,11 +2339,11 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v8.8h, v6.8h, v0.h[0]\n"
       "fmla v12.8h, v6.8h, v1.h[0]\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
+      "cmp x27, #0x10\n"
       "fmla v16.8h, v6.8h, v2.h[0]\n"
       "fmla v20.8h, v6.8h, v3.h[0]\n"
       "ldr q25, [x10, #0x0]\n"
-      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "add x25, x25, #0x10\n"
@@ -2691,16 +2689,16 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr h0, [x23], #0x2\n"
       "ldr q25, [x12, #0x0]\n"
       "ldr q24, [x11, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.8h, v25.8h, v3.h[0]\n"
       "fmla v12.8h, v25.8h, v2.h[0]\n"
       "fmla v16.8h, v25.8h, v1.h[0]\n"
       "fmla v20.8h, v25.8h, v0.h[0]\n"
       "ldr q25, [x10, #0x0]\n"
-      "add x10, x10, #0x10\n"
+      "add x12, x12, #0x10\n"
       "fmla v9.8h, v24.8h, v3.h[0]\n"
       "fmla v13.8h, v24.8h, v2.h[0]\n"
+      "add x11, x11, #0x10\n"
+      "add x10, x10, #0x10\n"
       "fmla v17.8h, v24.8h, v1.h[0]\n"
       "fmla v21.8h, v24.8h, v0.h[0]\n"
       "ldr q24, [x9, #0x0]\n"
@@ -2720,13 +2718,13 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 174b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "tbz %x[flags], #1, 182f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v25.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v25.8h }, [x21]\n"
       "ld1r { v24.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v25.8h\n"
       "fmin v9.8h, v9.8h, v25.8h\n"
@@ -2766,191 +2764,191 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "tbz x14, #4, 190f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
       "st1 { v9.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
       "tbz x14, #3, 186f\n"
       "st1 { v10.8h }, [x13], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
       "tbz x14, #2, 184f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
       "tbz x14, #1, 183f\n"
       "st1 { v11.s }[2], [x13], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
       "tbz x14, #0, 198f\n"
       "st1 { v11.h }[6], [x13]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
       "b 198f\n"
       "183:"  // Height 4: Partial direct writeback: partial_1_28
       "tbz x14, #0, 198f\n"
       "st1 { v11.h }[4], [x13]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
       "b 198f\n"
       "184:"  // Height 4: Partial direct writeback: partial_2_24
       "tbz x14, #1, 185f\n"
       "str s11, [x13], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
       "tbz x14, #0, 198f\n"
       "st1 { v11.h }[2], [x13]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
       "b 198f\n"
       "185:"  // Height 4: Partial direct writeback: partial_1_24
       "tbz x14, #0, 198f\n"
       "str h11, [x13, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
       "b 198f\n"
       "186:"  // Height 4: Partial direct writeback: partial_4_16
       "tbz x14, #2, 188f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
       "tbz x14, #1, 187f\n"
       "st1 { v10.s }[2], [x13], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
       "tbz x14, #0, 198f\n"
       "st1 { v10.h }[6], [x13]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
       "b 198f\n"
       "187:"  // Height 4: Partial direct writeback: partial_1_20
       "tbz x14, #0, 198f\n"
       "st1 { v10.h }[4], [x13]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
       "b 198f\n"
       "188:"  // Height 4: Partial direct writeback: partial_2_16
       "tbz x14, #1, 189f\n"
       "str s10, [x13], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
       "tbz x14, #0, 198f\n"
       "st1 { v10.h }[2], [x13]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
       "b 198f\n"
       "189:"  // Height 4: Partial direct writeback: partial_1_16
       "tbz x14, #0, 198f\n"
       "str h10, [x13, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
       "b 198f\n"
       "190:"  // Height 4: Partial direct writeback: partial_8_0
       "tbz x14, #3, 194f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
       "tbz x14, #2, 192f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
       "tbz x14, #1, 191f\n"
       "st1 { v9.s }[2], [x13], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
       "tbz x14, #0, 198f\n"
       "st1 { v9.h }[6], [x13]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
       "b 198f\n"
       "191:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x14, #0, 198f\n"
       "st1 { v9.h }[4], [x13]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
       "b 198f\n"
       "192:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x14, #1, 193f\n"
       "str s9, [x13], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
       "tbz x14, #0, 198f\n"
       "st1 { v9.h }[2], [x13]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
       "b 198f\n"
       "193:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x14, #0, 198f\n"
       "str h9, [x13, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
       "b 198f\n"
       "194:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x14, #2, 196f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x14, #1, 195f\n"
       "st1 { v8.s }[2], [x13], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
       "tbz x14, #0, 198f\n"
       "st1 { v8.h }[6], [x13]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
       "b 198f\n"
       "195:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x14, #0, 198f\n"
       "st1 { v8.h }[4], [x13]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
       "b 198f\n"
       "196:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x14, #1, 197f\n"
       "str s8, [x13], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
       "tbz x14, #0, 198f\n"
       "st1 { v8.h }[2], [x13]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
       "b 198f\n"
       "197:"  // Height 4: Partial direct writeback: partial_1_0
       "str h8, [x13, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
       "198:"  // Height 4: Partial direct writeback: Done
       "b 200f\n"
       "199:"  // Height 4: Full writeback
@@ -2959,37 +2957,37 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
       "200:"  // Height 4: Writeback done
       "subs x14, x14, #0x20\n"
       "bgt 152b\n"
       "b 302f\n"
       "201:"  // Height 5
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "202:"  // Height 5: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x18\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x18\n"
       "bgt 203f\n"
       "cmp x14, #0x10\n"
       "mov x9, x12\n"
@@ -3002,18 +3000,18 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cbz x15, 204f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x15, x15, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -3025,248 +3023,248 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "204:"  // Height 5: no bias
       "tbz %x[flags], #0, 222f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x20\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
+      "cmp x14, #0x20\n"
+      "add x22, x23, x20, LSL #1\n"
       "bge 221f\n"
       "tbz x14, #4, 212f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
       "ld1 { v9.8h }, [x13], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
-      "ld1 { v25.8h }, [x23], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
+      "ld1 { v25.8h }, [x22], #0x10\n"
       "tbz x14, #3, 208f\n"
       "ld1 { v10.8h }, [x13], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
-      "ld1 { v26.8h }, [x23], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
+      "ld1 { v26.8h }, [x22], #0x10\n"
       "tbz x14, #2, 206f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x14, #1, 205f\n"
       "ld1 { v11.s }[2], [x13], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
-      "ld1 { v27.s }[2], [x23], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
+      "ld1 { v27.s }[2], [x22], #0x4\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v11.h }[6], [x13]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
-      "ld1 { v27.h }[6], [x23]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
+      "ld1 { v27.h }[6], [x22]\n"
       "b 220f\n"
       "205:"  // Height 5: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v11.h }[4], [x13]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
-      "ld1 { v27.h }[4], [x23]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
+      "ld1 { v27.h }[4], [x22]\n"
       "b 220f\n"
       "206:"  // Height 5: Partial accumulate: partial_2_24
       "tbz x14, #1, 207f\n"
       "ldr s11, [x13], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
-      "ldr s27, [x23], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s27, [x22], #0x4\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v11.h }[2], [x13]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
-      "ld1 { v27.h }[2], [x23]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
+      "ld1 { v27.h }[2], [x22]\n"
       "b 220f\n"
       "207:"  // Height 5: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x14, #0, 220f\n"
       "ldr h11, [x13, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
-      "ldr h27, [x23, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
+      "ldr h27, [x22, #0x0]\n"
       "b 220f\n"
       "208:"  // Height 5: Partial accumulate: partial_4_16
       "tbz x14, #2, 210f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x14, #1, 209f\n"
       "ld1 { v10.s }[2], [x13], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
-      "ld1 { v26.s }[2], [x23], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
+      "ld1 { v26.s }[2], [x22], #0x4\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v10.h }[6], [x13]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
-      "ld1 { v26.h }[6], [x23]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
+      "ld1 { v26.h }[6], [x22]\n"
       "b 220f\n"
       "209:"  // Height 5: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v10.h }[4], [x13]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
-      "ld1 { v26.h }[4], [x23]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
+      "ld1 { v26.h }[4], [x22]\n"
       "b 220f\n"
       "210:"  // Height 5: Partial accumulate: partial_2_16
       "tbz x14, #1, 211f\n"
       "ldr s10, [x13], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr s26, [x23], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
+      "ldr s26, [x22], #0x4\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v10.h }[2], [x13]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
-      "ld1 { v26.h }[2], [x23]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
+      "ld1 { v26.h }[2], [x22]\n"
       "b 220f\n"
       "211:"  // Height 5: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x14, #0, 220f\n"
       "ldr h10, [x13, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
-      "ldr h26, [x23, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
+      "ldr h26, [x22, #0x0]\n"
       "b 220f\n"
       "212:"  // Height 5: Partial accumulate: partial_8_0
       "tbz x14, #3, 216f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
       "tbz x14, #2, 214f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x14, #1, 213f\n"
       "ld1 { v9.s }[2], [x13], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
-      "ld1 { v25.s }[2], [x23], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
+      "ld1 { v25.s }[2], [x22], #0x4\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v9.h }[6], [x13]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
-      "ld1 { v25.h }[6], [x23]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
+      "ld1 { v25.h }[6], [x22]\n"
       "b 220f\n"
       "213:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v9.h }[4], [x13]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
-      "ld1 { v25.h }[4], [x23]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
+      "ld1 { v25.h }[4], [x22]\n"
       "b 220f\n"
       "214:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x14, #1, 215f\n"
       "ldr s9, [x13], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
-      "ldr s25, [x23], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
+      "ldr s25, [x22], #0x4\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v9.h }[2], [x13]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
-      "ld1 { v25.h }[2], [x23]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
+      "ld1 { v25.h }[2], [x22]\n"
       "b 220f\n"
       "215:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x14, #0, 220f\n"
       "ldr h9, [x13, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
-      "ldr h25, [x23, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
+      "ldr h25, [x22, #0x0]\n"
       "b 220f\n"
       "216:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x14, #2, 218f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x14, #1, 217f\n"
       "ld1 { v8.s }[2], [x13], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
-      "ld1 { v24.s }[2], [x23], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
+      "ld1 { v24.s }[2], [x22], #0x4\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v8.h }[6], [x13]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
-      "ld1 { v24.h }[6], [x23]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
+      "ld1 { v24.h }[6], [x22]\n"
       "b 220f\n"
       "217:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v8.h }[4], [x13]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
-      "ld1 { v24.h }[4], [x23]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
+      "ld1 { v24.h }[4], [x22]\n"
       "b 220f\n"
       "218:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x14, #1, 219f\n"
       "ldr s8, [x13], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
+      "ldr s24, [x22], #0x4\n"
       "tbz x14, #0, 220f\n"
       "ld1 { v8.h }[2], [x13]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
-      "ld1 { v24.h }[2], [x23]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
+      "ld1 { v24.h }[2], [x22]\n"
       "b 220f\n"
       "219:"  // Height 5: Partial accumulate: partial_1_0
       "ldr h8, [x13, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
-      "ldr h24, [x23, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
+      "ldr h24, [x22, #0x0]\n"
       "220:"  // Height 5: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 223f\n"
@@ -3275,22 +3273,22 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
       "b 223f\n"
       "222:"  // Height 5: no accumulate
       "movi v8.16b, #0x0\n"
@@ -3317,8 +3315,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "224:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 225f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -3357,10 +3355,10 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v8.8h, v6.8h, v0.h[0]\n"
       "fmla v12.8h, v6.8h, v1.h[0]\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
+      "cmp x27, #0x10\n"
       "fmla v16.8h, v6.8h, v2.h[0]\n"
       "fmla v20.8h, v6.8h, v3.h[0]\n"
-      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "add x25, x25, #0x10\n"
       "fmla v24.8h, v6.8h, v4.h[0]\n"
       "ldr q29, [x10, #0x0]\n"
@@ -3774,19 +3772,19 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr h1, [x23], #0x2\n"
       "ldr h0, [x22], #0x2\n"
       "ldr q29, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "ldr q28, [x11, #0x0]\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.8h, v29.8h, v4.h[0]\n"
       "fmla v12.8h, v29.8h, v3.h[0]\n"
+      "ldr q28, [x11, #0x0]\n"
       "fmla v16.8h, v29.8h, v2.h[0]\n"
       "fmla v20.8h, v29.8h, v1.h[0]\n"
+      "add x12, x12, #0x10\n"
       "fmla v24.8h, v29.8h, v0.h[0]\n"
       "ldr q29, [x10, #0x0]\n"
-      "add x10, x10, #0x10\n"
       "fmla v9.8h, v28.8h, v4.h[0]\n"
+      "add x11, x11, #0x10\n"
       "fmla v13.8h, v28.8h, v3.h[0]\n"
       "fmla v17.8h, v28.8h, v2.h[0]\n"
+      "add x10, x10, #0x10\n"
       "fmla v21.8h, v28.8h, v1.h[0]\n"
       "fmla v25.8h, v28.8h, v0.h[0]\n"
       "ldr q28, [x9, #0x0]\n"
@@ -3808,14 +3806,14 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 224b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "tbz %x[flags], #1, 232f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v29.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v29.8h }, [x21]\n"
       "ld1r { v28.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v29.8h\n"
       "fmin v9.8h, v9.8h, v29.8h\n"
@@ -3863,223 +3861,223 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "tbz x14, #4, 240f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
       "st1 { v9.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v25.8h }, [x23], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v25.8h }, [x22], #0x10\n"
       "tbz x14, #3, 236f\n"
       "st1 { v10.8h }, [x13], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
-      "st1 { v26.8h }, [x23], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
+      "st1 { v26.8h }, [x22], #0x10\n"
       "tbz x14, #2, 234f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x14, #1, 233f\n"
       "st1 { v11.s }[2], [x13], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
-      "st1 { v27.s }[2], [x23], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
+      "st1 { v27.s }[2], [x22], #0x4\n"
       "tbz x14, #0, 248f\n"
       "st1 { v11.h }[6], [x13]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
-      "st1 { v27.h }[6], [x23]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
+      "st1 { v27.h }[6], [x22]\n"
       "b 248f\n"
       "233:"  // Height 5: Partial direct writeback: partial_1_28
       "tbz x14, #0, 248f\n"
       "st1 { v11.h }[4], [x13]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
-      "st1 { v27.h }[4], [x23]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
+      "st1 { v27.h }[4], [x22]\n"
       "b 248f\n"
       "234:"  // Height 5: Partial direct writeback: partial_2_24
       "tbz x14, #1, 235f\n"
       "str s11, [x13], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
-      "str s27, [x23], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
+      "str s27, [x22], #0x4\n"
       "tbz x14, #0, 248f\n"
       "st1 { v11.h }[2], [x13]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
-      "st1 { v27.h }[2], [x23]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
+      "st1 { v27.h }[2], [x22]\n"
       "b 248f\n"
       "235:"  // Height 5: Partial direct writeback: partial_1_24
       "tbz x14, #0, 248f\n"
       "str h11, [x13, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
-      "str h27, [x23, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
+      "str h27, [x22, #0x0]\n"
       "b 248f\n"
       "236:"  // Height 5: Partial direct writeback: partial_4_16
       "tbz x14, #2, 238f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x14, #1, 237f\n"
       "st1 { v10.s }[2], [x13], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
-      "st1 { v26.s }[2], [x23], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
+      "st1 { v26.s }[2], [x22], #0x4\n"
       "tbz x14, #0, 248f\n"
       "st1 { v10.h }[6], [x13]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
-      "st1 { v26.h }[6], [x23]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
+      "st1 { v26.h }[6], [x22]\n"
       "b 248f\n"
       "237:"  // Height 5: Partial direct writeback: partial_1_20
       "tbz x14, #0, 248f\n"
       "st1 { v10.h }[4], [x13]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
-      "st1 { v26.h }[4], [x23]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
+      "st1 { v26.h }[4], [x22]\n"
       "b 248f\n"
       "238:"  // Height 5: Partial direct writeback: partial_2_16
       "tbz x14, #1, 239f\n"
       "str s10, [x13], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
-      "str s26, [x23], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
+      "str s26, [x22], #0x4\n"
       "tbz x14, #0, 248f\n"
       "st1 { v10.h }[2], [x13]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
-      "st1 { v26.h }[2], [x23]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
+      "st1 { v26.h }[2], [x22]\n"
       "b 248f\n"
       "239:"  // Height 5: Partial direct writeback: partial_1_16
       "tbz x14, #0, 248f\n"
       "str h10, [x13, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
-      "str h26, [x23, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
+      "str h26, [x22, #0x0]\n"
       "b 248f\n"
       "240:"  // Height 5: Partial direct writeback: partial_8_0
       "tbz x14, #3, 244f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
       "tbz x14, #2, 242f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x14, #1, 241f\n"
       "st1 { v9.s }[2], [x13], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
-      "st1 { v25.s }[2], [x23], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
+      "st1 { v25.s }[2], [x22], #0x4\n"
       "tbz x14, #0, 248f\n"
       "st1 { v9.h }[6], [x13]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
-      "st1 { v25.h }[6], [x23]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
+      "st1 { v25.h }[6], [x22]\n"
       "b 248f\n"
       "241:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x14, #0, 248f\n"
       "st1 { v9.h }[4], [x13]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
-      "st1 { v25.h }[4], [x23]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
+      "st1 { v25.h }[4], [x22]\n"
       "b 248f\n"
       "242:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x14, #1, 243f\n"
       "str s9, [x13], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
-      "str s25, [x23], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
+      "str s25, [x22], #0x4\n"
       "tbz x14, #0, 248f\n"
       "st1 { v9.h }[2], [x13]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
-      "st1 { v25.h }[2], [x23]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
+      "st1 { v25.h }[2], [x22]\n"
       "b 248f\n"
       "243:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x14, #0, 248f\n"
       "str h9, [x13, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
-      "str h25, [x23, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
+      "str h25, [x22, #0x0]\n"
       "b 248f\n"
       "244:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x14, #2, 246f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x14, #1, 245f\n"
       "st1 { v8.s }[2], [x13], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
       "tbz x14, #0, 248f\n"
       "st1 { v8.h }[6], [x13]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
-      "st1 { v24.h }[6], [x23]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
+      "st1 { v24.h }[6], [x22]\n"
       "b 248f\n"
       "245:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x14, #0, 248f\n"
       "st1 { v8.h }[4], [x13]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
-      "st1 { v24.h }[4], [x23]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
+      "st1 { v24.h }[4], [x22]\n"
       "b 248f\n"
       "246:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x14, #1, 247f\n"
       "str s8, [x13], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
       "tbz x14, #0, 248f\n"
       "st1 { v8.h }[2], [x13]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
-      "st1 { v24.h }[2], [x23]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
+      "st1 { v24.h }[2], [x22]\n"
       "b 248f\n"
       "247:"  // Height 5: Partial direct writeback: partial_1_0
       "str h8, [x13, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
-      "str h24, [x23, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
+      "str h24, [x22, #0x0]\n"
       "248:"  // Height 5: Partial direct writeback: Done
       "b 250f\n"
       "249:"  // Height 5: Full writeback
@@ -4088,45 +4086,44 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "250:"  // Height 5: Writeback done
       "subs x14, x14, #0x20\n"
       "bgt 202b\n"
       "b 302f\n"
       "251:"  // Height 6
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0xc\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0xc\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "252:"  // Height 6: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x18\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x18\n"
       "bgt 253f\n"
       "cmp x14, #0x10\n"
       "mov x9, x12\n"
@@ -4139,18 +4136,18 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cbz x15, 254f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x15, x15, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -4166,281 +4163,281 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "254:"  // Height 6: no bias
       "tbz %x[flags], #0, 272f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x20\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "cmp x14, #0x20\n"
+      "add x21, x22, x20, LSL #1\n"
       "bge 271f\n"
       "tbz x14, #4, 262f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
-      "ld1 { v28.8h }, [x22], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
+      "ld1 { v28.8h }, [x21], #0x10\n"
       "ld1 { v9.8h }, [x13], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
-      "ld1 { v25.8h }, [x23], #0x10\n"
-      "ld1 { v29.8h }, [x22], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
+      "ld1 { v25.8h }, [x22], #0x10\n"
+      "ld1 { v29.8h }, [x21], #0x10\n"
       "tbz x14, #3, 258f\n"
       "ld1 { v10.8h }, [x13], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
-      "ld1 { v26.8h }, [x23], #0x10\n"
-      "ld1 { v30.8h }, [x22], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
+      "ld1 { v26.8h }, [x22], #0x10\n"
+      "ld1 { v30.8h }, [x21], #0x10\n"
       "tbz x14, #2, 256f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x14, #1, 255f\n"
       "ld1 { v11.s }[2], [x13], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
-      "ld1 { v27.s }[2], [x23], #0x4\n"
-      "ld1 { v31.s }[2], [x22], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
+      "ld1 { v27.s }[2], [x22], #0x4\n"
+      "ld1 { v31.s }[2], [x21], #0x4\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v11.h }[6], [x13]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
-      "ld1 { v27.h }[6], [x23]\n"
-      "ld1 { v31.h }[6], [x22]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
+      "ld1 { v27.h }[6], [x22]\n"
+      "ld1 { v31.h }[6], [x21]\n"
       "b 270f\n"
       "255:"  // Height 6: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v11.h }[4], [x13]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
-      "ld1 { v27.h }[4], [x23]\n"
-      "ld1 { v31.h }[4], [x22]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
+      "ld1 { v27.h }[4], [x22]\n"
+      "ld1 { v31.h }[4], [x21]\n"
       "b 270f\n"
       "256:"  // Height 6: Partial accumulate: partial_2_24
       "tbz x14, #1, 257f\n"
       "ldr s11, [x13], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
-      "ldr s27, [x23], #0x4\n"
-      "ldr s31, [x22], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s27, [x22], #0x4\n"
+      "ldr s31, [x21], #0x4\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v11.h }[2], [x13]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
-      "ld1 { v27.h }[2], [x23]\n"
-      "ld1 { v31.h }[2], [x22]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
+      "ld1 { v27.h }[2], [x22]\n"
+      "ld1 { v31.h }[2], [x21]\n"
       "b 270f\n"
       "257:"  // Height 6: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x14, #0, 270f\n"
       "ldr h11, [x13, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
-      "ldr h27, [x23, #0x0]\n"
-      "ldr h31, [x22, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
+      "ldr h27, [x22, #0x0]\n"
+      "ldr h31, [x21, #0x0]\n"
       "b 270f\n"
       "258:"  // Height 6: Partial accumulate: partial_4_16
       "tbz x14, #2, 260f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x14, #1, 259f\n"
       "ld1 { v10.s }[2], [x13], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
-      "ld1 { v26.s }[2], [x23], #0x4\n"
-      "ld1 { v30.s }[2], [x22], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
+      "ld1 { v26.s }[2], [x22], #0x4\n"
+      "ld1 { v30.s }[2], [x21], #0x4\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v10.h }[6], [x13]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
-      "ld1 { v26.h }[6], [x23]\n"
-      "ld1 { v30.h }[6], [x22]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
+      "ld1 { v26.h }[6], [x22]\n"
+      "ld1 { v30.h }[6], [x21]\n"
       "b 270f\n"
       "259:"  // Height 6: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v10.h }[4], [x13]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
-      "ld1 { v26.h }[4], [x23]\n"
-      "ld1 { v30.h }[4], [x22]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
+      "ld1 { v26.h }[4], [x22]\n"
+      "ld1 { v30.h }[4], [x21]\n"
       "b 270f\n"
       "260:"  // Height 6: Partial accumulate: partial_2_16
       "tbz x14, #1, 261f\n"
       "ldr s10, [x13], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr s26, [x23], #0x4\n"
-      "ldr s30, [x22], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
+      "ldr s26, [x22], #0x4\n"
+      "ldr s30, [x21], #0x4\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v10.h }[2], [x13]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
-      "ld1 { v26.h }[2], [x23]\n"
-      "ld1 { v30.h }[2], [x22]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
+      "ld1 { v26.h }[2], [x22]\n"
+      "ld1 { v30.h }[2], [x21]\n"
       "b 270f\n"
       "261:"  // Height 6: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x14, #0, 270f\n"
       "ldr h10, [x13, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
-      "ldr h26, [x23, #0x0]\n"
-      "ldr h30, [x22, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
+      "ldr h26, [x22, #0x0]\n"
+      "ldr h30, [x21, #0x0]\n"
       "b 270f\n"
       "262:"  // Height 6: Partial accumulate: partial_8_0
       "tbz x14, #3, 266f\n"
       "ld1 { v8.8h }, [x13], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
-      "ld1 { v28.8h }, [x22], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
+      "ld1 { v28.8h }, [x21], #0x10\n"
       "tbz x14, #2, 264f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x14, #1, 263f\n"
       "ld1 { v9.s }[2], [x13], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
-      "ld1 { v25.s }[2], [x23], #0x4\n"
-      "ld1 { v29.s }[2], [x22], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
+      "ld1 { v25.s }[2], [x22], #0x4\n"
+      "ld1 { v29.s }[2], [x21], #0x4\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v9.h }[6], [x13]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
-      "ld1 { v25.h }[6], [x23]\n"
-      "ld1 { v29.h }[6], [x22]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
+      "ld1 { v25.h }[6], [x22]\n"
+      "ld1 { v29.h }[6], [x21]\n"
       "b 270f\n"
       "263:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v9.h }[4], [x13]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
-      "ld1 { v25.h }[4], [x23]\n"
-      "ld1 { v29.h }[4], [x22]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
+      "ld1 { v25.h }[4], [x22]\n"
+      "ld1 { v29.h }[4], [x21]\n"
       "b 270f\n"
       "264:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x14, #1, 265f\n"
       "ldr s9, [x13], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
-      "ldr s25, [x23], #0x4\n"
-      "ldr s29, [x22], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
+      "ldr s25, [x22], #0x4\n"
+      "ldr s29, [x21], #0x4\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v9.h }[2], [x13]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
-      "ld1 { v25.h }[2], [x23]\n"
-      "ld1 { v29.h }[2], [x22]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
+      "ld1 { v25.h }[2], [x22]\n"
+      "ld1 { v29.h }[2], [x21]\n"
       "b 270f\n"
       "265:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x14, #0, 270f\n"
       "ldr h9, [x13, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
-      "ldr h25, [x23, #0x0]\n"
-      "ldr h29, [x22, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
+      "ldr h25, [x22, #0x0]\n"
+      "ldr h29, [x21, #0x0]\n"
       "b 270f\n"
       "266:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x14, #2, 268f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x14, #1, 267f\n"
       "ld1 { v8.s }[2], [x13], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
-      "ld1 { v24.s }[2], [x23], #0x4\n"
-      "ld1 { v28.s }[2], [x22], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
+      "ld1 { v24.s }[2], [x22], #0x4\n"
+      "ld1 { v28.s }[2], [x21], #0x4\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v8.h }[6], [x13]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
-      "ld1 { v24.h }[6], [x23]\n"
-      "ld1 { v28.h }[6], [x22]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
+      "ld1 { v24.h }[6], [x22]\n"
+      "ld1 { v28.h }[6], [x21]\n"
       "b 270f\n"
       "267:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v8.h }[4], [x13]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
-      "ld1 { v24.h }[4], [x23]\n"
-      "ld1 { v28.h }[4], [x22]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
+      "ld1 { v24.h }[4], [x22]\n"
+      "ld1 { v28.h }[4], [x21]\n"
       "b 270f\n"
       "268:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x14, #1, 269f\n"
       "ldr s8, [x13], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s28, [x22], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
+      "ldr s24, [x22], #0x4\n"
+      "ldr s28, [x21], #0x4\n"
       "tbz x14, #0, 270f\n"
       "ld1 { v8.h }[2], [x13]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
-      "ld1 { v24.h }[2], [x23]\n"
-      "ld1 { v28.h }[2], [x22]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
+      "ld1 { v24.h }[2], [x22]\n"
+      "ld1 { v28.h }[2], [x21]\n"
       "b 270f\n"
       "269:"  // Height 6: Partial accumulate: partial_1_0
       "ldr h8, [x13, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
-      "ldr h24, [x23, #0x0]\n"
-      "ldr h28, [x22, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
+      "ldr h24, [x22, #0x0]\n"
+      "ldr h28, [x21, #0x0]\n"
       "270:"  // Height 6: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 273f\n"
@@ -4449,26 +4446,26 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "b 273f\n"
       "272:"  // Height 6: no accumulate
       "movi v8.16b, #0x0\n"
@@ -4499,8 +4496,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "274:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 275f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -4543,10 +4540,10 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "fmla v8.8h, v6.8h, v0.h[0]\n"
       "fmla v12.8h, v6.8h, v1.h[0]\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
+      "cmp x27, #0x10\n"
       "fmla v16.8h, v6.8h, v2.h[0]\n"
       "fmla v20.8h, v6.8h, v3.h[0]\n"
-      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "add x25, x25, #0x10\n"
       "fmla v24.8h, v6.8h, v4.h[0]\n"
       "fmla v28.8h, v6.8h, v5.h[0]\n"
@@ -5029,12 +5026,12 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "ldr h2, [x21], #0x2\n"
       "ldr q1, [x12, #0x0]\n"
       "ldr q0, [x11, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.8h, v1.8h, v7.h[0]\n"
       "fmla v12.8h, v1.8h, v6.h[0]\n"
       "fmla v16.8h, v1.8h, v5.h[0]\n"
       "fmla v20.8h, v1.8h, v4.h[0]\n"
+      "add x12, x12, #0x10\n"
+      "add x11, x11, #0x10\n"
       "fmla v24.8h, v1.8h, v3.h[0]\n"
       "fmla v28.8h, v1.8h, v2.h[0]\n"
       "ldr q1, [x10, #0x0]\n"
@@ -5066,15 +5063,15 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 274b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "tbz %x[flags], #1, 282f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.8h }, [x21]\n"
       "ld1r { v0.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v1.8h\n"
       "fmin v9.8h, v9.8h, v1.8h\n"
@@ -5130,255 +5127,255 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "tbz x14, #4, 290f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
       "st1 { v9.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v25.8h }, [x23], #0x10\n"
-      "st1 { v28.8h }, [x22], #0x10\n"
-      "st1 { v29.8h }, [x22], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v25.8h }, [x22], #0x10\n"
+      "st1 { v28.8h }, [x21], #0x10\n"
+      "st1 { v29.8h }, [x21], #0x10\n"
       "tbz x14, #3, 286f\n"
       "st1 { v10.8h }, [x13], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
-      "st1 { v26.8h }, [x23], #0x10\n"
-      "st1 { v30.8h }, [x22], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
+      "st1 { v26.8h }, [x22], #0x10\n"
+      "st1 { v30.8h }, [x21], #0x10\n"
       "tbz x14, #2, 284f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x14, #1, 283f\n"
       "st1 { v11.s }[2], [x13], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
-      "st1 { v27.s }[2], [x23], #0x4\n"
-      "st1 { v31.s }[2], [x22], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
+      "st1 { v27.s }[2], [x22], #0x4\n"
+      "st1 { v31.s }[2], [x21], #0x4\n"
       "tbz x14, #0, 298f\n"
       "st1 { v11.h }[6], [x13]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
-      "st1 { v27.h }[6], [x23]\n"
-      "st1 { v31.h }[6], [x22]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
+      "st1 { v27.h }[6], [x22]\n"
+      "st1 { v31.h }[6], [x21]\n"
       "b 298f\n"
       "283:"  // Height 6: Partial direct writeback: partial_1_28
       "tbz x14, #0, 298f\n"
       "st1 { v11.h }[4], [x13]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
-      "st1 { v27.h }[4], [x23]\n"
-      "st1 { v31.h }[4], [x22]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
+      "st1 { v27.h }[4], [x22]\n"
+      "st1 { v31.h }[4], [x21]\n"
       "b 298f\n"
       "284:"  // Height 6: Partial direct writeback: partial_2_24
       "tbz x14, #1, 285f\n"
       "str s11, [x13], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
-      "str s27, [x23], #0x4\n"
-      "str s31, [x22], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
+      "str s27, [x22], #0x4\n"
+      "str s31, [x21], #0x4\n"
       "tbz x14, #0, 298f\n"
       "st1 { v11.h }[2], [x13]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
-      "st1 { v27.h }[2], [x23]\n"
-      "st1 { v31.h }[2], [x22]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
+      "st1 { v27.h }[2], [x22]\n"
+      "st1 { v31.h }[2], [x21]\n"
       "b 298f\n"
       "285:"  // Height 6: Partial direct writeback: partial_1_24
       "tbz x14, #0, 298f\n"
       "str h11, [x13, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
-      "str h27, [x23, #0x0]\n"
-      "str h31, [x22, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
+      "str h27, [x22, #0x0]\n"
+      "str h31, [x21, #0x0]\n"
       "b 298f\n"
       "286:"  // Height 6: Partial direct writeback: partial_4_16
       "tbz x14, #2, 288f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x14, #1, 287f\n"
       "st1 { v10.s }[2], [x13], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
-      "st1 { v26.s }[2], [x23], #0x4\n"
-      "st1 { v30.s }[2], [x22], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
+      "st1 { v26.s }[2], [x22], #0x4\n"
+      "st1 { v30.s }[2], [x21], #0x4\n"
       "tbz x14, #0, 298f\n"
       "st1 { v10.h }[6], [x13]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
-      "st1 { v26.h }[6], [x23]\n"
-      "st1 { v30.h }[6], [x22]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
+      "st1 { v26.h }[6], [x22]\n"
+      "st1 { v30.h }[6], [x21]\n"
       "b 298f\n"
       "287:"  // Height 6: Partial direct writeback: partial_1_20
       "tbz x14, #0, 298f\n"
       "st1 { v10.h }[4], [x13]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
-      "st1 { v26.h }[4], [x23]\n"
-      "st1 { v30.h }[4], [x22]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
+      "st1 { v26.h }[4], [x22]\n"
+      "st1 { v30.h }[4], [x21]\n"
       "b 298f\n"
       "288:"  // Height 6: Partial direct writeback: partial_2_16
       "tbz x14, #1, 289f\n"
       "str s10, [x13], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
-      "str s26, [x23], #0x4\n"
-      "str s30, [x22], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
+      "str s26, [x22], #0x4\n"
+      "str s30, [x21], #0x4\n"
       "tbz x14, #0, 298f\n"
       "st1 { v10.h }[2], [x13]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
-      "st1 { v26.h }[2], [x23]\n"
-      "st1 { v30.h }[2], [x22]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
+      "st1 { v26.h }[2], [x22]\n"
+      "st1 { v30.h }[2], [x21]\n"
       "b 298f\n"
       "289:"  // Height 6: Partial direct writeback: partial_1_16
       "tbz x14, #0, 298f\n"
       "str h10, [x13, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
-      "str h26, [x23, #0x0]\n"
-      "str h30, [x22, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
+      "str h26, [x22, #0x0]\n"
+      "str h30, [x21, #0x0]\n"
       "b 298f\n"
       "290:"  // Height 6: Partial direct writeback: partial_8_0
       "tbz x14, #3, 294f\n"
       "st1 { v8.8h }, [x13], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v28.8h }, [x22], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v28.8h }, [x21], #0x10\n"
       "tbz x14, #2, 292f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x14, #1, 291f\n"
       "st1 { v9.s }[2], [x13], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
-      "st1 { v25.s }[2], [x23], #0x4\n"
-      "st1 { v29.s }[2], [x22], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
+      "st1 { v25.s }[2], [x22], #0x4\n"
+      "st1 { v29.s }[2], [x21], #0x4\n"
       "tbz x14, #0, 298f\n"
       "st1 { v9.h }[6], [x13]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
-      "st1 { v25.h }[6], [x23]\n"
-      "st1 { v29.h }[6], [x22]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
+      "st1 { v25.h }[6], [x22]\n"
+      "st1 { v29.h }[6], [x21]\n"
       "b 298f\n"
       "291:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x14, #0, 298f\n"
       "st1 { v9.h }[4], [x13]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
-      "st1 { v25.h }[4], [x23]\n"
-      "st1 { v29.h }[4], [x22]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
+      "st1 { v25.h }[4], [x22]\n"
+      "st1 { v29.h }[4], [x21]\n"
       "b 298f\n"
       "292:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x14, #1, 293f\n"
       "str s9, [x13], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
-      "str s25, [x23], #0x4\n"
-      "str s29, [x22], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
+      "str s25, [x22], #0x4\n"
+      "str s29, [x21], #0x4\n"
       "tbz x14, #0, 298f\n"
       "st1 { v9.h }[2], [x13]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
-      "st1 { v25.h }[2], [x23]\n"
-      "st1 { v29.h }[2], [x22]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
+      "st1 { v25.h }[2], [x22]\n"
+      "st1 { v29.h }[2], [x21]\n"
       "b 298f\n"
       "293:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x14, #0, 298f\n"
       "str h9, [x13, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
-      "str h25, [x23, #0x0]\n"
-      "str h29, [x22, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
+      "str h25, [x22, #0x0]\n"
+      "str h29, [x21, #0x0]\n"
       "b 298f\n"
       "294:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x14, #2, 296f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x14, #1, 295f\n"
       "st1 { v8.s }[2], [x13], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
       "tbz x14, #0, 298f\n"
       "st1 { v8.h }[6], [x13]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
-      "st1 { v24.h }[6], [x23]\n"
-      "st1 { v28.h }[6], [x22]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
+      "st1 { v24.h }[6], [x22]\n"
+      "st1 { v28.h }[6], [x21]\n"
       "b 298f\n"
       "295:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x14, #0, 298f\n"
       "st1 { v8.h }[4], [x13]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
-      "st1 { v24.h }[4], [x23]\n"
-      "st1 { v28.h }[4], [x22]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
+      "st1 { v24.h }[4], [x22]\n"
+      "st1 { v28.h }[4], [x21]\n"
       "b 298f\n"
       "296:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x14, #1, 297f\n"
       "str s8, [x13], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
       "tbz x14, #0, 298f\n"
       "st1 { v8.h }[2], [x13]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
-      "st1 { v24.h }[2], [x23]\n"
-      "st1 { v28.h }[2], [x22]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
+      "st1 { v24.h }[2], [x22]\n"
+      "st1 { v28.h }[2], [x21]\n"
       "b 298f\n"
       "297:"  // Height 6: Partial direct writeback: partial_1_0
       "str h8, [x13, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
-      "str h24, [x23, #0x0]\n"
-      "str h28, [x22, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
+      "str h24, [x22, #0x0]\n"
+      "str h28, [x21, #0x0]\n"
       "298:"  // Height 6: Partial direct writeback: Done
       "b 300f\n"
       "299:"  // Height 6: Full writeback
@@ -5387,26 +5384,26 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q28, [x22, #0x0]\n"
-      "str q29, [x22, #0x10]\n"
-      "str q30, [x22, #0x20]\n"
-      "str q31, [x22, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
+      "str q28, [x21, #0x0]\n"
+      "str q29, [x21, #0x10]\n"
+      "str q30, [x21, #0x20]\n"
+      "str q31, [x21, #0x30]\n"
       "300:"  // Height 6: Writeback done
       "subs x14, x14, #0x20\n"
       "bgt 252b\n"
@@ -5422,8 +5419,8 @@ void a64_ffhybrid_fp16_mla_6x32 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "302:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16.hpp
index 658850e12c..8f2c7fdec5 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16.hpp
@@ -81,7 +81,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 1> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp
index 5dcaa9e5d5..c6c5b047ca 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp
@@ -49,19 +49,18 @@ void a64_ffhybrid_fp32_mla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -82,7 +81,6 @@ void a64_ffhybrid_fp32_mla_6x16 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -107,19 +105,19 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "bgt 69f\n"
       "beq 35f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #2\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 3f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -210,8 +208,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "16:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 17f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -235,9 +233,6 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr q17, [x10, #0x0]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "ldr q16, [x9, #0x0]\n"
-      "sub x27, x27, #0x4\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x8\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "ldr q17, [x12, #0x10]\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
@@ -256,29 +251,30 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr q16, [x9, #0x20]\n"
       "fmla v10.4s, v17.4s, v0.s[2]\n"
       "ldr q17, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "fmla v11.4s, v16.4s, v0.s[2]\n"
       "ldr q16, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       "fmla v8.4s, v17.4s, v0.s[3]\n"
       "ldr q17, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr q16, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x8\n"
       "fmla v10.4s, v17.4s, v0.s[3]\n"
-      "ldr q6, [x12, #0x0]\n"
       "fmla v11.4s, v16.4s, v0.s[3]\n"
+      "add x26, x26, #0x10\n"
       "ldr q0, [x26, #0x0]\n"
+      "add x12, x12, #0x40\n"
+      "ldr q6, [x12, #0x0]\n"
+      "add x11, x11, #0x40\n"
       "ldr q7, [x11, #0x0]\n"
+      "add x10, x10, #0x40\n"
+      "add x9, x9, #0x40\n"
       "bge 19b\n"
       "20:"  // Height 1: Multiply loop: Single iteration only
       "fmla v8.4s, v6.4s, v0.s[0]\n"
       "ldr q17, [x10, #0x0]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "ldr q16, [x9, #0x0]\n"
-      "sub x27, x27, #0x4\n"
-      "add x26, x26, #0x10\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "ldr q17, [x12, #0x10]\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
@@ -297,35 +293,37 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr q16, [x9, #0x20]\n"
       "fmla v10.4s, v17.4s, v0.s[2]\n"
       "ldr q17, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "fmla v11.4s, v16.4s, v0.s[2]\n"
       "ldr q16, [x11, #0x30]\n"
-      "add x11, x11, #0x40\n"
       "fmla v8.4s, v17.4s, v0.s[3]\n"
       "ldr q17, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr q16, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
+      "sub x27, x27, #0x4\n"
       "fmla v10.4s, v17.4s, v0.s[3]\n"
       "fmla v11.4s, v16.4s, v0.s[3]\n"
+      "add x26, x26, #0x10\n"
+      "add x12, x12, #0x40\n"
+      "add x11, x11, #0x40\n"
+      "add x10, x10, #0x40\n"
+      "add x9, x9, #0x40\n"
       "21:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 23f\n"
       "22:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x26], #0x4\n"
-      "ldr q17, [x12, #0x0]\n"
+      "ldr q16, [x12, #0x0]\n"
+      "fmla v8.4s, v16.4s, v18.s[0]\n"
       "sub x27, x27, #0x1\n"
+      "ldr q17, [x11, #0x0]\n"
+      "ldr q16, [x10, #0x0]\n"
+      "fmla v9.4s, v17.4s, v18.s[0]\n"
+      "fmla v10.4s, v16.4s, v18.s[0]\n"
+      "ldr q16, [x9, #0x0]\n"
+      "fmla v11.4s, v16.4s, v18.s[0]\n"
       "add x12, x12, #0x10\n"
-      "ldr q16, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "fmla v8.4s, v17.4s, v18.s[0]\n"
-      "ldr q17, [x10, #0x0]\n"
       "add x10, x10, #0x10\n"
-      "fmla v9.4s, v16.4s, v18.s[0]\n"
-      "ldr q16, [x9, #0x0]\n"
       "add x9, x9, #0x10\n"
-      "fmla v10.4s, v17.4s, v18.s[0]\n"
-      "fmla v11.4s, v16.4s, v18.s[0]\n"
       "cbnz x27, 22b\n"
       "23:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -333,9 +331,9 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 16b\n"
       "tbz %x[flags], #1, 24f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v17.4s\n"
       "fmin v9.4s, v9.4s, v17.4s\n"
@@ -406,19 +404,19 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "b 206f\n"
       "35:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "36:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #2\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 37f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -431,87 +429,87 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cbz x15, 38f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "add x15, x15, #0x40\n"
       "b 49f\n"
       "38:"  // Height 2: no bias
       "tbz %x[flags], #0, 48f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "bge 47f\n"
       "tbz x14, #3, 42f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
       "tbz x14, #2, 40f\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
       "tbz x14, #1, 39f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x14, #0, 46f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v15.s }[2], [x25]\n"
       "b 46f\n"
       "39:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 46f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
       "b 46f\n"
       "40:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x14, #1, 41f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x14, #0, 46f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v14.s }[2], [x25]\n"
       "b 46f\n"
       "41:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 46f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
       "b 46f\n"
       "42:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x14, #2, 44f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "tbz x14, #1, 43f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x14, #0, 46f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v13.s }[2], [x25]\n"
       "b 46f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 46f\n"
       "ldr s9, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
       "b 46f\n"
       "44:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x14, #1, 45f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x14, #0, 46f\n"
       "ld1 { v8.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v12.s }[2], [x25]\n"
       "b 46f\n"
       "45:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s8, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
       "46:"  // Height 2: Partial accumulate: Done
       "sub x13, x13, x20\n"
@@ -521,10 +519,10 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "b 49f\n"
       "48:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -539,8 +537,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "50:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 51f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -571,15 +569,15 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "ldr q16, [x9, #0x0]\n"
-      "add x26, x26, #0x10\n"
       "cmp x27, #0x8\n"
-      "add x25, x25, #0x10\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "fmla v14.4s, v17.4s, v1.s[0]\n"
       "ldr q17, [x12, #0x10]\n"
+      "add x26, x26, #0x10\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
       "fmla v15.4s, v16.4s, v1.s[0]\n"
       "ldr q16, [x11, #0x10]\n"
+      "add x25, x25, #0x10\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
       "fmla v12.4s, v17.4s, v1.s[1]\n"
       "ldr q17, [x10, #0x10]\n"
@@ -632,10 +630,10 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "ldr q16, [x9, #0x0]\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "fmla v14.4s, v17.4s, v1.s[0]\n"
       "ldr q17, [x12, #0x10]\n"
+      "add x25, x25, #0x10\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
       "fmla v15.4s, v16.4s, v1.s[0]\n"
       "ldr q16, [x11, #0x10]\n"
@@ -685,20 +683,20 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "sub x27, x27, #0x1\n"
       "ldr q17, [x12, #0x0]\n"
       "ldr q16, [x11, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.4s, v17.4s, v19.s[0]\n"
       "fmla v12.4s, v17.4s, v18.s[0]\n"
       "ldr q17, [x10, #0x0]\n"
-      "add x10, x10, #0x10\n"
       "fmla v9.4s, v16.4s, v19.s[0]\n"
       "fmla v13.4s, v16.4s, v18.s[0]\n"
       "ldr q16, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
       "fmla v10.4s, v17.4s, v19.s[0]\n"
       "fmla v14.4s, v17.4s, v18.s[0]\n"
+      "add x12, x12, #0x10\n"
+      "add x11, x11, #0x10\n"
       "fmla v11.4s, v16.4s, v19.s[0]\n"
       "fmla v15.4s, v16.4s, v18.s[0]\n"
+      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       "cbnz x27, 56b\n"
       "57:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -706,11 +704,11 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 50b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "tbz %x[flags], #1, 58f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v17.4s\n"
       "fmin v9.4s, v9.4s, v17.4s\n"
@@ -734,63 +732,63 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "tbz x14, #3, 62f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
       "st1 { v9.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
       "tbz x14, #2, 60f\n"
       "st1 { v10.4s }, [x13], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
       "tbz x14, #1, 59f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
       "tbz x14, #0, 66f\n"
       "st1 { v11.s }[2], [x13]\n"
-      "st1 { v15.s }[2], [x26]\n"
+      "st1 { v15.s }[2], [x25]\n"
       "b 66f\n"
       "59:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x14, #0, 66f\n"
       "str s11, [x13, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
       "b 66f\n"
       "60:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x14, #1, 61f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
+      "str d14, [x25], #0x8\n"
       "tbz x14, #0, 66f\n"
       "st1 { v10.s }[2], [x13]\n"
-      "st1 { v14.s }[2], [x26]\n"
+      "st1 { v14.s }[2], [x25]\n"
       "b 66f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x14, #0, 66f\n"
       "str s10, [x13, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
       "b 66f\n"
       "62:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x14, #2, 64f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
       "tbz x14, #1, 63f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
+      "str d13, [x25], #0x8\n"
       "tbz x14, #0, 66f\n"
       "st1 { v9.s }[2], [x13]\n"
-      "st1 { v13.s }[2], [x26]\n"
+      "st1 { v13.s }[2], [x25]\n"
       "b 66f\n"
       "63:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x14, #0, 66f\n"
       "str s9, [x13, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
       "b 66f\n"
       "64:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x14, #1, 65f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
+      "str d12, [x25], #0x8\n"
       "tbz x14, #0, 66f\n"
       "st1 { v8.s }[2], [x13]\n"
-      "st1 { v12.s }[2], [x26]\n"
+      "st1 { v12.s }[2], [x25]\n"
       "b 66f\n"
       "65:"  // Height 2: Partial direct writeback: partial_1_0
       "str s8, [x13, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
       "66:"  // Height 2: Partial direct writeback: Done
       "b 68f\n"
       "67:"  // Height 2: Full writeback
@@ -799,29 +797,29 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
       "68:"  // Height 2: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 36b\n"
       "b 206f\n"
       "69:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "70:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #2\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 71f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -834,109 +832,109 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cbz x15, 72f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x15, x15, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
       "b 83f\n"
       "72:"  // Height 3: no bias
       "tbz %x[flags], #0, 82f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
       "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "bge 81f\n"
       "tbz x14, #3, 76f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "tbz x14, #2, 74f\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
       "tbz x14, #1, 73f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x14, #0, 80f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
       "b 80f\n"
       "73:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 80f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
       "b 80f\n"
       "74:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x14, #1, 75f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x14, #0, 80f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
       "b 80f\n"
       "75:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 80f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
       "b 80f\n"
       "76:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x14, #2, 78f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
       "tbz x14, #1, 77f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x14, #0, 80f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
       "b 80f\n"
       "77:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 80f\n"
       "ldr s9, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
       "b 80f\n"
       "78:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x14, #1, 79f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
       "tbz x14, #0, 80f\n"
       "ld1 { v8.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
       "b 80f\n"
       "79:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s8, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
       "80:"  // Height 3: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 83f\n"
@@ -945,14 +943,14 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
       "b 83f\n"
       "82:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -971,8 +969,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "84:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 85f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1003,18 +1001,18 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v8.4s, v6.4s, v0.s[0]\n"
       "fmla v12.4s, v6.4s, v1.s[0]\n"
       "sub x27, x27, #0x4\n"
-      "add x26, x26, #0x10\n"
+      "cmp x27, #0x8\n"
       "fmla v16.4s, v6.4s, v2.s[0]\n"
       "ldr q21, [x10, #0x0]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
-      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "fmla v17.4s, v7.4s, v2.s[0]\n"
       "ldr q20, [x9, #0x0]\n"
       "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
       "fmla v10.4s, v21.4s, v0.s[0]\n"
       "fmla v14.4s, v21.4s, v1.s[0]\n"
+      "add x24, x24, #0x10\n"
       "fmla v18.4s, v21.4s, v2.s[0]\n"
       "ldr q21, [x12, #0x10]\n"
       "fmla v11.4s, v20.4s, v0.s[0]\n"
@@ -1049,8 +1047,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v14.4s, v21.4s, v1.s[2]\n"
       "fmla v18.4s, v21.4s, v2.s[2]\n"
       "ldr q21, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "fmla v11.4s, v20.4s, v0.s[2]\n"
+      "add x12, x12, #0x40\n"
       "fmla v15.4s, v20.4s, v1.s[2]\n"
       "fmla v19.4s, v20.4s, v2.s[2]\n"
       "ldr q20, [x11, #0x30]\n"
@@ -1059,8 +1057,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v12.4s, v21.4s, v1.s[3]\n"
       "fmla v16.4s, v21.4s, v2.s[3]\n"
       "ldr q21, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       "fmla v9.4s, v20.4s, v0.s[3]\n"
+      "add x10, x10, #0x40\n"
       "fmla v13.4s, v20.4s, v1.s[3]\n"
       "fmla v17.4s, v20.4s, v2.s[3]\n"
       "ldr q20, [x9, #0x30]\n"
@@ -1126,8 +1124,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v14.4s, v21.4s, v1.s[2]\n"
       "fmla v18.4s, v21.4s, v2.s[2]\n"
       "ldr q21, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "fmla v11.4s, v20.4s, v0.s[2]\n"
+      "add x12, x12, #0x40\n"
       "fmla v15.4s, v20.4s, v1.s[2]\n"
       "fmla v19.4s, v20.4s, v2.s[2]\n"
       "ldr q20, [x11, #0x30]\n"
@@ -1136,8 +1134,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v12.4s, v21.4s, v1.s[3]\n"
       "fmla v16.4s, v21.4s, v2.s[3]\n"
       "ldr q21, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       "fmla v9.4s, v20.4s, v0.s[3]\n"
+      "add x10, x10, #0x40\n"
       "fmla v13.4s, v20.4s, v1.s[3]\n"
       "fmla v17.4s, v20.4s, v2.s[3]\n"
       "ldr q20, [x9, #0x30]\n"
@@ -1156,23 +1154,23 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "sub x27, x27, #0x1\n"
       "ldr s22, [x24], #0x4\n"
       "ldr q21, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "ldr q20, [x11, #0x0]\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.4s, v21.4s, v24.s[0]\n"
       "fmla v12.4s, v21.4s, v23.s[0]\n"
+      "ldr q20, [x11, #0x0]\n"
       "fmla v16.4s, v21.4s, v22.s[0]\n"
       "ldr q21, [x10, #0x0]\n"
-      "add x10, x10, #0x10\n"
       "fmla v9.4s, v20.4s, v24.s[0]\n"
       "fmla v13.4s, v20.4s, v23.s[0]\n"
       "fmla v17.4s, v20.4s, v22.s[0]\n"
       "ldr q20, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
+      "add x12, x12, #0x10\n"
       "fmla v10.4s, v21.4s, v24.s[0]\n"
       "fmla v14.4s, v21.4s, v23.s[0]\n"
+      "add x11, x11, #0x10\n"
+      "add x10, x10, #0x10\n"
       "fmla v18.4s, v21.4s, v22.s[0]\n"
       "fmla v11.4s, v20.4s, v24.s[0]\n"
+      "add x9, x9, #0x10\n"
       "fmla v15.4s, v20.4s, v23.s[0]\n"
       "fmla v19.4s, v20.4s, v22.s[0]\n"
       "cbnz x27, 90b\n"
@@ -1182,12 +1180,12 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 84b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "tbz %x[flags], #1, 92f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.4s }, [x21]\n"
       "ld1r { v20.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v21.4s\n"
       "fmin v9.4s, v9.4s, v21.4s\n"
@@ -1219,79 +1217,79 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "tbz x14, #3, 96f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
       "st1 { v9.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
       "tbz x14, #2, 94f\n"
       "st1 { v10.4s }, [x13], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
       "tbz x14, #1, 93f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x14, #0, 100f\n"
       "st1 { v11.s }[2], [x13]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
       "b 100f\n"
       "93:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x14, #0, 100f\n"
       "str s11, [x13, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
       "b 100f\n"
       "94:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x14, #1, 95f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x14, #0, 100f\n"
       "st1 { v10.s }[2], [x13]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
       "b 100f\n"
       "95:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x14, #0, 100f\n"
       "str s10, [x13, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
       "b 100f\n"
       "96:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x14, #2, 98f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
       "tbz x14, #1, 97f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x14, #0, 100f\n"
       "st1 { v9.s }[2], [x13]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
       "b 100f\n"
       "97:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x14, #0, 100f\n"
       "str s9, [x13, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
       "b 100f\n"
       "98:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x14, #1, 99f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x14, #0, 100f\n"
       "st1 { v8.s }[2], [x13]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
       "b 100f\n"
       "99:"  // Height 3: Partial direct writeback: partial_1_0
       "str s8, [x13, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
       "100:"  // Height 3: Partial direct writeback: Done
       "b 102f\n"
       "101:"  // Height 3: Full writeback
@@ -1300,33 +1298,33 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "102:"  // Height 3: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 70b\n"
       "b 206f\n"
       "103:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "104:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #2\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 105f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -1339,18 +1337,18 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cbz x15, 106f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x15, x15, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1358,111 +1356,111 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "106:"  // Height 4: no bias
       "tbz %x[flags], #0, 116f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "cmp x14, #0x10\n"
+      "add x23, x24, x20, LSL #2\n"
       "bge 115f\n"
       "tbz x14, #3, 110f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
       "tbz x14, #2, 108f\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
       "tbz x14, #1, 107f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x14, #0, 114f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
       "b 114f\n"
       "107:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 114f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
       "b 114f\n"
       "108:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x14, #1, 109f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x14, #0, 114f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
       "b 114f\n"
       "109:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 114f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
       "b 114f\n"
       "110:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x14, #2, 112f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "tbz x14, #1, 111f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x14, #0, 114f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
       "b 114f\n"
       "111:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 114f\n"
       "ldr s9, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
       "b 114f\n"
       "112:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x14, #1, 113f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x14, #0, 114f\n"
       "ld1 { v8.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
       "b 114f\n"
       "113:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s8, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
       "114:"  // Height 4: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 117f\n"
@@ -1471,18 +1469,18 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "b 117f\n"
       "116:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1505,8 +1503,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "118:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 119f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1541,11 +1539,11 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v8.4s, v6.4s, v0.s[0]\n"
       "fmla v12.4s, v6.4s, v1.s[0]\n"
       "sub x27, x27, #0x4\n"
-      "add x26, x26, #0x10\n"
+      "cmp x27, #0x8\n"
       "fmla v16.4s, v6.4s, v2.s[0]\n"
       "fmla v20.4s, v6.4s, v3.s[0]\n"
       "ldr q25, [x10, #0x0]\n"
-      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "add x25, x25, #0x10\n"
@@ -1731,16 +1729,16 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr s26, [x23], #0x4\n"
       "ldr q25, [x12, #0x0]\n"
       "ldr q24, [x11, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.4s, v25.4s, v29.s[0]\n"
       "fmla v12.4s, v25.4s, v28.s[0]\n"
       "fmla v16.4s, v25.4s, v27.s[0]\n"
       "fmla v20.4s, v25.4s, v26.s[0]\n"
       "ldr q25, [x10, #0x0]\n"
-      "add x10, x10, #0x10\n"
+      "add x12, x12, #0x10\n"
       "fmla v9.4s, v24.4s, v29.s[0]\n"
       "fmla v13.4s, v24.4s, v28.s[0]\n"
+      "add x11, x11, #0x10\n"
+      "add x10, x10, #0x10\n"
       "fmla v17.4s, v24.4s, v27.s[0]\n"
       "fmla v21.4s, v24.4s, v26.s[0]\n"
       "ldr q24, [x9, #0x0]\n"
@@ -1760,13 +1758,13 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 118b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "tbz %x[flags], #1, 126f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v25.4s }, [x21]\n"
       "ld1r { v24.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v25.4s\n"
       "fmin v9.4s, v9.4s, v25.4s\n"
@@ -1806,95 +1804,95 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "tbz x14, #3, 130f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
       "st1 { v9.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
       "tbz x14, #2, 128f\n"
       "st1 { v10.4s }, [x13], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
       "tbz x14, #1, 127f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
       "tbz x14, #0, 134f\n"
       "st1 { v11.s }[2], [x13]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
       "b 134f\n"
       "127:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x14, #0, 134f\n"
       "str s11, [x13, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
       "b 134f\n"
       "128:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x14, #1, 129f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
       "tbz x14, #0, 134f\n"
       "st1 { v10.s }[2], [x13]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
       "b 134f\n"
       "129:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x14, #0, 134f\n"
       "str s10, [x13, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
       "b 134f\n"
       "130:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x14, #2, 132f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
       "tbz x14, #1, 131f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
       "tbz x14, #0, 134f\n"
       "st1 { v9.s }[2], [x13]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
       "b 134f\n"
       "131:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x14, #0, 134f\n"
       "str s9, [x13, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
       "b 134f\n"
       "132:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x14, #1, 133f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x14, #0, 134f\n"
       "st1 { v8.s }[2], [x13]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
       "b 134f\n"
       "133:"  // Height 4: Partial direct writeback: partial_1_0
       "str s8, [x13, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
       "134:"  // Height 4: Partial direct writeback: Done
       "b 136f\n"
       "135:"  // Height 4: Full writeback
@@ -1903,37 +1901,37 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
       "136:"  // Height 4: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 104b\n"
       "b 206f\n"
       "137:"  // Height 5
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "138:"  // Height 5: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #2\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 139f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -1946,18 +1944,18 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cbz x15, 140f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x15, x15, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1969,128 +1967,128 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "140:"  // Height 5: no bias
       "tbz %x[flags], #0, 150f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x14, #0x10\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 149f\n"
       "tbz x14, #3, 144f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x14, #2, 142f\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
       "tbz x14, #1, 141f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
       "b 148f\n"
       "141:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 148f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
       "b 148f\n"
       "142:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x14, #1, 143f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
       "b 148f\n"
       "143:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 148f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
       "b 148f\n"
       "144:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x14, #2, 146f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "tbz x14, #1, 145f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 148f\n"
       "145:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 148f\n"
       "ldr s9, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "b 148f\n"
       "146:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x14, #1, 147f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v8.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
       "b 148f\n"
       "147:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s8, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
       "148:"  // Height 5: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 151f\n"
@@ -2099,22 +2097,22 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
       "b 151f\n"
       "150:"  // Height 5: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2141,8 +2139,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "152:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 153f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2181,10 +2179,10 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v8.4s, v6.4s, v0.s[0]\n"
       "fmla v12.4s, v6.4s, v1.s[0]\n"
       "sub x27, x27, #0x4\n"
-      "add x26, x26, #0x10\n"
+      "cmp x27, #0x8\n"
       "fmla v16.4s, v6.4s, v2.s[0]\n"
       "fmla v20.4s, v6.4s, v3.s[0]\n"
-      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "add x25, x25, #0x10\n"
       "fmla v24.4s, v6.4s, v4.s[0]\n"
       "ldr q29, [x10, #0x0]\n"
@@ -2406,19 +2404,19 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr s31, [x23], #0x4\n"
       "ldr s30, [x22], #0x4\n"
       "ldr q29, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "ldr q28, [x11, #0x0]\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.4s, v29.4s, v2.s[0]\n"
       "fmla v12.4s, v29.4s, v1.s[0]\n"
+      "ldr q28, [x11, #0x0]\n"
       "fmla v16.4s, v29.4s, v0.s[0]\n"
       "fmla v20.4s, v29.4s, v31.s[0]\n"
+      "add x12, x12, #0x10\n"
       "fmla v24.4s, v29.4s, v30.s[0]\n"
       "ldr q29, [x10, #0x0]\n"
-      "add x10, x10, #0x10\n"
       "fmla v9.4s, v28.4s, v2.s[0]\n"
+      "add x11, x11, #0x10\n"
       "fmla v13.4s, v28.4s, v1.s[0]\n"
       "fmla v17.4s, v28.4s, v0.s[0]\n"
+      "add x10, x10, #0x10\n"
       "fmla v21.4s, v28.4s, v31.s[0]\n"
       "fmla v25.4s, v28.4s, v30.s[0]\n"
       "ldr q28, [x9, #0x0]\n"
@@ -2440,14 +2438,14 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 152b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "tbz %x[flags], #1, 160f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v29.4s }, [x21]\n"
       "ld1r { v28.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v29.4s\n"
       "fmin v9.4s, v9.4s, v29.4s\n"
@@ -2495,111 +2493,111 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "tbz x14, #3, 164f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
       "st1 { v9.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
       "tbz x14, #2, 162f\n"
       "st1 { v10.4s }, [x13], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
       "tbz x14, #1, 161f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x14, #0, 168f\n"
       "st1 { v11.s }[2], [x13]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
       "b 168f\n"
       "161:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x14, #0, 168f\n"
       "str s11, [x13, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
       "b 168f\n"
       "162:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x14, #1, 163f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x14, #0, 168f\n"
       "st1 { v10.s }[2], [x13]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
       "b 168f\n"
       "163:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x14, #0, 168f\n"
       "str s10, [x13, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
       "b 168f\n"
       "164:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x14, #2, 166f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x14, #1, 165f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x14, #0, 168f\n"
       "st1 { v9.s }[2], [x13]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 168f\n"
       "165:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x14, #0, 168f\n"
       "str s9, [x13, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 168f\n"
       "166:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x14, #1, 167f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x14, #0, 168f\n"
       "st1 { v8.s }[2], [x13]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 168f\n"
       "167:"  // Height 5: Partial direct writeback: partial_1_0
       "str s8, [x13, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "168:"  // Height 5: Partial direct writeback: Done
       "b 170f\n"
       "169:"  // Height 5: Full writeback
@@ -2608,45 +2606,44 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "170:"  // Height 5: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 138b\n"
       "b 206f\n"
       "171:"  // Height 6
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0x18\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0x18\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "172:"  // Height 6: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0xc\n"
       "add x11, x12, x20, LSL #2\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0xc\n"
       "bgt 173f\n"
       "cmp x14, #0x8\n"
       "mov x9, x12\n"
@@ -2659,18 +2656,18 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cbz x15, 174f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x15, x15, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -2686,145 +2683,145 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "174:"  // Height 6: no bias
       "tbz %x[flags], #0, 184f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x10\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x14, #0x10\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 183f\n"
       "tbz x14, #3, 178f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x14, #2, 176f\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x14, #1, 175f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x14, #0, 182f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 182f\n"
       "175:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 182f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 182f\n"
       "176:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x14, #1, 177f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x14, #0, 182f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 182f\n"
       "177:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 182f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 182f\n"
       "178:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x14, #2, 180f\n"
       "ld1 { v8.4s }, [x13], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x14, #1, 179f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x14, #0, 182f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 182f\n"
       "179:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 182f\n"
       "ldr s9, [x13, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 182f\n"
       "180:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x14, #1, 181f\n"
       "ldr d8, [x13], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x14, #0, 182f\n"
       "ld1 { v8.s }[2], [x13]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 182f\n"
       "181:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s8, [x13, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "182:"  // Height 6: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 185f\n"
@@ -2833,26 +2830,26 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr q9, [x13, #0x10]\n"
       "ldr q10, [x13, #0x20]\n"
       "ldr q11, [x13, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "b 185f\n"
       "184:"  // Height 6: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2883,8 +2880,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "186:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 187f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2927,10 +2924,10 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "fmla v8.4s, v6.4s, v0.s[0]\n"
       "fmla v12.4s, v6.4s, v1.s[0]\n"
       "sub x27, x27, #0x4\n"
-      "add x26, x26, #0x10\n"
+      "cmp x27, #0x8\n"
       "fmla v16.4s, v6.4s, v2.s[0]\n"
       "fmla v20.4s, v6.4s, v3.s[0]\n"
-      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "add x25, x25, #0x10\n"
       "fmla v24.4s, v6.4s, v4.s[0]\n"
       "fmla v28.4s, v6.4s, v5.s[0]\n"
@@ -3189,12 +3186,12 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "ldr s2, [x21], #0x4\n"
       "ldr q1, [x12, #0x0]\n"
       "ldr q0, [x11, #0x0]\n"
-      "add x12, x12, #0x10\n"
-      "add x11, x11, #0x10\n"
       "fmla v8.4s, v1.4s, v7.s[0]\n"
       "fmla v12.4s, v1.4s, v6.s[0]\n"
       "fmla v16.4s, v1.4s, v5.s[0]\n"
       "fmla v20.4s, v1.4s, v4.s[0]\n"
+      "add x12, x12, #0x10\n"
+      "add x11, x11, #0x10\n"
       "fmla v24.4s, v1.4s, v3.s[0]\n"
       "fmla v28.4s, v1.4s, v2.s[0]\n"
       "ldr q1, [x10, #0x0]\n"
@@ -3226,15 +3223,15 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 186b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "tbz %x[flags], #1, 194f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v1.4s\n"
       "fmin v9.4s, v9.4s, v1.4s\n"
@@ -3290,127 +3287,127 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "tbz x14, #3, 198f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
       "st1 { v9.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
-      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
+      "st1 { v29.4s }, [x21], #0x10\n"
       "tbz x14, #2, 196f\n"
       "st1 { v10.4s }, [x13], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
-      "st1 { v30.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v30.4s }, [x21], #0x10\n"
       "tbz x14, #1, 195f\n"
       "str d11, [x13], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x14, #0, 202f\n"
       "st1 { v11.s }[2], [x13]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
-      "st1 { v31.s }[2], [x22]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v31.s }[2], [x21]\n"
       "b 202f\n"
       "195:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x14, #0, 202f\n"
       "str s11, [x13, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
-      "str s31, [x22, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
+      "str s31, [x21, #0x0]\n"
       "b 202f\n"
       "196:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x14, #1, 197f\n"
       "str d10, [x13], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x14, #0, 202f\n"
       "st1 { v10.s }[2], [x13]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
-      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v30.s }[2], [x21]\n"
       "b 202f\n"
       "197:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x14, #0, 202f\n"
       "str s10, [x13, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
-      "str s30, [x22, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
+      "str s30, [x21, #0x0]\n"
       "b 202f\n"
       "198:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x14, #2, 200f\n"
       "st1 { v8.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
       "tbz x14, #1, 199f\n"
       "str d9, [x13], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x14, #0, 202f\n"
       "st1 { v9.s }[2], [x13]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
-      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v29.s }[2], [x21]\n"
       "b 202f\n"
       "199:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x14, #0, 202f\n"
       "str s9, [x13, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
-      "str s29, [x22, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
+      "str s29, [x21, #0x0]\n"
       "b 202f\n"
       "200:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x14, #1, 201f\n"
       "str d8, [x13], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x14, #0, 202f\n"
       "st1 { v8.s }[2], [x13]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
-      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v28.s }[2], [x21]\n"
       "b 202f\n"
       "201:"  // Height 6: Partial direct writeback: partial_1_0
       "str s8, [x13, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
-      "str s28, [x22, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
+      "str s28, [x21, #0x0]\n"
       "202:"  // Height 6: Partial direct writeback: Done
       "b 204f\n"
       "203:"  // Height 6: Full writeback
@@ -3419,26 +3416,26 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "str q10, [x13, #0x20]\n"
       "str q11, [x13, #0x30]\n"
       "add x13, x13, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q28, [x22, #0x0]\n"
-      "str q29, [x22, #0x10]\n"
-      "str q30, [x22, #0x20]\n"
-      "str q31, [x22, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
+      "str q28, [x21, #0x0]\n"
+      "str q29, [x21, #0x10]\n"
+      "str q30, [x21, #0x20]\n"
+      "str q31, [x21, #0x30]\n"
       "204:"  // Height 6: Writeback done
       "subs x14, x14, #0x10\n"
       "bgt 172b\n"
@@ -3454,8 +3451,8 @@ void a64_ffhybrid_fp32_mla_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "206:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp
index 73c096ca00..ac3cbf943f 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp
@@ -82,14 +82,16 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 24, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 4, 24, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
         if (std::is_same<T, float>::value) {
             switch (ci->get_cpu_model()) {
+                case CPUModel::V1:
+                    return { 23.64 };
                 default:
-                    return { 28.48 };
+                    return { 16.89 };
             }
         }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp
index 1f7804453c..88547ef3b3 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp
@@ -50,19 +50,18 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -83,7 +82,6 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -105,14 +103,13 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "bgt 89f\n"
       "beq 45f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x14\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
@@ -120,6 +117,7 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "add x27, x28, x20, LSL #1\n"
       "add x20, x27, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x14\n"
       "bgt 3f\n"
       "cmp x14, #0x10\n"
       "mov x27, x12\n"
@@ -138,19 +136,19 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "cbz x15, 4f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "ldr q12, [x15, #0x40]\n"
-      "ldr q13, [x15, #0x50]\n"
-      "add x15, x15, #0x60\n"
       "zip2 v14.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v15.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
+      "ldr q12, [x15, #0x40]\n"
+      "ldr q13, [x15, #0x50]\n"
       "zip2 v16.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
       "zip2 v17.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
+      "add x15, x15, #0x60\n"
       "zip2 v18.2d, v12.2d, v12.2d\n"
       "zip1 v12.2d, v12.2d, v12.2d\n"
       "zip2 v19.2d, v13.2d, v13.2d\n"
@@ -284,8 +282,8 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "mov x26, #0x0\n"
       "21:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 22f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -308,32 +306,28 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "blt 25f\n"
       "24:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x25, x25, #0x4\n"
-      "add x12, x12, #0x20\n"
-      "cmp x25, #0x8\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
       "ldr q24, [x10, #0x0]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
       "ldr q23, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       "ldr q22, [x9, #0x0]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       "ldr q21, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e58ec0a  // bfmmla v10.4s, v0.8h, v24.8h\n"
       "ldr q24, [x28, #0x0]\n"
       ".inst 0x6e57ec10  // bfmmla v16.4s, v0.8h, v23.8h\n"
       "ldr q23, [x28, #0x10]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e56ec0b  // bfmmla v11.4s, v0.8h, v22.8h\n"
       "ldr q22, [x27, #0x0]\n"
       ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
       "ldr q21, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
+      "sub x25, x25, #0x4\n"
+      "cmp x25, #0x8\n"
       ".inst 0x6e58ec0c  // bfmmla v12.4s, v0.8h, v24.8h\n"
+      "add x12, x12, #0x20\n"
       "ldr q4, [x12, #0x0]\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
       "ldr q5, [x12, #0x10]\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
@@ -341,36 +335,40 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
       "ld1 { v0.4s }, [x24], #0x10\n"
       "ldr q7, [x11, #0x10]\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
+      "add x28, x28, #0x20\n"
+      "add x27, x27, #0x20\n"
       "bge 24b\n"
       "25:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x25, x25, #0x4\n"
-      "add x12, x12, #0x20\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
-      "ldr q23, [x10, #0x0]\n"
+      "ldr q22, [x10, #0x0]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
-      "ldr q22, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
+      "ldr q25, [x10, #0x10]\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       "ldr q21, [x9, #0x0]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       "ldr q24, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
-      ".inst 0x6e57ec0a  // bfmmla v10.4s, v0.8h, v23.8h\n"
+      ".inst 0x6e56ec0a  // bfmmla v10.4s, v0.8h, v22.8h\n"
       "ldr q23, [x28, #0x0]\n"
-      ".inst 0x6e56ec10  // bfmmla v16.4s, v0.8h, v22.8h\n"
+      ".inst 0x6e59ec10  // bfmmla v16.4s, v0.8h, v25.8h\n"
       "ldr q22, [x28, #0x10]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e55ec0b  // bfmmla v11.4s, v0.8h, v21.8h\n"
       "ldr q21, [x27, #0x0]\n"
       ".inst 0x6e58ec11  // bfmmla v17.4s, v0.8h, v24.8h\n"
       "ldr q3, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
+      "sub x25, x25, #0x4\n"
       ".inst 0x6e57ec0c  // bfmmla v12.4s, v0.8h, v23.8h\n"
       ".inst 0x6e56ec12  // bfmmla v18.4s, v0.8h, v22.8h\n"
+      "add x12, x12, #0x20\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e55ec0d  // bfmmla v13.4s, v0.8h, v21.8h\n"
       ".inst 0x6e43ec13  // bfmmla v19.4s, v0.8h, v3.8h\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
+      "add x28, x28, #0x20\n"
+      "add x27, x27, #0x20\n"
       "26:"  // Height 1: Multiply loop: Main loop skip
       "cbz x25, 29f\n"
       "cbz x25, 29f\n"
@@ -382,37 +380,37 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "27:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr s0, [x24, #0x0]\n"
       "28:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q23, [x12, #0x0]\n"
-      "ldr q29, [x12, #0x10]\n"
+      "ldr q21, [x12, #0x0]\n"
+      "ldr q30, [x12, #0x10]\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "add x12, x12, #0x20\n"
-      "ldr q22, [x11, #0x0]\n"
-      "ldr q21, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
-      ".inst 0x6e57ec08  // bfmmla v8.4s, v0.8h, v23.8h\n"
-      "ldr q24, [x10, #0x0]\n"
-      ".inst 0x6e5dec0e  // bfmmla v14.4s, v0.8h, v29.8h\n"
+      ".inst 0x6e55ec08  // bfmmla v8.4s, v0.8h, v21.8h\n"
+      "ldr q21, [x11, #0x0]\n"
+      "ldr q22, [x11, #0x10]\n"
+      ".inst 0x6e5eec0e  // bfmmla v14.4s, v0.8h, v30.8h\n"
+      ".inst 0x6e55ec09  // bfmmla v9.4s, v0.8h, v21.8h\n"
+      "ldr q21, [x10, #0x0]\n"
       "ldr q23, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
-      ".inst 0x6e56ec09  // bfmmla v9.4s, v0.8h, v22.8h\n"
-      "ldr q22, [x9, #0x0]\n"
-      ".inst 0x6e55ec0f  // bfmmla v15.4s, v0.8h, v21.8h\n"
-      "ldr q21, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
-      ".inst 0x6e58ec0a  // bfmmla v10.4s, v0.8h, v24.8h\n"
-      "ldr q24, [x28, #0x0]\n"
+      ".inst 0x6e56ec0f  // bfmmla v15.4s, v0.8h, v22.8h\n"
+      ".inst 0x6e55ec0a  // bfmmla v10.4s, v0.8h, v21.8h\n"
+      "ldr q21, [x9, #0x0]\n"
+      "ldr q22, [x9, #0x10]\n"
       ".inst 0x6e57ec10  // bfmmla v16.4s, v0.8h, v23.8h\n"
+      ".inst 0x6e55ec0b  // bfmmla v11.4s, v0.8h, v21.8h\n"
+      "ldr q21, [x28, #0x0]\n"
       "ldr q23, [x28, #0x10]\n"
-      "add x28, x28, #0x20\n"
-      ".inst 0x6e56ec0b  // bfmmla v11.4s, v0.8h, v22.8h\n"
+      ".inst 0x6e56ec11  // bfmmla v17.4s, v0.8h, v22.8h\n"
+      ".inst 0x6e55ec0c  // bfmmla v12.4s, v0.8h, v21.8h\n"
       "ldr q22, [x27, #0x0]\n"
-      ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
       "ldr q21, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
-      ".inst 0x6e58ec0c  // bfmmla v12.4s, v0.8h, v24.8h\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
+      "add x12, x12, #0x20\n"
+      "add x11, x11, #0x20\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
+      "add x28, x28, #0x20\n"
+      "add x27, x27, #0x20\n"
       "29:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -425,9 +423,9 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "uzp1 v12.2d, v12.2d, v18.2d\n"
       "uzp1 v13.2d, v13.2d, v19.2d\n"
       "tbz %x[flags], #1, 30f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v22.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v22.4s }, [x21]\n"
       "ld1r { v21.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v22.4s\n"
       "fmin v9.4s, v9.4s, v22.4s\n"
@@ -531,14 +529,13 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "b 178f\n"
       "45:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "46:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x14\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
@@ -546,6 +543,7 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "add x27, x28, x20, LSL #1\n"
       "add x20, x27, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x14\n"
       "bgt 47f\n"
       "cmp x14, #0x10\n"
       "mov x27, x12\n"
@@ -564,19 +562,19 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "cbz x15, 48f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "ldr q12, [x15, #0x40]\n"
-      "ldr q13, [x15, #0x50]\n"
-      "add x15, x15, #0x60\n"
       "zip2 v14.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v15.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
+      "ldr q12, [x15, #0x40]\n"
+      "ldr q13, [x15, #0x50]\n"
       "zip2 v16.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
       "zip2 v17.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
+      "add x15, x15, #0x60\n"
       "zip2 v18.2d, v12.2d, v12.2d\n"
       "zip1 v12.2d, v12.2d, v12.2d\n"
       "zip2 v19.2d, v13.2d, v13.2d\n"
@@ -586,117 +584,117 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "tbz %x[flags], #0, 63f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x14, #0x18\n"
-      "add x24, x13, x20, LSL #2\n"
+      "add x23, x13, x20, LSL #2\n"
       "bge 61f\n"
       "tbz x14, #4, 52f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v12.4s }, [x13], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
       "tbz x14, #2, 50f\n"
       "ld1 { v13.4s }, [x13], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
       "tbz x14, #1, 49f\n"
       "ldr d20, [x13], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
       "tbz x14, #0, 60f\n"
       "ld1 { v20.s }[2], [x13]\n"
-      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v19.s }[2], [x23]\n"
       "b 60f\n"
       "49:"  // Height 2: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x14, #0, 60f\n"
       "ldr s20, [x13, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
       "b 60f\n"
       "50:"  // Height 2: Partial accumulate: partial_2_16
       "tbz x14, #1, 51f\n"
       "ldr d13, [x13], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
       "tbz x14, #0, 60f\n"
       "ld1 { v13.s }[2], [x13]\n"
-      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v18.s }[2], [x23]\n"
       "b 60f\n"
       "51:"  // Height 2: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x14, #0, 60f\n"
       "ldr s13, [x13, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
       "b 60f\n"
       "52:"  // Height 2: Partial accumulate: partial_8_0
       "tbz x14, #3, 56f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
       "tbz x14, #2, 54f\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
       "tbz x14, #1, 53f\n"
       "ldr d12, [x13], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x14, #0, 60f\n"
       "ld1 { v12.s }[2], [x13]\n"
-      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v17.s }[2], [x23]\n"
       "b 60f\n"
       "53:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 60f\n"
       "ldr s12, [x13, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
       "b 60f\n"
       "54:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x14, #1, 55f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x14, #0, 60f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v16.s }[2], [x23]\n"
       "b 60f\n"
       "55:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 60f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
       "b 60f\n"
       "56:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x14, #2, 58f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "tbz x14, #1, 57f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x14, #0, 60f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x23]\n"
       "b 60f\n"
       "57:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 60f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
       "b 60f\n"
       "58:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x14, #1, 59f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x14, #0, 60f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x23]\n"
       "b 60f\n"
       "59:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s9, [x13, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
       "60:"  // Height 2: Partial accumulate: Done
       "sub x13, x13, x20\n"
@@ -708,12 +706,12 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q12, [x13, #0x30]\n"
       "ldr q13, [x13, #0x40]\n"
       "ldr q20, [x13, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
       "62:"  // Height 2: MMLA fixup
       "zip1 v8.2d, v9.2d, v14.2d\n"
       "zip2 v14.2d, v9.2d, v14.2d\n"
@@ -745,8 +743,8 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "mov x26, #0x0\n"
       "65:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 66f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -773,72 +771,72 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "blt 69f\n"
       "68:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x25, x25, #0x4\n"
-      "add x12, x12, #0x20\n"
-      "cmp x25, #0x8\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       "ld1 { v1.4s }, [x23], #0x10\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
-      "ldr q29, [x10, #0x0]\n"
+      "ldr q30, [x10, #0x0]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
       "ldr q23, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       "ldr q22, [x9, #0x0]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       "ldr q21, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
-      ".inst 0x6e5dec0a  // bfmmla v10.4s, v0.8h, v29.8h\n"
-      "ldr q30, [x28, #0x0]\n"
+      ".inst 0x6e5eec0a  // bfmmla v10.4s, v0.8h, v30.8h\n"
+      "ldr q2, [x28, #0x0]\n"
       ".inst 0x6e57ec10  // bfmmla v16.4s, v0.8h, v23.8h\n"
       "ldr q23, [x28, #0x10]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e56ec0b  // bfmmla v11.4s, v0.8h, v22.8h\n"
       "ldr q22, [x27, #0x0]\n"
       ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
       "ldr q21, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
-      ".inst 0x6e5eec0c  // bfmmla v12.4s, v0.8h, v30.8h\n"
+      "sub x25, x25, #0x4\n"
+      "cmp x25, #0x8\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x6e42ec0c  // bfmmla v12.4s, v0.8h, v2.8h\n"
       "ldr q4, [x12, #0x0]\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
       "ldr q5, [x12, #0x10]\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       "ldr q6, [x11, #0x0]\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
       "ld1 { v0.4s }, [x24], #0x10\n"
+      "add x10, x10, #0x20\n"
       "ldr q7, [x11, #0x10]\n"
+      "add x9, x9, #0x20\n"
+      "add x28, x28, #0x20\n"
+      "add x27, x27, #0x20\n"
       "bge 68b\n"
       "69:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x25, x25, #0x4\n"
-      "add x12, x12, #0x20\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
       "ldr q24, [x10, #0x0]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
       "ldr q23, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       "ldr q22, [x9, #0x0]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       "ldr q21, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e58ec0a  // bfmmla v10.4s, v0.8h, v24.8h\n"
       "ldr q24, [x28, #0x0]\n"
       ".inst 0x6e57ec10  // bfmmla v16.4s, v0.8h, v23.8h\n"
       "ldr q23, [x28, #0x10]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e56ec0b  // bfmmla v11.4s, v0.8h, v22.8h\n"
       "ldr q22, [x27, #0x0]\n"
       ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
       "ldr q21, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
+      "sub x25, x25, #0x4\n"
       ".inst 0x6e58ec0c  // bfmmla v12.4s, v0.8h, v24.8h\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
+      "add x12, x12, #0x20\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
+      "add x28, x28, #0x20\n"
+      "add x27, x27, #0x20\n"
       "70:"  // Height 2: Multiply loop: Main loop skip
       "cbz x25, 73f\n"
       "cbz x25, 73f\n"
@@ -856,35 +854,35 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q24, [x12, #0x0]\n"
       "ldr q23, [x12, #0x10]\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "add x12, x12, #0x20\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       "ldr q22, [x11, #0x0]\n"
       "ldr q21, [x11, #0x10]\n"
-      "add x11, x11, #0x20\n"
-      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x6e58ec08  // bfmmla v8.4s, v0.8h, v24.8h\n"
-      "ldr q24, [x10, #0x0]\n"
       ".inst 0x6e57ec0e  // bfmmla v14.4s, v0.8h, v23.8h\n"
+      "ldr q24, [x10, #0x0]\n"
       "ldr q23, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
       ".inst 0x6e56ec09  // bfmmla v9.4s, v0.8h, v22.8h\n"
-      "ldr q22, [x9, #0x0]\n"
       ".inst 0x6e55ec0f  // bfmmla v15.4s, v0.8h, v21.8h\n"
+      "ldr q22, [x9, #0x0]\n"
       "ldr q21, [x9, #0x10]\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e58ec0a  // bfmmla v10.4s, v0.8h, v24.8h\n"
-      "ldr q24, [x28, #0x0]\n"
       ".inst 0x6e57ec10  // bfmmla v16.4s, v0.8h, v23.8h\n"
+      "ldr q24, [x28, #0x0]\n"
       "ldr q23, [x28, #0x10]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e56ec0b  // bfmmla v11.4s, v0.8h, v22.8h\n"
-      "ldr q22, [x27, #0x0]\n"
       ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
+      "ldr q22, [x27, #0x0]\n"
       "ldr q21, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
       ".inst 0x6e58ec0c  // bfmmla v12.4s, v0.8h, v24.8h\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
+      "add x12, x12, #0x20\n"
+      "add x11, x11, #0x20\n"
+      "add x10, x10, #0x20\n"
+      "add x9, x9, #0x20\n"
+      "add x28, x28, #0x20\n"
+      "add x27, x27, #0x20\n"
       "73:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -893,21 +891,21 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v4.2d, v8.2d, v14.2d\n"
       "uzp2 v8.2d, v8.2d, v14.2d\n"
+      "add x23, x13, x20, LSL #2\n"
       "uzp1 v14.2d, v9.2d, v15.2d\n"
       "uzp2 v9.2d, v9.2d, v15.2d\n"
       "uzp1 v15.2d, v10.2d, v16.2d\n"
       "uzp2 v10.2d, v10.2d, v16.2d\n"
       "uzp1 v16.2d, v11.2d, v17.2d\n"
       "uzp2 v11.2d, v11.2d, v17.2d\n"
-      "add x24, x13, x20, LSL #2\n"
       "uzp1 v17.2d, v12.2d, v18.2d\n"
       "uzp2 v12.2d, v12.2d, v18.2d\n"
       "uzp1 v18.2d, v13.2d, v19.2d\n"
       "uzp2 v13.2d, v13.2d, v19.2d\n"
       "tbz %x[flags], #1, 74f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v22.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v22.4s }, [x21]\n"
       "ld1r { v21.4s }, [x20]\n"
       "fmin v4.4s, v4.4s, v22.4s\n"
       "fmin v14.4s, v14.4s, v22.4s\n"
@@ -941,99 +939,99 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "st1 { v14.4s }, [x13], #0x10\n"
       "st1 { v15.4s }, [x13], #0x10\n"
       "st1 { v16.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v11.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v11.4s }, [x23], #0x10\n"
       "tbz x14, #2, 76f\n"
       "st1 { v17.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x23], #0x10\n"
       "tbz x14, #1, 75f\n"
       "str d18, [x13], #0x8\n"
-      "str d13, [x24], #0x8\n"
+      "str d13, [x23], #0x8\n"
       "tbz x14, #0, 86f\n"
       "st1 { v18.s }[2], [x13]\n"
-      "st1 { v13.s }[2], [x24]\n"
+      "st1 { v13.s }[2], [x23]\n"
       "b 86f\n"
       "75:"  // Height 2: Partial direct writeback: partial_1_20
       "tbz x14, #0, 86f\n"
       "str s18, [x13, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
+      "str s13, [x23, #0x0]\n"
       "b 86f\n"
       "76:"  // Height 2: Partial direct writeback: partial_2_16
       "tbz x14, #1, 77f\n"
       "str d17, [x13], #0x8\n"
-      "str d12, [x24], #0x8\n"
+      "str d12, [x23], #0x8\n"
       "tbz x14, #0, 86f\n"
       "st1 { v17.s }[2], [x13]\n"
-      "st1 { v12.s }[2], [x24]\n"
+      "st1 { v12.s }[2], [x23]\n"
       "b 86f\n"
       "77:"  // Height 2: Partial direct writeback: partial_1_16
       "tbz x14, #0, 86f\n"
       "str s17, [x13, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
+      "str s12, [x23, #0x0]\n"
       "b 86f\n"
       "78:"  // Height 2: Partial direct writeback: partial_8_0
       "tbz x14, #3, 82f\n"
       "st1 { v4.4s }, [x13], #0x10\n"
       "st1 { v14.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
       "tbz x14, #2, 80f\n"
       "st1 { v15.4s }, [x13], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
       "tbz x14, #1, 79f\n"
       "str d16, [x13], #0x8\n"
-      "str d11, [x24], #0x8\n"
+      "str d11, [x23], #0x8\n"
       "tbz x14, #0, 86f\n"
       "st1 { v16.s }[2], [x13]\n"
-      "st1 { v11.s }[2], [x24]\n"
+      "st1 { v11.s }[2], [x23]\n"
       "b 86f\n"
       "79:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x14, #0, 86f\n"
       "str s16, [x13, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
+      "str s11, [x23, #0x0]\n"
       "b 86f\n"
       "80:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x14, #1, 81f\n"
       "str d15, [x13], #0x8\n"
-      "str d10, [x24], #0x8\n"
+      "str d10, [x23], #0x8\n"
       "tbz x14, #0, 86f\n"
       "st1 { v15.s }[2], [x13]\n"
-      "st1 { v10.s }[2], [x24]\n"
+      "st1 { v10.s }[2], [x23]\n"
       "b 86f\n"
       "81:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x14, #0, 86f\n"
       "str s15, [x13, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
+      "str s10, [x23, #0x0]\n"
       "b 86f\n"
       "82:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x14, #2, 84f\n"
       "st1 { v4.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
       "tbz x14, #1, 83f\n"
       "str d14, [x13], #0x8\n"
-      "str d9, [x24], #0x8\n"
+      "str d9, [x23], #0x8\n"
       "tbz x14, #0, 86f\n"
       "st1 { v14.s }[2], [x13]\n"
-      "st1 { v9.s }[2], [x24]\n"
+      "st1 { v9.s }[2], [x23]\n"
       "b 86f\n"
       "83:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x14, #0, 86f\n"
       "str s14, [x13, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
+      "str s9, [x23, #0x0]\n"
       "b 86f\n"
       "84:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x14, #1, 85f\n"
       "str d4, [x13], #0x8\n"
-      "str d8, [x24], #0x8\n"
+      "str d8, [x23], #0x8\n"
       "tbz x14, #0, 86f\n"
       "st1 { v4.s }[2], [x13]\n"
-      "st1 { v8.s }[2], [x24]\n"
+      "st1 { v8.s }[2], [x23]\n"
       "b 86f\n"
       "85:"  // Height 2: Partial direct writeback: partial_1_0
       "str s4, [x13, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
+      "str s8, [x23, #0x0]\n"
       "86:"  // Height 2: Partial direct writeback: Done
       "b 88f\n"
       "87:"  // Height 2: Full writeback
@@ -1044,26 +1042,25 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "str q17, [x13, #0x40]\n"
       "str q18, [x13, #0x50]\n"
       "add x13, x13, #0x60\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q12, [x24, #0x40]\n"
-      "str q13, [x24, #0x50]\n"
+      "str q8, [x23, #0x0]\n"
+      "str q9, [x23, #0x10]\n"
+      "str q10, [x23, #0x20]\n"
+      "str q11, [x23, #0x30]\n"
+      "str q12, [x23, #0x40]\n"
+      "str q13, [x23, #0x50]\n"
       "88:"  // Height 2: Writeback done
       "subs x14, x14, #0x18\n"
       "bgt 46b\n"
       "b 178f\n"
       "89:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "90:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x14\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
@@ -1071,6 +1068,7 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "add x27, x28, x20, LSL #1\n"
       "add x20, x27, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x14\n"
       "bgt 91f\n"
       "cmp x14, #0x10\n"
       "mov x27, x12\n"
@@ -1089,19 +1087,19 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "cbz x15, 92f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "ldr q12, [x15, #0x40]\n"
-      "ldr q13, [x15, #0x50]\n"
-      "add x15, x15, #0x60\n"
       "zip2 v14.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v15.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
+      "ldr q12, [x15, #0x40]\n"
+      "ldr q13, [x15, #0x50]\n"
       "zip2 v16.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
       "zip2 v17.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
+      "add x15, x15, #0x60\n"
       "zip2 v18.2d, v12.2d, v12.2d\n"
       "zip1 v12.2d, v12.2d, v12.2d\n"
       "zip2 v19.2d, v13.2d, v13.2d\n"
@@ -1122,147 +1120,147 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "92:"  // Height 3: no bias
       "tbz %x[flags], #0, 107f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #2\n"
       "cmp x14, #0x18\n"
-      "add x24, x13, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 105f\n"
       "tbz x14, #4, 96f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
       "ld1 { v12.4s }, [x13], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "tbz x14, #2, 94f\n"
       "ld1 { v13.4s }, [x13], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x14, #1, 93f\n"
       "ldr d20, [x13], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d4, [x23], #0x8\n"
+      "ldr d4, [x22], #0x8\n"
       "tbz x14, #0, 104f\n"
       "ld1 { v20.s }[2], [x13]\n"
-      "ld1 { v19.s }[2], [x24]\n"
-      "ld1 { v4.s }[2], [x23]\n"
+      "ld1 { v19.s }[2], [x23]\n"
+      "ld1 { v4.s }[2], [x22]\n"
       "b 104f\n"
       "93:"  // Height 3: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x14, #0, 104f\n"
       "ldr s20, [x13, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
-      "ldr s4, [x23, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
+      "ldr s4, [x22, #0x0]\n"
       "b 104f\n"
       "94:"  // Height 3: Partial accumulate: partial_2_16
       "tbz x14, #1, 95f\n"
       "ldr d13, [x13], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x14, #0, 104f\n"
       "ld1 { v13.s }[2], [x13]\n"
-      "ld1 { v18.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v18.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 104f\n"
       "95:"  // Height 3: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x14, #0, 104f\n"
       "ldr s13, [x13, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "b 104f\n"
       "96:"  // Height 3: Partial accumulate: partial_8_0
       "tbz x14, #3, 100f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
       "tbz x14, #2, 98f\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
       "tbz x14, #1, 97f\n"
       "ldr d12, [x13], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x14, #0, 104f\n"
       "ld1 { v12.s }[2], [x13]\n"
-      "ld1 { v17.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v17.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
       "b 104f\n"
       "97:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 104f\n"
       "ldr s12, [x13, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
       "b 104f\n"
       "98:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x14, #1, 99f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d23, [x23], #0x8\n"
+      "ldr d23, [x22], #0x8\n"
       "tbz x14, #0, 104f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v16.s }[2], [x24]\n"
-      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v16.s }[2], [x23]\n"
+      "ld1 { v23.s }[2], [x22]\n"
       "b 104f\n"
       "99:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 104f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
-      "ldr s23, [x23, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
+      "ldr s23, [x22, #0x0]\n"
       "b 104f\n"
       "100:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x14, #2, 102f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "tbz x14, #1, 101f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d22, [x23], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
       "tbz x14, #0, 104f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
       "b 104f\n"
       "101:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 104f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
       "b 104f\n"
       "102:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x14, #1, 103f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d21, [x23], #0x8\n"
+      "ldr d21, [x22], #0x8\n"
       "tbz x14, #0, 104f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x23]\n"
+      "ld1 { v21.s }[2], [x22]\n"
       "b 104f\n"
       "103:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s9, [x13, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s21, [x23, #0x0]\n"
+      "ldr s21, [x22, #0x0]\n"
       "104:"  // Height 3: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 106f\n"
@@ -1273,18 +1271,18 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q12, [x13, #0x30]\n"
       "ldr q13, [x13, #0x40]\n"
       "ldr q20, [x13, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
-      "ldr q21, [x23, #0x0]\n"
-      "ldr q22, [x23, #0x10]\n"
-      "ldr q23, [x23, #0x20]\n"
-      "ldr q24, [x23, #0x30]\n"
-      "ldr q25, [x23, #0x40]\n"
-      "ldr q4, [x23, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
+      "ldr q21, [x22, #0x0]\n"
+      "ldr q22, [x22, #0x10]\n"
+      "ldr q23, [x22, #0x20]\n"
+      "ldr q24, [x22, #0x30]\n"
+      "ldr q25, [x22, #0x40]\n"
+      "ldr q4, [x22, #0x50]\n"
       "106:"  // Height 3: MMLA fixup
       "zip1 v8.2d, v9.2d, v14.2d\n"
       "zip2 v14.2d, v9.2d, v14.2d\n"
@@ -1340,8 +1338,8 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "mov x26, #0x0\n"
       "109:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 110f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1372,42 +1370,42 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "blt 113f\n"
       "112:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "sub x25, x25, #0x4\n"
-      "add x12, x12, #0x20\n"
-      "cmp x25, #0x8\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       "ld1 { v1.4s }, [x23], #0x10\n"
-      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
-      ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
-      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
       "ldr q4, [x10, #0x0]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
       "ldr q5, [x10, #0x10]\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      "sub x25, x25, #0x4\n"
+      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
       "ldr q6, [x9, #0x0]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "cmp x25, #0x8\n"
+      ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       "ldr q3, [x9, #0x10]\n"
-      "add x10, x10, #0x20\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e44ec0a  // bfmmla v10.4s, v0.8h, v4.8h\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e44ec56  // bfmmla v22.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x0]\n"
       ".inst 0x6e45ec10  // bfmmla v16.4s, v0.8h, v5.8h\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e45ec5c  // bfmmla v28.4s, v2.8h, v5.8h\n"
       "ldr q5, [x28, #0x10]\n"
       ".inst 0x6e46ec0b  // bfmmla v11.4s, v0.8h, v6.8h\n"
+      "add x10, x10, #0x20\n"
       ".inst 0x6e46ec57  // bfmmla v23.4s, v2.8h, v6.8h\n"
       "ldr q6, [x27, #0x0]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e43ec11  // bfmmla v17.4s, v0.8h, v3.8h\n"
+      "add x9, x9, #0x20\n"
       ".inst 0x6e43ec5d  // bfmmla v29.4s, v2.8h, v3.8h\n"
       "ldr q3, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
+      "add x28, x28, #0x20\n"
       ".inst 0x6e44ec0c  // bfmmla v12.4s, v0.8h, v4.8h\n"
+      "add x27, x27, #0x20\n"
       ".inst 0x6e44ec58  // bfmmla v24.4s, v2.8h, v4.8h\n"
       "ldr q4, [x12, #0x0]\n"
       ".inst 0x6e45ec12  // bfmmla v18.4s, v0.8h, v5.8h\n"
@@ -1424,35 +1422,35 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "bge 112b\n"
       "113:"  // Height 3: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "sub x25, x25, #0x4\n"
-      "add x12, x12, #0x20\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
-      ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
-      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
+      "sub x25, x25, #0x4\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
       "ldr q3, [x10, #0x0]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
       "ldr q4, [x10, #0x10]\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
       "ldr q6, [x9, #0x0]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       "ldr q1, [x9, #0x10]\n"
-      "add x10, x10, #0x20\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
+      "add x10, x10, #0x20\n"
       ".inst 0x6e43ec56  // bfmmla v22.4s, v2.8h, v3.8h\n"
       "ldr q5, [x28, #0x0]\n"
       ".inst 0x6e44ec10  // bfmmla v16.4s, v0.8h, v4.8h\n"
+      "add x9, x9, #0x20\n"
       ".inst 0x6e44ec5c  // bfmmla v28.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x10]\n"
       ".inst 0x6e46ec0b  // bfmmla v11.4s, v0.8h, v6.8h\n"
+      "add x28, x28, #0x20\n"
       ".inst 0x6e46ec57  // bfmmla v23.4s, v2.8h, v6.8h\n"
       "ldr q3, [x27, #0x0]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e41ec11  // bfmmla v17.4s, v0.8h, v1.8h\n"
       ".inst 0x6e41ec5d  // bfmmla v29.4s, v2.8h, v1.8h\n"
       "ldr q1, [x27, #0x10]\n"
@@ -1485,41 +1483,41 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q5, [x12, #0x0]\n"
       "ldr q4, [x12, #0x10]\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "ldr q3, [x11, #0x0]\n"
-      "ldr q6, [x11, #0x10]\n"
-      "add x12, x12, #0x20\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      ".inst 0x6e45ec54  // bfmmla v20.4s, v2.8h, v5.8h\n"
-      ".inst 0x6e44ec5a  // bfmmla v26.4s, v2.8h, v4.8h\n"
-      ".inst 0x6e43ec55  // bfmmla v21.4s, v2.8h, v3.8h\n"
-      ".inst 0x6e46ec5b  // bfmmla v27.4s, v2.8h, v6.8h\n"
+      "ldr q3, [x11, #0x0]\n"
+      "ldr q1, [x11, #0x10]\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       ".inst 0x6e45ec08  // bfmmla v8.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec54  // bfmmla v20.4s, v2.8h, v5.8h\n"
       "ldr q5, [x10, #0x0]\n"
       ".inst 0x6e44ec0e  // bfmmla v14.4s, v0.8h, v4.8h\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x6e44ec5a  // bfmmla v26.4s, v2.8h, v4.8h\n"
       "ldr q4, [x10, #0x10]\n"
       ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e43ec55  // bfmmla v21.4s, v2.8h, v3.8h\n"
       "ldr q3, [x9, #0x0]\n"
-      ".inst 0x6e46ec0f  // bfmmla v15.4s, v0.8h, v6.8h\n"
-      "ldr q1, [x9, #0x10]\n"
+      ".inst 0x6e41ec0f  // bfmmla v15.4s, v0.8h, v1.8h\n"
       "add x10, x10, #0x20\n"
-      "add x9, x9, #0x20\n"
+      ".inst 0x6e41ec5b  // bfmmla v27.4s, v2.8h, v1.8h\n"
+      "ldr q1, [x9, #0x10]\n"
       ".inst 0x6e45ec0a  // bfmmla v10.4s, v0.8h, v5.8h\n"
+      "add x9, x9, #0x20\n"
       ".inst 0x6e45ec56  // bfmmla v22.4s, v2.8h, v5.8h\n"
       "ldr q5, [x28, #0x0]\n"
       ".inst 0x6e44ec10  // bfmmla v16.4s, v0.8h, v4.8h\n"
       ".inst 0x6e44ec5c  // bfmmla v28.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x10]\n"
       ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
+      "add x28, x28, #0x20\n"
       ".inst 0x6e43ec57  // bfmmla v23.4s, v2.8h, v3.8h\n"
       "ldr q3, [x27, #0x0]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e41ec11  // bfmmla v17.4s, v0.8h, v1.8h\n"
       ".inst 0x6e41ec5d  // bfmmla v29.4s, v2.8h, v1.8h\n"
       "ldr q1, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
       ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
+      "add x27, x27, #0x20\n"
       ".inst 0x6e45ec58  // bfmmla v24.4s, v2.8h, v5.8h\n"
       ".inst 0x6e44ec12  // bfmmla v18.4s, v0.8h, v4.8h\n"
       ".inst 0x6e44ec5e  // bfmmla v30.4s, v2.8h, v4.8h\n"
@@ -1533,16 +1531,16 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "cmp x26, x20\n"
       "bne 109b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #2\n"
       "uzp1 v4.2d, v8.2d, v14.2d\n"
       "uzp2 v8.2d, v8.2d, v14.2d\n"
       "uzp1 v14.2d, v9.2d, v15.2d\n"
       "uzp2 v9.2d, v9.2d, v15.2d\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp1 v15.2d, v10.2d, v16.2d\n"
       "uzp2 v10.2d, v10.2d, v16.2d\n"
-      "add x24, x13, x20, LSL #2\n"
       "uzp1 v16.2d, v11.2d, v17.2d\n"
       "uzp2 v11.2d, v11.2d, v17.2d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v17.2d, v12.2d, v18.2d\n"
       "uzp2 v12.2d, v12.2d, v18.2d\n"
       "uzp1 v18.2d, v13.2d, v19.2d\n"
@@ -1554,9 +1552,9 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "uzp1 v24.2d, v24.2d, v30.2d\n"
       "uzp1 v25.2d, v25.2d, v31.2d\n"
       "tbz %x[flags], #1, 118f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v4.4s, v4.4s, v1.4s\n"
       "fmin v14.4s, v14.4s, v1.4s\n"
@@ -1602,126 +1600,126 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "st1 { v14.4s }, [x13], #0x10\n"
       "st1 { v15.4s }, [x13], #0x10\n"
       "st1 { v16.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v11.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v11.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
       "tbz x14, #2, 120f\n"
       "st1 { v17.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x14, #1, 119f\n"
       "str d18, [x13], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x14, #0, 130f\n"
       "st1 { v18.s }[2], [x13]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v13.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 130f\n"
       "119:"  // Height 3: Partial direct writeback: partial_1_20
       "tbz x14, #0, 130f\n"
       "str s18, [x13, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s13, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 130f\n"
       "120:"  // Height 3: Partial direct writeback: partial_2_16
       "tbz x14, #1, 121f\n"
       "str d17, [x13], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x14, #0, 130f\n"
       "st1 { v17.s }[2], [x13]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v12.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 130f\n"
       "121:"  // Height 3: Partial direct writeback: partial_1_16
       "tbz x14, #0, 130f\n"
       "str s17, [x13, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s12, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "b 130f\n"
       "122:"  // Height 3: Partial direct writeback: partial_8_0
       "tbz x14, #3, 126f\n"
       "st1 { v4.4s }, [x13], #0x10\n"
       "st1 { v14.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
       "tbz x14, #2, 124f\n"
       "st1 { v15.4s }, [x13], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
       "tbz x14, #1, 123f\n"
       "str d16, [x13], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
+      "str d11, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
       "tbz x14, #0, 130f\n"
       "st1 { v16.s }[2], [x13]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v11.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
       "b 130f\n"
       "123:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x14, #0, 130f\n"
       "str s16, [x13, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
+      "str s11, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
       "b 130f\n"
       "124:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x14, #1, 125f\n"
       "str d15, [x13], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
+      "str d10, [x23], #0x8\n"
+      "str d22, [x22], #0x8\n"
       "tbz x14, #0, 130f\n"
       "st1 { v15.s }[2], [x13]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v10.s }[2], [x23]\n"
+      "st1 { v22.s }[2], [x22]\n"
       "b 130f\n"
       "125:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x14, #0, 130f\n"
       "str s15, [x13, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
+      "str s10, [x23, #0x0]\n"
+      "str s22, [x22, #0x0]\n"
       "b 130f\n"
       "126:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x14, #2, 128f\n"
       "st1 { v4.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
       "tbz x14, #1, 127f\n"
       "str d14, [x13], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
+      "str d9, [x23], #0x8\n"
+      "str d21, [x22], #0x8\n"
       "tbz x14, #0, 130f\n"
       "st1 { v14.s }[2], [x13]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v9.s }[2], [x23]\n"
+      "st1 { v21.s }[2], [x22]\n"
       "b 130f\n"
       "127:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x14, #0, 130f\n"
       "str s14, [x13, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
+      "str s9, [x23, #0x0]\n"
+      "str s21, [x22, #0x0]\n"
       "b 130f\n"
       "128:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x14, #1, 129f\n"
       "str d4, [x13], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
+      "str d8, [x23], #0x8\n"
+      "str d20, [x22], #0x8\n"
       "tbz x14, #0, 130f\n"
       "st1 { v4.s }[2], [x13]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v8.s }[2], [x23]\n"
+      "st1 { v20.s }[2], [x22]\n"
       "b 130f\n"
       "129:"  // Height 3: Partial direct writeback: partial_1_0
       "str s4, [x13, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
+      "str s8, [x23, #0x0]\n"
+      "str s20, [x22, #0x0]\n"
       "130:"  // Height 3: Partial direct writeback: Done
       "b 132f\n"
       "131:"  // Height 3: Full writeback
@@ -1732,36 +1730,34 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "str q17, [x13, #0x40]\n"
       "str q18, [x13, #0x50]\n"
       "add x13, x13, #0x60\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q12, [x24, #0x40]\n"
-      "str q13, [x24, #0x50]\n"
-      "str q20, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q22, [x23, #0x20]\n"
-      "str q23, [x23, #0x30]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q25, [x23, #0x50]\n"
+      "str q8, [x23, #0x0]\n"
+      "str q9, [x23, #0x10]\n"
+      "str q10, [x23, #0x20]\n"
+      "str q11, [x23, #0x30]\n"
+      "str q12, [x23, #0x40]\n"
+      "str q13, [x23, #0x50]\n"
+      "str q20, [x22, #0x0]\n"
+      "str q21, [x22, #0x10]\n"
+      "str q22, [x22, #0x20]\n"
+      "str q23, [x22, #0x30]\n"
+      "str q24, [x22, #0x40]\n"
+      "str q25, [x22, #0x50]\n"
       "132:"  // Height 3: Writeback done
       "subs x14, x14, #0x18\n"
       "bgt 90b\n"
       "b 178f\n"
       "133:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0x10\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0x10\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "134:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x14, #0x14\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
@@ -1769,6 +1765,7 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "add x27, x28, x20, LSL #1\n"
       "add x20, x27, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x14, #0x14\n"
       "bgt 135f\n"
       "cmp x14, #0x10\n"
       "mov x27, x12\n"
@@ -1787,19 +1784,19 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "cbz x15, 136f\n"
       "ldr q8, [x15, #0x0]\n"
       "ldr q9, [x15, #0x10]\n"
-      "ldr q10, [x15, #0x20]\n"
-      "ldr q11, [x15, #0x30]\n"
-      "ldr q12, [x15, #0x40]\n"
-      "ldr q13, [x15, #0x50]\n"
-      "add x15, x15, #0x60\n"
       "zip2 v14.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x15, #0x20]\n"
+      "ldr q11, [x15, #0x30]\n"
       "zip2 v15.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
+      "ldr q12, [x15, #0x40]\n"
+      "ldr q13, [x15, #0x50]\n"
       "zip2 v16.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
       "zip2 v17.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
+      "add x15, x15, #0x60\n"
       "zip2 v18.2d, v12.2d, v12.2d\n"
       "zip1 v12.2d, v12.2d, v12.2d\n"
       "zip2 v19.2d, v13.2d, v13.2d\n"
@@ -1820,175 +1817,175 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "136:"  // Height 4: no bias
       "tbz %x[flags], #0, 151f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x14, #0x18\n"
-      "add x24, x13, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x23, x13, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x14, #0x18\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 149f\n"
       "tbz x14, #4, 140f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
-      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
+      "ld1 { v27.4s }, [x21], #0x10\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v12.4s }, [x13], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x14, #2, 138f\n"
       "ld1 { v13.4s }, [x13], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x14, #1, 137f\n"
       "ldr d20, [x13], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d4, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d4, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v20.s }[2], [x13]\n"
-      "ld1 { v19.s }[2], [x24]\n"
-      "ld1 { v4.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v19.s }[2], [x23]\n"
+      "ld1 { v4.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 148f\n"
       "137:"  // Height 4: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x14, #0, 148f\n"
       "ldr s20, [x13, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
-      "ldr s4, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
+      "ldr s4, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 148f\n"
       "138:"  // Height 4: Partial accumulate: partial_2_16
       "tbz x14, #1, 139f\n"
       "ldr d13, [x13], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v13.s }[2], [x13]\n"
-      "ld1 { v18.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v18.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 148f\n"
       "139:"  // Height 4: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x14, #0, 148f\n"
       "ldr s13, [x13, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 148f\n"
       "140:"  // Height 4: Partial accumulate: partial_8_0
       "tbz x14, #3, 144f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x13], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
-      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
+      "ld1 { v27.4s }, [x21], #0x10\n"
       "tbz x14, #2, 142f\n"
       "ld1 { v11.4s }, [x13], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x14, #1, 141f\n"
       "ldr d12, [x13], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v12.s }[2], [x13]\n"
-      "ld1 { v17.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v17.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 148f\n"
       "141:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x14, #0, 148f\n"
       "ldr s12, [x13, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 148f\n"
       "142:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x14, #1, 143f\n"
       "ldr d11, [x13], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d23, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d23, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v11.s }[2], [x13]\n"
-      "ld1 { v16.s }[2], [x24]\n"
-      "ld1 { v23.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v16.s }[2], [x23]\n"
+      "ld1 { v23.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 148f\n"
       "143:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x14, #0, 148f\n"
       "ldr s11, [x13, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
-      "ldr s23, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
+      "ldr s23, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "b 148f\n"
       "144:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x14, #2, 146f\n"
       "ld1 { v9.4s }, [x13], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "tbz x14, #1, 145f\n"
       "ldr d10, [x13], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d27, [x22], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
+      "ldr d27, [x21], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v10.s }[2], [x13]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
-      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
+      "ld1 { v27.s }[2], [x21]\n"
       "b 148f\n"
       "145:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x14, #0, 148f\n"
       "ldr s10, [x13, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
-      "ldr s27, [x22, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
+      "ldr s27, [x21, #0x0]\n"
       "b 148f\n"
       "146:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x14, #1, 147f\n"
       "ldr d9, [x13], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d21, [x23], #0x8\n"
-      "ldr d26, [x22], #0x8\n"
+      "ldr d21, [x22], #0x8\n"
+      "ldr d26, [x21], #0x8\n"
       "tbz x14, #0, 148f\n"
       "ld1 { v9.s }[2], [x13]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v21.s }[2], [x23]\n"
-      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x23]\n"
+      "ld1 { v21.s }[2], [x22]\n"
+      "ld1 { v26.s }[2], [x21]\n"
       "b 148f\n"
       "147:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s9, [x13, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s21, [x23, #0x0]\n"
-      "ldr s26, [x22, #0x0]\n"
+      "ldr s21, [x22, #0x0]\n"
+      "ldr s26, [x21, #0x0]\n"
       "148:"  // Height 4: Partial accumulate: Done
       "sub x13, x13, x20\n"
       "b 150f\n"
@@ -1999,24 +1996,24 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q12, [x13, #0x30]\n"
       "ldr q13, [x13, #0x40]\n"
       "ldr q20, [x13, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
-      "ldr q21, [x23, #0x0]\n"
-      "ldr q22, [x23, #0x10]\n"
-      "ldr q23, [x23, #0x20]\n"
-      "ldr q24, [x23, #0x30]\n"
-      "ldr q25, [x23, #0x40]\n"
-      "ldr q4, [x23, #0x50]\n"
-      "ldr q26, [x22, #0x0]\n"
-      "ldr q27, [x22, #0x10]\n"
-      "ldr q28, [x22, #0x20]\n"
-      "ldr q29, [x22, #0x30]\n"
-      "ldr q30, [x22, #0x40]\n"
-      "ldr q31, [x22, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
+      "ldr q21, [x22, #0x0]\n"
+      "ldr q22, [x22, #0x10]\n"
+      "ldr q23, [x22, #0x20]\n"
+      "ldr q24, [x22, #0x30]\n"
+      "ldr q25, [x22, #0x40]\n"
+      "ldr q4, [x22, #0x50]\n"
+      "ldr q26, [x21, #0x0]\n"
+      "ldr q27, [x21, #0x10]\n"
+      "ldr q28, [x21, #0x20]\n"
+      "ldr q29, [x21, #0x30]\n"
+      "ldr q30, [x21, #0x40]\n"
+      "ldr q31, [x21, #0x50]\n"
       "150:"  // Height 4: MMLA fixup
       "zip1 v8.2d, v9.2d, v14.2d\n"
       "zip2 v14.2d, v9.2d, v14.2d\n"
@@ -2072,8 +2069,8 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "mov x26, #0x0\n"
       "153:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 154f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2110,26 +2107,26 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x25, x25, #0x4\n"
-      "add x12, x12, #0x20\n"
       "cmp x25, #0x8\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       "ld1 { v1.4s }, [x23], #0x10\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
       "ld1 { v3.4s }, [x21], #0x10\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
-      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
-      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
       "ldr q4, [x10, #0x0]\n"
+      "add x12, x12, #0x20\n"
+      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
       "ldr q5, [x10, #0x10]\n"
+      "add x11, x11, #0x20\n"
+      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
       "ldr q6, [x9, #0x0]\n"
+      "add x10, x10, #0x20\n"
+      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       "ldr q7, [x9, #0x10]\n"
-      "add x10, x10, #0x20\n"
       "add x9, x9, #0x20\n"
       ".inst 0x6e44ec0a  // bfmmla v10.4s, v0.8h, v4.8h\n"
       ".inst 0x6e44ec56  // bfmmla v22.4s, v2.8h, v4.8h\n"
@@ -2137,10 +2134,10 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       ".inst 0x6e45ec10  // bfmmla v16.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec5c  // bfmmla v28.4s, v2.8h, v5.8h\n"
       "ldr q5, [x28, #0x10]\n"
+      "add x28, x28, #0x20\n"
       ".inst 0x6e46ec0b  // bfmmla v11.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec57  // bfmmla v23.4s, v2.8h, v6.8h\n"
       "ldr q6, [x27, #0x0]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e47ec11  // bfmmla v17.4s, v0.8h, v7.8h\n"
       ".inst 0x6e47ec5d  // bfmmla v29.4s, v2.8h, v7.8h\n"
       "ldr q7, [x27, #0x10]\n"
@@ -2165,31 +2162,31 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x25, x25, #0x4\n"
       "add x12, x12, #0x20\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
-      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
-      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
       "ldr q3, [x10, #0x0]\n"
+      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
       "ldr q4, [x10, #0x10]\n"
+      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      "add x10, x10, #0x20\n"
       ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
       "ldr q6, [x9, #0x0]\n"
+      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       "ldr q1, [x9, #0x10]\n"
-      "add x10, x10, #0x20\n"
-      "add x9, x9, #0x20\n"
       ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
+      "add x9, x9, #0x20\n"
       ".inst 0x6e43ec56  // bfmmla v22.4s, v2.8h, v3.8h\n"
       "ldr q5, [x28, #0x0]\n"
       ".inst 0x6e44ec10  // bfmmla v16.4s, v0.8h, v4.8h\n"
       ".inst 0x6e44ec5c  // bfmmla v28.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x10]\n"
-      "add x28, x28, #0x20\n"
       ".inst 0x6e46ec0b  // bfmmla v11.4s, v0.8h, v6.8h\n"
+      "add x28, x28, #0x20\n"
       ".inst 0x6e46ec57  // bfmmla v23.4s, v2.8h, v6.8h\n"
       "ldr q3, [x27, #0x0]\n"
       ".inst 0x6e41ec11  // bfmmla v17.4s, v0.8h, v1.8h\n"
@@ -2230,39 +2227,39 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "ldr q7, [x11, #0x0]\n"
       "ldr q6, [x11, #0x10]\n"
-      "add x12, x12, #0x20\n"
-      "add x11, x11, #0x20\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
       ".inst 0x6e45ec08  // bfmmla v8.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec54  // bfmmla v20.4s, v2.8h, v5.8h\n"
       "ldr q5, [x10, #0x0]\n"
+      "add x12, x12, #0x20\n"
       ".inst 0x6e44ec0e  // bfmmla v14.4s, v0.8h, v4.8h\n"
       ".inst 0x6e44ec5a  // bfmmla v26.4s, v2.8h, v4.8h\n"
       "ldr q4, [x10, #0x10]\n"
-      "add x10, x10, #0x20\n"
+      "add x11, x11, #0x20\n"
       ".inst 0x6e47ec09  // bfmmla v9.4s, v0.8h, v7.8h\n"
       ".inst 0x6e47ec55  // bfmmla v21.4s, v2.8h, v7.8h\n"
       "ldr q3, [x9, #0x0]\n"
+      "add x10, x10, #0x20\n"
       ".inst 0x6e46ec0f  // bfmmla v15.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec5b  // bfmmla v27.4s, v2.8h, v6.8h\n"
       "ldr q1, [x9, #0x10]\n"
-      ".inst 0x6e45ec0a  // bfmmla v10.4s, v0.8h, v5.8h\n"
       "add x9, x9, #0x20\n"
+      ".inst 0x6e45ec0a  // bfmmla v10.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec56  // bfmmla v22.4s, v2.8h, v5.8h\n"
       "ldr q5, [x28, #0x0]\n"
       ".inst 0x6e44ec10  // bfmmla v16.4s, v0.8h, v4.8h\n"
       ".inst 0x6e44ec5c  // bfmmla v28.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x10]\n"
-      ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
       "add x28, x28, #0x20\n"
+      ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
       ".inst 0x6e43ec57  // bfmmla v23.4s, v2.8h, v3.8h\n"
       "ldr q3, [x27, #0x0]\n"
       ".inst 0x6e41ec11  // bfmmla v17.4s, v0.8h, v1.8h\n"
       ".inst 0x6e41ec5d  // bfmmla v29.4s, v2.8h, v1.8h\n"
       "ldr q1, [x27, #0x10]\n"
-      ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
       "add x27, x27, #0x20\n"
+      ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec58  // bfmmla v24.4s, v2.8h, v5.8h\n"
       ".inst 0x6e44ec12  // bfmmla v18.4s, v0.8h, v4.8h\n"
       ".inst 0x6e44ec5e  // bfmmla v30.4s, v2.8h, v4.8h\n"
@@ -2276,17 +2273,17 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "cmp x26, x20\n"
       "bne 153b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp1 v4.2d, v8.2d, v14.2d\n"
       "uzp2 v8.2d, v8.2d, v14.2d\n"
       "uzp1 v14.2d, v9.2d, v15.2d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v9.2d, v9.2d, v15.2d\n"
       "uzp1 v15.2d, v10.2d, v16.2d\n"
       "uzp2 v10.2d, v10.2d, v16.2d\n"
-      "add x24, x13, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v16.2d, v11.2d, v17.2d\n"
       "uzp2 v11.2d, v11.2d, v17.2d\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v17.2d, v12.2d, v18.2d\n"
       "uzp2 v12.2d, v12.2d, v18.2d\n"
       "uzp1 v18.2d, v13.2d, v19.2d\n"
@@ -2304,9 +2301,9 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "uzp1 v30.2d, v25.2d, v31.2d\n"
       "uzp2 v25.2d, v25.2d, v31.2d\n"
       "tbz %x[flags], #1, 162f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v4.4s, v4.4s, v1.4s\n"
       "fmin v14.4s, v14.4s, v1.4s\n"
@@ -2364,153 +2361,153 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "st1 { v14.4s }, [x13], #0x10\n"
       "st1 { v15.4s }, [x13], #0x10\n"
       "st1 { v16.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v11.4s }, [x24], #0x10\n"
-      "st1 { v19.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
-      "st1 { v27.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v21.4s }, [x22], #0x10\n"
-      "st1 { v22.4s }, [x22], #0x10\n"
-      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v11.4s }, [x23], #0x10\n"
+      "st1 { v19.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v27.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v20.4s }, [x21], #0x10\n"
+      "st1 { v21.4s }, [x21], #0x10\n"
+      "st1 { v22.4s }, [x21], #0x10\n"
+      "st1 { v23.4s }, [x21], #0x10\n"
       "tbz x14, #2, 164f\n"
       "st1 { v17.4s }, [x13], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v29.4s }, [x23], #0x10\n"
-      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x23], #0x10\n"
+      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v24.4s }, [x21], #0x10\n"
       "tbz x14, #1, 163f\n"
       "str d18, [x13], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
-      "str d25, [x22], #0x8\n"
+      "str d13, [x23], #0x8\n"
+      "str d30, [x22], #0x8\n"
+      "str d25, [x21], #0x8\n"
       "tbz x14, #0, 174f\n"
       "st1 { v18.s }[2], [x13]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
-      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v13.s }[2], [x23]\n"
+      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v25.s }[2], [x21]\n"
       "b 174f\n"
       "163:"  // Height 4: Partial direct writeback: partial_1_20
       "tbz x14, #0, 174f\n"
       "str s18, [x13, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
-      "str s25, [x22, #0x0]\n"
+      "str s13, [x23, #0x0]\n"
+      "str s30, [x22, #0x0]\n"
+      "str s25, [x21, #0x0]\n"
       "b 174f\n"
       "164:"  // Height 4: Partial direct writeback: partial_2_16
       "tbz x14, #1, 165f\n"
       "str d17, [x13], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "str d29, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
+      "str d12, [x23], #0x8\n"
+      "str d29, [x22], #0x8\n"
+      "str d24, [x21], #0x8\n"
       "tbz x14, #0, 174f\n"
       "st1 { v17.s }[2], [x13]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "st1 { v29.s }[2], [x23]\n"
-      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v12.s }[2], [x23]\n"
+      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v24.s }[2], [x21]\n"
       "b 174f\n"
       "165:"  // Height 4: Partial direct writeback: partial_1_16
       "tbz x14, #0, 174f\n"
       "str s17, [x13, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "str s29, [x23, #0x0]\n"
-      "str s24, [x22, #0x0]\n"
+      "str s12, [x23, #0x0]\n"
+      "str s29, [x22, #0x0]\n"
+      "str s24, [x21, #0x0]\n"
       "b 174f\n"
       "166:"  // Height 4: Partial direct writeback: partial_8_0
       "tbz x14, #3, 170f\n"
       "st1 { v4.4s }, [x13], #0x10\n"
       "st1 { v14.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v19.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v19.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v20.4s }, [x21], #0x10\n"
+      "st1 { v21.4s }, [x21], #0x10\n"
       "tbz x14, #2, 168f\n"
       "st1 { v15.4s }, [x13], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v27.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v27.4s }, [x22], #0x10\n"
+      "st1 { v22.4s }, [x21], #0x10\n"
       "tbz x14, #1, 167f\n"
       "str d16, [x13], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "str d28, [x23], #0x8\n"
-      "str d23, [x22], #0x8\n"
+      "str d11, [x23], #0x8\n"
+      "str d28, [x22], #0x8\n"
+      "str d23, [x21], #0x8\n"
       "tbz x14, #0, 174f\n"
       "st1 { v16.s }[2], [x13]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "st1 { v28.s }[2], [x23]\n"
-      "st1 { v23.s }[2], [x22]\n"
+      "st1 { v11.s }[2], [x23]\n"
+      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v23.s }[2], [x21]\n"
       "b 174f\n"
       "167:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x14, #0, 174f\n"
       "str s16, [x13, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "str s28, [x23, #0x0]\n"
-      "str s23, [x22, #0x0]\n"
+      "str s11, [x23, #0x0]\n"
+      "str s28, [x22, #0x0]\n"
+      "str s23, [x21, #0x0]\n"
       "b 174f\n"
       "168:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x14, #1, 169f\n"
       "str d15, [x13], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d22, [x22], #0x8\n"
+      "str d10, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d22, [x21], #0x8\n"
       "tbz x14, #0, 174f\n"
       "st1 { v15.s }[2], [x13]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
-      "st1 { v22.s }[2], [x22]\n"
+      "st1 { v10.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v22.s }[2], [x21]\n"
       "b 174f\n"
       "169:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x14, #0, 174f\n"
       "str s15, [x13, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
-      "str s22, [x22, #0x0]\n"
+      "str s10, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
+      "str s22, [x21, #0x0]\n"
       "b 174f\n"
       "170:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x14, #2, 172f\n"
       "st1 { v4.4s }, [x13], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v19.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v19.4s }, [x22], #0x10\n"
+      "st1 { v20.4s }, [x21], #0x10\n"
       "tbz x14, #1, 171f\n"
       "str d14, [x13], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d21, [x22], #0x8\n"
+      "str d9, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d21, [x21], #0x8\n"
       "tbz x14, #0, 174f\n"
       "st1 { v14.s }[2], [x13]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
-      "st1 { v21.s }[2], [x22]\n"
+      "st1 { v9.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v21.s }[2], [x21]\n"
       "b 174f\n"
       "171:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x14, #0, 174f\n"
       "str s14, [x13, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
-      "str s21, [x22, #0x0]\n"
+      "str s9, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
+      "str s21, [x21, #0x0]\n"
       "b 174f\n"
       "172:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x14, #1, 173f\n"
       "str d4, [x13], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "str d19, [x23], #0x8\n"
-      "str d20, [x22], #0x8\n"
+      "str d8, [x23], #0x8\n"
+      "str d19, [x22], #0x8\n"
+      "str d20, [x21], #0x8\n"
       "tbz x14, #0, 174f\n"
       "st1 { v4.s }[2], [x13]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "st1 { v19.s }[2], [x23]\n"
-      "st1 { v20.s }[2], [x22]\n"
+      "st1 { v8.s }[2], [x23]\n"
+      "st1 { v19.s }[2], [x22]\n"
+      "st1 { v20.s }[2], [x21]\n"
       "b 174f\n"
       "173:"  // Height 4: Partial direct writeback: partial_1_0
       "str s4, [x13, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "str s19, [x23, #0x0]\n"
-      "str s20, [x22, #0x0]\n"
+      "str s8, [x23, #0x0]\n"
+      "str s19, [x22, #0x0]\n"
+      "str s20, [x21, #0x0]\n"
       "174:"  // Height 4: Partial direct writeback: Done
       "b 176f\n"
       "175:"  // Height 4: Full writeback
@@ -2521,24 +2518,24 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "str q17, [x13, #0x40]\n"
       "str q18, [x13, #0x50]\n"
       "add x13, x13, #0x60\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q12, [x24, #0x40]\n"
-      "str q13, [x24, #0x50]\n"
-      "str q19, [x23, #0x0]\n"
-      "str q26, [x23, #0x10]\n"
-      "str q27, [x23, #0x20]\n"
-      "str q28, [x23, #0x30]\n"
-      "str q29, [x23, #0x40]\n"
-      "str q30, [x23, #0x50]\n"
-      "str q20, [x22, #0x0]\n"
-      "str q21, [x22, #0x10]\n"
-      "str q22, [x22, #0x20]\n"
-      "str q23, [x22, #0x30]\n"
-      "str q24, [x22, #0x40]\n"
-      "str q25, [x22, #0x50]\n"
+      "str q8, [x23, #0x0]\n"
+      "str q9, [x23, #0x10]\n"
+      "str q10, [x23, #0x20]\n"
+      "str q11, [x23, #0x30]\n"
+      "str q12, [x23, #0x40]\n"
+      "str q13, [x23, #0x50]\n"
+      "str q19, [x22, #0x0]\n"
+      "str q26, [x22, #0x10]\n"
+      "str q27, [x22, #0x20]\n"
+      "str q28, [x22, #0x30]\n"
+      "str q29, [x22, #0x40]\n"
+      "str q30, [x22, #0x50]\n"
+      "str q20, [x21, #0x0]\n"
+      "str q21, [x21, #0x10]\n"
+      "str q22, [x21, #0x20]\n"
+      "str q23, [x21, #0x30]\n"
+      "str q24, [x21, #0x40]\n"
+      "str q25, [x21, #0x50]\n"
       "176:"  // Height 4: Writeback done
       "subs x14, x14, #0x18\n"
       "bgt 134b\n"
@@ -2554,8 +2551,8 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "178:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp
index 376920a17b..98f7fc9403 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp
@@ -82,7 +82,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp
index 9497508289..9ab4aa98f9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16/generic.cpp
@@ -50,8 +50,8 @@ void a64_ffhybrid_fp32bf16fp32_mmla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
+        void *output_ptr = nullptr;
+        const float *bias = nullptr;
     } ka;
 
     unsigned long flags=0;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12.hpp
index eb08de0ade..49973ddb92 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12.hpp
@@ -41,8 +41,7 @@ void a64_ffinterleaved_bf16fp32_dot_8x12( ARGLIST );
 class cls_a64_ffinterleaved_bf16fp32_dot_8x12
 {
 public:
-    typedef bfloat16 lhs_operand_type;
-    typedef bfloat16 rhs_operand_type;
+    typedef bfloat16 operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -73,8 +72,8 @@ public:
     }
 
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 2> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 2, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 2> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 2, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp
index 349ad1c985..6fd5fd4b6c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp
@@ -54,17 +54,17 @@ void a64_ffinterleaved_bf16fp32_dot_8x12(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x25, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x24, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x24, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x23, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x25, #0x8\n"
-      "mov %x[Apanel], x24\n"
       "add x22, x23, x20, LSL #1\n"
       "add x21, x22, x20, LSL #1\n"
       "add x20, x21, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x25, #0x8\n"
+      "mov %x[Apanel], x24\n"
       "bgt 3f\n"
       "cmp x25, #0x4\n"
       "mov x21, x23\n"
@@ -79,12 +79,12 @@ void a64_ffinterleaved_bf16fp32_dot_8x12(
       "movi v9.16b, #0x0\n"
       "ldr q6, [x21, #0x0]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v10.16b, #0x0\n"
       "movi v11.16b, #0x0\n"
       "movi v12.16b, #0x0\n"
       "movi v13.16b, #0x0\n"
       "movi v14.16b, #0x0\n"
-      "cmp x20, #0x2\n"
       "movi v15.16b, #0x0\n"
       "movi v16.16b, #0x0\n"
       "movi v17.16b, #0x0\n"
@@ -170,18 +170,18 @@ void a64_ffinterleaved_bf16fp32_dot_8x12(
       "ldr q6, [x21, #0x0]\n"
       "bge 4b\n"
       "5:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "add x23, x23, #0x10\n"
       ".inst 0x4f40f088  // bfdot v8.4s, v4.8h, v0.h[0]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
       ".inst 0x4f60f08b  // bfdot v11.4s, v4.8h, v0.h[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       ".inst 0x4f40f88e  // bfdot v14.4s, v4.8h, v0.h[2]\n"
       ".inst 0x4f60f891  // bfdot v17.4s, v4.8h, v0.h[3]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4f41f094  // bfdot v20.4s, v4.8h, v1.h[0]\n"
       ".inst 0x4f61f097  // bfdot v23.4s, v4.8h, v1.h[1]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4f41f89a  // bfdot v26.4s, v4.8h, v1.h[2]\n"
       ".inst 0x4f61f89d  // bfdot v29.4s, v4.8h, v1.h[3]\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x4f40f0a9  // bfdot v9.4s, v5.8h, v0.h[0]\n"
       ".inst 0x4f60f0ac  // bfdot v12.4s, v5.8h, v0.h[1]\n"
       ".inst 0x4f40f8af  // bfdot v15.4s, v5.8h, v0.h[2]\n"
@@ -204,8 +204,8 @@ void a64_ffinterleaved_bf16fp32_dot_8x12(
       "add %x[Apanel], %x[Apanel], #0x20\n"
       "ldr q2, [x23, #0x0]\n"
       "ldr q1, [x22, #0x0]\n"
-      "ldr q0, [x21, #0x0]\n"
       ".inst 0x4f44f048  // bfdot v8.4s, v2.8h, v4.h[0]\n"
+      "ldr q0, [x21, #0x0]\n"
       ".inst 0x4f64f04b  // bfdot v11.4s, v2.8h, v4.h[1]\n"
       ".inst 0x4f44f84e  // bfdot v14.4s, v2.8h, v4.h[2]\n"
       ".inst 0x4f64f851  // bfdot v17.4s, v2.8h, v4.h[3]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp
index eb382952fa..1a8b0fd630 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp
@@ -41,8 +41,7 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12( ARGLIST );
 class cls_a64_ffinterleaved_bf16fp32_mmla_8x12
 {
 public:
-    typedef bfloat16 lhs_operand_type;
-    typedef bfloat16 rhs_operand_type;
+    typedef bfloat16 operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -73,8 +72,8 @@ public:
     }
 
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 4> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
@@ -89,8 +88,10 @@ public:
 
         if (std::is_same<T, float>::value) {
             switch (ci->get_cpu_model()) {
+                case CPUModel::V1:
+                    return { 45.25, 4.29, 4.80 };
                 default:
-                    return { 38.10, 5.23, 3.15 };
+                    return { 29.85, 2.60, 5.49 };
             }
         }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp
index 5331f9e652..658491571d 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_mmla_8x12/generic.cpp
@@ -54,17 +54,17 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x25, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x24, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x24, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x23, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x25, #0x8\n"
-      "mov %x[Apanel], x24\n"
       "add x22, x23, x20, LSL #1\n"
       "add x21, x22, x20, LSL #1\n"
       "add x20, x21, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x25, #0x8\n"
+      "mov %x[Apanel], x24\n"
       "bgt 3f\n"
       "cmp x25, #0x4\n"
       "mov x21, x23\n"
@@ -79,14 +79,14 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12(
       "movi v9.16b, #0x0\n"
       "ldr q2, [%x[Apanel], #0x20]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v10.16b, #0x0\n"
       "movi v11.16b, #0x0\n"
-      "movi v12.16b, #0x0\n"
       "add x23, x23, #0x20\n"
+      "movi v12.16b, #0x0\n"
       "movi v13.16b, #0x0\n"
-      "movi v14.16b, #0x0\n"
       "add %x[Apanel], %x[Apanel], #0x30\n"
-      "cmp x20, #0x2\n"
+      "movi v14.16b, #0x0\n"
       "movi v15.16b, #0x0\n"
       "movi v16.16b, #0x0\n"
       "movi v17.16b, #0x0\n"
@@ -217,19 +217,19 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12(
       "cbz x20, 6f\n"
       "ldr q1, [x23, #0x0]\n"
       "ldr q7, [%x[Apanel], #0x0]\n"
+      ".inst 0x6e41ece8  // bfmmla v8.4s, v7.8h, v1.8h\n"
       "ldr q6, [%x[Apanel], #0x10]\n"
       "ldr q0, [x23, #0x10]\n"
+      ".inst 0x6e40eceb  // bfmmla v11.4s, v7.8h, v0.8h\n"
       "ldr q5, [%x[Apanel], #0x20]\n"
       "ldr q4, [%x[Apanel], #0x30]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
+      ".inst 0x6e41ecce  // bfmmla v14.4s, v6.8h, v1.8h\n"
       "ldr q3, [x22, #0x0]\n"
       "ldr q2, [x22, #0x10]\n"
-      ".inst 0x6e41ece8  // bfmmla v8.4s, v7.8h, v1.8h\n"
-      ".inst 0x6e40eceb  // bfmmla v11.4s, v7.8h, v0.8h\n"
-      ".inst 0x6e41ecce  // bfmmla v14.4s, v6.8h, v1.8h\n"
       ".inst 0x6e40ecd1  // bfmmla v17.4s, v6.8h, v0.8h\n"
       ".inst 0x6e41ecb4  // bfmmla v20.4s, v5.8h, v1.8h\n"
       ".inst 0x6e40ecb7  // bfmmla v23.4s, v5.8h, v0.8h\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
       ".inst 0x6e41ec9a  // bfmmla v26.4s, v4.8h, v1.8h\n"
       "ldr q1, [x21, #0x0]\n"
       ".inst 0x6e40ec9d  // bfmmla v29.4s, v4.8h, v0.8h\n"
@@ -252,41 +252,41 @@ void a64_ffinterleaved_bf16fp32_mmla_8x12(
       ".inst 0x6e40ec9f  // bfmmla v31.4s, v4.8h, v0.8h\n"
       "6:"  // multiply loop done
       "subs x25, x25, #0xc\n"
-      "uzp1 v2.2d, v8.2d, v11.2d\n"
+      "uzp1 v0.2d, v8.2d, v11.2d\n"
       "uzp2 v8.2d, v8.2d, v11.2d\n"
       "uzp1 v1.2d, v9.2d, v12.2d\n"
       "uzp2 v9.2d, v9.2d, v12.2d\n"
+      "str q0, [%x[Cpanel], #0x0]\n"
       "uzp1 v0.2d, v10.2d, v13.2d\n"
       "uzp2 v10.2d, v10.2d, v13.2d\n"
-      "str q2, [%x[Cpanel], #0x0]\n"
-      "uzp1 v3.2d, v14.2d, v17.2d\n"
-      "uzp2 v14.2d, v14.2d, v17.2d\n"
       "str q1, [%x[Cpanel], #0x10]\n"
+      "str q0, [%x[Cpanel], #0x20]\n"
+      "uzp1 v0.2d, v14.2d, v17.2d\n"
+      "uzp2 v14.2d, v14.2d, v17.2d\n"
+      "str q8, [%x[Cpanel], #0x30]\n"
       "uzp1 v2.2d, v15.2d, v18.2d\n"
       "uzp2 v15.2d, v15.2d, v18.2d\n"
-      "str q0, [%x[Cpanel], #0x20]\n"
+      "str q9, [%x[Cpanel], #0x40]\n"
       "uzp1 v17.2d, v16.2d, v19.2d\n"
       "uzp2 v16.2d, v16.2d, v19.2d\n"
-      "str q8, [%x[Cpanel], #0x30]\n"
+      "str q10, [%x[Cpanel], #0x50]\n"
       "uzp1 v1.2d, v20.2d, v23.2d\n"
       "uzp2 v20.2d, v20.2d, v23.2d\n"
-      "str q9, [%x[Cpanel], #0x40]\n"
+      "str q0, [%x[Cpanel], #0x60]\n"
       "uzp1 v0.2d, v21.2d, v24.2d\n"
       "uzp2 v21.2d, v21.2d, v24.2d\n"
-      "str q10, [%x[Cpanel], #0x50]\n"
+      "str q2, [%x[Cpanel], #0x70]\n"
       "uzp1 v23.2d, v22.2d, v25.2d\n"
       "uzp2 v22.2d, v22.2d, v25.2d\n"
-      "str q3, [%x[Cpanel], #0x60]\n"
+      "str q17, [%x[Cpanel], #0x80]\n"
       "uzp1 v19.2d, v26.2d, v29.2d\n"
       "uzp2 v26.2d, v26.2d, v29.2d\n"
-      "str q2, [%x[Cpanel], #0x70]\n"
+      "str q14, [%x[Cpanel], #0x90]\n"
       "uzp1 v18.2d, v27.2d, v30.2d\n"
       "uzp2 v27.2d, v27.2d, v30.2d\n"
-      "str q17, [%x[Cpanel], #0x80]\n"
+      "str q15, [%x[Cpanel], #0xa0]\n"
       "uzp1 v17.2d, v28.2d, v31.2d\n"
       "uzp2 v28.2d, v28.2d, v31.2d\n"
-      "str q14, [%x[Cpanel], #0x90]\n"
-      "str q15, [%x[Cpanel], #0xa0]\n"
       "str q16, [%x[Cpanel], #0xb0]\n"
       "str q1, [%x[Cpanel], #0xc0]\n"
       "str q0, [%x[Cpanel], #0xd0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24.hpp
index 42136ce085..29e524a89b 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24.hpp
@@ -40,8 +40,7 @@ void a64_ffinterleaved_fp16_mla_8x24( ARGLIST );
 class cls_a64_ffinterleaved_fp16_mla_8x24
 {
 public:
-    typedef __fp16 lhs_operand_type;
-    typedef __fp16 rhs_operand_type;
+    typedef __fp16 operand_type;
     typedef __fp16 result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -72,8 +71,8 @@ public:
     }
 
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 24, 1> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 24, 1, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 24, 1> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 24, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp
index 2ad85ad424..744d05dbdf 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp
@@ -53,17 +53,17 @@ void a64_ffinterleaved_fp16_mla_8x24(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x25, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x24, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x24, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x23, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x25, #0x10\n"
-      "mov %x[Apanel], x24\n"
       "add x22, x23, x20, LSL #1\n"
       "add x21, x22, x20, LSL #1\n"
       "add x20, x21, x20, LSL #1\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x25, #0x10\n"
+      "mov %x[Apanel], x24\n"
       "bgt 3f\n"
       "cmp x25, #0x8\n"
       "mov x21, x23\n"
@@ -77,13 +77,13 @@ void a64_ffinterleaved_fp16_mla_8x24(
       "ldr q4, [x21, #0x0]\n"
       "movi v9.16b, #0x0\n"
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v10.16b, #0x0\n"
       "movi v11.16b, #0x0\n"
       "movi v12.16b, #0x0\n"
       "movi v13.16b, #0x0\n"
       "movi v14.16b, #0x0\n"
       "movi v15.16b, #0x0\n"
-      "cmp x20, #0x2\n"
       "movi v16.16b, #0x0\n"
       "movi v17.16b, #0x0\n"
       "movi v18.16b, #0x0\n"
@@ -166,18 +166,18 @@ void a64_ffinterleaved_fp16_mla_8x24(
       "fmla v31.8h, v1.8h, v7.h[7]\n"
       "bge 4b\n"
       "5:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      "add x23, x23, #0x10\n"
       "fmla v8.8h, v2.8h, v0.h[0]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
       "fmla v11.8h, v2.8h, v0.h[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       "fmla v14.8h, v2.8h, v0.h[2]\n"
       "fmla v17.8h, v2.8h, v0.h[3]\n"
+      "add x23, x23, #0x10\n"
       "fmla v20.8h, v2.8h, v0.h[4]\n"
       "fmla v23.8h, v2.8h, v0.h[5]\n"
+      "add x22, x22, #0x10\n"
       "fmla v26.8h, v2.8h, v0.h[6]\n"
       "fmla v29.8h, v2.8h, v0.h[7]\n"
+      "add x21, x21, #0x10\n"
       "fmla v9.8h, v3.8h, v0.h[0]\n"
       "fmla v12.8h, v3.8h, v0.h[1]\n"
       "fmla v15.8h, v3.8h, v0.h[2]\n"
@@ -197,13 +197,13 @@ void a64_ffinterleaved_fp16_mla_8x24(
       "cbz x20, 6f\n"
       "ldr q3, [%x[Apanel], #0x0]\n"
       "ldr q2, [x23, #0x0]\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
+      "fmla v8.8h, v2.8h, v3.h[0]\n"
       "ldr q1, [x22, #0x0]\n"
       "ldr q0, [x21, #0x0]\n"
-      "fmla v8.8h, v2.8h, v3.h[0]\n"
       "fmla v11.8h, v2.8h, v3.h[1]\n"
       "fmla v14.8h, v2.8h, v3.h[2]\n"
       "fmla v17.8h, v2.8h, v3.h[3]\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       "fmla v20.8h, v2.8h, v3.h[4]\n"
       "fmla v23.8h, v2.8h, v3.h[5]\n"
       "fmla v26.8h, v2.8h, v3.h[6]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12.hpp
index bb6deaf68f..6b01ffe63b 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12.hpp
@@ -40,8 +40,7 @@ void a64_ffinterleaved_fp32_mla_8x12( ARGLIST );
 class cls_a64_ffinterleaved_fp32_mla_8x12
 {
 public:
-    typedef float lhs_operand_type;
-    typedef float rhs_operand_type;
+    typedef float operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -72,8 +71,8 @@ public:
     }
 
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 1> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 1, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 1> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp
index 45970fdc0b..f93bc6c719 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp
@@ -53,17 +53,17 @@ void a64_ffinterleaved_fp32_mla_8x12(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x25, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x24, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x24, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x23, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cmp x25, #0x8\n"
-      "mov %x[Apanel], x24\n"
       "add x22, x23, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "cmp x25, #0x8\n"
+      "mov %x[Apanel], x24\n"
       "bgt 3f\n"
       "cmp x25, #0x4\n"
       "mov x21, x23\n"
@@ -78,12 +78,12 @@ void a64_ffinterleaved_fp32_mla_8x12(
       "movi v9.16b, #0x0\n"
       "ldr q6, [x21, #0x0]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x4\n"
       "movi v10.16b, #0x0\n"
       "movi v11.16b, #0x0\n"
       "movi v12.16b, #0x0\n"
       "movi v13.16b, #0x0\n"
       "movi v14.16b, #0x0\n"
-      "cmp x20, #0x4\n"
       "movi v15.16b, #0x0\n"
       "movi v16.16b, #0x0\n"
       "movi v17.16b, #0x0\n"
@@ -227,18 +227,18 @@ void a64_ffinterleaved_fp32_mla_8x12(
       "fmla v31.4s, v2.4s, v7.s[3]\n"
       "bge 4b\n"
       "5:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "add x23, x23, #0x10\n"
       "fmla v8.4s, v4.4s, v0.s[0]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
       "fmla v11.4s, v4.4s, v0.s[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "fmla v14.4s, v4.4s, v0.s[2]\n"
       "fmla v17.4s, v4.4s, v0.s[3]\n"
+      "add x23, x23, #0x10\n"
       "fmla v20.4s, v4.4s, v1.s[0]\n"
       "fmla v23.4s, v4.4s, v1.s[1]\n"
+      "add x22, x22, #0x10\n"
       "fmla v26.4s, v4.4s, v1.s[2]\n"
       "fmla v29.4s, v4.4s, v1.s[3]\n"
+      "add x21, x21, #0x10\n"
       "fmla v9.4s, v5.4s, v0.s[0]\n"
       "fmla v12.4s, v5.4s, v0.s[1]\n"
       "fmla v15.4s, v5.4s, v0.s[2]\n"
@@ -262,22 +262,22 @@ void a64_ffinterleaved_fp32_mla_8x12(
       "subs x20, x20, #0x1\n"
       "ldr q2, [x23, #0x0]\n"
       "ldr q1, [x22, #0x0]\n"
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "ldr q0, [x21, #0x0]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
       "fmla v8.4s, v2.4s, v4.s[0]\n"
+      "ldr q0, [x21, #0x0]\n"
       "fmla v11.4s, v2.4s, v4.s[1]\n"
       "fmla v14.4s, v2.4s, v4.s[2]\n"
       "fmla v17.4s, v2.4s, v4.s[3]\n"
       "fmla v20.4s, v2.4s, v3.s[0]\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "fmla v23.4s, v2.4s, v3.s[1]\n"
       "fmla v26.4s, v2.4s, v3.s[2]\n"
+      "add x23, x23, #0x10\n"
       "fmla v29.4s, v2.4s, v3.s[3]\n"
       "fmla v9.4s, v1.4s, v4.s[0]\n"
+      "add x22, x22, #0x10\n"
       "fmla v12.4s, v1.4s, v4.s[1]\n"
       "fmla v15.4s, v1.4s, v4.s[2]\n"
+      "add x21, x21, #0x10\n"
       "fmla v18.4s, v1.4s, v4.s[3]\n"
       "fmla v21.4s, v1.4s, v3.s[0]\n"
       "fmla v24.4s, v1.4s, v3.s[1]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12.hpp
index 8dc9112ebe..059c1f6d4d 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12.hpp
@@ -42,8 +42,7 @@ void a64_gemm_s16_asimd_8x12(const int16_t *, const int16_t *, int32_t *, int, i
 // structure.
 class cls_a64_gemm_s16_8x12 {
 public:
-    typedef int16_t lhs_operand_type;
-    typedef int16_t rhs_operand_type;
+    typedef int16_t operand_type;
     typedef int32_t result_type;
 
     typedef void (*kern_type)(const int16_t *, const int16_t *, int32_t *, int, int, int);
@@ -62,8 +61,8 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 1, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 1, true> transforms_quantized = {};
 
     kern_type kernel = a64_gemm_s16_asimd_8x12;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp
index d0edcfcb5e..226c13b400 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp
@@ -37,8 +37,7 @@ void a64_gemm_s8_4x4(const int8_t *, const int8_t *, int32_t *, int, int, int);
 
 class cls_a64_gemm_s8_4x4 {
 public:
-    typedef int8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
+    typedef int8_t operand_type;
     typedef int32_t result_type;
 
     typedef void (*kern_type)(const int8_t *, const int8_t *, int32_t *, int, int, int);
@@ -57,8 +56,8 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 4, 16> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 4, 16, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 4, 4, 16> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 4, 4, 16, true> transforms_quantized = {};
 
     template<typename T>
     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp
index b55ac61403..1396c6ae8c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp
@@ -55,37 +55,37 @@ void a64_gemm_s8_4x4(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpanel
             register int8x16_t b3a asm("v11");
 
             __asm __volatile (
-                "movi    v16.4s, #0x0\n"
-                "ldr    q0, [%[a_ptr]]\n"
-                "movi    v17.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v18.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v19.4s, #0x0\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "movi    v20.4s, #0x0\n"
-                "ldr    %q[b3], [%[b_ptr], #48]\n"
-                "movi    v21.4s, #0x0\n"
-                "ldr    q1, [%[a_ptr], #16]\n"
-                "movi    v22.4s, #0x0\n"
-                "ldr    q2, [%[a_ptr], #32]\n"
-                "movi    v23.4s, #0x0\n"
-                "ldr    q3, [%[a_ptr], #48]\n"
-                "movi    v24.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
+                "ldr	q0, [%[a_ptr]]\n"
+                "movi	v17.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v18.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v19.4s, #0x0\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "movi	v20.4s, #0x0\n"
+                "ldr	%q[b3], [%[b_ptr], #48]\n"
+                "movi	v21.4s, #0x0\n"
+                "ldr	q1, [%[a_ptr], #16]\n"
+                "movi	v22.4s, #0x0\n"
+                "ldr	q2, [%[a_ptr], #32]\n"
+                "movi	v23.4s, #0x0\n"
+                "ldr	q3, [%[a_ptr], #48]\n"
+                "movi	v24.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v25.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v26.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v27.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v28.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v29.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v30.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v31.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
 
                 // Loop structure optimized for A57 (after r0).
@@ -107,351 +107,351 @@ void a64_gemm_s8_4x4(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpanel
                 // of multiplies that need to be pulled out.
 
                 // Start of unroll 0 (first iteration)
-                "smull    v12.8h, v0.8b, %[b0].8b\n"
-                "smull    v13.8h, v0.8b, %[b1].8b\n"
+                "smull	v12.8h, v0.8b, %[b0].8b\n"
+                "smull	v13.8h, v0.8b, %[b1].8b\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 // Unroll 0 continuation (branch target)
                 "1:\n"
-                "smull    v14.8h, v0.8b, %[b2].8b\n"
-                "subs    %w[k], %w[k], #1\n"
-                "smull    v15.8h, v0.8b, %[b3].8b\n"
-                "ldr    %q[b0a], [%[b_ptr], #64]\n"
-                "smlal2    v12.8h, v0.16b, %[b0].16b\n"
-                "smlal2    v13.8h, v0.16b, %[b1].16b\n"
-                "ldr    %q[b1a], [%[b_ptr], #80]\n"
-                "smlal2    v14.8h, v0.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v0.16b, %[b3].16b\n"
-                "ldr     q0, [%[a_ptr], #64]\n"
-
-                "sadalp    v16.4s, v12.8h\n"
-                "smull    v12.8h, v1.8b, %[b0].8b\n"
-                "sadalp    v17.4s, v13.8h\n"
-                "sadalp    v18.4s, v14.8h\n"
-                "smull    v13.8h, v1.8b, %[b1].8b\n"
-                "sadalp    v19.4s, v15.8h\n"
-                "smull    v14.8h, v1.8b, %[b2].8b\n"
-                "ldr    %q[b2a], [%[b_ptr], #96]\n"
-                "smull    v15.8h, v1.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v1.16b, %[b0].16b\n"
-                "ldr    %q[b3a], [%[b_ptr], #112]\n"
-                "smlal2    v13.8h, v1.16b, %[b1].16b\n"
-                "add    %[b_ptr], %[b_ptr], #128\n"
-                "smlal2    v14.8h, v1.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v1.16b, %[b3].16b\n"
-                "ldr     q1, [%[a_ptr], #80]\n"
-
-                "sadalp    v20.4s, v12.8h\n"
-                "smull    v12.8h, v2.8b, %[b0].8b\n"
-                "sadalp    v21.4s, v13.8h\n"
-                "sadalp    v22.4s, v14.8h\n"
-                "smull    v13.8h, v2.8b, %[b1].8b\n"
-                "sadalp    v23.4s, v15.8h\n"
-                "smull    v14.8h, v2.8b, %[b2].8b\n"
-                "smull    v15.8h, v2.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v2.16b, %[b0].16b\n"
+                "smull	v14.8h, v0.8b, %[b2].8b\n"
+                "subs	%w[k], %w[k], #1\n"
+                "smull	v15.8h, v0.8b, %[b3].8b\n"
+                "ldr	%q[b0a], [%[b_ptr], #64]\n"
+                "smlal2	v12.8h, v0.16b, %[b0].16b\n"
+                "smlal2	v13.8h, v0.16b, %[b1].16b\n"
+                "ldr	%q[b1a], [%[b_ptr], #80]\n"
+                "smlal2	v14.8h, v0.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v0.16b, %[b3].16b\n"
+                "ldr 	q0, [%[a_ptr], #64]\n"
+
+                "sadalp	v16.4s, v12.8h\n"
+                "smull	v12.8h, v1.8b, %[b0].8b\n"
+                "sadalp	v17.4s, v13.8h\n"
+                "sadalp	v18.4s, v14.8h\n"
+                "smull	v13.8h, v1.8b, %[b1].8b\n"
+                "sadalp	v19.4s, v15.8h\n"
+                "smull	v14.8h, v1.8b, %[b2].8b\n"
+                "ldr	%q[b2a], [%[b_ptr], #96]\n"
+                "smull	v15.8h, v1.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v1.16b, %[b0].16b\n"
+                "ldr	%q[b3a], [%[b_ptr], #112]\n"
+                "smlal2	v13.8h, v1.16b, %[b1].16b\n"
+                "add	%[b_ptr], %[b_ptr], #128\n"
+                "smlal2	v14.8h, v1.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v1.16b, %[b3].16b\n"
+                "ldr 	q1, [%[a_ptr], #80]\n"
+
+                "sadalp	v20.4s, v12.8h\n"
+                "smull	v12.8h, v2.8b, %[b0].8b\n"
+                "sadalp	v21.4s, v13.8h\n"
+                "sadalp	v22.4s, v14.8h\n"
+                "smull	v13.8h, v2.8b, %[b1].8b\n"
+                "sadalp	v23.4s, v15.8h\n"
+                "smull	v14.8h, v2.8b, %[b2].8b\n"
+                "smull	v15.8h, v2.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v2.16b, %[b0].16b\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "smlal2    v13.8h, v2.16b, %[b1].16b\n"
-                "smlal2    v14.8h, v2.16b, %[b2].16b\n"
+                "smlal2	v13.8h, v2.16b, %[b1].16b\n"
+                "smlal2	v14.8h, v2.16b, %[b2].16b\n"
                 ASM_PREFETCH("[%[a_ptr], #320]")
-                "smlal2    v15.8h, v2.16b, %[b3].16b\n"
-                "ldr     q2, [%[a_ptr], #96]\n"
-
-                "sadalp    v24.4s, v12.8h\n"
-                "smull    v12.8h, v3.8b, %[b0].8b\n"
-                "sadalp    v25.4s, v13.8h\n"
-                "sadalp    v26.4s, v14.8h\n"
-                "smull    v13.8h, v3.8b, %[b1].8b\n"
-                "sadalp    v27.4s, v15.8h\n"
-                "smull    v14.8h, v3.8b, %[b2].8b\n"
-                "smull    v15.8h, v3.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v3.16b, %[b0].16b\n"
-                "ldr     %q[b0], [%[b_ptr], #0]\n"
-                "smlal2    v13.8h, v3.16b, %[b1].16b\n"
-                "smlal2    v14.8h, v3.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v3.16b, %[b3].16b\n"
-                "ldr     q3, [%[a_ptr], #112]\n"
+                "smlal2	v15.8h, v2.16b, %[b3].16b\n"
+                "ldr 	q2, [%[a_ptr], #96]\n"
+
+                "sadalp	v24.4s, v12.8h\n"
+                "smull	v12.8h, v3.8b, %[b0].8b\n"
+                "sadalp	v25.4s, v13.8h\n"
+                "sadalp	v26.4s, v14.8h\n"
+                "smull	v13.8h, v3.8b, %[b1].8b\n"
+                "sadalp	v27.4s, v15.8h\n"
+                "smull	v14.8h, v3.8b, %[b2].8b\n"
+                "smull	v15.8h, v3.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v3.16b, %[b0].16b\n"
+                "ldr 	%q[b0], [%[b_ptr], #0]\n"
+                "smlal2	v13.8h, v3.16b, %[b1].16b\n"
+                "smlal2	v14.8h, v3.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v3.16b, %[b3].16b\n"
+                "ldr 	q3, [%[a_ptr], #112]\n"
 
                 // Unroll 1
-                "sadalp    v28.4s, v12.8h\n"
-                "smull    v12.8h, v0.8b, %[b0a].8b\n"
-                "sadalp    v29.4s, v13.8h\n"
-                "sadalp    v30.4s, v14.8h\n"
-                "smull    v13.8h, v0.8b, %[b1a].8b\n"
-                "sadalp    v31.4s, v15.8h\n"
-                "smull    v14.8h, v0.8b, %[b2a].8b\n"
-                "smull    v15.8h, v0.8b, %[b3a].8b\n"
-                "ldr     %q[b1], [%[b_ptr], #16]\n"
-                "smlal2    v12.8h, v0.16b, %[b0a].16b\n"
-                "smlal2    v13.8h, v0.16b, %[b1a].16b\n"
-                "ldr     %q[b2], [%[b_ptr], #32]\n"
-                "smlal2    v14.8h, v0.16b, %[b2a].16b\n"
-                "smlal2    v15.8h, v0.16b, %[b3a].16b\n"
-                "ldr     q0, [%[a_ptr], #128]\n"
-
-                "sadalp    v16.4s, v12.8h\n"
-                "smull    v12.8h, v1.8b, %[b0a].8b\n"
-                "sadalp    v17.4s, v13.8h\n"
-                "sadalp    v18.4s, v14.8h\n"
-                "smull    v13.8h, v1.8b, %[b1a].8b\n"
-                "sadalp    v19.4s, v15.8h\n"
-                "add    %[a_ptr], %[a_ptr], #128\n"
-                "smull    v14.8h, v1.8b, %[b2a].8b\n"
-                "smull    v15.8h, v1.8b, %[b3a].8b\n"
-                "ldr     %q[b3], [%[b_ptr], #48]\n"
-                "smlal2    v12.8h, v1.16b, %[b0a].16b\n"
-                "smlal2    v13.8h, v1.16b, %[b1a].16b\n"
-                "smlal2    v14.8h, v1.16b, %[b2a].16b\n"
-                "smlal2    v15.8h, v1.16b, %[b3a].16b\n"
-                "ldr     q1, [%[a_ptr], #16]\n"
-
-                "sadalp    v20.4s, v12.8h\n"
-                "smull    v12.8h, v2.8b, %[b0a].8b\n"
-                "sadalp    v21.4s, v13.8h\n"
-                "sadalp    v22.4s, v14.8h\n"
-                "smull    v13.8h, v2.8b, %[b1a].8b\n"
-                "sadalp    v23.4s, v15.8h\n"
-                "smull    v14.8h, v2.8b, %[b2a].8b\n"
-                "smull    v15.8h, v2.8b, %[b3a].8b\n"
-                "smlal2    v12.8h, v2.16b, %[b0a].16b\n"
+                "sadalp	v28.4s, v12.8h\n"
+                "smull	v12.8h, v0.8b, %[b0a].8b\n"
+                "sadalp	v29.4s, v13.8h\n"
+                "sadalp	v30.4s, v14.8h\n"
+                "smull	v13.8h, v0.8b, %[b1a].8b\n"
+                "sadalp	v31.4s, v15.8h\n"
+                "smull	v14.8h, v0.8b, %[b2a].8b\n"
+                "smull	v15.8h, v0.8b, %[b3a].8b\n"
+                "ldr 	%q[b1], [%[b_ptr], #16]\n"
+                "smlal2	v12.8h, v0.16b, %[b0a].16b\n"
+                "smlal2	v13.8h, v0.16b, %[b1a].16b\n"
+                "ldr 	%q[b2], [%[b_ptr], #32]\n"
+                "smlal2	v14.8h, v0.16b, %[b2a].16b\n"
+                "smlal2	v15.8h, v0.16b, %[b3a].16b\n"
+                "ldr 	q0, [%[a_ptr], #128]\n"
+
+                "sadalp	v16.4s, v12.8h\n"
+                "smull	v12.8h, v1.8b, %[b0a].8b\n"
+                "sadalp	v17.4s, v13.8h\n"
+                "sadalp	v18.4s, v14.8h\n"
+                "smull	v13.8h, v1.8b, %[b1a].8b\n"
+                "sadalp	v19.4s, v15.8h\n"
+                "add	%[a_ptr], %[a_ptr], #128\n"
+                "smull	v14.8h, v1.8b, %[b2a].8b\n"
+                "smull	v15.8h, v1.8b, %[b3a].8b\n"
+                "ldr 	%q[b3], [%[b_ptr], #48]\n"
+                "smlal2	v12.8h, v1.16b, %[b0a].16b\n"
+                "smlal2	v13.8h, v1.16b, %[b1a].16b\n"
+                "smlal2	v14.8h, v1.16b, %[b2a].16b\n"
+                "smlal2	v15.8h, v1.16b, %[b3a].16b\n"
+                "ldr 	q1, [%[a_ptr], #16]\n"
+
+                "sadalp	v20.4s, v12.8h\n"
+                "smull	v12.8h, v2.8b, %[b0a].8b\n"
+                "sadalp	v21.4s, v13.8h\n"
+                "sadalp	v22.4s, v14.8h\n"
+                "smull	v13.8h, v2.8b, %[b1a].8b\n"
+                "sadalp	v23.4s, v15.8h\n"
+                "smull	v14.8h, v2.8b, %[b2a].8b\n"
+                "smull	v15.8h, v2.8b, %[b3a].8b\n"
+                "smlal2	v12.8h, v2.16b, %[b0a].16b\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "smlal2    v13.8h, v2.16b, %[b1a].16b\n"
-                "smlal2    v14.8h, v2.16b, %[b2a].16b\n"
+                "smlal2	v13.8h, v2.16b, %[b1a].16b\n"
+                "smlal2	v14.8h, v2.16b, %[b2a].16b\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "smlal2    v15.8h, v2.16b, %[b3a].16b\n"
-                "ldr     q2, [%[a_ptr], #32]\n"
-
-                "sadalp    v24.4s, v12.8h\n"
-                "smull    v12.8h, v3.8b, %[b0a].8b\n"
-                "sadalp    v25.4s, v13.8h\n"
-                "sadalp    v26.4s, v14.8h\n"
-                "smull    v13.8h, v3.8b, %[b1a].8b\n"
-                "sadalp    v27.4s, v15.8h\n"
-                "smull    v14.8h, v3.8b, %[b2a].8b\n"
-                "smull    v15.8h, v3.8b, %[b3a].8b\n"
-                "smlal2    v12.8h, v3.16b, %[b0a].16b\n"
-                "smlal2    v13.8h, v3.16b, %[b1a].16b\n"
-                "smlal2    v14.8h, v3.16b, %[b2a].16b\n"
-                "smlal2    v15.8h, v3.16b, %[b3a].16b\n"
-                "ldr     q3, [%[a_ptr], #48]\n"
+                "smlal2	v15.8h, v2.16b, %[b3a].16b\n"
+                "ldr 	q2, [%[a_ptr], #32]\n"
+
+                "sadalp	v24.4s, v12.8h\n"
+                "smull	v12.8h, v3.8b, %[b0a].8b\n"
+                "sadalp	v25.4s, v13.8h\n"
+                "sadalp	v26.4s, v14.8h\n"
+                "smull	v13.8h, v3.8b, %[b1a].8b\n"
+                "sadalp	v27.4s, v15.8h\n"
+                "smull	v14.8h, v3.8b, %[b2a].8b\n"
+                "smull	v15.8h, v3.8b, %[b3a].8b\n"
+                "smlal2	v12.8h, v3.16b, %[b0a].16b\n"
+                "smlal2	v13.8h, v3.16b, %[b1a].16b\n"
+                "smlal2	v14.8h, v3.16b, %[b2a].16b\n"
+                "smlal2	v15.8h, v3.16b, %[b3a].16b\n"
+                "ldr 	q3, [%[a_ptr], #48]\n"
 
                 // Start of unroll 0 for next iteration.
-                "sadalp    v28.4s, v12.8h\n"
-                "smull    v12.8h, v0.8b, %[b0].8b\n"
-                "sadalp    v29.4s, v13.8h\n"
-                "sadalp    v30.4s, v14.8h\n"
-                "smull    v13.8h, v0.8b, %[b1].8b\n"
-                "sadalp    v31.4s, v15.8h\n"
-                "bne    1b\n"
+                "sadalp	v28.4s, v12.8h\n"
+                "smull	v12.8h, v0.8b, %[b0].8b\n"
+                "sadalp	v29.4s, v13.8h\n"
+                "sadalp	v30.4s, v14.8h\n"
+                "smull	v13.8h, v0.8b, %[b1].8b\n"
+                "sadalp	v31.4s, v15.8h\n"
+                "bne	1b\n"
 
                 // Target to use when K=1 or 2 (i.e. zero iterations of main loop)
                 "4:\n"
 
                 // Branch to alternative tail for odd K
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration (even K)
-                "smull    v14.8h, v0.8b, %[b2].8b\n"
-                "smull    v15.8h, v0.8b, %[b3].8b\n"
-                "ldr    %q[b0a], [%[b_ptr], #64]\n"
-                "smlal2    v12.8h, v0.16b, %[b0].16b\n"
-                "smlal2    v13.8h, v0.16b, %[b1].16b\n"
-                "ldr    %q[b1a], [%[b_ptr], #80]\n"
-                "smlal2    v14.8h, v0.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v0.16b, %[b3].16b\n"
-                "ldr     q0, [%[a_ptr], #64]\n"
-
-                "sadalp    v16.4s, v12.8h\n"
-                "smull    v12.8h, v1.8b, %[b0].8b\n"
-                "sadalp    v17.4s, v13.8h\n"
-                "sadalp    v18.4s, v14.8h\n"
-                "smull    v13.8h, v1.8b, %[b1].8b\n"
-                "sadalp    v19.4s, v15.8h\n"
-                "smull    v14.8h, v1.8b, %[b2].8b\n"
-                "ldr    %q[b2a], [%[b_ptr], #96]\n"
-                "smull    v15.8h, v1.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v1.16b, %[b0].16b\n"
-                "ldr    %q[b3a], [%[b_ptr], #112]\n"
-                "smlal2    v13.8h, v1.16b, %[b1].16b\n"
-                "add    %[b_ptr], %[b_ptr], #128\n"
-                "smlal2    v14.8h, v1.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v1.16b, %[b3].16b\n"
-                "ldr     q1, [%[a_ptr], #80]\n"
-
-                "sadalp    v20.4s, v12.8h\n"
-                "smull    v12.8h, v2.8b, %[b0].8b\n"
-                "sadalp    v21.4s, v13.8h\n"
-                "sadalp    v22.4s, v14.8h\n"
-                "smull    v13.8h, v2.8b, %[b1].8b\n"
-                "sadalp    v23.4s, v15.8h\n"
-                "smull    v14.8h, v2.8b, %[b2].8b\n"
-                "smull    v15.8h, v2.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v2.16b, %[b0].16b\n"
-                "smlal2    v13.8h, v2.16b, %[b1].16b\n"
-                "smlal2    v14.8h, v2.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v2.16b, %[b3].16b\n"
-                "ldr     q2, [%[a_ptr], #96]\n"
-
-                "sadalp    v24.4s, v12.8h\n"
-                "smull    v12.8h, v3.8b, %[b0].8b\n"
-                "sadalp    v25.4s, v13.8h\n"
-                "sadalp    v26.4s, v14.8h\n"
-                "smull    v13.8h, v3.8b, %[b1].8b\n"
-                "sadalp    v27.4s, v15.8h\n"
-                "smull    v14.8h, v3.8b, %[b2].8b\n"
-                "smull    v15.8h, v3.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v3.16b, %[b0].16b\n"
-                "smlal2    v13.8h, v3.16b, %[b1].16b\n"
-                "smlal2    v14.8h, v3.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v3.16b, %[b3].16b\n"
-                "ldr     q3, [%[a_ptr], #112]\n"
+                "smull	v14.8h, v0.8b, %[b2].8b\n"
+                "smull	v15.8h, v0.8b, %[b3].8b\n"
+                "ldr	%q[b0a], [%[b_ptr], #64]\n"
+                "smlal2	v12.8h, v0.16b, %[b0].16b\n"
+                "smlal2	v13.8h, v0.16b, %[b1].16b\n"
+                "ldr	%q[b1a], [%[b_ptr], #80]\n"
+                "smlal2	v14.8h, v0.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v0.16b, %[b3].16b\n"
+                "ldr 	q0, [%[a_ptr], #64]\n"
+
+                "sadalp	v16.4s, v12.8h\n"
+                "smull	v12.8h, v1.8b, %[b0].8b\n"
+                "sadalp	v17.4s, v13.8h\n"
+                "sadalp	v18.4s, v14.8h\n"
+                "smull	v13.8h, v1.8b, %[b1].8b\n"
+                "sadalp	v19.4s, v15.8h\n"
+                "smull	v14.8h, v1.8b, %[b2].8b\n"
+                "ldr	%q[b2a], [%[b_ptr], #96]\n"
+                "smull	v15.8h, v1.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v1.16b, %[b0].16b\n"
+                "ldr	%q[b3a], [%[b_ptr], #112]\n"
+                "smlal2	v13.8h, v1.16b, %[b1].16b\n"
+                "add	%[b_ptr], %[b_ptr], #128\n"
+                "smlal2	v14.8h, v1.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v1.16b, %[b3].16b\n"
+                "ldr 	q1, [%[a_ptr], #80]\n"
+
+                "sadalp	v20.4s, v12.8h\n"
+                "smull	v12.8h, v2.8b, %[b0].8b\n"
+                "sadalp	v21.4s, v13.8h\n"
+                "sadalp	v22.4s, v14.8h\n"
+                "smull	v13.8h, v2.8b, %[b1].8b\n"
+                "sadalp	v23.4s, v15.8h\n"
+                "smull	v14.8h, v2.8b, %[b2].8b\n"
+                "smull	v15.8h, v2.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v2.16b, %[b0].16b\n"
+                "smlal2	v13.8h, v2.16b, %[b1].16b\n"
+                "smlal2	v14.8h, v2.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v2.16b, %[b3].16b\n"
+                "ldr 	q2, [%[a_ptr], #96]\n"
+
+                "sadalp	v24.4s, v12.8h\n"
+                "smull	v12.8h, v3.8b, %[b0].8b\n"
+                "sadalp	v25.4s, v13.8h\n"
+                "sadalp	v26.4s, v14.8h\n"
+                "smull	v13.8h, v3.8b, %[b1].8b\n"
+                "sadalp	v27.4s, v15.8h\n"
+                "smull	v14.8h, v3.8b, %[b2].8b\n"
+                "smull	v15.8h, v3.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v3.16b, %[b0].16b\n"
+                "smlal2	v13.8h, v3.16b, %[b1].16b\n"
+                "smlal2	v14.8h, v3.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v3.16b, %[b3].16b\n"
+                "ldr 	q3, [%[a_ptr], #112]\n"
 
                 // Unroll 1
-                "sadalp    v28.4s, v12.8h\n"
-                "smull    v12.8h, v0.8b, %[b0a].8b\n"
-                "sadalp    v29.4s, v13.8h\n"
-                "sadalp    v30.4s, v14.8h\n"
-                "smull    v13.8h, v0.8b, %[b1a].8b\n"
-                "sadalp    v31.4s, v15.8h\n"
-                "smull    v14.8h, v0.8b, %[b2a].8b\n"
-                "add    %[a_ptr], %[a_ptr], #128\n"
-                "smull    v15.8h, v0.8b, %[b3a].8b\n"
-                "smlal2    v12.8h, v0.16b, %[b0a].16b\n"
-                "smlal2    v13.8h, v0.16b, %[b1a].16b\n"
-                "smlal2    v14.8h, v0.16b, %[b2a].16b\n"
-                "smlal2    v15.8h, v0.16b, %[b3a].16b\n"
-
-                "sadalp    v16.4s, v12.8h\n"
-                "smull    v12.8h, v1.8b, %[b0a].8b\n"
-                "sadalp    v17.4s, v13.8h\n"
-                "sadalp    v18.4s, v14.8h\n"
-                "smull    v13.8h, v1.8b, %[b1a].8b\n"
-                "sadalp    v19.4s, v15.8h\n"
-                "smull    v14.8h, v1.8b, %[b2a].8b\n"
-                "smull    v15.8h, v1.8b, %[b3a].8b\n"
-                "smlal2    v12.8h, v1.16b, %[b0a].16b\n"
-                "addp    v16.4s, v16.4s, v17.4s\n"
-                "smlal2    v13.8h, v1.16b, %[b1a].16b\n"
-                "addp    v17.4s, v18.4s, v19.4s\n"
-                "smlal2    v14.8h, v1.16b, %[b2a].16b\n"
-                "smlal2    v15.8h, v1.16b, %[b3a].16b\n"
-
-                "sadalp    v20.4s, v12.8h\n"
-                "smull    v12.8h, v2.8b, %[b0a].8b\n"
-                "sadalp    v21.4s, v13.8h\n"
-                "sadalp    v22.4s, v14.8h\n"
-                "smull    v13.8h, v2.8b, %[b1a].8b\n"
-                "sadalp    v23.4s, v15.8h\n"
-                "addp    v16.4s, v16.4s, v17.4s\n"
-                "smull    v14.8h, v2.8b, %[b2a].8b\n"
-                "addp    v18.4s, v20.4s, v21.4s\n"
-                "addp    v19.4s, v22.4s, v23.4s\n"
-                "smull    v15.8h, v2.8b, %[b3a].8b\n"
-                "smlal2    v12.8h, v2.16b, %[b0a].16b\n"
-                "str    q16, [%[c_ptr]]\n"
-                "smlal2    v13.8h, v2.16b, %[b1a].16b\n"
-                "smlal2    v14.8h, v2.16b, %[b2a].16b\n"
-                "smlal2    v15.8h, v2.16b, %[b3a].16b\n"
-
-                "sadalp    v24.4s, v12.8h\n"
-                "smull    v12.8h, v3.8b, %[b0a].8b\n"
-                "sadalp    v25.4s, v13.8h\n"
-                "sadalp    v26.4s, v14.8h\n"
-                "smull    v13.8h, v3.8b, %[b1a].8b\n"
-                "sadalp    v27.4s, v15.8h\n"
-                "addp    v17.4s, v18.4s, v19.4s\n"
-                "smull    v14.8h, v3.8b, %[b2a].8b\n"
-                "addp    v20.4s, v24.4s, v25.4s\n"
-                "addp    v21.4s, v26.4s, v27.4s\n"
-                "smull    v15.8h, v3.8b, %[b3a].8b\n"
-                "smlal2    v12.8h, v3.16b, %[b0a].16b\n"
-                "str    q17, [%[c_ptr], #16]\n"
-                "smlal2    v13.8h, v3.16b, %[b1a].16b\n"
-                "smlal2    v14.8h, v3.16b, %[b2a].16b\n"
-                "addp    v18.4s, v20.4s, v21.4s\n"
-                "smlal2    v15.8h, v3.16b, %[b3a].16b\n"
-                "b    3f\n"
+                "sadalp	v28.4s, v12.8h\n"
+                "smull	v12.8h, v0.8b, %[b0a].8b\n"
+                "sadalp	v29.4s, v13.8h\n"
+                "sadalp	v30.4s, v14.8h\n"
+                "smull	v13.8h, v0.8b, %[b1a].8b\n"
+                "sadalp	v31.4s, v15.8h\n"
+                "smull	v14.8h, v0.8b, %[b2a].8b\n"
+                "add	%[a_ptr], %[a_ptr], #128\n"
+                "smull	v15.8h, v0.8b, %[b3a].8b\n"
+                "smlal2	v12.8h, v0.16b, %[b0a].16b\n"
+                "smlal2	v13.8h, v0.16b, %[b1a].16b\n"
+                "smlal2	v14.8h, v0.16b, %[b2a].16b\n"
+                "smlal2	v15.8h, v0.16b, %[b3a].16b\n"
+
+                "sadalp	v16.4s, v12.8h\n"
+                "smull	v12.8h, v1.8b, %[b0a].8b\n"
+                "sadalp	v17.4s, v13.8h\n"
+                "sadalp	v18.4s, v14.8h\n"
+                "smull	v13.8h, v1.8b, %[b1a].8b\n"
+                "sadalp	v19.4s, v15.8h\n"
+                "smull	v14.8h, v1.8b, %[b2a].8b\n"
+                "smull	v15.8h, v1.8b, %[b3a].8b\n"
+                "smlal2	v12.8h, v1.16b, %[b0a].16b\n"
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "smlal2	v13.8h, v1.16b, %[b1a].16b\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "smlal2	v14.8h, v1.16b, %[b2a].16b\n"
+                "smlal2	v15.8h, v1.16b, %[b3a].16b\n"
+
+                "sadalp	v20.4s, v12.8h\n"
+                "smull	v12.8h, v2.8b, %[b0a].8b\n"
+                "sadalp	v21.4s, v13.8h\n"
+                "sadalp	v22.4s, v14.8h\n"
+                "smull	v13.8h, v2.8b, %[b1a].8b\n"
+                "sadalp	v23.4s, v15.8h\n"
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "smull	v14.8h, v2.8b, %[b2a].8b\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "addp	v19.4s, v22.4s, v23.4s\n"
+                "smull	v15.8h, v2.8b, %[b3a].8b\n"
+                "smlal2	v12.8h, v2.16b, %[b0a].16b\n"
+                "str	q16, [%[c_ptr]]\n"
+                "smlal2	v13.8h, v2.16b, %[b1a].16b\n"
+                "smlal2	v14.8h, v2.16b, %[b2a].16b\n"
+                "smlal2	v15.8h, v2.16b, %[b3a].16b\n"
+
+                "sadalp	v24.4s, v12.8h\n"
+                "smull	v12.8h, v3.8b, %[b0a].8b\n"
+                "sadalp	v25.4s, v13.8h\n"
+                "sadalp	v26.4s, v14.8h\n"
+                "smull	v13.8h, v3.8b, %[b1a].8b\n"
+                "sadalp	v27.4s, v15.8h\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "smull	v14.8h, v3.8b, %[b2a].8b\n"
+                "addp	v20.4s, v24.4s, v25.4s\n"
+                "addp	v21.4s, v26.4s, v27.4s\n"
+                "smull	v15.8h, v3.8b, %[b3a].8b\n"
+                "smlal2	v12.8h, v3.16b, %[b0a].16b\n"
+                "str	q17, [%[c_ptr], #16]\n"
+                "smlal2	v13.8h, v3.16b, %[b1a].16b\n"
+                "smlal2	v14.8h, v3.16b, %[b2a].16b\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "smlal2	v15.8h, v3.16b, %[b3a].16b\n"
+                "b	3f\n"
 
                 // Detached final iteration (odd K)
                 "2:\n"
-                "smull    v14.8h, v0.8b, %[b2].8b\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "smull    v15.8h, v0.8b, %[b3].8b\n"
-                "add    %[b_ptr], %[b_ptr], #64\n"
-                "smlal2    v12.8h, v0.16b, %[b0].16b\n"
-                "smlal2    v13.8h, v0.16b, %[b1].16b\n"
-                "smlal2    v14.8h, v0.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v0.16b, %[b3].16b\n"
-
-                "sadalp    v16.4s, v12.8h\n"
-                "smull    v12.8h, v1.8b, %[b0].8b\n"
-                "sadalp    v17.4s, v13.8h\n"
-                "sadalp    v18.4s, v14.8h\n"
-                "smull    v13.8h, v1.8b, %[b1].8b\n"
-                "sadalp    v19.4s, v15.8h\n"
-                "smull    v14.8h, v1.8b, %[b2].8b\n"
-                "smull    v15.8h, v1.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v1.16b, %[b0].16b\n"
-                "addp    v16.4s, v16.4s, v17.4s\n"
-                "smlal2    v13.8h, v1.16b, %[b1].16b\n"
-                "addp    v17.4s, v18.4s, v19.4s\n"
-                "smlal2    v14.8h, v1.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v1.16b, %[b3].16b\n"
-
-                "sadalp    v20.4s, v12.8h\n"
-                "smull    v12.8h, v2.8b, %[b0].8b\n"
-                "sadalp    v21.4s, v13.8h\n"
-                "sadalp    v22.4s, v14.8h\n"
-                "smull    v13.8h, v2.8b, %[b1].8b\n"
-                "sadalp    v23.4s, v15.8h\n"
-                "addp    v16.4s, v16.4s, v17.4s\n"
-                "smull    v14.8h, v2.8b, %[b2].8b\n"
-                "addp    v18.4s, v20.4s, v21.4s\n"
-                "addp    v19.4s, v22.4s, v23.4s\n"
-                "smull    v15.8h, v2.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v2.16b, %[b0].16b\n"
-                "str    q16, [%[c_ptr]]\n"
-                "smlal2    v13.8h, v2.16b, %[b1].16b\n"
-                "smlal2    v14.8h, v2.16b, %[b2].16b\n"
-                "smlal2    v15.8h, v2.16b, %[b3].16b\n"
-
-                "sadalp    v24.4s, v12.8h\n"
-                "smull    v12.8h, v3.8b, %[b0].8b\n"
-                "sadalp    v25.4s, v13.8h\n"
-                "sadalp    v26.4s, v14.8h\n"
-                "smull    v13.8h, v3.8b, %[b1].8b\n"
-                "sadalp    v27.4s, v15.8h\n"
-                "addp    v17.4s, v18.4s, v19.4s\n"
-                "smull    v14.8h, v3.8b, %[b2].8b\n"
-                "addp    v20.4s, v24.4s, v25.4s\n"
-                "addp    v21.4s, v26.4s, v27.4s\n"
-                "smull    v15.8h, v3.8b, %[b3].8b\n"
-                "smlal2    v12.8h, v3.16b, %[b0].16b\n"
-                "str    q17, [%[c_ptr], #16]\n"
-                "smlal2    v13.8h, v3.16b, %[b1].16b\n"
-                "smlal2    v14.8h, v3.16b, %[b2].16b\n"
-                "addp    v18.4s, v20.4s, v21.4s\n"
-                "smlal2    v15.8h, v3.16b, %[b3].16b\n"
+                "smull	v14.8h, v0.8b, %[b2].8b\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "smull	v15.8h, v0.8b, %[b3].8b\n"
+                "add	%[b_ptr], %[b_ptr], #64\n"
+                "smlal2	v12.8h, v0.16b, %[b0].16b\n"
+                "smlal2	v13.8h, v0.16b, %[b1].16b\n"
+                "smlal2	v14.8h, v0.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v0.16b, %[b3].16b\n"
+
+                "sadalp	v16.4s, v12.8h\n"
+                "smull	v12.8h, v1.8b, %[b0].8b\n"
+                "sadalp	v17.4s, v13.8h\n"
+                "sadalp	v18.4s, v14.8h\n"
+                "smull	v13.8h, v1.8b, %[b1].8b\n"
+                "sadalp	v19.4s, v15.8h\n"
+                "smull	v14.8h, v1.8b, %[b2].8b\n"
+                "smull	v15.8h, v1.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v1.16b, %[b0].16b\n"
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "smlal2	v13.8h, v1.16b, %[b1].16b\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "smlal2	v14.8h, v1.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v1.16b, %[b3].16b\n"
+
+                "sadalp	v20.4s, v12.8h\n"
+                "smull	v12.8h, v2.8b, %[b0].8b\n"
+                "sadalp	v21.4s, v13.8h\n"
+                "sadalp	v22.4s, v14.8h\n"
+                "smull	v13.8h, v2.8b, %[b1].8b\n"
+                "sadalp	v23.4s, v15.8h\n"
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "smull	v14.8h, v2.8b, %[b2].8b\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "addp	v19.4s, v22.4s, v23.4s\n"
+                "smull	v15.8h, v2.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v2.16b, %[b0].16b\n"
+                "str	q16, [%[c_ptr]]\n"
+                "smlal2	v13.8h, v2.16b, %[b1].16b\n"
+                "smlal2	v14.8h, v2.16b, %[b2].16b\n"
+                "smlal2	v15.8h, v2.16b, %[b3].16b\n"
+
+                "sadalp	v24.4s, v12.8h\n"
+                "smull	v12.8h, v3.8b, %[b0].8b\n"
+                "sadalp	v25.4s, v13.8h\n"
+                "sadalp	v26.4s, v14.8h\n"
+                "smull	v13.8h, v3.8b, %[b1].8b\n"
+                "sadalp	v27.4s, v15.8h\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "smull	v14.8h, v3.8b, %[b2].8b\n"
+                "addp	v20.4s, v24.4s, v25.4s\n"
+                "addp	v21.4s, v26.4s, v27.4s\n"
+                "smull	v15.8h, v3.8b, %[b3].8b\n"
+                "smlal2	v12.8h, v3.16b, %[b0].16b\n"
+                "str	q17, [%[c_ptr], #16]\n"
+                "smlal2	v13.8h, v3.16b, %[b1].16b\n"
+                "smlal2	v14.8h, v3.16b, %[b2].16b\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "smlal2	v15.8h, v3.16b, %[b3].16b\n"
 
                 "3:\n"
 
                 // Final additions
-                "sadalp    v28.4s, v12.8h\n"
-                "str    q18, [%[c_ptr], #32]\n"
-                "sadalp    v29.4s, v13.8h\n"
-                "sadalp    v30.4s, v14.8h\n"
-                "sadalp    v31.4s, v15.8h\n"
+                "sadalp	v28.4s, v12.8h\n"
+                "str	q18, [%[c_ptr], #32]\n"
+                "sadalp	v29.4s, v13.8h\n"
+                "sadalp	v30.4s, v14.8h\n"
+                "sadalp	v31.4s, v15.8h\n"
 
                 // Horizontal reduction, phase 1
-                "addp    v22.4s, v28.4s, v29.4s\n"
-                "addp    v23.4s, v30.4s, v31.4s\n"
+                "addp	v22.4s, v28.4s, v29.4s\n"
+                "addp	v23.4s, v30.4s, v31.4s\n"
 
                 // Horizontal reduction, phase 2
-                "addp    v19.4s, v22.4s, v23.4s\n"
-                "str    q19, [%[c_ptr], #48]\n"
-                "add    %[c_ptr], %[c_ptr], #64\n"
+                "addp	v19.4s, v22.4s, v23.4s\n"
+                "str	q19, [%[c_ptr], #48]\n"
+                "add	%[c_ptr], %[c_ptr], #64\n"
 
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp
index d0c64b2f6c..cc4b81ecb5 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12.hpp
@@ -39,8 +39,7 @@ void a64_gemm_s8_8x12_x1(const int8_t *, const int8_t *, int32_t *, int, int, in
 
 class cls_a64_gemm_s8_8x12 {
 public:
-    typedef int8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
+    typedef int8_t operand_type;
     typedef int32_t result_type;
 
     typedef void (*kern_type)(const int8_t *, const int8_t *, int32_t *, int, int, int);
@@ -59,8 +58,8 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 4> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
 
     template<typename T>
     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp
index fc46781100..5ba14d2409 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp
@@ -111,11 +111,11 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
 
                 "1:\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
-                "subs    %w[k], %w[k], #1\n"
+                "subs	%w[k], %w[k], #1\n"
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
 
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 "ins    %[b2].d[1], x20\n"
@@ -123,7 +123,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[a_ptr], #40]\n"
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
 
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 "ins    %[a0a].d[1], x20\n"
@@ -131,7 +131,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[a_ptr], #56]\n"
                 ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
                 "ins    %[a1a].d[1], x20\n"
@@ -139,7 +139,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[b_ptr], #56]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 "ins    %[b0].d[1], x20\n"
@@ -155,8 +155,8 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
 
-        // Unroll 1
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
+		// Unroll 1
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x4f85e048 // sdot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
                 "ins    %[b1].d[1], x20\n"
@@ -164,7 +164,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[b_ptr], #88]\n"
                 ".word 0x4f85e84a // sdot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
                 ".word 0x4fa5e84b // sdot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
-                "ldr    %d[a0], [%[a_ptr], #64]\n"
+                "ldr	%d[a0], [%[a_ptr], #64]\n"
 
                 ".word 0x4f86e04c // sdot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
                 "ins    %[b2].d[1], x20\n"
@@ -172,7 +172,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[a_ptr], #72]\n"
                 ".word 0x4f86e84e // sdot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
                 ".word 0x4fa6e84f // sdot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
-                "ldr    %d[a1], [%[a_ptr], #80]\n"
+                "ldr	%d[a1], [%[a_ptr], #80]\n"
 
                 ".word 0x4f85e070 // sdot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
                 "ins    %[a0].d[1], x20\n"
@@ -180,7 +180,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[a_ptr], #88]\n"
                 ".word 0x4f85e872 // sdot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
                 ".word 0x4fa5e873 // sdot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #96]\n"
+                "ldr	%d[b0], [%[b_ptr], #96]\n"
 
                 ".word 0x4f86e074 // sdot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
                 "ins    %[a1].d[1], x20\n"
@@ -188,7 +188,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[b_ptr], #104]\n"
                 ".word 0x4f86e876 // sdot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
                 ".word 0x4fa6e877 // sdot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #112]\n"
+                "ldr	%d[b1], [%[b_ptr], #112]\n"
 
                 ".word 0x4f85e098 // sdot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
                 "ins    %[b0].d[1], x20\n"
@@ -196,19 +196,19 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[b_ptr], #120]\n"
                 ".word 0x4f85e89a // sdot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
                 ".word 0x4fa5e89b // sdot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
 
                 ".word 0x4f86e09c // sdot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
                 ASM_PREFETCH("[%[b_ptr], #640]")
                 ".word 0x4fa6e09d // sdot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x4f86e89e // sdot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
                 "ins    %[b1].d[1], x20\n"
                 ".word 0x4fa6e89f // sdot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
                 "ldr    %d[b2], [%[b_ptr], #32]\n"
 
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
-                "b.ne    1b\n"
+                "b.ne	1b\n"
 
                 // Branch here if K=1 or 2.  Do the right thing for odd/even at the end.
                 "4:\n"
@@ -221,7 +221,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
 
                 // Even K continuation
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
 
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 "ins    %[b2].d[1], x20\n"
@@ -230,7 +230,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ASM_PREFETCHW("[%[c_ptr]]")
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
 
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 "ins    %[a0a].d[1], x20\n"
@@ -238,7 +238,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 "ldr    x20, [%[a_ptr], #56]\n"
                 ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
                 "ins    %[a1a].d[1], x20\n"
@@ -253,7 +253,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 ASM_PREFETCHW("[%[c_ptr], #128]")
                 ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
                 "ins    %[b0].d[1], x20\n"
@@ -262,7 +262,7 @@ void a64_gemm_s8_8x12_a55r1(const int8_t *Apanel, const int8_t *Bpanel, int32_t
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ASM_PREFETCHW("[%[c_ptr], #192]")
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x4f85e048 // sdot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
                 "ins    %[b1].d[1], x20\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp
index f25947da26..98bff14104 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp
@@ -53,63 +53,63 @@ void a64_gemm_s8_8x12(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
             register int32x4_t a1a asm("v6");
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.4s, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.4s, #0x0\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "movi    v11.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.4s, #0x0\n"
+                "movi	v8.4s, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.4s, #0x0\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "movi	v11.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v13.4s, #0x0\n"
+                "movi	v13.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v14.4s, #0x0\n"
+                "movi	v14.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.4s, #0x0\n"
+                "movi	v15.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v16.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v17.4s, #0x0\n"
+                "movi	v17.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v18.4s, #0x0\n"
+                "movi	v18.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v19.4s, #0x0\n"
+                "movi	v19.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.4s, #0x0\n"
+                "movi	v20.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "movi    v21.4s, #0x0\n"
+                "movi	v21.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "movi    v22.4s, #0x0\n"
-                "movi    v23.4s, #0x0\n"
-                "movi    v24.4s, #0x0\n"
-                "movi    v25.4s, #0x0\n"
-                "movi    v26.4s, #0x0\n"
-                "movi    v27.4s, #0x0\n"
-                "movi    v28.4s, #0x0\n"
-                "movi    v29.4s, #0x0\n"
-                "movi    v30.4s, #0x0\n"
-                "movi    v31.4s, #0x0\n"
+                "movi	v22.4s, #0x0\n"
+                "movi	v23.4s, #0x0\n"
+                "movi	v24.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 // Loop proper
                 "1:\n"
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
 
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0a], [%[a_ptr], #32]\n"
+                "ldr	%q[a0a], [%[a_ptr], #32]\n"
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "ldr    %q[a1a], [%[a_ptr], #48]\n"
+                "ldr	%q[a1a], [%[a_ptr], #48]\n"
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -120,7 +120,7 @@ void a64_gemm_s8_8x12(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
                 ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
@@ -131,19 +131,19 @@ void a64_gemm_s8_8x12(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
                 ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x4f85e048 // sdot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
                 ".word 0x4fa5e049 // sdot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
-                "ldr    %q[a0], [%[a_ptr], #64]\n"
+                "ldr	%q[a0], [%[a_ptr], #64]\n"
                 ".word 0x4f85e84a // sdot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
                 ".word 0x4fa5e84b // sdot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
                 ".word 0x4f86e04c // sdot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
-                "ldr    %q[a1], [%[a_ptr], #80]\n"
+                "ldr	%q[a1], [%[a_ptr], #80]\n"
                 ".word 0x4fa6e04d // sdot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
                 ".word 0x4f86e84e // sdot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
                 ".word 0x4fa6e84f // sdot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #96]\n"
+                "ldr	%q[b0], [%[b_ptr], #96]\n"
 
                 ".word 0x4f85e070 // sdot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
                 ".word 0x4fa5e071 // sdot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
@@ -154,40 +154,40 @@ void a64_gemm_s8_8x12(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
                 ".word 0x4fa6e075 // sdot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
                 ".word 0x4f86e876 // sdot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
                 ".word 0x4fa6e877 // sdot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #112]\n"
+                "ldr	%q[b1], [%[b_ptr], #112]\n"
 
                 ".word 0x4f85e098 // sdot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
                 ".word 0x4fa5e099 // sdot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
                 ".word 0x4f85e89a // sdot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
                 ".word 0x4fa5e89b // sdot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x4f86e09c // sdot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
                 ".word 0x4fa6e09d // sdot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
-                "subs    %w[k], %w[k], #1\n"
+                "subs	%w[k], %w[k], #1\n"
                 ".word 0x4f86e89e // sdot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
                 ".word 0x4fa6e89f // sdot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
-                "bne    1b\n"
+                "bne	1b\n"
 
                 // Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
                 "4:\n"
 
                 // Branch to alternative tail for odd K
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration (even K)
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0a], [%[a_ptr], #32]\n"
+                "ldr	%q[a0a], [%[a_ptr], #32]\n"
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "ldr    %q[a1a], [%[a_ptr], #48]\n"
+                "ldr	%q[a1a], [%[a_ptr], #48]\n"
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -197,140 +197,140 @@ void a64_gemm_s8_8x12(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cpane
                 ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
                 ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
                 ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x4f85e048 // sdot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
 
                 ".word 0x4f85e070 // sdot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x4fa5e049 // sdot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
+                "str	q8, [%[c_ptr], #0]\n"
                 ".word 0x4fa5e071 // sdot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
+                "str	q16, [%[c_ptr], #16]\n"
                 ".word 0x4f85e098 // sdot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
-                "str    q24, [%[c_ptr], #32]\n"
+                "str	q24, [%[c_ptr], #32]\n"
 
                 ".word 0x4fa5e099 // sdot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
+                "str	q9, [%[c_ptr], #48]\n"
                 ".word 0x4f85e84a // sdot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
+                "str	q17, [%[c_ptr], #64]\n"
                 ".word 0x4f85e872 // sdot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
+                "str	q25, [%[c_ptr], #80]\n"
                 ".word 0x4f85e89a // sdot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
+                "str	q10, [%[c_ptr], #96]\n"
 
                 ".word 0x4fa5e84b // sdot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
+                "str	q18, [%[c_ptr], #112]\n"
                 ".word 0x4fa5e873 // sdot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
+                "str	q26, [%[c_ptr], #128]\n"
                 ".word 0x4fa5e89b // sdot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
+                "str	q11, [%[c_ptr], #144]\n"
 
                 ".word 0x4f86e04c // sdot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
+                "str	q19, [%[c_ptr], #160]\n"
                 ".word 0x4f86e074 // sdot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
+                "str	q27, [%[c_ptr], #176]\n"
                 ".word 0x4f86e09c // sdot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 ".word 0x4fa6e04d // sdot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
+                "str	q20, [%[c_ptr], #208]\n"
                 ".word 0x4fa6e075 // sdot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
+                "str	q28, [%[c_ptr], #224]\n"
                 ".word 0x4fa6e09d // sdot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
+                "str	q13, [%[c_ptr], #240]\n"
 
                 ".word 0x4f86e84e // sdot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
+                "str	q21, [%[c_ptr], #256]\n"
                 ".word 0x4f86e876 // sdot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
+                "str	q29, [%[c_ptr], #272]\n"
                 ".word 0x4f86e89e // sdot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
+                "str	q14, [%[c_ptr], #288]\n"
 
                 ".word 0x4fa6e84f // sdot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
+                "str	q22, [%[c_ptr], #304]\n"
                 ".word 0x4fa6e877 // sdot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
+                "str	q30, [%[c_ptr], #320]\n"
                 ".word 0x4fa6e89f // sdot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
-                "b    3f\n"
+                "b	3f\n"
 
                 // Detached final iteration (odd K)
                 "2:\n"
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
+                "str	q8, [%[c_ptr], #0]\n"
                 ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
+                "str	q16, [%[c_ptr], #16]\n"
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "str    q24, [%[c_ptr], #32]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "str	q24, [%[c_ptr], #32]\n"
                 ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
+                "str	q9, [%[c_ptr], #48]\n"
 
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
+                "str	q17, [%[c_ptr], #64]\n"
                 ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
+                "str	q25, [%[c_ptr], #80]\n"
                 ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
+                "str	q10, [%[c_ptr], #96]\n"
 
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
+                "str	q18, [%[c_ptr], #112]\n"
                 ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
+                "str	q26, [%[c_ptr], #128]\n"
                 ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
+                "str	q11, [%[c_ptr], #144]\n"
 
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
+                "str	q19, [%[c_ptr], #160]\n"
                 ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
+                "str	q27, [%[c_ptr], #176]\n"
                 ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
+                "str	q20, [%[c_ptr], #208]\n"
                 ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
+                "str	q28, [%[c_ptr], #224]\n"
                 ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
+                "str	q13, [%[c_ptr], #240]\n"
 
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
+                "str	q21, [%[c_ptr], #256]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
+                "str	q29, [%[c_ptr], #272]\n"
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
+                "str	q14, [%[c_ptr], #288]\n"
 
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
+                "str	q22, [%[c_ptr], #304]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
+                "str	q30, [%[c_ptr], #320]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
 
                 // Common tail
                 "3:\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
 
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp
index 30f819d45e..1f9d001553 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp
@@ -52,60 +52,60 @@ void a64_gemm_s8_8x12_x1(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cp
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.4s, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.4s, #0x0\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "movi    v11.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.4s, #0x0\n"
+                "movi	v8.4s, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.4s, #0x0\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "movi	v11.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v13.4s, #0x0\n"
+                "movi	v13.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v14.4s, #0x0\n"
+                "movi	v14.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.4s, #0x0\n"
+                "movi	v15.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v16.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v17.4s, #0x0\n"
+                "movi	v17.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v18.4s, #0x0\n"
+                "movi	v18.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v19.4s, #0x0\n"
+                "movi	v19.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.4s, #0x0\n"
+                "movi	v20.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "movi    v21.4s, #0x0\n"
+                "movi	v21.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "movi    v22.4s, #0x0\n"
-                "movi    v23.4s, #0x0\n"
-                "movi    v24.4s, #0x0\n"
-                "movi    v25.4s, #0x0\n"
-                "movi    v26.4s, #0x0\n"
-                "movi    v27.4s, #0x0\n"
-                "movi    v28.4s, #0x0\n"
-                "movi    v29.4s, #0x0\n"
-                "movi    v30.4s, #0x0\n"
-                "movi    v31.4s, #0x0\n"
+                "movi	v22.4s, #0x0\n"
+                "movi	v23.4s, #0x0\n"
+                "movi	v24.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 // Loop proper
                 "1:\n"
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -116,74 +116,74 @@ void a64_gemm_s8_8x12_x1(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cp
                 ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #448]")
                 ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0], [%[a_ptr], #32]\n"
+                "ldr	%q[a0], [%[a_ptr], #32]\n"
                 ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #48]\n"
+                "ldr	%q[a1], [%[a_ptr], #48]\n"
 
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #96]\n"
+                "ldr	%q[b0], [%[b_ptr], #96]\n"
 
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #512]")
                 ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "subs    %w[k], %w[k], #1\n"
+                "subs	%w[k], %w[k], #1\n"
                 ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #112]\n"
+                "ldr	%q[b1], [%[b_ptr], #112]\n"
 
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
                 ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
                 ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "bne    1b\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "bne	1b\n"
 
                 // Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
                 "4:\n"
 
                 // Branch to alternative tail for odd K
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration (even K)
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -193,142 +193,142 @@ void a64_gemm_s8_8x12_x1(const int8_t *Apanel, const int8_t *Bpanel, int32_t *Cp
                 ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
                 ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0], [%[a_ptr], #-32]\n"
+                "ldr	%q[a0], [%[a_ptr], #-32]\n"
                 ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
                 ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #-16]\n"
+                "ldr	%q[a1], [%[a_ptr], #-16]\n"
 
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
+                "str	q8, [%[c_ptr], #0]\n"
                 ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
+                "str	q16, [%[c_ptr], #16]\n"
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
-                "str    q24, [%[c_ptr], #32]\n"
+                "str	q24, [%[c_ptr], #32]\n"
 
                 ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
+                "str	q9, [%[c_ptr], #48]\n"
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
+                "str	q17, [%[c_ptr], #64]\n"
                 ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
+                "str	q25, [%[c_ptr], #80]\n"
                 ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
+                "str	q10, [%[c_ptr], #96]\n"
 
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
+                "str	q18, [%[c_ptr], #112]\n"
                 ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
+                "str	q26, [%[c_ptr], #128]\n"
                 ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
+                "str	q11, [%[c_ptr], #144]\n"
 
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
+                "str	q19, [%[c_ptr], #160]\n"
                 ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
+                "str	q27, [%[c_ptr], #176]\n"
                 ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
+                "str	q20, [%[c_ptr], #208]\n"
                 ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
+                "str	q28, [%[c_ptr], #224]\n"
                 ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
+                "str	q13, [%[c_ptr], #240]\n"
 
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
+                "str	q21, [%[c_ptr], #256]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
+                "str	q29, [%[c_ptr], #272]\n"
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
+                "str	q14, [%[c_ptr], #288]\n"
 
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
+                "str	q22, [%[c_ptr], #304]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
+                "str	q30, [%[c_ptr], #320]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
-                "b    3f\n"
+                "b	3f\n"
 
                 // Detached final iteration (odd K)
                 "2:\n"
                 ".word 0x4f80e048 // sdot v8.4s , %[b0].16b, %[a0].4b[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x4f80e070 // sdot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x4fa0e049 // sdot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
+                "str	q8, [%[c_ptr], #0]\n"
                 ".word 0x4fa0e071 // sdot v17.4s, %[b1].16b, %[a0].4b[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
+                "str	q16, [%[c_ptr], #16]\n"
                 ".word 0x4f80e098 // sdot v24.4s, %[b2].16b, %[a0].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "str    q24, [%[c_ptr], #32]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "str	q24, [%[c_ptr], #32]\n"
                 ".word 0x4fa0e099 // sdot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
+                "str	q9, [%[c_ptr], #48]\n"
 
                 ".word 0x4f80e84a // sdot v10.4s, %[b0].16b, %[a0].4b[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
+                "str	q17, [%[c_ptr], #64]\n"
                 ".word 0x4f80e872 // sdot v18.4s, %[b1].16b, %[a0].4b[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
+                "str	q25, [%[c_ptr], #80]\n"
                 ".word 0x4f80e89a // sdot v26.4s, %[b2].16b, %[a0].4b[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
+                "str	q10, [%[c_ptr], #96]\n"
 
                 ".word 0x4fa0e84b // sdot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
+                "str	q18, [%[c_ptr], #112]\n"
                 ".word 0x4fa0e873 // sdot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
+                "str	q26, [%[c_ptr], #128]\n"
                 ".word 0x4fa0e89b // sdot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
+                "str	q11, [%[c_ptr], #144]\n"
 
                 ".word 0x4f81e04c // sdot v12.4s, %[b0].16b, %[a1].4b[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
+                "str	q19, [%[c_ptr], #160]\n"
                 ".word 0x4f81e074 // sdot v20.4s, %[b1].16b, %[a1].4b[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
+                "str	q27, [%[c_ptr], #176]\n"
                 ".word 0x4f81e09c // sdot v28.4s, %[b2].16b, %[a1].4b[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 ".word 0x4fa1e04d // sdot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
+                "str	q20, [%[c_ptr], #208]\n"
                 ".word 0x4fa1e075 // sdot v21.4s, %[b1].16b, %[a1].4b[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
+                "str	q28, [%[c_ptr], #224]\n"
                 ".word 0x4fa1e09d // sdot v29.4s, %[b2].16b, %[a1].4b[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
+                "str	q13, [%[c_ptr], #240]\n"
 
                 ".word 0x4f81e84e // sdot v14.4s, %[b0].16b, %[a1].4b[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
+                "str	q21, [%[c_ptr], #256]\n"
                 ".word 0x4f81e876 // sdot v22.4s, %[b1].16b, %[a1].4b[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
+                "str	q29, [%[c_ptr], #272]\n"
                 ".word 0x4f81e89e // sdot v30.4s, %[b2].16b, %[a1].4b[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
+                "str	q14, [%[c_ptr], #288]\n"
 
                 ".word 0x4fa1e84f // sdot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
+                "str	q22, [%[c_ptr], #304]\n"
                 ".word 0x4fa1e877 // sdot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
+                "str	q30, [%[c_ptr], #320]\n"
                 ".word 0x4fa1e89f // sdot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
 
                 // Common tail
                 "3:\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "+w" (a0), [a1] "+w" (a1),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12.hpp
index af13fbd4e9..c7295275e3 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12.hpp
@@ -34,8 +34,7 @@ void a64_gemm_u16_asimd_8x12(const uint16_t *, const uint16_t *, uint32_t *, int
 
 class cls_a64_gemm_u16_8x12 {
 public:
-    typedef uint16_t lhs_operand_type;
-    typedef uint16_t rhs_operand_type;
+    typedef uint16_t operand_type;
     typedef uint32_t result_type;
 
     typedef void (*kern_type)(const uint16_t *, const uint16_t *, uint32_t *, int, int, int);
@@ -54,8 +53,8 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 1, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 1, true> transforms_quantized = {};
 
     kern_type kernel = a64_gemm_u16_asimd_8x12;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp
index cdc902f2cc..6d27dd73f1 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp
@@ -35,8 +35,7 @@ void a64_gemm_u8_4x4(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpa
 
 class cls_a64_gemm_u8_4x4 {
 public:
-    typedef uint8_t lhs_operand_type;
-    typedef uint8_t rhs_operand_type;
+    typedef uint8_t operand_type;
     typedef uint32_t result_type;
 
     typedef void (*kern_type)(const uint8_t *, const uint8_t *, uint32_t *, int, int, int);
@@ -65,8 +64,8 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 4, 16> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 4, 16, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 4, 4, 16> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 4, 4, 16, true> transforms_quantized = {};
 
     template<typename T>
     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp
index c227f21702..495a81692d 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp
@@ -49,225 +49,225 @@ void a64_gemm_u8_4x4(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpa
             register uint8x16_t b3  asm("v7");
 
             __asm __volatile (
-                "movi    v16.4s, #0x0\n"
-                "ldr    q0, [%[a_ptr]]\n"
-                "movi    v17.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v18.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v19.4s, #0x0\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "movi    v20.4s, #0x0\n"
-                "ldr    %q[b3], [%[b_ptr], #48]\n"
-                "movi    v21.4s, #0x0\n"
-                "ldr    q1, [%[a_ptr], #16]\n"
-                "movi    v22.4s, #0x0\n"
-                "ldr    q2, [%[a_ptr], #32]\n"
-                "movi    v23.4s, #0x0\n"
-                "ldr    q3, [%[a_ptr], #48]\n"
-                "movi    v24.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
+                "ldr	q0, [%[a_ptr]]\n"
+                "movi	v17.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v18.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v19.4s, #0x0\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "movi	v20.4s, #0x0\n"
+                "ldr	%q[b3], [%[b_ptr], #48]\n"
+                "movi	v21.4s, #0x0\n"
+                "ldr	q1, [%[a_ptr], #16]\n"
+                "movi	v22.4s, #0x0\n"
+                "ldr	q2, [%[a_ptr], #32]\n"
+                "movi	v23.4s, #0x0\n"
+                "ldr	q3, [%[a_ptr], #48]\n"
+                "movi	v24.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v25.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v26.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v27.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v28.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v29.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v30.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v31.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
 
-                "umull    v12.8h, v0.8b, %[b0].8b\n"
-        "add    %[a_ptr], %[a_ptr], #64\n"
-                "umull    v13.8h, v0.8b, %[b1].8b\n"
-                "umull    v14.8h, v0.8b, %[b2].8b\n"
-        "add    %[b_ptr], %[b_ptr], #64\n"
-                "umull    v15.8h, v0.8b, %[b3].8b\n"
+                "umull	v12.8h, v0.8b, %[b0].8b\n"
+		"add	%[a_ptr], %[a_ptr], #64\n"
+                "umull	v13.8h, v0.8b, %[b1].8b\n"
+                "umull	v14.8h, v0.8b, %[b2].8b\n"
+		"add	%[b_ptr], %[b_ptr], #64\n"
+                "umull	v15.8h, v0.8b, %[b3].8b\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 2f\n"
+                "cbz	%w[k], 2f\n"
 
                 "1:\n"
-                "uadalp    v16.4s, v12.8h\n"
-                "umull2    v12.8h, v0.16b, %[b0].16b\n"
-                "uadalp    v17.4s, v13.8h\n"
-                "umull2    v13.8h, v0.16b, %[b1].16b\n"
-                "uadalp    v18.4s, v14.8h\n"
-                "umull2    v14.8h, v0.16b, %[b2].16b\n"
-                "uadalp    v19.4s, v15.8h\n"
-                "umull2    v15.8h, v0.16b, %[b3].16b\n"
-                "ldr     q0, [%[a_ptr]]\n"
-
-                "uadalp    v16.4s, v12.8h\n"
-                "umull    v12.8h, v1.8b, %[b0].8b\n"
-                "uadalp    v17.4s, v13.8h\n"
-                "umull    v13.8h, v1.8b, %[b1].8b\n"
-                "subs    %w[k], %w[k], #1\n"
-                "uadalp    v18.4s, v14.8h\n"
-                "umull    v14.8h, v1.8b, %[b2].8b\n"
-                "uadalp    v19.4s, v15.8h\n"
-                "umull    v15.8h, v1.8b, %[b3].8b\n"
-
-                "uadalp    v20.4s, v12.8h\n"
-                "umull2    v12.8h, v1.16b, %[b0].16b\n"
-                "uadalp    v21.4s, v13.8h\n"
-                "umull2    v13.8h, v1.16b, %[b1].16b\n"
+                "uadalp	v16.4s, v12.8h\n"
+                "umull2	v12.8h, v0.16b, %[b0].16b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "umull2	v13.8h, v0.16b, %[b1].16b\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull2	v14.8h, v0.16b, %[b2].16b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "umull2	v15.8h, v0.16b, %[b3].16b\n"
+                "ldr 	q0, [%[a_ptr]]\n"
+
+                "uadalp	v16.4s, v12.8h\n"
+                "umull	v12.8h, v1.8b, %[b0].8b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "umull	v13.8h, v1.8b, %[b1].8b\n"
+                "subs	%w[k], %w[k], #1\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull	v14.8h, v1.8b, %[b2].8b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "umull	v15.8h, v1.8b, %[b3].8b\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull2	v12.8h, v1.16b, %[b0].16b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "umull2	v13.8h, v1.16b, %[b1].16b\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "uadalp    v22.4s, v14.8h\n"
-                "umull2    v14.8h, v1.16b, %[b2].16b\n"
-                "uadalp    v23.4s, v15.8h\n"
-                "umull2    v15.8h, v1.16b, %[b3].16b\n"
-                "ldr     q1, [%[a_ptr], #16]\n"
-
-                "uadalp    v20.4s, v12.8h\n"
-                "umull    v12.8h, v2.8b, %[b0].8b\n"
-                "uadalp    v21.4s, v13.8h\n"
-                "umull    v13.8h, v2.8b, %[b1].8b\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull2	v14.8h, v1.16b, %[b2].16b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "umull2	v15.8h, v1.16b, %[b3].16b\n"
+                "ldr 	q1, [%[a_ptr], #16]\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull	v12.8h, v2.8b, %[b0].8b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "umull	v13.8h, v2.8b, %[b1].8b\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "uadalp    v22.4s, v14.8h\n"
-                "umull    v14.8h, v2.8b, %[b2].8b\n"
-                "uadalp    v23.4s, v15.8h\n"
-                "umull    v15.8h, v2.8b, %[b3].8b\n"
-
-                "uadalp    v24.4s, v12.8h\n"
-                "umull2    v12.8h, v2.16b, %[b0].16b\n"
-                "uadalp    v25.4s, v13.8h\n"
-                "umull2    v13.8h, v2.16b, %[b1].16b\n"
-                "uadalp    v26.4s, v14.8h\n"
-                "umull2    v14.8h, v2.16b, %[b2].16b\n"
-                "uadalp    v27.4s, v15.8h\n"
-                "umull2    v15.8h, v2.16b, %[b3].16b\n"
-                "ldr    q2, [%[a_ptr], #32]\n"
-
-                "uadalp    v24.4s, v12.8h\n"
-                "umull    v12.8h, v3.8b, %[b0].8b\n"
-                "uadalp    v25.4s, v13.8h\n"
-                "umull    v13.8h, v3.8b, %[b1].8b\n"
-                "uadalp    v26.4s, v14.8h\n"
-                "umull    v14.8h, v3.8b, %[b2].8b\n"
-                "uadalp    v27.4s, v15.8h\n"
-                "umull    v15.8h, v3.8b, %[b3].8b\n"
-
-                "uadalp    v28.4s, v12.8h\n"
-                "umull2    v12.8h, v3.16b, %[b0].16b\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "uadalp    v29.4s, v13.8h\n"
-                "umull2    v13.8h, v3.16b, %[b1].16b\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "uadalp    v30.4s, v14.8h\n"
-                "umull2    v14.8h, v3.16b, %[b2].16b\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "uadalp    v31.4s, v15.8h\n"
-                "umull2    v15.8h, v3.16b, %[b3].16b\n"
-                "ldr    %q[b3], [%[b_ptr], #48]\n"
-
-                "uadalp    v28.4s, v12.8h\n"
-                "umull    v12.8h, v0.8b, %[b0].8b\n"
-                "add    %[b_ptr], %[b_ptr], #64\n"
-                "uadalp    v29.4s, v13.8h\n"
-                "umull    v13.8h, v0.8b, %[b1].8b\n"
-                "ldr    q3, [%[a_ptr], #48]\n"
-                "uadalp    v30.4s, v14.8h\n"
-                "umull    v14.8h, v0.8b, %[b2].8b\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "uadalp    v31.4s, v15.8h\n"
-                "umull    v15.8h, v0.8b, %[b3].8b\n"
-                "bne    1b\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull	v14.8h, v2.8b, %[b2].8b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "umull	v15.8h, v2.8b, %[b3].8b\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull2	v12.8h, v2.16b, %[b0].16b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "umull2	v13.8h, v2.16b, %[b1].16b\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull2	v14.8h, v2.16b, %[b2].16b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "umull2	v15.8h, v2.16b, %[b3].16b\n"
+                "ldr	q2, [%[a_ptr], #32]\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull	v12.8h, v3.8b, %[b0].8b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "umull	v13.8h, v3.8b, %[b1].8b\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull	v14.8h, v3.8b, %[b2].8b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "umull	v15.8h, v3.8b, %[b3].8b\n"
+
+                "uadalp	v28.4s, v12.8h\n"
+                "umull2	v12.8h, v3.16b, %[b0].16b\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "uadalp	v29.4s, v13.8h\n"
+                "umull2	v13.8h, v3.16b, %[b1].16b\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "uadalp	v30.4s, v14.8h\n"
+                "umull2	v14.8h, v3.16b, %[b2].16b\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "uadalp	v31.4s, v15.8h\n"
+                "umull2	v15.8h, v3.16b, %[b3].16b\n"
+                "ldr	%q[b3], [%[b_ptr], #48]\n"
+
+                "uadalp	v28.4s, v12.8h\n"
+                "umull	v12.8h, v0.8b, %[b0].8b\n"
+                "add	%[b_ptr], %[b_ptr], #64\n"
+                "uadalp	v29.4s, v13.8h\n"
+                "umull	v13.8h, v0.8b, %[b1].8b\n"
+                "ldr	q3, [%[a_ptr], #48]\n"
+                "uadalp	v30.4s, v14.8h\n"
+                "umull	v14.8h, v0.8b, %[b2].8b\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "uadalp	v31.4s, v15.8h\n"
+                "umull	v15.8h, v0.8b, %[b3].8b\n"
+                "bne	1b\n"
 
                 // Branch target
                 "2:\n"
-                "uadalp    v16.4s, v12.8h\n"
-                "umull2    v12.8h, v0.16b, %[b0].16b\n"
-                "uadalp    v17.4s, v13.8h\n"
-                "umull2    v13.8h, v0.16b, %[b1].16b\n"
-                "uadalp    v18.4s, v14.8h\n"
-                "umull2    v14.8h, v0.16b, %[b2].16b\n"
-                "uadalp    v19.4s, v15.8h\n"
-                "umull2    v15.8h, v0.16b, %[b3].16b\n"
-
-                "uadalp    v16.4s, v12.8h\n"
-                "umull    v12.8h, v1.8b, %[b0].8b\n"
-                "uadalp    v17.4s, v13.8h\n"
-                "umull    v13.8h, v1.8b, %[b1].8b\n"
-                "uadalp    v18.4s, v14.8h\n"
-                "umull    v14.8h, v1.8b, %[b2].8b\n"
-                "uadalp    v19.4s, v15.8h\n"
-                "umull    v15.8h, v1.8b, %[b3].8b\n"
-
-                "uadalp    v20.4s, v12.8h\n"
-                "umull2    v12.8h, v1.16b, %[b0].16b\n"
-                "uadalp    v21.4s, v13.8h\n"
-                "umull2    v13.8h, v1.16b, %[b1].16b\n"
-                "uadalp    v22.4s, v14.8h\n"
-                "umull2    v14.8h, v1.16b, %[b2].16b\n"
-                "uadalp    v23.4s, v15.8h\n"
-                "umull2    v15.8h, v1.16b, %[b3].16b\n"
-
-                "uadalp    v20.4s, v12.8h\n"
-                "umull    v12.8h, v2.8b, %[b0].8b\n"
-                "uadalp    v21.4s, v13.8h\n"
-                "umull    v13.8h, v2.8b, %[b1].8b\n"
-                "uadalp    v22.4s, v14.8h\n"
-                "umull    v14.8h, v2.8b, %[b2].8b\n"
-                "uadalp    v23.4s, v15.8h\n"
-                "umull    v15.8h, v2.8b, %[b3].8b\n"
-
-                "uadalp    v24.4s, v12.8h\n"
-                "umull2    v12.8h, v2.16b, %[b0].16b\n"
-                "uadalp    v25.4s, v13.8h\n"
-                "umull2    v13.8h, v2.16b, %[b1].16b\n"
-                "uadalp    v26.4s, v14.8h\n"
-                "umull2    v14.8h, v2.16b, %[b2].16b\n"
-                "uadalp    v27.4s, v15.8h\n"
-                "umull2    v15.8h, v2.16b, %[b3].16b\n"
-
-                "uadalp    v24.4s, v12.8h\n"
-                "umull    v12.8h, v3.8b, %[b0].8b\n"
-                "uadalp    v25.4s, v13.8h\n"
-                "umull    v13.8h, v3.8b, %[b1].8b\n"
-                "uadalp    v26.4s, v14.8h\n"
-                "umull    v14.8h, v3.8b, %[b2].8b\n"
-                "uadalp    v27.4s, v15.8h\n"
-                "umull    v15.8h, v3.8b, %[b3].8b\n"
-
-                "uadalp    v28.4s, v12.8h\n"
-                "umull2    v12.8h, v3.16b, %[b0].16b\n"
-                "uadalp    v29.4s, v13.8h\n"
-                "umull2    v13.8h, v3.16b, %[b1].16b\n"
-                "uadalp    v30.4s, v14.8h\n"
-                "umull2    v14.8h, v3.16b, %[b2].16b\n"
-                "uadalp    v31.4s, v15.8h\n"
-                "umull2    v15.8h, v3.16b, %[b3].16b\n"
-
-                "uadalp    v28.4s, v12.8h\n"
-                "uadalp    v29.4s, v13.8h\n"
-                "uadalp    v30.4s, v14.8h\n"
-                "uadalp    v31.4s, v15.8h\n"
-
-                "addp    v16.4s, v16.4s, v17.4s\n"
-                "addp    v17.4s, v18.4s, v19.4s\n"
-                "addp    v18.4s, v20.4s, v21.4s\n"
-                "addp    v19.4s, v22.4s, v23.4s\n"
-                "addp    v20.4s, v24.4s, v25.4s\n"
-                "addp    v21.4s, v26.4s, v27.4s\n"
-                "addp    v22.4s, v28.4s, v29.4s\n"
-                "addp    v23.4s, v30.4s, v31.4s\n"
-
-                "addp    v16.4s, v16.4s, v17.4s\n"
-                "addp    v17.4s, v18.4s, v19.4s\n"
-                "addp    v18.4s, v20.4s, v21.4s\n"
-                "addp    v19.4s, v22.4s, v23.4s\n"
-
-                "str    q16, [%[c_ptr]]\n"
-                "str    q17, [%[c_ptr], #16]\n"
-                "str    q18, [%[c_ptr], #32]\n"
-                "str    q19, [%[c_ptr], #48]\n"
-                "add    %[c_ptr], %[c_ptr], #64\n"
+                "uadalp	v16.4s, v12.8h\n"
+                "umull2	v12.8h, v0.16b, %[b0].16b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "umull2	v13.8h, v0.16b, %[b1].16b\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull2	v14.8h, v0.16b, %[b2].16b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "umull2	v15.8h, v0.16b, %[b3].16b\n"
+
+                "uadalp	v16.4s, v12.8h\n"
+                "umull	v12.8h, v1.8b, %[b0].8b\n"
+                "uadalp	v17.4s, v13.8h\n"
+                "umull	v13.8h, v1.8b, %[b1].8b\n"
+                "uadalp	v18.4s, v14.8h\n"
+                "umull	v14.8h, v1.8b, %[b2].8b\n"
+                "uadalp	v19.4s, v15.8h\n"
+                "umull	v15.8h, v1.8b, %[b3].8b\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull2	v12.8h, v1.16b, %[b0].16b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "umull2	v13.8h, v1.16b, %[b1].16b\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull2	v14.8h, v1.16b, %[b2].16b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "umull2	v15.8h, v1.16b, %[b3].16b\n"
+
+                "uadalp	v20.4s, v12.8h\n"
+                "umull	v12.8h, v2.8b, %[b0].8b\n"
+                "uadalp	v21.4s, v13.8h\n"
+                "umull	v13.8h, v2.8b, %[b1].8b\n"
+                "uadalp	v22.4s, v14.8h\n"
+                "umull	v14.8h, v2.8b, %[b2].8b\n"
+                "uadalp	v23.4s, v15.8h\n"
+                "umull	v15.8h, v2.8b, %[b3].8b\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull2	v12.8h, v2.16b, %[b0].16b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "umull2	v13.8h, v2.16b, %[b1].16b\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull2	v14.8h, v2.16b, %[b2].16b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "umull2	v15.8h, v2.16b, %[b3].16b\n"
+
+                "uadalp	v24.4s, v12.8h\n"
+                "umull	v12.8h, v3.8b, %[b0].8b\n"
+                "uadalp	v25.4s, v13.8h\n"
+                "umull	v13.8h, v3.8b, %[b1].8b\n"
+                "uadalp	v26.4s, v14.8h\n"
+                "umull	v14.8h, v3.8b, %[b2].8b\n"
+                "uadalp	v27.4s, v15.8h\n"
+                "umull	v15.8h, v3.8b, %[b3].8b\n"
+
+                "uadalp	v28.4s, v12.8h\n"
+                "umull2	v12.8h, v3.16b, %[b0].16b\n"
+                "uadalp	v29.4s, v13.8h\n"
+                "umull2	v13.8h, v3.16b, %[b1].16b\n"
+                "uadalp	v30.4s, v14.8h\n"
+                "umull2	v14.8h, v3.16b, %[b2].16b\n"
+                "uadalp	v31.4s, v15.8h\n"
+                "umull2	v15.8h, v3.16b, %[b3].16b\n"
+
+                "uadalp	v28.4s, v12.8h\n"
+                "uadalp	v29.4s, v13.8h\n"
+                "uadalp	v30.4s, v14.8h\n"
+                "uadalp	v31.4s, v15.8h\n"
+
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "addp	v19.4s, v22.4s, v23.4s\n"
+                "addp	v20.4s, v24.4s, v25.4s\n"
+                "addp	v21.4s, v26.4s, v27.4s\n"
+                "addp	v22.4s, v28.4s, v29.4s\n"
+                "addp	v23.4s, v30.4s, v31.4s\n"
+
+                "addp	v16.4s, v16.4s, v17.4s\n"
+                "addp	v17.4s, v18.4s, v19.4s\n"
+                "addp	v18.4s, v20.4s, v21.4s\n"
+                "addp	v19.4s, v22.4s, v23.4s\n"
+
+                "str	q16, [%[c_ptr]]\n"
+                "str	q17, [%[c_ptr], #16]\n"
+                "str	q18, [%[c_ptr], #32]\n"
+                "str	q19, [%[c_ptr], #48]\n"
+                "add	%[c_ptr], %[c_ptr], #64\n"
 
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp
index f4e43407df..37f6dde3ab 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp
@@ -37,8 +37,7 @@ void a64_gemm_u8_8x12_x1(const uint8_t *, const uint8_t *, uint32_t *, int, int,
 
 class cls_a64_gemm_u8_8x12 {
 public:
-    typedef uint8_t lhs_operand_type;
-    typedef uint8_t rhs_operand_type;
+    typedef uint8_t operand_type;
     typedef uint32_t result_type;
 
     typedef void (*kern_type)(const uint8_t *, const uint8_t *, uint32_t *, int, int, int);
@@ -67,8 +66,8 @@ public:
     }
 
     // Use the standard fixed sized transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 4> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
 
     template<typename T>
     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp
index 8203db21d0..7d433d1a26 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp
@@ -111,11 +111,11 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
 
                 "1:\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
-                "subs    %w[k], %w[k], #1\n"
+                "subs	%w[k], %w[k], #1\n"
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
 
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 "ins    %[b2].d[1], x20\n"
@@ -123,7 +123,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[a_ptr], #40]\n"
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
 
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 "ins    %[a0a].d[1], x20\n"
@@ -131,7 +131,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[a_ptr], #56]\n"
                 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
                 "ins    %[a1a].d[1], x20\n"
@@ -139,7 +139,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[b_ptr], #56]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 "ins    %[b0].d[1], x20\n"
@@ -155,8 +155,8 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
 
-        // Unroll 1
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
+		// Unroll 1
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x6f85e048 // udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
                 "ins    %[b1].d[1], x20\n"
@@ -164,7 +164,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[b_ptr], #88]\n"
                 ".word 0x6f85e84a // udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
                 ".word 0x6fa5e84b // udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
-                "ldr    %d[a0], [%[a_ptr], #64]\n"
+                "ldr	%d[a0], [%[a_ptr], #64]\n"
 
                 ".word 0x6f86e04c // udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
                 "ins    %[b2].d[1], x20\n"
@@ -172,7 +172,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[a_ptr], #72]\n"
                 ".word 0x6f86e84e // udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
                 ".word 0x6fa6e84f // udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
-                "ldr    %d[a1], [%[a_ptr], #80]\n"
+                "ldr	%d[a1], [%[a_ptr], #80]\n"
 
                 ".word 0x6f85e070 // udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
                 "ins    %[a0].d[1], x20\n"
@@ -180,7 +180,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[a_ptr], #88]\n"
                 ".word 0x6f85e872 // udot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
                 ".word 0x6fa5e873 // udot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #96]\n"
+                "ldr	%d[b0], [%[b_ptr], #96]\n"
 
                 ".word 0x6f86e074 // udot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
                 "ins    %[a1].d[1], x20\n"
@@ -188,7 +188,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[b_ptr], #104]\n"
                 ".word 0x6f86e876 // udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
                 ".word 0x6fa6e877 // udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #112]\n"
+                "ldr	%d[b1], [%[b_ptr], #112]\n"
 
                 ".word 0x6f85e098 // udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
                 "ins    %[b0].d[1], x20\n"
@@ -196,19 +196,19 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[b_ptr], #120]\n"
                 ".word 0x6f85e89a // udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
                 ".word 0x6fa5e89b // udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
 
                 ".word 0x6f86e09c // udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
                 ASM_PREFETCH("[%[b_ptr], #640]")
                 ".word 0x6fa6e09d // udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x6f86e89e // udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
                 "ins    %[b1].d[1], x20\n"
                 ".word 0x6fa6e89f // udot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
                 "ldr    %d[b2], [%[b_ptr], #32]\n"
 
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
-                "b.ne    1b\n"
+                "b.ne	1b\n"
 
                 // Branch here if K=1 or 2.  Do the right thing for odd/even at the end.
                 "4:\n"
@@ -221,7 +221,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
 
                 // Even K continuation
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
 
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 "ins    %[b2].d[1], x20\n"
@@ -230,7 +230,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ASM_PREFETCHW("[%[c_ptr]]")
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
 
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 "ins    %[a0a].d[1], x20\n"
@@ -238,7 +238,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 "ldr    x20, [%[a_ptr], #56]\n"
                 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
                 "ins    %[a1a].d[1], x20\n"
@@ -253,7 +253,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 ASM_PREFETCHW("[%[c_ptr], #128]")
                 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
                 "ins    %[b0].d[1], x20\n"
@@ -262,7 +262,7 @@ void a64_gemm_u8_8x12_a55r1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ASM_PREFETCHW("[%[c_ptr], #192]")
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x6f85e048 // udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
                 "ins    %[b1].d[1], x20\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp
index 956ad9448e..1c5e8472e6 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp
@@ -53,63 +53,63 @@ void a64_gemm_u8_8x12(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
             register uint8x16_t a1a asm("v6");
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.4s, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.4s, #0x0\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "movi    v11.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.4s, #0x0\n"
+                "movi	v8.4s, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.4s, #0x0\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "movi	v11.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v13.4s, #0x0\n"
+                "movi	v13.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v14.4s, #0x0\n"
+                "movi	v14.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.4s, #0x0\n"
+                "movi	v15.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v16.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v17.4s, #0x0\n"
+                "movi	v17.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v18.4s, #0x0\n"
+                "movi	v18.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v19.4s, #0x0\n"
+                "movi	v19.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.4s, #0x0\n"
+                "movi	v20.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "movi    v21.4s, #0x0\n"
+                "movi	v21.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "movi    v22.4s, #0x0\n"
-                "movi    v23.4s, #0x0\n"
-                "movi    v24.4s, #0x0\n"
-                "movi    v25.4s, #0x0\n"
-                "movi    v26.4s, #0x0\n"
-                "movi    v27.4s, #0x0\n"
-                "movi    v28.4s, #0x0\n"
-                "movi    v29.4s, #0x0\n"
-                "movi    v30.4s, #0x0\n"
-                "movi    v31.4s, #0x0\n"
+                "movi	v22.4s, #0x0\n"
+                "movi	v23.4s, #0x0\n"
+                "movi	v24.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 // Loop proper
                 "1:\n"
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
 
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0a], [%[a_ptr], #32]\n"
+                "ldr	%q[a0a], [%[a_ptr], #32]\n"
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "ldr    %q[a1a], [%[a_ptr], #48]\n"
+                "ldr	%q[a1a], [%[a_ptr], #48]\n"
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -120,7 +120,7 @@ void a64_gemm_u8_8x12(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
                 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
@@ -131,19 +131,19 @@ void a64_gemm_u8_8x12(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
                 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x6f85e048 // udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
                 ".word 0x6fa5e049 // udot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
-                "ldr    %q[a0], [%[a_ptr], #64]\n"
+                "ldr	%q[a0], [%[a_ptr], #64]\n"
                 ".word 0x6f85e84a // udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
                 ".word 0x6fa5e84b // udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
                 ".word 0x6f86e04c // udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
-                "ldr    %q[a1], [%[a_ptr], #80]\n"
+                "ldr	%q[a1], [%[a_ptr], #80]\n"
                 ".word 0x6fa6e04d // udot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
                 ".word 0x6f86e84e // udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
                 ".word 0x6fa6e84f // udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #96]\n"
+                "ldr	%q[b0], [%[b_ptr], #96]\n"
 
                 ".word 0x6f85e070 // udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
                 ".word 0x6fa5e071 // udot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
@@ -154,40 +154,40 @@ void a64_gemm_u8_8x12(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
                 ".word 0x6fa6e075 // udot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
                 ".word 0x6f86e876 // udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
                 ".word 0x6fa6e877 // udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #112]\n"
+                "ldr	%q[b1], [%[b_ptr], #112]\n"
 
                 ".word 0x6f85e098 // udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
                 ".word 0x6fa5e099 // udot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
                 ".word 0x6f85e89a // udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
                 ".word 0x6fa5e89b // udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x6f86e09c // udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
                 ".word 0x6fa6e09d // udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
-                "subs    %w[k], %w[k], #1\n"
+                "subs	%w[k], %w[k], #1\n"
                 ".word 0x6f86e89e // udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
                 ".word 0x6fa6e89f // udot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
-                "bne    1b\n"
+                "bne	1b\n"
 
                 // Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
                 "4:\n"
 
                 // Branch to alternative tail for odd K
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration (even K)
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0a], [%[a_ptr], #32]\n"
+                "ldr	%q[a0a], [%[a_ptr], #32]\n"
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "ldr    %q[a1a], [%[a_ptr], #48]\n"
+                "ldr	%q[a1a], [%[a_ptr], #48]\n"
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -197,140 +197,140 @@ void a64_gemm_u8_8x12(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cp
                 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
                 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
                 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x6f85e048 // udot v8.4s , %[b0].16b, %[a0a].4b[0]\n"
 
                 ".word 0x6f85e070 // udot v16.4s, %[b1].16b, %[a0a].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x6fa5e049 // udot v9.4s , %[b0].16b, %[a0a].4b[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
+                "str	q8, [%[c_ptr], #0]\n"
                 ".word 0x6fa5e071 // udot v17.4s, %[b1].16b, %[a0a].4b[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
+                "str	q16, [%[c_ptr], #16]\n"
                 ".word 0x6f85e098 // udot v24.4s, %[b2].16b, %[a0a].4b[0]\n"
-                "str    q24, [%[c_ptr], #32]\n"
+                "str	q24, [%[c_ptr], #32]\n"
 
                 ".word 0x6fa5e099 // udot v25.4s, %[b2].16b, %[a0a].4b[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
+                "str	q9, [%[c_ptr], #48]\n"
                 ".word 0x6f85e84a // udot v10.4s, %[b0].16b, %[a0a].4b[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
+                "str	q17, [%[c_ptr], #64]\n"
                 ".word 0x6f85e872 // udot v18.4s, %[b1].16b, %[a0a].4b[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
+                "str	q25, [%[c_ptr], #80]\n"
                 ".word 0x6f85e89a // udot v26.4s, %[b2].16b, %[a0a].4b[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
+                "str	q10, [%[c_ptr], #96]\n"
 
                 ".word 0x6fa5e84b // udot v11.4s, %[b0].16b, %[a0a].4b[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
+                "str	q18, [%[c_ptr], #112]\n"
                 ".word 0x6fa5e873 // udot v19.4s, %[b1].16b, %[a0a].4b[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
+                "str	q26, [%[c_ptr], #128]\n"
                 ".word 0x6fa5e89b // udot v27.4s, %[b2].16b, %[a0a].4b[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
+                "str	q11, [%[c_ptr], #144]\n"
 
                 ".word 0x6f86e04c // udot v12.4s, %[b0].16b, %[a1a].4b[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
+                "str	q19, [%[c_ptr], #160]\n"
                 ".word 0x6f86e074 // udot v20.4s, %[b1].16b, %[a1a].4b[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
+                "str	q27, [%[c_ptr], #176]\n"
                 ".word 0x6f86e09c // udot v28.4s, %[b2].16b, %[a1a].4b[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 ".word 0x6fa6e04d // udot v13.4s, %[b0].16b, %[a1a].4b[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
+                "str	q20, [%[c_ptr], #208]\n"
                 ".word 0x6fa6e075 // udot v21.4s, %[b1].16b, %[a1a].4b[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
+                "str	q28, [%[c_ptr], #224]\n"
                 ".word 0x6fa6e09d // udot v29.4s, %[b2].16b, %[a1a].4b[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
+                "str	q13, [%[c_ptr], #240]\n"
 
                 ".word 0x6f86e84e // udot v14.4s, %[b0].16b, %[a1a].4b[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
+                "str	q21, [%[c_ptr], #256]\n"
                 ".word 0x6f86e876 // udot v22.4s, %[b1].16b, %[a1a].4b[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
+                "str	q29, [%[c_ptr], #272]\n"
                 ".word 0x6f86e89e // udot v30.4s, %[b2].16b, %[a1a].4b[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
+                "str	q14, [%[c_ptr], #288]\n"
 
                 ".word 0x6fa6e84f // udot v15.4s, %[b0].16b, %[a1a].4b[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
+                "str	q22, [%[c_ptr], #304]\n"
                 ".word 0x6fa6e877 // udot v23.4s, %[b1].16b, %[a1a].4b[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
+                "str	q30, [%[c_ptr], #320]\n"
                 ".word 0x6fa6e89f // udot v31.4s, %[b2].16b, %[a1a].4b[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
-                "b    3f\n"
+                "b	3f\n"
 
                 // Detached final iteration (odd K)
                 "2:\n"
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
+                "str	q8, [%[c_ptr], #0]\n"
                 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
+                "str	q16, [%[c_ptr], #16]\n"
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "str    q24, [%[c_ptr], #32]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "str	q24, [%[c_ptr], #32]\n"
                 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
+                "str	q9, [%[c_ptr], #48]\n"
 
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
+                "str	q17, [%[c_ptr], #64]\n"
                 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
+                "str	q25, [%[c_ptr], #80]\n"
                 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
+                "str	q10, [%[c_ptr], #96]\n"
 
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
+                "str	q18, [%[c_ptr], #112]\n"
                 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
+                "str	q26, [%[c_ptr], #128]\n"
                 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
+                "str	q11, [%[c_ptr], #144]\n"
 
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
+                "str	q19, [%[c_ptr], #160]\n"
                 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
+                "str	q27, [%[c_ptr], #176]\n"
                 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
+                "str	q20, [%[c_ptr], #208]\n"
                 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
+                "str	q28, [%[c_ptr], #224]\n"
                 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
+                "str	q13, [%[c_ptr], #240]\n"
 
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
+                "str	q21, [%[c_ptr], #256]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
+                "str	q29, [%[c_ptr], #272]\n"
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
+                "str	q14, [%[c_ptr], #288]\n"
 
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
+                "str	q22, [%[c_ptr], #304]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
+                "str	q30, [%[c_ptr], #320]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
 
                 // Common tail
                 "3:\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
 
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp
index 182d246d7a..63581a6008 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp
@@ -52,60 +52,60 @@ void a64_gemm_u8_8x12_x1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.4s, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.4s, #0x0\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "movi    v11.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.4s, #0x0\n"
+                "movi	v8.4s, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.4s, #0x0\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "movi	v11.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v13.4s, #0x0\n"
+                "movi	v13.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v14.4s, #0x0\n"
+                "movi	v14.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.4s, #0x0\n"
+                "movi	v15.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v16.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v17.4s, #0x0\n"
+                "movi	v17.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v18.4s, #0x0\n"
+                "movi	v18.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v19.4s, #0x0\n"
+                "movi	v19.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.4s, #0x0\n"
+                "movi	v20.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "movi    v21.4s, #0x0\n"
+                "movi	v21.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "movi    v22.4s, #0x0\n"
-                "movi    v23.4s, #0x0\n"
-                "movi    v24.4s, #0x0\n"
-                "movi    v25.4s, #0x0\n"
-                "movi    v26.4s, #0x0\n"
-                "movi    v27.4s, #0x0\n"
-                "movi    v28.4s, #0x0\n"
-                "movi    v29.4s, #0x0\n"
-                "movi    v30.4s, #0x0\n"
-                "movi    v31.4s, #0x0\n"
+                "movi	v22.4s, #0x0\n"
+                "movi	v23.4s, #0x0\n"
+                "movi	v24.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 // Loop proper
                 "1:\n"
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -116,74 +116,74 @@ void a64_gemm_u8_8x12_x1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t
                 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #448]")
                 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0], [%[a_ptr], #32]\n"
+                "ldr	%q[a0], [%[a_ptr], #32]\n"
                 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #48]\n"
+                "ldr	%q[a1], [%[a_ptr], #48]\n"
 
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #96]\n"
+                "ldr	%q[b0], [%[b_ptr], #96]\n"
 
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #512]")
                 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "subs    %w[k], %w[k], #1\n"
+                "subs	%w[k], %w[k], #1\n"
                 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #112]\n"
+                "ldr	%q[b1], [%[b_ptr], #112]\n"
 
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
                 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
                 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "bne    1b\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "bne	1b\n"
 
                 // Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
                 "4:\n"
 
                 // Branch to alternative tail for odd K
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration (even K)
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
 
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -193,142 +193,142 @@ void a64_gemm_u8_8x12_x1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t
                 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
 
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
                 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
                 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "ldr    %q[a0], [%[a_ptr], #-32]\n"
+                "ldr	%q[a0], [%[a_ptr], #-32]\n"
                 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
                 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #-16]\n"
+                "ldr	%q[a1], [%[a_ptr], #-16]\n"
 
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
 
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
+                "str	q8, [%[c_ptr], #0]\n"
                 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
+                "str	q16, [%[c_ptr], #16]\n"
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
-                "str    q24, [%[c_ptr], #32]\n"
+                "str	q24, [%[c_ptr], #32]\n"
 
                 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
+                "str	q9, [%[c_ptr], #48]\n"
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
+                "str	q17, [%[c_ptr], #64]\n"
                 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
+                "str	q25, [%[c_ptr], #80]\n"
                 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
+                "str	q10, [%[c_ptr], #96]\n"
 
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
+                "str	q18, [%[c_ptr], #112]\n"
                 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
+                "str	q26, [%[c_ptr], #128]\n"
                 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
+                "str	q11, [%[c_ptr], #144]\n"
 
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
+                "str	q19, [%[c_ptr], #160]\n"
                 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
+                "str	q27, [%[c_ptr], #176]\n"
                 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
+                "str	q20, [%[c_ptr], #208]\n"
                 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
+                "str	q28, [%[c_ptr], #224]\n"
                 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
+                "str	q13, [%[c_ptr], #240]\n"
 
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
+                "str	q21, [%[c_ptr], #256]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
+                "str	q29, [%[c_ptr], #272]\n"
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
+                "str	q14, [%[c_ptr], #288]\n"
 
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
+                "str	q22, [%[c_ptr], #304]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
+                "str	q30, [%[c_ptr], #320]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
-                "b    3f\n"
+                "b	3f\n"
 
                 // Detached final iteration (odd K)
                 "2:\n"
                 ".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
                 ".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
                 ".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
+                "str	q8, [%[c_ptr], #0]\n"
                 ".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
+                "str	q16, [%[c_ptr], #16]\n"
                 ".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "str    q24, [%[c_ptr], #32]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "str	q24, [%[c_ptr], #32]\n"
                 ".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
+                "str	q9, [%[c_ptr], #48]\n"
 
                 ".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
+                "str	q17, [%[c_ptr], #64]\n"
                 ".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
+                "str	q25, [%[c_ptr], #80]\n"
                 ".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
+                "str	q10, [%[c_ptr], #96]\n"
 
                 ".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
+                "str	q18, [%[c_ptr], #112]\n"
                 ".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
+                "str	q26, [%[c_ptr], #128]\n"
                 ".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
+                "str	q11, [%[c_ptr], #144]\n"
 
                 ".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
+                "str	q19, [%[c_ptr], #160]\n"
                 ".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
+                "str	q27, [%[c_ptr], #176]\n"
                 ".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 ".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
+                "str	q20, [%[c_ptr], #208]\n"
                 ".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
+                "str	q28, [%[c_ptr], #224]\n"
                 ".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
+                "str	q13, [%[c_ptr], #240]\n"
 
                 ".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
+                "str	q21, [%[c_ptr], #256]\n"
                 ".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
+                "str	q29, [%[c_ptr], #272]\n"
                 ".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
+                "str	q14, [%[c_ptr], #288]\n"
 
                 ".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
+                "str	q22, [%[c_ptr], #304]\n"
                 ".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
+                "str	q30, [%[c_ptr], #320]\n"
                 ".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
 
                 // Common tail
                 "3:\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
 
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24.hpp
index 6bc40b4ac8..d9668aae02 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24.hpp
@@ -41,8 +41,7 @@ void a64_hgemm_asimd_8x24_x1(const __fp16 *, const __fp16 *, __fp16 *, int, int,
 // the constructor to pick a kernel implementation).
 class cls_a64_hgemm_8x24 {
 public:
-    typedef __fp16 lhs_operand_type;
-    typedef __fp16 rhs_operand_type;
+    typedef __fp16 operand_type;
     typedef __fp16 result_type;
 
     typedef void (*kern_type)(const __fp16 *, const __fp16 *, __fp16 *, int, int, int);
@@ -61,7 +60,7 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 24> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 24> transforms = {};
 
     template<typename T>
     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp
index 742b406438..0686589f5c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp
@@ -22,8 +22,8 @@
  * SOFTWARE.
  */
 
-// Build on AArch64 where either ENABLE_FP16_KERNELS is set or FP16 is explicitly supported.
-#if defined(__aarch64__) && (defined(ENABLE_FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
+// Build on AArch64 where either FP16_KERNELS is set or FP16 is explicitly supported.
+#if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
 
 #include <arm_neon.h>
 
@@ -72,311 +72,311 @@ void a64_hgemm_asimd_8x24_a55r1(const __fp16 *Apanel, const __fp16 *Bpanel, __fp
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.8h, #0x0\n"
-                "ldr    %d[a0], [%[a_ptr]]\n"
-                "movi    v9.8h, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.8h, #0x0\n"
-                "ldr    %d[a1], [%[a_ptr], #8]\n"
-                "movi    v11.8h, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.8h, #0x0\n"
-                "movi    v13.8h, #0x0\n"
+                "movi	v8.8h, #0x0\n"
+                "ldr	%d[a0], [%[a_ptr]]\n"
+                "movi	v9.8h, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.8h, #0x0\n"
+                "ldr	%d[a1], [%[a_ptr], #8]\n"
+                "movi	v11.8h, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.8h, #0x0\n"
+                "movi	v13.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v14.8h, #0x0\n"
-                "movi    v15.8h, #0x0\n"
+                "movi	v14.8h, #0x0\n"
+                "movi	v15.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v16.8h, #0x0\n"
-                "movi    v17.8h, #0x0\n"
+                "movi	v16.8h, #0x0\n"
+                "movi	v17.8h, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v18.8h, #0x0\n"
-                "movi    v19.8h, #0x0\n"
+                "movi	v18.8h, #0x0\n"
+                "movi	v19.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v20.8h, #0x0\n"
-                "movi    v21.8h, #0x0\n"
+                "movi	v20.8h, #0x0\n"
+                "movi	v21.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v22.8h, #0x0\n"
-                "movi    v23.8h, #0x0\n"
+                "movi	v22.8h, #0x0\n"
+                "movi	v23.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v24.8h, #0x0\n"
-                "movi    v25.8h, #0x0\n"
-                "movi    v26.8h, #0x0\n"
-                "movi    v27.8h, #0x0\n"
-                "movi    v28.8h, #0x0\n"
-                "movi    v29.8h, #0x0\n"
-                "movi    v30.8h, #0x0\n"
-                "movi    v31.8h, #0x0\n"
+                "movi	v24.8h, #0x0\n"
+                "movi	v25.8h, #0x0\n"
+                "movi	v26.8h, #0x0\n"
+                "movi	v27.8h, #0x0\n"
+                "movi	v28.8h, #0x0\n"
+                "movi	v29.8h, #0x0\n"
+                "movi	v30.8h, #0x0\n"
+                "movi	v31.8h, #0x0\n"
 
                 // The loop is offset by these two instructions which must
                 // always be executed.
-                "fmla     v8.8h , %[b0].8h, %[a0].h[0]\n"
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0].h[0]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 "1:\n"
-                "fmla      v9.8h , %[b0].8h, %[a0].h[1]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0].h[2]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
-                "ldr    %d[a0a], [%[a_ptr], #16]\n"
-
-                "fmla     v12.8h, %[b0].8h, %[a1].h[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.8h, %[b0].8h, %[a1].h[1]\n"
-                "fmla    v14.8h, %[b0].8h, %[a1].h[2]\n"
-                "fmla    v15.8h, %[b0].8h, %[a1].h[3]\n"
-                "ldr    %d[a1a], [%[a_ptr], #24]\n"
-
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v20.8h, %[b1].8h, %[a1].h[0]\n"
-                "fmla    v21.8h, %[b1].8h, %[a1].h[1]\n"
-                "ldr    x20, [%[b_ptr], #56]\n"
-                "fmla    v22.8h, %[b1].8h, %[a1].h[2]\n"
-                "fmla    v23.8h, %[b1].8h, %[a1].h[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
-
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
-                "ldr    x20, [%[b_ptr], #72]\n"
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "fmla  	v9.8h , %[b0].8h, %[a0].h[1]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0].h[2]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "ldr	%d[a0a], [%[a_ptr], #16]\n"
+
+                "fmla 	v12.8h, %[b0].8h, %[a1].h[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.8h, %[b0].8h, %[a1].h[1]\n"
+                "fmla	v14.8h, %[b0].8h, %[a1].h[2]\n"
+                "fmla	v15.8h, %[b0].8h, %[a1].h[3]\n"
+                "ldr	%d[a1a], [%[a_ptr], #24]\n"
+
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v20.8h, %[b1].8h, %[a1].h[0]\n"
+                "fmla	v21.8h, %[b1].8h, %[a1].h[1]\n"
+                "ldr	x20, [%[b_ptr], #56]\n"
+                "fmla	v22.8h, %[b1].8h, %[a1].h[2]\n"
+                "fmla	v23.8h, %[b1].8h, %[a1].h[3]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
+
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "ldr	x20, [%[b_ptr], #72]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
 
-                "fmla    v28.8h, %[b2].8h, %[a1].h[0]\n"
-                "fmla    v29.8h, %[b2].8h, %[a1].h[1]\n"
+                "fmla	v28.8h, %[b2].8h, %[a1].h[0]\n"
+                "fmla	v29.8h, %[b2].8h, %[a1].h[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "fmla    v30.8h, %[b2].8h, %[a1].h[2]\n"
-                "fmla    v31.8h, %[b2].8h, %[a1].h[3]\n"
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
+                "fmla	v30.8h, %[b2].8h, %[a1].h[2]\n"
+                "fmla	v31.8h, %[b2].8h, %[a1].h[3]\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
 
                 // Unroll 1
-                "fmla     v8.8h , %[b0].8h, %[a0a].h[0]\n"
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v9.8h , %[b0].8h, %[a0a].h[1]\n"
-                "ldr    x20, [%[b_ptr], #88]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0a].h[2]\n"
-                "fmla    v11.8h, %[b0].8h, %[a0a].h[3]\n"
-                "ldr    %d[a0], [%[a_ptr], #32]\n"
-
-                "fmla     v12.8h, %[b0].8h, %[a1a].h[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.8h, %[b0].8h, %[a1a].h[1]\n"
-                "fmla    v14.8h, %[b0].8h, %[a1a].h[2]\n"
-                "fmla    v15.8h, %[b0].8h, %[a1a].h[3]\n"
-                "ldr    %d[a1], [%[a_ptr], #40]\n"
-
-                "fmla    v16.8h, %[b1].8h, %[a0a].h[0]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v17.8h, %[b1].8h, %[a0a].h[1]\n"
-                "fmla    v18.8h, %[b1].8h, %[a0a].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0a].h[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #96]\n"
-
-                "fmla    v20.8h, %[b1].8h, %[a1a].h[0]\n"
-                "fmla    v21.8h, %[b1].8h, %[a1a].h[1]\n"
-                "ldr    x20, [%[b_ptr], #104]\n"
-                "fmla    v22.8h, %[b1].8h, %[a1a].h[2]\n"
-                "fmla    v23.8h, %[b1].8h, %[a1a].h[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #112]\n"
-
-                "fmla    v24.8h, %[b2].8h, %[a0a].h[0]\n"
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v25.8h, %[b2].8h, %[a0a].h[1]\n"
-                "ldr    x20, [%[b_ptr], #120]\n"
-                "fmla    v26.8h, %[b2].8h, %[a0a].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0a].h[3]\n"
-
-                "fmla    v28.8h, %[b2].8h, %[a1a].h[0]\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0a].h[0]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v9.8h , %[b0].8h, %[a0a].h[1]\n"
+                "ldr	x20, [%[b_ptr], #88]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0a].h[2]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0a].h[3]\n"
+                "ldr	%d[a0], [%[a_ptr], #32]\n"
+
+                "fmla 	v12.8h, %[b0].8h, %[a1a].h[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.8h, %[b0].8h, %[a1a].h[1]\n"
+                "fmla	v14.8h, %[b0].8h, %[a1a].h[2]\n"
+                "fmla	v15.8h, %[b0].8h, %[a1a].h[3]\n"
+                "ldr	%d[a1], [%[a_ptr], #40]\n"
+
+                "fmla	v16.8h, %[b1].8h, %[a0a].h[0]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v17.8h, %[b1].8h, %[a0a].h[1]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0a].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0a].h[3]\n"
+                "ldr	%d[b0], [%[b_ptr], #96]\n"
+
+                "fmla	v20.8h, %[b1].8h, %[a1a].h[0]\n"
+                "fmla	v21.8h, %[b1].8h, %[a1a].h[1]\n"
+                "ldr	x20, [%[b_ptr], #104]\n"
+                "fmla	v22.8h, %[b1].8h, %[a1a].h[2]\n"
+                "fmla	v23.8h, %[b1].8h, %[a1a].h[3]\n"
+                "ldr	%d[b1], [%[b_ptr], #112]\n"
+
+                "fmla	v24.8h, %[b2].8h, %[a0a].h[0]\n"
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v25.8h, %[b2].8h, %[a0a].h[1]\n"
+                "ldr	x20, [%[b_ptr], #120]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0a].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0a].h[3]\n"
+
+                "fmla	v28.8h, %[b2].8h, %[a1a].h[0]\n"
                 ASM_PREFETCH("[%[b_ptr], #448]")
-                "fmla    v29.8h, %[b2].8h, %[a1a].h[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v30.8h, %[b2].8h, %[a1a].h[2]\n"
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v31.8h, %[b2].8h, %[a1a].h[3]\n"
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
+                "fmla	v29.8h, %[b2].8h, %[a1a].h[1]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v30.8h, %[b2].8h, %[a1a].h[2]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v31.8h, %[b2].8h, %[a1a].h[3]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
 
-                "fmla     v8.8h , %[b0].8h, %[a0].h[0]\n"
-                "bne    1b\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0].h[0]\n"
+                "bne	1b\n"
 
                 "4:\n"
 
                 // Start final iteration - branch off to "odd" code before we load a0a
-                "fmla      v9.8h , %[b0].8h, %[a0].h[1]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0].h[2]\n"
-                "cbnz    %w[oddk], 2f\n"
+                "fmla  	v9.8h , %[b0].8h, %[a0].h[1]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0].h[2]\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Even K continuation
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
-                "ldr    %d[a0a], [%[a_ptr], #16]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "ldr	%d[a0a], [%[a_ptr], #16]\n"
 
-                "fmla     v12.8h, %[b0].8h, %[a1].h[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.8h, %[b0].8h, %[a1].h[1]\n"
+                "fmla 	v12.8h, %[b0].8h, %[a1].h[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.8h, %[b0].8h, %[a1].h[1]\n"
                 ASM_PREFETCHW("[%[c_ptr]]")
-                "fmla    v14.8h, %[b0].8h, %[a1].h[2]\n"
-                "fmla    v15.8h, %[b0].8h, %[a1].h[3]\n"
-                "ldr    %d[a1a], [%[a_ptr], #24]\n"
+                "fmla	v14.8h, %[b0].8h, %[a1].h[2]\n"
+                "fmla	v15.8h, %[b0].8h, %[a1].h[3]\n"
+                "ldr	%d[a1a], [%[a_ptr], #24]\n"
 
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #64]")
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v20.8h, %[b1].8h, %[a1].h[0]\n"
-                "fmla    v21.8h, %[b1].8h, %[a1].h[1]\n"
-                "ldr    x20, [%[b_ptr], #56]\n"
-                "fmla    v22.8h, %[b1].8h, %[a1].h[2]\n"
-                "fmla    v23.8h, %[b1].8h, %[a1].h[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
-
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
-                "ldr    x20, [%[b_ptr], #72]\n"
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v20.8h, %[b1].8h, %[a1].h[0]\n"
+                "fmla	v21.8h, %[b1].8h, %[a1].h[1]\n"
+                "ldr	x20, [%[b_ptr], #56]\n"
+                "fmla	v22.8h, %[b1].8h, %[a1].h[2]\n"
+                "fmla	v23.8h, %[b1].8h, %[a1].h[3]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
+
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "ldr	x20, [%[b_ptr], #72]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
                 ASM_PREFETCHW("[%[c_ptr], #128]")
 
-                "fmla    v28.8h, %[b2].8h, %[a1].h[0]\n"
-                "fmla    v29.8h, %[b2].8h, %[a1].h[1]\n"
+                "fmla	v28.8h, %[b2].8h, %[a1].h[0]\n"
+                "fmla	v29.8h, %[b2].8h, %[a1].h[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #192]")
-                "fmla    v30.8h, %[b2].8h, %[a1].h[2]\n"
-                "fmla    v31.8h, %[b2].8h, %[a1].h[3]\n"
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
-
-                "fmla     v8.8h , %[b0].8h, %[a0a].h[0]\n"
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v9.8h , %[b0].8h, %[a0a].h[1]\n"
-                "ldr    x20, [%[b_ptr], #88]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0a].h[2]\n"
-                "fmla    v11.8h, %[b0].8h, %[a0a].h[3]\n"
+                "fmla	v30.8h, %[b2].8h, %[a1].h[2]\n"
+                "fmla	v31.8h, %[b2].8h, %[a1].h[3]\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
+
+                "fmla 	v8.8h , %[b0].8h, %[a0a].h[0]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v9.8h , %[b0].8h, %[a0a].h[1]\n"
+                "ldr	x20, [%[b_ptr], #88]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0a].h[2]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0a].h[3]\n"
                 ASM_PREFETCHW("[%[c_ptr], #256]")
 
-                "fmla     v12.8h, %[b0].8h, %[a1a].h[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.8h, %[b0].8h, %[a1a].h[1]\n"
+                "fmla 	v12.8h, %[b0].8h, %[a1a].h[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.8h, %[b0].8h, %[a1a].h[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #320]")
-                "fmla    v14.8h, %[b0].8h, %[a1a].h[2]\n"
-                "fmla    v15.8h, %[b0].8h, %[a1a].h[3]\n"
-                "ldr    %d[a1], [%[a_ptr], #40]\n"
+                "fmla	v14.8h, %[b0].8h, %[a1a].h[2]\n"
+                "fmla	v15.8h, %[b0].8h, %[a1a].h[3]\n"
+                "ldr	%d[a1], [%[a_ptr], #40]\n"
 
-                "fmla    v16.8h, %[b1].8h, %[a0a].h[0]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v17.8h, %[b1].8h, %[a0a].h[1]\n"
+                "fmla	v16.8h, %[b1].8h, %[a0a].h[0]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v17.8h, %[b1].8h, %[a0a].h[1]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #384]")
-                "fmla    v18.8h, %[b1].8h, %[a0a].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0a].h[3]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0a].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0a].h[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #448]")
 
-                "fmla    v20.8h, %[b1].8h, %[a1a].h[0]\n"
-                "fmla    v21.8h, %[b1].8h, %[a1a].h[1]\n"
+                "fmla	v20.8h, %[b1].8h, %[a1a].h[0]\n"
+                "fmla	v21.8h, %[b1].8h, %[a1a].h[1]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #512]")
-                "fmla    v22.8h, %[b1].8h, %[a1a].h[2]\n"
-                "fmla    v23.8h, %[b1].8h, %[a1a].h[3]\n"
+                "fmla	v22.8h, %[b1].8h, %[a1a].h[2]\n"
+                "fmla	v23.8h, %[b1].8h, %[a1a].h[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #576]")
 
-                "fmla    v24.8h, %[b2].8h, %[a0a].h[0]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0a].h[1]\n"
+                "fmla	v24.8h, %[b2].8h, %[a0a].h[0]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0a].h[1]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #640]")
-                "fmla    v26.8h, %[b2].8h, %[a0a].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0a].h[3]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0a].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0a].h[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #704]")
 
-                "fmla    v28.8h, %[b2].8h, %[a1a].h[0]\n"
-                "fmla    v29.8h, %[b2].8h, %[a1a].h[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v30.8h, %[b2].8h, %[a1a].h[2]\n"
-                "fmla    v31.8h, %[b2].8h, %[a1a].h[3]\n"
-                "b    3f\n"
+                "fmla	v28.8h, %[b2].8h, %[a1a].h[0]\n"
+                "fmla	v29.8h, %[b2].8h, %[a1a].h[1]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v30.8h, %[b2].8h, %[a1a].h[2]\n"
+                "fmla	v31.8h, %[b2].8h, %[a1a].h[3]\n"
+                "b	3f\n"
 
                 "2:\n"
 
                 // Odd tail
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
                 ASM_PREFETCHW("[%[c_ptr]]")
 
-                "fmla     v12.8h, %[b0].8h, %[a1].h[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.8h, %[b0].8h, %[a1].h[1]\n"
+                "fmla 	v12.8h, %[b0].8h, %[a1].h[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.8h, %[b0].8h, %[a1].h[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #64]")
-                "fmla    v14.8h, %[b0].8h, %[a1].h[2]\n"
-                "add    %[a_ptr], %[a_ptr], #16\n"
-                "fmla    v15.8h, %[b0].8h, %[a1].h[3]\n"
+                "fmla	v14.8h, %[b0].8h, %[a1].h[2]\n"
+                "add	%[a_ptr], %[a_ptr], #16\n"
+                "fmla	v15.8h, %[b0].8h, %[a1].h[3]\n"
                 ASM_PREFETCHW("[%[c_ptr], #128]")
 
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #192]")
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
                 ASM_PREFETCHW("[%[c_ptr], #256]")
 
-                "fmla    v20.8h, %[b1].8h, %[a1].h[0]\n"
-                "fmla    v21.8h, %[b1].8h, %[a1].h[1]\n"
+                "fmla	v20.8h, %[b1].8h, %[a1].h[0]\n"
+                "fmla	v21.8h, %[b1].8h, %[a1].h[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #320]")
-                "fmla    v22.8h, %[b1].8h, %[a1].h[2]\n"
-                "fmla    v23.8h, %[b1].8h, %[a1].h[3]\n"
+                "fmla	v22.8h, %[b1].8h, %[a1].h[2]\n"
+                "fmla	v23.8h, %[b1].8h, %[a1].h[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #384]")
 
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #384]")
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #448]")
 
-                "fmla    v28.8h, %[b2].8h, %[a1].h[0]\n"
+                "fmla	v28.8h, %[b2].8h, %[a1].h[0]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #512]")
-                "fmla    v29.8h, %[b2].8h, %[a1].h[1]\n"
+                "fmla	v29.8h, %[b2].8h, %[a1].h[1]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #576]")
-                "fmla    v30.8h, %[b2].8h, %[a1].h[2]\n"
+                "fmla	v30.8h, %[b2].8h, %[a1].h[2]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #640]")
-                "fmla    v31.8h, %[b2].8h, %[a1].h[3]\n"
+                "fmla	v31.8h, %[b2].8h, %[a1].h[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #704]")
 
                 // Common tail
                 // A55 won't dual issue these stores with anything else, so
                 // simplest to do them all in this common code.
                 "3:\n"
-                "str    q8,  [%[c_ptr]]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-                "str    q24, [%[c_ptr], #32]\n"
-                "str    q9,  [%[c_ptr], #48]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "str    q12, [%[c_ptr], #192]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "str    q15, [%[c_ptr], #336]\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
+                "str	q8,  [%[c_ptr]]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+                "str	q24, [%[c_ptr], #32]\n"
+                "str	q9,  [%[c_ptr], #48]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "str	q12, [%[c_ptr], #192]\n"
+                "str	q20, [%[c_ptr], #208]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "str	q15, [%[c_ptr], #336]\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
                 "5:\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "=w" (a0), [a0a] "=w" (a0a), [a1] "=w" (a1), [a1a] "=w" (a1a),
@@ -391,4 +391,4 @@ void a64_hgemm_asimd_8x24_a55r1(const __fp16 *Apanel, const __fp16 *Bpanel, __fp
 
 } // namespace arm_gemm
 
-#endif // __aarch64__ && (ENABLE_FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+#endif // __aarch64__ && (FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp
index be97c0d2d7..be92554dce 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp
@@ -22,8 +22,8 @@
  * SOFTWARE.
  */
 
-// Build on AArch64 where either ENABLE_FP16_KERNELS is set or FP16 is explicitly supported.
-#if defined(__aarch64__) && (defined(ENABLE_FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
+// Build on AArch64 where either FP16_KERNELS is set or FP16 is explicitly supported.
+#if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
 
 #include <arm_neon.h>
 
@@ -67,270 +67,270 @@ void a64_hgemm_asimd_8x24(const __fp16 *Apanel, const __fp16 *Bpanel, __fp16 *Cp
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.8h, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.8h, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.8h, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v11.8h, #0x0\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "movi    v12.8h, #0x0\n"
-                "ldr    %q[b0a], [%[b_ptr], #48]\n"
-                "movi    v13.8h, #0x0\n"
-                "ldr    %q[b1a], [%[b_ptr], #64]\n"
-                "movi    v14.8h, #0x0\n"
+                "movi	v8.8h, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.8h, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.8h, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v11.8h, #0x0\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "movi	v12.8h, #0x0\n"
+                "ldr	%q[b0a], [%[b_ptr], #48]\n"
+                "movi	v13.8h, #0x0\n"
+                "ldr	%q[b1a], [%[b_ptr], #64]\n"
+                "movi	v14.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v15.8h, #0x0\n"
+                "movi	v15.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v16.8h, #0x0\n"
+                "movi	v16.8h, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v17.8h, #0x0\n"
+                "movi	v17.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v18.8h, #0x0\n"
+                "movi	v18.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v19.8h, #0x0\n"
+                "movi	v19.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.8h, #0x0\n"
-                "movi    v21.8h, #0x0\n"
-                "movi    v22.8h, #0x0\n"
-                "movi    v23.8h, #0x0\n"
-                "movi    v24.8h, #0x0\n"
-                "movi    v25.8h, #0x0\n"
-                "movi    v26.8h, #0x0\n"
-                "movi    v27.8h, #0x0\n"
-                "movi    v28.8h, #0x0\n"
-                "movi    v29.8h, #0x0\n"
-                "movi    v30.8h, #0x0\n"
-                "movi    v31.8h, #0x0\n"
+                "movi	v20.8h, #0x0\n"
+                "movi	v21.8h, #0x0\n"
+                "movi	v22.8h, #0x0\n"
+                "movi	v23.8h, #0x0\n"
+                "movi	v24.8h, #0x0\n"
+                "movi	v25.8h, #0x0\n"
+                "movi	v26.8h, #0x0\n"
+                "movi	v27.8h, #0x0\n"
+                "movi	v28.8h, #0x0\n"
+                "movi	v29.8h, #0x0\n"
+                "movi	v30.8h, #0x0\n"
+                "movi	v31.8h, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 "1:\n"
-                "fmla     v8.8h , %[b0].8h, %[a0].h[0]\n"
-                "fmla      v9.8h , %[b0].8h, %[a0].h[1]\n"
-                "ldr    %q[a0a], [%[a_ptr], #16]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0].h[2]\n"
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
-                "ldr    %q[b2a], [%[b_ptr], #80]\n"
-                "fmla     v12.8h, %[b0].8h, %[a0].h[4]\n"
-                "fmla    v13.8h, %[b0].8h, %[a0].h[5]\n"
-                "fmla    v14.8h, %[b0].8h, %[a0].h[6]\n"
-                "fmla    v15.8h, %[b0].8h, %[a0].h[7]\n"
-                "ldr    %q[b0], [%[b_ptr], #96]\n"
-
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0].h[0]\n"
+                "fmla  	v9.8h , %[b0].8h, %[a0].h[1]\n"
+                "ldr	%q[a0a], [%[a_ptr], #16]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0].h[2]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "ldr	%q[b2a], [%[b_ptr], #80]\n"
+                "fmla 	v12.8h, %[b0].8h, %[a0].h[4]\n"
+                "fmla	v13.8h, %[b0].8h, %[a0].h[5]\n"
+                "fmla	v14.8h, %[b0].8h, %[a0].h[6]\n"
+                "fmla	v15.8h, %[b0].8h, %[a0].h[7]\n"
+                "ldr	%q[b0], [%[b_ptr], #96]\n"
+
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v20.8h, %[b1].8h, %[a0].h[4]\n"
-                "fmla    v21.8h, %[b1].8h, %[a0].h[5]\n"
-                "fmla    v22.8h, %[b1].8h, %[a0].h[6]\n"
-                "fmla    v23.8h, %[b1].8h, %[a0].h[7]\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v20.8h, %[b1].8h, %[a0].h[4]\n"
+                "fmla	v21.8h, %[b1].8h, %[a0].h[5]\n"
+                "fmla	v22.8h, %[b1].8h, %[a0].h[6]\n"
+                "fmla	v23.8h, %[b1].8h, %[a0].h[7]\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #288]")
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
-                "fmla    v28.8h, %[b2].8h, %[a0].h[4]\n"
-                "fmla    v29.8h, %[b2].8h, %[a0].h[5]\n"
-                "fmla    v30.8h, %[b2].8h, %[a0].h[6]\n"
-                "fmla    v31.8h, %[b2].8h, %[a0].h[7]\n"
-                "ldr    %q[a0], [%[a_ptr], #32]\n"
-
-                "fmla     v8.8h , %[b0a].8h, %[a0a].h[0]\n"
-                "fmla    v9.8h , %[b0a].8h, %[a0a].h[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla    v10.8h, %[b0a].8h, %[a0a].h[2]\n"
-                "fmla    v11.8h, %[b0a].8h, %[a0a].h[3]\n"
-                "fmla     v12.8h, %[b0a].8h, %[a0a].h[4]\n"
-                "fmla    v13.8h, %[b0a].8h, %[a0a].h[5]\n"
-                "fmla    v14.8h, %[b0a].8h, %[a0a].h[6]\n"
-                "fmla    v15.8h, %[b0a].8h, %[a0a].h[7]\n"
-                "ldr    %q[b0a], [%[b_ptr], #48]\n"
-
-                "fmla    v16.8h, %[b1a].8h, %[a0a].h[0]\n"
-                "fmla    v17.8h, %[b1a].8h, %[a0a].h[1]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "fmla	v28.8h, %[b2].8h, %[a0].h[4]\n"
+                "fmla	v29.8h, %[b2].8h, %[a0].h[5]\n"
+                "fmla	v30.8h, %[b2].8h, %[a0].h[6]\n"
+                "fmla	v31.8h, %[b2].8h, %[a0].h[7]\n"
+                "ldr	%q[a0], [%[a_ptr], #32]\n"
+
+                "fmla 	v8.8h , %[b0a].8h, %[a0a].h[0]\n"
+                "fmla	v9.8h , %[b0a].8h, %[a0a].h[1]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla	v10.8h, %[b0a].8h, %[a0a].h[2]\n"
+                "fmla	v11.8h, %[b0a].8h, %[a0a].h[3]\n"
+                "fmla 	v12.8h, %[b0a].8h, %[a0a].h[4]\n"
+                "fmla	v13.8h, %[b0a].8h, %[a0a].h[5]\n"
+                "fmla	v14.8h, %[b0a].8h, %[a0a].h[6]\n"
+                "fmla	v15.8h, %[b0a].8h, %[a0a].h[7]\n"
+                "ldr	%q[b0a], [%[b_ptr], #48]\n"
+
+                "fmla	v16.8h, %[b1a].8h, %[a0a].h[0]\n"
+                "fmla	v17.8h, %[b1a].8h, %[a0a].h[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #352]")
-                "fmla    v18.8h, %[b1a].8h, %[a0a].h[2]\n"
-                "fmla    v19.8h, %[b1a].8h, %[a0a].h[3]\n"
-                "fmla    v20.8h, %[b1a].8h, %[a0a].h[4]\n"
-                "fmla    v21.8h, %[b1a].8h, %[a0a].h[5]\n"
-                "fmla    v22.8h, %[b1a].8h, %[a0a].h[6]\n"
-                "fmla    v23.8h, %[b1a].8h, %[a0a].h[7]\n"
-                "ldr    %q[b1a], [%[b_ptr], #64]\n"
-
-                "fmla    v24.8h, %[b2a].8h, %[a0a].h[0]\n"
-                "fmla    v25.8h, %[b2a].8h, %[a0a].h[1]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v26.8h, %[b2a].8h, %[a0a].h[2]\n"
-                "fmla    v27.8h, %[b2a].8h, %[a0a].h[3]\n"
-                "fmla    v28.8h, %[b2a].8h, %[a0a].h[4]\n"
-                "fmla    v29.8h, %[b2a].8h, %[a0a].h[5]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v30.8h, %[b2a].8h, %[a0a].h[6]\n"
-                "fmla    v31.8h, %[b2a].8h, %[a0a].h[7]\n"
-
-                "bne    1b\n"
+                "fmla	v18.8h, %[b1a].8h, %[a0a].h[2]\n"
+                "fmla	v19.8h, %[b1a].8h, %[a0a].h[3]\n"
+                "fmla	v20.8h, %[b1a].8h, %[a0a].h[4]\n"
+                "fmla	v21.8h, %[b1a].8h, %[a0a].h[5]\n"
+                "fmla	v22.8h, %[b1a].8h, %[a0a].h[6]\n"
+                "fmla	v23.8h, %[b1a].8h, %[a0a].h[7]\n"
+                "ldr	%q[b1a], [%[b_ptr], #64]\n"
+
+                "fmla	v24.8h, %[b2a].8h, %[a0a].h[0]\n"
+                "fmla	v25.8h, %[b2a].8h, %[a0a].h[1]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v26.8h, %[b2a].8h, %[a0a].h[2]\n"
+                "fmla	v27.8h, %[b2a].8h, %[a0a].h[3]\n"
+                "fmla	v28.8h, %[b2a].8h, %[a0a].h[4]\n"
+                "fmla	v29.8h, %[b2a].8h, %[a0a].h[5]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v30.8h, %[b2a].8h, %[a0a].h[6]\n"
+                "fmla	v31.8h, %[b2a].8h, %[a0a].h[7]\n"
+
+                "bne	1b\n"
                 "4:\n"
 
                 // Jump to odd tail if necessary.
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Even tail.
-                "fmla     v8.8h , %[b0].8h, %[a0].h[0]\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0].h[0]\n"
                 "fmla   v9.8h , %[b0].8h, %[a0].h[1]\n"
-                "ldr    %q[a0a], [%[a_ptr], #16]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0].h[2]\n"
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
-                "ldr    %q[b2a], [%[b_ptr], #80]\n"
-                "fmla     v12.8h, %[b0].8h, %[a0].h[4]\n"
+                "ldr	%q[a0a], [%[a_ptr], #16]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0].h[2]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "ldr	%q[b2a], [%[b_ptr], #80]\n"
+                "fmla 	v12.8h, %[b0].8h, %[a0].h[4]\n"
                 "fmla   v13.8h, %[b0].8h, %[a0].h[5]\n"
-                "fmla    v14.8h, %[b0].8h, %[a0].h[6]\n"
-                "fmla    v15.8h, %[b0].8h, %[a0].h[7]\n"
-
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
-                "fmla    v20.8h, %[b1].8h, %[a0].h[4]\n"
-                "fmla    v21.8h, %[b1].8h, %[a0].h[5]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v22.8h, %[b1].8h, %[a0].h[6]\n"
-                "fmla    v23.8h, %[b1].8h, %[a0].h[7]\n"
-
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
-                "fmla    v28.8h, %[b2].8h, %[a0].h[4]\n"
-                "fmla    v29.8h, %[b2].8h, %[a0].h[5]\n"
-                "fmla    v30.8h, %[b2].8h, %[a0].h[6]\n"
-                "fmla    v31.8h, %[b2].8h, %[a0].h[7]\n"
-
-                "fmla     v8.8h , %[b0a].8h, %[a0a].h[0]\n"
-                "fmla    v16.8h, %[b1a].8h, %[a0a].h[0]\n"
-                "str    q8, [%[c_ptr]]\n"
-                "fmla    v24.8h, %[b2a].8h, %[a0a].h[0]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-
-                "fmla      v9.8h , %[b0a].8h, %[a0a].h[1]\n"
-                "str    q24, [%[c_ptr], #32]\n"
-                "fmla    v17.8h, %[b1a].8h, %[a0a].h[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
-                "fmla    v25.8h, %[b2a].8h, %[a0a].h[1]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-
-                "fmla    v10.8h, %[b0a].8h, %[a0a].h[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "fmla    v18.8h, %[b1a].8h, %[a0a].h[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-                "fmla    v26.8h, %[b2a].8h, %[a0a].h[2]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-
-                "fmla    v11.8h, %[b0a].8h, %[a0a].h[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "fmla    v19.8h, %[b1a].8h, %[a0a].h[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-                "fmla    v27.8h, %[b2a].8h, %[a0a].h[3]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-
-                "fmla     v12.8h, %[b0a].8h, %[a0a].h[4]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "fmla    v20.8h, %[b1a].8h, %[a0a].h[4]\n"
-                "str    q12, [%[c_ptr], #192]\n"
-                "fmla    v28.8h, %[b2a].8h, %[a0a].h[4]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-
-                "fmla      v13.8h, %[b0a].8h, %[a0a].h[5]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "fmla    v21.8h, %[b1a].8h, %[a0a].h[5]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-                "fmla    v29.8h, %[b2a].8h, %[a0a].h[5]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-
-                "fmla    v14.8h, %[b0a].8h, %[a0a].h[6]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "fmla    v22.8h, %[b1a].8h, %[a0a].h[6]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-                "fmla    v30.8h, %[b2a].8h, %[a0a].h[6]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-
-                "fmla    v15.8h, %[b0a].8h, %[a0a].h[7]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "fmla    v23.8h, %[b1a].8h, %[a0a].h[7]\n"
-                "str    q15, [%[c_ptr], #336]\n"
-                "fmla    v31.8h, %[b2a].8h, %[a0a].h[7]\n"
-                "b    3f\n"
+                "fmla	v14.8h, %[b0].8h, %[a0].h[6]\n"
+                "fmla	v15.8h, %[b0].8h, %[a0].h[7]\n"
+
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "fmla	v20.8h, %[b1].8h, %[a0].h[4]\n"
+                "fmla	v21.8h, %[b1].8h, %[a0].h[5]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v22.8h, %[b1].8h, %[a0].h[6]\n"
+                "fmla	v23.8h, %[b1].8h, %[a0].h[7]\n"
+
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "fmla	v28.8h, %[b2].8h, %[a0].h[4]\n"
+                "fmla	v29.8h, %[b2].8h, %[a0].h[5]\n"
+                "fmla	v30.8h, %[b2].8h, %[a0].h[6]\n"
+                "fmla	v31.8h, %[b2].8h, %[a0].h[7]\n"
+
+                "fmla 	v8.8h , %[b0a].8h, %[a0a].h[0]\n"
+                "fmla	v16.8h, %[b1a].8h, %[a0a].h[0]\n"
+                "str	q8, [%[c_ptr]]\n"
+                "fmla	v24.8h, %[b2a].8h, %[a0a].h[0]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+
+                "fmla  	v9.8h , %[b0a].8h, %[a0a].h[1]\n"
+                "str	q24, [%[c_ptr], #32]\n"
+                "fmla	v17.8h, %[b1a].8h, %[a0a].h[1]\n"
+                "str	q9, [%[c_ptr], #48]\n"
+                "fmla	v25.8h, %[b2a].8h, %[a0a].h[1]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+
+                "fmla	v10.8h, %[b0a].8h, %[a0a].h[2]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "fmla	v18.8h, %[b1a].8h, %[a0a].h[2]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+                "fmla	v26.8h, %[b2a].8h, %[a0a].h[2]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+
+                "fmla	v11.8h, %[b0a].8h, %[a0a].h[3]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "fmla	v19.8h, %[b1a].8h, %[a0a].h[3]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+                "fmla	v27.8h, %[b2a].8h, %[a0a].h[3]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+
+                "fmla 	v12.8h, %[b0a].8h, %[a0a].h[4]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "fmla	v20.8h, %[b1a].8h, %[a0a].h[4]\n"
+                "str	q12, [%[c_ptr], #192]\n"
+                "fmla	v28.8h, %[b2a].8h, %[a0a].h[4]\n"
+                "str	q20, [%[c_ptr], #208]\n"
+
+                "fmla  	v13.8h, %[b0a].8h, %[a0a].h[5]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "fmla	v21.8h, %[b1a].8h, %[a0a].h[5]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+                "fmla	v29.8h, %[b2a].8h, %[a0a].h[5]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+
+                "fmla	v14.8h, %[b0a].8h, %[a0a].h[6]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "fmla	v22.8h, %[b1a].8h, %[a0a].h[6]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+                "fmla	v30.8h, %[b2a].8h, %[a0a].h[6]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+
+                "fmla	v15.8h, %[b0a].8h, %[a0a].h[7]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "fmla	v23.8h, %[b1a].8h, %[a0a].h[7]\n"
+                "str	q15, [%[c_ptr], #336]\n"
+                "fmla	v31.8h, %[b2a].8h, %[a0a].h[7]\n"
+                "b	3f\n"
 
                 // Odd tail
                 "2:\n"
-                "fmla     v8.8h , %[b0].8h, %[a0].h[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "add    %[a_ptr], %[a_ptr], #16\n"
-                "str    q8, [%[c_ptr]]\n"
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-
-                "fmla      v9.8h , %[b0].8h, %[a0].h[1]\n"
-                "str    q24, [%[c_ptr], #32]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-
-                "fmla    v10.8h, %[b0].8h, %[a0].h[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-
-                "fmla     v12.8h, %[b0].8h, %[a0].h[4]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "fmla    v20.8h, %[b1].8h, %[a0].h[4]\n"
-                "str    q12, [%[c_ptr], #192]\n"
-                "fmla    v28.8h, %[b2].8h, %[a0].h[4]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-
-                "fmla      v13.8h, %[b0].8h, %[a0].h[5]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "fmla    v21.8h, %[b1].8h, %[a0].h[5]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-                "fmla    v29.8h, %[b2].8h, %[a0].h[5]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-
-                "fmla    v14.8h, %[b0].8h, %[a0].h[6]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "fmla    v22.8h, %[b1].8h, %[a0].h[6]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-                "fmla    v30.8h, %[b2].8h, %[a0].h[6]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-
-                "fmla    v15.8h, %[b0].8h, %[a0].h[7]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "fmla    v23.8h, %[b1].8h, %[a0].h[7]\n"
-                "str    q15, [%[c_ptr], #336]\n"
-                "fmla    v31.8h, %[b2].8h, %[a0].h[7]\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0].h[0]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "add	%[a_ptr], %[a_ptr], #16\n"
+                "str	q8, [%[c_ptr]]\n"
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+
+                "fmla  	v9.8h , %[b0].8h, %[a0].h[1]\n"
+                "str	q24, [%[c_ptr], #32]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "str	q9, [%[c_ptr], #48]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+
+                "fmla	v10.8h, %[b0].8h, %[a0].h[2]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+
+                "fmla 	v12.8h, %[b0].8h, %[a0].h[4]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "fmla	v20.8h, %[b1].8h, %[a0].h[4]\n"
+                "str	q12, [%[c_ptr], #192]\n"
+                "fmla	v28.8h, %[b2].8h, %[a0].h[4]\n"
+                "str	q20, [%[c_ptr], #208]\n"
+
+                "fmla  	v13.8h, %[b0].8h, %[a0].h[5]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "fmla	v21.8h, %[b1].8h, %[a0].h[5]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+                "fmla	v29.8h, %[b2].8h, %[a0].h[5]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+
+                "fmla	v14.8h, %[b0].8h, %[a0].h[6]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "fmla	v22.8h, %[b1].8h, %[a0].h[6]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+                "fmla	v30.8h, %[b2].8h, %[a0].h[6]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+
+                "fmla	v15.8h, %[b0].8h, %[a0].h[7]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "fmla	v23.8h, %[b1].8h, %[a0].h[7]\n"
+                "str	q15, [%[c_ptr], #336]\n"
+                "fmla	v31.8h, %[b2].8h, %[a0].h[7]\n"
 
                 "3:\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "+w" (a0), [a0a] "+w" (a0a),
@@ -346,4 +346,4 @@ void a64_hgemm_asimd_8x24(const __fp16 *Apanel, const __fp16 *Bpanel, __fp16 *Cp
 
 } // namespace arm_gemm
 
-#endif // __aarch64__ && (ENABLE_FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+#endif // __aarch64__ && (FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp
index 6e9349fac2..9aa5a2a9cc 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp
@@ -22,8 +22,8 @@
  * SOFTWARE.
  */
 
-// Build on AArch64 where either ENABLE_FP16_KERNELS is set or FP16 is explicitly supported.
-#if defined(__aarch64__) && (defined(ENABLE_FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
+// Build on AArch64 where either FP16_KERNELS is set or FP16 is explicitly supported.
+#if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
 
 #include <arm_neon.h>
 
@@ -64,271 +64,271 @@ void a64_hgemm_asimd_8x24_x1(const __fp16 *Apanel, const __fp16 *Bpanel, __fp16
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.8h, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.8h, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.8h, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v11.8h, #0x0\n"
-                "movi    v12.8h, #0x0\n"
-                "movi    v13.8h, #0x0\n"
-                "movi    v14.8h, #0x0\n"
+                "movi	v8.8h, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.8h, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.8h, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v11.8h, #0x0\n"
+                "movi	v12.8h, #0x0\n"
+                "movi	v13.8h, #0x0\n"
+                "movi	v14.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v15.8h, #0x0\n"
+                "movi	v15.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v16.8h, #0x0\n"
+                "movi	v16.8h, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v17.8h, #0x0\n"
+                "movi	v17.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v18.8h, #0x0\n"
+                "movi	v18.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v19.8h, #0x0\n"
+                "movi	v19.8h, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.8h, #0x0\n"
-                "movi    v21.8h, #0x0\n"
-                "movi    v22.8h, #0x0\n"
-                "movi    v23.8h, #0x0\n"
-                "movi    v24.8h, #0x0\n"
-                "movi    v25.8h, #0x0\n"
-                "movi    v26.8h, #0x0\n"
-                "movi    v27.8h, #0x0\n"
-                "movi    v28.8h, #0x0\n"
-                "movi    v29.8h, #0x0\n"
-                "movi    v30.8h, #0x0\n"
-                "movi    v31.8h, #0x0\n"
+                "movi	v20.8h, #0x0\n"
+                "movi	v21.8h, #0x0\n"
+                "movi	v22.8h, #0x0\n"
+                "movi	v23.8h, #0x0\n"
+                "movi	v24.8h, #0x0\n"
+                "movi	v25.8h, #0x0\n"
+                "movi	v26.8h, #0x0\n"
+                "movi	v27.8h, #0x0\n"
+                "movi	v28.8h, #0x0\n"
+                "movi	v29.8h, #0x0\n"
+                "movi	v30.8h, #0x0\n"
+                "movi	v31.8h, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 "1:\n"
-                "fmla     v8.8h , %[b0].8h, %[a0].h[0]\n"
-                "fmla      v9.8h , %[b0].8h, %[a0].h[1]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0].h[2]\n"
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla     v12.8h, %[b0].8h, %[a0].h[4]\n"
-                "fmla    v13.8h, %[b0].8h, %[a0].h[5]\n"
-                "fmla    v14.8h, %[b0].8h, %[a0].h[6]\n"
-                "fmla    v15.8h, %[b0].8h, %[a0].h[7]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0].h[0]\n"
+                "fmla  	v9.8h , %[b0].8h, %[a0].h[1]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0].h[2]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla 	v12.8h, %[b0].8h, %[a0].h[4]\n"
+                "fmla	v13.8h, %[b0].8h, %[a0].h[5]\n"
+                "fmla	v14.8h, %[b0].8h, %[a0].h[6]\n"
+                "fmla	v15.8h, %[b0].8h, %[a0].h[7]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v20.8h, %[b1].8h, %[a0].h[4]\n"
-                "fmla    v21.8h, %[b1].8h, %[a0].h[5]\n"
-                "fmla    v22.8h, %[b1].8h, %[a0].h[6]\n"
-                "fmla    v23.8h, %[b1].8h, %[a0].h[7]\n"
-                "ldr    %q[b1], [%[b_ptr], #-32]\n"
-
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v20.8h, %[b1].8h, %[a0].h[4]\n"
+                "fmla	v21.8h, %[b1].8h, %[a0].h[5]\n"
+                "fmla	v22.8h, %[b1].8h, %[a0].h[6]\n"
+                "fmla	v23.8h, %[b1].8h, %[a0].h[7]\n"
+                "ldr	%q[b1], [%[b_ptr], #-32]\n"
+
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #288]")
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
-                "ldr    %q[a0a], [%[a_ptr], #16]\n"
-                "fmla    v28.8h, %[b2].8h, %[a0].h[4]\n"
-                "fmla    v29.8h, %[b2].8h, %[a0].h[5]\n"
-                "fmla    v30.8h, %[b2].8h, %[a0].h[6]\n"
-                "fmla    v31.8h, %[b2].8h, %[a0].h[7]\n"
-                "ldr    %q[b2], [%[b_ptr], #-16]\n"
-
-                "fmla     v8.8h , %[b0].8h, %[a0a].h[0]\n"
-                "fmla    v9.8h , %[b0].8h, %[a0a].h[1]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0a].h[2]\n"
-                "fmla    v11.8h, %[b0].8h, %[a0a].h[3]\n"
-                "fmla     v12.8h, %[b0].8h, %[a0a].h[4]\n"
-                "fmla    v13.8h, %[b0].8h, %[a0a].h[5]\n"
-                "fmla    v14.8h, %[b0].8h, %[a0a].h[6]\n"
-                "fmla    v15.8h, %[b0].8h, %[a0a].h[7]\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-
-                "fmla    v16.8h, %[b1].8h, %[a0a].h[0]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0a].h[1]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "ldr	%q[a0a], [%[a_ptr], #16]\n"
+                "fmla	v28.8h, %[b2].8h, %[a0].h[4]\n"
+                "fmla	v29.8h, %[b2].8h, %[a0].h[5]\n"
+                "fmla	v30.8h, %[b2].8h, %[a0].h[6]\n"
+                "fmla	v31.8h, %[b2].8h, %[a0].h[7]\n"
+                "ldr	%q[b2], [%[b_ptr], #-16]\n"
+
+                "fmla 	v8.8h , %[b0].8h, %[a0a].h[0]\n"
+                "fmla	v9.8h , %[b0].8h, %[a0a].h[1]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0a].h[2]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0a].h[3]\n"
+                "fmla 	v12.8h, %[b0].8h, %[a0a].h[4]\n"
+                "fmla	v13.8h, %[b0].8h, %[a0a].h[5]\n"
+                "fmla	v14.8h, %[b0].8h, %[a0a].h[6]\n"
+                "fmla	v15.8h, %[b0].8h, %[a0a].h[7]\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+
+                "fmla	v16.8h, %[b1].8h, %[a0a].h[0]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0a].h[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #352]")
-                "fmla    v18.8h, %[b1].8h, %[a0a].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0a].h[3]\n"
-                "fmla    v20.8h, %[b1].8h, %[a0a].h[4]\n"
-                "fmla    v21.8h, %[b1].8h, %[a0a].h[5]\n"
-                "fmla    v22.8h, %[b1].8h, %[a0a].h[6]\n"
-                "fmla    v23.8h, %[b1].8h, %[a0a].h[7]\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-
-                "fmla    v24.8h, %[b2].8h, %[a0a].h[0]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0a].h[1]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v26.8h, %[b2].8h, %[a0a].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0a].h[3]\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "fmla    v28.8h, %[b2].8h, %[a0a].h[4]\n"
-                "fmla    v29.8h, %[b2].8h, %[a0a].h[5]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v30.8h, %[b2].8h, %[a0a].h[6]\n"
-                "fmla    v31.8h, %[b2].8h, %[a0a].h[7]\n"
-
-                "bne    1b\n"
+                "fmla	v18.8h, %[b1].8h, %[a0a].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0a].h[3]\n"
+                "fmla	v20.8h, %[b1].8h, %[a0a].h[4]\n"
+                "fmla	v21.8h, %[b1].8h, %[a0a].h[5]\n"
+                "fmla	v22.8h, %[b1].8h, %[a0a].h[6]\n"
+                "fmla	v23.8h, %[b1].8h, %[a0a].h[7]\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+
+                "fmla	v24.8h, %[b2].8h, %[a0a].h[0]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0a].h[1]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v26.8h, %[b2].8h, %[a0a].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0a].h[3]\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "fmla	v28.8h, %[b2].8h, %[a0a].h[4]\n"
+                "fmla	v29.8h, %[b2].8h, %[a0a].h[5]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v30.8h, %[b2].8h, %[a0a].h[6]\n"
+                "fmla	v31.8h, %[b2].8h, %[a0a].h[7]\n"
+
+                "bne	1b\n"
                 "4:\n"
 
                 // Jump to odd tail if necessary.
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Even tail.
-                "fmla     v8.8h , %[b0].8h, %[a0].h[0]\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0].h[0]\n"
                 "fmla   v9.8h , %[b0].8h, %[a0].h[1]\n"
-                "fmla    v10.8h, %[b0].8h, %[a0].h[2]\n"
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla     v12.8h, %[b0].8h, %[a0].h[4]\n"
+                "fmla	v10.8h, %[b0].8h, %[a0].h[2]\n"
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla 	v12.8h, %[b0].8h, %[a0].h[4]\n"
                 "fmla   v13.8h, %[b0].8h, %[a0].h[5]\n"
-                "fmla    v14.8h, %[b0].8h, %[a0].h[6]\n"
-                "fmla    v15.8h, %[b0].8h, %[a0].h[7]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
-                "fmla    v20.8h, %[b1].8h, %[a0].h[4]\n"
-                "fmla    v21.8h, %[b1].8h, %[a0].h[5]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v22.8h, %[b1].8h, %[a0].h[6]\n"
-                "fmla    v23.8h, %[b1].8h, %[a0].h[7]\n"
-                "ldr    %q[b1], [%[b_ptr], #-32]\n"
-
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
-                "ldr    %q[a0a], [%[a_ptr], #-16]\n"
-                "fmla    v28.8h, %[b2].8h, %[a0].h[4]\n"
-                "fmla    v29.8h, %[b2].8h, %[a0].h[5]\n"
-                "fmla    v30.8h, %[b2].8h, %[a0].h[6]\n"
-                "fmla    v31.8h, %[b2].8h, %[a0].h[7]\n"
-                "ldr    %q[b2], [%[b_ptr], #-16]\n"
-
-                "fmla     v8.8h , %[b0].8h, %[a0a].h[0]\n"
-                "fmla    v16.8h, %[b1].8h, %[a0a].h[0]\n"
-                "str    q8, [%[c_ptr]]\n"
-                "fmla    v24.8h, %[b2].8h, %[a0a].h[0]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-
-                "fmla      v9.8h , %[b0].8h, %[a0a].h[1]\n"
-                "str    q24, [%[c_ptr], #32]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0a].h[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0a].h[1]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-
-                "fmla    v10.8h, %[b0].8h, %[a0a].h[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "fmla    v18.8h, %[b1].8h, %[a0a].h[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-                "fmla    v26.8h, %[b2].8h, %[a0a].h[2]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-
-                "fmla    v11.8h, %[b0].8h, %[a0a].h[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0a].h[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0a].h[3]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-
-                "fmla     v12.8h, %[b0].8h, %[a0a].h[4]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "fmla    v20.8h, %[b1].8h, %[a0a].h[4]\n"
-                "str    q12, [%[c_ptr], #192]\n"
-                "fmla    v28.8h, %[b2].8h, %[a0a].h[4]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-
-                "fmla      v13.8h, %[b0].8h, %[a0a].h[5]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "fmla    v21.8h, %[b1].8h, %[a0a].h[5]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-                "fmla    v29.8h, %[b2].8h, %[a0a].h[5]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-
-                "fmla    v14.8h, %[b0].8h, %[a0a].h[6]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "fmla    v22.8h, %[b1].8h, %[a0a].h[6]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-                "fmla    v30.8h, %[b2].8h, %[a0a].h[6]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-
-                "fmla    v15.8h, %[b0].8h, %[a0a].h[7]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "fmla    v23.8h, %[b1].8h, %[a0a].h[7]\n"
-                "str    q15, [%[c_ptr], #336]\n"
-                "fmla    v31.8h, %[b2].8h, %[a0a].h[7]\n"
-                "b    3f\n"
+                "fmla	v14.8h, %[b0].8h, %[a0].h[6]\n"
+                "fmla	v15.8h, %[b0].8h, %[a0].h[7]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "fmla	v20.8h, %[b1].8h, %[a0].h[4]\n"
+                "fmla	v21.8h, %[b1].8h, %[a0].h[5]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v22.8h, %[b1].8h, %[a0].h[6]\n"
+                "fmla	v23.8h, %[b1].8h, %[a0].h[7]\n"
+                "ldr	%q[b1], [%[b_ptr], #-32]\n"
+
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "ldr	%q[a0a], [%[a_ptr], #-16]\n"
+                "fmla	v28.8h, %[b2].8h, %[a0].h[4]\n"
+                "fmla	v29.8h, %[b2].8h, %[a0].h[5]\n"
+                "fmla	v30.8h, %[b2].8h, %[a0].h[6]\n"
+                "fmla	v31.8h, %[b2].8h, %[a0].h[7]\n"
+                "ldr	%q[b2], [%[b_ptr], #-16]\n"
+
+                "fmla 	v8.8h , %[b0].8h, %[a0a].h[0]\n"
+                "fmla	v16.8h, %[b1].8h, %[a0a].h[0]\n"
+                "str	q8, [%[c_ptr]]\n"
+                "fmla	v24.8h, %[b2].8h, %[a0a].h[0]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+
+                "fmla  	v9.8h , %[b0].8h, %[a0a].h[1]\n"
+                "str	q24, [%[c_ptr], #32]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0a].h[1]\n"
+                "str	q9, [%[c_ptr], #48]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0a].h[1]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+
+                "fmla	v10.8h, %[b0].8h, %[a0a].h[2]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0a].h[2]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0a].h[2]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+
+                "fmla	v11.8h, %[b0].8h, %[a0a].h[3]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0a].h[3]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0a].h[3]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+
+                "fmla 	v12.8h, %[b0].8h, %[a0a].h[4]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "fmla	v20.8h, %[b1].8h, %[a0a].h[4]\n"
+                "str	q12, [%[c_ptr], #192]\n"
+                "fmla	v28.8h, %[b2].8h, %[a0a].h[4]\n"
+                "str	q20, [%[c_ptr], #208]\n"
+
+                "fmla  	v13.8h, %[b0].8h, %[a0a].h[5]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "fmla	v21.8h, %[b1].8h, %[a0a].h[5]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+                "fmla	v29.8h, %[b2].8h, %[a0a].h[5]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+
+                "fmla	v14.8h, %[b0].8h, %[a0a].h[6]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "fmla	v22.8h, %[b1].8h, %[a0a].h[6]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+                "fmla	v30.8h, %[b2].8h, %[a0a].h[6]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+
+                "fmla	v15.8h, %[b0].8h, %[a0a].h[7]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "fmla	v23.8h, %[b1].8h, %[a0a].h[7]\n"
+                "str	q15, [%[c_ptr], #336]\n"
+                "fmla	v31.8h, %[b2].8h, %[a0a].h[7]\n"
+                "b	3f\n"
 
                 // Odd tail
                 "2:\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla     v8.8h , %[b0].8h, %[a0].h[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "fmla    v16.8h, %[b1].8h, %[a0].h[0]\n"
-                "add    %[a_ptr], %[a_ptr], #16\n"
-                "str    q8, [%[c_ptr]]\n"
-                "fmla    v24.8h, %[b2].8h, %[a0].h[0]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-
-                "fmla      v9.8h , %[b0].8h, %[a0].h[1]\n"
-                "str    q24, [%[c_ptr], #32]\n"
-                "fmla    v17.8h, %[b1].8h, %[a0].h[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
-                "fmla    v25.8h, %[b2].8h, %[a0].h[1]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-
-                "fmla    v10.8h, %[b0].8h, %[a0].h[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "fmla    v18.8h, %[b1].8h, %[a0].h[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-                "fmla    v26.8h, %[b2].8h, %[a0].h[2]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-
-                "fmla    v11.8h, %[b0].8h, %[a0].h[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "fmla    v19.8h, %[b1].8h, %[a0].h[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-                "fmla    v27.8h, %[b2].8h, %[a0].h[3]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-
-                "fmla     v12.8h, %[b0].8h, %[a0].h[4]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "fmla    v20.8h, %[b1].8h, %[a0].h[4]\n"
-                "str    q12, [%[c_ptr], #192]\n"
-                "fmla    v28.8h, %[b2].8h, %[a0].h[4]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-
-                "fmla      v13.8h, %[b0].8h, %[a0].h[5]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "fmla    v21.8h, %[b1].8h, %[a0].h[5]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-                "fmla    v29.8h, %[b2].8h, %[a0].h[5]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-
-                "fmla    v14.8h, %[b0].8h, %[a0].h[6]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "fmla    v22.8h, %[b1].8h, %[a0].h[6]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-                "fmla    v30.8h, %[b2].8h, %[a0].h[6]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-
-                "fmla    v15.8h, %[b0].8h, %[a0].h[7]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "fmla    v23.8h, %[b1].8h, %[a0].h[7]\n"
-                "str    q15, [%[c_ptr], #336]\n"
-                "fmla    v31.8h, %[b2].8h, %[a0].h[7]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla 	v8.8h , %[b0].8h, %[a0].h[0]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "fmla	v16.8h, %[b1].8h, %[a0].h[0]\n"
+                "add	%[a_ptr], %[a_ptr], #16\n"
+                "str	q8, [%[c_ptr]]\n"
+                "fmla	v24.8h, %[b2].8h, %[a0].h[0]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+
+                "fmla  	v9.8h , %[b0].8h, %[a0].h[1]\n"
+                "str	q24, [%[c_ptr], #32]\n"
+                "fmla	v17.8h, %[b1].8h, %[a0].h[1]\n"
+                "str	q9, [%[c_ptr], #48]\n"
+                "fmla	v25.8h, %[b2].8h, %[a0].h[1]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+
+                "fmla	v10.8h, %[b0].8h, %[a0].h[2]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "fmla	v18.8h, %[b1].8h, %[a0].h[2]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+                "fmla	v26.8h, %[b2].8h, %[a0].h[2]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+
+                "fmla	v11.8h, %[b0].8h, %[a0].h[3]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "fmla	v19.8h, %[b1].8h, %[a0].h[3]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+                "fmla	v27.8h, %[b2].8h, %[a0].h[3]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+
+                "fmla 	v12.8h, %[b0].8h, %[a0].h[4]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "fmla	v20.8h, %[b1].8h, %[a0].h[4]\n"
+                "str	q12, [%[c_ptr], #192]\n"
+                "fmla	v28.8h, %[b2].8h, %[a0].h[4]\n"
+                "str	q20, [%[c_ptr], #208]\n"
+
+                "fmla  	v13.8h, %[b0].8h, %[a0].h[5]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "fmla	v21.8h, %[b1].8h, %[a0].h[5]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+                "fmla	v29.8h, %[b2].8h, %[a0].h[5]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+
+                "fmla	v14.8h, %[b0].8h, %[a0].h[6]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "fmla	v22.8h, %[b1].8h, %[a0].h[6]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+                "fmla	v30.8h, %[b2].8h, %[a0].h[6]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+
+                "fmla	v15.8h, %[b0].8h, %[a0].h[7]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "fmla	v23.8h, %[b1].8h, %[a0].h[7]\n"
+                "str	q15, [%[c_ptr], #336]\n"
+                "fmla	v31.8h, %[b2].8h, %[a0].h[7]\n"
 
                 "3:\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "+w" (a0), [a0a] "+w" (a0a),
@@ -343,4 +343,4 @@ void a64_hgemm_asimd_8x24_x1(const __fp16 *Apanel, const __fp16 *Bpanel, __fp16
 
 } // namespace arm_gemm
 
-#endif // __aarch64__ && (ENABLE_FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+#endif // __aarch64__ && (FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16.hpp
index 278d869afb..22a80885b3 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 2> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 2> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp
index 4494e2ac13..10f5dd4716 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp
@@ -48,19 +48,18 @@ void a64_hybrid_bf16fp32_dot_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -81,7 +80,6 @@ void a64_hybrid_bf16fp32_dot_6x16 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -104,10 +102,10 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "cmp %x[M], #0x2\n"
       "bgt 71f\n"
       "beq 36f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x12, 3f\n"
       "ldr q8, [x12, #0x0]\n"
@@ -190,8 +188,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "mov x28, #0x0\n"
       "15:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 16f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -215,10 +213,6 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f40f0e9  // bfdot v9.4s, v7.8h, v0.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f40f22a  // bfdot v10.4s, v17.8h, v0.h[0]\n"
       "ldr q17, [x10, #0x40]\n"
       ".inst 0x4f40f20b  // bfdot v11.4s, v16.8h, v0.h[0]\n"
@@ -243,21 +237,22 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q17, [x10, #0xe0]\n"
       ".inst 0x4f60fa09  // bfdot v9.4s, v16.8h, v0.h[3]\n"
       "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4f60fa2a  // bfdot v10.4s, v17.8h, v0.h[3]\n"
-      "ldr q6, [x10, #0x0]\n"
       ".inst 0x4f60fa0b  // bfdot v11.4s, v16.8h, v0.h[3]\n"
       "ldr q0, [x26, #0x0]\n"
+      "cmp x27, #0x10\n"
+      "add x10, x10, #0x100\n"
+      "ldr q6, [x10, #0x0]\n"
       "ldr q7, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 18b\n"
       "19:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x4f40f0c8  // bfdot v8.4s, v6.8h, v0.h[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f40f0e9  // bfdot v9.4s, v7.8h, v0.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f40f22a  // bfdot v10.4s, v17.8h, v0.h[0]\n"
       "ldr q17, [x10, #0x40]\n"
       ".inst 0x4f40f20b  // bfdot v11.4s, v16.8h, v0.h[0]\n"
@@ -282,26 +277,29 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q17, [x10, #0xe0]\n"
       ".inst 0x4f60fa09  // bfdot v9.4s, v16.8h, v0.h[3]\n"
       "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x4f60fa2a  // bfdot v10.4s, v17.8h, v0.h[3]\n"
       ".inst 0x4f60fa0b  // bfdot v11.4s, v16.8h, v0.h[3]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "20:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 24f\n"
       "cmp x27, #0x2\n"
       "blt 22f\n"
       "21:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x26], #0x4\n"
-      "ldr q17, [x10, #0x0]\n"
+      "ldr q16, [x10, #0x0]\n"
+      ".inst 0x4f52f208  // bfdot v8.4s, v16.8h, v18.h[0]\n"
       "sub x27, x27, #0x2\n"
       "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x2\n"
-      ".inst 0x4f52f228  // bfdot v8.4s, v17.8h, v18.h[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f52f209  // bfdot v9.4s, v16.8h, v18.h[0]\n"
+      "cmp x27, #0x2\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x4f52f22a  // bfdot v10.4s, v17.8h, v18.h[0]\n"
       ".inst 0x4f52f20b  // bfdot v11.4s, v16.8h, v18.h[0]\n"
+      "add x10, x10, #0x40\n"
       "bge 21b\n"
       "22:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 24f\n"
@@ -310,12 +308,12 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q17, [x10, #0x0]\n"
       "ldr q16, [x10, #0x10]\n"
       ".inst 0x4f40f228  // bfdot v8.4s, v17.8h, v0.h[0]\n"
-      "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f40f209  // bfdot v9.4s, v16.8h, v0.h[0]\n"
+      "ldr q17, [x10, #0x20]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x4f40f22a  // bfdot v10.4s, v17.8h, v0.h[0]\n"
       ".inst 0x4f40f20b  // bfdot v11.4s, v16.8h, v0.h[0]\n"
+      "add x10, x10, #0x40\n"
       "24:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -323,9 +321,9 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "bne 15b\n"
       "prfm pstl1keep, [x9, #0x0]\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v17.4s\n"
       "fmin v9.4s, v9.4s, v17.4s\n"
@@ -395,95 +393,95 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "bgt 2b\n"
       "b 212f\n"
       "36:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "37:"  // Height 2: Column loop
       "cbz x12, 38f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "add x12, x12, #0x40\n"
       "b 49f\n"
       "38:"  // Height 2: no bias
       "tbz %x[flags], #0, 48f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "bge 47f\n"
       "tbz x11, #3, 42f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
       "tbz x11, #2, 40f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
       "tbz x11, #1, 39f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x11, #0, 46f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v15.s }[2], [x25]\n"
       "b 46f\n"
       "39:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 46f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
       "b 46f\n"
       "40:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x11, #1, 41f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x11, #0, 46f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v14.s }[2], [x25]\n"
       "b 46f\n"
       "41:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 46f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
       "b 46f\n"
       "42:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x11, #2, 44f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "tbz x11, #1, 43f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x11, #0, 46f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v13.s }[2], [x25]\n"
       "b 46f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 46f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
       "b 46f\n"
       "44:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x11, #1, 45f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x11, #0, 46f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v12.s }[2], [x25]\n"
       "b 46f\n"
       "45:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
       "46:"  // Height 2: Partial accumulate: Done
       "sub x9, x9, x20\n"
@@ -493,10 +491,10 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "b 49f\n"
       "48:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -511,8 +509,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "mov x28, #0x0\n"
       "50:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 51f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -544,22 +542,22 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       ".inst 0x4f41f0ed  // bfdot v13.4s, v7.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f40f22a  // bfdot v10.4s, v17.8h, v0.h[0]\n"
       ".inst 0x4f41f22e  // bfdot v14.4s, v17.8h, v1.h[0]\n"
       "ldr q17, [x10, #0x40]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4f40f20b  // bfdot v11.4s, v16.8h, v0.h[0]\n"
       ".inst 0x4f41f20f  // bfdot v15.4s, v16.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "cmp x27, #0x10\n"
       ".inst 0x4f60f228  // bfdot v8.4s, v17.8h, v0.h[1]\n"
       ".inst 0x4f61f22c  // bfdot v12.4s, v17.8h, v1.h[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f60f209  // bfdot v9.4s, v16.8h, v0.h[1]\n"
       ".inst 0x4f61f20d  // bfdot v13.4s, v16.8h, v1.h[1]\n"
       "ldr q16, [x10, #0x70]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f60f22a  // bfdot v10.4s, v17.8h, v0.h[1]\n"
       ".inst 0x4f61f22e  // bfdot v14.4s, v17.8h, v1.h[1]\n"
       "ldr q17, [x10, #0x80]\n"
@@ -603,18 +601,18 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       ".inst 0x4f41f0ed  // bfdot v13.4s, v7.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f40f22a  // bfdot v10.4s, v17.8h, v0.h[0]\n"
       ".inst 0x4f41f22e  // bfdot v14.4s, v17.8h, v1.h[0]\n"
       "ldr q17, [x10, #0x40]\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x4f40f20b  // bfdot v11.4s, v16.8h, v0.h[0]\n"
       ".inst 0x4f41f20f  // bfdot v15.4s, v16.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f60f228  // bfdot v8.4s, v17.8h, v0.h[1]\n"
       ".inst 0x4f61f22c  // bfdot v12.4s, v17.8h, v1.h[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f60f209  // bfdot v9.4s, v16.8h, v0.h[1]\n"
       ".inst 0x4f61f20d  // bfdot v13.4s, v16.8h, v1.h[1]\n"
       "ldr q16, [x10, #0x70]\n"
@@ -655,18 +653,18 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr s19, [x26], #0x4\n"
       "ldr s18, [x25], #0x4\n"
       "sub x27, x27, #0x2\n"
+      "cmp x27, #0x2\n"
       "ldr q17, [x10, #0x0]\n"
       "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x2\n"
       ".inst 0x4f53f228  // bfdot v8.4s, v17.8h, v19.h[0]\n"
       ".inst 0x4f52f22c  // bfdot v12.4s, v17.8h, v18.h[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f53f209  // bfdot v9.4s, v16.8h, v19.h[0]\n"
       ".inst 0x4f52f20d  // bfdot v13.4s, v16.8h, v18.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x4f53f22a  // bfdot v10.4s, v17.8h, v19.h[0]\n"
       ".inst 0x4f52f22e  // bfdot v14.4s, v17.8h, v18.h[0]\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x4f53f20b  // bfdot v11.4s, v16.8h, v19.h[0]\n"
       ".inst 0x4f52f20f  // bfdot v15.4s, v16.8h, v18.h[0]\n"
       "bge 56b\n"
@@ -683,9 +681,9 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       ".inst 0x4f40f209  // bfdot v9.4s, v16.8h, v0.h[0]\n"
       ".inst 0x4f41f20d  // bfdot v13.4s, v16.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x4f40f22a  // bfdot v10.4s, v17.8h, v0.h[0]\n"
       ".inst 0x4f41f22e  // bfdot v14.4s, v17.8h, v1.h[0]\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x4f40f20b  // bfdot v11.4s, v16.8h, v0.h[0]\n"
       ".inst 0x4f41f20f  // bfdot v15.4s, v16.8h, v1.h[0]\n"
       "59:"  // Height 2: Multiply loop: No odd multiplies
@@ -694,13 +692,13 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 50b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "tbz %x[flags], #1, 60f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v17.4s\n"
       "fmin v9.4s, v9.4s, v17.4s\n"
@@ -724,63 +722,63 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "tbz x11, #3, 64f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
       "tbz x11, #2, 62f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
       "tbz x11, #1, 61f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
       "tbz x11, #0, 68f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
+      "st1 { v15.s }[2], [x25]\n"
       "b 68f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x11, #0, 68f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
       "b 68f\n"
       "62:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x11, #1, 63f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
+      "str d14, [x25], #0x8\n"
       "tbz x11, #0, 68f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
+      "st1 { v14.s }[2], [x25]\n"
       "b 68f\n"
       "63:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x11, #0, 68f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
       "b 68f\n"
       "64:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x11, #2, 66f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
       "tbz x11, #1, 65f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
+      "str d13, [x25], #0x8\n"
       "tbz x11, #0, 68f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
+      "st1 { v13.s }[2], [x25]\n"
       "b 68f\n"
       "65:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x11, #0, 68f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
       "b 68f\n"
       "66:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x11, #1, 67f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
+      "str d12, [x25], #0x8\n"
       "tbz x11, #0, 68f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
+      "st1 { v12.s }[2], [x25]\n"
       "b 68f\n"
       "67:"  // Height 2: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
       "68:"  // Height 2: Partial direct writeback: Done
       "b 70f\n"
       "69:"  // Height 2: Full writeback
@@ -789,126 +787,126 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
       "70:"  // Height 2: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 37b\n"
       "b 212f\n"
       "71:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "72:"  // Height 3: Column loop
       "cbz x12, 73f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
       "b 84f\n"
       "73:"  // Height 3: no bias
       "tbz %x[flags], #0, 83f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "bge 82f\n"
       "tbz x11, #3, 77f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "tbz x11, #2, 75f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
       "tbz x11, #1, 74f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x11, #0, 81f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
       "b 81f\n"
       "74:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 81f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
       "b 81f\n"
       "75:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x11, #1, 76f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x11, #0, 81f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
       "b 81f\n"
       "76:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 81f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
       "b 81f\n"
       "77:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x11, #2, 79f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
       "tbz x11, #1, 78f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x11, #0, 81f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
       "b 81f\n"
       "78:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 81f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
       "b 81f\n"
       "79:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x11, #1, 80f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
       "tbz x11, #0, 81f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
       "b 81f\n"
       "80:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
       "81:"  // Height 3: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 84f\n"
@@ -917,14 +915,14 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
       "b 84f\n"
       "83:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -943,8 +941,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "mov x28, #0x0\n"
       "85:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 86f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -984,18 +982,18 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       ".inst 0x4f42f0f1  // bfdot v17.4s, v7.8h, v2.h[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "add x24, x24, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f40f2aa  // bfdot v10.4s, v21.8h, v0.h[0]\n"
       ".inst 0x4f41f2ae  // bfdot v14.4s, v21.8h, v1.h[0]\n"
+      "cmp x27, #0x10\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f42f2b2  // bfdot v18.4s, v21.8h, v2.h[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f40f28b  // bfdot v11.4s, v20.8h, v0.h[0]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f41f28f  // bfdot v15.4s, v20.8h, v1.h[0]\n"
       ".inst 0x4f42f293  // bfdot v19.4s, v20.8h, v2.h[0]\n"
       "ldr q20, [x10, #0x50]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f60f2a8  // bfdot v8.4s, v21.8h, v0.h[1]\n"
       ".inst 0x4f61f2ac  // bfdot v12.4s, v21.8h, v1.h[1]\n"
       ".inst 0x4f62f2b0  // bfdot v16.4s, v21.8h, v2.h[1]\n"
@@ -1062,14 +1060,14 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       ".inst 0x4f42f0f1  // bfdot v17.4s, v7.8h, v2.h[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f40f2aa  // bfdot v10.4s, v21.8h, v0.h[0]\n"
       ".inst 0x4f41f2ae  // bfdot v14.4s, v21.8h, v1.h[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f42f2b2  // bfdot v18.4s, v21.8h, v2.h[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f40f28b  // bfdot v11.4s, v20.8h, v0.h[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f41f28f  // bfdot v15.4s, v20.8h, v1.h[0]\n"
       ".inst 0x4f42f293  // bfdot v19.4s, v20.8h, v2.h[0]\n"
       "ldr q20, [x10, #0x50]\n"
@@ -1128,12 +1126,12 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr s24, [x26], #0x4\n"
       "ldr s23, [x25], #0x4\n"
       "sub x27, x27, #0x2\n"
+      "cmp x27, #0x2\n"
       "ldr s22, [x24], #0x4\n"
       "ldr q21, [x10, #0x0]\n"
-      "cmp x27, #0x2\n"
-      "ldr q20, [x10, #0x10]\n"
       ".inst 0x4f58f2a8  // bfdot v8.4s, v21.8h, v24.h[0]\n"
       ".inst 0x4f57f2ac  // bfdot v12.4s, v21.8h, v23.h[0]\n"
+      "ldr q20, [x10, #0x10]\n"
       ".inst 0x4f56f2b0  // bfdot v16.4s, v21.8h, v22.h[0]\n"
       "ldr q21, [x10, #0x20]\n"
       ".inst 0x4f58f289  // bfdot v9.4s, v20.8h, v24.h[0]\n"
@@ -1177,15 +1175,15 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 85b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "tbz %x[flags], #1, 95f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.4s }, [x21]\n"
       "ld1r { v20.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v21.4s\n"
       "fmin v9.4s, v9.4s, v21.4s\n"
@@ -1217,79 +1215,79 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "tbz x11, #3, 99f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
       "tbz x11, #2, 97f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
       "tbz x11, #1, 96f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x11, #0, 103f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
       "b 103f\n"
       "96:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x11, #0, 103f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
       "b 103f\n"
       "97:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x11, #1, 98f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x11, #0, 103f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
       "b 103f\n"
       "98:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x11, #0, 103f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
       "b 103f\n"
       "99:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x11, #2, 101f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
       "tbz x11, #1, 100f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x11, #0, 103f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
       "b 103f\n"
       "100:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x11, #0, 103f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
       "b 103f\n"
       "101:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x11, #1, 102f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x11, #0, 103f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
       "b 103f\n"
       "102:"  // Height 3: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
       "103:"  // Height 3: Partial direct writeback: Done
       "b 105f\n"
       "104:"  // Height 3: Full writeback
@@ -1298,39 +1296,39 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "105:"  // Height 3: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 72b\n"
       "b 212f\n"
       "106:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "107:"  // Height 4: Column loop
       "cbz x12, 108f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1338,111 +1336,111 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "108:"  // Height 4: no bias
       "tbz %x[flags], #0, 118f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x23, x24, x20, LSL #2\n"
       "bge 117f\n"
       "tbz x11, #3, 112f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
       "tbz x11, #2, 110f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
       "tbz x11, #1, 109f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x11, #0, 116f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
       "b 116f\n"
       "109:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 116f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
       "b 116f\n"
       "110:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x11, #1, 111f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x11, #0, 116f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
       "b 116f\n"
       "111:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 116f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
       "b 116f\n"
       "112:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x11, #2, 114f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "tbz x11, #1, 113f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x11, #0, 116f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
       "b 116f\n"
       "113:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 116f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
       "b 116f\n"
       "114:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x11, #1, 115f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x11, #0, 116f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
       "b 116f\n"
       "115:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
       "116:"  // Height 4: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 119f\n"
@@ -1451,18 +1449,18 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "b 119f\n"
       "118:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1485,8 +1483,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "mov x28, #0x0\n"
       "120:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 121f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1626,14 +1624,14 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       ".inst 0x4f40f0e9  // bfdot v9.4s, v7.8h, v0.h[0]\n"
       ".inst 0x4f41f0ed  // bfdot v13.4s, v7.8h, v1.h[0]\n"
       "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x4f42f0f1  // bfdot v17.4s, v7.8h, v2.h[0]\n"
       ".inst 0x4f43f0f5  // bfdot v21.4s, v7.8h, v3.h[0]\n"
       "ldr q24, [x10, #0x30]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f40f32a  // bfdot v10.4s, v25.8h, v0.h[0]\n"
       ".inst 0x4f41f32e  // bfdot v14.4s, v25.8h, v1.h[0]\n"
-      "sub x27, x27, #0x8\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f42f332  // bfdot v18.4s, v25.8h, v2.h[0]\n"
       ".inst 0x4f43f336  // bfdot v22.4s, v25.8h, v3.h[0]\n"
@@ -1711,9 +1709,9 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr s29, [x26], #0x4\n"
       "ldr s28, [x25], #0x4\n"
       "sub x27, x27, #0x2\n"
+      "cmp x27, #0x2\n"
       "ldr s27, [x24], #0x4\n"
       "ldr s26, [x23], #0x4\n"
-      "cmp x27, #0x2\n"
       "ldr q25, [x10, #0x0]\n"
       "ldr q24, [x10, #0x10]\n"
       ".inst 0x4f5df328  // bfdot v8.4s, v25.8h, v29.h[0]\n"
@@ -1770,17 +1768,17 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 120b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
-      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 130f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v25.4s }, [x21]\n"
       "ld1r { v24.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v25.4s\n"
       "fmin v9.4s, v9.4s, v25.4s\n"
@@ -1820,95 +1818,95 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "tbz x11, #3, 134f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
       "tbz x11, #2, 132f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
       "tbz x11, #1, 131f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
       "tbz x11, #0, 138f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
       "b 138f\n"
       "131:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x11, #0, 138f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
       "b 138f\n"
       "132:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x11, #1, 133f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
       "tbz x11, #0, 138f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
       "b 138f\n"
       "133:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x11, #0, 138f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
       "b 138f\n"
       "134:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x11, #2, 136f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
       "tbz x11, #1, 135f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
       "tbz x11, #0, 138f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
       "b 138f\n"
       "135:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x11, #0, 138f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
       "b 138f\n"
       "136:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x11, #1, 137f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x11, #0, 138f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
       "b 138f\n"
       "137:"  // Height 4: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
       "138:"  // Height 4: Partial direct writeback: Done
       "b 140f\n"
       "139:"  // Height 4: Full writeback
@@ -1917,43 +1915,43 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
       "140:"  // Height 4: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 107b\n"
       "b 212f\n"
       "141:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "142:"  // Height 5: Column loop
       "cbz x12, 143f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1965,128 +1963,128 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "143:"  // Height 5: no bias
       "tbz %x[flags], #0, 153f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 152f\n"
       "tbz x11, #3, 147f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x11, #2, 145f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
       "tbz x11, #1, 144f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x11, #0, 151f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
       "b 151f\n"
       "144:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 151f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
       "b 151f\n"
       "145:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x11, #1, 146f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x11, #0, 151f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
       "b 151f\n"
       "146:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 151f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
       "b 151f\n"
       "147:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x11, #2, 149f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "tbz x11, #1, 148f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x11, #0, 151f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 151f\n"
       "148:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 151f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "b 151f\n"
       "149:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x11, #1, 150f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x11, #0, 151f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
       "b 151f\n"
       "150:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
       "151:"  // Height 5: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 154f\n"
@@ -2095,22 +2093,22 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
       "b 154f\n"
       "153:"  // Height 5: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2137,8 +2135,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "mov x28, #0x0\n"
       "155:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 156f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2304,12 +2302,12 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "add x22, x22, #0x10\n"
       ".inst 0x4f41f0ed  // bfdot v13.4s, v7.8h, v1.h[0]\n"
       ".inst 0x4f42f0f1  // bfdot v17.4s, v7.8h, v2.h[0]\n"
+      "sub x27, x27, #0x8\n"
       "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f43f0f5  // bfdot v21.4s, v7.8h, v3.h[0]\n"
       ".inst 0x4f44f0f9  // bfdot v25.4s, v7.8h, v4.h[0]\n"
       "ldr q28, [x10, #0x30]\n"
-      "sub x27, x27, #0x8\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f40f3aa  // bfdot v10.4s, v29.8h, v0.h[0]\n"
       ".inst 0x4f41f3ae  // bfdot v14.4s, v29.8h, v1.h[0]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
@@ -2404,14 +2402,14 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr s2, [x26], #0x4\n"
       "ldr s1, [x25], #0x4\n"
       "sub x27, x27, #0x2\n"
+      "cmp x27, #0x2\n"
       "ldr s0, [x24], #0x4\n"
       "ldr s31, [x23], #0x4\n"
-      "cmp x27, #0x2\n"
       "ldr s30, [x22], #0x4\n"
       "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
       ".inst 0x4f42f3a8  // bfdot v8.4s, v29.8h, v2.h[0]\n"
       ".inst 0x4f41f3ac  // bfdot v12.4s, v29.8h, v1.h[0]\n"
+      "ldr q28, [x10, #0x10]\n"
       ".inst 0x4f40f3b0  // bfdot v16.4s, v29.8h, v0.h[0]\n"
       ".inst 0x4f5ff3b4  // bfdot v20.4s, v29.8h, v31.h[0]\n"
       ".inst 0x4f5ef3b8  // bfdot v24.4s, v29.8h, v30.h[0]\n"
@@ -2473,19 +2471,19 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 155b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
-      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 165f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v29.4s }, [x21]\n"
       "ld1r { v28.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v29.4s\n"
       "fmin v9.4s, v9.4s, v29.4s\n"
@@ -2533,111 +2531,111 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "tbz x11, #3, 169f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
       "tbz x11, #2, 167f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
       "tbz x11, #1, 166f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x11, #0, 173f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
       "b 173f\n"
       "166:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x11, #0, 173f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
       "b 173f\n"
       "167:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x11, #1, 168f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x11, #0, 173f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
       "b 173f\n"
       "168:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x11, #0, 173f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
       "b 173f\n"
       "169:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x11, #2, 171f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x11, #1, 170f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x11, #0, 173f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 173f\n"
       "170:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x11, #0, 173f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 173f\n"
       "171:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x11, #1, 172f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x11, #0, 173f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 173f\n"
       "172:"  // Height 5: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "173:"  // Height 5: Partial direct writeback: Done
       "b 175f\n"
       "174:"  // Height 5: Full writeback
@@ -2646,51 +2644,50 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "175:"  // Height 5: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 142b\n"
       "b 212f\n"
       "176:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "177:"  // Height 6: Column loop
       "cbz x12, 178f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -2706,145 +2703,145 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "178:"  // Height 6: no bias
       "tbz %x[flags], #0, 188f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 187f\n"
       "tbz x11, #3, 182f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x11, #2, 180f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x11, #1, 179f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x11, #0, 186f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 186f\n"
       "179:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 186f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 186f\n"
       "180:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x11, #1, 181f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x11, #0, 186f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 186f\n"
       "181:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 186f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 186f\n"
       "182:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x11, #2, 184f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x11, #1, 183f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x11, #0, 186f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 186f\n"
       "183:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 186f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 186f\n"
       "184:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x11, #1, 185f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x11, #0, 186f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 186f\n"
       "185:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "186:"  // Height 6: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 189f\n"
@@ -2853,26 +2850,26 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "b 189f\n"
       "188:"  // Height 6: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2903,8 +2900,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "mov x28, #0x0\n"
       "190:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 191f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -3094,18 +3091,18 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       ".inst 0x4f40f0e9  // bfdot v9.4s, v7.8h, v0.h[0]\n"
       ".inst 0x4f41f0ed  // bfdot v13.4s, v7.8h, v1.h[0]\n"
       "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x4f42f0f1  // bfdot v17.4s, v7.8h, v2.h[0]\n"
       ".inst 0x4f43f0f5  // bfdot v21.4s, v7.8h, v3.h[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f44f0f9  // bfdot v25.4s, v7.8h, v4.h[0]\n"
       ".inst 0x4f45f0fd  // bfdot v29.4s, v7.8h, v5.h[0]\n"
       "ldr q7, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f40f0ca  // bfdot v10.4s, v6.8h, v0.h[0]\n"
       ".inst 0x4f41f0ce  // bfdot v14.4s, v6.8h, v1.h[0]\n"
-      "sub x27, x27, #0x8\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4f42f0d2  // bfdot v18.4s, v6.8h, v2.h[0]\n"
       ".inst 0x4f43f0d6  // bfdot v22.4s, v6.8h, v3.h[0]\n"
@@ -3211,9 +3208,9 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "ldr s7, [x26], #0x4\n"
       "ldr s6, [x25], #0x4\n"
       "sub x27, x27, #0x2\n"
+      "cmp x27, #0x2\n"
       "ldr s5, [x24], #0x4\n"
       "ldr s4, [x23], #0x4\n"
-      "cmp x27, #0x2\n"
       "ldr s3, [x22], #0x4\n"
       "ldr s2, [x21], #0x4\n"
       "ldr q1, [x10, #0x0]\n"
@@ -3290,21 +3287,21 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 190b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "add x21, x22, x20, LSL #2\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbz %x[flags], #1, 200f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v1.4s\n"
       "fmin v9.4s, v9.4s, v1.4s\n"
@@ -3360,127 +3357,127 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "tbz x11, #3, 204f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
-      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
+      "st1 { v29.4s }, [x21], #0x10\n"
       "tbz x11, #2, 202f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
-      "st1 { v30.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v30.4s }, [x21], #0x10\n"
       "tbz x11, #1, 201f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x11, #0, 208f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
-      "st1 { v31.s }[2], [x22]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v31.s }[2], [x21]\n"
       "b 208f\n"
       "201:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x11, #0, 208f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
-      "str s31, [x22, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
+      "str s31, [x21, #0x0]\n"
       "b 208f\n"
       "202:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x11, #1, 203f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x11, #0, 208f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
-      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v30.s }[2], [x21]\n"
       "b 208f\n"
       "203:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x11, #0, 208f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
-      "str s30, [x22, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
+      "str s30, [x21, #0x0]\n"
       "b 208f\n"
       "204:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x11, #2, 206f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
       "tbz x11, #1, 205f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x11, #0, 208f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
-      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v29.s }[2], [x21]\n"
       "b 208f\n"
       "205:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x11, #0, 208f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
-      "str s29, [x22, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
+      "str s29, [x21, #0x0]\n"
       "b 208f\n"
       "206:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x11, #1, 207f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x11, #0, 208f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
-      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v28.s }[2], [x21]\n"
       "b 208f\n"
       "207:"  // Height 6: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
-      "str s28, [x22, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
+      "str s28, [x21, #0x0]\n"
       "208:"  // Height 6: Partial direct writeback: Done
       "b 210f\n"
       "209:"  // Height 6: Full writeback
@@ -3489,26 +3486,26 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q28, [x22, #0x0]\n"
-      "str q29, [x22, #0x10]\n"
-      "str q30, [x22, #0x20]\n"
-      "str q31, [x22, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
+      "str q28, [x21, #0x0]\n"
+      "str q29, [x21, #0x10]\n"
+      "str q30, [x21, #0x20]\n"
+      "str q31, [x21, #0x30]\n"
       "210:"  // Height 6: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 177b\n"
@@ -3524,8 +3521,8 @@ void a64_hybrid_bf16fp32_dot_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "212:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16.hpp
index ee57113f9b..e0a1ed62d9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp
index 47a85803d0..a804d5c102 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp
@@ -48,19 +48,18 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -81,7 +80,6 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -104,23 +102,23 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp %x[M], #0x2\n"
       "bgt 75f\n"
       "beq 38f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x12, 3f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "b 15f\n"
@@ -211,8 +209,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "16:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 17f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -232,12 +230,7 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q6, [x10, #0x10]\n"
       "blt 20f\n"
       "19:"  // Height 1: Multiply loop: Main loop head
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "trn1 v20.2d, v1.2d, v21.2d\n"
-      "trn2 v1.2d, v1.2d, v21.2d\n"
       ".inst 0x6e47ee88  // bfmmla v8.4s, v20.8h, v7.8h\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x6e46ee8c  // bfmmla v12.4s, v20.8h, v6.8h\n"
@@ -250,6 +243,7 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ee8e  // bfmmla v14.4s, v20.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v21.2d\n"
       ".inst 0x6e52ee8b  // bfmmla v11.4s, v20.8h, v18.8h\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x6e51ee8f  // bfmmla v15.4s, v20.8h, v17.8h\n"
@@ -266,19 +260,19 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x6e51ec2e  // bfmmla v14.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x8\n"
+      "add x26, x26, #0x10\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e52ec2b  // bfmmla v11.4s, v1.8h, v18.8h\n"
-      "ldr q7, [x10, #0x0]\n"
       ".inst 0x6e51ec2f  // bfmmla v15.4s, v1.8h, v17.8h\n"
       "ldr q1, [x26, #0x0]\n"
+      "add x10, x10, #0x100\n"
+      "ldr q7, [x10, #0x0]\n"
       "ldr q6, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 19b\n"
       "20:"  // Height 1: Multiply loop: Single iteration only
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "trn1 v19.2d, v1.2d, v17.2d\n"
-      "trn2 v1.2d, v1.2d, v17.2d\n"
+      "trn1 v19.2d, v1.2d, v20.2d\n"
       ".inst 0x6e47ee68  // bfmmla v8.4s, v19.8h, v7.8h\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x6e46ee6c  // bfmmla v12.4s, v19.8h, v6.8h\n"
@@ -290,16 +284,17 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e51ee6a  // bfmmla v10.4s, v19.8h, v17.8h\n"
       "ldr q17, [x10, #0x60]\n"
       ".inst 0x6e52ee6e  // bfmmla v14.4s, v19.8h, v18.8h\n"
-      "ldr q25, [x10, #0x70]\n"
+      "ldr q24, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v20.2d\n"
       ".inst 0x6e51ee6b  // bfmmla v11.4s, v19.8h, v17.8h\n"
       "ldr q17, [x10, #0x80]\n"
-      ".inst 0x6e59ee6f  // bfmmla v15.4s, v19.8h, v25.8h\n"
-      "ldr q3, [x10, #0x90]\n"
+      ".inst 0x6e58ee6f  // bfmmla v15.4s, v19.8h, v24.8h\n"
+      "ldr q2, [x10, #0x90]\n"
       ".inst 0x6e51ec28  // bfmmla v8.4s, v1.8h, v17.8h\n"
-      "ldr q19, [x10, #0xa0]\n"
-      ".inst 0x6e43ec2c  // bfmmla v12.4s, v1.8h, v3.8h\n"
+      "ldr q18, [x10, #0xa0]\n"
+      ".inst 0x6e42ec2c  // bfmmla v12.4s, v1.8h, v2.8h\n"
       "ldr q17, [x10, #0xb0]\n"
-      ".inst 0x6e53ec29  // bfmmla v9.4s, v1.8h, v19.8h\n"
+      ".inst 0x6e52ec29  // bfmmla v9.4s, v1.8h, v18.8h\n"
       "ldr q18, [x10, #0xc0]\n"
       ".inst 0x6e51ec2d  // bfmmla v13.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0xd0]\n"
@@ -307,21 +302,22 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x6e51ec2e  // bfmmla v14.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e52ec2b  // bfmmla v11.4s, v1.8h, v18.8h\n"
       ".inst 0x6e51ec2f  // bfmmla v15.4s, v1.8h, v17.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "21:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 26f\n"
       "cmp x27, #0x4\n"
       "blt 23f\n"
       "22:"  // Height 1: Multiply loop: Odd block loop
       "ldr d19, [x26], #0x8\n"
-      "ldr q20, [x10, #0x0]\n"
-      "sub x27, x27, #0x4\n"
+      "ldr q18, [x10, #0x0]\n"
+      "trn1 v19.2d, v19.2d, v17.2d\n"
       "ldr q17, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
-      "trn1 v19.2d, v19.2d, v18.2d\n"
-      ".inst 0x6e54ee68  // bfmmla v8.4s, v19.8h, v20.8h\n"
+      ".inst 0x6e52ee68  // bfmmla v8.4s, v19.8h, v18.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e51ee6c  // bfmmla v12.4s, v19.8h, v17.8h\n"
       "ldr q17, [x10, #0x30]\n"
@@ -333,9 +329,11 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ee6e  // bfmmla v14.4s, v19.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
+      "add x10, x10, #0x80\n"
       "bge 22b\n"
       "23:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 26f\n"
@@ -362,9 +360,9 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e42ee6e  // bfmmla v14.4s, v19.8h, v2.8h\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
+      "add x10, x10, #0x80\n"
       "26:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -376,9 +374,9 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "uzp1 v10.2d, v10.2d, v14.2d\n"
       "uzp1 v11.2d, v11.2d, v15.2d\n"
       "tbz %x[flags], #1, 27f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v18.4s }, [x21]\n"
       "ld1r { v17.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v18.4s\n"
       "fmin v9.4s, v9.4s, v18.4s\n"
@@ -448,23 +446,23 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "bgt 2b\n"
       "b 224f\n"
       "38:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "39:"  // Height 2: Column loop
       "cbz x12, 40f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "b 52f\n"
@@ -472,75 +470,75 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "tbz %x[flags], #0, 51f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "bge 49f\n"
       "tbz x11, #3, 44f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
       "tbz x11, #2, 42f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
       "tbz x11, #1, 41f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x11, #0, 48f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v15.s }[2], [x25]\n"
       "b 48f\n"
       "41:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 48f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
       "b 48f\n"
       "42:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x11, #1, 43f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x11, #0, 48f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v14.s }[2], [x25]\n"
       "b 48f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 48f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
       "b 48f\n"
       "44:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x11, #2, 46f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "tbz x11, #1, 45f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x11, #0, 48f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v13.s }[2], [x25]\n"
       "b 48f\n"
       "45:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 48f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
       "b 48f\n"
       "46:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x11, #1, 47f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x11, #0, 48f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v12.s }[2], [x25]\n"
       "b 48f\n"
       "47:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
       "48:"  // Height 2: Partial accumulate: Done
       "sub x9, x9, x20\n"
@@ -550,10 +548,10 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "50:"  // Height 2: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -577,8 +575,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "53:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 54f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -603,14 +601,6 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "blt 57f\n"
       "56:"  // Height 2: Multiply loop: Main loop head
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e47ee68  // bfmmla v8.4s, v19.8h, v7.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e46ee6c  // bfmmla v12.4s, v19.8h, v6.8h\n"
@@ -623,6 +613,7 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ee6e  // bfmmla v14.4s, v19.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
@@ -639,21 +630,22 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x6e51ec2e  // bfmmla v14.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x8\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "ldr q2, [x25, #0x0]\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e52ec2b  // bfmmla v11.4s, v1.8h, v18.8h\n"
+      "add x10, x10, #0x100\n"
       "ldr q7, [x10, #0x0]\n"
       ".inst 0x6e51ec2f  // bfmmla v15.4s, v1.8h, v17.8h\n"
       "ldr q1, [x26, #0x0]\n"
       "ldr q6, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "bge 56b\n"
       "57:"  // Height 2: Multiply loop: Single iteration only
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e47ee68  // bfmmla v8.4s, v19.8h, v7.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e46ee6c  // bfmmla v12.4s, v19.8h, v6.8h\n"
@@ -666,6 +658,7 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ee6e  // bfmmla v14.4s, v19.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
@@ -682,36 +675,41 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x6e51ec2e  // bfmmla v14.4s, v1.8h, v17.8h\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e52ec2b  // bfmmla v11.4s, v1.8h, v18.8h\n"
       ".inst 0x6e51ec2f  // bfmmla v15.4s, v1.8h, v17.8h\n"
+      "sub x27, x27, #0x8\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "58:"  // Height 2: Multiply loop: Main loop skip
       "cbz x27, 63f\n"
       "cmp x27, #0x4\n"
       "blt 60f\n"
       "59:"  // Height 2: Multiply loop: Odd block loop
-      "ldr d20, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d18, [x26], #0x8\n"
+      "ldr d17, [x25], #0x8\n"
+      "trn1 v19.2d, v18.2d, v17.2d\n"
       "sub x27, x27, #0x4\n"
       "ldr q18, [x10, #0x0]\n"
       "ldr q17, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
-      "trn1 v19.2d, v20.2d, v19.2d\n"
       ".inst 0x6e52ee68  // bfmmla v8.4s, v19.8h, v18.8h\n"
-      "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e51ee6c  // bfmmla v12.4s, v19.8h, v17.8h\n"
-      "ldr q6, [x10, #0x30]\n"
+      "ldr q26, [x10, #0x20]\n"
+      "ldr q5, [x10, #0x30]\n"
       ".inst 0x6e5aee69  // bfmmla v9.4s, v19.8h, v26.8h\n"
+      ".inst 0x6e45ee6d  // bfmmla v13.4s, v19.8h, v5.8h\n"
       "ldr q18, [x10, #0x40]\n"
-      ".inst 0x6e46ee6d  // bfmmla v13.4s, v19.8h, v6.8h\n"
       "ldr q17, [x10, #0x50]\n"
       ".inst 0x6e52ee6a  // bfmmla v10.4s, v19.8h, v18.8h\n"
-      "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ee6e  // bfmmla v14.4s, v19.8h, v17.8h\n"
+      "ldr q18, [x10, #0x60]\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
+      "add x10, x10, #0x80\n"
       "bge 59b\n"
       "60:"  // Height 2: Multiply loop: Skip odd blocks
       "cbz x27, 63f\n"
@@ -736,35 +734,35 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e52ee69  // bfmmla v9.4s, v19.8h, v18.8h\n"
       "ldr q30, [x10, #0x40]\n"
       ".inst 0x6e51ee6d  // bfmmla v13.4s, v19.8h, v17.8h\n"
-      "ldr q31, [x10, #0x50]\n"
+      "ldr q26, [x10, #0x50]\n"
       ".inst 0x6e5eee6a  // bfmmla v10.4s, v19.8h, v30.8h\n"
       "ldr q18, [x10, #0x60]\n"
-      ".inst 0x6e5fee6e  // bfmmla v14.4s, v19.8h, v31.8h\n"
+      ".inst 0x6e5aee6e  // bfmmla v14.4s, v19.8h, v26.8h\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e52ee6b  // bfmmla v11.4s, v19.8h, v18.8h\n"
       ".inst 0x6e51ee6f  // bfmmla v15.4s, v19.8h, v17.8h\n"
+      "add x10, x10, #0x80\n"
       "63:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 53b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v18.4s }, [x21]\n"
       "ld1r { v17.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v18.4s\n"
       "fmin v12.4s, v12.4s, v18.4s\n"
@@ -788,63 +786,63 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "tbz x11, #3, 68f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
       "tbz x11, #2, 66f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
       "tbz x11, #1, 65f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
+      "str d11, [x25], #0x8\n"
       "tbz x11, #0, 72f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
+      "st1 { v11.s }[2], [x25]\n"
       "b 72f\n"
       "65:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x11, #0, 72f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
       "b 72f\n"
       "66:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x11, #1, 67f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
+      "str d10, [x25], #0x8\n"
       "tbz x11, #0, 72f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
+      "st1 { v10.s }[2], [x25]\n"
       "b 72f\n"
       "67:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x11, #0, 72f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
       "b 72f\n"
       "68:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x11, #2, 70f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
       "tbz x11, #1, 69f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
+      "str d9, [x25], #0x8\n"
       "tbz x11, #0, 72f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
+      "st1 { v9.s }[2], [x25]\n"
       "b 72f\n"
       "69:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x11, #0, 72f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
       "b 72f\n"
       "70:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x11, #1, 71f\n"
       "str d7, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
+      "str d8, [x25], #0x8\n"
       "tbz x11, #0, 72f\n"
       "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
+      "st1 { v8.s }[2], [x25]\n"
       "b 72f\n"
       "71:"  // Height 2: Partial direct writeback: partial_1_0
       "str s7, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
       "72:"  // Height 2: Partial direct writeback: Done
       "b 74f\n"
       "73:"  // Height 2: Full writeback
@@ -853,32 +851,32 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
       "74:"  // Height 2: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 39b\n"
       "b 224f\n"
       "75:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "76:"  // Height 3: Column loop
       "cbz x12, 77f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -893,94 +891,94 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "77:"  // Height 3: no bias
       "tbz %x[flags], #0, 88f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "bge 86f\n"
       "tbz x11, #3, 81f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
       "tbz x11, #2, 79f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
       "tbz x11, #1, 78f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
       "tbz x11, #0, 85f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
       "b 85f\n"
       "78:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 85f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
       "b 85f\n"
       "79:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x11, #1, 80f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x11, #0, 85f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
       "b 85f\n"
       "80:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 85f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
       "b 85f\n"
       "81:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x11, #2, 83f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "tbz x11, #1, 82f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x11, #0, 85f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
       "b 85f\n"
       "82:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 85f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
       "b 85f\n"
       "83:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x11, #1, 84f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x11, #0, 85f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
       "b 85f\n"
       "84:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
       "85:"  // Height 3: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 87f\n"
@@ -989,14 +987,14 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
       "87:"  // Height 3: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -1036,8 +1034,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "90:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 91f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1067,38 +1065,35 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "93:"  // Height 3: Multiply loop: Main loop head
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
-      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
+      "trn1 v27.2d, v3.2d, v29.2d\n"
       ".inst 0x6e47ef70  // bfmmla v16.4s, v27.8h, v7.8h\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e46ef74  // bfmmla v20.4s, v27.8h, v6.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "trn2 v3.2d, v3.2d, v29.2d\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x90]\n"
       "ldr q2, [x25, #0x0]\n"
@@ -1106,12 +1101,15 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e5aec70  // bfmmla v16.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x6e59ec2c  // bfmmla v12.4s, v1.8h, v25.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e59ec74  // bfmmla v20.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x6e5aec29  // bfmmla v9.4s, v1.8h, v26.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e5aec71  // bfmmla v17.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x6e59ec2d  // bfmmla v13.4s, v1.8h, v25.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e59ec75  // bfmmla v21.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0xd0]\n"
       ".inst 0x6e5aec2a  // bfmmla v10.4s, v1.8h, v26.8h\n"
@@ -1133,43 +1131,43 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "94:"  // Height 3: Multiply loop: Single iteration only
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
-      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
+      "trn1 v27.2d, v3.2d, v29.2d\n"
       ".inst 0x6e47ef70  // bfmmla v16.4s, v27.8h, v7.8h\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e46ef74  // bfmmla v20.4s, v27.8h, v6.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "trn2 v3.2d, v3.2d, v29.2d\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x90]\n"
       ".inst 0x6e5aec28  // bfmmla v8.4s, v1.8h, v26.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e5aec70  // bfmmla v16.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x6e59ec2c  // bfmmla v12.4s, v1.8h, v25.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e59ec74  // bfmmla v20.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x6e5aec29  // bfmmla v9.4s, v1.8h, v26.8h\n"
@@ -1194,25 +1192,25 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp x27, #0x4\n"
       "blt 97f\n"
       "96:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "sub x27, x27, #0x4\n"
-      "ldr d27, [x24], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "trn1 v28.2d, v26.2d, v25.2d\n"
+      "ldr d25, [x24], #0x8\n"
       "ldr q26, [x10, #0x0]\n"
-      "cmp x27, #0x4\n"
-      "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v27.2d, v29.2d\n"
+      "trn1 v27.2d, v25.2d, v27.2d\n"
       ".inst 0x6e5aef88  // bfmmla v8.4s, v28.8h, v26.8h\n"
+      "ldr q25, [x10, #0x10]\n"
       ".inst 0x6e5aef70  // bfmmla v16.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e59ef8c  // bfmmla v12.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef74  // bfmmla v20.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "sub x27, x27, #0x4\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
@@ -1221,8 +1219,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
@@ -1248,9 +1246,9 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e5aef88  // bfmmla v8.4s, v28.8h, v26.8h\n"
-      ".inst 0x6e5def8c  // bfmmla v12.4s, v28.8h, v29.8h\n"
       ".inst 0x6e5aef70  // bfmmla v16.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e5def8c  // bfmmla v12.4s, v28.8h, v29.8h\n"
       ".inst 0x6e5def74  // bfmmla v20.4s, v27.8h, v29.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
@@ -1276,27 +1274,27 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 90b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp1 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v18.2d, v18.2d, v22.2d\n"
       "uzp1 v19.2d, v19.2d, v23.2d\n"
       "tbz %x[flags], #1, 101f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v26.4s }, [x21]\n"
       "ld1r { v25.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v26.4s\n"
       "fmin v12.4s, v12.4s, v26.4s\n"
@@ -1328,79 +1326,79 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "tbz x11, #3, 105f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
       "tbz x11, #2, 103f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
       "tbz x11, #1, 102f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x11, #0, 109f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
       "b 109f\n"
       "102:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x11, #0, 109f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
       "b 109f\n"
       "103:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x11, #1, 104f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x11, #0, 109f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
       "b 109f\n"
       "104:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x11, #0, 109f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
       "b 109f\n"
       "105:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x11, #2, 107f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
       "tbz x11, #1, 106f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x11, #0, 109f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
       "b 109f\n"
       "106:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x11, #0, 109f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
       "b 109f\n"
       "107:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x11, #1, 108f\n"
       "str d7, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x11, #0, 109f\n"
       "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
       "b 109f\n"
       "108:"  // Height 3: Partial direct writeback: partial_1_0
       "str s7, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
       "109:"  // Height 3: Partial direct writeback: Done
       "b 111f\n"
       "110:"  // Height 3: Full writeback
@@ -1409,36 +1407,36 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "111:"  // Height 3: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 76b\n"
       "b 224f\n"
       "112:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "113:"  // Height 4: Column loop
       "cbz x12, 114f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -1453,111 +1451,111 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "114:"  // Height 4: no bias
       "tbz %x[flags], #0, 125f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x23, x24, x20, LSL #2\n"
       "bge 123f\n"
       "tbz x11, #3, 118f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
       "tbz x11, #2, 116f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
       "tbz x11, #1, 115f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x11, #0, 122f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
       "b 122f\n"
       "115:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 122f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
       "b 122f\n"
       "116:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x11, #1, 117f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x11, #0, 122f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
       "b 122f\n"
       "117:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 122f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
       "b 122f\n"
       "118:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x11, #2, 120f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "tbz x11, #1, 119f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x11, #0, 122f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
       "b 122f\n"
       "119:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 122f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
       "b 122f\n"
       "120:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x11, #1, 121f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x11, #0, 122f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
       "b 122f\n"
       "121:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
       "122:"  // Height 4: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 124f\n"
@@ -1566,18 +1564,18 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "124:"  // Height 4: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -1617,8 +1615,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "127:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 128f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1652,38 +1650,33 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "130:"  // Height 4: Multiply loop: Main loop head
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "ldr q4, [x23, #0x0]\n"
-      ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
-      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e47ef70  // bfmmla v16.4s, v27.8h, v7.8h\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e46ef74  // bfmmla v20.4s, v27.8h, v6.8h\n"
       "ldr q25, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "add x23, x23, #0x10\n"
+      "ldr q4, [x23, #0x0]\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
@@ -1694,18 +1687,23 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e5aec70  // bfmmla v16.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x6e59ec2c  // bfmmla v12.4s, v1.8h, v25.8h\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e59ec74  // bfmmla v20.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x6e5aec29  // bfmmla v9.4s, v1.8h, v26.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e5aec71  // bfmmla v17.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x6e59ec2d  // bfmmla v13.4s, v1.8h, v25.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e59ec75  // bfmmla v21.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0xd0]\n"
       ".inst 0x6e5aec2a  // bfmmla v10.4s, v1.8h, v26.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e5aec72  // bfmmla v18.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0xe0]\n"
       ".inst 0x6e59ec2e  // bfmmla v14.4s, v1.8h, v25.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e59ec76  // bfmmla v22.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0xf0]\n"
       "add x10, x10, #0x100\n"
@@ -1721,48 +1719,48 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "131:"  // Height 4: Multiply loop: Single iteration only
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "sub x27, x27, #0x8\n"
-      ".inst 0x6e47ef88  // bfmmla v8.4s, v28.8h, v7.8h\n"
-      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e47ef70  // bfmmla v16.4s, v27.8h, v7.8h\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e46ef8c  // bfmmla v12.4s, v28.8h, v6.8h\n"
       ".inst 0x6e46ef74  // bfmmla v20.4s, v27.8h, v6.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aef89  // bfmmla v9.4s, v28.8h, v26.8h\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x6e5aef71  // bfmmla v17.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ef8d  // bfmmla v13.4s, v28.8h, v25.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e59ef75  // bfmmla v21.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e5aef8a  // bfmmla v10.4s, v28.8h, v26.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e5aef72  // bfmmla v18.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e59ef8e  // bfmmla v14.4s, v28.8h, v25.8h\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e59ef76  // bfmmla v22.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x6e5aef8b  // bfmmla v11.4s, v28.8h, v26.8h\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e5aef73  // bfmmla v19.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x6e59ef8f  // bfmmla v15.4s, v28.8h, v25.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e59ef77  // bfmmla v23.4s, v27.8h, v25.8h\n"
       "ldr q25, [x10, #0x90]\n"
       ".inst 0x6e5aec28  // bfmmla v8.4s, v1.8h, v26.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e5aec70  // bfmmla v16.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x6e59ec2c  // bfmmla v12.4s, v1.8h, v25.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e59ec74  // bfmmla v20.4s, v3.8h, v25.8h\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x6e5aec29  // bfmmla v9.4s, v1.8h, v26.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e5aec71  // bfmmla v17.4s, v3.8h, v26.8h\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x6e59ec2d  // bfmmla v13.4s, v1.8h, v25.8h\n"
@@ -1784,16 +1782,16 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp x27, #0x4\n"
       "blt 134f\n"
       "133:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "trn1 v28.2d, v26.2d, v25.2d\n"
       "sub x27, x27, #0x4\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d26, [x24], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "trn1 v27.2d, v26.2d, v25.2d\n"
       "cmp x27, #0x4\n"
       "ldr q26, [x10, #0x0]\n"
       "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v29.2d, v27.2d\n"
       ".inst 0x6e5aef88  // bfmmla v8.4s, v28.8h, v26.8h\n"
       ".inst 0x6e5aef70  // bfmmla v16.4s, v27.8h, v26.8h\n"
       "ldr q26, [x10, #0x20]\n"
@@ -1870,23 +1868,23 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 127b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
@@ -1894,9 +1892,9 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
       "tbz %x[flags], #1, 138f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v26.4s }, [x21]\n"
       "ld1r { v25.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v26.4s\n"
       "fmin v12.4s, v12.4s, v26.4s\n"
@@ -1936,95 +1934,95 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "tbz x11, #3, 142f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
       "tbz x11, #2, 140f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
       "tbz x11, #1, 139f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
       "tbz x11, #0, 146f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
       "b 146f\n"
       "139:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x11, #0, 146f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
       "b 146f\n"
       "140:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x11, #1, 141f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
       "tbz x11, #0, 146f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
       "b 146f\n"
       "141:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x11, #0, 146f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
       "b 146f\n"
       "142:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x11, #2, 144f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
       "tbz x11, #1, 143f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
       "tbz x11, #0, 146f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
       "b 146f\n"
       "143:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x11, #0, 146f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
       "b 146f\n"
       "144:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x11, #1, 145f\n"
       "str d7, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
       "tbz x11, #0, 146f\n"
       "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
       "b 146f\n"
       "145:"  // Height 4: Partial direct writeback: partial_1_0
       "str s7, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
       "146:"  // Height 4: Partial direct writeback: Done
       "b 148f\n"
       "147:"  // Height 4: Full writeback
@@ -2033,40 +2031,40 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
       "148:"  // Height 4: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 113b\n"
       "b 224f\n"
       "149:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "150:"  // Height 5: Column loop
       "cbz x12, 151f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -2089,128 +2087,128 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "151:"  // Height 5: no bias
       "tbz %x[flags], #0, 162f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 160f\n"
       "tbz x11, #3, 155f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
       "tbz x11, #2, 153f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v27.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v27.4s }, [x22], #0x10\n"
       "tbz x11, #1, 152f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d6, [x23], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d6, [x22], #0x8\n"
       "tbz x11, #0, 159f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v6.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v6.s }[2], [x22]\n"
       "b 159f\n"
       "152:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 159f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s6, [x23, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s6, [x22, #0x0]\n"
       "b 159f\n"
       "153:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x11, #1, 154f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x11, #0, 159f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
       "b 159f\n"
       "154:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 159f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
       "b 159f\n"
       "155:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x11, #2, 157f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x11, #1, 156f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x11, #0, 159f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
       "b 159f\n"
       "156:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 159f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
       "b 159f\n"
       "157:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x11, #1, 158f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x11, #0, 159f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 159f\n"
       "158:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "159:"  // Height 5: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 161f\n"
@@ -2219,22 +2217,22 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q25, [x23, #0x0]\n"
-      "ldr q26, [x23, #0x10]\n"
-      "ldr q27, [x23, #0x20]\n"
-      "ldr q6, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q25, [x22, #0x0]\n"
+      "ldr q26, [x22, #0x10]\n"
+      "ldr q27, [x22, #0x20]\n"
+      "ldr q6, [x22, #0x30]\n"
       "161:"  // Height 5: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -2290,8 +2288,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "164:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 165f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2328,51 +2326,51 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "167:"  // Height 5: Multiply loop: Main loop head
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x8\n"
+      ".inst 0x6e47ecc8  // bfmmla v8.4s, v6.8h, v7.8h\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
+      "sub x27, x27, #0x8\n"
       "trn1 v4.2d, v5.2d, v0.2d\n"
       "trn2 v5.2d, v5.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
-      ".inst 0x6e47ecc8  // bfmmla v8.4s, v6.8h, v7.8h\n"
-      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
-      "add x22, x22, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x6e40eccc  // bfmmla v12.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec54  // bfmmla v20.4s, v2.8h, v0.8h\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e40ec9c  // bfmmla v28.4s, v4.8h, v0.8h\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x6e47ecc9  // bfmmla v9.4s, v6.8h, v7.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e47ec51  // bfmmla v17.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec99  // bfmmla v25.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e40eccd  // bfmmla v13.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec55  // bfmmla v21.4s, v2.8h, v0.8h\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e40ec9d  // bfmmla v29.4s, v4.8h, v0.8h\n"
       "ldr q0, [x10, #0x50]\n"
       ".inst 0x6e47ecca  // bfmmla v10.4s, v6.8h, v7.8h\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e47ec52  // bfmmla v18.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9a  // bfmmla v26.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e40ecce  // bfmmla v14.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec56  // bfmmla v22.4s, v2.8h, v0.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e40ec9e  // bfmmla v30.4s, v4.8h, v0.8h\n"
       "ldr q0, [x10, #0x70]\n"
       ".inst 0x6e47eccb  // bfmmla v11.4s, v6.8h, v7.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e47ec53  // bfmmla v19.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9b  // bfmmla v27.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e40eccf  // bfmmla v15.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec57  // bfmmla v23.4s, v2.8h, v0.8h\n"
       "ldr q2, [x25, #0x0]\n"
@@ -2418,47 +2416,47 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "168:"  // Height 5: Multiply loop: Single iteration only
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
+      ".inst 0x6e47ecc8  // bfmmla v8.4s, v6.8h, v7.8h\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v0.2d\n"
       "trn2 v5.2d, v5.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
-      ".inst 0x6e47ecc8  // bfmmla v8.4s, v6.8h, v7.8h\n"
-      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x6e40eccc  // bfmmla v12.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec54  // bfmmla v20.4s, v2.8h, v0.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e40ec9c  // bfmmla v28.4s, v4.8h, v0.8h\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x6e47ecc9  // bfmmla v9.4s, v6.8h, v7.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e47ec51  // bfmmla v17.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec99  // bfmmla v25.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e40eccd  // bfmmla v13.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec55  // bfmmla v21.4s, v2.8h, v0.8h\n"
+      "add x22, x22, #0x10\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e40ec9d  // bfmmla v29.4s, v4.8h, v0.8h\n"
       "ldr q0, [x10, #0x50]\n"
       ".inst 0x6e47ecca  // bfmmla v10.4s, v6.8h, v7.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e47ec52  // bfmmla v18.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9a  // bfmmla v26.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e40ecce  // bfmmla v14.4s, v6.8h, v0.8h\n"
       ".inst 0x6e40ec56  // bfmmla v22.4s, v2.8h, v0.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e40ec9e  // bfmmla v30.4s, v4.8h, v0.8h\n"
       "ldr q0, [x10, #0x70]\n"
       ".inst 0x6e47eccb  // bfmmla v11.4s, v6.8h, v7.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e47ec53  // bfmmla v19.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9b  // bfmmla v27.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x80]\n"
@@ -2502,24 +2500,24 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp x27, #0x4\n"
       "blt 171f\n"
       "170:"  // Height 5: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x4\n"
-      "ldr d3, [x24], #0x8\n"
-      "ldr d2, [x23], #0x8\n"
-      "cmp x27, #0x4\n"
       "ldr d0, [x22], #0x8\n"
       "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v3.2d, v2.2d\n"
-      "trn1 v2.2d, v0.2d, v5.2d\n"
-      "ldr q0, [x10, #0x10]\n"
+      "trn1 v2.2d, v0.2d, v2.2d\n"
       ".inst 0x6e41ec88  // bfmmla v8.4s, v4.8h, v1.8h\n"
+      "ldr q0, [x10, #0x10]\n"
       ".inst 0x6e41ec70  // bfmmla v16.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec58  // bfmmla v24.4s, v2.8h, v1.8h\n"
       "ldr q1, [x10, #0x20]\n"
       ".inst 0x6e40ec8c  // bfmmla v12.4s, v4.8h, v0.8h\n"
       ".inst 0x6e40ec74  // bfmmla v20.4s, v3.8h, v0.8h\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6e40ec5c  // bfmmla v28.4s, v2.8h, v0.8h\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x6e41ec89  // bfmmla v9.4s, v4.8h, v1.8h\n"
@@ -2538,8 +2536,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e40ec76  // bfmmla v22.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec5e  // bfmmla v30.4s, v2.8h, v0.8h\n"
       "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e46ec8b  // bfmmla v11.4s, v4.8h, v6.8h\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x6e46ec73  // bfmmla v19.4s, v3.8h, v6.8h\n"
       ".inst 0x6e46ec5b  // bfmmla v27.4s, v2.8h, v6.8h\n"
       ".inst 0x6e40ec8f  // bfmmla v15.4s, v4.8h, v0.8h\n"
@@ -2610,27 +2608,27 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 164b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
@@ -2640,9 +2638,9 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "uzp1 v26.2d, v26.2d, v30.2d\n"
       "uzp1 v27.2d, v27.2d, v31.2d\n"
       "tbz %x[flags], #1, 175f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v1.4s\n"
       "fmin v12.4s, v12.4s, v1.4s\n"
@@ -2690,111 +2688,111 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "tbz x11, #3, 179f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
       "tbz x11, #2, 177f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
       "tbz x11, #1, 176f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x11, #0, 183f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
       "b 183f\n"
       "176:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x11, #0, 183f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
       "b 183f\n"
       "177:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x11, #1, 178f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x11, #0, 183f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
       "b 183f\n"
       "178:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x11, #0, 183f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
       "b 183f\n"
       "179:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x11, #2, 181f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x11, #1, 180f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x11, #0, 183f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 183f\n"
       "180:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x11, #0, 183f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 183f\n"
       "181:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x11, #1, 182f\n"
       "str d7, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x11, #0, 183f\n"
       "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 183f\n"
       "182:"  // Height 5: Partial direct writeback: partial_1_0
       "str s7, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "183:"  // Height 5: Partial direct writeback: Done
       "b 185f\n"
       "184:"  // Height 5: Full writeback
@@ -2803,48 +2801,47 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "185:"  // Height 5: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 150b\n"
       "b 224f\n"
       "186:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "187:"  // Height 6: Column loop
       "cbz x12, 188f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -2867,145 +2864,145 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "188:"  // Height 6: no bias
       "tbz %x[flags], #0, 199f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 197f\n"
       "tbz x11, #3, 192f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x11, #2, 190f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v27.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x11, #1, 189f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d6, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d6, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x11, #0, 196f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v6.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v6.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 196f\n"
       "189:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 196f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s6, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s6, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 196f\n"
       "190:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x11, #1, 191f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x11, #0, 196f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 196f\n"
       "191:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 196f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 196f\n"
       "192:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x11, #2, 194f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x11, #1, 193f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x11, #0, 196f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 196f\n"
       "193:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 196f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 196f\n"
       "194:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x11, #1, 195f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x11, #0, 196f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 196f\n"
       "195:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "196:"  // Height 6: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 198f\n"
@@ -3014,26 +3011,26 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q25, [x23, #0x0]\n"
-      "ldr q26, [x23, #0x10]\n"
-      "ldr q27, [x23, #0x20]\n"
-      "ldr q6, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q25, [x22, #0x0]\n"
+      "ldr q26, [x22, #0x10]\n"
+      "ldr q27, [x22, #0x20]\n"
+      "ldr q6, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "198:"  // Height 6: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -3089,8 +3086,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "201:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 202f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -3131,56 +3128,56 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "204:"  // Height 6: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
       "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
       "ldr q6, [x10, #0x10]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x6e46ec0c  // bfmmla v12.4s, v0.8h, v6.8h\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e46ec54  // bfmmla v20.4s, v2.8h, v6.8h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e46ec9c  // bfmmla v28.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e47ec09  // bfmmla v9.4s, v0.8h, v7.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e47ec51  // bfmmla v17.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec99  // bfmmla v25.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e46ec0d  // bfmmla v13.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
+      "add x22, x22, #0x10\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x6e46ec9d  // bfmmla v29.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x50]\n"
       ".inst 0x6e47ec0a  // bfmmla v10.4s, v0.8h, v7.8h\n"
+      "cmp x27, #0x10\n"
       ".inst 0x6e47ec52  // bfmmla v18.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9a  // bfmmla v26.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e46ec0e  // bfmmla v14.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec56  // bfmmla v22.4s, v2.8h, v6.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e46ec9e  // bfmmla v30.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x70]\n"
       ".inst 0x6e47ec0b  // bfmmla v11.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e47ec53  // bfmmla v19.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9b  // bfmmla v27.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e46ec0f  // bfmmla v15.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec57  // bfmmla v23.4s, v2.8h, v6.8h\n"
       "ldr q2, [x25, #0x0]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e46ec9f  // bfmmla v31.4s, v4.8h, v6.8h\n"
       "ldr q0, [x10, #0x90]\n"
       "ldr q4, [x23, #0x0]\n"
@@ -3224,52 +3221,52 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "205:"  // Height 6: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
+      "add x25, x25, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
       "ldr q6, [x10, #0x10]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x6e46ec0c  // bfmmla v12.4s, v0.8h, v6.8h\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e46ec54  // bfmmla v20.4s, v2.8h, v6.8h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e46ec9c  // bfmmla v28.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e47ec09  // bfmmla v9.4s, v0.8h, v7.8h\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e47ec51  // bfmmla v17.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec99  // bfmmla v25.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e46ec0d  // bfmmla v13.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
+      "add x21, x21, #0x10\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e46ec9d  // bfmmla v29.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x50]\n"
       ".inst 0x6e47ec0a  // bfmmla v10.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e47ec52  // bfmmla v18.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9a  // bfmmla v26.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e46ec0e  // bfmmla v14.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec56  // bfmmla v22.4s, v2.8h, v6.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e46ec9e  // bfmmla v30.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x70]\n"
       ".inst 0x6e47ec0b  // bfmmla v11.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e47ec53  // bfmmla v19.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9b  // bfmmla v27.4s, v4.8h, v7.8h\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e46ec0f  // bfmmla v15.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec57  // bfmmla v23.4s, v2.8h, v6.8h\n"
       ".inst 0x6e46ec9f  // bfmmla v31.4s, v4.8h, v6.8h\n"
@@ -3310,18 +3307,18 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp x27, #0x4\n"
       "blt 208f\n"
       "207:"  // Height 6: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x4\n"
-      "ldr d5, [x24], #0x8\n"
-      "ldr d3, [x23], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "cmp x27, #0x4\n"
-      "ldr d2, [x22], #0x8\n"
+      "ldr d1, [x22], #0x8\n"
       "ldr d0, [x21], #0x8\n"
+      "trn1 v2.2d, v1.2d, v0.2d\n"
       "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v5.2d, v3.2d\n"
-      "trn1 v2.2d, v2.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
       ".inst 0x6e41ec88  // bfmmla v8.4s, v4.8h, v1.8h\n"
       ".inst 0x6e41ec70  // bfmmla v16.4s, v3.8h, v1.8h\n"
@@ -3383,9 +3380,9 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "ldr q0, [x10, #0x0]\n"
       "trn1 v7.2d, v1.2d, v2.2d\n"
       "trn1 v3.2d, v3.2d, v4.2d\n"
+      ".inst 0x6e40ece8  // bfmmla v8.4s, v7.8h, v0.8h\n"
       "trn1 v2.2d, v5.2d, v6.2d\n"
       "ldr q1, [x10, #0x10]\n"
-      ".inst 0x6e40ece8  // bfmmla v8.4s, v7.8h, v0.8h\n"
       ".inst 0x6e40ec70  // bfmmla v16.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec58  // bfmmla v24.4s, v2.8h, v0.8h\n"
       "ldr q0, [x10, #0x20]\n"
@@ -3409,8 +3406,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       ".inst 0x6e41ec76  // bfmmla v22.4s, v3.8h, v1.8h\n"
       ".inst 0x6e41ec5e  // bfmmla v30.4s, v2.8h, v1.8h\n"
       "ldr q6, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e40eceb  // bfmmla v11.4s, v7.8h, v0.8h\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x6e40ec73  // bfmmla v19.4s, v3.8h, v0.8h\n"
       ".inst 0x6e40ec5b  // bfmmla v27.4s, v2.8h, v0.8h\n"
       ".inst 0x6e46ecef  // bfmmla v15.4s, v7.8h, v6.8h\n"
@@ -3422,31 +3419,31 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 201b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
       "uzp1 v23.2d, v24.2d, v28.2d\n"
@@ -3458,9 +3455,9 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "uzp1 v30.2d, v27.2d, v31.2d\n"
       "uzp2 v27.2d, v27.2d, v31.2d\n"
       "tbz %x[flags], #1, 212f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v7.4s, v7.4s, v1.4s\n"
       "fmin v12.4s, v12.4s, v1.4s\n"
@@ -3516,127 +3513,127 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "tbz x11, #3, 216f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x23], #0x10\n"
-      "st1 { v24.4s }, [x22], #0x10\n"
-      "st1 { v25.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v24.4s }, [x21], #0x10\n"
+      "st1 { v25.4s }, [x21], #0x10\n"
       "tbz x11, #2, 214f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v29.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
       "tbz x11, #1, 213f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
-      "str d27, [x22], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d30, [x22], #0x8\n"
+      "str d27, [x21], #0x8\n"
       "tbz x11, #0, 220f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
-      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v27.s }[2], [x21]\n"
       "b 220f\n"
       "213:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x11, #0, 220f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
-      "str s27, [x22, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s30, [x22, #0x0]\n"
+      "str s27, [x21, #0x0]\n"
       "b 220f\n"
       "214:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x11, #1, 215f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d29, [x23], #0x8\n"
-      "str d26, [x22], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d29, [x22], #0x8\n"
+      "str d26, [x21], #0x8\n"
       "tbz x11, #0, 220f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v29.s }[2], [x23]\n"
-      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v26.s }[2], [x21]\n"
       "b 220f\n"
       "215:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x11, #0, 220f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s29, [x23, #0x0]\n"
-      "str s26, [x22, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s29, [x22, #0x0]\n"
+      "str s26, [x21, #0x0]\n"
       "b 220f\n"
       "216:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x11, #2, 218f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
-      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v24.4s }, [x21], #0x10\n"
       "tbz x11, #1, 217f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d28, [x23], #0x8\n"
-      "str d25, [x22], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d28, [x22], #0x8\n"
+      "str d25, [x21], #0x8\n"
       "tbz x11, #0, 220f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v28.s }[2], [x23]\n"
-      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v25.s }[2], [x21]\n"
       "b 220f\n"
       "217:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x11, #0, 220f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s28, [x23, #0x0]\n"
-      "str s25, [x22, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s28, [x22, #0x0]\n"
+      "str s25, [x21, #0x0]\n"
       "b 220f\n"
       "218:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x11, #1, 219f\n"
       "str d7, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
+      "str d24, [x21], #0x8\n"
       "tbz x11, #0, 220f\n"
       "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
-      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
+      "st1 { v24.s }[2], [x21]\n"
       "b 220f\n"
       "219:"  // Height 6: Partial direct writeback: partial_1_0
       "str s7, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
-      "str s24, [x22, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
+      "str s24, [x21, #0x0]\n"
       "220:"  // Height 6: Partial direct writeback: Done
       "b 222f\n"
       "221:"  // Height 6: Full writeback
@@ -3645,26 +3642,26 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
-      "str q23, [x23, #0x0]\n"
-      "str q28, [x23, #0x10]\n"
-      "str q29, [x23, #0x20]\n"
-      "str q30, [x23, #0x30]\n"
-      "str q24, [x22, #0x0]\n"
-      "str q25, [x22, #0x10]\n"
-      "str q26, [x22, #0x20]\n"
-      "str q27, [x22, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
+      "str q23, [x22, #0x0]\n"
+      "str q28, [x22, #0x10]\n"
+      "str q29, [x22, #0x20]\n"
+      "str q30, [x22, #0x30]\n"
+      "str q24, [x21, #0x0]\n"
+      "str q25, [x21, #0x10]\n"
+      "str q26, [x21, #0x20]\n"
+      "str q27, [x21, #0x30]\n"
       "222:"  // Height 6: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 187b\n"
@@ -3680,8 +3677,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "224:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32.hpp
index 12244a2e99..d012e992e6 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 32, 1> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 32, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp
index aae6322b59..faf34f7fc0 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp
@@ -47,19 +47,18 @@ void a64_hybrid_fp16_mla_6x32_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const __fp16 *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void a64_hybrid_fp16_mla_6x32_a55 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -103,10 +101,10 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "cmp %x[M], #0x2\n"
       "bgt 99f\n"
       "beq 50f\n"
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x7, 3f\n"
       "ldr q8, [x7, #0x0]\n"
@@ -245,8 +243,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "mov x15, #0x0\n"
       "23:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 24f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -268,153 +266,150 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "26:"  // Height 1: Multiply loop: Main loop head
       "fmla v8.8h, v6.8h, v0.h[0]\n"
       "ldr d17, [x17, #0x20]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr x20, [x17, #0x28]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "ldr d16, [x17, #0x30]\n"
-      "add x13, x13, #0x10\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x38]\n"
-      "sub x14, x14, #0x8\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x48]\n"
-      "ldr x22, [x13, #0x8]\n"
-      "cmp x14, #0x10\n"
       "mov v16.d[1], x20\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "ldr d17, [x17, #0x40]\n"
+      "ldr x20, [x17, #0x48]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
       "ldr d16, [x17, #0x50]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x58]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x68]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[1]\n"
       "ldr d17, [x17, #0x60]\n"
+      "ldr x20, [x17, #0x68]\n"
       "fmla v9.8h, v16.8h, v0.h[1]\n"
       "ldr d16, [x17, #0x70]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x78]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x88]\n"
       "mov v16.d[1], x20\n"
       "fmla v10.8h, v17.8h, v0.h[1]\n"
       "ldr d17, [x17, #0x80]\n"
+      "ldr x20, [x17, #0x88]\n"
       "fmla v11.8h, v16.8h, v0.h[1]\n"
       "ldr d16, [x17, #0x90]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x98]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xa8]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[2]\n"
       "ldr d17, [x17, #0xa0]\n"
+      "ldr x20, [x17, #0xa8]\n"
       "fmla v9.8h, v16.8h, v0.h[2]\n"
       "ldr d16, [x17, #0xb0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0xb8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xc8]\n"
       "mov v16.d[1], x20\n"
       "fmla v10.8h, v17.8h, v0.h[2]\n"
       "ldr d17, [x17, #0xc0]\n"
+      "ldr x20, [x17, #0xc8]\n"
       "fmla v11.8h, v16.8h, v0.h[2]\n"
       "ldr d16, [x17, #0xd0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0xd8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xe8]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[3]\n"
       "ldr d17, [x17, #0xe0]\n"
+      "ldr x20, [x17, #0xe8]\n"
       "fmla v9.8h, v16.8h, v0.h[3]\n"
       "ldr d16, [x17, #0xf0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0xf8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x108]\n"
       "mov v16.d[1], x20\n"
       "fmla v10.8h, v17.8h, v0.h[3]\n"
       "ldr d17, [x17, #0x100]\n"
+      "ldr x20, [x17, #0x108]\n"
       "fmla v11.8h, v16.8h, v0.h[3]\n"
       "ldr d16, [x17, #0x110]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x118]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x128]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[4]\n"
       "ldr d17, [x17, #0x120]\n"
+      "ldr x20, [x17, #0x128]\n"
       "fmla v9.8h, v16.8h, v0.h[4]\n"
       "ldr d16, [x17, #0x130]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x138]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x148]\n"
       "mov v16.d[1], x20\n"
       "fmla v10.8h, v17.8h, v0.h[4]\n"
       "ldr d17, [x17, #0x140]\n"
+      "ldr x20, [x17, #0x148]\n"
       "fmla v11.8h, v16.8h, v0.h[4]\n"
       "ldr d16, [x17, #0x150]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x158]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x168]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[5]\n"
       "ldr d17, [x17, #0x160]\n"
+      "ldr x20, [x17, #0x168]\n"
       "fmla v9.8h, v16.8h, v0.h[5]\n"
       "ldr d16, [x17, #0x170]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x178]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x188]\n"
       "mov v16.d[1], x20\n"
       "fmla v10.8h, v17.8h, v0.h[5]\n"
       "ldr d17, [x17, #0x180]\n"
+      "ldr x20, [x17, #0x188]\n"
       "fmla v11.8h, v16.8h, v0.h[5]\n"
       "ldr d16, [x17, #0x190]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x198]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x1a8]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[6]\n"
       "ldr d17, [x17, #0x1a0]\n"
+      "ldr x20, [x17, #0x1a8]\n"
       "fmla v9.8h, v16.8h, v0.h[6]\n"
       "ldr d16, [x17, #0x1b0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x1b8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x1c8]\n"
       "mov v16.d[1], x20\n"
       "fmla v10.8h, v17.8h, v0.h[6]\n"
       "ldr d17, [x17, #0x1c0]\n"
+      "ldr x20, [x17, #0x1c8]\n"
       "fmla v11.8h, v16.8h, v0.h[6]\n"
       "ldr d16, [x17, #0x1d0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x1d8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x1e8]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[7]\n"
       "ldr d17, [x17, #0x1e0]\n"
+      "ldr x20, [x17, #0x1e8]\n"
       "fmla v9.8h, v16.8h, v0.h[7]\n"
       "ldr d16, [x17, #0x1f0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x1f8]\n"
-      "add x17, x17, #0x200\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x8]\n"
       "mov v16.d[1], x20\n"
+      "add x13, x13, #0x10\n"
+      "add x17, x17, #0x200\n"
       "fmla v10.8h, v17.8h, v0.h[7]\n"
       "ldr d6, [x17, #0x0]\n"
+      "ldr x20, [x17, #0x8]\n"
       "fmla v11.8h, v16.8h, v0.h[7]\n"
       "ldr d0, [x13, #0x0]\n"
+      "sub x14, x14, #0x8\n"
       "ldr d7, [x17, #0x10]\n"
+      "cmp x14, #0x10\n"
+      "ldr x21, [x13, #0x8]\n"
+      "mov v6.d[1], x20\n"
       "ldr x20, [x17, #0x18]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x22\n"
+      "mov v0.d[1], x21\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "bge 26b\n"
       "27:"  // Height 1: Multiply loop: Single iteration only
       "fmla v8.8h, v6.8h, v0.h[0]\n"
       "ldr q17, [x17, #0x20]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "ldr q16, [x17, #0x30]\n"
-      "add x13, x13, #0x10\n"
-      "sub x14, x14, #0x8\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "ldr q17, [x17, #0x40]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
       "ldr q16, [x17, #0x50]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v8.8h, v17.8h, v0.h[1]\n"
       "ldr q17, [x17, #0x60]\n"
       "fmla v9.8h, v16.8h, v0.h[1]\n"
@@ -467,23 +462,26 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr q17, [x17, #0x1e0]\n"
       "fmla v9.8h, v16.8h, v0.h[7]\n"
       "ldr q16, [x17, #0x1f0]\n"
-      "add x17, x17, #0x200\n"
+      "add x13, x13, #0x10\n"
+      "sub x14, x14, #0x8\n"
       "fmla v10.8h, v17.8h, v0.h[7]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v11.8h, v16.8h, v0.h[7]\n"
+      "add x17, x17, #0x200\n"
       "28:"  // Height 1: Multiply loop: Main loop skip
       "cbz x14, 30f\n"
       "29:"  // Height 1: Multiply loop: Odd block loop
       "ldr h0, [x13], #0x2\n"
       "sub x14, x14, #0x1\n"
-      "ldr q17, [x17, #0x0]\n"
+      "ldr q16, [x17, #0x0]\n"
+      "fmla v8.8h, v16.8h, v0.h[0]\n"
       "ldr q16, [x17, #0x10]\n"
-      "fmla v8.8h, v17.8h, v0.h[0]\n"
-      "ldr q17, [x17, #0x20]\n"
       "fmla v9.8h, v16.8h, v0.h[0]\n"
+      "ldr q16, [x17, #0x20]\n"
+      "fmla v10.8h, v16.8h, v0.h[0]\n"
       "ldr q16, [x17, #0x30]\n"
-      "add x17, x17, #0x40\n"
-      "fmla v10.8h, v17.8h, v0.h[0]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
+      "add x17, x17, #0x40\n"
       "cbnz x14, 29b\n"
       "30:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -492,14 +490,14 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "bne 23b\n"
       "prfm pstl1keep, [x16, #0x0]\n"
       "tbz %x[flags], #1, 31f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.8h }, [x20]\n"
+      "fmin v8.8h, v8.8h, v16.8h\n"
+      "fmin v9.8h, v9.8h, v16.8h\n"
+      "fmin v10.8h, v10.8h, v16.8h\n"
+      "fmin v11.8h, v11.8h, v16.8h\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.8h }, [x21]\n"
       "ld1r { v16.8h }, [x20]\n"
-      "fmin v8.8h, v8.8h, v17.8h\n"
-      "fmin v9.8h, v9.8h, v17.8h\n"
-      "fmin v10.8h, v10.8h, v17.8h\n"
-      "fmin v11.8h, v11.8h, v17.8h\n"
       "fmax v8.8h, v8.8h, v16.8h\n"
       "fmax v9.8h, v9.8h, v16.8h\n"
       "fmax v10.8h, v10.8h, v16.8h\n"
@@ -612,168 +610,168 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "bgt 2b\n"
       "b 296f\n"
       "50:"  // Height 2
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "51:"  // Height 2: Column loop
       "cbz x7, 52f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
+      "add x7, x7, #0x40\n"
       "b 71f\n"
       "52:"  // Height 2: no bias
       "tbz %x[flags], #0, 70f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x8, #0x20\n"
-      "add x26, x16, x20, LSL #1\n"
+      "add x25, x16, x20, LSL #1\n"
       "bge 69f\n"
       "tbz x8, #4, 60f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
       "ld1 { v9.8h }, [x16], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
       "tbz x8, #3, 56f\n"
       "ld1 { v10.8h }, [x16], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
       "tbz x8, #2, 54f\n"
       "ldr d11, [x16], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "tbz x8, #1, 53f\n"
       "ld1 { v11.s }[2], [x16], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v11.h }[6], [x16]\n"
-      "ld1 { v15.h }[6], [x26]\n"
+      "ld1 { v15.h }[6], [x25]\n"
       "b 68f\n"
       "53:"  // Height 2: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v11.h }[4], [x16]\n"
-      "ld1 { v15.h }[4], [x26]\n"
+      "ld1 { v15.h }[4], [x25]\n"
       "b 68f\n"
       "54:"  // Height 2: Partial accumulate: partial_2_24
       "tbz x8, #1, 55f\n"
       "ldr s11, [x16], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v11.h }[2], [x16]\n"
-      "ld1 { v15.h }[2], [x26]\n"
+      "ld1 { v15.h }[2], [x25]\n"
       "b 68f\n"
       "55:"  // Height 2: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x8, #0, 68f\n"
       "ldr h11, [x16, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
       "b 68f\n"
       "56:"  // Height 2: Partial accumulate: partial_4_16
       "tbz x8, #2, 58f\n"
       "ldr d10, [x16], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "tbz x8, #1, 57f\n"
       "ld1 { v10.s }[2], [x16], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v10.h }[6], [x16]\n"
-      "ld1 { v14.h }[6], [x26]\n"
+      "ld1 { v14.h }[6], [x25]\n"
       "b 68f\n"
       "57:"  // Height 2: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v10.h }[4], [x16]\n"
-      "ld1 { v14.h }[4], [x26]\n"
+      "ld1 { v14.h }[4], [x25]\n"
       "b 68f\n"
       "58:"  // Height 2: Partial accumulate: partial_2_16
       "tbz x8, #1, 59f\n"
       "ldr s10, [x16], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v10.h }[2], [x16]\n"
-      "ld1 { v14.h }[2], [x26]\n"
+      "ld1 { v14.h }[2], [x25]\n"
       "b 68f\n"
       "59:"  // Height 2: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x8, #0, 68f\n"
       "ldr h10, [x16, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
       "b 68f\n"
       "60:"  // Height 2: Partial accumulate: partial_8_0
       "tbz x8, #3, 64f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
       "tbz x8, #2, 62f\n"
       "ldr d9, [x16], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "tbz x8, #1, 61f\n"
       "ld1 { v9.s }[2], [x16], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v9.h }[6], [x16]\n"
-      "ld1 { v13.h }[6], [x26]\n"
+      "ld1 { v13.h }[6], [x25]\n"
       "b 68f\n"
       "61:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v9.h }[4], [x16]\n"
-      "ld1 { v13.h }[4], [x26]\n"
+      "ld1 { v13.h }[4], [x25]\n"
       "b 68f\n"
       "62:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x8, #1, 63f\n"
       "ldr s9, [x16], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v9.h }[2], [x16]\n"
-      "ld1 { v13.h }[2], [x26]\n"
+      "ld1 { v13.h }[2], [x25]\n"
       "b 68f\n"
       "63:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x8, #0, 68f\n"
       "ldr h9, [x16, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
       "b 68f\n"
       "64:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x8, #2, 66f\n"
       "ldr d8, [x16], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "tbz x8, #1, 65f\n"
       "ld1 { v8.s }[2], [x16], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v8.h }[6], [x16]\n"
-      "ld1 { v12.h }[6], [x26]\n"
+      "ld1 { v12.h }[6], [x25]\n"
       "b 68f\n"
       "65:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v8.h }[4], [x16]\n"
-      "ld1 { v12.h }[4], [x26]\n"
+      "ld1 { v12.h }[4], [x25]\n"
       "b 68f\n"
       "66:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x8, #1, 67f\n"
       "ldr s8, [x16], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "tbz x8, #0, 68f\n"
       "ld1 { v8.h }[2], [x16]\n"
-      "ld1 { v12.h }[2], [x26]\n"
+      "ld1 { v12.h }[2], [x25]\n"
       "b 68f\n"
       "67:"  // Height 2: Partial accumulate: partial_1_0
       "ldr h8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "68:"  // Height 2: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 71f\n"
@@ -782,10 +780,10 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "b 71f\n"
       "70:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -800,8 +798,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "mov x15, #0x0\n"
       "72:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 73f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -826,178 +824,178 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "blt 76f\n"
       "75:"  // Height 2: Multiply loop: Main loop head
       "fmla v8.8h, v6.8h, v0.h[0]\n"
-      "ldr x20, [x17, #0x28]\n"
+      "ldr x21, [x17, #0x28]\n"
       "fmla v12.8h, v6.8h, v1.h[0]\n"
       "ldr d17, [x17, #0x20]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
-      "ldr x21, [x17, #0x38]\n"
+      "ldr x20, [x17, #0x38]\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "ldr d16, [x17, #0x30]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x48]\n"
-      "add x13, x13, #0x10\n"
-      "add x12, x12, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.8h, v17.8h, v1.h[0]\n"
       "ldr d17, [x17, #0x40]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
-      "ldr x21, [x17, #0x58]\n"
+      "ldr x20, [x17, #0x48]\n"
       "fmla v15.8h, v16.8h, v1.h[0]\n"
       "ldr d16, [x17, #0x50]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x68]\n"
-      "ldr x23, [x13, #0x8]\n"
-      "sub x14, x14, #0x8\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0x58]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[1]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v12.8h, v17.8h, v1.h[1]\n"
       "ldr d17, [x17, #0x60]\n"
       "fmla v9.8h, v16.8h, v0.h[1]\n"
-      "ldr x21, [x17, #0x78]\n"
+      "ldr x20, [x17, #0x78]\n"
       "fmla v13.8h, v16.8h, v1.h[1]\n"
       "ldr d16, [x17, #0x70]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x88]\n"
-      "ldr x22, [x12, #0x8]\n"
-      "cmp x14, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.8h, v17.8h, v0.h[1]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.8h, v17.8h, v1.h[1]\n"
       "ldr d17, [x17, #0x80]\n"
       "fmla v11.8h, v16.8h, v0.h[1]\n"
-      "ldr x21, [x17, #0x98]\n"
+      "ldr x20, [x17, #0x88]\n"
       "fmla v15.8h, v16.8h, v1.h[1]\n"
       "ldr d16, [x17, #0x90]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xa8]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0x98]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[2]\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v12.8h, v17.8h, v1.h[2]\n"
       "ldr d17, [x17, #0xa0]\n"
       "fmla v9.8h, v16.8h, v0.h[2]\n"
-      "ldr x21, [x17, #0xb8]\n"
+      "ldr x20, [x17, #0xb8]\n"
       "fmla v13.8h, v16.8h, v1.h[2]\n"
       "ldr d16, [x17, #0xb0]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xc8]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.8h, v17.8h, v0.h[2]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.8h, v17.8h, v1.h[2]\n"
       "ldr d17, [x17, #0xc0]\n"
       "fmla v11.8h, v16.8h, v0.h[2]\n"
-      "ldr x21, [x17, #0xd8]\n"
+      "ldr x20, [x17, #0xc8]\n"
       "fmla v15.8h, v16.8h, v1.h[2]\n"
       "ldr d16, [x17, #0xd0]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xe8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0xd8]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[3]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v12.8h, v17.8h, v1.h[3]\n"
       "ldr d17, [x17, #0xe0]\n"
       "fmla v9.8h, v16.8h, v0.h[3]\n"
-      "ldr x21, [x17, #0xf8]\n"
+      "ldr x20, [x17, #0xf8]\n"
       "fmla v13.8h, v16.8h, v1.h[3]\n"
       "ldr d16, [x17, #0xf0]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x108]\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.8h, v17.8h, v0.h[3]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.8h, v17.8h, v1.h[3]\n"
       "ldr d17, [x17, #0x100]\n"
       "fmla v11.8h, v16.8h, v0.h[3]\n"
-      "ldr x21, [x17, #0x118]\n"
+      "ldr x20, [x17, #0x108]\n"
       "fmla v15.8h, v16.8h, v1.h[3]\n"
       "ldr d16, [x17, #0x110]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x128]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0x118]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[4]\n"
+      "ldr x21, [x17, #0x128]\n"
       "fmla v12.8h, v17.8h, v1.h[4]\n"
       "ldr d17, [x17, #0x120]\n"
       "fmla v9.8h, v16.8h, v0.h[4]\n"
-      "ldr x21, [x17, #0x138]\n"
+      "ldr x20, [x17, #0x138]\n"
       "fmla v13.8h, v16.8h, v1.h[4]\n"
       "ldr d16, [x17, #0x130]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x148]\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.8h, v17.8h, v0.h[4]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.8h, v17.8h, v1.h[4]\n"
       "ldr d17, [x17, #0x140]\n"
       "fmla v11.8h, v16.8h, v0.h[4]\n"
-      "ldr x21, [x17, #0x158]\n"
+      "ldr x20, [x17, #0x148]\n"
       "fmla v15.8h, v16.8h, v1.h[4]\n"
       "ldr d16, [x17, #0x150]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x168]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0x158]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[5]\n"
+      "ldr x21, [x17, #0x168]\n"
       "fmla v12.8h, v17.8h, v1.h[5]\n"
       "ldr d17, [x17, #0x160]\n"
       "fmla v9.8h, v16.8h, v0.h[5]\n"
-      "ldr x21, [x17, #0x178]\n"
+      "ldr x20, [x17, #0x178]\n"
       "fmla v13.8h, v16.8h, v1.h[5]\n"
       "ldr d16, [x17, #0x170]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x188]\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.8h, v17.8h, v0.h[5]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.8h, v17.8h, v1.h[5]\n"
       "ldr d17, [x17, #0x180]\n"
       "fmla v11.8h, v16.8h, v0.h[5]\n"
-      "ldr x21, [x17, #0x198]\n"
+      "ldr x20, [x17, #0x188]\n"
       "fmla v15.8h, v16.8h, v1.h[5]\n"
       "ldr d16, [x17, #0x190]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x1a8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0x198]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[6]\n"
+      "ldr x21, [x17, #0x1a8]\n"
       "fmla v12.8h, v17.8h, v1.h[6]\n"
       "ldr d17, [x17, #0x1a0]\n"
       "fmla v9.8h, v16.8h, v0.h[6]\n"
-      "ldr x21, [x17, #0x1b8]\n"
+      "ldr x20, [x17, #0x1b8]\n"
       "fmla v13.8h, v16.8h, v1.h[6]\n"
       "ldr d16, [x17, #0x1b0]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x1c8]\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.8h, v17.8h, v0.h[6]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.8h, v17.8h, v1.h[6]\n"
       "ldr d17, [x17, #0x1c0]\n"
       "fmla v11.8h, v16.8h, v0.h[6]\n"
-      "ldr x21, [x17, #0x1d8]\n"
+      "ldr x20, [x17, #0x1c8]\n"
       "fmla v15.8h, v16.8h, v1.h[6]\n"
       "ldr d16, [x17, #0x1d0]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x1e8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0x1d8]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.8h, v17.8h, v0.h[7]\n"
+      "ldr x21, [x17, #0x1e8]\n"
       "fmla v12.8h, v17.8h, v1.h[7]\n"
       "ldr d17, [x17, #0x1e0]\n"
       "fmla v9.8h, v16.8h, v0.h[7]\n"
-      "ldr x21, [x17, #0x1f8]\n"
+      "ldr x20, [x17, #0x1f8]\n"
       "fmla v13.8h, v16.8h, v1.h[7]\n"
       "ldr d16, [x17, #0x1f0]\n"
-      "mov v17.d[1], x20\n"
+      "mov v17.d[1], x21\n"
+      "add x13, x13, #0x10\n"
+      "mov v16.d[1], x20\n"
+      "add x12, x12, #0x10\n"
       "add x17, x17, #0x200\n"
-      "ldr x20, [x17, #0x8]\n"
-      "mov v16.d[1], x21\n"
       "fmla v10.8h, v17.8h, v0.h[7]\n"
       "fmla v14.8h, v17.8h, v1.h[7]\n"
       "ldr d6, [x17, #0x0]\n"
+      "ldr x21, [x17, #0x8]\n"
       "fmla v11.8h, v16.8h, v0.h[7]\n"
       "ldr d0, [x13, #0x0]\n"
       "fmla v15.8h, v16.8h, v1.h[7]\n"
       "ldr d1, [x12, #0x0]\n"
+      "sub x14, x14, #0x8\n"
       "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x20\n"
+      "cmp x14, #0x10\n"
+      "ldr x20, [x13, #0x8]\n"
+      "mov v6.d[1], x21\n"
+      "ldr x21, [x12, #0x8]\n"
+      "mov v0.d[1], x20\n"
       "ldr x20, [x17, #0x18]\n"
-      "mov v0.d[1], x23\n"
-      "mov v1.d[1], x22\n"
+      "mov v1.d[1], x21\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "bge 75b\n"
       "76:"  // Height 2: Multiply loop: Single iteration only
       "fmla v8.8h, v6.8h, v0.h[0]\n"
@@ -1107,8 +1105,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "sub x14, x14, #0x1\n"
       "ldr h0, [x12], #0x2\n"
       "ldr q17, [x17, #0x0]\n"
-      "ldr q16, [x17, #0x10]\n"
       "fmla v8.8h, v17.8h, v1.h[0]\n"
+      "ldr q16, [x17, #0x10]\n"
       "fmla v12.8h, v17.8h, v0.h[0]\n"
       "ldr q17, [x17, #0x20]\n"
       "fmla v9.8h, v16.8h, v1.h[0]\n"
@@ -1126,22 +1124,22 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "cmp x15, x20\n"
       "bne 72b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x16, x20, LSL #1\n"
       "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "tbz %x[flags], #1, 80f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.8h }, [x20]\n"
+      "fmin v8.8h, v8.8h, v16.8h\n"
+      "fmin v9.8h, v9.8h, v16.8h\n"
+      "fmin v10.8h, v10.8h, v16.8h\n"
+      "fmin v11.8h, v11.8h, v16.8h\n"
+      "fmin v12.8h, v12.8h, v16.8h\n"
+      "fmin v13.8h, v13.8h, v16.8h\n"
+      "fmin v14.8h, v14.8h, v16.8h\n"
+      "fmin v15.8h, v15.8h, v16.8h\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.8h }, [x21]\n"
       "ld1r { v16.8h }, [x20]\n"
-      "fmin v8.8h, v8.8h, v17.8h\n"
-      "fmin v9.8h, v9.8h, v17.8h\n"
-      "fmin v10.8h, v10.8h, v17.8h\n"
-      "fmin v11.8h, v11.8h, v17.8h\n"
-      "fmin v12.8h, v12.8h, v17.8h\n"
-      "fmin v13.8h, v13.8h, v17.8h\n"
-      "fmin v14.8h, v14.8h, v17.8h\n"
-      "fmin v15.8h, v15.8h, v17.8h\n"
       "fmax v8.8h, v8.8h, v16.8h\n"
       "fmax v9.8h, v9.8h, v16.8h\n"
       "fmax v10.8h, v10.8h, v16.8h\n"
@@ -1156,127 +1154,127 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "tbz x8, #4, 88f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
       "st1 { v9.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
       "tbz x8, #3, 84f\n"
       "st1 { v10.8h }, [x16], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
       "tbz x8, #2, 82f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
       "tbz x8, #1, 81f\n"
       "st1 { v11.s }[2], [x16], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
       "tbz x8, #0, 96f\n"
       "st1 { v11.h }[6], [x16]\n"
-      "st1 { v15.h }[6], [x26]\n"
+      "st1 { v15.h }[6], [x25]\n"
       "b 96f\n"
       "81:"  // Height 2: Partial direct writeback: partial_1_28
       "tbz x8, #0, 96f\n"
       "st1 { v11.h }[4], [x16]\n"
-      "st1 { v15.h }[4], [x26]\n"
+      "st1 { v15.h }[4], [x25]\n"
       "b 96f\n"
       "82:"  // Height 2: Partial direct writeback: partial_2_24
       "tbz x8, #1, 83f\n"
       "str s11, [x16], #0x4\n"
-      "str s15, [x26], #0x4\n"
+      "str s15, [x25], #0x4\n"
       "tbz x8, #0, 96f\n"
       "st1 { v11.h }[2], [x16]\n"
-      "st1 { v15.h }[2], [x26]\n"
+      "st1 { v15.h }[2], [x25]\n"
       "b 96f\n"
       "83:"  // Height 2: Partial direct writeback: partial_1_24
       "tbz x8, #0, 96f\n"
       "str h11, [x16, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
       "b 96f\n"
       "84:"  // Height 2: Partial direct writeback: partial_4_16
       "tbz x8, #2, 86f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
+      "str d14, [x25], #0x8\n"
       "tbz x8, #1, 85f\n"
       "st1 { v10.s }[2], [x16], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
       "tbz x8, #0, 96f\n"
       "st1 { v10.h }[6], [x16]\n"
-      "st1 { v14.h }[6], [x26]\n"
+      "st1 { v14.h }[6], [x25]\n"
       "b 96f\n"
       "85:"  // Height 2: Partial direct writeback: partial_1_20
       "tbz x8, #0, 96f\n"
       "st1 { v10.h }[4], [x16]\n"
-      "st1 { v14.h }[4], [x26]\n"
+      "st1 { v14.h }[4], [x25]\n"
       "b 96f\n"
       "86:"  // Height 2: Partial direct writeback: partial_2_16
       "tbz x8, #1, 87f\n"
       "str s10, [x16], #0x4\n"
-      "str s14, [x26], #0x4\n"
+      "str s14, [x25], #0x4\n"
       "tbz x8, #0, 96f\n"
       "st1 { v10.h }[2], [x16]\n"
-      "st1 { v14.h }[2], [x26]\n"
+      "st1 { v14.h }[2], [x25]\n"
       "b 96f\n"
       "87:"  // Height 2: Partial direct writeback: partial_1_16
       "tbz x8, #0, 96f\n"
       "str h10, [x16, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
       "b 96f\n"
       "88:"  // Height 2: Partial direct writeback: partial_8_0
       "tbz x8, #3, 92f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
       "tbz x8, #2, 90f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
+      "str d13, [x25], #0x8\n"
       "tbz x8, #1, 89f\n"
       "st1 { v9.s }[2], [x16], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
       "tbz x8, #0, 96f\n"
       "st1 { v9.h }[6], [x16]\n"
-      "st1 { v13.h }[6], [x26]\n"
+      "st1 { v13.h }[6], [x25]\n"
       "b 96f\n"
       "89:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x8, #0, 96f\n"
       "st1 { v9.h }[4], [x16]\n"
-      "st1 { v13.h }[4], [x26]\n"
+      "st1 { v13.h }[4], [x25]\n"
       "b 96f\n"
       "90:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x8, #1, 91f\n"
       "str s9, [x16], #0x4\n"
-      "str s13, [x26], #0x4\n"
+      "str s13, [x25], #0x4\n"
       "tbz x8, #0, 96f\n"
       "st1 { v9.h }[2], [x16]\n"
-      "st1 { v13.h }[2], [x26]\n"
+      "st1 { v13.h }[2], [x25]\n"
       "b 96f\n"
       "91:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x8, #0, 96f\n"
       "str h9, [x16, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
       "b 96f\n"
       "92:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x8, #2, 94f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
+      "str d12, [x25], #0x8\n"
       "tbz x8, #1, 93f\n"
       "st1 { v8.s }[2], [x16], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
       "tbz x8, #0, 96f\n"
       "st1 { v8.h }[6], [x16]\n"
-      "st1 { v12.h }[6], [x26]\n"
+      "st1 { v12.h }[6], [x25]\n"
       "b 96f\n"
       "93:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x8, #0, 96f\n"
       "st1 { v8.h }[4], [x16]\n"
-      "st1 { v12.h }[4], [x26]\n"
+      "st1 { v12.h }[4], [x25]\n"
       "b 96f\n"
       "94:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x8, #1, 95f\n"
       "str s8, [x16], #0x4\n"
-      "str s12, [x26], #0x4\n"
+      "str s12, [x25], #0x4\n"
       "tbz x8, #0, 96f\n"
       "st1 { v8.h }[2], [x16]\n"
-      "st1 { v12.h }[2], [x26]\n"
+      "st1 { v12.h }[2], [x25]\n"
       "b 96f\n"
       "95:"  // Height 2: Partial direct writeback: partial_1_0
       "str h8, [x16, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
       "96:"  // Height 2: Partial direct writeback: Done
       "b 98f\n"
       "97:"  // Height 2: Full writeback
@@ -1285,31 +1283,31 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
       "98:"  // Height 2: Writeback done
       "subs x8, x8, #0x20\n"
       "bgt 51b\n"
       "b 296f\n"
       "99:"  // Height 3
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "100:"  // Height 3: Column loop
       "cbz x7, 101f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
       "mov v16.16b, v8.16b\n"
+      "add x7, x7, #0x40\n"
       "mov v17.16b, v9.16b\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
@@ -1317,182 +1315,182 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "101:"  // Height 3: no bias
       "tbz %x[flags], #0, 119f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x16, x20, LSL #1\n"
       "cmp x8, #0x20\n"
-      "add x26, x16, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "bge 118f\n"
       "tbz x8, #4, 109f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
       "ld1 { v9.8h }, [x16], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
       "tbz x8, #3, 105f\n"
       "ld1 { v10.8h }, [x16], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
       "tbz x8, #2, 103f\n"
       "ldr d11, [x16], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x8, #1, 102f\n"
       "ld1 { v11.s }[2], [x16], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v11.h }[6], [x16]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
       "b 117f\n"
       "102:"  // Height 3: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v11.h }[4], [x16]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
       "b 117f\n"
       "103:"  // Height 3: Partial accumulate: partial_2_24
       "tbz x8, #1, 104f\n"
       "ldr s11, [x16], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s15, [x26], #0x4\n"
-      "ldr s19, [x25], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v11.h }[2], [x16]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
       "b 117f\n"
       "104:"  // Height 3: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x8, #0, 117f\n"
       "ldr h11, [x16, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
       "b 117f\n"
       "105:"  // Height 3: Partial accumulate: partial_4_16
       "tbz x8, #2, 107f\n"
       "ldr d10, [x16], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x8, #1, 106f\n"
       "ld1 { v10.s }[2], [x16], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v10.h }[6], [x16]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
       "b 117f\n"
       "106:"  // Height 3: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v10.h }[4], [x16]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
       "b 117f\n"
       "107:"  // Height 3: Partial accumulate: partial_2_16
       "tbz x8, #1, 108f\n"
       "ldr s10, [x16], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s14, [x26], #0x4\n"
-      "ldr s18, [x25], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v10.h }[2], [x16]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
       "b 117f\n"
       "108:"  // Height 3: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x8, #0, 117f\n"
       "ldr h10, [x16, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
       "b 117f\n"
       "109:"  // Height 3: Partial accumulate: partial_8_0
       "tbz x8, #3, 113f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
       "tbz x8, #2, 111f\n"
       "ldr d9, [x16], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x8, #1, 110f\n"
       "ld1 { v9.s }[2], [x16], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v9.h }[6], [x16]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
       "b 117f\n"
       "110:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v9.h }[4], [x16]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
       "b 117f\n"
       "111:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x8, #1, 112f\n"
       "ldr s9, [x16], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s13, [x26], #0x4\n"
-      "ldr s17, [x25], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v9.h }[2], [x16]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
       "b 117f\n"
       "112:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x8, #0, 117f\n"
       "ldr h9, [x16, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
       "b 117f\n"
       "113:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x8, #2, 115f\n"
       "ldr d8, [x16], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
       "tbz x8, #1, 114f\n"
       "ld1 { v8.s }[2], [x16], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v8.h }[6], [x16]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
       "b 117f\n"
       "114:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v8.h }[4], [x16]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
       "b 117f\n"
       "115:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x8, #1, 116f\n"
       "ldr s8, [x16], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s12, [x26], #0x4\n"
-      "ldr s16, [x25], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
       "tbz x8, #0, 117f\n"
       "ld1 { v8.h }[2], [x16]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
       "b 117f\n"
       "116:"  // Height 3: Partial accumulate: partial_1_0
       "ldr h8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h12, [x26, #0x0]\n"
-      "ldr h16, [x25, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
       "117:"  // Height 3: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 120f\n"
@@ -1501,14 +1499,14 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
       "b 120f\n"
       "119:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1527,8 +1525,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "mov x15, #0x0\n"
       "121:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 122f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1563,15 +1561,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v16.8h, v6.8h, v2.h[0]\n"
       "ldr d21, [x17, #0x20]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
-      "add x13, x13, #0x10\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
-      "add x12, x12, #0x10\n"
+      "ldr x21, [x17, #0x48]\n"
       "fmla v17.8h, v7.8h, v2.h[0]\n"
       "ldr d20, [x17, #0x30]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x48]\n"
-      "add x11, x11, #0x10\n"
-      "ldr x24, [x13, #0x8]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.8h, v21.8h, v0.h[0]\n"
       "fmla v14.8h, v21.8h, v1.h[0]\n"
@@ -1579,15 +1573,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v18.8h, v21.8h, v2.h[0]\n"
       "ldr d21, [x17, #0x40]\n"
       "fmla v11.8h, v20.8h, v0.h[0]\n"
-      "ldr x23, [x12, #0x8]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.8h, v20.8h, v1.h[0]\n"
-      "ldr x22, [x11, #0x8]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v19.8h, v20.8h, v2.h[0]\n"
       "ldr d20, [x17, #0x50]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x68]\n"
-      "sub x14, x14, #0x8\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.8h, v21.8h, v0.h[1]\n"
       "fmla v12.8h, v21.8h, v1.h[1]\n"
@@ -1595,14 +1585,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v16.8h, v21.8h, v2.h[1]\n"
       "ldr d21, [x17, #0x60]\n"
       "fmla v9.8h, v20.8h, v0.h[1]\n"
-      "cmp x14, #0x10\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.8h, v20.8h, v1.h[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "ldr x21, [x17, #0x88]\n"
       "fmla v17.8h, v20.8h, v2.h[1]\n"
       "ldr d20, [x17, #0x70]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x88]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.8h, v21.8h, v0.h[1]\n"
       "fmla v14.8h, v21.8h, v1.h[1]\n"
@@ -1610,11 +1597,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v18.8h, v21.8h, v2.h[1]\n"
       "ldr d21, [x17, #0x80]\n"
       "fmla v11.8h, v20.8h, v0.h[1]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.8h, v20.8h, v1.h[1]\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v19.8h, v20.8h, v2.h[1]\n"
       "ldr d20, [x17, #0x90]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xa8]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.8h, v21.8h, v0.h[2]\n"
       "fmla v12.8h, v21.8h, v1.h[2]\n"
@@ -1622,11 +1609,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v16.8h, v21.8h, v2.h[2]\n"
       "ldr d21, [x17, #0xa0]\n"
       "fmla v9.8h, v20.8h, v0.h[2]\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.8h, v20.8h, v1.h[2]\n"
+      "ldr x21, [x17, #0xc8]\n"
       "fmla v17.8h, v20.8h, v2.h[2]\n"
       "ldr d20, [x17, #0xb0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xc8]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.8h, v21.8h, v0.h[2]\n"
       "fmla v14.8h, v21.8h, v1.h[2]\n"
@@ -1634,11 +1621,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v18.8h, v21.8h, v2.h[2]\n"
       "ldr d21, [x17, #0xc0]\n"
       "fmla v11.8h, v20.8h, v0.h[2]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.8h, v20.8h, v1.h[2]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v19.8h, v20.8h, v2.h[2]\n"
       "ldr d20, [x17, #0xd0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xe8]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.8h, v21.8h, v0.h[3]\n"
       "fmla v12.8h, v21.8h, v1.h[3]\n"
@@ -1646,11 +1633,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v16.8h, v21.8h, v2.h[3]\n"
       "ldr d21, [x17, #0xe0]\n"
       "fmla v9.8h, v20.8h, v0.h[3]\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.8h, v20.8h, v1.h[3]\n"
+      "ldr x21, [x17, #0x108]\n"
       "fmla v17.8h, v20.8h, v2.h[3]\n"
       "ldr d20, [x17, #0xf0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x108]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.8h, v21.8h, v0.h[3]\n"
       "fmla v14.8h, v21.8h, v1.h[3]\n"
@@ -1658,11 +1645,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v18.8h, v21.8h, v2.h[3]\n"
       "ldr d21, [x17, #0x100]\n"
       "fmla v11.8h, v20.8h, v0.h[3]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.8h, v20.8h, v1.h[3]\n"
+      "ldr x21, [x17, #0x128]\n"
       "fmla v19.8h, v20.8h, v2.h[3]\n"
       "ldr d20, [x17, #0x110]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x128]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.8h, v21.8h, v0.h[4]\n"
       "fmla v12.8h, v21.8h, v1.h[4]\n"
@@ -1670,11 +1657,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v16.8h, v21.8h, v2.h[4]\n"
       "ldr d21, [x17, #0x120]\n"
       "fmla v9.8h, v20.8h, v0.h[4]\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.8h, v20.8h, v1.h[4]\n"
+      "ldr x21, [x17, #0x148]\n"
       "fmla v17.8h, v20.8h, v2.h[4]\n"
       "ldr d20, [x17, #0x130]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x148]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.8h, v21.8h, v0.h[4]\n"
       "fmla v14.8h, v21.8h, v1.h[4]\n"
@@ -1682,11 +1669,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v18.8h, v21.8h, v2.h[4]\n"
       "ldr d21, [x17, #0x140]\n"
       "fmla v11.8h, v20.8h, v0.h[4]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.8h, v20.8h, v1.h[4]\n"
+      "ldr x21, [x17, #0x168]\n"
       "fmla v19.8h, v20.8h, v2.h[4]\n"
       "ldr d20, [x17, #0x150]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x168]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.8h, v21.8h, v0.h[5]\n"
       "fmla v12.8h, v21.8h, v1.h[5]\n"
@@ -1694,11 +1681,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v16.8h, v21.8h, v2.h[5]\n"
       "ldr d21, [x17, #0x160]\n"
       "fmla v9.8h, v20.8h, v0.h[5]\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.8h, v20.8h, v1.h[5]\n"
+      "ldr x21, [x17, #0x188]\n"
       "fmla v17.8h, v20.8h, v2.h[5]\n"
       "ldr d20, [x17, #0x170]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x188]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.8h, v21.8h, v0.h[5]\n"
       "fmla v14.8h, v21.8h, v1.h[5]\n"
@@ -1706,11 +1693,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v18.8h, v21.8h, v2.h[5]\n"
       "ldr d21, [x17, #0x180]\n"
       "fmla v11.8h, v20.8h, v0.h[5]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.8h, v20.8h, v1.h[5]\n"
+      "ldr x21, [x17, #0x1a8]\n"
       "fmla v19.8h, v20.8h, v2.h[5]\n"
       "ldr d20, [x17, #0x190]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x1a8]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.8h, v21.8h, v0.h[6]\n"
       "fmla v12.8h, v21.8h, v1.h[6]\n"
@@ -1718,11 +1705,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v16.8h, v21.8h, v2.h[6]\n"
       "ldr d21, [x17, #0x1a0]\n"
       "fmla v9.8h, v20.8h, v0.h[6]\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.8h, v20.8h, v1.h[6]\n"
+      "ldr x21, [x17, #0x1c8]\n"
       "fmla v17.8h, v20.8h, v2.h[6]\n"
       "ldr d20, [x17, #0x1b0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x1c8]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.8h, v21.8h, v0.h[6]\n"
       "fmla v14.8h, v21.8h, v1.h[6]\n"
@@ -1730,11 +1717,11 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v18.8h, v21.8h, v2.h[6]\n"
       "ldr d21, [x17, #0x1c0]\n"
       "fmla v11.8h, v20.8h, v0.h[6]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.8h, v20.8h, v1.h[6]\n"
+      "ldr x21, [x17, #0x1e8]\n"
       "fmla v19.8h, v20.8h, v2.h[6]\n"
       "ldr d20, [x17, #0x1d0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x1e8]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.8h, v21.8h, v0.h[7]\n"
       "fmla v12.8h, v21.8h, v1.h[7]\n"
@@ -1742,29 +1729,40 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v16.8h, v21.8h, v2.h[7]\n"
       "ldr d21, [x17, #0x1e0]\n"
       "fmla v9.8h, v20.8h, v0.h[7]\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.8h, v20.8h, v1.h[7]\n"
+      "add x13, x13, #0x10\n"
       "fmla v17.8h, v20.8h, v2.h[7]\n"
       "ldr d20, [x17, #0x1f0]\n"
-      "mov v21.d[1], x21\n"
-      "add x17, x17, #0x200\n"
-      "ldr x21, [x17, #0x8]\n"
       "mov v20.d[1], x20\n"
+      "add x12, x12, #0x10\n"
+      "add x11, x11, #0x10\n"
+      "add x17, x17, #0x200\n"
       "fmla v10.8h, v21.8h, v0.h[7]\n"
+      "ldr x20, [x17, #0x8]\n"
       "fmla v14.8h, v21.8h, v1.h[7]\n"
-      "ldr x20, [x17, #0x18]\n"
+      "ldr x23, [x13, #0x8]\n"
       "fmla v18.8h, v21.8h, v2.h[7]\n"
       "ldr d6, [x17, #0x0]\n"
       "fmla v11.8h, v20.8h, v0.h[7]\n"
       "ldr d0, [x13, #0x0]\n"
       "fmla v15.8h, v20.8h, v1.h[7]\n"
       "ldr d1, [x12, #0x0]\n"
+      "ldr x22, [x12, #0x8]\n"
       "fmla v19.8h, v20.8h, v2.h[7]\n"
       "ldr d2, [x11, #0x0]\n"
+      "sub x14, x14, #0x8\n"
       "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x24\n"
-      "mov v1.d[1], x23\n"
-      "mov v2.d[1], x22\n"
+      "cmp x14, #0x10\n"
+      "ldr x21, [x11, #0x8]\n"
+      "mov v6.d[1], x20\n"
+      "ldr x20, [x17, #0x18]\n"
+      "mov v0.d[1], x23\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
+      "mov v1.d[1], x22\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v2.d[1], x21\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "mov v7.d[1], x20\n"
       "bge 124b\n"
       "125:"  // Height 3: Multiply loop: Single iteration only
@@ -1910,8 +1908,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr h1, [x12], #0x2\n"
       "ldr h0, [x11], #0x2\n"
       "ldr q21, [x17, #0x0]\n"
-      "ldr q20, [x17, #0x10]\n"
       "fmla v8.8h, v21.8h, v2.h[0]\n"
+      "ldr q20, [x17, #0x10]\n"
       "fmla v12.8h, v21.8h, v1.h[0]\n"
       "fmla v16.8h, v21.8h, v0.h[0]\n"
       "ldr q21, [x17, #0x20]\n"
@@ -1933,28 +1931,28 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "cmp x15, x20\n"
       "bne 121b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x16, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "tbz %x[flags], #1, 129f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v20.8h }, [x20]\n"
+      "fmin v8.8h, v8.8h, v20.8h\n"
+      "fmin v9.8h, v9.8h, v20.8h\n"
+      "fmin v10.8h, v10.8h, v20.8h\n"
+      "fmin v11.8h, v11.8h, v20.8h\n"
+      "fmin v12.8h, v12.8h, v20.8h\n"
+      "fmin v13.8h, v13.8h, v20.8h\n"
+      "fmin v14.8h, v14.8h, v20.8h\n"
+      "fmin v15.8h, v15.8h, v20.8h\n"
+      "fmin v16.8h, v16.8h, v20.8h\n"
+      "fmin v17.8h, v17.8h, v20.8h\n"
+      "fmin v18.8h, v18.8h, v20.8h\n"
+      "fmin v19.8h, v19.8h, v20.8h\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.8h }, [x21]\n"
       "ld1r { v20.8h }, [x20]\n"
-      "fmin v8.8h, v8.8h, v21.8h\n"
-      "fmin v9.8h, v9.8h, v21.8h\n"
-      "fmin v10.8h, v10.8h, v21.8h\n"
-      "fmin v11.8h, v11.8h, v21.8h\n"
-      "fmin v12.8h, v12.8h, v21.8h\n"
-      "fmin v13.8h, v13.8h, v21.8h\n"
-      "fmin v14.8h, v14.8h, v21.8h\n"
-      "fmin v15.8h, v15.8h, v21.8h\n"
-      "fmin v16.8h, v16.8h, v21.8h\n"
-      "fmin v17.8h, v17.8h, v21.8h\n"
-      "fmin v18.8h, v18.8h, v21.8h\n"
-      "fmin v19.8h, v19.8h, v21.8h\n"
       "fmax v8.8h, v8.8h, v20.8h\n"
       "fmax v9.8h, v9.8h, v20.8h\n"
       "fmax v10.8h, v10.8h, v20.8h\n"
@@ -1973,159 +1971,159 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "tbz x8, #4, 137f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
       "st1 { v9.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
       "tbz x8, #3, 133f\n"
       "st1 { v10.8h }, [x16], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
       "tbz x8, #2, 131f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x8, #1, 130f\n"
       "st1 { v11.s }[2], [x16], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
       "tbz x8, #0, 145f\n"
       "st1 { v11.h }[6], [x16]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
       "b 145f\n"
       "130:"  // Height 3: Partial direct writeback: partial_1_28
       "tbz x8, #0, 145f\n"
       "st1 { v11.h }[4], [x16]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
       "b 145f\n"
       "131:"  // Height 3: Partial direct writeback: partial_2_24
       "tbz x8, #1, 132f\n"
       "str s11, [x16], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
       "tbz x8, #0, 145f\n"
       "st1 { v11.h }[2], [x16]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
       "b 145f\n"
       "132:"  // Height 3: Partial direct writeback: partial_1_24
       "tbz x8, #0, 145f\n"
       "str h11, [x16, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
       "b 145f\n"
       "133:"  // Height 3: Partial direct writeback: partial_4_16
       "tbz x8, #2, 135f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x8, #1, 134f\n"
       "st1 { v10.s }[2], [x16], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
       "tbz x8, #0, 145f\n"
       "st1 { v10.h }[6], [x16]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
       "b 145f\n"
       "134:"  // Height 3: Partial direct writeback: partial_1_20
       "tbz x8, #0, 145f\n"
       "st1 { v10.h }[4], [x16]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
       "b 145f\n"
       "135:"  // Height 3: Partial direct writeback: partial_2_16
       "tbz x8, #1, 136f\n"
       "str s10, [x16], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
       "tbz x8, #0, 145f\n"
       "st1 { v10.h }[2], [x16]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
       "b 145f\n"
       "136:"  // Height 3: Partial direct writeback: partial_1_16
       "tbz x8, #0, 145f\n"
       "str h10, [x16, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
       "b 145f\n"
       "137:"  // Height 3: Partial direct writeback: partial_8_0
       "tbz x8, #3, 141f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
       "tbz x8, #2, 139f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x8, #1, 138f\n"
       "st1 { v9.s }[2], [x16], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
       "tbz x8, #0, 145f\n"
       "st1 { v9.h }[6], [x16]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
       "b 145f\n"
       "138:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x8, #0, 145f\n"
       "st1 { v9.h }[4], [x16]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
       "b 145f\n"
       "139:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x8, #1, 140f\n"
       "str s9, [x16], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
       "tbz x8, #0, 145f\n"
       "st1 { v9.h }[2], [x16]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
       "b 145f\n"
       "140:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x8, #0, 145f\n"
       "str h9, [x16, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
       "b 145f\n"
       "141:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x8, #2, 143f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x8, #1, 142f\n"
       "st1 { v8.s }[2], [x16], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
       "tbz x8, #0, 145f\n"
       "st1 { v8.h }[6], [x16]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
       "b 145f\n"
       "142:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x8, #0, 145f\n"
       "st1 { v8.h }[4], [x16]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
       "b 145f\n"
       "143:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x8, #1, 144f\n"
       "str s8, [x16], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
       "tbz x8, #0, 145f\n"
       "st1 { v8.h }[2], [x16]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
       "b 145f\n"
       "144:"  // Height 3: Partial direct writeback: partial_1_0
       "str h8, [x16, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
       "145:"  // Height 3: Partial direct writeback: Done
       "b 147f\n"
       "146:"  // Height 3: Full writeback
@@ -2134,35 +2132,35 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "147:"  // Height 3: Writeback done
       "subs x8, x8, #0x20\n"
       "bgt 100b\n"
       "b 296f\n"
       "148:"  // Height 4
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "149:"  // Height 4: Column loop
       "cbz x7, 150f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
       "mov v16.16b, v8.16b\n"
+      "add x7, x7, #0x40\n"
       "mov v17.16b, v9.16b\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
@@ -2174,215 +2172,215 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "150:"  // Height 4: no bias
       "tbz %x[flags], #0, 168f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x20\n"
-      "add x26, x16, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x16, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "cmp x8, #0x20\n"
+      "add x23, x24, x20, LSL #1\n"
       "bge 167f\n"
       "tbz x8, #4, 158f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
       "ld1 { v9.8h }, [x16], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
       "tbz x8, #3, 154f\n"
       "ld1 { v10.8h }, [x16], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
       "tbz x8, #2, 152f\n"
       "ldr d11, [x16], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x8, #1, 151f\n"
       "ld1 { v11.s }[2], [x16], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v11.h }[6], [x16]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
       "b 166f\n"
       "151:"  // Height 4: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v11.h }[4], [x16]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
       "b 166f\n"
       "152:"  // Height 4: Partial accumulate: partial_2_24
       "tbz x8, #1, 153f\n"
       "ldr s11, [x16], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s15, [x26], #0x4\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v11.h }[2], [x16]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
       "b 166f\n"
       "153:"  // Height 4: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x8, #0, 166f\n"
       "ldr h11, [x16, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
       "b 166f\n"
       "154:"  // Height 4: Partial accumulate: partial_4_16
       "tbz x8, #2, 156f\n"
       "ldr d10, [x16], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x8, #1, 155f\n"
       "ld1 { v10.s }[2], [x16], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v10.h }[6], [x16]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
       "b 166f\n"
       "155:"  // Height 4: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v10.h }[4], [x16]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
       "b 166f\n"
       "156:"  // Height 4: Partial accumulate: partial_2_16
       "tbz x8, #1, 157f\n"
       "ldr s10, [x16], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s14, [x26], #0x4\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v10.h }[2], [x16]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
       "b 166f\n"
       "157:"  // Height 4: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x8, #0, 166f\n"
       "ldr h10, [x16, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
       "b 166f\n"
       "158:"  // Height 4: Partial accumulate: partial_8_0
       "tbz x8, #3, 162f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
       "tbz x8, #2, 160f\n"
       "ldr d9, [x16], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x8, #1, 159f\n"
       "ld1 { v9.s }[2], [x16], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v9.h }[6], [x16]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
       "b 166f\n"
       "159:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v9.h }[4], [x16]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
       "b 166f\n"
       "160:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x8, #1, 161f\n"
       "ldr s9, [x16], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s13, [x26], #0x4\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v9.h }[2], [x16]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
       "b 166f\n"
       "161:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x8, #0, 166f\n"
       "ldr h9, [x16, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
       "b 166f\n"
       "162:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x8, #2, 164f\n"
       "ldr d8, [x16], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x8, #1, 163f\n"
       "ld1 { v8.s }[2], [x16], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v8.h }[6], [x16]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
       "b 166f\n"
       "163:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v8.h }[4], [x16]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
       "b 166f\n"
       "164:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x8, #1, 165f\n"
       "ldr s8, [x16], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s12, [x26], #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
       "tbz x8, #0, 166f\n"
       "ld1 { v8.h }[2], [x16]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
       "b 166f\n"
       "165:"  // Height 4: Partial accumulate: partial_1_0
       "ldr h8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h12, [x26, #0x0]\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
       "166:"  // Height 4: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 169f\n"
@@ -2391,18 +2389,18 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "b 169f\n"
       "168:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2425,8 +2423,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "mov x15, #0x0\n"
       "170:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 171f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2459,233 +2457,234 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "blt 174f\n"
       "173:"  // Height 4: Multiply loop: Main loop head
       "fmla v8.8h, v6.8h, v0.h[0]\n"
-      "ldr x20, [x17, #0x28]\n"
+      "ldr x21, [x17, #0x28]\n"
       "fmla v12.8h, v6.8h, v1.h[0]\n"
-      "ldr x21, [x17, #0x38]\n"
+      "ldr x20, [x17, #0x38]\n"
       "fmla v16.8h, v6.8h, v2.h[0]\n"
       "add x13, x13, #0x10\n"
       "fmla v20.8h, v6.8h, v3.h[0]\n"
       "ldr d25, [x17, #0x20]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
-      "add x12, x12, #0x10\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
-      "add x11, x11, #0x10\n"
+      "ldr x21, [x17, #0x48]\n"
       "fmla v17.8h, v7.8h, v2.h[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x12, x12, #0x10\n"
       "fmla v21.8h, v7.8h, v3.h[0]\n"
       "ldr d24, [x17, #0x30]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.8h, v25.8h, v0.h[0]\n"
-      "ldr x20, [x17, #0x48]\n"
       "fmla v14.8h, v25.8h, v1.h[0]\n"
-      "add x10, x10, #0x10\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x58]\n"
       "fmla v18.8h, v25.8h, v2.h[0]\n"
+      "add x11, x11, #0x10\n"
       "fmla v22.8h, v25.8h, v3.h[0]\n"
       "ldr d25, [x17, #0x40]\n"
       "fmla v11.8h, v24.8h, v0.h[0]\n"
-      "ldr x21, [x17, #0x58]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.8h, v24.8h, v1.h[0]\n"
-      "ldr x25, [x13, #0x8]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v19.8h, v24.8h, v2.h[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x10, x10, #0x10\n"
       "fmla v23.8h, v24.8h, v3.h[0]\n"
       "ldr d24, [x17, #0x50]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.8h, v25.8h, v0.h[1]\n"
-      "ldr x20, [x17, #0x68]\n"
       "fmla v12.8h, v25.8h, v1.h[1]\n"
-      "ldr x24, [x12, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x78]\n"
       "fmla v16.8h, v25.8h, v2.h[1]\n"
+      "ldr x25, [x13, #0x8]\n"
       "fmla v20.8h, v25.8h, v3.h[1]\n"
       "ldr d25, [x17, #0x60]\n"
       "fmla v9.8h, v24.8h, v0.h[1]\n"
-      "ldr x21, [x17, #0x78]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.8h, v24.8h, v1.h[1]\n"
-      "ldr x23, [x11, #0x8]\n"
+      "ldr x21, [x17, #0x88]\n"
       "fmla v17.8h, v24.8h, v2.h[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x24, [x12, #0x8]\n"
       "fmla v21.8h, v24.8h, v3.h[1]\n"
       "ldr d24, [x17, #0x70]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.8h, v25.8h, v0.h[1]\n"
-      "ldr x20, [x17, #0x88]\n"
       "fmla v14.8h, v25.8h, v1.h[1]\n"
-      "ldr x22, [x10, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x98]\n"
       "fmla v18.8h, v25.8h, v2.h[1]\n"
+      "ldr x23, [x11, #0x8]\n"
       "fmla v22.8h, v25.8h, v3.h[1]\n"
       "ldr d25, [x17, #0x80]\n"
       "fmla v11.8h, v24.8h, v0.h[1]\n"
-      "ldr x21, [x17, #0x98]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.8h, v24.8h, v1.h[1]\n"
-      "sub x14, x14, #0x8\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v19.8h, v24.8h, v2.h[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x22, [x10, #0x8]\n"
       "fmla v23.8h, v24.8h, v3.h[1]\n"
       "ldr d24, [x17, #0x90]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.8h, v25.8h, v0.h[2]\n"
-      "ldr x20, [x17, #0xa8]\n"
       "fmla v12.8h, v25.8h, v1.h[2]\n"
-      "cmp x14, #0x10\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0xb8]\n"
       "fmla v16.8h, v25.8h, v2.h[2]\n"
+      "sub x14, x14, #0x8\n"
       "fmla v20.8h, v25.8h, v3.h[2]\n"
       "ldr d25, [x17, #0xa0]\n"
       "fmla v9.8h, v24.8h, v0.h[2]\n"
-      "ldr x21, [x17, #0xb8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.8h, v24.8h, v1.h[2]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr x21, [x17, #0xc8]\n"
       "fmla v17.8h, v24.8h, v2.h[2]\n"
-      "mov v25.d[1], x20\n"
+      "cmp x14, #0x10\n"
       "fmla v21.8h, v24.8h, v3.h[2]\n"
       "ldr d24, [x17, #0xb0]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.8h, v25.8h, v0.h[2]\n"
-      "ldr x20, [x17, #0xc8]\n"
       "fmla v14.8h, v25.8h, v1.h[2]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0xd8]\n"
       "fmla v18.8h, v25.8h, v2.h[2]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v22.8h, v25.8h, v3.h[2]\n"
       "ldr d25, [x17, #0xc0]\n"
       "fmla v11.8h, v24.8h, v0.h[2]\n"
-      "ldr x21, [x17, #0xd8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.8h, v24.8h, v1.h[2]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v19.8h, v24.8h, v2.h[2]\n"
-      "mov v25.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "fmla v23.8h, v24.8h, v3.h[2]\n"
       "ldr d24, [x17, #0xd0]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.8h, v25.8h, v0.h[3]\n"
-      "ldr x20, [x17, #0xe8]\n"
       "fmla v12.8h, v25.8h, v1.h[3]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0xf8]\n"
       "fmla v16.8h, v25.8h, v2.h[3]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "fmla v20.8h, v25.8h, v3.h[3]\n"
       "ldr d25, [x17, #0xe0]\n"
       "fmla v9.8h, v24.8h, v0.h[3]\n"
-      "ldr x21, [x17, #0xf8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.8h, v24.8h, v1.h[3]\n"
+      "ldr x21, [x17, #0x108]\n"
       "fmla v17.8h, v24.8h, v2.h[3]\n"
-      "mov v25.d[1], x20\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       "fmla v21.8h, v24.8h, v3.h[3]\n"
       "ldr d24, [x17, #0xf0]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.8h, v25.8h, v0.h[3]\n"
-      "ldr x20, [x17, #0x108]\n"
       "fmla v14.8h, v25.8h, v1.h[3]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x118]\n"
       "fmla v18.8h, v25.8h, v2.h[3]\n"
       "fmla v22.8h, v25.8h, v3.h[3]\n"
       "ldr d25, [x17, #0x100]\n"
       "fmla v11.8h, v24.8h, v0.h[3]\n"
-      "ldr x21, [x17, #0x118]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.8h, v24.8h, v1.h[3]\n"
+      "ldr x21, [x17, #0x128]\n"
       "fmla v19.8h, v24.8h, v2.h[3]\n"
-      "mov v25.d[1], x20\n"
       "fmla v23.8h, v24.8h, v3.h[3]\n"
       "ldr d24, [x17, #0x110]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.8h, v25.8h, v0.h[4]\n"
-      "ldr x20, [x17, #0x128]\n"
       "fmla v12.8h, v25.8h, v1.h[4]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x138]\n"
       "fmla v16.8h, v25.8h, v2.h[4]\n"
       "fmla v20.8h, v25.8h, v3.h[4]\n"
       "ldr d25, [x17, #0x120]\n"
       "fmla v9.8h, v24.8h, v0.h[4]\n"
-      "ldr x21, [x17, #0x138]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.8h, v24.8h, v1.h[4]\n"
+      "ldr x21, [x17, #0x148]\n"
       "fmla v17.8h, v24.8h, v2.h[4]\n"
-      "mov v25.d[1], x20\n"
       "fmla v21.8h, v24.8h, v3.h[4]\n"
       "ldr d24, [x17, #0x130]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.8h, v25.8h, v0.h[4]\n"
-      "ldr x20, [x17, #0x148]\n"
       "fmla v14.8h, v25.8h, v1.h[4]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x158]\n"
       "fmla v18.8h, v25.8h, v2.h[4]\n"
       "fmla v22.8h, v25.8h, v3.h[4]\n"
       "ldr d25, [x17, #0x140]\n"
       "fmla v11.8h, v24.8h, v0.h[4]\n"
-      "ldr x21, [x17, #0x158]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.8h, v24.8h, v1.h[4]\n"
+      "ldr x21, [x17, #0x168]\n"
       "fmla v19.8h, v24.8h, v2.h[4]\n"
-      "mov v25.d[1], x20\n"
       "fmla v23.8h, v24.8h, v3.h[4]\n"
       "ldr d24, [x17, #0x150]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.8h, v25.8h, v0.h[5]\n"
-      "ldr x20, [x17, #0x168]\n"
       "fmla v12.8h, v25.8h, v1.h[5]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x178]\n"
       "fmla v16.8h, v25.8h, v2.h[5]\n"
       "fmla v20.8h, v25.8h, v3.h[5]\n"
       "ldr d25, [x17, #0x160]\n"
       "fmla v9.8h, v24.8h, v0.h[5]\n"
-      "ldr x21, [x17, #0x178]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.8h, v24.8h, v1.h[5]\n"
+      "ldr x21, [x17, #0x188]\n"
       "fmla v17.8h, v24.8h, v2.h[5]\n"
-      "mov v25.d[1], x20\n"
       "fmla v21.8h, v24.8h, v3.h[5]\n"
       "ldr d24, [x17, #0x170]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.8h, v25.8h, v0.h[5]\n"
-      "ldr x20, [x17, #0x188]\n"
       "fmla v14.8h, v25.8h, v1.h[5]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x198]\n"
       "fmla v18.8h, v25.8h, v2.h[5]\n"
       "fmla v22.8h, v25.8h, v3.h[5]\n"
       "ldr d25, [x17, #0x180]\n"
       "fmla v11.8h, v24.8h, v0.h[5]\n"
-      "ldr x21, [x17, #0x198]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.8h, v24.8h, v1.h[5]\n"
+      "ldr x21, [x17, #0x1a8]\n"
       "fmla v19.8h, v24.8h, v2.h[5]\n"
-      "mov v25.d[1], x20\n"
       "fmla v23.8h, v24.8h, v3.h[5]\n"
       "ldr d24, [x17, #0x190]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.8h, v25.8h, v0.h[6]\n"
-      "ldr x20, [x17, #0x1a8]\n"
       "fmla v12.8h, v25.8h, v1.h[6]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x1b8]\n"
       "fmla v16.8h, v25.8h, v2.h[6]\n"
       "fmla v20.8h, v25.8h, v3.h[6]\n"
       "ldr d25, [x17, #0x1a0]\n"
       "fmla v9.8h, v24.8h, v0.h[6]\n"
-      "ldr x21, [x17, #0x1b8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.8h, v24.8h, v1.h[6]\n"
+      "ldr x21, [x17, #0x1c8]\n"
       "fmla v17.8h, v24.8h, v2.h[6]\n"
-      "mov v25.d[1], x20\n"
       "fmla v21.8h, v24.8h, v3.h[6]\n"
       "ldr d24, [x17, #0x1b0]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.8h, v25.8h, v0.h[6]\n"
-      "ldr x20, [x17, #0x1c8]\n"
       "fmla v14.8h, v25.8h, v1.h[6]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x1d8]\n"
       "fmla v18.8h, v25.8h, v2.h[6]\n"
       "fmla v22.8h, v25.8h, v3.h[6]\n"
       "ldr d25, [x17, #0x1c0]\n"
       "fmla v11.8h, v24.8h, v0.h[6]\n"
-      "ldr x21, [x17, #0x1d8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.8h, v24.8h, v1.h[6]\n"
+      "ldr x21, [x17, #0x1e8]\n"
       "fmla v19.8h, v24.8h, v2.h[6]\n"
-      "mov v25.d[1], x20\n"
       "fmla v23.8h, v24.8h, v3.h[6]\n"
       "ldr d24, [x17, #0x1d0]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.8h, v25.8h, v0.h[7]\n"
-      "ldr x20, [x17, #0x1e8]\n"
       "fmla v12.8h, v25.8h, v1.h[7]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x1f8]\n"
       "fmla v16.8h, v25.8h, v2.h[7]\n"
       "fmla v20.8h, v25.8h, v3.h[7]\n"
       "ldr d25, [x17, #0x1e0]\n"
       "fmla v9.8h, v24.8h, v0.h[7]\n"
-      "ldr x21, [x17, #0x1f8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.8h, v24.8h, v1.h[7]\n"
       "fmla v17.8h, v24.8h, v2.h[7]\n"
-      "mov v25.d[1], x20\n"
       "fmla v21.8h, v24.8h, v3.h[7]\n"
       "ldr d24, [x17, #0x1f0]\n"
+      "mov v24.d[1], x20\n"
       "add x17, x17, #0x200\n"
       "fmla v10.8h, v25.8h, v0.h[7]\n"
+      "ldr x21, [x17, #0x8]\n"
       "fmla v14.8h, v25.8h, v1.h[7]\n"
-      "ldr x20, [x17, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x18]\n"
       "fmla v18.8h, v25.8h, v2.h[7]\n"
       "fmla v22.8h, v25.8h, v3.h[7]\n"
       "ldr d6, [x17, #0x0]\n"
@@ -2698,8 +2697,7 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v23.8h, v24.8h, v3.h[7]\n"
       "ldr d3, [x10, #0x0]\n"
       "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x20\n"
-      "ldr x20, [x17, #0x18]\n"
+      "mov v6.d[1], x21\n"
       "mov v0.d[1], x25\n"
       "mov v1.d[1], x24\n"
       "mov v2.d[1], x23\n"
@@ -2884,8 +2882,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr h1, [x11], #0x2\n"
       "ldr h0, [x10], #0x2\n"
       "ldr q25, [x17, #0x0]\n"
-      "ldr q24, [x17, #0x10]\n"
       "fmla v8.8h, v25.8h, v3.h[0]\n"
+      "ldr q24, [x17, #0x10]\n"
       "fmla v12.8h, v25.8h, v2.h[0]\n"
       "fmla v16.8h, v25.8h, v1.h[0]\n"
       "fmla v20.8h, v25.8h, v0.h[0]\n"
@@ -2911,34 +2909,34 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "cmp x15, x20\n"
       "bne 170b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "add x25, x16, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
+      "prfm pstl1keep, [x16, #0x0]\n"
       "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 178f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v24.8h }, [x20]\n"
+      "fmin v8.8h, v8.8h, v24.8h\n"
+      "fmin v9.8h, v9.8h, v24.8h\n"
+      "fmin v10.8h, v10.8h, v24.8h\n"
+      "fmin v11.8h, v11.8h, v24.8h\n"
+      "fmin v12.8h, v12.8h, v24.8h\n"
+      "fmin v13.8h, v13.8h, v24.8h\n"
+      "fmin v14.8h, v14.8h, v24.8h\n"
+      "fmin v15.8h, v15.8h, v24.8h\n"
+      "fmin v16.8h, v16.8h, v24.8h\n"
+      "fmin v17.8h, v17.8h, v24.8h\n"
+      "fmin v18.8h, v18.8h, v24.8h\n"
+      "fmin v19.8h, v19.8h, v24.8h\n"
+      "fmin v20.8h, v20.8h, v24.8h\n"
+      "fmin v21.8h, v21.8h, v24.8h\n"
+      "fmin v22.8h, v22.8h, v24.8h\n"
+      "fmin v23.8h, v23.8h, v24.8h\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v25.8h }, [x21]\n"
       "ld1r { v24.8h }, [x20]\n"
-      "fmin v8.8h, v8.8h, v25.8h\n"
-      "fmin v9.8h, v9.8h, v25.8h\n"
-      "fmin v10.8h, v10.8h, v25.8h\n"
-      "fmin v11.8h, v11.8h, v25.8h\n"
-      "fmin v12.8h, v12.8h, v25.8h\n"
-      "fmin v13.8h, v13.8h, v25.8h\n"
-      "fmin v14.8h, v14.8h, v25.8h\n"
-      "fmin v15.8h, v15.8h, v25.8h\n"
-      "fmin v16.8h, v16.8h, v25.8h\n"
-      "fmin v17.8h, v17.8h, v25.8h\n"
-      "fmin v18.8h, v18.8h, v25.8h\n"
-      "fmin v19.8h, v19.8h, v25.8h\n"
-      "fmin v20.8h, v20.8h, v25.8h\n"
-      "fmin v21.8h, v21.8h, v25.8h\n"
-      "fmin v22.8h, v22.8h, v25.8h\n"
-      "fmin v23.8h, v23.8h, v25.8h\n"
       "fmax v8.8h, v8.8h, v24.8h\n"
       "fmax v9.8h, v9.8h, v24.8h\n"
       "fmax v10.8h, v10.8h, v24.8h\n"
@@ -2961,191 +2959,191 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "tbz x8, #4, 186f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
       "st1 { v9.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
       "tbz x8, #3, 182f\n"
       "st1 { v10.8h }, [x16], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
       "tbz x8, #2, 180f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
       "tbz x8, #1, 179f\n"
       "st1 { v11.s }[2], [x16], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
       "tbz x8, #0, 194f\n"
       "st1 { v11.h }[6], [x16]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
       "b 194f\n"
       "179:"  // Height 4: Partial direct writeback: partial_1_28
       "tbz x8, #0, 194f\n"
       "st1 { v11.h }[4], [x16]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
       "b 194f\n"
       "180:"  // Height 4: Partial direct writeback: partial_2_24
       "tbz x8, #1, 181f\n"
       "str s11, [x16], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
       "tbz x8, #0, 194f\n"
       "st1 { v11.h }[2], [x16]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
       "b 194f\n"
       "181:"  // Height 4: Partial direct writeback: partial_1_24
       "tbz x8, #0, 194f\n"
       "str h11, [x16, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
       "b 194f\n"
       "182:"  // Height 4: Partial direct writeback: partial_4_16
       "tbz x8, #2, 184f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
       "tbz x8, #1, 183f\n"
       "st1 { v10.s }[2], [x16], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
       "tbz x8, #0, 194f\n"
       "st1 { v10.h }[6], [x16]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
       "b 194f\n"
       "183:"  // Height 4: Partial direct writeback: partial_1_20
       "tbz x8, #0, 194f\n"
       "st1 { v10.h }[4], [x16]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
       "b 194f\n"
       "184:"  // Height 4: Partial direct writeback: partial_2_16
       "tbz x8, #1, 185f\n"
       "str s10, [x16], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
       "tbz x8, #0, 194f\n"
       "st1 { v10.h }[2], [x16]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
       "b 194f\n"
       "185:"  // Height 4: Partial direct writeback: partial_1_16
       "tbz x8, #0, 194f\n"
       "str h10, [x16, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
       "b 194f\n"
       "186:"  // Height 4: Partial direct writeback: partial_8_0
       "tbz x8, #3, 190f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
       "tbz x8, #2, 188f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
       "tbz x8, #1, 187f\n"
       "st1 { v9.s }[2], [x16], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
       "tbz x8, #0, 194f\n"
       "st1 { v9.h }[6], [x16]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
       "b 194f\n"
       "187:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x8, #0, 194f\n"
       "st1 { v9.h }[4], [x16]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
       "b 194f\n"
       "188:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x8, #1, 189f\n"
       "str s9, [x16], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
       "tbz x8, #0, 194f\n"
       "st1 { v9.h }[2], [x16]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
       "b 194f\n"
       "189:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x8, #0, 194f\n"
       "str h9, [x16, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
       "b 194f\n"
       "190:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x8, #2, 192f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x8, #1, 191f\n"
       "st1 { v8.s }[2], [x16], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
       "tbz x8, #0, 194f\n"
       "st1 { v8.h }[6], [x16]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
       "b 194f\n"
       "191:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x8, #0, 194f\n"
       "st1 { v8.h }[4], [x16]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
       "b 194f\n"
       "192:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x8, #1, 193f\n"
       "str s8, [x16], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
       "tbz x8, #0, 194f\n"
       "st1 { v8.h }[2], [x16]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
       "b 194f\n"
       "193:"  // Height 4: Partial direct writeback: partial_1_0
       "str h8, [x16, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
       "194:"  // Height 4: Partial direct writeback: Done
       "b 196f\n"
       "195:"  // Height 4: Full writeback
@@ -3154,39 +3152,39 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
       "196:"  // Height 4: Writeback done
       "subs x8, x8, #0x20\n"
       "bgt 149b\n"
       "b 296f\n"
       "197:"  // Height 5
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "198:"  // Height 5: Column loop
       "cbz x7, 199f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
       "mov v16.16b, v8.16b\n"
+      "add x7, x7, #0x40\n"
       "mov v17.16b, v9.16b\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
@@ -3202,248 +3200,248 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "199:"  // Height 5: no bias
       "tbz %x[flags], #0, 217f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x20\n"
-      "add x26, x16, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x16, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
+      "cmp x8, #0x20\n"
+      "add x22, x23, x20, LSL #1\n"
       "bge 216f\n"
       "tbz x8, #4, 207f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
       "ld1 { v9.8h }, [x16], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
-      "ld1 { v25.8h }, [x23], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
+      "ld1 { v25.8h }, [x22], #0x10\n"
       "tbz x8, #3, 203f\n"
       "ld1 { v10.8h }, [x16], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
-      "ld1 { v26.8h }, [x23], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
+      "ld1 { v26.8h }, [x22], #0x10\n"
       "tbz x8, #2, 201f\n"
       "ldr d11, [x16], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x8, #1, 200f\n"
       "ld1 { v11.s }[2], [x16], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
-      "ld1 { v27.s }[2], [x23], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
+      "ld1 { v27.s }[2], [x22], #0x4\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v11.h }[6], [x16]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
-      "ld1 { v27.h }[6], [x23]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
+      "ld1 { v27.h }[6], [x22]\n"
       "b 215f\n"
       "200:"  // Height 5: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v11.h }[4], [x16]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
-      "ld1 { v27.h }[4], [x23]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
+      "ld1 { v27.h }[4], [x22]\n"
       "b 215f\n"
       "201:"  // Height 5: Partial accumulate: partial_2_24
       "tbz x8, #1, 202f\n"
       "ldr s11, [x16], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s15, [x26], #0x4\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
-      "ldr s27, [x23], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s27, [x22], #0x4\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v11.h }[2], [x16]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
-      "ld1 { v27.h }[2], [x23]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
+      "ld1 { v27.h }[2], [x22]\n"
       "b 215f\n"
       "202:"  // Height 5: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x8, #0, 215f\n"
       "ldr h11, [x16, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
-      "ldr h27, [x23, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
+      "ldr h27, [x22, #0x0]\n"
       "b 215f\n"
       "203:"  // Height 5: Partial accumulate: partial_4_16
       "tbz x8, #2, 205f\n"
       "ldr d10, [x16], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x8, #1, 204f\n"
       "ld1 { v10.s }[2], [x16], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
-      "ld1 { v26.s }[2], [x23], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
+      "ld1 { v26.s }[2], [x22], #0x4\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v10.h }[6], [x16]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
-      "ld1 { v26.h }[6], [x23]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
+      "ld1 { v26.h }[6], [x22]\n"
       "b 215f\n"
       "204:"  // Height 5: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v10.h }[4], [x16]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
-      "ld1 { v26.h }[4], [x23]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
+      "ld1 { v26.h }[4], [x22]\n"
       "b 215f\n"
       "205:"  // Height 5: Partial accumulate: partial_2_16
       "tbz x8, #1, 206f\n"
       "ldr s10, [x16], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s14, [x26], #0x4\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr s26, [x23], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
+      "ldr s26, [x22], #0x4\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v10.h }[2], [x16]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
-      "ld1 { v26.h }[2], [x23]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
+      "ld1 { v26.h }[2], [x22]\n"
       "b 215f\n"
       "206:"  // Height 5: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x8, #0, 215f\n"
       "ldr h10, [x16, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
-      "ldr h26, [x23, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
+      "ldr h26, [x22, #0x0]\n"
       "b 215f\n"
       "207:"  // Height 5: Partial accumulate: partial_8_0
       "tbz x8, #3, 211f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
       "tbz x8, #2, 209f\n"
       "ldr d9, [x16], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x8, #1, 208f\n"
       "ld1 { v9.s }[2], [x16], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
-      "ld1 { v25.s }[2], [x23], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
+      "ld1 { v25.s }[2], [x22], #0x4\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v9.h }[6], [x16]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
-      "ld1 { v25.h }[6], [x23]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
+      "ld1 { v25.h }[6], [x22]\n"
       "b 215f\n"
       "208:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v9.h }[4], [x16]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
-      "ld1 { v25.h }[4], [x23]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
+      "ld1 { v25.h }[4], [x22]\n"
       "b 215f\n"
       "209:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x8, #1, 210f\n"
       "ldr s9, [x16], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s13, [x26], #0x4\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
-      "ldr s25, [x23], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
+      "ldr s25, [x22], #0x4\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v9.h }[2], [x16]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
-      "ld1 { v25.h }[2], [x23]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
+      "ld1 { v25.h }[2], [x22]\n"
       "b 215f\n"
       "210:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x8, #0, 215f\n"
       "ldr h9, [x16, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
-      "ldr h25, [x23, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
+      "ldr h25, [x22, #0x0]\n"
       "b 215f\n"
       "211:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x8, #2, 213f\n"
       "ldr d8, [x16], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x8, #1, 212f\n"
       "ld1 { v8.s }[2], [x16], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
-      "ld1 { v24.s }[2], [x23], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
+      "ld1 { v24.s }[2], [x22], #0x4\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v8.h }[6], [x16]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
-      "ld1 { v24.h }[6], [x23]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
+      "ld1 { v24.h }[6], [x22]\n"
       "b 215f\n"
       "212:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v8.h }[4], [x16]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
-      "ld1 { v24.h }[4], [x23]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
+      "ld1 { v24.h }[4], [x22]\n"
       "b 215f\n"
       "213:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x8, #1, 214f\n"
       "ldr s8, [x16], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s12, [x26], #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
+      "ldr s24, [x22], #0x4\n"
       "tbz x8, #0, 215f\n"
       "ld1 { v8.h }[2], [x16]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
-      "ld1 { v24.h }[2], [x23]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
+      "ld1 { v24.h }[2], [x22]\n"
       "b 215f\n"
       "214:"  // Height 5: Partial accumulate: partial_1_0
       "ldr h8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h12, [x26, #0x0]\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
-      "ldr h24, [x23, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
+      "ldr h24, [x22, #0x0]\n"
       "215:"  // Height 5: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 218f\n"
@@ -3452,22 +3450,22 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
       "b 218f\n"
       "217:"  // Height 5: no accumulate
       "movi v8.16b, #0x0\n"
@@ -3494,8 +3492,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "mov x15, #0x0\n"
       "219:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 220f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -3542,259 +3540,259 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v24.8h, v6.8h, v4.h[0]\n"
       "ldr d29, [x17, #0x20]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
-      "add x11, x11, #0x10\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
-      "add x10, x10, #0x10\n"
+      "ldr x21, [x17, #0x48]\n"
       "fmla v17.8h, v7.8h, v2.h[0]\n"
-      "mov v29.d[1], x21\n"
+      "add x11, x11, #0x10\n"
       "fmla v21.8h, v7.8h, v3.h[0]\n"
-      "ldr x21, [x17, #0x48]\n"
+      "add x10, x10, #0x10\n"
       "fmla v25.8h, v7.8h, v4.h[0]\n"
       "ldr d28, [x17, #0x30]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.8h, v29.8h, v0.h[0]\n"
-      "add x9, x9, #0x10\n"
       "fmla v14.8h, v29.8h, v1.h[0]\n"
-      "ldr x26, [x13, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x58]\n"
       "fmla v18.8h, v29.8h, v2.h[0]\n"
+      "add x9, x9, #0x10\n"
       "fmla v22.8h, v29.8h, v3.h[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x26, [x13, #0x8]\n"
       "fmla v26.8h, v29.8h, v4.h[0]\n"
       "ldr d29, [x17, #0x40]\n"
       "fmla v11.8h, v28.8h, v0.h[0]\n"
-      "ldr x25, [x12, #0x8]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.8h, v28.8h, v1.h[0]\n"
-      "ldr x24, [x11, #0x8]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v19.8h, v28.8h, v2.h[0]\n"
-      "mov v29.d[1], x21\n"
+      "ldr x25, [x12, #0x8]\n"
       "fmla v23.8h, v28.8h, v3.h[0]\n"
-      "ldr x21, [x17, #0x68]\n"
+      "ldr x24, [x11, #0x8]\n"
       "fmla v27.8h, v28.8h, v4.h[0]\n"
       "ldr d28, [x17, #0x50]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.8h, v29.8h, v0.h[1]\n"
-      "ldr x23, [x10, #0x8]\n"
       "fmla v12.8h, v29.8h, v1.h[1]\n"
-      "ldr x22, [x9, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x78]\n"
       "fmla v16.8h, v29.8h, v2.h[1]\n"
+      "ldr x23, [x10, #0x8]\n"
       "fmla v20.8h, v29.8h, v3.h[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "ldr x22, [x9, #0x8]\n"
       "fmla v24.8h, v29.8h, v4.h[1]\n"
       "ldr d29, [x17, #0x60]\n"
       "fmla v9.8h, v28.8h, v0.h[1]\n"
-      "sub x14, x14, #0x8\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.8h, v28.8h, v1.h[1]\n"
-      "cmp x14, #0x10\n"
+      "ldr x21, [x17, #0x88]\n"
       "fmla v17.8h, v28.8h, v2.h[1]\n"
-      "mov v29.d[1], x21\n"
+      "sub x14, x14, #0x8\n"
       "fmla v21.8h, v28.8h, v3.h[1]\n"
-      "ldr x21, [x17, #0x88]\n"
+      "cmp x14, #0x10\n"
       "fmla v25.8h, v28.8h, v4.h[1]\n"
       "ldr d28, [x17, #0x70]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.8h, v29.8h, v0.h[1]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v14.8h, v29.8h, v1.h[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x98]\n"
       "fmla v18.8h, v29.8h, v2.h[1]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v22.8h, v29.8h, v3.h[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "fmla v26.8h, v29.8h, v4.h[1]\n"
       "ldr d29, [x17, #0x80]\n"
       "fmla v11.8h, v28.8h, v0.h[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.8h, v28.8h, v1.h[1]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v19.8h, v28.8h, v2.h[1]\n"
-      "mov v29.d[1], x21\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "fmla v23.8h, v28.8h, v3.h[1]\n"
-      "ldr x21, [x17, #0xa8]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       "fmla v27.8h, v28.8h, v4.h[1]\n"
       "ldr d28, [x17, #0x90]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.8h, v29.8h, v0.h[2]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v12.8h, v29.8h, v1.h[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0xb8]\n"
       "fmla v16.8h, v29.8h, v2.h[2]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v20.8h, v29.8h, v3.h[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
       "fmla v24.8h, v29.8h, v4.h[2]\n"
       "ldr d29, [x17, #0xa0]\n"
       "fmla v9.8h, v28.8h, v0.h[2]\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.8h, v28.8h, v1.h[2]\n"
+      "ldr x21, [x17, #0xc8]\n"
       "fmla v17.8h, v28.8h, v2.h[2]\n"
-      "mov v29.d[1], x21\n"
       "fmla v21.8h, v28.8h, v3.h[2]\n"
-      "ldr x21, [x17, #0xc8]\n"
       "fmla v25.8h, v28.8h, v4.h[2]\n"
       "ldr d28, [x17, #0xb0]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.8h, v29.8h, v0.h[2]\n"
       "fmla v14.8h, v29.8h, v1.h[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0xd8]\n"
       "fmla v18.8h, v29.8h, v2.h[2]\n"
       "fmla v22.8h, v29.8h, v3.h[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
       "fmla v26.8h, v29.8h, v4.h[2]\n"
       "ldr d29, [x17, #0xc0]\n"
       "fmla v11.8h, v28.8h, v0.h[2]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.8h, v28.8h, v1.h[2]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v19.8h, v28.8h, v2.h[2]\n"
-      "mov v29.d[1], x21\n"
       "fmla v23.8h, v28.8h, v3.h[2]\n"
-      "ldr x21, [x17, #0xe8]\n"
       "fmla v27.8h, v28.8h, v4.h[2]\n"
       "ldr d28, [x17, #0xd0]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.8h, v29.8h, v0.h[3]\n"
       "fmla v12.8h, v29.8h, v1.h[3]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0xf8]\n"
       "fmla v16.8h, v29.8h, v2.h[3]\n"
       "fmla v20.8h, v29.8h, v3.h[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
       "fmla v24.8h, v29.8h, v4.h[3]\n"
       "ldr d29, [x17, #0xe0]\n"
       "fmla v9.8h, v28.8h, v0.h[3]\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.8h, v28.8h, v1.h[3]\n"
+      "ldr x21, [x17, #0x108]\n"
       "fmla v17.8h, v28.8h, v2.h[3]\n"
-      "mov v29.d[1], x21\n"
       "fmla v21.8h, v28.8h, v3.h[3]\n"
-      "ldr x21, [x17, #0x108]\n"
       "fmla v25.8h, v28.8h, v4.h[3]\n"
       "ldr d28, [x17, #0xf0]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.8h, v29.8h, v0.h[3]\n"
       "fmla v14.8h, v29.8h, v1.h[3]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x118]\n"
       "fmla v18.8h, v29.8h, v2.h[3]\n"
       "fmla v22.8h, v29.8h, v3.h[3]\n"
-      "ldr x20, [x17, #0x118]\n"
       "fmla v26.8h, v29.8h, v4.h[3]\n"
       "ldr d29, [x17, #0x100]\n"
       "fmla v11.8h, v28.8h, v0.h[3]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.8h, v28.8h, v1.h[3]\n"
+      "ldr x21, [x17, #0x128]\n"
       "fmla v19.8h, v28.8h, v2.h[3]\n"
-      "mov v29.d[1], x21\n"
       "fmla v23.8h, v28.8h, v3.h[3]\n"
-      "ldr x21, [x17, #0x128]\n"
       "fmla v27.8h, v28.8h, v4.h[3]\n"
       "ldr d28, [x17, #0x110]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.8h, v29.8h, v0.h[4]\n"
       "fmla v12.8h, v29.8h, v1.h[4]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x138]\n"
       "fmla v16.8h, v29.8h, v2.h[4]\n"
       "fmla v20.8h, v29.8h, v3.h[4]\n"
-      "ldr x20, [x17, #0x138]\n"
       "fmla v24.8h, v29.8h, v4.h[4]\n"
       "ldr d29, [x17, #0x120]\n"
       "fmla v9.8h, v28.8h, v0.h[4]\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.8h, v28.8h, v1.h[4]\n"
+      "ldr x21, [x17, #0x148]\n"
       "fmla v17.8h, v28.8h, v2.h[4]\n"
-      "mov v29.d[1], x21\n"
       "fmla v21.8h, v28.8h, v3.h[4]\n"
-      "ldr x21, [x17, #0x148]\n"
       "fmla v25.8h, v28.8h, v4.h[4]\n"
       "ldr d28, [x17, #0x130]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.8h, v29.8h, v0.h[4]\n"
       "fmla v14.8h, v29.8h, v1.h[4]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x158]\n"
       "fmla v18.8h, v29.8h, v2.h[4]\n"
       "fmla v22.8h, v29.8h, v3.h[4]\n"
-      "ldr x20, [x17, #0x158]\n"
       "fmla v26.8h, v29.8h, v4.h[4]\n"
       "ldr d29, [x17, #0x140]\n"
       "fmla v11.8h, v28.8h, v0.h[4]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.8h, v28.8h, v1.h[4]\n"
+      "ldr x21, [x17, #0x168]\n"
       "fmla v19.8h, v28.8h, v2.h[4]\n"
-      "mov v29.d[1], x21\n"
       "fmla v23.8h, v28.8h, v3.h[4]\n"
-      "ldr x21, [x17, #0x168]\n"
       "fmla v27.8h, v28.8h, v4.h[4]\n"
       "ldr d28, [x17, #0x150]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.8h, v29.8h, v0.h[5]\n"
       "fmla v12.8h, v29.8h, v1.h[5]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x178]\n"
       "fmla v16.8h, v29.8h, v2.h[5]\n"
       "fmla v20.8h, v29.8h, v3.h[5]\n"
-      "ldr x20, [x17, #0x178]\n"
       "fmla v24.8h, v29.8h, v4.h[5]\n"
       "ldr d29, [x17, #0x160]\n"
       "fmla v9.8h, v28.8h, v0.h[5]\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.8h, v28.8h, v1.h[5]\n"
+      "ldr x21, [x17, #0x188]\n"
       "fmla v17.8h, v28.8h, v2.h[5]\n"
-      "mov v29.d[1], x21\n"
       "fmla v21.8h, v28.8h, v3.h[5]\n"
-      "ldr x21, [x17, #0x188]\n"
       "fmla v25.8h, v28.8h, v4.h[5]\n"
       "ldr d28, [x17, #0x170]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.8h, v29.8h, v0.h[5]\n"
       "fmla v14.8h, v29.8h, v1.h[5]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x198]\n"
       "fmla v18.8h, v29.8h, v2.h[5]\n"
       "fmla v22.8h, v29.8h, v3.h[5]\n"
-      "ldr x20, [x17, #0x198]\n"
       "fmla v26.8h, v29.8h, v4.h[5]\n"
       "ldr d29, [x17, #0x180]\n"
       "fmla v11.8h, v28.8h, v0.h[5]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.8h, v28.8h, v1.h[5]\n"
+      "ldr x21, [x17, #0x1a8]\n"
       "fmla v19.8h, v28.8h, v2.h[5]\n"
-      "mov v29.d[1], x21\n"
       "fmla v23.8h, v28.8h, v3.h[5]\n"
-      "ldr x21, [x17, #0x1a8]\n"
       "fmla v27.8h, v28.8h, v4.h[5]\n"
       "ldr d28, [x17, #0x190]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.8h, v29.8h, v0.h[6]\n"
       "fmla v12.8h, v29.8h, v1.h[6]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x1b8]\n"
       "fmla v16.8h, v29.8h, v2.h[6]\n"
       "fmla v20.8h, v29.8h, v3.h[6]\n"
-      "ldr x20, [x17, #0x1b8]\n"
       "fmla v24.8h, v29.8h, v4.h[6]\n"
       "ldr d29, [x17, #0x1a0]\n"
       "fmla v9.8h, v28.8h, v0.h[6]\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.8h, v28.8h, v1.h[6]\n"
+      "ldr x21, [x17, #0x1c8]\n"
       "fmla v17.8h, v28.8h, v2.h[6]\n"
-      "mov v29.d[1], x21\n"
       "fmla v21.8h, v28.8h, v3.h[6]\n"
-      "ldr x21, [x17, #0x1c8]\n"
       "fmla v25.8h, v28.8h, v4.h[6]\n"
       "ldr d28, [x17, #0x1b0]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.8h, v29.8h, v0.h[6]\n"
       "fmla v14.8h, v29.8h, v1.h[6]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x1d8]\n"
       "fmla v18.8h, v29.8h, v2.h[6]\n"
       "fmla v22.8h, v29.8h, v3.h[6]\n"
-      "ldr x20, [x17, #0x1d8]\n"
       "fmla v26.8h, v29.8h, v4.h[6]\n"
       "ldr d29, [x17, #0x1c0]\n"
       "fmla v11.8h, v28.8h, v0.h[6]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.8h, v28.8h, v1.h[6]\n"
+      "ldr x21, [x17, #0x1e8]\n"
       "fmla v19.8h, v28.8h, v2.h[6]\n"
-      "mov v29.d[1], x21\n"
       "fmla v23.8h, v28.8h, v3.h[6]\n"
-      "ldr x21, [x17, #0x1e8]\n"
       "fmla v27.8h, v28.8h, v4.h[6]\n"
       "ldr d28, [x17, #0x1d0]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.8h, v29.8h, v0.h[7]\n"
       "fmla v12.8h, v29.8h, v1.h[7]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x1f8]\n"
       "fmla v16.8h, v29.8h, v2.h[7]\n"
       "fmla v20.8h, v29.8h, v3.h[7]\n"
-      "ldr x20, [x17, #0x1f8]\n"
       "fmla v24.8h, v29.8h, v4.h[7]\n"
       "ldr d29, [x17, #0x1e0]\n"
       "fmla v9.8h, v28.8h, v0.h[7]\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.8h, v28.8h, v1.h[7]\n"
       "fmla v17.8h, v28.8h, v2.h[7]\n"
-      "mov v29.d[1], x21\n"
       "fmla v21.8h, v28.8h, v3.h[7]\n"
       "fmla v25.8h, v28.8h, v4.h[7]\n"
       "ldr d28, [x17, #0x1f0]\n"
+      "mov v28.d[1], x20\n"
       "add x17, x17, #0x200\n"
       "fmla v10.8h, v29.8h, v0.h[7]\n"
-      "fmla v14.8h, v29.8h, v1.h[7]\n"
       "ldr x21, [x17, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "fmla v14.8h, v29.8h, v1.h[7]\n"
+      "ldr x20, [x17, #0x18]\n"
       "fmla v18.8h, v29.8h, v2.h[7]\n"
       "fmla v22.8h, v29.8h, v3.h[7]\n"
-      "ldr x20, [x17, #0x18]\n"
       "fmla v26.8h, v29.8h, v4.h[7]\n"
       "ldr d6, [x17, #0x0]\n"
       "fmla v11.8h, v28.8h, v0.h[7]\n"
@@ -4029,8 +4027,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr h1, [x10], #0x2\n"
       "ldr h0, [x9], #0x2\n"
       "ldr q29, [x17, #0x0]\n"
-      "ldr q28, [x17, #0x10]\n"
       "fmla v8.8h, v29.8h, v4.h[0]\n"
+      "ldr q28, [x17, #0x10]\n"
       "fmla v12.8h, v29.8h, v3.h[0]\n"
       "fmla v16.8h, v29.8h, v2.h[0]\n"
       "fmla v20.8h, v29.8h, v1.h[0]\n"
@@ -4060,40 +4058,40 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "cmp x15, x20\n"
       "bne 219b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "add x25, x16, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "add x23, x24, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
+      "prfm pstl1keep, [x16, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 227f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v28.8h }, [x20]\n"
+      "fmin v8.8h, v8.8h, v28.8h\n"
+      "fmin v9.8h, v9.8h, v28.8h\n"
+      "fmin v10.8h, v10.8h, v28.8h\n"
+      "fmin v11.8h, v11.8h, v28.8h\n"
+      "fmin v12.8h, v12.8h, v28.8h\n"
+      "fmin v13.8h, v13.8h, v28.8h\n"
+      "fmin v14.8h, v14.8h, v28.8h\n"
+      "fmin v15.8h, v15.8h, v28.8h\n"
+      "fmin v16.8h, v16.8h, v28.8h\n"
+      "fmin v17.8h, v17.8h, v28.8h\n"
+      "fmin v18.8h, v18.8h, v28.8h\n"
+      "fmin v19.8h, v19.8h, v28.8h\n"
+      "fmin v20.8h, v20.8h, v28.8h\n"
+      "fmin v21.8h, v21.8h, v28.8h\n"
+      "fmin v22.8h, v22.8h, v28.8h\n"
+      "fmin v23.8h, v23.8h, v28.8h\n"
+      "fmin v24.8h, v24.8h, v28.8h\n"
+      "fmin v25.8h, v25.8h, v28.8h\n"
+      "fmin v26.8h, v26.8h, v28.8h\n"
+      "fmin v27.8h, v27.8h, v28.8h\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v29.8h }, [x21]\n"
       "ld1r { v28.8h }, [x20]\n"
-      "fmin v8.8h, v8.8h, v29.8h\n"
-      "fmin v9.8h, v9.8h, v29.8h\n"
-      "fmin v10.8h, v10.8h, v29.8h\n"
-      "fmin v11.8h, v11.8h, v29.8h\n"
-      "fmin v12.8h, v12.8h, v29.8h\n"
-      "fmin v13.8h, v13.8h, v29.8h\n"
-      "fmin v14.8h, v14.8h, v29.8h\n"
-      "fmin v15.8h, v15.8h, v29.8h\n"
-      "fmin v16.8h, v16.8h, v29.8h\n"
-      "fmin v17.8h, v17.8h, v29.8h\n"
-      "fmin v18.8h, v18.8h, v29.8h\n"
-      "fmin v19.8h, v19.8h, v29.8h\n"
-      "fmin v20.8h, v20.8h, v29.8h\n"
-      "fmin v21.8h, v21.8h, v29.8h\n"
-      "fmin v22.8h, v22.8h, v29.8h\n"
-      "fmin v23.8h, v23.8h, v29.8h\n"
-      "fmin v24.8h, v24.8h, v29.8h\n"
-      "fmin v25.8h, v25.8h, v29.8h\n"
-      "fmin v26.8h, v26.8h, v29.8h\n"
-      "fmin v27.8h, v27.8h, v29.8h\n"
       "fmax v8.8h, v8.8h, v28.8h\n"
       "fmax v9.8h, v9.8h, v28.8h\n"
       "fmax v10.8h, v10.8h, v28.8h\n"
@@ -4120,223 +4118,223 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "tbz x8, #4, 235f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
       "st1 { v9.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v25.8h }, [x23], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v25.8h }, [x22], #0x10\n"
       "tbz x8, #3, 231f\n"
       "st1 { v10.8h }, [x16], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
-      "st1 { v26.8h }, [x23], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
+      "st1 { v26.8h }, [x22], #0x10\n"
       "tbz x8, #2, 229f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x8, #1, 228f\n"
       "st1 { v11.s }[2], [x16], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
-      "st1 { v27.s }[2], [x23], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
+      "st1 { v27.s }[2], [x22], #0x4\n"
       "tbz x8, #0, 243f\n"
       "st1 { v11.h }[6], [x16]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
-      "st1 { v27.h }[6], [x23]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
+      "st1 { v27.h }[6], [x22]\n"
       "b 243f\n"
       "228:"  // Height 5: Partial direct writeback: partial_1_28
       "tbz x8, #0, 243f\n"
       "st1 { v11.h }[4], [x16]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
-      "st1 { v27.h }[4], [x23]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
+      "st1 { v27.h }[4], [x22]\n"
       "b 243f\n"
       "229:"  // Height 5: Partial direct writeback: partial_2_24
       "tbz x8, #1, 230f\n"
       "str s11, [x16], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
-      "str s27, [x23], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
+      "str s27, [x22], #0x4\n"
       "tbz x8, #0, 243f\n"
       "st1 { v11.h }[2], [x16]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
-      "st1 { v27.h }[2], [x23]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
+      "st1 { v27.h }[2], [x22]\n"
       "b 243f\n"
       "230:"  // Height 5: Partial direct writeback: partial_1_24
       "tbz x8, #0, 243f\n"
       "str h11, [x16, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
-      "str h27, [x23, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
+      "str h27, [x22, #0x0]\n"
       "b 243f\n"
       "231:"  // Height 5: Partial direct writeback: partial_4_16
       "tbz x8, #2, 233f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x8, #1, 232f\n"
       "st1 { v10.s }[2], [x16], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
-      "st1 { v26.s }[2], [x23], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
+      "st1 { v26.s }[2], [x22], #0x4\n"
       "tbz x8, #0, 243f\n"
       "st1 { v10.h }[6], [x16]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
-      "st1 { v26.h }[6], [x23]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
+      "st1 { v26.h }[6], [x22]\n"
       "b 243f\n"
       "232:"  // Height 5: Partial direct writeback: partial_1_20
       "tbz x8, #0, 243f\n"
       "st1 { v10.h }[4], [x16]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
-      "st1 { v26.h }[4], [x23]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
+      "st1 { v26.h }[4], [x22]\n"
       "b 243f\n"
       "233:"  // Height 5: Partial direct writeback: partial_2_16
       "tbz x8, #1, 234f\n"
       "str s10, [x16], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
-      "str s26, [x23], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
+      "str s26, [x22], #0x4\n"
       "tbz x8, #0, 243f\n"
       "st1 { v10.h }[2], [x16]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
-      "st1 { v26.h }[2], [x23]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
+      "st1 { v26.h }[2], [x22]\n"
       "b 243f\n"
       "234:"  // Height 5: Partial direct writeback: partial_1_16
       "tbz x8, #0, 243f\n"
       "str h10, [x16, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
-      "str h26, [x23, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
+      "str h26, [x22, #0x0]\n"
       "b 243f\n"
       "235:"  // Height 5: Partial direct writeback: partial_8_0
       "tbz x8, #3, 239f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
       "tbz x8, #2, 237f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x8, #1, 236f\n"
       "st1 { v9.s }[2], [x16], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
-      "st1 { v25.s }[2], [x23], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
+      "st1 { v25.s }[2], [x22], #0x4\n"
       "tbz x8, #0, 243f\n"
       "st1 { v9.h }[6], [x16]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
-      "st1 { v25.h }[6], [x23]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
+      "st1 { v25.h }[6], [x22]\n"
       "b 243f\n"
       "236:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x8, #0, 243f\n"
       "st1 { v9.h }[4], [x16]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
-      "st1 { v25.h }[4], [x23]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
+      "st1 { v25.h }[4], [x22]\n"
       "b 243f\n"
       "237:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x8, #1, 238f\n"
       "str s9, [x16], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
-      "str s25, [x23], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
+      "str s25, [x22], #0x4\n"
       "tbz x8, #0, 243f\n"
       "st1 { v9.h }[2], [x16]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
-      "st1 { v25.h }[2], [x23]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
+      "st1 { v25.h }[2], [x22]\n"
       "b 243f\n"
       "238:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x8, #0, 243f\n"
       "str h9, [x16, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
-      "str h25, [x23, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
+      "str h25, [x22, #0x0]\n"
       "b 243f\n"
       "239:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x8, #2, 241f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x8, #1, 240f\n"
       "st1 { v8.s }[2], [x16], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
       "tbz x8, #0, 243f\n"
       "st1 { v8.h }[6], [x16]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
-      "st1 { v24.h }[6], [x23]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
+      "st1 { v24.h }[6], [x22]\n"
       "b 243f\n"
       "240:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x8, #0, 243f\n"
       "st1 { v8.h }[4], [x16]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
-      "st1 { v24.h }[4], [x23]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
+      "st1 { v24.h }[4], [x22]\n"
       "b 243f\n"
       "241:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x8, #1, 242f\n"
       "str s8, [x16], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
       "tbz x8, #0, 243f\n"
       "st1 { v8.h }[2], [x16]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
-      "st1 { v24.h }[2], [x23]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
+      "st1 { v24.h }[2], [x22]\n"
       "b 243f\n"
       "242:"  // Height 5: Partial direct writeback: partial_1_0
       "str h8, [x16, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
-      "str h24, [x23, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
+      "str h24, [x22, #0x0]\n"
       "243:"  // Height 5: Partial direct writeback: Done
       "b 245f\n"
       "244:"  // Height 5: Full writeback
@@ -4345,22 +4343,22 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "245:"  // Height 5: Writeback done
       "subs x8, x8, #0x20\n"
       "bgt 198b\n"
@@ -4368,24 +4366,23 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "246:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "mov x20, #0xc\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "madd x20, x21, x20, x16\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "247:"  // Height 6: Column loop
       "cbz x7, 248f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
       "mov v16.16b, v8.16b\n"
+      "add x7, x7, #0x40\n"
       "mov v17.16b, v9.16b\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
@@ -4405,281 +4402,281 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "248:"  // Height 6: no bias
       "tbz %x[flags], #0, 266f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x20\n"
-      "add x26, x16, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x16, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "cmp x8, #0x20\n"
+      "add x21, x22, x20, LSL #1\n"
       "bge 265f\n"
       "tbz x8, #4, 256f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
-      "ld1 { v28.8h }, [x22], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
+      "ld1 { v28.8h }, [x21], #0x10\n"
       "ld1 { v9.8h }, [x16], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
-      "ld1 { v25.8h }, [x23], #0x10\n"
-      "ld1 { v29.8h }, [x22], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
+      "ld1 { v25.8h }, [x22], #0x10\n"
+      "ld1 { v29.8h }, [x21], #0x10\n"
       "tbz x8, #3, 252f\n"
       "ld1 { v10.8h }, [x16], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
-      "ld1 { v26.8h }, [x23], #0x10\n"
-      "ld1 { v30.8h }, [x22], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
+      "ld1 { v26.8h }, [x22], #0x10\n"
+      "ld1 { v30.8h }, [x21], #0x10\n"
       "tbz x8, #2, 250f\n"
       "ldr d11, [x16], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x8, #1, 249f\n"
       "ld1 { v11.s }[2], [x16], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
-      "ld1 { v27.s }[2], [x23], #0x4\n"
-      "ld1 { v31.s }[2], [x22], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
+      "ld1 { v27.s }[2], [x22], #0x4\n"
+      "ld1 { v31.s }[2], [x21], #0x4\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v11.h }[6], [x16]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
-      "ld1 { v27.h }[6], [x23]\n"
-      "ld1 { v31.h }[6], [x22]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
+      "ld1 { v27.h }[6], [x22]\n"
+      "ld1 { v31.h }[6], [x21]\n"
       "b 264f\n"
       "249:"  // Height 6: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v11.h }[4], [x16]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
-      "ld1 { v27.h }[4], [x23]\n"
-      "ld1 { v31.h }[4], [x22]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
+      "ld1 { v27.h }[4], [x22]\n"
+      "ld1 { v31.h }[4], [x21]\n"
       "b 264f\n"
       "250:"  // Height 6: Partial accumulate: partial_2_24
       "tbz x8, #1, 251f\n"
       "ldr s11, [x16], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s15, [x26], #0x4\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
-      "ldr s27, [x23], #0x4\n"
-      "ldr s31, [x22], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s27, [x22], #0x4\n"
+      "ldr s31, [x21], #0x4\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v11.h }[2], [x16]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
-      "ld1 { v27.h }[2], [x23]\n"
-      "ld1 { v31.h }[2], [x22]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
+      "ld1 { v27.h }[2], [x22]\n"
+      "ld1 { v31.h }[2], [x21]\n"
       "b 264f\n"
       "251:"  // Height 6: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x8, #0, 264f\n"
       "ldr h11, [x16, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
-      "ldr h27, [x23, #0x0]\n"
-      "ldr h31, [x22, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
+      "ldr h27, [x22, #0x0]\n"
+      "ldr h31, [x21, #0x0]\n"
       "b 264f\n"
       "252:"  // Height 6: Partial accumulate: partial_4_16
       "tbz x8, #2, 254f\n"
       "ldr d10, [x16], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x8, #1, 253f\n"
       "ld1 { v10.s }[2], [x16], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
-      "ld1 { v26.s }[2], [x23], #0x4\n"
-      "ld1 { v30.s }[2], [x22], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
+      "ld1 { v26.s }[2], [x22], #0x4\n"
+      "ld1 { v30.s }[2], [x21], #0x4\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v10.h }[6], [x16]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
-      "ld1 { v26.h }[6], [x23]\n"
-      "ld1 { v30.h }[6], [x22]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
+      "ld1 { v26.h }[6], [x22]\n"
+      "ld1 { v30.h }[6], [x21]\n"
       "b 264f\n"
       "253:"  // Height 6: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v10.h }[4], [x16]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
-      "ld1 { v26.h }[4], [x23]\n"
-      "ld1 { v30.h }[4], [x22]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
+      "ld1 { v26.h }[4], [x22]\n"
+      "ld1 { v30.h }[4], [x21]\n"
       "b 264f\n"
       "254:"  // Height 6: Partial accumulate: partial_2_16
       "tbz x8, #1, 255f\n"
       "ldr s10, [x16], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s14, [x26], #0x4\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr s26, [x23], #0x4\n"
-      "ldr s30, [x22], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
+      "ldr s26, [x22], #0x4\n"
+      "ldr s30, [x21], #0x4\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v10.h }[2], [x16]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
-      "ld1 { v26.h }[2], [x23]\n"
-      "ld1 { v30.h }[2], [x22]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
+      "ld1 { v26.h }[2], [x22]\n"
+      "ld1 { v30.h }[2], [x21]\n"
       "b 264f\n"
       "255:"  // Height 6: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x8, #0, 264f\n"
       "ldr h10, [x16, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
-      "ldr h26, [x23, #0x0]\n"
-      "ldr h30, [x22, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
+      "ldr h26, [x22, #0x0]\n"
+      "ldr h30, [x21, #0x0]\n"
       "b 264f\n"
       "256:"  // Height 6: Partial accumulate: partial_8_0
       "tbz x8, #3, 260f\n"
       "ld1 { v8.8h }, [x16], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
-      "ld1 { v28.8h }, [x22], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
+      "ld1 { v28.8h }, [x21], #0x10\n"
       "tbz x8, #2, 258f\n"
       "ldr d9, [x16], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x8, #1, 257f\n"
       "ld1 { v9.s }[2], [x16], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
-      "ld1 { v25.s }[2], [x23], #0x4\n"
-      "ld1 { v29.s }[2], [x22], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
+      "ld1 { v25.s }[2], [x22], #0x4\n"
+      "ld1 { v29.s }[2], [x21], #0x4\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v9.h }[6], [x16]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
-      "ld1 { v25.h }[6], [x23]\n"
-      "ld1 { v29.h }[6], [x22]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
+      "ld1 { v25.h }[6], [x22]\n"
+      "ld1 { v29.h }[6], [x21]\n"
       "b 264f\n"
       "257:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v9.h }[4], [x16]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
-      "ld1 { v25.h }[4], [x23]\n"
-      "ld1 { v29.h }[4], [x22]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
+      "ld1 { v25.h }[4], [x22]\n"
+      "ld1 { v29.h }[4], [x21]\n"
       "b 264f\n"
       "258:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x8, #1, 259f\n"
       "ldr s9, [x16], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s13, [x26], #0x4\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
-      "ldr s25, [x23], #0x4\n"
-      "ldr s29, [x22], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
+      "ldr s25, [x22], #0x4\n"
+      "ldr s29, [x21], #0x4\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v9.h }[2], [x16]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
-      "ld1 { v25.h }[2], [x23]\n"
-      "ld1 { v29.h }[2], [x22]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
+      "ld1 { v25.h }[2], [x22]\n"
+      "ld1 { v29.h }[2], [x21]\n"
       "b 264f\n"
       "259:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x8, #0, 264f\n"
       "ldr h9, [x16, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
-      "ldr h25, [x23, #0x0]\n"
-      "ldr h29, [x22, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
+      "ldr h25, [x22, #0x0]\n"
+      "ldr h29, [x21, #0x0]\n"
       "b 264f\n"
       "260:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x8, #2, 262f\n"
       "ldr d8, [x16], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x8, #1, 261f\n"
       "ld1 { v8.s }[2], [x16], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
-      "ld1 { v24.s }[2], [x23], #0x4\n"
-      "ld1 { v28.s }[2], [x22], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
+      "ld1 { v24.s }[2], [x22], #0x4\n"
+      "ld1 { v28.s }[2], [x21], #0x4\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v8.h }[6], [x16]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
-      "ld1 { v24.h }[6], [x23]\n"
-      "ld1 { v28.h }[6], [x22]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
+      "ld1 { v24.h }[6], [x22]\n"
+      "ld1 { v28.h }[6], [x21]\n"
       "b 264f\n"
       "261:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v8.h }[4], [x16]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
-      "ld1 { v24.h }[4], [x23]\n"
-      "ld1 { v28.h }[4], [x22]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
+      "ld1 { v24.h }[4], [x22]\n"
+      "ld1 { v28.h }[4], [x21]\n"
       "b 264f\n"
       "262:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x8, #1, 263f\n"
       "ldr s8, [x16], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s12, [x26], #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s28, [x22], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
+      "ldr s24, [x22], #0x4\n"
+      "ldr s28, [x21], #0x4\n"
       "tbz x8, #0, 264f\n"
       "ld1 { v8.h }[2], [x16]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
-      "ld1 { v24.h }[2], [x23]\n"
-      "ld1 { v28.h }[2], [x22]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
+      "ld1 { v24.h }[2], [x22]\n"
+      "ld1 { v28.h }[2], [x21]\n"
       "b 264f\n"
       "263:"  // Height 6: Partial accumulate: partial_1_0
       "ldr h8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h12, [x26, #0x0]\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
-      "ldr h24, [x23, #0x0]\n"
-      "ldr h28, [x22, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
+      "ldr h24, [x22, #0x0]\n"
+      "ldr h28, [x21, #0x0]\n"
       "264:"  // Height 6: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 267f\n"
@@ -4688,26 +4685,26 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "b 267f\n"
       "266:"  // Height 6: no accumulate
       "movi v8.16b, #0x0\n"
@@ -4738,8 +4735,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "mov x15, #0x0\n"
       "268:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 269f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -4792,290 +4789,290 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "fmla v28.8h, v6.8h, v5.h[0]\n"
       "ldr d6, [x17, #0x20]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
-      "add x10, x10, #0x10\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
-      "add x9, x9, #0x10\n"
+      "ldr x21, [x17, #0x48]\n"
       "fmla v17.8h, v7.8h, v2.h[0]\n"
-      "mov v6.d[1], x21\n"
+      "add x10, x10, #0x10\n"
       "fmla v21.8h, v7.8h, v3.h[0]\n"
-      "ldr x21, [x17, #0x48]\n"
+      "add x9, x9, #0x10\n"
       "fmla v25.8h, v7.8h, v4.h[0]\n"
       "add x28, x28, #0x10\n"
       "fmla v29.8h, v7.8h, v5.h[0]\n"
       "ldr d7, [x17, #0x30]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.8h, v6.8h, v0.h[0]\n"
-      "ldr x27, [x13, #0x8]\n"
       "fmla v14.8h, v6.8h, v1.h[0]\n"
-      "ldr x26, [x12, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x58]\n"
       "fmla v18.8h, v6.8h, v2.h[0]\n"
+      "ldr x27, [x13, #0x8]\n"
       "fmla v22.8h, v6.8h, v3.h[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x26, [x12, #0x8]\n"
       "fmla v26.8h, v6.8h, v4.h[0]\n"
       "ldr x25, [x11, #0x8]\n"
       "fmla v30.8h, v6.8h, v5.h[0]\n"
       "ldr d6, [x17, #0x40]\n"
       "fmla v11.8h, v7.8h, v0.h[0]\n"
-      "ldr x24, [x10, #0x8]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.8h, v7.8h, v1.h[0]\n"
-      "ldr x23, [x9, #0x8]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v19.8h, v7.8h, v2.h[0]\n"
-      "mov v6.d[1], x21\n"
+      "ldr x24, [x10, #0x8]\n"
       "fmla v23.8h, v7.8h, v3.h[0]\n"
-      "ldr x21, [x17, #0x68]\n"
+      "ldr x23, [x9, #0x8]\n"
       "fmla v27.8h, v7.8h, v4.h[0]\n"
       "ldr x22, [x28, #0x8]\n"
       "fmla v31.8h, v7.8h, v5.h[0]\n"
       "ldr d7, [x17, #0x50]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.8h, v6.8h, v0.h[1]\n"
-      "sub x14, x14, #0x8\n"
       "fmla v12.8h, v6.8h, v1.h[1]\n"
-      "cmp x14, #0x10\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x78]\n"
       "fmla v16.8h, v6.8h, v2.h[1]\n"
+      "sub x14, x14, #0x8\n"
       "fmla v20.8h, v6.8h, v3.h[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "cmp x14, #0x10\n"
       "fmla v24.8h, v6.8h, v4.h[1]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v28.8h, v6.8h, v5.h[1]\n"
       "ldr d6, [x17, #0x60]\n"
       "fmla v9.8h, v7.8h, v0.h[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x17, #0x88]\n"
       "fmla v17.8h, v7.8h, v2.h[1]\n"
-      "mov v6.d[1], x21\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "fmla v21.8h, v7.8h, v3.h[1]\n"
-      "ldr x21, [x17, #0x88]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "fmla v25.8h, v7.8h, v4.h[1]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       "fmla v29.8h, v7.8h, v5.h[1]\n"
       "ldr d7, [x17, #0x70]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.8h, v6.8h, v0.h[1]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v14.8h, v6.8h, v1.h[1]\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x98]\n"
       "fmla v18.8h, v6.8h, v2.h[1]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v22.8h, v6.8h, v3.h[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v26.8h, v6.8h, v4.h[1]\n"
       "fmla v30.8h, v6.8h, v5.h[1]\n"
       "ldr d6, [x17, #0x80]\n"
       "fmla v11.8h, v7.8h, v0.h[1]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.8h, v7.8h, v1.h[1]\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v19.8h, v7.8h, v2.h[1]\n"
-      "mov v6.d[1], x21\n"
       "fmla v23.8h, v7.8h, v3.h[1]\n"
-      "ldr x21, [x17, #0xa8]\n"
       "fmla v27.8h, v7.8h, v4.h[1]\n"
       "fmla v31.8h, v7.8h, v5.h[1]\n"
       "ldr d7, [x17, #0x90]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.8h, v6.8h, v0.h[2]\n"
       "fmla v12.8h, v6.8h, v1.h[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0xb8]\n"
       "fmla v16.8h, v6.8h, v2.h[2]\n"
       "fmla v20.8h, v6.8h, v3.h[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
       "fmla v24.8h, v6.8h, v4.h[2]\n"
       "fmla v28.8h, v6.8h, v5.h[2]\n"
       "ldr d6, [x17, #0xa0]\n"
       "fmla v9.8h, v7.8h, v0.h[2]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[2]\n"
+      "ldr x21, [x17, #0xc8]\n"
       "fmla v17.8h, v7.8h, v2.h[2]\n"
-      "mov v6.d[1], x21\n"
       "fmla v21.8h, v7.8h, v3.h[2]\n"
-      "ldr x21, [x17, #0xc8]\n"
       "fmla v25.8h, v7.8h, v4.h[2]\n"
       "fmla v29.8h, v7.8h, v5.h[2]\n"
       "ldr d7, [x17, #0xb0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.8h, v6.8h, v0.h[2]\n"
       "fmla v14.8h, v6.8h, v1.h[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0xd8]\n"
       "fmla v18.8h, v6.8h, v2.h[2]\n"
       "fmla v22.8h, v6.8h, v3.h[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
       "fmla v26.8h, v6.8h, v4.h[2]\n"
       "fmla v30.8h, v6.8h, v5.h[2]\n"
       "ldr d6, [x17, #0xc0]\n"
       "fmla v11.8h, v7.8h, v0.h[2]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.8h, v7.8h, v1.h[2]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v19.8h, v7.8h, v2.h[2]\n"
-      "mov v6.d[1], x21\n"
       "fmla v23.8h, v7.8h, v3.h[2]\n"
-      "ldr x21, [x17, #0xe8]\n"
       "fmla v27.8h, v7.8h, v4.h[2]\n"
       "fmla v31.8h, v7.8h, v5.h[2]\n"
       "ldr d7, [x17, #0xd0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.8h, v6.8h, v0.h[3]\n"
       "fmla v12.8h, v6.8h, v1.h[3]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0xf8]\n"
       "fmla v16.8h, v6.8h, v2.h[3]\n"
       "fmla v20.8h, v6.8h, v3.h[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
       "fmla v24.8h, v6.8h, v4.h[3]\n"
       "fmla v28.8h, v6.8h, v5.h[3]\n"
       "ldr d6, [x17, #0xe0]\n"
       "fmla v9.8h, v7.8h, v0.h[3]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[3]\n"
+      "ldr x21, [x17, #0x108]\n"
       "fmla v17.8h, v7.8h, v2.h[3]\n"
-      "mov v6.d[1], x21\n"
       "fmla v21.8h, v7.8h, v3.h[3]\n"
-      "ldr x21, [x17, #0x108]\n"
       "fmla v25.8h, v7.8h, v4.h[3]\n"
       "fmla v29.8h, v7.8h, v5.h[3]\n"
       "ldr d7, [x17, #0xf0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.8h, v6.8h, v0.h[3]\n"
       "fmla v14.8h, v6.8h, v1.h[3]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x118]\n"
       "fmla v18.8h, v6.8h, v2.h[3]\n"
       "fmla v22.8h, v6.8h, v3.h[3]\n"
-      "ldr x20, [x17, #0x118]\n"
       "fmla v26.8h, v6.8h, v4.h[3]\n"
       "fmla v30.8h, v6.8h, v5.h[3]\n"
       "ldr d6, [x17, #0x100]\n"
       "fmla v11.8h, v7.8h, v0.h[3]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.8h, v7.8h, v1.h[3]\n"
+      "ldr x21, [x17, #0x128]\n"
       "fmla v19.8h, v7.8h, v2.h[3]\n"
-      "mov v6.d[1], x21\n"
       "fmla v23.8h, v7.8h, v3.h[3]\n"
-      "ldr x21, [x17, #0x128]\n"
       "fmla v27.8h, v7.8h, v4.h[3]\n"
       "fmla v31.8h, v7.8h, v5.h[3]\n"
       "ldr d7, [x17, #0x110]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.8h, v6.8h, v0.h[4]\n"
       "fmla v12.8h, v6.8h, v1.h[4]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x138]\n"
       "fmla v16.8h, v6.8h, v2.h[4]\n"
       "fmla v20.8h, v6.8h, v3.h[4]\n"
-      "ldr x20, [x17, #0x138]\n"
       "fmla v24.8h, v6.8h, v4.h[4]\n"
       "fmla v28.8h, v6.8h, v5.h[4]\n"
       "ldr d6, [x17, #0x120]\n"
       "fmla v9.8h, v7.8h, v0.h[4]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[4]\n"
+      "ldr x21, [x17, #0x148]\n"
       "fmla v17.8h, v7.8h, v2.h[4]\n"
-      "mov v6.d[1], x21\n"
       "fmla v21.8h, v7.8h, v3.h[4]\n"
-      "ldr x21, [x17, #0x148]\n"
       "fmla v25.8h, v7.8h, v4.h[4]\n"
       "fmla v29.8h, v7.8h, v5.h[4]\n"
       "ldr d7, [x17, #0x130]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.8h, v6.8h, v0.h[4]\n"
       "fmla v14.8h, v6.8h, v1.h[4]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x158]\n"
       "fmla v18.8h, v6.8h, v2.h[4]\n"
       "fmla v22.8h, v6.8h, v3.h[4]\n"
-      "ldr x20, [x17, #0x158]\n"
       "fmla v26.8h, v6.8h, v4.h[4]\n"
       "fmla v30.8h, v6.8h, v5.h[4]\n"
       "ldr d6, [x17, #0x140]\n"
       "fmla v11.8h, v7.8h, v0.h[4]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.8h, v7.8h, v1.h[4]\n"
+      "ldr x21, [x17, #0x168]\n"
       "fmla v19.8h, v7.8h, v2.h[4]\n"
-      "mov v6.d[1], x21\n"
       "fmla v23.8h, v7.8h, v3.h[4]\n"
-      "ldr x21, [x17, #0x168]\n"
       "fmla v27.8h, v7.8h, v4.h[4]\n"
       "fmla v31.8h, v7.8h, v5.h[4]\n"
       "ldr d7, [x17, #0x150]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.8h, v6.8h, v0.h[5]\n"
       "fmla v12.8h, v6.8h, v1.h[5]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x178]\n"
       "fmla v16.8h, v6.8h, v2.h[5]\n"
       "fmla v20.8h, v6.8h, v3.h[5]\n"
-      "ldr x20, [x17, #0x178]\n"
       "fmla v24.8h, v6.8h, v4.h[5]\n"
       "fmla v28.8h, v6.8h, v5.h[5]\n"
       "ldr d6, [x17, #0x160]\n"
       "fmla v9.8h, v7.8h, v0.h[5]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[5]\n"
+      "ldr x21, [x17, #0x188]\n"
       "fmla v17.8h, v7.8h, v2.h[5]\n"
-      "mov v6.d[1], x21\n"
       "fmla v21.8h, v7.8h, v3.h[5]\n"
-      "ldr x21, [x17, #0x188]\n"
       "fmla v25.8h, v7.8h, v4.h[5]\n"
       "fmla v29.8h, v7.8h, v5.h[5]\n"
       "ldr d7, [x17, #0x170]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.8h, v6.8h, v0.h[5]\n"
       "fmla v14.8h, v6.8h, v1.h[5]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x198]\n"
       "fmla v18.8h, v6.8h, v2.h[5]\n"
       "fmla v22.8h, v6.8h, v3.h[5]\n"
-      "ldr x20, [x17, #0x198]\n"
       "fmla v26.8h, v6.8h, v4.h[5]\n"
       "fmla v30.8h, v6.8h, v5.h[5]\n"
       "ldr d6, [x17, #0x180]\n"
       "fmla v11.8h, v7.8h, v0.h[5]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.8h, v7.8h, v1.h[5]\n"
+      "ldr x21, [x17, #0x1a8]\n"
       "fmla v19.8h, v7.8h, v2.h[5]\n"
-      "mov v6.d[1], x21\n"
       "fmla v23.8h, v7.8h, v3.h[5]\n"
-      "ldr x21, [x17, #0x1a8]\n"
       "fmla v27.8h, v7.8h, v4.h[5]\n"
       "fmla v31.8h, v7.8h, v5.h[5]\n"
       "ldr d7, [x17, #0x190]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.8h, v6.8h, v0.h[6]\n"
       "fmla v12.8h, v6.8h, v1.h[6]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x1b8]\n"
       "fmla v16.8h, v6.8h, v2.h[6]\n"
       "fmla v20.8h, v6.8h, v3.h[6]\n"
-      "ldr x20, [x17, #0x1b8]\n"
       "fmla v24.8h, v6.8h, v4.h[6]\n"
       "fmla v28.8h, v6.8h, v5.h[6]\n"
       "ldr d6, [x17, #0x1a0]\n"
       "fmla v9.8h, v7.8h, v0.h[6]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[6]\n"
+      "ldr x21, [x17, #0x1c8]\n"
       "fmla v17.8h, v7.8h, v2.h[6]\n"
-      "mov v6.d[1], x21\n"
       "fmla v21.8h, v7.8h, v3.h[6]\n"
-      "ldr x21, [x17, #0x1c8]\n"
       "fmla v25.8h, v7.8h, v4.h[6]\n"
       "fmla v29.8h, v7.8h, v5.h[6]\n"
       "ldr d7, [x17, #0x1b0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.8h, v6.8h, v0.h[6]\n"
       "fmla v14.8h, v6.8h, v1.h[6]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x1d8]\n"
       "fmla v18.8h, v6.8h, v2.h[6]\n"
       "fmla v22.8h, v6.8h, v3.h[6]\n"
-      "ldr x20, [x17, #0x1d8]\n"
       "fmla v26.8h, v6.8h, v4.h[6]\n"
       "fmla v30.8h, v6.8h, v5.h[6]\n"
       "ldr d6, [x17, #0x1c0]\n"
       "fmla v11.8h, v7.8h, v0.h[6]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.8h, v7.8h, v1.h[6]\n"
+      "ldr x21, [x17, #0x1e8]\n"
       "fmla v19.8h, v7.8h, v2.h[6]\n"
-      "mov v6.d[1], x21\n"
       "fmla v23.8h, v7.8h, v3.h[6]\n"
-      "ldr x21, [x17, #0x1e8]\n"
       "fmla v27.8h, v7.8h, v4.h[6]\n"
       "fmla v31.8h, v7.8h, v5.h[6]\n"
       "ldr d7, [x17, #0x1d0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.8h, v6.8h, v0.h[7]\n"
       "fmla v12.8h, v6.8h, v1.h[7]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x1f8]\n"
       "fmla v16.8h, v6.8h, v2.h[7]\n"
       "fmla v20.8h, v6.8h, v3.h[7]\n"
-      "ldr x20, [x17, #0x1f8]\n"
       "fmla v24.8h, v6.8h, v4.h[7]\n"
       "fmla v28.8h, v6.8h, v5.h[7]\n"
       "ldr d6, [x17, #0x1e0]\n"
       "fmla v9.8h, v7.8h, v0.h[7]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.8h, v7.8h, v1.h[7]\n"
       "fmla v17.8h, v7.8h, v2.h[7]\n"
-      "mov v6.d[1], x21\n"
       "fmla v21.8h, v7.8h, v3.h[7]\n"
       "fmla v25.8h, v7.8h, v4.h[7]\n"
       "fmla v29.8h, v7.8h, v5.h[7]\n"
       "ldr d7, [x17, #0x1f0]\n"
+      "mov v7.d[1], x20\n"
       "add x17, x17, #0x200\n"
       "fmla v10.8h, v6.8h, v0.h[7]\n"
-      "fmla v14.8h, v6.8h, v1.h[7]\n"
       "ldr x21, [x17, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "fmla v14.8h, v6.8h, v1.h[7]\n"
+      "ldr x20, [x17, #0x18]\n"
       "fmla v18.8h, v6.8h, v2.h[7]\n"
       "fmla v22.8h, v6.8h, v3.h[7]\n"
-      "ldr x20, [x17, #0x18]\n"
       "fmla v26.8h, v6.8h, v4.h[7]\n"
       "fmla v30.8h, v6.8h, v5.h[7]\n"
       "ldr d6, [x17, #0x0]\n"
@@ -5349,8 +5346,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "ldr h3, [x9], #0x2\n"
       "ldr h2, [x28], #0x2\n"
       "ldr q1, [x17, #0x0]\n"
-      "ldr q0, [x17, #0x10]\n"
       "fmla v8.8h, v1.8h, v7.h[0]\n"
+      "ldr q0, [x17, #0x10]\n"
       "fmla v12.8h, v1.8h, v6.h[0]\n"
       "fmla v16.8h, v1.8h, v5.h[0]\n"
       "fmla v20.8h, v1.8h, v4.h[0]\n"
@@ -5384,46 +5381,46 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "cmp x15, x20\n"
       "bne 268b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "add x25, x16, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "add x23, x24, x20, LSL #1\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
+      "prfm pstl1keep, [x16, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbz %x[flags], #1, 276f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.8h }, [x20]\n"
+      "fmin v8.8h, v8.8h, v0.8h\n"
+      "fmin v9.8h, v9.8h, v0.8h\n"
+      "fmin v10.8h, v10.8h, v0.8h\n"
+      "fmin v11.8h, v11.8h, v0.8h\n"
+      "fmin v12.8h, v12.8h, v0.8h\n"
+      "fmin v13.8h, v13.8h, v0.8h\n"
+      "fmin v14.8h, v14.8h, v0.8h\n"
+      "fmin v15.8h, v15.8h, v0.8h\n"
+      "fmin v16.8h, v16.8h, v0.8h\n"
+      "fmin v17.8h, v17.8h, v0.8h\n"
+      "fmin v18.8h, v18.8h, v0.8h\n"
+      "fmin v19.8h, v19.8h, v0.8h\n"
+      "fmin v20.8h, v20.8h, v0.8h\n"
+      "fmin v21.8h, v21.8h, v0.8h\n"
+      "fmin v22.8h, v22.8h, v0.8h\n"
+      "fmin v23.8h, v23.8h, v0.8h\n"
+      "fmin v24.8h, v24.8h, v0.8h\n"
+      "fmin v25.8h, v25.8h, v0.8h\n"
+      "fmin v26.8h, v26.8h, v0.8h\n"
+      "fmin v27.8h, v27.8h, v0.8h\n"
+      "fmin v28.8h, v28.8h, v0.8h\n"
+      "fmin v29.8h, v29.8h, v0.8h\n"
+      "fmin v30.8h, v30.8h, v0.8h\n"
+      "fmin v31.8h, v31.8h, v0.8h\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.8h }, [x21]\n"
       "ld1r { v0.8h }, [x20]\n"
-      "fmin v8.8h, v8.8h, v1.8h\n"
-      "fmin v9.8h, v9.8h, v1.8h\n"
-      "fmin v10.8h, v10.8h, v1.8h\n"
-      "fmin v11.8h, v11.8h, v1.8h\n"
-      "fmin v12.8h, v12.8h, v1.8h\n"
-      "fmin v13.8h, v13.8h, v1.8h\n"
-      "fmin v14.8h, v14.8h, v1.8h\n"
-      "fmin v15.8h, v15.8h, v1.8h\n"
-      "fmin v16.8h, v16.8h, v1.8h\n"
-      "fmin v17.8h, v17.8h, v1.8h\n"
-      "fmin v18.8h, v18.8h, v1.8h\n"
-      "fmin v19.8h, v19.8h, v1.8h\n"
-      "fmin v20.8h, v20.8h, v1.8h\n"
-      "fmin v21.8h, v21.8h, v1.8h\n"
-      "fmin v22.8h, v22.8h, v1.8h\n"
-      "fmin v23.8h, v23.8h, v1.8h\n"
-      "fmin v24.8h, v24.8h, v1.8h\n"
-      "fmin v25.8h, v25.8h, v1.8h\n"
-      "fmin v26.8h, v26.8h, v1.8h\n"
-      "fmin v27.8h, v27.8h, v1.8h\n"
-      "fmin v28.8h, v28.8h, v1.8h\n"
-      "fmin v29.8h, v29.8h, v1.8h\n"
-      "fmin v30.8h, v30.8h, v1.8h\n"
-      "fmin v31.8h, v31.8h, v1.8h\n"
       "fmax v8.8h, v8.8h, v0.8h\n"
       "fmax v9.8h, v9.8h, v0.8h\n"
       "fmax v10.8h, v10.8h, v0.8h\n"
@@ -5454,255 +5451,255 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "tbz x8, #4, 284f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
       "st1 { v9.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v25.8h }, [x23], #0x10\n"
-      "st1 { v28.8h }, [x22], #0x10\n"
-      "st1 { v29.8h }, [x22], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v25.8h }, [x22], #0x10\n"
+      "st1 { v28.8h }, [x21], #0x10\n"
+      "st1 { v29.8h }, [x21], #0x10\n"
       "tbz x8, #3, 280f\n"
       "st1 { v10.8h }, [x16], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
-      "st1 { v26.8h }, [x23], #0x10\n"
-      "st1 { v30.8h }, [x22], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
+      "st1 { v26.8h }, [x22], #0x10\n"
+      "st1 { v30.8h }, [x21], #0x10\n"
       "tbz x8, #2, 278f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x8, #1, 277f\n"
       "st1 { v11.s }[2], [x16], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
-      "st1 { v27.s }[2], [x23], #0x4\n"
-      "st1 { v31.s }[2], [x22], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
+      "st1 { v27.s }[2], [x22], #0x4\n"
+      "st1 { v31.s }[2], [x21], #0x4\n"
       "tbz x8, #0, 292f\n"
       "st1 { v11.h }[6], [x16]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
-      "st1 { v27.h }[6], [x23]\n"
-      "st1 { v31.h }[6], [x22]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
+      "st1 { v27.h }[6], [x22]\n"
+      "st1 { v31.h }[6], [x21]\n"
       "b 292f\n"
       "277:"  // Height 6: Partial direct writeback: partial_1_28
       "tbz x8, #0, 292f\n"
       "st1 { v11.h }[4], [x16]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
-      "st1 { v27.h }[4], [x23]\n"
-      "st1 { v31.h }[4], [x22]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
+      "st1 { v27.h }[4], [x22]\n"
+      "st1 { v31.h }[4], [x21]\n"
       "b 292f\n"
       "278:"  // Height 6: Partial direct writeback: partial_2_24
       "tbz x8, #1, 279f\n"
       "str s11, [x16], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
-      "str s27, [x23], #0x4\n"
-      "str s31, [x22], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
+      "str s27, [x22], #0x4\n"
+      "str s31, [x21], #0x4\n"
       "tbz x8, #0, 292f\n"
       "st1 { v11.h }[2], [x16]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
-      "st1 { v27.h }[2], [x23]\n"
-      "st1 { v31.h }[2], [x22]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
+      "st1 { v27.h }[2], [x22]\n"
+      "st1 { v31.h }[2], [x21]\n"
       "b 292f\n"
       "279:"  // Height 6: Partial direct writeback: partial_1_24
       "tbz x8, #0, 292f\n"
       "str h11, [x16, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
-      "str h27, [x23, #0x0]\n"
-      "str h31, [x22, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
+      "str h27, [x22, #0x0]\n"
+      "str h31, [x21, #0x0]\n"
       "b 292f\n"
       "280:"  // Height 6: Partial direct writeback: partial_4_16
       "tbz x8, #2, 282f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x8, #1, 281f\n"
       "st1 { v10.s }[2], [x16], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
-      "st1 { v26.s }[2], [x23], #0x4\n"
-      "st1 { v30.s }[2], [x22], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
+      "st1 { v26.s }[2], [x22], #0x4\n"
+      "st1 { v30.s }[2], [x21], #0x4\n"
       "tbz x8, #0, 292f\n"
       "st1 { v10.h }[6], [x16]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
-      "st1 { v26.h }[6], [x23]\n"
-      "st1 { v30.h }[6], [x22]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
+      "st1 { v26.h }[6], [x22]\n"
+      "st1 { v30.h }[6], [x21]\n"
       "b 292f\n"
       "281:"  // Height 6: Partial direct writeback: partial_1_20
       "tbz x8, #0, 292f\n"
       "st1 { v10.h }[4], [x16]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
-      "st1 { v26.h }[4], [x23]\n"
-      "st1 { v30.h }[4], [x22]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
+      "st1 { v26.h }[4], [x22]\n"
+      "st1 { v30.h }[4], [x21]\n"
       "b 292f\n"
       "282:"  // Height 6: Partial direct writeback: partial_2_16
       "tbz x8, #1, 283f\n"
       "str s10, [x16], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
-      "str s26, [x23], #0x4\n"
-      "str s30, [x22], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
+      "str s26, [x22], #0x4\n"
+      "str s30, [x21], #0x4\n"
       "tbz x8, #0, 292f\n"
       "st1 { v10.h }[2], [x16]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
-      "st1 { v26.h }[2], [x23]\n"
-      "st1 { v30.h }[2], [x22]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
+      "st1 { v26.h }[2], [x22]\n"
+      "st1 { v30.h }[2], [x21]\n"
       "b 292f\n"
       "283:"  // Height 6: Partial direct writeback: partial_1_16
       "tbz x8, #0, 292f\n"
       "str h10, [x16, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
-      "str h26, [x23, #0x0]\n"
-      "str h30, [x22, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
+      "str h26, [x22, #0x0]\n"
+      "str h30, [x21, #0x0]\n"
       "b 292f\n"
       "284:"  // Height 6: Partial direct writeback: partial_8_0
       "tbz x8, #3, 288f\n"
       "st1 { v8.8h }, [x16], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v28.8h }, [x22], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v28.8h }, [x21], #0x10\n"
       "tbz x8, #2, 286f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x8, #1, 285f\n"
       "st1 { v9.s }[2], [x16], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
-      "st1 { v25.s }[2], [x23], #0x4\n"
-      "st1 { v29.s }[2], [x22], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
+      "st1 { v25.s }[2], [x22], #0x4\n"
+      "st1 { v29.s }[2], [x21], #0x4\n"
       "tbz x8, #0, 292f\n"
       "st1 { v9.h }[6], [x16]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
-      "st1 { v25.h }[6], [x23]\n"
-      "st1 { v29.h }[6], [x22]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
+      "st1 { v25.h }[6], [x22]\n"
+      "st1 { v29.h }[6], [x21]\n"
       "b 292f\n"
       "285:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x8, #0, 292f\n"
       "st1 { v9.h }[4], [x16]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
-      "st1 { v25.h }[4], [x23]\n"
-      "st1 { v29.h }[4], [x22]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
+      "st1 { v25.h }[4], [x22]\n"
+      "st1 { v29.h }[4], [x21]\n"
       "b 292f\n"
       "286:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x8, #1, 287f\n"
       "str s9, [x16], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
-      "str s25, [x23], #0x4\n"
-      "str s29, [x22], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
+      "str s25, [x22], #0x4\n"
+      "str s29, [x21], #0x4\n"
       "tbz x8, #0, 292f\n"
       "st1 { v9.h }[2], [x16]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
-      "st1 { v25.h }[2], [x23]\n"
-      "st1 { v29.h }[2], [x22]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
+      "st1 { v25.h }[2], [x22]\n"
+      "st1 { v29.h }[2], [x21]\n"
       "b 292f\n"
       "287:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x8, #0, 292f\n"
       "str h9, [x16, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
-      "str h25, [x23, #0x0]\n"
-      "str h29, [x22, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
+      "str h25, [x22, #0x0]\n"
+      "str h29, [x21, #0x0]\n"
       "b 292f\n"
       "288:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x8, #2, 290f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x8, #1, 289f\n"
       "st1 { v8.s }[2], [x16], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
       "tbz x8, #0, 292f\n"
       "st1 { v8.h }[6], [x16]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
-      "st1 { v24.h }[6], [x23]\n"
-      "st1 { v28.h }[6], [x22]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
+      "st1 { v24.h }[6], [x22]\n"
+      "st1 { v28.h }[6], [x21]\n"
       "b 292f\n"
       "289:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x8, #0, 292f\n"
       "st1 { v8.h }[4], [x16]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
-      "st1 { v24.h }[4], [x23]\n"
-      "st1 { v28.h }[4], [x22]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
+      "st1 { v24.h }[4], [x22]\n"
+      "st1 { v28.h }[4], [x21]\n"
       "b 292f\n"
       "290:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x8, #1, 291f\n"
       "str s8, [x16], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
       "tbz x8, #0, 292f\n"
       "st1 { v8.h }[2], [x16]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
-      "st1 { v24.h }[2], [x23]\n"
-      "st1 { v28.h }[2], [x22]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
+      "st1 { v24.h }[2], [x22]\n"
+      "st1 { v28.h }[2], [x21]\n"
       "b 292f\n"
       "291:"  // Height 6: Partial direct writeback: partial_1_0
       "str h8, [x16, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
-      "str h24, [x23, #0x0]\n"
-      "str h28, [x22, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
+      "str h24, [x22, #0x0]\n"
+      "str h28, [x21, #0x0]\n"
       "292:"  // Height 6: Partial direct writeback: Done
       "b 294f\n"
       "293:"  // Height 6: Full writeback
@@ -5711,26 +5708,26 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q28, [x22, #0x0]\n"
-      "str q29, [x22, #0x10]\n"
-      "str q30, [x22, #0x20]\n"
-      "str q31, [x22, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
+      "str q28, [x21, #0x0]\n"
+      "str q29, [x21, #0x10]\n"
+      "str q30, [x21, #0x20]\n"
+      "str q31, [x21, #0x30]\n"
       "294:"  // Height 6: Writeback done
       "subs x8, x8, #0x20\n"
       "bgt 247b\n"
@@ -5746,8 +5743,8 @@ void a64_hybrid_fp16_mla_6x32_a55 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "296:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp
index 4e81b724eb..978b344f1a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp
@@ -47,19 +47,18 @@ void a64_hybrid_fp16_mla_6x32 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const __fp16 *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void a64_hybrid_fp16_mla_6x32 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -103,10 +101,10 @@ void a64_hybrid_fp16_mla_6x32 (
       "cmp %x[M], #0x2\n"
       "bgt 99f\n"
       "beq 50f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x12, 3f\n"
       "ldr q8, [x12, #0x0]\n"
@@ -245,8 +243,8 @@ void a64_hybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "23:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 24f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -270,10 +268,6 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr q17, [x10, #0x20]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "sub x27, x27, #0x8\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "ldr q17, [x10, #0x40]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
@@ -330,21 +324,22 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr q17, [x10, #0x1e0]\n"
       "fmla v9.8h, v16.8h, v0.h[7]\n"
       "ldr q16, [x10, #0x1f0]\n"
-      "add x10, x10, #0x200\n"
+      "sub x27, x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "fmla v10.8h, v17.8h, v0.h[7]\n"
-      "ldr q6, [x10, #0x0]\n"
       "fmla v11.8h, v16.8h, v0.h[7]\n"
       "ldr q0, [x26, #0x0]\n"
+      "cmp x27, #0x10\n"
+      "add x10, x10, #0x200\n"
+      "ldr q6, [x10, #0x0]\n"
       "ldr q7, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 26b\n"
       "27:"  // Height 1: Multiply loop: Single iteration only
       "fmla v8.8h, v6.8h, v0.h[0]\n"
       "ldr q17, [x10, #0x20]\n"
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "ldr q17, [x10, #0x40]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
@@ -401,23 +396,26 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr q17, [x10, #0x1e0]\n"
       "fmla v9.8h, v16.8h, v0.h[7]\n"
       "ldr q16, [x10, #0x1f0]\n"
-      "add x10, x10, #0x200\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "fmla v10.8h, v17.8h, v0.h[7]\n"
       "fmla v11.8h, v16.8h, v0.h[7]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x200\n"
       "28:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 30f\n"
       "29:"  // Height 1: Multiply loop: Odd block loop
       "ldr h0, [x26], #0x2\n"
-      "ldr q17, [x10, #0x0]\n"
+      "ldr q16, [x10, #0x0]\n"
+      "fmla v8.8h, v16.8h, v0.h[0]\n"
       "sub x27, x27, #0x1\n"
-      "ldr q16, [x10, #0x10]\n"
-      "fmla v8.8h, v17.8h, v0.h[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      "fmla v9.8h, v16.8h, v0.h[0]\n"
+      "ldr q17, [x10, #0x10]\n"
+      "ldr q16, [x10, #0x20]\n"
+      "fmla v9.8h, v17.8h, v0.h[0]\n"
+      "fmla v10.8h, v16.8h, v0.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      "fmla v10.8h, v17.8h, v0.h[0]\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
+      "add x10, x10, #0x40\n"
       "cbnz x27, 29b\n"
       "30:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -426,9 +424,9 @@ void a64_hybrid_fp16_mla_6x32 (
       "bne 23b\n"
       "prfm pstl1keep, [x9, #0x0]\n"
       "tbz %x[flags], #1, 31f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.8h }, [x21]\n"
       "ld1r { v16.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v17.8h\n"
       "fmin v9.8h, v9.8h, v17.8h\n"
@@ -546,167 +544,167 @@ void a64_hybrid_fp16_mla_6x32 (
       "bgt 2b\n"
       "b 296f\n"
       "50:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "51:"  // Height 2: Column loop
       "cbz x12, 52f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "add x12, x12, #0x40\n"
       "b 71f\n"
       "52:"  // Height 2: no bias
       "tbz %x[flags], #0, 70f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x11, #0x20\n"
-      "add x26, x9, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "bge 69f\n"
       "tbz x11, #4, 60f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
       "ld1 { v9.8h }, [x9], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
       "tbz x11, #3, 56f\n"
       "ld1 { v10.8h }, [x9], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
       "tbz x11, #2, 54f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "tbz x11, #1, 53f\n"
       "ld1 { v11.s }[2], [x9], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v11.h }[6], [x9]\n"
-      "ld1 { v15.h }[6], [x26]\n"
+      "ld1 { v15.h }[6], [x25]\n"
       "b 68f\n"
       "53:"  // Height 2: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v11.h }[4], [x9]\n"
-      "ld1 { v15.h }[4], [x26]\n"
+      "ld1 { v15.h }[4], [x25]\n"
       "b 68f\n"
       "54:"  // Height 2: Partial accumulate: partial_2_24
       "tbz x11, #1, 55f\n"
       "ldr s11, [x9], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v11.h }[2], [x9]\n"
-      "ld1 { v15.h }[2], [x26]\n"
+      "ld1 { v15.h }[2], [x25]\n"
       "b 68f\n"
       "55:"  // Height 2: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x11, #0, 68f\n"
       "ldr h11, [x9, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
       "b 68f\n"
       "56:"  // Height 2: Partial accumulate: partial_4_16
       "tbz x11, #2, 58f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "tbz x11, #1, 57f\n"
       "ld1 { v10.s }[2], [x9], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v10.h }[6], [x9]\n"
-      "ld1 { v14.h }[6], [x26]\n"
+      "ld1 { v14.h }[6], [x25]\n"
       "b 68f\n"
       "57:"  // Height 2: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v10.h }[4], [x9]\n"
-      "ld1 { v14.h }[4], [x26]\n"
+      "ld1 { v14.h }[4], [x25]\n"
       "b 68f\n"
       "58:"  // Height 2: Partial accumulate: partial_2_16
       "tbz x11, #1, 59f\n"
       "ldr s10, [x9], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v10.h }[2], [x9]\n"
-      "ld1 { v14.h }[2], [x26]\n"
+      "ld1 { v14.h }[2], [x25]\n"
       "b 68f\n"
       "59:"  // Height 2: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x11, #0, 68f\n"
       "ldr h10, [x9, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
       "b 68f\n"
       "60:"  // Height 2: Partial accumulate: partial_8_0
       "tbz x11, #3, 64f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
       "tbz x11, #2, 62f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "tbz x11, #1, 61f\n"
       "ld1 { v9.s }[2], [x9], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v9.h }[6], [x9]\n"
-      "ld1 { v13.h }[6], [x26]\n"
+      "ld1 { v13.h }[6], [x25]\n"
       "b 68f\n"
       "61:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v9.h }[4], [x9]\n"
-      "ld1 { v13.h }[4], [x26]\n"
+      "ld1 { v13.h }[4], [x25]\n"
       "b 68f\n"
       "62:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x11, #1, 63f\n"
       "ldr s9, [x9], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v9.h }[2], [x9]\n"
-      "ld1 { v13.h }[2], [x26]\n"
+      "ld1 { v13.h }[2], [x25]\n"
       "b 68f\n"
       "63:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x11, #0, 68f\n"
       "ldr h9, [x9, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
       "b 68f\n"
       "64:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x11, #2, 66f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "tbz x11, #1, 65f\n"
       "ld1 { v8.s }[2], [x9], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v8.h }[6], [x9]\n"
-      "ld1 { v12.h }[6], [x26]\n"
+      "ld1 { v12.h }[6], [x25]\n"
       "b 68f\n"
       "65:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v8.h }[4], [x9]\n"
-      "ld1 { v12.h }[4], [x26]\n"
+      "ld1 { v12.h }[4], [x25]\n"
       "b 68f\n"
       "66:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x11, #1, 67f\n"
       "ldr s8, [x9], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
       "tbz x11, #0, 68f\n"
       "ld1 { v8.h }[2], [x9]\n"
-      "ld1 { v12.h }[2], [x26]\n"
+      "ld1 { v12.h }[2], [x25]\n"
       "b 68f\n"
       "67:"  // Height 2: Partial accumulate: partial_1_0
       "ldr h8, [x9, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
       "68:"  // Height 2: Partial accumulate: Done
       "sub x9, x9, x20\n"
@@ -716,10 +714,10 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "b 71f\n"
       "70:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -734,8 +732,8 @@ void a64_hybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "72:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 73f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -767,22 +765,22 @@ void a64_hybrid_fp16_mla_6x32 (
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "fmla v14.8h, v17.8h, v1.h[0]\n"
       "ldr q17, [x10, #0x40]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x25, x25, #0x10\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
       "fmla v15.8h, v16.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "cmp x27, #0x10\n"
       "fmla v8.8h, v17.8h, v0.h[1]\n"
       "fmla v12.8h, v17.8h, v1.h[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v9.8h, v16.8h, v0.h[1]\n"
       "fmla v13.8h, v16.8h, v1.h[1]\n"
       "ldr q16, [x10, #0x70]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.8h, v17.8h, v0.h[1]\n"
       "fmla v14.8h, v17.8h, v1.h[1]\n"
       "ldr q17, [x10, #0x80]\n"
@@ -874,18 +872,18 @@ void a64_hybrid_fp16_mla_6x32 (
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.8h, v17.8h, v0.h[0]\n"
       "fmla v14.8h, v17.8h, v1.h[0]\n"
       "ldr q17, [x10, #0x40]\n"
+      "sub x27, x27, #0x8\n"
       "fmla v11.8h, v16.8h, v0.h[0]\n"
       "fmla v15.8h, v16.8h, v1.h[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v8.8h, v17.8h, v0.h[1]\n"
       "fmla v12.8h, v17.8h, v1.h[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v9.8h, v16.8h, v0.h[1]\n"
       "fmla v13.8h, v16.8h, v1.h[1]\n"
       "ldr q16, [x10, #0x70]\n"
@@ -980,9 +978,9 @@ void a64_hybrid_fp16_mla_6x32 (
       "fmla v9.8h, v16.8h, v1.h[0]\n"
       "fmla v13.8h, v16.8h, v0.h[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       "fmla v10.8h, v17.8h, v1.h[0]\n"
       "fmla v14.8h, v17.8h, v0.h[0]\n"
+      "add x10, x10, #0x40\n"
       "fmla v11.8h, v16.8h, v1.h[0]\n"
       "fmla v15.8h, v16.8h, v0.h[0]\n"
       "cbnz x27, 78b\n"
@@ -992,13 +990,13 @@ void a64_hybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 72b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #1\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "tbz %x[flags], #1, 80f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.8h }, [x21]\n"
       "ld1r { v16.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v17.8h\n"
       "fmin v9.8h, v9.8h, v17.8h\n"
@@ -1022,127 +1020,127 @@ void a64_hybrid_fp16_mla_6x32 (
       "tbz x11, #4, 88f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
       "st1 { v9.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
       "tbz x11, #3, 84f\n"
       "st1 { v10.8h }, [x9], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
       "tbz x11, #2, 82f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
       "tbz x11, #1, 81f\n"
       "st1 { v11.s }[2], [x9], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
       "tbz x11, #0, 96f\n"
       "st1 { v11.h }[6], [x9]\n"
-      "st1 { v15.h }[6], [x26]\n"
+      "st1 { v15.h }[6], [x25]\n"
       "b 96f\n"
       "81:"  // Height 2: Partial direct writeback: partial_1_28
       "tbz x11, #0, 96f\n"
       "st1 { v11.h }[4], [x9]\n"
-      "st1 { v15.h }[4], [x26]\n"
+      "st1 { v15.h }[4], [x25]\n"
       "b 96f\n"
       "82:"  // Height 2: Partial direct writeback: partial_2_24
       "tbz x11, #1, 83f\n"
       "str s11, [x9], #0x4\n"
-      "str s15, [x26], #0x4\n"
+      "str s15, [x25], #0x4\n"
       "tbz x11, #0, 96f\n"
       "st1 { v11.h }[2], [x9]\n"
-      "st1 { v15.h }[2], [x26]\n"
+      "st1 { v15.h }[2], [x25]\n"
       "b 96f\n"
       "83:"  // Height 2: Partial direct writeback: partial_1_24
       "tbz x11, #0, 96f\n"
       "str h11, [x9, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
       "b 96f\n"
       "84:"  // Height 2: Partial direct writeback: partial_4_16
       "tbz x11, #2, 86f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
+      "str d14, [x25], #0x8\n"
       "tbz x11, #1, 85f\n"
       "st1 { v10.s }[2], [x9], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
       "tbz x11, #0, 96f\n"
       "st1 { v10.h }[6], [x9]\n"
-      "st1 { v14.h }[6], [x26]\n"
+      "st1 { v14.h }[6], [x25]\n"
       "b 96f\n"
       "85:"  // Height 2: Partial direct writeback: partial_1_20
       "tbz x11, #0, 96f\n"
       "st1 { v10.h }[4], [x9]\n"
-      "st1 { v14.h }[4], [x26]\n"
+      "st1 { v14.h }[4], [x25]\n"
       "b 96f\n"
       "86:"  // Height 2: Partial direct writeback: partial_2_16
       "tbz x11, #1, 87f\n"
       "str s10, [x9], #0x4\n"
-      "str s14, [x26], #0x4\n"
+      "str s14, [x25], #0x4\n"
       "tbz x11, #0, 96f\n"
       "st1 { v10.h }[2], [x9]\n"
-      "st1 { v14.h }[2], [x26]\n"
+      "st1 { v14.h }[2], [x25]\n"
       "b 96f\n"
       "87:"  // Height 2: Partial direct writeback: partial_1_16
       "tbz x11, #0, 96f\n"
       "str h10, [x9, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
       "b 96f\n"
       "88:"  // Height 2: Partial direct writeback: partial_8_0
       "tbz x11, #3, 92f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
       "tbz x11, #2, 90f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
+      "str d13, [x25], #0x8\n"
       "tbz x11, #1, 89f\n"
       "st1 { v9.s }[2], [x9], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
       "tbz x11, #0, 96f\n"
       "st1 { v9.h }[6], [x9]\n"
-      "st1 { v13.h }[6], [x26]\n"
+      "st1 { v13.h }[6], [x25]\n"
       "b 96f\n"
       "89:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x11, #0, 96f\n"
       "st1 { v9.h }[4], [x9]\n"
-      "st1 { v13.h }[4], [x26]\n"
+      "st1 { v13.h }[4], [x25]\n"
       "b 96f\n"
       "90:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x11, #1, 91f\n"
       "str s9, [x9], #0x4\n"
-      "str s13, [x26], #0x4\n"
+      "str s13, [x25], #0x4\n"
       "tbz x11, #0, 96f\n"
       "st1 { v9.h }[2], [x9]\n"
-      "st1 { v13.h }[2], [x26]\n"
+      "st1 { v13.h }[2], [x25]\n"
       "b 96f\n"
       "91:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x11, #0, 96f\n"
       "str h9, [x9, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
       "b 96f\n"
       "92:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x11, #2, 94f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
+      "str d12, [x25], #0x8\n"
       "tbz x11, #1, 93f\n"
       "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
       "tbz x11, #0, 96f\n"
       "st1 { v8.h }[6], [x9]\n"
-      "st1 { v12.h }[6], [x26]\n"
+      "st1 { v12.h }[6], [x25]\n"
       "b 96f\n"
       "93:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x11, #0, 96f\n"
       "st1 { v8.h }[4], [x9]\n"
-      "st1 { v12.h }[4], [x26]\n"
+      "st1 { v12.h }[4], [x25]\n"
       "b 96f\n"
       "94:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x11, #1, 95f\n"
       "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
+      "str s12, [x25], #0x4\n"
       "tbz x11, #0, 96f\n"
       "st1 { v8.h }[2], [x9]\n"
-      "st1 { v12.h }[2], [x26]\n"
+      "st1 { v12.h }[2], [x25]\n"
       "b 96f\n"
       "95:"  // Height 2: Partial direct writeback: partial_1_0
       "str h8, [x9, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
       "96:"  // Height 2: Partial direct writeback: Done
       "b 98f\n"
       "97:"  // Height 2: Full writeback
@@ -1151,214 +1149,214 @@ void a64_hybrid_fp16_mla_6x32 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
       "98:"  // Height 2: Writeback done
       "subs x11, x11, #0x20\n"
       "bgt 51b\n"
       "b 296f\n"
       "99:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "100:"  // Height 3: Column loop
       "cbz x12, 101f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
       "b 120f\n"
       "101:"  // Height 3: no bias
       "tbz %x[flags], #0, 119f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #1\n"
       "cmp x11, #0x20\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "bge 118f\n"
       "tbz x11, #4, 109f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
       "ld1 { v9.8h }, [x9], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
       "tbz x11, #3, 105f\n"
       "ld1 { v10.8h }, [x9], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
       "tbz x11, #2, 103f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x11, #1, 102f\n"
       "ld1 { v11.s }[2], [x9], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v11.h }[6], [x9]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
       "b 117f\n"
       "102:"  // Height 3: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v11.h }[4], [x9]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
       "b 117f\n"
       "103:"  // Height 3: Partial accumulate: partial_2_24
       "tbz x11, #1, 104f\n"
       "ldr s11, [x9], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s19, [x25], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v11.h }[2], [x9]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
       "b 117f\n"
       "104:"  // Height 3: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x11, #0, 117f\n"
       "ldr h11, [x9, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
       "b 117f\n"
       "105:"  // Height 3: Partial accumulate: partial_4_16
       "tbz x11, #2, 107f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x11, #1, 106f\n"
       "ld1 { v10.s }[2], [x9], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v10.h }[6], [x9]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
       "b 117f\n"
       "106:"  // Height 3: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v10.h }[4], [x9]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
       "b 117f\n"
       "107:"  // Height 3: Partial accumulate: partial_2_16
       "tbz x11, #1, 108f\n"
       "ldr s10, [x9], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s18, [x25], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v10.h }[2], [x9]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
       "b 117f\n"
       "108:"  // Height 3: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x11, #0, 117f\n"
       "ldr h10, [x9, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
       "b 117f\n"
       "109:"  // Height 3: Partial accumulate: partial_8_0
       "tbz x11, #3, 113f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
       "tbz x11, #2, 111f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x11, #1, 110f\n"
       "ld1 { v9.s }[2], [x9], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v9.h }[6], [x9]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
       "b 117f\n"
       "110:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v9.h }[4], [x9]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
       "b 117f\n"
       "111:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x11, #1, 112f\n"
       "ldr s9, [x9], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s17, [x25], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v9.h }[2], [x9]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
       "b 117f\n"
       "112:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x11, #0, 117f\n"
       "ldr h9, [x9, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
       "b 117f\n"
       "113:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x11, #2, 115f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
       "tbz x11, #1, 114f\n"
       "ld1 { v8.s }[2], [x9], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v8.h }[6], [x9]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
       "b 117f\n"
       "114:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v8.h }[4], [x9]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
       "b 117f\n"
       "115:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x11, #1, 116f\n"
       "ldr s8, [x9], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s16, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
       "tbz x11, #0, 117f\n"
       "ld1 { v8.h }[2], [x9]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
       "b 117f\n"
       "116:"  // Height 3: Partial accumulate: partial_1_0
       "ldr h8, [x9, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h16, [x25, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
       "117:"  // Height 3: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 120f\n"
@@ -1367,14 +1365,14 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
       "b 120f\n"
       "119:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1393,8 +1391,8 @@ void a64_hybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "121:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 122f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1434,18 +1432,18 @@ void a64_hybrid_fp16_mla_6x32 (
       "fmla v17.8h, v7.8h, v2.h[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "add x24, x24, #0x10\n"
-      "cmp x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.8h, v21.8h, v0.h[0]\n"
       "fmla v14.8h, v21.8h, v1.h[0]\n"
+      "cmp x27, #0x10\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v18.8h, v21.8h, v2.h[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v11.8h, v20.8h, v0.h[0]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v15.8h, v20.8h, v1.h[0]\n"
       "fmla v19.8h, v20.8h, v2.h[0]\n"
       "ldr q20, [x10, #0x50]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v8.8h, v21.8h, v0.h[1]\n"
       "fmla v12.8h, v21.8h, v1.h[1]\n"
       "fmla v16.8h, v21.8h, v2.h[1]\n"
@@ -1576,14 +1574,14 @@ void a64_hybrid_fp16_mla_6x32 (
       "fmla v17.8h, v7.8h, v2.h[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "sub x27, x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.8h, v21.8h, v0.h[0]\n"
       "fmla v14.8h, v21.8h, v1.h[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v18.8h, v21.8h, v2.h[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v11.8h, v20.8h, v0.h[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v15.8h, v20.8h, v1.h[0]\n"
       "fmla v19.8h, v20.8h, v2.h[0]\n"
       "ldr q20, [x10, #0x50]\n"
@@ -1706,9 +1704,9 @@ void a64_hybrid_fp16_mla_6x32 (
       "sub x27, x27, #0x1\n"
       "ldr h0, [x24], #0x2\n"
       "ldr q21, [x10, #0x0]\n"
-      "ldr q20, [x10, #0x10]\n"
       "fmla v8.8h, v21.8h, v2.h[0]\n"
       "fmla v12.8h, v21.8h, v1.h[0]\n"
+      "ldr q20, [x10, #0x10]\n"
       "fmla v16.8h, v21.8h, v0.h[0]\n"
       "ldr q21, [x10, #0x20]\n"
       "fmla v9.8h, v20.8h, v2.h[0]\n"
@@ -1729,15 +1727,15 @@ void a64_hybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 121b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #1\n"
       "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "tbz %x[flags], #1, 129f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v21.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.8h }, [x21]\n"
       "ld1r { v20.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v21.8h\n"
       "fmin v9.8h, v9.8h, v21.8h\n"
@@ -1769,159 +1767,159 @@ void a64_hybrid_fp16_mla_6x32 (
       "tbz x11, #4, 137f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
       "st1 { v9.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
       "tbz x11, #3, 133f\n"
       "st1 { v10.8h }, [x9], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
       "tbz x11, #2, 131f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x11, #1, 130f\n"
       "st1 { v11.s }[2], [x9], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
       "tbz x11, #0, 145f\n"
       "st1 { v11.h }[6], [x9]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
       "b 145f\n"
       "130:"  // Height 3: Partial direct writeback: partial_1_28
       "tbz x11, #0, 145f\n"
       "st1 { v11.h }[4], [x9]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
       "b 145f\n"
       "131:"  // Height 3: Partial direct writeback: partial_2_24
       "tbz x11, #1, 132f\n"
       "str s11, [x9], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
       "tbz x11, #0, 145f\n"
       "st1 { v11.h }[2], [x9]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
       "b 145f\n"
       "132:"  // Height 3: Partial direct writeback: partial_1_24
       "tbz x11, #0, 145f\n"
       "str h11, [x9, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
       "b 145f\n"
       "133:"  // Height 3: Partial direct writeback: partial_4_16
       "tbz x11, #2, 135f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x11, #1, 134f\n"
       "st1 { v10.s }[2], [x9], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
       "tbz x11, #0, 145f\n"
       "st1 { v10.h }[6], [x9]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
       "b 145f\n"
       "134:"  // Height 3: Partial direct writeback: partial_1_20
       "tbz x11, #0, 145f\n"
       "st1 { v10.h }[4], [x9]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
       "b 145f\n"
       "135:"  // Height 3: Partial direct writeback: partial_2_16
       "tbz x11, #1, 136f\n"
       "str s10, [x9], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
       "tbz x11, #0, 145f\n"
       "st1 { v10.h }[2], [x9]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
       "b 145f\n"
       "136:"  // Height 3: Partial direct writeback: partial_1_16
       "tbz x11, #0, 145f\n"
       "str h10, [x9, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
       "b 145f\n"
       "137:"  // Height 3: Partial direct writeback: partial_8_0
       "tbz x11, #3, 141f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
       "tbz x11, #2, 139f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x11, #1, 138f\n"
       "st1 { v9.s }[2], [x9], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
       "tbz x11, #0, 145f\n"
       "st1 { v9.h }[6], [x9]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
       "b 145f\n"
       "138:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x11, #0, 145f\n"
       "st1 { v9.h }[4], [x9]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
       "b 145f\n"
       "139:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x11, #1, 140f\n"
       "str s9, [x9], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
       "tbz x11, #0, 145f\n"
       "st1 { v9.h }[2], [x9]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
       "b 145f\n"
       "140:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x11, #0, 145f\n"
       "str h9, [x9, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
       "b 145f\n"
       "141:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x11, #2, 143f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x11, #1, 142f\n"
       "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
       "tbz x11, #0, 145f\n"
       "st1 { v8.h }[6], [x9]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
       "b 145f\n"
       "142:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x11, #0, 145f\n"
       "st1 { v8.h }[4], [x9]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
       "b 145f\n"
       "143:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x11, #1, 144f\n"
       "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
       "tbz x11, #0, 145f\n"
       "st1 { v8.h }[2], [x9]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
       "b 145f\n"
       "144:"  // Height 3: Partial direct writeback: partial_1_0
       "str h8, [x9, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
       "145:"  // Height 3: Partial direct writeback: Done
       "b 147f\n"
       "146:"  // Height 3: Full writeback
@@ -1930,39 +1928,39 @@ void a64_hybrid_fp16_mla_6x32 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "147:"  // Height 3: Writeback done
       "subs x11, x11, #0x20\n"
       "bgt 100b\n"
       "b 296f\n"
       "148:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "149:"  // Height 4: Column loop
       "cbz x12, 150f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1970,215 +1968,215 @@ void a64_hybrid_fp16_mla_6x32 (
       "150:"  // Height 4: no bias
       "tbz %x[flags], #0, 168f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x20\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "cmp x11, #0x20\n"
+      "add x23, x24, x20, LSL #1\n"
       "bge 167f\n"
       "tbz x11, #4, 158f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
       "ld1 { v9.8h }, [x9], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
       "tbz x11, #3, 154f\n"
       "ld1 { v10.8h }, [x9], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
       "tbz x11, #2, 152f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x11, #1, 151f\n"
       "ld1 { v11.s }[2], [x9], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v11.h }[6], [x9]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
       "b 166f\n"
       "151:"  // Height 4: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v11.h }[4], [x9]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
       "b 166f\n"
       "152:"  // Height 4: Partial accumulate: partial_2_24
       "tbz x11, #1, 153f\n"
       "ldr s11, [x9], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v11.h }[2], [x9]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
       "b 166f\n"
       "153:"  // Height 4: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x11, #0, 166f\n"
       "ldr h11, [x9, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
       "b 166f\n"
       "154:"  // Height 4: Partial accumulate: partial_4_16
       "tbz x11, #2, 156f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x11, #1, 155f\n"
       "ld1 { v10.s }[2], [x9], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v10.h }[6], [x9]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
       "b 166f\n"
       "155:"  // Height 4: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v10.h }[4], [x9]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
       "b 166f\n"
       "156:"  // Height 4: Partial accumulate: partial_2_16
       "tbz x11, #1, 157f\n"
       "ldr s10, [x9], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v10.h }[2], [x9]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
       "b 166f\n"
       "157:"  // Height 4: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x11, #0, 166f\n"
       "ldr h10, [x9, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
       "b 166f\n"
       "158:"  // Height 4: Partial accumulate: partial_8_0
       "tbz x11, #3, 162f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
       "tbz x11, #2, 160f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x11, #1, 159f\n"
       "ld1 { v9.s }[2], [x9], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v9.h }[6], [x9]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
       "b 166f\n"
       "159:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v9.h }[4], [x9]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
       "b 166f\n"
       "160:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x11, #1, 161f\n"
       "ldr s9, [x9], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v9.h }[2], [x9]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
       "b 166f\n"
       "161:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x11, #0, 166f\n"
       "ldr h9, [x9, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
       "b 166f\n"
       "162:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x11, #2, 164f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x11, #1, 163f\n"
       "ld1 { v8.s }[2], [x9], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v8.h }[6], [x9]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
       "b 166f\n"
       "163:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v8.h }[4], [x9]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
       "b 166f\n"
       "164:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x11, #1, 165f\n"
       "ldr s8, [x9], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
       "tbz x11, #0, 166f\n"
       "ld1 { v8.h }[2], [x9]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
       "b 166f\n"
       "165:"  // Height 4: Partial accumulate: partial_1_0
       "ldr h8, [x9, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
       "166:"  // Height 4: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 169f\n"
@@ -2187,18 +2185,18 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "b 169f\n"
       "168:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2221,8 +2219,8 @@ void a64_hybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "170:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 171f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2442,14 +2440,14 @@ void a64_hybrid_fp16_mla_6x32 (
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x8\n"
       "fmla v17.8h, v7.8h, v2.h[0]\n"
       "fmla v21.8h, v7.8h, v3.h[0]\n"
       "ldr q24, [x10, #0x30]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v10.8h, v25.8h, v0.h[0]\n"
       "fmla v14.8h, v25.8h, v1.h[0]\n"
-      "sub x27, x27, #0x8\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v18.8h, v25.8h, v2.h[0]\n"
       "fmla v22.8h, v25.8h, v3.h[0]\n"
@@ -2635,17 +2633,17 @@ void a64_hybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 170b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "prfm pstl1keep, [x25, #0x0]\n"
-      "add x24, x25, x20, LSL #1\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 178f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v25.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v25.8h }, [x21]\n"
       "ld1r { v24.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v25.8h\n"
       "fmin v9.8h, v9.8h, v25.8h\n"
@@ -2685,191 +2683,191 @@ void a64_hybrid_fp16_mla_6x32 (
       "tbz x11, #4, 186f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
       "st1 { v9.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
       "tbz x11, #3, 182f\n"
       "st1 { v10.8h }, [x9], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
       "tbz x11, #2, 180f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
       "tbz x11, #1, 179f\n"
       "st1 { v11.s }[2], [x9], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
       "tbz x11, #0, 194f\n"
       "st1 { v11.h }[6], [x9]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
       "b 194f\n"
       "179:"  // Height 4: Partial direct writeback: partial_1_28
       "tbz x11, #0, 194f\n"
       "st1 { v11.h }[4], [x9]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
       "b 194f\n"
       "180:"  // Height 4: Partial direct writeback: partial_2_24
       "tbz x11, #1, 181f\n"
       "str s11, [x9], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
       "tbz x11, #0, 194f\n"
       "st1 { v11.h }[2], [x9]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
       "b 194f\n"
       "181:"  // Height 4: Partial direct writeback: partial_1_24
       "tbz x11, #0, 194f\n"
       "str h11, [x9, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
       "b 194f\n"
       "182:"  // Height 4: Partial direct writeback: partial_4_16
       "tbz x11, #2, 184f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
       "tbz x11, #1, 183f\n"
       "st1 { v10.s }[2], [x9], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
       "tbz x11, #0, 194f\n"
       "st1 { v10.h }[6], [x9]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
       "b 194f\n"
       "183:"  // Height 4: Partial direct writeback: partial_1_20
       "tbz x11, #0, 194f\n"
       "st1 { v10.h }[4], [x9]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
       "b 194f\n"
       "184:"  // Height 4: Partial direct writeback: partial_2_16
       "tbz x11, #1, 185f\n"
       "str s10, [x9], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
       "tbz x11, #0, 194f\n"
       "st1 { v10.h }[2], [x9]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
       "b 194f\n"
       "185:"  // Height 4: Partial direct writeback: partial_1_16
       "tbz x11, #0, 194f\n"
       "str h10, [x9, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
       "b 194f\n"
       "186:"  // Height 4: Partial direct writeback: partial_8_0
       "tbz x11, #3, 190f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
       "tbz x11, #2, 188f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
       "tbz x11, #1, 187f\n"
       "st1 { v9.s }[2], [x9], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
       "tbz x11, #0, 194f\n"
       "st1 { v9.h }[6], [x9]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
       "b 194f\n"
       "187:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x11, #0, 194f\n"
       "st1 { v9.h }[4], [x9]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
       "b 194f\n"
       "188:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x11, #1, 189f\n"
       "str s9, [x9], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
       "tbz x11, #0, 194f\n"
       "st1 { v9.h }[2], [x9]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
       "b 194f\n"
       "189:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x11, #0, 194f\n"
       "str h9, [x9, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
       "b 194f\n"
       "190:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x11, #2, 192f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x11, #1, 191f\n"
       "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
       "tbz x11, #0, 194f\n"
       "st1 { v8.h }[6], [x9]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
       "b 194f\n"
       "191:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x11, #0, 194f\n"
       "st1 { v8.h }[4], [x9]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
       "b 194f\n"
       "192:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x11, #1, 193f\n"
       "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
       "tbz x11, #0, 194f\n"
       "st1 { v8.h }[2], [x9]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
       "b 194f\n"
       "193:"  // Height 4: Partial direct writeback: partial_1_0
       "str h8, [x9, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
       "194:"  // Height 4: Partial direct writeback: Done
       "b 196f\n"
       "195:"  // Height 4: Full writeback
@@ -2878,43 +2876,43 @@ void a64_hybrid_fp16_mla_6x32 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
       "196:"  // Height 4: Writeback done
       "subs x11, x11, #0x20\n"
       "bgt 149b\n"
       "b 296f\n"
       "197:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "198:"  // Height 5: Column loop
       "cbz x12, 199f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -2926,248 +2924,248 @@ void a64_hybrid_fp16_mla_6x32 (
       "199:"  // Height 5: no bias
       "tbz %x[flags], #0, 217f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x20\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
+      "cmp x11, #0x20\n"
+      "add x22, x23, x20, LSL #1\n"
       "bge 216f\n"
       "tbz x11, #4, 207f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
       "ld1 { v9.8h }, [x9], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
-      "ld1 { v25.8h }, [x23], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
+      "ld1 { v25.8h }, [x22], #0x10\n"
       "tbz x11, #3, 203f\n"
       "ld1 { v10.8h }, [x9], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
-      "ld1 { v26.8h }, [x23], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
+      "ld1 { v26.8h }, [x22], #0x10\n"
       "tbz x11, #2, 201f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x11, #1, 200f\n"
       "ld1 { v11.s }[2], [x9], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
-      "ld1 { v27.s }[2], [x23], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
+      "ld1 { v27.s }[2], [x22], #0x4\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v11.h }[6], [x9]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
-      "ld1 { v27.h }[6], [x23]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
+      "ld1 { v27.h }[6], [x22]\n"
       "b 215f\n"
       "200:"  // Height 5: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v11.h }[4], [x9]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
-      "ld1 { v27.h }[4], [x23]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
+      "ld1 { v27.h }[4], [x22]\n"
       "b 215f\n"
       "201:"  // Height 5: Partial accumulate: partial_2_24
       "tbz x11, #1, 202f\n"
       "ldr s11, [x9], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
-      "ldr s27, [x23], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s27, [x22], #0x4\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v11.h }[2], [x9]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
-      "ld1 { v27.h }[2], [x23]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
+      "ld1 { v27.h }[2], [x22]\n"
       "b 215f\n"
       "202:"  // Height 5: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x11, #0, 215f\n"
       "ldr h11, [x9, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
-      "ldr h27, [x23, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
+      "ldr h27, [x22, #0x0]\n"
       "b 215f\n"
       "203:"  // Height 5: Partial accumulate: partial_4_16
       "tbz x11, #2, 205f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x11, #1, 204f\n"
       "ld1 { v10.s }[2], [x9], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
-      "ld1 { v26.s }[2], [x23], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
+      "ld1 { v26.s }[2], [x22], #0x4\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v10.h }[6], [x9]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
-      "ld1 { v26.h }[6], [x23]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
+      "ld1 { v26.h }[6], [x22]\n"
       "b 215f\n"
       "204:"  // Height 5: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v10.h }[4], [x9]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
-      "ld1 { v26.h }[4], [x23]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
+      "ld1 { v26.h }[4], [x22]\n"
       "b 215f\n"
       "205:"  // Height 5: Partial accumulate: partial_2_16
       "tbz x11, #1, 206f\n"
       "ldr s10, [x9], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr s26, [x23], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
+      "ldr s26, [x22], #0x4\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v10.h }[2], [x9]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
-      "ld1 { v26.h }[2], [x23]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
+      "ld1 { v26.h }[2], [x22]\n"
       "b 215f\n"
       "206:"  // Height 5: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x11, #0, 215f\n"
       "ldr h10, [x9, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
-      "ldr h26, [x23, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
+      "ldr h26, [x22, #0x0]\n"
       "b 215f\n"
       "207:"  // Height 5: Partial accumulate: partial_8_0
       "tbz x11, #3, 211f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
       "tbz x11, #2, 209f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x11, #1, 208f\n"
       "ld1 { v9.s }[2], [x9], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
-      "ld1 { v25.s }[2], [x23], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
+      "ld1 { v25.s }[2], [x22], #0x4\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v9.h }[6], [x9]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
-      "ld1 { v25.h }[6], [x23]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
+      "ld1 { v25.h }[6], [x22]\n"
       "b 215f\n"
       "208:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v9.h }[4], [x9]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
-      "ld1 { v25.h }[4], [x23]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
+      "ld1 { v25.h }[4], [x22]\n"
       "b 215f\n"
       "209:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x11, #1, 210f\n"
       "ldr s9, [x9], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
-      "ldr s25, [x23], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
+      "ldr s25, [x22], #0x4\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v9.h }[2], [x9]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
-      "ld1 { v25.h }[2], [x23]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
+      "ld1 { v25.h }[2], [x22]\n"
       "b 215f\n"
       "210:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x11, #0, 215f\n"
       "ldr h9, [x9, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
-      "ldr h25, [x23, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
+      "ldr h25, [x22, #0x0]\n"
       "b 215f\n"
       "211:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x11, #2, 213f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x11, #1, 212f\n"
       "ld1 { v8.s }[2], [x9], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
-      "ld1 { v24.s }[2], [x23], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
+      "ld1 { v24.s }[2], [x22], #0x4\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v8.h }[6], [x9]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
-      "ld1 { v24.h }[6], [x23]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
+      "ld1 { v24.h }[6], [x22]\n"
       "b 215f\n"
       "212:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v8.h }[4], [x9]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
-      "ld1 { v24.h }[4], [x23]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
+      "ld1 { v24.h }[4], [x22]\n"
       "b 215f\n"
       "213:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x11, #1, 214f\n"
       "ldr s8, [x9], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
+      "ldr s24, [x22], #0x4\n"
       "tbz x11, #0, 215f\n"
       "ld1 { v8.h }[2], [x9]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
-      "ld1 { v24.h }[2], [x23]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
+      "ld1 { v24.h }[2], [x22]\n"
       "b 215f\n"
       "214:"  // Height 5: Partial accumulate: partial_1_0
       "ldr h8, [x9, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
-      "ldr h24, [x23, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
+      "ldr h24, [x22, #0x0]\n"
       "215:"  // Height 5: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 218f\n"
@@ -3176,22 +3174,22 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
       "b 218f\n"
       "217:"  // Height 5: no accumulate
       "movi v8.16b, #0x0\n"
@@ -3218,8 +3216,8 @@ void a64_hybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "219:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 220f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -3481,12 +3479,12 @@ void a64_hybrid_fp16_mla_6x32 (
       "add x22, x22, #0x10\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "fmla v17.8h, v7.8h, v2.h[0]\n"
+      "sub x27, x27, #0x8\n"
       "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v21.8h, v7.8h, v3.h[0]\n"
       "fmla v25.8h, v7.8h, v4.h[0]\n"
       "ldr q28, [x10, #0x30]\n"
-      "sub x27, x27, #0x8\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.8h, v29.8h, v0.h[0]\n"
       "fmla v14.8h, v29.8h, v1.h[0]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
@@ -3679,9 +3677,9 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr h1, [x23], #0x2\n"
       "ldr h0, [x22], #0x2\n"
       "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
       "fmla v8.8h, v29.8h, v4.h[0]\n"
       "fmla v12.8h, v29.8h, v3.h[0]\n"
+      "ldr q28, [x10, #0x10]\n"
       "fmla v16.8h, v29.8h, v2.h[0]\n"
       "fmla v20.8h, v29.8h, v1.h[0]\n"
       "fmla v24.8h, v29.8h, v0.h[0]\n"
@@ -3710,19 +3708,19 @@ void a64_hybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 219b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "prfm pstl1keep, [x25, #0x0]\n"
-      "add x24, x25, x20, LSL #1\n"
       "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #1\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 227f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v29.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v29.8h }, [x21]\n"
       "ld1r { v28.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v29.8h\n"
       "fmin v9.8h, v9.8h, v29.8h\n"
@@ -3770,223 +3768,223 @@ void a64_hybrid_fp16_mla_6x32 (
       "tbz x11, #4, 235f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
       "st1 { v9.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v25.8h }, [x23], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v25.8h }, [x22], #0x10\n"
       "tbz x11, #3, 231f\n"
       "st1 { v10.8h }, [x9], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
-      "st1 { v26.8h }, [x23], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
+      "st1 { v26.8h }, [x22], #0x10\n"
       "tbz x11, #2, 229f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x11, #1, 228f\n"
       "st1 { v11.s }[2], [x9], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
-      "st1 { v27.s }[2], [x23], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
+      "st1 { v27.s }[2], [x22], #0x4\n"
       "tbz x11, #0, 243f\n"
       "st1 { v11.h }[6], [x9]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
-      "st1 { v27.h }[6], [x23]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
+      "st1 { v27.h }[6], [x22]\n"
       "b 243f\n"
       "228:"  // Height 5: Partial direct writeback: partial_1_28
       "tbz x11, #0, 243f\n"
       "st1 { v11.h }[4], [x9]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
-      "st1 { v27.h }[4], [x23]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
+      "st1 { v27.h }[4], [x22]\n"
       "b 243f\n"
       "229:"  // Height 5: Partial direct writeback: partial_2_24
       "tbz x11, #1, 230f\n"
       "str s11, [x9], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
-      "str s27, [x23], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
+      "str s27, [x22], #0x4\n"
       "tbz x11, #0, 243f\n"
       "st1 { v11.h }[2], [x9]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
-      "st1 { v27.h }[2], [x23]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
+      "st1 { v27.h }[2], [x22]\n"
       "b 243f\n"
       "230:"  // Height 5: Partial direct writeback: partial_1_24
       "tbz x11, #0, 243f\n"
       "str h11, [x9, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
-      "str h27, [x23, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
+      "str h27, [x22, #0x0]\n"
       "b 243f\n"
       "231:"  // Height 5: Partial direct writeback: partial_4_16
       "tbz x11, #2, 233f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x11, #1, 232f\n"
       "st1 { v10.s }[2], [x9], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
-      "st1 { v26.s }[2], [x23], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
+      "st1 { v26.s }[2], [x22], #0x4\n"
       "tbz x11, #0, 243f\n"
       "st1 { v10.h }[6], [x9]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
-      "st1 { v26.h }[6], [x23]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
+      "st1 { v26.h }[6], [x22]\n"
       "b 243f\n"
       "232:"  // Height 5: Partial direct writeback: partial_1_20
       "tbz x11, #0, 243f\n"
       "st1 { v10.h }[4], [x9]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
-      "st1 { v26.h }[4], [x23]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
+      "st1 { v26.h }[4], [x22]\n"
       "b 243f\n"
       "233:"  // Height 5: Partial direct writeback: partial_2_16
       "tbz x11, #1, 234f\n"
       "str s10, [x9], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
-      "str s26, [x23], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
+      "str s26, [x22], #0x4\n"
       "tbz x11, #0, 243f\n"
       "st1 { v10.h }[2], [x9]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
-      "st1 { v26.h }[2], [x23]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
+      "st1 { v26.h }[2], [x22]\n"
       "b 243f\n"
       "234:"  // Height 5: Partial direct writeback: partial_1_16
       "tbz x11, #0, 243f\n"
       "str h10, [x9, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
-      "str h26, [x23, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
+      "str h26, [x22, #0x0]\n"
       "b 243f\n"
       "235:"  // Height 5: Partial direct writeback: partial_8_0
       "tbz x11, #3, 239f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
       "tbz x11, #2, 237f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x11, #1, 236f\n"
       "st1 { v9.s }[2], [x9], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
-      "st1 { v25.s }[2], [x23], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
+      "st1 { v25.s }[2], [x22], #0x4\n"
       "tbz x11, #0, 243f\n"
       "st1 { v9.h }[6], [x9]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
-      "st1 { v25.h }[6], [x23]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
+      "st1 { v25.h }[6], [x22]\n"
       "b 243f\n"
       "236:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x11, #0, 243f\n"
       "st1 { v9.h }[4], [x9]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
-      "st1 { v25.h }[4], [x23]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
+      "st1 { v25.h }[4], [x22]\n"
       "b 243f\n"
       "237:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x11, #1, 238f\n"
       "str s9, [x9], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
-      "str s25, [x23], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
+      "str s25, [x22], #0x4\n"
       "tbz x11, #0, 243f\n"
       "st1 { v9.h }[2], [x9]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
-      "st1 { v25.h }[2], [x23]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
+      "st1 { v25.h }[2], [x22]\n"
       "b 243f\n"
       "238:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x11, #0, 243f\n"
       "str h9, [x9, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
-      "str h25, [x23, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
+      "str h25, [x22, #0x0]\n"
       "b 243f\n"
       "239:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x11, #2, 241f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x11, #1, 240f\n"
       "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
       "tbz x11, #0, 243f\n"
       "st1 { v8.h }[6], [x9]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
-      "st1 { v24.h }[6], [x23]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
+      "st1 { v24.h }[6], [x22]\n"
       "b 243f\n"
       "240:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x11, #0, 243f\n"
       "st1 { v8.h }[4], [x9]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
-      "st1 { v24.h }[4], [x23]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
+      "st1 { v24.h }[4], [x22]\n"
       "b 243f\n"
       "241:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x11, #1, 242f\n"
       "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
       "tbz x11, #0, 243f\n"
       "st1 { v8.h }[2], [x9]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
-      "st1 { v24.h }[2], [x23]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
+      "st1 { v24.h }[2], [x22]\n"
       "b 243f\n"
       "242:"  // Height 5: Partial direct writeback: partial_1_0
       "str h8, [x9, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
-      "str h24, [x23, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
+      "str h24, [x22, #0x0]\n"
       "243:"  // Height 5: Partial direct writeback: Done
       "b 245f\n"
       "244:"  // Height 5: Full writeback
@@ -3995,51 +3993,50 @@ void a64_hybrid_fp16_mla_6x32 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "245:"  // Height 5: Writeback done
       "subs x11, x11, #0x20\n"
       "bgt 198b\n"
       "b 296f\n"
       "246:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0xc\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "247:"  // Height 6: Column loop
       "cbz x12, 248f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -4055,281 +4052,281 @@ void a64_hybrid_fp16_mla_6x32 (
       "248:"  // Height 6: no bias
       "tbz %x[flags], #0, 266f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x20\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "cmp x11, #0x20\n"
+      "add x21, x22, x20, LSL #1\n"
       "bge 265f\n"
       "tbz x11, #4, 256f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
-      "ld1 { v28.8h }, [x22], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
+      "ld1 { v28.8h }, [x21], #0x10\n"
       "ld1 { v9.8h }, [x9], #0x10\n"
-      "ld1 { v13.8h }, [x26], #0x10\n"
-      "ld1 { v17.8h }, [x25], #0x10\n"
-      "ld1 { v21.8h }, [x24], #0x10\n"
-      "ld1 { v25.8h }, [x23], #0x10\n"
-      "ld1 { v29.8h }, [x22], #0x10\n"
+      "ld1 { v13.8h }, [x25], #0x10\n"
+      "ld1 { v17.8h }, [x24], #0x10\n"
+      "ld1 { v21.8h }, [x23], #0x10\n"
+      "ld1 { v25.8h }, [x22], #0x10\n"
+      "ld1 { v29.8h }, [x21], #0x10\n"
       "tbz x11, #3, 252f\n"
       "ld1 { v10.8h }, [x9], #0x10\n"
-      "ld1 { v14.8h }, [x26], #0x10\n"
-      "ld1 { v18.8h }, [x25], #0x10\n"
-      "ld1 { v22.8h }, [x24], #0x10\n"
-      "ld1 { v26.8h }, [x23], #0x10\n"
-      "ld1 { v30.8h }, [x22], #0x10\n"
+      "ld1 { v14.8h }, [x25], #0x10\n"
+      "ld1 { v18.8h }, [x24], #0x10\n"
+      "ld1 { v22.8h }, [x23], #0x10\n"
+      "ld1 { v26.8h }, [x22], #0x10\n"
+      "ld1 { v30.8h }, [x21], #0x10\n"
       "tbz x11, #2, 250f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x11, #1, 249f\n"
       "ld1 { v11.s }[2], [x9], #0x4\n"
-      "ld1 { v15.s }[2], [x26], #0x4\n"
+      "ld1 { v15.s }[2], [x25], #0x4\n"
       "mov x20, #0x3c\n"
-      "ld1 { v19.s }[2], [x25], #0x4\n"
-      "ld1 { v23.s }[2], [x24], #0x4\n"
-      "ld1 { v27.s }[2], [x23], #0x4\n"
-      "ld1 { v31.s }[2], [x22], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "ld1 { v23.s }[2], [x23], #0x4\n"
+      "ld1 { v27.s }[2], [x22], #0x4\n"
+      "ld1 { v31.s }[2], [x21], #0x4\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v11.h }[6], [x9]\n"
-      "ld1 { v15.h }[6], [x26]\n"
-      "ld1 { v19.h }[6], [x25]\n"
-      "ld1 { v23.h }[6], [x24]\n"
-      "ld1 { v27.h }[6], [x23]\n"
-      "ld1 { v31.h }[6], [x22]\n"
+      "ld1 { v15.h }[6], [x25]\n"
+      "ld1 { v19.h }[6], [x24]\n"
+      "ld1 { v23.h }[6], [x23]\n"
+      "ld1 { v27.h }[6], [x22]\n"
+      "ld1 { v31.h }[6], [x21]\n"
       "b 264f\n"
       "249:"  // Height 6: Partial accumulate: partial_1_28
       "mov x20, #0x38\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v11.h }[4], [x9]\n"
-      "ld1 { v15.h }[4], [x26]\n"
-      "ld1 { v19.h }[4], [x25]\n"
-      "ld1 { v23.h }[4], [x24]\n"
-      "ld1 { v27.h }[4], [x23]\n"
-      "ld1 { v31.h }[4], [x22]\n"
+      "ld1 { v15.h }[4], [x25]\n"
+      "ld1 { v19.h }[4], [x24]\n"
+      "ld1 { v23.h }[4], [x23]\n"
+      "ld1 { v27.h }[4], [x22]\n"
+      "ld1 { v31.h }[4], [x21]\n"
       "b 264f\n"
       "250:"  // Height 6: Partial accumulate: partial_2_24
       "tbz x11, #1, 251f\n"
       "ldr s11, [x9], #0x4\n"
-      "ldr s15, [x26], #0x4\n"
+      "ldr s15, [x25], #0x4\n"
       "mov x20, #0x34\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s23, [x24], #0x4\n"
-      "ldr s27, [x23], #0x4\n"
-      "ldr s31, [x22], #0x4\n"
+      "ldr s19, [x24], #0x4\n"
+      "ldr s23, [x23], #0x4\n"
+      "ldr s27, [x22], #0x4\n"
+      "ldr s31, [x21], #0x4\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v11.h }[2], [x9]\n"
-      "ld1 { v15.h }[2], [x26]\n"
-      "ld1 { v19.h }[2], [x25]\n"
-      "ld1 { v23.h }[2], [x24]\n"
-      "ld1 { v27.h }[2], [x23]\n"
-      "ld1 { v31.h }[2], [x22]\n"
+      "ld1 { v15.h }[2], [x25]\n"
+      "ld1 { v19.h }[2], [x24]\n"
+      "ld1 { v23.h }[2], [x23]\n"
+      "ld1 { v27.h }[2], [x22]\n"
+      "ld1 { v31.h }[2], [x21]\n"
       "b 264f\n"
       "251:"  // Height 6: Partial accumulate: partial_1_24
       "mov x20, #0x30\n"
       "tbz x11, #0, 264f\n"
       "ldr h11, [x9, #0x0]\n"
-      "ldr h15, [x26, #0x0]\n"
-      "ldr h19, [x25, #0x0]\n"
-      "ldr h23, [x24, #0x0]\n"
-      "ldr h27, [x23, #0x0]\n"
-      "ldr h31, [x22, #0x0]\n"
+      "ldr h15, [x25, #0x0]\n"
+      "ldr h19, [x24, #0x0]\n"
+      "ldr h23, [x23, #0x0]\n"
+      "ldr h27, [x22, #0x0]\n"
+      "ldr h31, [x21, #0x0]\n"
       "b 264f\n"
       "252:"  // Height 6: Partial accumulate: partial_4_16
       "tbz x11, #2, 254f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x11, #1, 253f\n"
       "ld1 { v10.s }[2], [x9], #0x4\n"
-      "ld1 { v14.s }[2], [x26], #0x4\n"
+      "ld1 { v14.s }[2], [x25], #0x4\n"
       "mov x20, #0x2c\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v22.s }[2], [x24], #0x4\n"
-      "ld1 { v26.s }[2], [x23], #0x4\n"
-      "ld1 { v30.s }[2], [x22], #0x4\n"
+      "ld1 { v18.s }[2], [x24], #0x4\n"
+      "ld1 { v22.s }[2], [x23], #0x4\n"
+      "ld1 { v26.s }[2], [x22], #0x4\n"
+      "ld1 { v30.s }[2], [x21], #0x4\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v10.h }[6], [x9]\n"
-      "ld1 { v14.h }[6], [x26]\n"
-      "ld1 { v18.h }[6], [x25]\n"
-      "ld1 { v22.h }[6], [x24]\n"
-      "ld1 { v26.h }[6], [x23]\n"
-      "ld1 { v30.h }[6], [x22]\n"
+      "ld1 { v14.h }[6], [x25]\n"
+      "ld1 { v18.h }[6], [x24]\n"
+      "ld1 { v22.h }[6], [x23]\n"
+      "ld1 { v26.h }[6], [x22]\n"
+      "ld1 { v30.h }[6], [x21]\n"
       "b 264f\n"
       "253:"  // Height 6: Partial accumulate: partial_1_20
       "mov x20, #0x28\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v10.h }[4], [x9]\n"
-      "ld1 { v14.h }[4], [x26]\n"
-      "ld1 { v18.h }[4], [x25]\n"
-      "ld1 { v22.h }[4], [x24]\n"
-      "ld1 { v26.h }[4], [x23]\n"
-      "ld1 { v30.h }[4], [x22]\n"
+      "ld1 { v14.h }[4], [x25]\n"
+      "ld1 { v18.h }[4], [x24]\n"
+      "ld1 { v22.h }[4], [x23]\n"
+      "ld1 { v26.h }[4], [x22]\n"
+      "ld1 { v30.h }[4], [x21]\n"
       "b 264f\n"
       "254:"  // Height 6: Partial accumulate: partial_2_16
       "tbz x11, #1, 255f\n"
       "ldr s10, [x9], #0x4\n"
-      "ldr s14, [x26], #0x4\n"
+      "ldr s14, [x25], #0x4\n"
       "mov x20, #0x24\n"
-      "ldr s18, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr s26, [x23], #0x4\n"
-      "ldr s30, [x22], #0x4\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr s22, [x23], #0x4\n"
+      "ldr s26, [x22], #0x4\n"
+      "ldr s30, [x21], #0x4\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v10.h }[2], [x9]\n"
-      "ld1 { v14.h }[2], [x26]\n"
-      "ld1 { v18.h }[2], [x25]\n"
-      "ld1 { v22.h }[2], [x24]\n"
-      "ld1 { v26.h }[2], [x23]\n"
-      "ld1 { v30.h }[2], [x22]\n"
+      "ld1 { v14.h }[2], [x25]\n"
+      "ld1 { v18.h }[2], [x24]\n"
+      "ld1 { v22.h }[2], [x23]\n"
+      "ld1 { v26.h }[2], [x22]\n"
+      "ld1 { v30.h }[2], [x21]\n"
       "b 264f\n"
       "255:"  // Height 6: Partial accumulate: partial_1_16
       "mov x20, #0x20\n"
       "tbz x11, #0, 264f\n"
       "ldr h10, [x9, #0x0]\n"
-      "ldr h14, [x26, #0x0]\n"
-      "ldr h18, [x25, #0x0]\n"
-      "ldr h22, [x24, #0x0]\n"
-      "ldr h26, [x23, #0x0]\n"
-      "ldr h30, [x22, #0x0]\n"
+      "ldr h14, [x25, #0x0]\n"
+      "ldr h18, [x24, #0x0]\n"
+      "ldr h22, [x23, #0x0]\n"
+      "ldr h26, [x22, #0x0]\n"
+      "ldr h30, [x21, #0x0]\n"
       "b 264f\n"
       "256:"  // Height 6: Partial accumulate: partial_8_0
       "tbz x11, #3, 260f\n"
       "ld1 { v8.8h }, [x9], #0x10\n"
-      "ld1 { v12.8h }, [x26], #0x10\n"
-      "ld1 { v16.8h }, [x25], #0x10\n"
-      "ld1 { v20.8h }, [x24], #0x10\n"
-      "ld1 { v24.8h }, [x23], #0x10\n"
-      "ld1 { v28.8h }, [x22], #0x10\n"
+      "ld1 { v12.8h }, [x25], #0x10\n"
+      "ld1 { v16.8h }, [x24], #0x10\n"
+      "ld1 { v20.8h }, [x23], #0x10\n"
+      "ld1 { v24.8h }, [x22], #0x10\n"
+      "ld1 { v28.8h }, [x21], #0x10\n"
       "tbz x11, #2, 258f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x11, #1, 257f\n"
       "ld1 { v9.s }[2], [x9], #0x4\n"
-      "ld1 { v13.s }[2], [x26], #0x4\n"
+      "ld1 { v13.s }[2], [x25], #0x4\n"
       "mov x20, #0x1c\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v21.s }[2], [x24], #0x4\n"
-      "ld1 { v25.s }[2], [x23], #0x4\n"
-      "ld1 { v29.s }[2], [x22], #0x4\n"
+      "ld1 { v17.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x23], #0x4\n"
+      "ld1 { v25.s }[2], [x22], #0x4\n"
+      "ld1 { v29.s }[2], [x21], #0x4\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v9.h }[6], [x9]\n"
-      "ld1 { v13.h }[6], [x26]\n"
-      "ld1 { v17.h }[6], [x25]\n"
-      "ld1 { v21.h }[6], [x24]\n"
-      "ld1 { v25.h }[6], [x23]\n"
-      "ld1 { v29.h }[6], [x22]\n"
+      "ld1 { v13.h }[6], [x25]\n"
+      "ld1 { v17.h }[6], [x24]\n"
+      "ld1 { v21.h }[6], [x23]\n"
+      "ld1 { v25.h }[6], [x22]\n"
+      "ld1 { v29.h }[6], [x21]\n"
       "b 264f\n"
       "257:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x18\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v9.h }[4], [x9]\n"
-      "ld1 { v13.h }[4], [x26]\n"
-      "ld1 { v17.h }[4], [x25]\n"
-      "ld1 { v21.h }[4], [x24]\n"
-      "ld1 { v25.h }[4], [x23]\n"
-      "ld1 { v29.h }[4], [x22]\n"
+      "ld1 { v13.h }[4], [x25]\n"
+      "ld1 { v17.h }[4], [x24]\n"
+      "ld1 { v21.h }[4], [x23]\n"
+      "ld1 { v25.h }[4], [x22]\n"
+      "ld1 { v29.h }[4], [x21]\n"
       "b 264f\n"
       "258:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x11, #1, 259f\n"
       "ldr s9, [x9], #0x4\n"
-      "ldr s13, [x26], #0x4\n"
+      "ldr s13, [x25], #0x4\n"
       "mov x20, #0x14\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s21, [x24], #0x4\n"
-      "ldr s25, [x23], #0x4\n"
-      "ldr s29, [x22], #0x4\n"
+      "ldr s17, [x24], #0x4\n"
+      "ldr s21, [x23], #0x4\n"
+      "ldr s25, [x22], #0x4\n"
+      "ldr s29, [x21], #0x4\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v9.h }[2], [x9]\n"
-      "ld1 { v13.h }[2], [x26]\n"
-      "ld1 { v17.h }[2], [x25]\n"
-      "ld1 { v21.h }[2], [x24]\n"
-      "ld1 { v25.h }[2], [x23]\n"
-      "ld1 { v29.h }[2], [x22]\n"
+      "ld1 { v13.h }[2], [x25]\n"
+      "ld1 { v17.h }[2], [x24]\n"
+      "ld1 { v21.h }[2], [x23]\n"
+      "ld1 { v25.h }[2], [x22]\n"
+      "ld1 { v29.h }[2], [x21]\n"
       "b 264f\n"
       "259:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x10\n"
       "tbz x11, #0, 264f\n"
       "ldr h9, [x9, #0x0]\n"
-      "ldr h13, [x26, #0x0]\n"
-      "ldr h17, [x25, #0x0]\n"
-      "ldr h21, [x24, #0x0]\n"
-      "ldr h25, [x23, #0x0]\n"
-      "ldr h29, [x22, #0x0]\n"
+      "ldr h13, [x25, #0x0]\n"
+      "ldr h17, [x24, #0x0]\n"
+      "ldr h21, [x23, #0x0]\n"
+      "ldr h25, [x22, #0x0]\n"
+      "ldr h29, [x21, #0x0]\n"
       "b 264f\n"
       "260:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x11, #2, 262f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x11, #1, 261f\n"
       "ld1 { v8.s }[2], [x9], #0x4\n"
-      "ld1 { v12.s }[2], [x26], #0x4\n"
+      "ld1 { v12.s }[2], [x25], #0x4\n"
       "mov x20, #0xc\n"
-      "ld1 { v16.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
-      "ld1 { v24.s }[2], [x23], #0x4\n"
-      "ld1 { v28.s }[2], [x22], #0x4\n"
+      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v20.s }[2], [x23], #0x4\n"
+      "ld1 { v24.s }[2], [x22], #0x4\n"
+      "ld1 { v28.s }[2], [x21], #0x4\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v8.h }[6], [x9]\n"
-      "ld1 { v12.h }[6], [x26]\n"
-      "ld1 { v16.h }[6], [x25]\n"
-      "ld1 { v20.h }[6], [x24]\n"
-      "ld1 { v24.h }[6], [x23]\n"
-      "ld1 { v28.h }[6], [x22]\n"
+      "ld1 { v12.h }[6], [x25]\n"
+      "ld1 { v16.h }[6], [x24]\n"
+      "ld1 { v20.h }[6], [x23]\n"
+      "ld1 { v24.h }[6], [x22]\n"
+      "ld1 { v28.h }[6], [x21]\n"
       "b 264f\n"
       "261:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x8\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v8.h }[4], [x9]\n"
-      "ld1 { v12.h }[4], [x26]\n"
-      "ld1 { v16.h }[4], [x25]\n"
-      "ld1 { v20.h }[4], [x24]\n"
-      "ld1 { v24.h }[4], [x23]\n"
-      "ld1 { v28.h }[4], [x22]\n"
+      "ld1 { v12.h }[4], [x25]\n"
+      "ld1 { v16.h }[4], [x24]\n"
+      "ld1 { v20.h }[4], [x23]\n"
+      "ld1 { v24.h }[4], [x22]\n"
+      "ld1 { v28.h }[4], [x21]\n"
       "b 264f\n"
       "262:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x11, #1, 263f\n"
       "ldr s8, [x9], #0x4\n"
-      "ldr s12, [x26], #0x4\n"
+      "ldr s12, [x25], #0x4\n"
       "mov x20, #0x4\n"
-      "ldr s16, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s28, [x22], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "ldr s20, [x23], #0x4\n"
+      "ldr s24, [x22], #0x4\n"
+      "ldr s28, [x21], #0x4\n"
       "tbz x11, #0, 264f\n"
       "ld1 { v8.h }[2], [x9]\n"
-      "ld1 { v12.h }[2], [x26]\n"
-      "ld1 { v16.h }[2], [x25]\n"
-      "ld1 { v20.h }[2], [x24]\n"
-      "ld1 { v24.h }[2], [x23]\n"
-      "ld1 { v28.h }[2], [x22]\n"
+      "ld1 { v12.h }[2], [x25]\n"
+      "ld1 { v16.h }[2], [x24]\n"
+      "ld1 { v20.h }[2], [x23]\n"
+      "ld1 { v24.h }[2], [x22]\n"
+      "ld1 { v28.h }[2], [x21]\n"
       "b 264f\n"
       "263:"  // Height 6: Partial accumulate: partial_1_0
       "ldr h8, [x9, #0x0]\n"
-      "ldr h12, [x26, #0x0]\n"
+      "ldr h12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr h16, [x25, #0x0]\n"
-      "ldr h20, [x24, #0x0]\n"
-      "ldr h24, [x23, #0x0]\n"
-      "ldr h28, [x22, #0x0]\n"
+      "ldr h16, [x24, #0x0]\n"
+      "ldr h20, [x23, #0x0]\n"
+      "ldr h24, [x22, #0x0]\n"
+      "ldr h28, [x21, #0x0]\n"
       "264:"  // Height 6: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 267f\n"
@@ -4338,26 +4335,26 @@ void a64_hybrid_fp16_mla_6x32 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "b 267f\n"
       "266:"  // Height 6: no accumulate
       "movi v8.16b, #0x0\n"
@@ -4388,8 +4385,8 @@ void a64_hybrid_fp16_mla_6x32 (
       "mov x28, #0x0\n"
       "268:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 269f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -4691,18 +4688,18 @@ void a64_hybrid_fp16_mla_6x32 (
       "fmla v9.8h, v7.8h, v0.h[0]\n"
       "fmla v13.8h, v7.8h, v1.h[0]\n"
       "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x8\n"
       "fmla v17.8h, v7.8h, v2.h[0]\n"
       "fmla v21.8h, v7.8h, v3.h[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v25.8h, v7.8h, v4.h[0]\n"
       "fmla v29.8h, v7.8h, v5.h[0]\n"
       "ldr q7, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v10.8h, v6.8h, v0.h[0]\n"
       "fmla v14.8h, v6.8h, v1.h[0]\n"
-      "sub x27, x27, #0x8\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       "prfm pldl1keep, [x22, #0x80]\n"
       "fmla v18.8h, v6.8h, v2.h[0]\n"
       "fmla v22.8h, v6.8h, v3.h[0]\n"
@@ -4958,21 +4955,21 @@ void a64_hybrid_fp16_mla_6x32 (
       "cmp x28, x20\n"
       "bne 268b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #1\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "add x21, x22, x20, LSL #1\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbz %x[flags], #1, 276f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.8h }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.8h }, [x21]\n"
       "ld1r { v0.8h }, [x20]\n"
       "fmin v8.8h, v8.8h, v1.8h\n"
       "fmin v9.8h, v9.8h, v1.8h\n"
@@ -5028,255 +5025,255 @@ void a64_hybrid_fp16_mla_6x32 (
       "tbz x11, #4, 284f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
       "st1 { v9.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v13.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v17.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v21.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v25.8h }, [x23], #0x10\n"
-      "st1 { v28.8h }, [x22], #0x10\n"
-      "st1 { v29.8h }, [x22], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v13.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v17.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v21.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v25.8h }, [x22], #0x10\n"
+      "st1 { v28.8h }, [x21], #0x10\n"
+      "st1 { v29.8h }, [x21], #0x10\n"
       "tbz x11, #3, 280f\n"
       "st1 { v10.8h }, [x9], #0x10\n"
-      "st1 { v14.8h }, [x26], #0x10\n"
-      "st1 { v18.8h }, [x25], #0x10\n"
-      "st1 { v22.8h }, [x24], #0x10\n"
-      "st1 { v26.8h }, [x23], #0x10\n"
-      "st1 { v30.8h }, [x22], #0x10\n"
+      "st1 { v14.8h }, [x25], #0x10\n"
+      "st1 { v18.8h }, [x24], #0x10\n"
+      "st1 { v22.8h }, [x23], #0x10\n"
+      "st1 { v26.8h }, [x22], #0x10\n"
+      "st1 { v30.8h }, [x21], #0x10\n"
       "tbz x11, #2, 278f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x11, #1, 277f\n"
       "st1 { v11.s }[2], [x9], #0x4\n"
-      "st1 { v15.s }[2], [x26], #0x4\n"
-      "st1 { v19.s }[2], [x25], #0x4\n"
-      "st1 { v23.s }[2], [x24], #0x4\n"
-      "st1 { v27.s }[2], [x23], #0x4\n"
-      "st1 { v31.s }[2], [x22], #0x4\n"
+      "st1 { v15.s }[2], [x25], #0x4\n"
+      "st1 { v19.s }[2], [x24], #0x4\n"
+      "st1 { v23.s }[2], [x23], #0x4\n"
+      "st1 { v27.s }[2], [x22], #0x4\n"
+      "st1 { v31.s }[2], [x21], #0x4\n"
       "tbz x11, #0, 292f\n"
       "st1 { v11.h }[6], [x9]\n"
-      "st1 { v15.h }[6], [x26]\n"
-      "st1 { v19.h }[6], [x25]\n"
-      "st1 { v23.h }[6], [x24]\n"
-      "st1 { v27.h }[6], [x23]\n"
-      "st1 { v31.h }[6], [x22]\n"
+      "st1 { v15.h }[6], [x25]\n"
+      "st1 { v19.h }[6], [x24]\n"
+      "st1 { v23.h }[6], [x23]\n"
+      "st1 { v27.h }[6], [x22]\n"
+      "st1 { v31.h }[6], [x21]\n"
       "b 292f\n"
       "277:"  // Height 6: Partial direct writeback: partial_1_28
       "tbz x11, #0, 292f\n"
       "st1 { v11.h }[4], [x9]\n"
-      "st1 { v15.h }[4], [x26]\n"
-      "st1 { v19.h }[4], [x25]\n"
-      "st1 { v23.h }[4], [x24]\n"
-      "st1 { v27.h }[4], [x23]\n"
-      "st1 { v31.h }[4], [x22]\n"
+      "st1 { v15.h }[4], [x25]\n"
+      "st1 { v19.h }[4], [x24]\n"
+      "st1 { v23.h }[4], [x23]\n"
+      "st1 { v27.h }[4], [x22]\n"
+      "st1 { v31.h }[4], [x21]\n"
       "b 292f\n"
       "278:"  // Height 6: Partial direct writeback: partial_2_24
       "tbz x11, #1, 279f\n"
       "str s11, [x9], #0x4\n"
-      "str s15, [x26], #0x4\n"
-      "str s19, [x25], #0x4\n"
-      "str s23, [x24], #0x4\n"
-      "str s27, [x23], #0x4\n"
-      "str s31, [x22], #0x4\n"
+      "str s15, [x25], #0x4\n"
+      "str s19, [x24], #0x4\n"
+      "str s23, [x23], #0x4\n"
+      "str s27, [x22], #0x4\n"
+      "str s31, [x21], #0x4\n"
       "tbz x11, #0, 292f\n"
       "st1 { v11.h }[2], [x9]\n"
-      "st1 { v15.h }[2], [x26]\n"
-      "st1 { v19.h }[2], [x25]\n"
-      "st1 { v23.h }[2], [x24]\n"
-      "st1 { v27.h }[2], [x23]\n"
-      "st1 { v31.h }[2], [x22]\n"
+      "st1 { v15.h }[2], [x25]\n"
+      "st1 { v19.h }[2], [x24]\n"
+      "st1 { v23.h }[2], [x23]\n"
+      "st1 { v27.h }[2], [x22]\n"
+      "st1 { v31.h }[2], [x21]\n"
       "b 292f\n"
       "279:"  // Height 6: Partial direct writeback: partial_1_24
       "tbz x11, #0, 292f\n"
       "str h11, [x9, #0x0]\n"
-      "str h15, [x26, #0x0]\n"
-      "str h19, [x25, #0x0]\n"
-      "str h23, [x24, #0x0]\n"
-      "str h27, [x23, #0x0]\n"
-      "str h31, [x22, #0x0]\n"
+      "str h15, [x25, #0x0]\n"
+      "str h19, [x24, #0x0]\n"
+      "str h23, [x23, #0x0]\n"
+      "str h27, [x22, #0x0]\n"
+      "str h31, [x21, #0x0]\n"
       "b 292f\n"
       "280:"  // Height 6: Partial direct writeback: partial_4_16
       "tbz x11, #2, 282f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x11, #1, 281f\n"
       "st1 { v10.s }[2], [x9], #0x4\n"
-      "st1 { v14.s }[2], [x26], #0x4\n"
-      "st1 { v18.s }[2], [x25], #0x4\n"
-      "st1 { v22.s }[2], [x24], #0x4\n"
-      "st1 { v26.s }[2], [x23], #0x4\n"
-      "st1 { v30.s }[2], [x22], #0x4\n"
+      "st1 { v14.s }[2], [x25], #0x4\n"
+      "st1 { v18.s }[2], [x24], #0x4\n"
+      "st1 { v22.s }[2], [x23], #0x4\n"
+      "st1 { v26.s }[2], [x22], #0x4\n"
+      "st1 { v30.s }[2], [x21], #0x4\n"
       "tbz x11, #0, 292f\n"
       "st1 { v10.h }[6], [x9]\n"
-      "st1 { v14.h }[6], [x26]\n"
-      "st1 { v18.h }[6], [x25]\n"
-      "st1 { v22.h }[6], [x24]\n"
-      "st1 { v26.h }[6], [x23]\n"
-      "st1 { v30.h }[6], [x22]\n"
+      "st1 { v14.h }[6], [x25]\n"
+      "st1 { v18.h }[6], [x24]\n"
+      "st1 { v22.h }[6], [x23]\n"
+      "st1 { v26.h }[6], [x22]\n"
+      "st1 { v30.h }[6], [x21]\n"
       "b 292f\n"
       "281:"  // Height 6: Partial direct writeback: partial_1_20
       "tbz x11, #0, 292f\n"
       "st1 { v10.h }[4], [x9]\n"
-      "st1 { v14.h }[4], [x26]\n"
-      "st1 { v18.h }[4], [x25]\n"
-      "st1 { v22.h }[4], [x24]\n"
-      "st1 { v26.h }[4], [x23]\n"
-      "st1 { v30.h }[4], [x22]\n"
+      "st1 { v14.h }[4], [x25]\n"
+      "st1 { v18.h }[4], [x24]\n"
+      "st1 { v22.h }[4], [x23]\n"
+      "st1 { v26.h }[4], [x22]\n"
+      "st1 { v30.h }[4], [x21]\n"
       "b 292f\n"
       "282:"  // Height 6: Partial direct writeback: partial_2_16
       "tbz x11, #1, 283f\n"
       "str s10, [x9], #0x4\n"
-      "str s14, [x26], #0x4\n"
-      "str s18, [x25], #0x4\n"
-      "str s22, [x24], #0x4\n"
-      "str s26, [x23], #0x4\n"
-      "str s30, [x22], #0x4\n"
+      "str s14, [x25], #0x4\n"
+      "str s18, [x24], #0x4\n"
+      "str s22, [x23], #0x4\n"
+      "str s26, [x22], #0x4\n"
+      "str s30, [x21], #0x4\n"
       "tbz x11, #0, 292f\n"
       "st1 { v10.h }[2], [x9]\n"
-      "st1 { v14.h }[2], [x26]\n"
-      "st1 { v18.h }[2], [x25]\n"
-      "st1 { v22.h }[2], [x24]\n"
-      "st1 { v26.h }[2], [x23]\n"
-      "st1 { v30.h }[2], [x22]\n"
+      "st1 { v14.h }[2], [x25]\n"
+      "st1 { v18.h }[2], [x24]\n"
+      "st1 { v22.h }[2], [x23]\n"
+      "st1 { v26.h }[2], [x22]\n"
+      "st1 { v30.h }[2], [x21]\n"
       "b 292f\n"
       "283:"  // Height 6: Partial direct writeback: partial_1_16
       "tbz x11, #0, 292f\n"
       "str h10, [x9, #0x0]\n"
-      "str h14, [x26, #0x0]\n"
-      "str h18, [x25, #0x0]\n"
-      "str h22, [x24, #0x0]\n"
-      "str h26, [x23, #0x0]\n"
-      "str h30, [x22, #0x0]\n"
+      "str h14, [x25, #0x0]\n"
+      "str h18, [x24, #0x0]\n"
+      "str h22, [x23, #0x0]\n"
+      "str h26, [x22, #0x0]\n"
+      "str h30, [x21, #0x0]\n"
       "b 292f\n"
       "284:"  // Height 6: Partial direct writeback: partial_8_0
       "tbz x11, #3, 288f\n"
       "st1 { v8.8h }, [x9], #0x10\n"
-      "st1 { v12.8h }, [x26], #0x10\n"
-      "st1 { v16.8h }, [x25], #0x10\n"
-      "st1 { v20.8h }, [x24], #0x10\n"
-      "st1 { v24.8h }, [x23], #0x10\n"
-      "st1 { v28.8h }, [x22], #0x10\n"
+      "st1 { v12.8h }, [x25], #0x10\n"
+      "st1 { v16.8h }, [x24], #0x10\n"
+      "st1 { v20.8h }, [x23], #0x10\n"
+      "st1 { v24.8h }, [x22], #0x10\n"
+      "st1 { v28.8h }, [x21], #0x10\n"
       "tbz x11, #2, 286f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x11, #1, 285f\n"
       "st1 { v9.s }[2], [x9], #0x4\n"
-      "st1 { v13.s }[2], [x26], #0x4\n"
-      "st1 { v17.s }[2], [x25], #0x4\n"
-      "st1 { v21.s }[2], [x24], #0x4\n"
-      "st1 { v25.s }[2], [x23], #0x4\n"
-      "st1 { v29.s }[2], [x22], #0x4\n"
+      "st1 { v13.s }[2], [x25], #0x4\n"
+      "st1 { v17.s }[2], [x24], #0x4\n"
+      "st1 { v21.s }[2], [x23], #0x4\n"
+      "st1 { v25.s }[2], [x22], #0x4\n"
+      "st1 { v29.s }[2], [x21], #0x4\n"
       "tbz x11, #0, 292f\n"
       "st1 { v9.h }[6], [x9]\n"
-      "st1 { v13.h }[6], [x26]\n"
-      "st1 { v17.h }[6], [x25]\n"
-      "st1 { v21.h }[6], [x24]\n"
-      "st1 { v25.h }[6], [x23]\n"
-      "st1 { v29.h }[6], [x22]\n"
+      "st1 { v13.h }[6], [x25]\n"
+      "st1 { v17.h }[6], [x24]\n"
+      "st1 { v21.h }[6], [x23]\n"
+      "st1 { v25.h }[6], [x22]\n"
+      "st1 { v29.h }[6], [x21]\n"
       "b 292f\n"
       "285:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x11, #0, 292f\n"
       "st1 { v9.h }[4], [x9]\n"
-      "st1 { v13.h }[4], [x26]\n"
-      "st1 { v17.h }[4], [x25]\n"
-      "st1 { v21.h }[4], [x24]\n"
-      "st1 { v25.h }[4], [x23]\n"
-      "st1 { v29.h }[4], [x22]\n"
+      "st1 { v13.h }[4], [x25]\n"
+      "st1 { v17.h }[4], [x24]\n"
+      "st1 { v21.h }[4], [x23]\n"
+      "st1 { v25.h }[4], [x22]\n"
+      "st1 { v29.h }[4], [x21]\n"
       "b 292f\n"
       "286:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x11, #1, 287f\n"
       "str s9, [x9], #0x4\n"
-      "str s13, [x26], #0x4\n"
-      "str s17, [x25], #0x4\n"
-      "str s21, [x24], #0x4\n"
-      "str s25, [x23], #0x4\n"
-      "str s29, [x22], #0x4\n"
+      "str s13, [x25], #0x4\n"
+      "str s17, [x24], #0x4\n"
+      "str s21, [x23], #0x4\n"
+      "str s25, [x22], #0x4\n"
+      "str s29, [x21], #0x4\n"
       "tbz x11, #0, 292f\n"
       "st1 { v9.h }[2], [x9]\n"
-      "st1 { v13.h }[2], [x26]\n"
-      "st1 { v17.h }[2], [x25]\n"
-      "st1 { v21.h }[2], [x24]\n"
-      "st1 { v25.h }[2], [x23]\n"
-      "st1 { v29.h }[2], [x22]\n"
+      "st1 { v13.h }[2], [x25]\n"
+      "st1 { v17.h }[2], [x24]\n"
+      "st1 { v21.h }[2], [x23]\n"
+      "st1 { v25.h }[2], [x22]\n"
+      "st1 { v29.h }[2], [x21]\n"
       "b 292f\n"
       "287:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x11, #0, 292f\n"
       "str h9, [x9, #0x0]\n"
-      "str h13, [x26, #0x0]\n"
-      "str h17, [x25, #0x0]\n"
-      "str h21, [x24, #0x0]\n"
-      "str h25, [x23, #0x0]\n"
-      "str h29, [x22, #0x0]\n"
+      "str h13, [x25, #0x0]\n"
+      "str h17, [x24, #0x0]\n"
+      "str h21, [x23, #0x0]\n"
+      "str h25, [x22, #0x0]\n"
+      "str h29, [x21, #0x0]\n"
       "b 292f\n"
       "288:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x11, #2, 290f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x11, #1, 289f\n"
       "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
       "tbz x11, #0, 292f\n"
       "st1 { v8.h }[6], [x9]\n"
-      "st1 { v12.h }[6], [x26]\n"
-      "st1 { v16.h }[6], [x25]\n"
-      "st1 { v20.h }[6], [x24]\n"
-      "st1 { v24.h }[6], [x23]\n"
-      "st1 { v28.h }[6], [x22]\n"
+      "st1 { v12.h }[6], [x25]\n"
+      "st1 { v16.h }[6], [x24]\n"
+      "st1 { v20.h }[6], [x23]\n"
+      "st1 { v24.h }[6], [x22]\n"
+      "st1 { v28.h }[6], [x21]\n"
       "b 292f\n"
       "289:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x11, #0, 292f\n"
       "st1 { v8.h }[4], [x9]\n"
-      "st1 { v12.h }[4], [x26]\n"
-      "st1 { v16.h }[4], [x25]\n"
-      "st1 { v20.h }[4], [x24]\n"
-      "st1 { v24.h }[4], [x23]\n"
-      "st1 { v28.h }[4], [x22]\n"
+      "st1 { v12.h }[4], [x25]\n"
+      "st1 { v16.h }[4], [x24]\n"
+      "st1 { v20.h }[4], [x23]\n"
+      "st1 { v24.h }[4], [x22]\n"
+      "st1 { v28.h }[4], [x21]\n"
       "b 292f\n"
       "290:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x11, #1, 291f\n"
       "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
       "tbz x11, #0, 292f\n"
       "st1 { v8.h }[2], [x9]\n"
-      "st1 { v12.h }[2], [x26]\n"
-      "st1 { v16.h }[2], [x25]\n"
-      "st1 { v20.h }[2], [x24]\n"
-      "st1 { v24.h }[2], [x23]\n"
-      "st1 { v28.h }[2], [x22]\n"
+      "st1 { v12.h }[2], [x25]\n"
+      "st1 { v16.h }[2], [x24]\n"
+      "st1 { v20.h }[2], [x23]\n"
+      "st1 { v24.h }[2], [x22]\n"
+      "st1 { v28.h }[2], [x21]\n"
       "b 292f\n"
       "291:"  // Height 6: Partial direct writeback: partial_1_0
       "str h8, [x9, #0x0]\n"
-      "str h12, [x26, #0x0]\n"
-      "str h16, [x25, #0x0]\n"
-      "str h20, [x24, #0x0]\n"
-      "str h24, [x23, #0x0]\n"
-      "str h28, [x22, #0x0]\n"
+      "str h12, [x25, #0x0]\n"
+      "str h16, [x24, #0x0]\n"
+      "str h20, [x23, #0x0]\n"
+      "str h24, [x22, #0x0]\n"
+      "str h28, [x21, #0x0]\n"
       "292:"  // Height 6: Partial direct writeback: Done
       "b 294f\n"
       "293:"  // Height 6: Full writeback
@@ -5285,26 +5282,26 @@ void a64_hybrid_fp16_mla_6x32 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q28, [x22, #0x0]\n"
-      "str q29, [x22, #0x10]\n"
-      "str q30, [x22, #0x20]\n"
-      "str q31, [x22, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
+      "str q28, [x21, #0x0]\n"
+      "str q29, [x21, #0x10]\n"
+      "str q30, [x21, #0x20]\n"
+      "str q31, [x21, #0x30]\n"
       "294:"  // Height 6: Writeback done
       "subs x11, x11, #0x20\n"
       "bgt 247b\n"
@@ -5320,8 +5317,8 @@ void a64_hybrid_fp16_mla_6x32 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "296:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24.hpp
index 7a025a5deb..bce4de74f7 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24.hpp
@@ -71,8 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixedTRB<lhs_operand_type, rhs_operand_type, result_type, 4, 24, 1> transforms = {};
-
+    StdTransformsFixedTRB<rhs_operand_type, result_type, 4, 24, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp
index 23587e6317..52c4e1be65 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp
@@ -47,19 +47,18 @@ void a64_hybrid_fp32_mla_4x24_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void a64_hybrid_fp32_mla_4x24_a55 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -100,10 +98,10 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "cmp %x[M], #0x2\n"
       "bgt 83f\n"
       "beq 42f\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x17, %x[bias]\n"
       "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x14, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x17, 3f\n"
       "ldr q8, [x17, #0x0]\n"
@@ -223,8 +221,8 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "mov x13, #0x0\n"
       "19:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w12, [x20, x13, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -256,50 +254,47 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v11.4s, v7.4s, v0.s[0]\n"
       "ldr d16, [x15, #0x70]\n"
       "mov v19.d[1], x20\n"
-      "ldr x22, [x15, #0x58]\n"
-      "add x11, x11, #0x10\n"
+      "ldr x20, [x15, #0x58]\n"
+      "mov v18.d[1], x20\n"
       "ldr x20, [x15, #0x68]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0x78]\n"
+      "mov v16.d[1], x20\n"
       "fmla v12.4s, v19.4s, v0.s[0]\n"
       "ldr d19, [x15, #0x80]\n"
-      "sub x12, x12, #0x4\n"
-      "ldr x21, [x15, #0x78]\n"
-      "mov v18.d[1], x22\n"
-      "mov v17.d[1], x20\n"
       "ldr x20, [x15, #0x88]\n"
       "fmla v13.4s, v18.4s, v0.s[0]\n"
       "ldr d18, [x15, #0x90]\n"
-      "mov v16.d[1], x21\n"
-      "ldr x22, [x15, #0x98]\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
       "ldr d17, [x15, #0xa0]\n"
-      "ldr x21, [x15, #0xa8]\n"
       "fmla v9.4s, v16.4s, v0.s[1]\n"
       "ldr d16, [x15, #0xb0]\n"
       "mov v19.d[1], x20\n"
+      "ldr x20, [x15, #0x98]\n"
+      "mov v18.d[1], x20\n"
+      "ldr x20, [x15, #0xa8]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x15, #0xb8]\n"
-      "mov v18.d[1], x22\n"
-      "mov v17.d[1], x21\n"
+      "mov v16.d[1], x20\n"
       "fmla v10.4s, v19.4s, v0.s[1]\n"
       "ldr d19, [x15, #0xc0]\n"
+      "ldr x20, [x15, #0xc8]\n"
       "fmla v11.4s, v18.4s, v0.s[1]\n"
       "ldr d18, [x15, #0xd0]\n"
-      "mov v16.d[1], x20\n"
-      "ldr x20, [x15, #0xc8]\n"
       "fmla v12.4s, v17.4s, v0.s[1]\n"
       "ldr d17, [x15, #0xe0]\n"
       "fmla v13.4s, v16.4s, v0.s[1]\n"
       "ldr d16, [x15, #0xf0]\n"
-      "cmp x12, #0x8\n"
-      "ldr x22, [x15, #0xd8]\n"
       "mov v19.d[1], x20\n"
+      "ldr x20, [x15, #0xd8]\n"
+      "mov v18.d[1], x20\n"
       "ldr x20, [x15, #0xe8]\n"
-      "ldr x21, [x15, #0xf8]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0xf8]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.4s, v19.4s, v0.s[2]\n"
       "ldr d19, [x15, #0x100]\n"
-      "mov v18.d[1], x22\n"
-      "mov v17.d[1], x20\n"
       "ldr x20, [x15, #0x108]\n"
-      "mov v16.d[1], x21\n"
       "fmla v9.4s, v18.4s, v0.s[2]\n"
       "ldr d18, [x15, #0x110]\n"
       "fmla v10.4s, v17.4s, v0.s[2]\n"
@@ -307,49 +302,52 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v11.4s, v16.4s, v0.s[2]\n"
       "ldr d16, [x15, #0x130]\n"
       "mov v19.d[1], x20\n"
-      "ldr x22, [x15, #0x118]\n"
+      "ldr x20, [x15, #0x118]\n"
+      "mov v18.d[1], x20\n"
       "ldr x20, [x15, #0x128]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0x138]\n"
+      "mov v16.d[1], x20\n"
       "fmla v12.4s, v19.4s, v0.s[2]\n"
       "ldr d19, [x15, #0x140]\n"
-      "ldr x21, [x15, #0x138]\n"
-      "mov v18.d[1], x22\n"
-      "mov v17.d[1], x20\n"
       "ldr x20, [x15, #0x148]\n"
       "fmla v13.4s, v18.4s, v0.s[2]\n"
       "ldr d18, [x15, #0x150]\n"
-      "mov v16.d[1], x21\n"
-      "ldr x22, [x15, #0x158]\n"
       "fmla v8.4s, v17.4s, v0.s[3]\n"
       "ldr d17, [x15, #0x160]\n"
-      "ldr x21, [x15, #0x168]\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr d16, [x15, #0x170]\n"
       "mov v19.d[1], x20\n"
+      "ldr x20, [x15, #0x158]\n"
+      "mov v18.d[1], x20\n"
+      "ldr x20, [x15, #0x168]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x15, #0x178]\n"
-      "mov v18.d[1], x22\n"
-      "mov v17.d[1], x21\n"
+      "mov v16.d[1], x20\n"
+      "add x11, x11, #0x10\n"
       "add x15, x15, #0x180\n"
       "fmla v10.4s, v19.4s, v0.s[3]\n"
       "ldr d4, [x15, #0x0]\n"
-      "mov v16.d[1], x20\n"
-      "ldr x21, [x15, #0x8]\n"
+      "ldr x20, [x15, #0x8]\n"
       "fmla v11.4s, v18.4s, v0.s[3]\n"
       "ldr d5, [x15, #0x10]\n"
-      "ldr x20, [x15, #0x18]\n"
       "fmla v12.4s, v17.4s, v0.s[3]\n"
       "ldr d6, [x15, #0x20]\n"
       "fmla v13.4s, v16.4s, v0.s[3]\n"
       "ldr d0, [x11, #0x0]\n"
-      "mov v4.d[1], x21\n"
+      "sub x12, x12, #0x4\n"
       "ldr d7, [x15, #0x30]\n"
-      "mov v5.d[1], x20\n"
-      "ldr x22, [x15, #0x28]\n"
+      "cmp x12, #0x8\n"
+      "ldr x21, [x15, #0x18]\n"
+      "mov v4.d[1], x20\n"
+      "ldr x20, [x15, #0x28]\n"
+      "mov v5.d[1], x21\n"
       "ldr x21, [x11, #0x8]\n"
+      "mov v6.d[1], x20\n"
       "ldr x20, [x15, #0x38]\n"
-      "mov v6.d[1], x22\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
       "mov v0.d[1], x21\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "bge 22b\n"
       "23:"  // Height 1: Multiply loop: Single iteration only
       "fmla v8.4s, v4.4s, v0.s[0]\n"
@@ -403,21 +401,21 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "24:"  // Height 1: Multiply loop: Main loop skip
       "cbz x12, 26f\n"
       "25:"  // Height 1: Multiply loop: Odd block loop
-      "ldr s20, [x11], #0x4\n"
+      "ldr s17, [x11], #0x4\n"
       "sub x12, x12, #0x1\n"
-      "ldr q17, [x15, #0x0]\n"
+      "ldr q16, [x15, #0x0]\n"
+      "fmla v8.4s, v16.4s, v17.s[0]\n"
       "ldr q16, [x15, #0x10]\n"
-      "ldr q19, [x15, #0x20]\n"
-      "ldr q18, [x15, #0x30]\n"
-      "fmla v8.4s, v17.4s, v20.s[0]\n"
-      "ldr q17, [x15, #0x40]\n"
-      "fmla v9.4s, v16.4s, v20.s[0]\n"
+      "fmla v9.4s, v16.4s, v17.s[0]\n"
+      "ldr q16, [x15, #0x20]\n"
+      "fmla v10.4s, v16.4s, v17.s[0]\n"
+      "ldr q16, [x15, #0x30]\n"
+      "fmla v11.4s, v16.4s, v17.s[0]\n"
+      "ldr q16, [x15, #0x40]\n"
+      "fmla v12.4s, v16.4s, v17.s[0]\n"
       "ldr q16, [x15, #0x50]\n"
-      "fmla v10.4s, v19.4s, v20.s[0]\n"
-      "fmla v11.4s, v18.4s, v20.s[0]\n"
+      "fmla v13.4s, v16.4s, v17.s[0]\n"
       "add x15, x15, #0x60\n"
-      "fmla v12.4s, v17.4s, v20.s[0]\n"
-      "fmla v13.4s, v16.4s, v20.s[0]\n"
       "cbnz x12, 25b\n"
       "26:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -426,16 +424,16 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "bne 19b\n"
       "prfm pstl1keep, [x14, #0x0]\n"
       "tbz %x[flags], #1, 27f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v16.4s\n"
+      "fmin v9.4s, v9.4s, v16.4s\n"
+      "fmin v10.4s, v10.4s, v16.4s\n"
+      "fmin v11.4s, v11.4s, v16.4s\n"
+      "fmin v12.4s, v12.4s, v16.4s\n"
+      "fmin v13.4s, v13.4s, v16.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v17.4s\n"
-      "fmin v9.4s, v9.4s, v17.4s\n"
-      "fmin v10.4s, v10.4s, v17.4s\n"
-      "fmin v11.4s, v11.4s, v17.4s\n"
-      "fmin v12.4s, v12.4s, v17.4s\n"
-      "fmin v13.4s, v13.4s, v17.4s\n"
       "fmax v8.4s, v8.4s, v16.4s\n"
       "fmax v9.4s, v9.4s, v16.4s\n"
       "fmax v10.4s, v10.4s, v16.4s\n"
@@ -531,142 +529,142 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "bgt 2b\n"
       "b 166f\n"
       "42:"  // Height 2
-      "ldr x17, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x17, %x[bias]\n"
       "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x14, %x[output_ptr]\n"
       "43:"  // Height 2: Column loop
       "cbz x17, 44f\n"
       "ldr q8, [x17, #0x0]\n"
-      "ldr q9, [x17, #0x10]\n"
-      "ldr q10, [x17, #0x20]\n"
-      "ldr q11, [x17, #0x30]\n"
       "mov v14.16b, v8.16b\n"
-      "ldr q12, [x17, #0x40]\n"
+      "ldr q9, [x17, #0x10]\n"
       "mov v15.16b, v9.16b\n"
-      "ldr q13, [x17, #0x50]\n"
+      "ldr q10, [x17, #0x20]\n"
       "mov v16.16b, v10.16b\n"
+      "ldr q11, [x17, #0x30]\n"
       "mov v17.16b, v11.16b\n"
-      "add x17, x17, #0x60\n"
+      "ldr q12, [x17, #0x40]\n"
       "mov v18.16b, v12.16b\n"
+      "ldr q13, [x17, #0x50]\n"
       "mov v19.16b, v13.16b\n"
+      "add x17, x17, #0x60\n"
       "b 59f\n"
       "44:"  // Height 2: no bias
       "tbz %x[flags], #0, 58f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x16, #0x18\n"
-      "add x24, x14, x20, LSL #2\n"
+      "add x23, x14, x20, LSL #2\n"
       "bge 57f\n"
       "tbz x16, #4, 48f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "ld1 { v9.4s }, [x14], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x14], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v11.4s }, [x14], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
       "tbz x16, #2, 46f\n"
       "ld1 { v12.4s }, [x14], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
       "tbz x16, #1, 45f\n"
       "ldr d13, [x14], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "tbz x16, #0, 56f\n"
       "ld1 { v13.s }[2], [x14]\n"
-      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v19.s }[2], [x23]\n"
       "b 56f\n"
       "45:"  // Height 2: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x16, #0, 56f\n"
       "ldr s13, [x14, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
       "b 56f\n"
       "46:"  // Height 2: Partial accumulate: partial_2_16
       "tbz x16, #1, 47f\n"
       "ldr d12, [x14], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "tbz x16, #0, 56f\n"
       "ld1 { v12.s }[2], [x14]\n"
-      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v18.s }[2], [x23]\n"
       "b 56f\n"
       "47:"  // Height 2: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x16, #0, 56f\n"
       "ldr s12, [x14, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
       "b 56f\n"
       "48:"  // Height 2: Partial accumulate: partial_8_0
       "tbz x16, #3, 52f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "ld1 { v9.4s }, [x14], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
       "tbz x16, #2, 50f\n"
       "ld1 { v10.4s }, [x14], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
       "tbz x16, #1, 49f\n"
       "ldr d11, [x14], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "tbz x16, #0, 56f\n"
       "ld1 { v11.s }[2], [x14]\n"
-      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v17.s }[2], [x23]\n"
       "b 56f\n"
       "49:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x16, #0, 56f\n"
       "ldr s11, [x14, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
       "b 56f\n"
       "50:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x16, #1, 51f\n"
       "ldr d10, [x14], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "tbz x16, #0, 56f\n"
       "ld1 { v10.s }[2], [x14]\n"
-      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v16.s }[2], [x23]\n"
       "b 56f\n"
       "51:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x16, #0, 56f\n"
       "ldr s10, [x14, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
       "b 56f\n"
       "52:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x16, #2, 54f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "tbz x16, #1, 53f\n"
       "ldr d9, [x14], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "tbz x16, #0, 56f\n"
       "ld1 { v9.s }[2], [x14]\n"
-      "ld1 { v15.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x23]\n"
       "b 56f\n"
       "53:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x16, #0, 56f\n"
       "ldr s9, [x14, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
       "b 56f\n"
       "54:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x16, #1, 55f\n"
       "ldr d8, [x14], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "tbz x16, #0, 56f\n"
       "ld1 { v8.s }[2], [x14]\n"
-      "ld1 { v14.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x23]\n"
       "b 56f\n"
       "55:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s8, [x14, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "56:"  // Height 2: Partial accumulate: Done
       "sub x14, x14, x20\n"
       "b 59f\n"
@@ -677,12 +675,12 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "ldr q11, [x14, #0x30]\n"
       "ldr q12, [x14, #0x40]\n"
       "ldr q13, [x14, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
       "b 59f\n"
       "58:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -701,8 +699,8 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "mov x13, #0x0\n"
       "60:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w12, [x20, x13, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 61f\n"
       "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -745,15 +743,15 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v17.4s, v7.4s, v1.s[0]\n"
       "ldr d20, [x15, #0x70]\n"
       "mov v23.d[1], x23\n"
-      "ldr x23, [x15, #0x88]\n"
-      "mov v22.d[1], x22\n"
-      "ldr x22, [x15, #0x98]\n"
-      "mov v21.d[1], x21\n"
       "fmla v12.4s, v23.4s, v0.s[0]\n"
-      "mov v20.d[1], x20\n"
+      "mov v22.d[1], x22\n"
       "fmla v18.4s, v23.4s, v1.s[0]\n"
       "ldr d23, [x15, #0x80]\n"
+      "mov v21.d[1], x21\n"
+      "mov v20.d[1], x20\n"
+      "ldr x23, [x15, #0x88]\n"
       "fmla v13.4s, v22.4s, v0.s[0]\n"
+      "ldr x22, [x15, #0x98]\n"
       "fmla v19.4s, v22.4s, v1.s[0]\n"
       "ldr d22, [x15, #0x90]\n"
       "fmla v8.4s, v21.4s, v0.s[1]\n"
@@ -765,15 +763,15 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v15.4s, v20.4s, v1.s[1]\n"
       "ldr d20, [x15, #0xb0]\n"
       "mov v23.d[1], x23\n"
-      "ldr x23, [x15, #0xc8]\n"
-      "mov v22.d[1], x22\n"
-      "ldr x22, [x15, #0xd8]\n"
-      "mov v21.d[1], x21\n"
       "fmla v10.4s, v23.4s, v0.s[1]\n"
-      "mov v20.d[1], x20\n"
+      "mov v22.d[1], x22\n"
       "fmla v16.4s, v23.4s, v1.s[1]\n"
       "ldr d23, [x15, #0xc0]\n"
+      "mov v21.d[1], x21\n"
+      "mov v20.d[1], x20\n"
+      "ldr x23, [x15, #0xc8]\n"
       "fmla v11.4s, v22.4s, v0.s[1]\n"
+      "ldr x22, [x15, #0xd8]\n"
       "fmla v17.4s, v22.4s, v1.s[1]\n"
       "ldr d22, [x15, #0xd0]\n"
       "fmla v12.4s, v21.4s, v0.s[1]\n"
@@ -785,15 +783,15 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v19.4s, v20.4s, v1.s[1]\n"
       "ldr d20, [x15, #0xf0]\n"
       "mov v23.d[1], x23\n"
-      "ldr x23, [x15, #0x108]\n"
-      "mov v22.d[1], x22\n"
-      "ldr x22, [x15, #0x118]\n"
-      "mov v21.d[1], x21\n"
       "fmla v8.4s, v23.4s, v0.s[2]\n"
-      "mov v20.d[1], x20\n"
+      "mov v22.d[1], x22\n"
       "fmla v14.4s, v23.4s, v1.s[2]\n"
       "ldr d23, [x15, #0x100]\n"
+      "mov v21.d[1], x21\n"
+      "mov v20.d[1], x20\n"
+      "ldr x23, [x15, #0x108]\n"
       "fmla v9.4s, v22.4s, v0.s[2]\n"
+      "ldr x22, [x15, #0x118]\n"
       "fmla v15.4s, v22.4s, v1.s[2]\n"
       "ldr d22, [x15, #0x110]\n"
       "fmla v10.4s, v21.4s, v0.s[2]\n"
@@ -805,15 +803,15 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v17.4s, v20.4s, v1.s[2]\n"
       "ldr d20, [x15, #0x130]\n"
       "mov v23.d[1], x23\n"
-      "ldr x23, [x15, #0x148]\n"
-      "mov v22.d[1], x22\n"
-      "ldr x22, [x15, #0x158]\n"
-      "mov v21.d[1], x21\n"
       "fmla v12.4s, v23.4s, v0.s[2]\n"
-      "mov v20.d[1], x20\n"
+      "mov v22.d[1], x22\n"
       "fmla v18.4s, v23.4s, v1.s[2]\n"
       "ldr d23, [x15, #0x140]\n"
+      "mov v21.d[1], x21\n"
+      "mov v20.d[1], x20\n"
+      "ldr x23, [x15, #0x148]\n"
       "fmla v13.4s, v22.4s, v0.s[2]\n"
+      "ldr x22, [x15, #0x158]\n"
       "fmla v19.4s, v22.4s, v1.s[2]\n"
       "ldr d22, [x15, #0x150]\n"
       "fmla v8.4s, v21.4s, v0.s[3]\n"
@@ -944,19 +942,19 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "sub x12, x12, #0x1\n"
       "ldr s24, [x10], #0x4\n"
       "ldr q21, [x15, #0x0]\n"
-      "ldr q20, [x15, #0x10]\n"
-      "ldr q23, [x15, #0x20]\n"
-      "ldr q22, [x15, #0x30]\n"
       "fmla v8.4s, v21.4s, v25.s[0]\n"
+      "ldr q20, [x15, #0x10]\n"
       "fmla v14.4s, v21.4s, v24.s[0]\n"
-      "ldr q21, [x15, #0x40]\n"
+      "ldr q23, [x15, #0x20]\n"
       "fmla v9.4s, v20.4s, v25.s[0]\n"
+      "ldr q22, [x15, #0x30]\n"
       "fmla v15.4s, v20.4s, v24.s[0]\n"
-      "ldr q20, [x15, #0x50]\n"
+      "ldr q21, [x15, #0x40]\n"
       "fmla v10.4s, v23.4s, v25.s[0]\n"
-      "add x15, x15, #0x60\n"
+      "ldr q20, [x15, #0x50]\n"
       "fmla v16.4s, v23.4s, v24.s[0]\n"
       "fmla v11.4s, v22.4s, v25.s[0]\n"
+      "add x15, x15, #0x60\n"
       "fmla v17.4s, v22.4s, v24.s[0]\n"
       "fmla v12.4s, v21.4s, v25.s[0]\n"
       "fmla v18.4s, v21.4s, v24.s[0]\n"
@@ -969,26 +967,26 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "cmp x13, x20\n"
       "bne 60b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x14, x20, LSL #2\n"
       "prfm pstl1keep, [x14, #0x0]\n"
-      "add x24, x14, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 68f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v20.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v20.4s\n"
+      "fmin v9.4s, v9.4s, v20.4s\n"
+      "fmin v10.4s, v10.4s, v20.4s\n"
+      "fmin v11.4s, v11.4s, v20.4s\n"
+      "fmin v12.4s, v12.4s, v20.4s\n"
+      "fmin v13.4s, v13.4s, v20.4s\n"
+      "fmin v14.4s, v14.4s, v20.4s\n"
+      "fmin v15.4s, v15.4s, v20.4s\n"
+      "fmin v16.4s, v16.4s, v20.4s\n"
+      "fmin v17.4s, v17.4s, v20.4s\n"
+      "fmin v18.4s, v18.4s, v20.4s\n"
+      "fmin v19.4s, v19.4s, v20.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.4s }, [x21]\n"
       "ld1r { v20.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v21.4s\n"
-      "fmin v9.4s, v9.4s, v21.4s\n"
-      "fmin v10.4s, v10.4s, v21.4s\n"
-      "fmin v11.4s, v11.4s, v21.4s\n"
-      "fmin v12.4s, v12.4s, v21.4s\n"
-      "fmin v13.4s, v13.4s, v21.4s\n"
-      "fmin v14.4s, v14.4s, v21.4s\n"
-      "fmin v15.4s, v15.4s, v21.4s\n"
-      "fmin v16.4s, v16.4s, v21.4s\n"
-      "fmin v17.4s, v17.4s, v21.4s\n"
-      "fmin v18.4s, v18.4s, v21.4s\n"
-      "fmin v19.4s, v19.4s, v21.4s\n"
       "fmax v8.4s, v8.4s, v20.4s\n"
       "fmax v9.4s, v9.4s, v20.4s\n"
       "fmax v10.4s, v10.4s, v20.4s\n"
@@ -1009,99 +1007,99 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "st1 { v9.4s }, [x14], #0x10\n"
       "st1 { v10.4s }, [x14], #0x10\n"
       "st1 { v11.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
       "tbz x16, #2, 70f\n"
       "st1 { v12.4s }, [x14], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
       "tbz x16, #1, 69f\n"
       "str d13, [x14], #0x8\n"
-      "str d19, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
       "tbz x16, #0, 80f\n"
       "st1 { v13.s }[2], [x14]\n"
-      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
       "b 80f\n"
       "69:"  // Height 2: Partial direct writeback: partial_1_20
       "tbz x16, #0, 80f\n"
       "str s13, [x14, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
       "b 80f\n"
       "70:"  // Height 2: Partial direct writeback: partial_2_16
       "tbz x16, #1, 71f\n"
       "str d12, [x14], #0x8\n"
-      "str d18, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
       "tbz x16, #0, 80f\n"
       "st1 { v12.s }[2], [x14]\n"
-      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
       "b 80f\n"
       "71:"  // Height 2: Partial direct writeback: partial_1_16
       "tbz x16, #0, 80f\n"
       "str s12, [x14, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
       "b 80f\n"
       "72:"  // Height 2: Partial direct writeback: partial_8_0
       "tbz x16, #3, 76f\n"
       "st1 { v8.4s }, [x14], #0x10\n"
       "st1 { v9.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
       "tbz x16, #2, 74f\n"
       "st1 { v10.4s }, [x14], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
       "tbz x16, #1, 73f\n"
       "str d11, [x14], #0x8\n"
-      "str d17, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
       "tbz x16, #0, 80f\n"
       "st1 { v11.s }[2], [x14]\n"
-      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
       "b 80f\n"
       "73:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x16, #0, 80f\n"
       "str s11, [x14, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
       "b 80f\n"
       "74:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x16, #1, 75f\n"
       "str d10, [x14], #0x8\n"
-      "str d16, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
       "tbz x16, #0, 80f\n"
       "st1 { v10.s }[2], [x14]\n"
-      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
       "b 80f\n"
       "75:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x16, #0, 80f\n"
       "str s10, [x14, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
       "b 80f\n"
       "76:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x16, #2, 78f\n"
       "st1 { v8.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
       "tbz x16, #1, 77f\n"
       "str d9, [x14], #0x8\n"
-      "str d15, [x24], #0x8\n"
+      "str d15, [x23], #0x8\n"
       "tbz x16, #0, 80f\n"
       "st1 { v9.s }[2], [x14]\n"
-      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v15.s }[2], [x23]\n"
       "b 80f\n"
       "77:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x16, #0, 80f\n"
       "str s9, [x14, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
+      "str s15, [x23, #0x0]\n"
       "b 80f\n"
       "78:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x16, #1, 79f\n"
       "str d8, [x14], #0x8\n"
-      "str d14, [x24], #0x8\n"
+      "str d14, [x23], #0x8\n"
       "tbz x16, #0, 80f\n"
       "st1 { v8.s }[2], [x14]\n"
-      "st1 { v14.s }[2], [x24]\n"
+      "st1 { v14.s }[2], [x23]\n"
       "b 80f\n"
       "79:"  // Height 2: Partial direct writeback: partial_1_0
       "str s8, [x14, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
+      "str s14, [x23, #0x0]\n"
       "80:"  // Height 2: Partial direct writeback: Done
       "b 82f\n"
       "81:"  // Height 2: Full writeback
@@ -1112,37 +1110,37 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "str q12, [x14, #0x40]\n"
       "str q13, [x14, #0x50]\n"
       "add x14, x14, #0x60\n"
-      "str q14, [x24, #0x0]\n"
-      "str q15, [x24, #0x10]\n"
-      "str q16, [x24, #0x20]\n"
-      "str q17, [x24, #0x30]\n"
-      "str q18, [x24, #0x40]\n"
-      "str q19, [x24, #0x50]\n"
+      "str q14, [x23, #0x0]\n"
+      "str q15, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "str q17, [x23, #0x30]\n"
+      "str q18, [x23, #0x40]\n"
+      "str q19, [x23, #0x50]\n"
       "82:"  // Height 2: Writeback done
       "subs x16, x16, #0x18\n"
       "bgt 43b\n"
       "b 166f\n"
       "83:"  // Height 3
-      "ldr x17, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x17, %x[bias]\n"
       "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x14, %x[output_ptr]\n"
       "84:"  // Height 3: Column loop
       "cbz x17, 85f\n"
       "ldr q8, [x17, #0x0]\n"
-      "ldr q9, [x17, #0x10]\n"
-      "ldr q10, [x17, #0x20]\n"
-      "ldr q11, [x17, #0x30]\n"
       "mov v14.16b, v8.16b\n"
-      "ldr q12, [x17, #0x40]\n"
+      "ldr q9, [x17, #0x10]\n"
       "mov v15.16b, v9.16b\n"
-      "ldr q13, [x17, #0x50]\n"
+      "ldr q10, [x17, #0x20]\n"
       "mov v16.16b, v10.16b\n"
+      "ldr q11, [x17, #0x30]\n"
       "mov v17.16b, v11.16b\n"
-      "add x17, x17, #0x60\n"
+      "ldr q12, [x17, #0x40]\n"
       "mov v18.16b, v12.16b\n"
+      "ldr q13, [x17, #0x50]\n"
       "mov v19.16b, v13.16b\n"
       "mov v20.16b, v8.16b\n"
+      "add x17, x17, #0x60\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1152,147 +1150,147 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "85:"  // Height 3: no bias
       "tbz %x[flags], #0, 99f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x14, x20, LSL #2\n"
       "cmp x16, #0x18\n"
-      "add x24, x14, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 98f\n"
       "tbz x16, #4, 89f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v9.4s }, [x14], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x14], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
       "ld1 { v11.4s }, [x14], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
       "tbz x16, #2, 87f\n"
       "ld1 { v12.4s }, [x14], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "tbz x16, #1, 86f\n"
       "ldr d13, [x14], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d19, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x16, #0, 97f\n"
       "ld1 { v13.s }[2], [x14]\n"
-      "ld1 { v19.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v19.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 97f\n"
       "86:"  // Height 3: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x16, #0, 97f\n"
       "ldr s13, [x14, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "b 97f\n"
       "87:"  // Height 3: Partial accumulate: partial_2_16
       "tbz x16, #1, 88f\n"
       "ldr d12, [x14], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d18, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x16, #0, 97f\n"
       "ld1 { v12.s }[2], [x14]\n"
-      "ld1 { v18.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v18.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
       "b 97f\n"
       "88:"  // Height 3: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x16, #0, 97f\n"
       "ldr s12, [x14, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
       "b 97f\n"
       "89:"  // Height 3: Partial accumulate: partial_8_0
       "tbz x16, #3, 93f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v9.4s }, [x14], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "tbz x16, #2, 91f\n"
       "ld1 { v10.4s }, [x14], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
       "tbz x16, #1, 90f\n"
       "ldr d11, [x14], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d17, [x24], #0x8\n"
-      "ldr d23, [x23], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
+      "ldr d23, [x22], #0x8\n"
       "tbz x16, #0, 97f\n"
       "ld1 { v11.s }[2], [x14]\n"
-      "ld1 { v17.s }[2], [x24]\n"
-      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v17.s }[2], [x23]\n"
+      "ld1 { v23.s }[2], [x22]\n"
       "b 97f\n"
       "90:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x16, #0, 97f\n"
       "ldr s11, [x14, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
-      "ldr s23, [x23, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
+      "ldr s23, [x22, #0x0]\n"
       "b 97f\n"
       "91:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x16, #1, 92f\n"
       "ldr d10, [x14], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d16, [x24], #0x8\n"
-      "ldr d22, [x23], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
       "tbz x16, #0, 97f\n"
       "ld1 { v10.s }[2], [x14]\n"
-      "ld1 { v16.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v16.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
       "b 97f\n"
       "92:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x16, #0, 97f\n"
       "ldr s10, [x14, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
       "b 97f\n"
       "93:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x16, #2, 95f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
       "tbz x16, #1, 94f\n"
       "ldr d9, [x14], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d15, [x24], #0x8\n"
-      "ldr d21, [x23], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
+      "ldr d21, [x22], #0x8\n"
       "tbz x16, #0, 97f\n"
       "ld1 { v9.s }[2], [x14]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x23]\n"
+      "ld1 { v21.s }[2], [x22]\n"
       "b 97f\n"
       "94:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x16, #0, 97f\n"
       "ldr s9, [x14, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s21, [x23, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
+      "ldr s21, [x22, #0x0]\n"
       "b 97f\n"
       "95:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x16, #1, 96f\n"
       "ldr d8, [x14], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "ldr d20, [x23], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
+      "ldr d20, [x22], #0x8\n"
       "tbz x16, #0, 97f\n"
       "ld1 { v8.s }[2], [x14]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x23]\n"
+      "ld1 { v20.s }[2], [x22]\n"
       "b 97f\n"
       "96:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s8, [x14, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s20, [x23, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
+      "ldr s20, [x22, #0x0]\n"
       "97:"  // Height 3: Partial accumulate: Done
       "sub x14, x14, x20\n"
       "b 100f\n"
@@ -1303,18 +1301,18 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "ldr q11, [x14, #0x30]\n"
       "ldr q12, [x14, #0x40]\n"
       "ldr q13, [x14, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
-      "ldr q20, [x23, #0x0]\n"
-      "ldr q21, [x23, #0x10]\n"
-      "ldr q22, [x23, #0x20]\n"
-      "ldr q23, [x23, #0x30]\n"
-      "ldr q24, [x23, #0x40]\n"
-      "ldr q25, [x23, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
+      "ldr q20, [x22, #0x0]\n"
+      "ldr q21, [x22, #0x10]\n"
+      "ldr q22, [x22, #0x20]\n"
+      "ldr q23, [x22, #0x30]\n"
+      "ldr q24, [x22, #0x40]\n"
+      "ldr q25, [x22, #0x50]\n"
       "b 100f\n"
       "99:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1339,8 +1337,8 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "mov x13, #0x0\n"
       "101:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w12, [x20, x13, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 102f\n"
       "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1385,23 +1383,23 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v10.4s, v6.4s, v0.s[0]\n"
       "mov v29.d[1], x23\n"
       "fmla v16.4s, v6.4s, v1.s[0]\n"
-      "ldr x23, [x15, #0x88]\n"
+      "mov v28.d[1], x22\n"
       "fmla v22.4s, v6.4s, v2.s[0]\n"
       "ldr d27, [x15, #0x60]\n"
       "fmla v11.4s, v7.4s, v0.s[0]\n"
-      "mov v28.d[1], x22\n"
+      "mov v27.d[1], x21\n"
       "fmla v17.4s, v7.4s, v1.s[0]\n"
-      "ldr x22, [x15, #0x98]\n"
+      "ldr x23, [x15, #0x88]\n"
       "fmla v23.4s, v7.4s, v2.s[0]\n"
       "ldr d26, [x15, #0x70]\n"
-      "mov v27.d[1], x21\n"
+      "mov v26.d[1], x20\n"
       "fmla v12.4s, v29.4s, v0.s[0]\n"
       "fmla v18.4s, v29.4s, v1.s[0]\n"
-      "ldr x21, [x15, #0xa8]\n"
-      "mov v26.d[1], x20\n"
+      "ldr x22, [x15, #0x98]\n"
       "fmla v24.4s, v29.4s, v2.s[0]\n"
       "ldr d29, [x15, #0x80]\n"
       "fmla v13.4s, v28.4s, v0.s[0]\n"
+      "ldr x21, [x15, #0xa8]\n"
       "fmla v19.4s, v28.4s, v1.s[0]\n"
       "ldr x20, [x15, #0xb8]\n"
       "fmla v25.4s, v28.4s, v2.s[0]\n"
@@ -1409,23 +1407,23 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v8.4s, v27.4s, v0.s[1]\n"
       "mov v29.d[1], x23\n"
       "fmla v14.4s, v27.4s, v1.s[1]\n"
-      "ldr x23, [x15, #0xc8]\n"
+      "mov v28.d[1], x22\n"
       "fmla v20.4s, v27.4s, v2.s[1]\n"
       "ldr d27, [x15, #0xa0]\n"
       "fmla v9.4s, v26.4s, v0.s[1]\n"
-      "mov v28.d[1], x22\n"
+      "mov v27.d[1], x21\n"
       "fmla v15.4s, v26.4s, v1.s[1]\n"
-      "ldr x22, [x15, #0xd8]\n"
+      "ldr x23, [x15, #0xc8]\n"
       "fmla v21.4s, v26.4s, v2.s[1]\n"
       "ldr d26, [x15, #0xb0]\n"
-      "mov v27.d[1], x21\n"
+      "mov v26.d[1], x20\n"
       "fmla v10.4s, v29.4s, v0.s[1]\n"
       "fmla v16.4s, v29.4s, v1.s[1]\n"
-      "ldr x21, [x15, #0xe8]\n"
-      "mov v26.d[1], x20\n"
+      "ldr x22, [x15, #0xd8]\n"
       "fmla v22.4s, v29.4s, v2.s[1]\n"
       "ldr d29, [x15, #0xc0]\n"
       "fmla v11.4s, v28.4s, v0.s[1]\n"
+      "ldr x21, [x15, #0xe8]\n"
       "fmla v17.4s, v28.4s, v1.s[1]\n"
       "ldr x20, [x15, #0xf8]\n"
       "fmla v23.4s, v28.4s, v2.s[1]\n"
@@ -1433,23 +1431,23 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v12.4s, v27.4s, v0.s[1]\n"
       "mov v29.d[1], x23\n"
       "fmla v18.4s, v27.4s, v1.s[1]\n"
-      "ldr x23, [x15, #0x108]\n"
+      "mov v28.d[1], x22\n"
       "fmla v24.4s, v27.4s, v2.s[1]\n"
       "ldr d27, [x15, #0xe0]\n"
       "fmla v13.4s, v26.4s, v0.s[1]\n"
-      "mov v28.d[1], x22\n"
+      "mov v27.d[1], x21\n"
       "fmla v19.4s, v26.4s, v1.s[1]\n"
-      "ldr x22, [x15, #0x118]\n"
+      "ldr x23, [x15, #0x108]\n"
       "fmla v25.4s, v26.4s, v2.s[1]\n"
       "ldr d26, [x15, #0xf0]\n"
-      "mov v27.d[1], x21\n"
+      "mov v26.d[1], x20\n"
       "fmla v8.4s, v29.4s, v0.s[2]\n"
       "fmla v14.4s, v29.4s, v1.s[2]\n"
-      "ldr x21, [x15, #0x128]\n"
-      "mov v26.d[1], x20\n"
+      "ldr x22, [x15, #0x118]\n"
       "fmla v20.4s, v29.4s, v2.s[2]\n"
       "ldr d29, [x15, #0x100]\n"
       "fmla v9.4s, v28.4s, v0.s[2]\n"
+      "ldr x21, [x15, #0x128]\n"
       "fmla v15.4s, v28.4s, v1.s[2]\n"
       "ldr x20, [x15, #0x138]\n"
       "fmla v21.4s, v28.4s, v2.s[2]\n"
@@ -1457,23 +1455,23 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v10.4s, v27.4s, v0.s[2]\n"
       "mov v29.d[1], x23\n"
       "fmla v16.4s, v27.4s, v1.s[2]\n"
-      "ldr x23, [x15, #0x148]\n"
+      "mov v28.d[1], x22\n"
       "fmla v22.4s, v27.4s, v2.s[2]\n"
       "ldr d27, [x15, #0x120]\n"
       "fmla v11.4s, v26.4s, v0.s[2]\n"
-      "mov v28.d[1], x22\n"
+      "mov v27.d[1], x21\n"
       "fmla v17.4s, v26.4s, v1.s[2]\n"
-      "ldr x22, [x15, #0x158]\n"
+      "ldr x23, [x15, #0x148]\n"
       "fmla v23.4s, v26.4s, v2.s[2]\n"
       "ldr d26, [x15, #0x130]\n"
-      "mov v27.d[1], x21\n"
+      "mov v26.d[1], x20\n"
       "fmla v12.4s, v29.4s, v0.s[2]\n"
       "fmla v18.4s, v29.4s, v1.s[2]\n"
-      "ldr x21, [x15, #0x168]\n"
-      "mov v26.d[1], x20\n"
+      "ldr x22, [x15, #0x158]\n"
       "fmla v24.4s, v29.4s, v2.s[2]\n"
       "ldr d29, [x15, #0x140]\n"
       "fmla v13.4s, v28.4s, v0.s[2]\n"
+      "ldr x21, [x15, #0x168]\n"
       "fmla v19.4s, v28.4s, v1.s[2]\n"
       "ldr x20, [x15, #0x178]\n"
       "fmla v25.4s, v28.4s, v2.s[2]\n"
@@ -1481,54 +1479,54 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v8.4s, v27.4s, v0.s[3]\n"
       "mov v29.d[1], x23\n"
       "fmla v14.4s, v27.4s, v1.s[3]\n"
-      "add x11, x11, #0x10\n"
+      "mov v28.d[1], x22\n"
       "fmla v20.4s, v27.4s, v2.s[3]\n"
       "ldr d27, [x15, #0x160]\n"
       "fmla v9.4s, v26.4s, v0.s[3]\n"
-      "mov v28.d[1], x22\n"
+      "mov v27.d[1], x21\n"
       "fmla v15.4s, v26.4s, v1.s[3]\n"
-      "add x10, x10, #0x10\n"
+      "add x11, x11, #0x10\n"
       "fmla v21.4s, v26.4s, v2.s[3]\n"
       "ldr d26, [x15, #0x170]\n"
-      "mov v27.d[1], x21\n"
+      "mov v26.d[1], x20\n"
+      "add x10, x10, #0x10\n"
       "add x9, x9, #0x10\n"
       "add x15, x15, #0x180\n"
       "fmla v10.4s, v29.4s, v0.s[3]\n"
-      "mov v26.d[1], x20\n"
+      "ldr x26, [x15, #0x8]\n"
       "fmla v16.4s, v29.4s, v1.s[3]\n"
+      "ldr x25, [x15, #0x18]\n"
       "fmla v22.4s, v29.4s, v2.s[3]\n"
       "ldr d4, [x15, #0x0]\n"
-      "ldr x20, [x15, #0x8]\n"
       "fmla v11.4s, v28.4s, v0.s[3]\n"
+      "ldr x24, [x15, #0x28]\n"
       "fmla v17.4s, v28.4s, v1.s[3]\n"
-      "ldr x25, [x15, #0x18]\n"
+      "ldr x23, [x11, #0x8]\n"
       "fmla v23.4s, v28.4s, v2.s[3]\n"
       "ldr d5, [x15, #0x10]\n"
       "fmla v12.4s, v27.4s, v0.s[3]\n"
-      "ldr x24, [x15, #0x28]\n"
+      "ldr x22, [x10, #0x8]\n"
       "fmla v18.4s, v27.4s, v1.s[3]\n"
-      "ldr x23, [x11, #0x8]\n"
+      "ldr x21, [x9, #0x8]\n"
       "fmla v24.4s, v27.4s, v2.s[3]\n"
       "ldr d6, [x15, #0x20]\n"
       "fmla v13.4s, v26.4s, v0.s[3]\n"
       "ldr d0, [x11, #0x0]\n"
       "fmla v19.4s, v26.4s, v1.s[3]\n"
       "ldr d1, [x10, #0x0]\n"
-      "ldr x22, [x10, #0x8]\n"
       "fmla v25.4s, v26.4s, v2.s[3]\n"
       "ldr d2, [x9, #0x0]\n"
-      "sub x12, x12, #0x4\n"
       "ldr d7, [x15, #0x30]\n"
-      "cmp x12, #0x8\n"
-      "ldr x21, [x9, #0x8]\n"
-      "mov v4.d[1], x20\n"
+      "sub x12, x12, #0x4\n"
       "ldr x20, [x15, #0x38]\n"
-      "mov v5.d[1], x25\n"
+      "cmp x12, #0x8\n"
       "prfm pldl1keep, [x11, #0x80]\n"
-      "mov v6.d[1], x24\n"
+      "mov v4.d[1], x26\n"
       "prfm pldl1keep, [x10, #0x80]\n"
-      "mov v0.d[1], x23\n"
+      "mov v5.d[1], x25\n"
       "prfm pldl1keep, [x9, #0x80]\n"
+      "mov v6.d[1], x24\n"
+      "mov v0.d[1], x23\n"
       "mov v1.d[1], x22\n"
       "mov v2.d[1], x21\n"
       "mov v7.d[1], x20\n"
@@ -1642,14 +1640,14 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "ldr s31, [x10], #0x4\n"
       "ldr s30, [x9], #0x4\n"
       "ldr q27, [x15, #0x0]\n"
-      "ldr q26, [x15, #0x10]\n"
-      "ldr q29, [x15, #0x20]\n"
-      "ldr q28, [x15, #0x30]\n"
       "fmla v8.4s, v27.4s, v0.s[0]\n"
+      "ldr q26, [x15, #0x10]\n"
       "fmla v14.4s, v27.4s, v31.s[0]\n"
+      "ldr q29, [x15, #0x20]\n"
       "fmla v20.4s, v27.4s, v30.s[0]\n"
-      "ldr q27, [x15, #0x40]\n"
+      "ldr q28, [x15, #0x30]\n"
       "fmla v9.4s, v26.4s, v0.s[0]\n"
+      "ldr q27, [x15, #0x40]\n"
       "fmla v15.4s, v26.4s, v31.s[0]\n"
       "fmla v21.4s, v26.4s, v30.s[0]\n"
       "ldr q26, [x15, #0x50]\n"
@@ -1673,34 +1671,34 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "cmp x13, x20\n"
       "bne 101b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x14, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "prfm pstl1keep, [x14, #0x0]\n"
-      "add x24, x14, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 109f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v26.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v26.4s\n"
+      "fmin v9.4s, v9.4s, v26.4s\n"
+      "fmin v10.4s, v10.4s, v26.4s\n"
+      "fmin v11.4s, v11.4s, v26.4s\n"
+      "fmin v12.4s, v12.4s, v26.4s\n"
+      "fmin v13.4s, v13.4s, v26.4s\n"
+      "fmin v14.4s, v14.4s, v26.4s\n"
+      "fmin v15.4s, v15.4s, v26.4s\n"
+      "fmin v16.4s, v16.4s, v26.4s\n"
+      "fmin v17.4s, v17.4s, v26.4s\n"
+      "fmin v18.4s, v18.4s, v26.4s\n"
+      "fmin v19.4s, v19.4s, v26.4s\n"
+      "fmin v20.4s, v20.4s, v26.4s\n"
+      "fmin v21.4s, v21.4s, v26.4s\n"
+      "fmin v22.4s, v22.4s, v26.4s\n"
+      "fmin v23.4s, v23.4s, v26.4s\n"
+      "fmin v24.4s, v24.4s, v26.4s\n"
+      "fmin v25.4s, v25.4s, v26.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v27.4s }, [x21]\n"
       "ld1r { v26.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v27.4s\n"
-      "fmin v9.4s, v9.4s, v27.4s\n"
-      "fmin v10.4s, v10.4s, v27.4s\n"
-      "fmin v11.4s, v11.4s, v27.4s\n"
-      "fmin v12.4s, v12.4s, v27.4s\n"
-      "fmin v13.4s, v13.4s, v27.4s\n"
-      "fmin v14.4s, v14.4s, v27.4s\n"
-      "fmin v15.4s, v15.4s, v27.4s\n"
-      "fmin v16.4s, v16.4s, v27.4s\n"
-      "fmin v17.4s, v17.4s, v27.4s\n"
-      "fmin v18.4s, v18.4s, v27.4s\n"
-      "fmin v19.4s, v19.4s, v27.4s\n"
-      "fmin v20.4s, v20.4s, v27.4s\n"
-      "fmin v21.4s, v21.4s, v27.4s\n"
-      "fmin v22.4s, v22.4s, v27.4s\n"
-      "fmin v23.4s, v23.4s, v27.4s\n"
-      "fmin v24.4s, v24.4s, v27.4s\n"
-      "fmin v25.4s, v25.4s, v27.4s\n"
       "fmax v8.4s, v8.4s, v26.4s\n"
       "fmax v9.4s, v9.4s, v26.4s\n"
       "fmax v10.4s, v10.4s, v26.4s\n"
@@ -1727,126 +1725,126 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "st1 { v9.4s }, [x14], #0x10\n"
       "st1 { v10.4s }, [x14], #0x10\n"
       "st1 { v11.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
       "tbz x16, #2, 111f\n"
       "st1 { v12.4s }, [x14], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x16, #1, 110f\n"
       "str d13, [x14], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x16, #0, 121f\n"
       "st1 { v13.s }[2], [x14]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 121f\n"
       "110:"  // Height 3: Partial direct writeback: partial_1_20
       "tbz x16, #0, 121f\n"
       "str s13, [x14, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 121f\n"
       "111:"  // Height 3: Partial direct writeback: partial_2_16
       "tbz x16, #1, 112f\n"
       "str d12, [x14], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x16, #0, 121f\n"
       "st1 { v12.s }[2], [x14]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 121f\n"
       "112:"  // Height 3: Partial direct writeback: partial_1_16
       "tbz x16, #0, 121f\n"
       "str s12, [x14, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "b 121f\n"
       "113:"  // Height 3: Partial direct writeback: partial_8_0
       "tbz x16, #3, 117f\n"
       "st1 { v8.4s }, [x14], #0x10\n"
       "st1 { v9.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
       "tbz x16, #2, 115f\n"
       "st1 { v10.4s }, [x14], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
       "tbz x16, #1, 114f\n"
       "str d11, [x14], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
       "tbz x16, #0, 121f\n"
       "st1 { v11.s }[2], [x14]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
       "b 121f\n"
       "114:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x16, #0, 121f\n"
       "str s11, [x14, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
       "b 121f\n"
       "115:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x16, #1, 116f\n"
       "str d10, [x14], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d22, [x22], #0x8\n"
       "tbz x16, #0, 121f\n"
       "st1 { v10.s }[2], [x14]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v22.s }[2], [x22]\n"
       "b 121f\n"
       "116:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x16, #0, 121f\n"
       "str s10, [x14, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s22, [x22, #0x0]\n"
       "b 121f\n"
       "117:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x16, #2, 119f\n"
       "st1 { v8.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
       "tbz x16, #1, 118f\n"
       "str d9, [x14], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
+      "str d15, [x23], #0x8\n"
+      "str d21, [x22], #0x8\n"
       "tbz x16, #0, 121f\n"
       "st1 { v9.s }[2], [x14]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v15.s }[2], [x23]\n"
+      "st1 { v21.s }[2], [x22]\n"
       "b 121f\n"
       "118:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x16, #0, 121f\n"
       "str s9, [x14, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
+      "str s15, [x23, #0x0]\n"
+      "str s21, [x22, #0x0]\n"
       "b 121f\n"
       "119:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x16, #1, 120f\n"
       "str d8, [x14], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
+      "str d14, [x23], #0x8\n"
+      "str d20, [x22], #0x8\n"
       "tbz x16, #0, 121f\n"
       "st1 { v8.s }[2], [x14]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v14.s }[2], [x23]\n"
+      "st1 { v20.s }[2], [x22]\n"
       "b 121f\n"
       "120:"  // Height 3: Partial direct writeback: partial_1_0
       "str s8, [x14, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
+      "str s14, [x23, #0x0]\n"
+      "str s20, [x22, #0x0]\n"
       "121:"  // Height 3: Partial direct writeback: Done
       "b 123f\n"
       "122:"  // Height 3: Full writeback
@@ -1857,18 +1855,18 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "str q12, [x14, #0x40]\n"
       "str q13, [x14, #0x50]\n"
       "add x14, x14, #0x60\n"
-      "str q14, [x24, #0x0]\n"
-      "str q15, [x24, #0x10]\n"
-      "str q16, [x24, #0x20]\n"
-      "str q17, [x24, #0x30]\n"
-      "str q18, [x24, #0x40]\n"
-      "str q19, [x24, #0x50]\n"
-      "str q20, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q22, [x23, #0x20]\n"
-      "str q23, [x23, #0x30]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q25, [x23, #0x50]\n"
+      "str q14, [x23, #0x0]\n"
+      "str q15, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "str q17, [x23, #0x30]\n"
+      "str q18, [x23, #0x40]\n"
+      "str q19, [x23, #0x50]\n"
+      "str q20, [x22, #0x0]\n"
+      "str q21, [x22, #0x10]\n"
+      "str q22, [x22, #0x20]\n"
+      "str q23, [x22, #0x30]\n"
+      "str q24, [x22, #0x40]\n"
+      "str q25, [x22, #0x50]\n"
       "123:"  // Height 3: Writeback done
       "subs x16, x16, #0x18\n"
       "bgt 84b\n"
@@ -1876,28 +1874,27 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "124:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "mov x20, #0x10\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x17, %x[bias]\n"
       "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
-      "madd x20, x21, x20, x14\n"
       "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x14, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "125:"  // Height 4: Column loop
       "cbz x17, 126f\n"
       "ldr q8, [x17, #0x0]\n"
-      "ldr q9, [x17, #0x10]\n"
-      "ldr q10, [x17, #0x20]\n"
-      "ldr q11, [x17, #0x30]\n"
       "mov v14.16b, v8.16b\n"
-      "ldr q12, [x17, #0x40]\n"
+      "ldr q9, [x17, #0x10]\n"
       "mov v15.16b, v9.16b\n"
-      "ldr q13, [x17, #0x50]\n"
+      "ldr q10, [x17, #0x20]\n"
       "mov v16.16b, v10.16b\n"
+      "ldr q11, [x17, #0x30]\n"
       "mov v17.16b, v11.16b\n"
-      "add x17, x17, #0x60\n"
+      "ldr q12, [x17, #0x40]\n"
       "mov v18.16b, v12.16b\n"
+      "ldr q13, [x17, #0x50]\n"
       "mov v19.16b, v13.16b\n"
       "mov v20.16b, v8.16b\n"
+      "add x17, x17, #0x60\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1913,175 +1910,175 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "126:"  // Height 4: no bias
       "tbz %x[flags], #0, 140f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x16, #0x18\n"
-      "add x24, x14, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x23, x14, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x16, #0x18\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 139f\n"
       "tbz x16, #4, 130f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v9.4s }, [x14], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v27.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x14], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v11.4s }, [x14], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x16, #2, 128f\n"
       "ld1 { v12.4s }, [x14], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x16, #1, 127f\n"
       "ldr d13, [x14], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d19, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x16, #0, 138f\n"
       "ld1 { v13.s }[2], [x14]\n"
-      "ld1 { v19.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v19.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 138f\n"
       "127:"  // Height 4: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x16, #0, 138f\n"
       "ldr s13, [x14, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 138f\n"
       "128:"  // Height 4: Partial accumulate: partial_2_16
       "tbz x16, #1, 129f\n"
       "ldr d12, [x14], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d18, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x16, #0, 138f\n"
       "ld1 { v12.s }[2], [x14]\n"
-      "ld1 { v18.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v18.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 138f\n"
       "129:"  // Height 4: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x16, #0, 138f\n"
       "ldr s12, [x14, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 138f\n"
       "130:"  // Height 4: Partial accumulate: partial_8_0
       "tbz x16, #3, 134f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v9.4s }, [x14], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v27.4s }, [x21], #0x10\n"
       "tbz x16, #2, 132f\n"
       "ld1 { v10.4s }, [x14], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x16, #1, 131f\n"
       "ldr d11, [x14], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d17, [x24], #0x8\n"
-      "ldr d23, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
+      "ldr d23, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x16, #0, 138f\n"
       "ld1 { v11.s }[2], [x14]\n"
-      "ld1 { v17.s }[2], [x24]\n"
-      "ld1 { v23.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v17.s }[2], [x23]\n"
+      "ld1 { v23.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 138f\n"
       "131:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x16, #0, 138f\n"
       "ldr s11, [x14, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
-      "ldr s23, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
+      "ldr s23, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 138f\n"
       "132:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x16, #1, 133f\n"
       "ldr d10, [x14], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d16, [x24], #0x8\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x16, #0, 138f\n"
       "ld1 { v10.s }[2], [x14]\n"
-      "ld1 { v16.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v16.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 138f\n"
       "133:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x16, #0, 138f\n"
       "ldr s10, [x14, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "b 138f\n"
       "134:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x16, #2, 136f\n"
       "ld1 { v8.4s }, [x14], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "tbz x16, #1, 135f\n"
       "ldr d9, [x14], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d15, [x24], #0x8\n"
-      "ldr d21, [x23], #0x8\n"
-      "ldr d27, [x22], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
+      "ldr d21, [x22], #0x8\n"
+      "ldr d27, [x21], #0x8\n"
       "tbz x16, #0, 138f\n"
       "ld1 { v9.s }[2], [x14]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v21.s }[2], [x23]\n"
-      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x23]\n"
+      "ld1 { v21.s }[2], [x22]\n"
+      "ld1 { v27.s }[2], [x21]\n"
       "b 138f\n"
       "135:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x16, #0, 138f\n"
       "ldr s9, [x14, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s21, [x23, #0x0]\n"
-      "ldr s27, [x22, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
+      "ldr s21, [x22, #0x0]\n"
+      "ldr s27, [x21, #0x0]\n"
       "b 138f\n"
       "136:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x16, #1, 137f\n"
       "ldr d8, [x14], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "ldr d20, [x23], #0x8\n"
-      "ldr d26, [x22], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
+      "ldr d20, [x22], #0x8\n"
+      "ldr d26, [x21], #0x8\n"
       "tbz x16, #0, 138f\n"
       "ld1 { v8.s }[2], [x14]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v20.s }[2], [x23]\n"
-      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x23]\n"
+      "ld1 { v20.s }[2], [x22]\n"
+      "ld1 { v26.s }[2], [x21]\n"
       "b 138f\n"
       "137:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s8, [x14, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s20, [x23, #0x0]\n"
-      "ldr s26, [x22, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
+      "ldr s20, [x22, #0x0]\n"
+      "ldr s26, [x21, #0x0]\n"
       "138:"  // Height 4: Partial accumulate: Done
       "sub x14, x14, x20\n"
       "b 141f\n"
@@ -2092,24 +2089,24 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "ldr q11, [x14, #0x30]\n"
       "ldr q12, [x14, #0x40]\n"
       "ldr q13, [x14, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
-      "ldr q20, [x23, #0x0]\n"
-      "ldr q21, [x23, #0x10]\n"
-      "ldr q22, [x23, #0x20]\n"
-      "ldr q23, [x23, #0x30]\n"
-      "ldr q24, [x23, #0x40]\n"
-      "ldr q25, [x23, #0x50]\n"
-      "ldr q26, [x22, #0x0]\n"
-      "ldr q27, [x22, #0x10]\n"
-      "ldr q28, [x22, #0x20]\n"
-      "ldr q29, [x22, #0x30]\n"
-      "ldr q30, [x22, #0x40]\n"
-      "ldr q31, [x22, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
+      "ldr q20, [x22, #0x0]\n"
+      "ldr q21, [x22, #0x10]\n"
+      "ldr q22, [x22, #0x20]\n"
+      "ldr q23, [x22, #0x30]\n"
+      "ldr q24, [x22, #0x40]\n"
+      "ldr q25, [x22, #0x50]\n"
+      "ldr q26, [x21, #0x0]\n"
+      "ldr q27, [x21, #0x10]\n"
+      "ldr q28, [x21, #0x20]\n"
+      "ldr q29, [x21, #0x30]\n"
+      "ldr q30, [x21, #0x40]\n"
+      "ldr q31, [x21, #0x50]\n"
       "b 141f\n"
       "140:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2140,8 +2137,8 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "mov x13, #0x0\n"
       "142:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w12, [x20, x13, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 143f\n"
       "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2186,169 +2183,169 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "fmla v9.4s, v5.4s, v0.s[0]\n"
       "ldr x20, [x15, #0x78]\n"
       "fmla v15.4s, v5.4s, v1.s[0]\n"
-      "add x11, x11, #0x10\n"
-      "fmla v21.4s, v5.4s, v2.s[0]\n"
       "mov v4.d[1], x23\n"
+      "fmla v21.4s, v5.4s, v2.s[0]\n"
+      "ldr x23, [x15, #0x88]\n"
       "fmla v27.4s, v5.4s, v3.s[0]\n"
       "ldr d5, [x15, #0x50]\n"
       "fmla v10.4s, v6.4s, v0.s[0]\n"
-      "ldr x23, [x15, #0x88]\n"
+      "mov v5.d[1], x22\n"
       "fmla v16.4s, v6.4s, v1.s[0]\n"
-      "add x10, x10, #0x10\n"
+      "ldr x22, [x15, #0x98]\n"
       "fmla v22.4s, v6.4s, v2.s[0]\n"
-      "mov v5.d[1], x22\n"
+      "add x11, x11, #0x10\n"
       "fmla v28.4s, v6.4s, v3.s[0]\n"
       "ldr d6, [x15, #0x60]\n"
       "fmla v11.4s, v7.4s, v0.s[0]\n"
-      "ldr x22, [x15, #0x98]\n"
+      "mov v6.d[1], x21\n"
       "fmla v17.4s, v7.4s, v1.s[0]\n"
-      "add x9, x9, #0x10\n"
+      "ldr x21, [x15, #0xa8]\n"
       "fmla v23.4s, v7.4s, v2.s[0]\n"
-      "mov v6.d[1], x21\n"
+      "add x10, x10, #0x10\n"
       "fmla v29.4s, v7.4s, v3.s[0]\n"
       "ldr d7, [x15, #0x70]\n"
+      "mov v7.d[1], x20\n"
       "fmla v12.4s, v4.4s, v0.s[0]\n"
-      "ldr x21, [x15, #0xa8]\n"
       "fmla v18.4s, v4.4s, v1.s[0]\n"
-      "add x28, x28, #0x10\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0xb8]\n"
       "fmla v24.4s, v4.4s, v2.s[0]\n"
+      "add x9, x9, #0x10\n"
       "fmla v30.4s, v4.4s, v3.s[0]\n"
       "ldr d4, [x15, #0x80]\n"
       "fmla v13.4s, v5.4s, v0.s[0]\n"
-      "ldr x20, [x15, #0xb8]\n"
+      "mov v4.d[1], x23\n"
       "fmla v19.4s, v5.4s, v1.s[0]\n"
-      "ldr x27, [x11, #0x8]\n"
+      "ldr x23, [x15, #0xc8]\n"
       "fmla v25.4s, v5.4s, v2.s[0]\n"
-      "mov v4.d[1], x23\n"
+      "add x28, x28, #0x10\n"
       "fmla v31.4s, v5.4s, v3.s[0]\n"
       "ldr d5, [x15, #0x90]\n"
       "fmla v8.4s, v6.4s, v0.s[1]\n"
-      "ldr x23, [x15, #0xc8]\n"
+      "mov v5.d[1], x22\n"
       "fmla v14.4s, v6.4s, v1.s[1]\n"
-      "ldr x26, [x10, #0x8]\n"
+      "ldr x22, [x15, #0xd8]\n"
       "fmla v20.4s, v6.4s, v2.s[1]\n"
-      "mov v5.d[1], x22\n"
+      "ldr x27, [x11, #0x8]\n"
       "fmla v26.4s, v6.4s, v3.s[1]\n"
       "ldr d6, [x15, #0xa0]\n"
       "fmla v9.4s, v7.4s, v0.s[1]\n"
-      "ldr x22, [x15, #0xd8]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.4s, v7.4s, v1.s[1]\n"
-      "ldr x25, [x9, #0x8]\n"
+      "ldr x21, [x15, #0xe8]\n"
       "fmla v21.4s, v7.4s, v2.s[1]\n"
-      "mov v6.d[1], x21\n"
+      "ldr x26, [x10, #0x8]\n"
       "fmla v27.4s, v7.4s, v3.s[1]\n"
       "ldr d7, [x15, #0xb0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.4s, v4.4s, v0.s[1]\n"
-      "ldr x21, [x15, #0xe8]\n"
       "fmla v16.4s, v4.4s, v1.s[1]\n"
-      "ldr x24, [x28, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0xf8]\n"
       "fmla v22.4s, v4.4s, v2.s[1]\n"
+      "ldr x25, [x9, #0x8]\n"
       "fmla v28.4s, v4.4s, v3.s[1]\n"
       "ldr d4, [x15, #0xc0]\n"
       "fmla v11.4s, v5.4s, v0.s[1]\n"
-      "ldr x20, [x15, #0xf8]\n"
+      "mov v4.d[1], x23\n"
       "fmla v17.4s, v5.4s, v1.s[1]\n"
-      "sub x12, x12, #0x4\n"
+      "ldr x23, [x15, #0x108]\n"
       "fmla v23.4s, v5.4s, v2.s[1]\n"
-      "mov v4.d[1], x23\n"
+      "ldr x24, [x28, #0x8]\n"
       "fmla v29.4s, v5.4s, v3.s[1]\n"
       "ldr d5, [x15, #0xd0]\n"
       "fmla v12.4s, v6.4s, v0.s[1]\n"
-      "ldr x23, [x15, #0x108]\n"
+      "mov v5.d[1], x22\n"
       "fmla v18.4s, v6.4s, v1.s[1]\n"
-      "cmp x12, #0x8\n"
+      "ldr x22, [x15, #0x118]\n"
       "fmla v24.4s, v6.4s, v2.s[1]\n"
-      "mov v5.d[1], x22\n"
+      "sub x12, x12, #0x4\n"
       "fmla v30.4s, v6.4s, v3.s[1]\n"
       "ldr d6, [x15, #0xe0]\n"
       "fmla v13.4s, v7.4s, v0.s[1]\n"
-      "ldr x22, [x15, #0x118]\n"
+      "mov v6.d[1], x21\n"
       "fmla v19.4s, v7.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x15, #0x128]\n"
       "fmla v25.4s, v7.4s, v2.s[1]\n"
-      "mov v6.d[1], x21\n"
+      "cmp x12, #0x8\n"
       "fmla v31.4s, v7.4s, v3.s[1]\n"
       "ldr d7, [x15, #0xf0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.4s, v4.4s, v0.s[2]\n"
-      "ldr x21, [x15, #0x128]\n"
       "fmla v14.4s, v4.4s, v1.s[2]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0x138]\n"
       "fmla v20.4s, v4.4s, v2.s[2]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "fmla v26.4s, v4.4s, v3.s[2]\n"
       "ldr d4, [x15, #0x100]\n"
       "fmla v9.4s, v5.4s, v0.s[2]\n"
-      "ldr x20, [x15, #0x138]\n"
+      "mov v4.d[1], x23\n"
       "fmla v15.4s, v5.4s, v1.s[2]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
+      "ldr x23, [x15, #0x148]\n"
       "fmla v21.4s, v5.4s, v2.s[2]\n"
-      "mov v4.d[1], x23\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       "fmla v27.4s, v5.4s, v3.s[2]\n"
       "ldr d5, [x15, #0x110]\n"
       "fmla v10.4s, v6.4s, v0.s[2]\n"
-      "ldr x23, [x15, #0x148]\n"
+      "mov v5.d[1], x22\n"
       "fmla v16.4s, v6.4s, v1.s[2]\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
+      "ldr x22, [x15, #0x158]\n"
       "fmla v22.4s, v6.4s, v2.s[2]\n"
-      "mov v5.d[1], x22\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v28.4s, v6.4s, v3.s[2]\n"
       "ldr d6, [x15, #0x120]\n"
       "fmla v11.4s, v7.4s, v0.s[2]\n"
-      "ldr x22, [x15, #0x158]\n"
+      "mov v6.d[1], x21\n"
       "fmla v17.4s, v7.4s, v1.s[2]\n"
+      "ldr x21, [x15, #0x168]\n"
       "fmla v23.4s, v7.4s, v2.s[2]\n"
-      "mov v6.d[1], x21\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v29.4s, v7.4s, v3.s[2]\n"
       "ldr d7, [x15, #0x130]\n"
+      "mov v7.d[1], x20\n"
       "fmla v12.4s, v4.4s, v0.s[2]\n"
-      "ldr x21, [x15, #0x168]\n"
       "fmla v18.4s, v4.4s, v1.s[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0x178]\n"
       "fmla v24.4s, v4.4s, v2.s[2]\n"
       "fmla v30.4s, v4.4s, v3.s[2]\n"
       "ldr d4, [x15, #0x140]\n"
       "fmla v13.4s, v5.4s, v0.s[2]\n"
-      "ldr x20, [x15, #0x178]\n"
+      "mov v4.d[1], x23\n"
       "fmla v19.4s, v5.4s, v1.s[2]\n"
       "fmla v25.4s, v5.4s, v2.s[2]\n"
-      "mov v4.d[1], x23\n"
       "fmla v31.4s, v5.4s, v3.s[2]\n"
       "ldr d5, [x15, #0x150]\n"
       "fmla v8.4s, v6.4s, v0.s[3]\n"
+      "mov v5.d[1], x22\n"
       "fmla v14.4s, v6.4s, v1.s[3]\n"
       "fmla v20.4s, v6.4s, v2.s[3]\n"
-      "mov v5.d[1], x22\n"
       "fmla v26.4s, v6.4s, v3.s[3]\n"
       "ldr d6, [x15, #0x160]\n"
       "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.4s, v7.4s, v1.s[3]\n"
       "fmla v21.4s, v7.4s, v2.s[3]\n"
-      "mov v6.d[1], x21\n"
       "fmla v27.4s, v7.4s, v3.s[3]\n"
       "ldr d7, [x15, #0x170]\n"
+      "mov v7.d[1], x20\n"
       "add x15, x15, #0x180\n"
       "fmla v10.4s, v4.4s, v0.s[3]\n"
-      "fmla v16.4s, v4.4s, v1.s[3]\n"
       "ldr x23, [x15, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "fmla v16.4s, v4.4s, v1.s[3]\n"
+      "ldr x22, [x15, #0x18]\n"
       "fmla v22.4s, v4.4s, v2.s[3]\n"
+      "ldr x21, [x15, #0x28]\n"
       "fmla v28.4s, v4.4s, v3.s[3]\n"
       "ldr d4, [x15, #0x0]\n"
       "fmla v11.4s, v5.4s, v0.s[3]\n"
-      "ldr x22, [x15, #0x18]\n"
+      "ldr x20, [x15, #0x38]\n"
       "fmla v17.4s, v5.4s, v1.s[3]\n"
-      "ldr x21, [x15, #0x28]\n"
+      "mov v4.d[1], x23\n"
       "fmla v23.4s, v5.4s, v2.s[3]\n"
-      "ldr x20, [x15, #0x38]\n"
       "fmla v29.4s, v5.4s, v3.s[3]\n"
       "ldr d5, [x15, #0x10]\n"
       "fmla v12.4s, v6.4s, v0.s[3]\n"
-      "mov v4.d[1], x23\n"
+      "mov v5.d[1], x22\n"
       "fmla v18.4s, v6.4s, v1.s[3]\n"
       "fmla v24.4s, v6.4s, v2.s[3]\n"
-      "mov v5.d[1], x22\n"
       "fmla v30.4s, v6.4s, v3.s[3]\n"
       "ldr d6, [x15, #0x20]\n"
       "fmla v13.4s, v7.4s, v0.s[3]\n"
@@ -2503,12 +2500,12 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "ldr s5, [x9], #0x4\n"
       "ldr s4, [x28], #0x4\n"
       "ldr q1, [x15, #0x0]\n"
-      "ldr q0, [x15, #0x10]\n"
-      "ldr q3, [x15, #0x20]\n"
-      "ldr q2, [x15, #0x30]\n"
       "fmla v8.4s, v1.4s, v7.s[0]\n"
+      "ldr q0, [x15, #0x10]\n"
       "fmla v14.4s, v1.4s, v6.s[0]\n"
+      "ldr q3, [x15, #0x20]\n"
       "fmla v20.4s, v1.4s, v5.s[0]\n"
+      "ldr q2, [x15, #0x30]\n"
       "fmla v26.4s, v1.4s, v4.s[0]\n"
       "ldr q1, [x15, #0x40]\n"
       "fmla v9.4s, v0.4s, v7.s[0]\n"
@@ -2540,42 +2537,42 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "cmp x13, x20\n"
       "bne 142b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x14, #0x0]\n"
-      "add x24, x14, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "add x23, x14, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
+      "prfm pstl1keep, [x14, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbz %x[flags], #1, 150f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v0.4s\n"
+      "fmin v9.4s, v9.4s, v0.4s\n"
+      "fmin v10.4s, v10.4s, v0.4s\n"
+      "fmin v11.4s, v11.4s, v0.4s\n"
+      "fmin v12.4s, v12.4s, v0.4s\n"
+      "fmin v13.4s, v13.4s, v0.4s\n"
+      "fmin v14.4s, v14.4s, v0.4s\n"
+      "fmin v15.4s, v15.4s, v0.4s\n"
+      "fmin v16.4s, v16.4s, v0.4s\n"
+      "fmin v17.4s, v17.4s, v0.4s\n"
+      "fmin v18.4s, v18.4s, v0.4s\n"
+      "fmin v19.4s, v19.4s, v0.4s\n"
+      "fmin v20.4s, v20.4s, v0.4s\n"
+      "fmin v21.4s, v21.4s, v0.4s\n"
+      "fmin v22.4s, v22.4s, v0.4s\n"
+      "fmin v23.4s, v23.4s, v0.4s\n"
+      "fmin v24.4s, v24.4s, v0.4s\n"
+      "fmin v25.4s, v25.4s, v0.4s\n"
+      "fmin v26.4s, v26.4s, v0.4s\n"
+      "fmin v27.4s, v27.4s, v0.4s\n"
+      "fmin v28.4s, v28.4s, v0.4s\n"
+      "fmin v29.4s, v29.4s, v0.4s\n"
+      "fmin v30.4s, v30.4s, v0.4s\n"
+      "fmin v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v1.4s\n"
-      "fmin v9.4s, v9.4s, v1.4s\n"
-      "fmin v10.4s, v10.4s, v1.4s\n"
-      "fmin v11.4s, v11.4s, v1.4s\n"
-      "fmin v12.4s, v12.4s, v1.4s\n"
-      "fmin v13.4s, v13.4s, v1.4s\n"
-      "fmin v14.4s, v14.4s, v1.4s\n"
-      "fmin v15.4s, v15.4s, v1.4s\n"
-      "fmin v16.4s, v16.4s, v1.4s\n"
-      "fmin v17.4s, v17.4s, v1.4s\n"
-      "fmin v18.4s, v18.4s, v1.4s\n"
-      "fmin v19.4s, v19.4s, v1.4s\n"
-      "fmin v20.4s, v20.4s, v1.4s\n"
-      "fmin v21.4s, v21.4s, v1.4s\n"
-      "fmin v22.4s, v22.4s, v1.4s\n"
-      "fmin v23.4s, v23.4s, v1.4s\n"
-      "fmin v24.4s, v24.4s, v1.4s\n"
-      "fmin v25.4s, v25.4s, v1.4s\n"
-      "fmin v26.4s, v26.4s, v1.4s\n"
-      "fmin v27.4s, v27.4s, v1.4s\n"
-      "fmin v28.4s, v28.4s, v1.4s\n"
-      "fmin v29.4s, v29.4s, v1.4s\n"
-      "fmin v30.4s, v30.4s, v1.4s\n"
-      "fmin v31.4s, v31.4s, v1.4s\n"
       "fmax v8.4s, v8.4s, v0.4s\n"
       "fmax v9.4s, v9.4s, v0.4s\n"
       "fmax v10.4s, v10.4s, v0.4s\n"
@@ -2608,153 +2605,153 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "st1 { v9.4s }, [x14], #0x10\n"
       "st1 { v10.4s }, [x14], #0x10\n"
       "st1 { v11.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
-      "st1 { v27.4s }, [x22], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
-      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
+      "st1 { v27.4s }, [x21], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
+      "st1 { v29.4s }, [x21], #0x10\n"
       "tbz x16, #2, 152f\n"
       "st1 { v12.4s }, [x14], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v30.4s }, [x22], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v30.4s }, [x21], #0x10\n"
       "tbz x16, #1, 151f\n"
       "str d13, [x14], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x16, #0, 162f\n"
       "st1 { v13.s }[2], [x14]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
-      "st1 { v31.s }[2], [x22]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v31.s }[2], [x21]\n"
       "b 162f\n"
       "151:"  // Height 4: Partial direct writeback: partial_1_20
       "tbz x16, #0, 162f\n"
       "str s13, [x14, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
-      "str s31, [x22, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
+      "str s31, [x21, #0x0]\n"
       "b 162f\n"
       "152:"  // Height 4: Partial direct writeback: partial_2_16
       "tbz x16, #1, 153f\n"
       "str d12, [x14], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x16, #0, 162f\n"
       "st1 { v12.s }[2], [x14]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
-      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v30.s }[2], [x21]\n"
       "b 162f\n"
       "153:"  // Height 4: Partial direct writeback: partial_1_16
       "tbz x16, #0, 162f\n"
       "str s12, [x14, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
-      "str s30, [x22, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
+      "str s30, [x21, #0x0]\n"
       "b 162f\n"
       "154:"  // Height 4: Partial direct writeback: partial_8_0
       "tbz x16, #3, 158f\n"
       "st1 { v8.4s }, [x14], #0x10\n"
       "st1 { v9.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
-      "st1 { v27.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
+      "st1 { v27.4s }, [x21], #0x10\n"
       "tbz x16, #2, 156f\n"
       "st1 { v10.4s }, [x14], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
       "tbz x16, #1, 155f\n"
       "str d11, [x14], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x16, #0, 162f\n"
       "st1 { v11.s }[2], [x14]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
-      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
+      "st1 { v29.s }[2], [x21]\n"
       "b 162f\n"
       "155:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x16, #0, 162f\n"
       "str s11, [x14, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
-      "str s29, [x22, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
+      "str s29, [x21, #0x0]\n"
       "b 162f\n"
       "156:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x16, #1, 157f\n"
       "str d10, [x14], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d22, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x16, #0, 162f\n"
       "st1 { v10.s }[2], [x14]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
-      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v22.s }[2], [x22]\n"
+      "st1 { v28.s }[2], [x21]\n"
       "b 162f\n"
       "157:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x16, #0, 162f\n"
       "str s10, [x14, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
-      "str s28, [x22, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s22, [x22, #0x0]\n"
+      "str s28, [x21, #0x0]\n"
       "b 162f\n"
       "158:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x16, #2, 160f\n"
       "st1 { v8.4s }, [x14], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
       "tbz x16, #1, 159f\n"
       "str d9, [x14], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
-      "str d27, [x22], #0x8\n"
+      "str d15, [x23], #0x8\n"
+      "str d21, [x22], #0x8\n"
+      "str d27, [x21], #0x8\n"
       "tbz x16, #0, 162f\n"
       "st1 { v9.s }[2], [x14]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
-      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v15.s }[2], [x23]\n"
+      "st1 { v21.s }[2], [x22]\n"
+      "st1 { v27.s }[2], [x21]\n"
       "b 162f\n"
       "159:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x16, #0, 162f\n"
       "str s9, [x14, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
-      "str s27, [x22, #0x0]\n"
+      "str s15, [x23, #0x0]\n"
+      "str s21, [x22, #0x0]\n"
+      "str s27, [x21, #0x0]\n"
       "b 162f\n"
       "160:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x16, #1, 161f\n"
       "str d8, [x14], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
-      "str d26, [x22], #0x8\n"
+      "str d14, [x23], #0x8\n"
+      "str d20, [x22], #0x8\n"
+      "str d26, [x21], #0x8\n"
       "tbz x16, #0, 162f\n"
       "st1 { v8.s }[2], [x14]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
-      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v14.s }[2], [x23]\n"
+      "st1 { v20.s }[2], [x22]\n"
+      "st1 { v26.s }[2], [x21]\n"
       "b 162f\n"
       "161:"  // Height 4: Partial direct writeback: partial_1_0
       "str s8, [x14, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
-      "str s26, [x22, #0x0]\n"
+      "str s14, [x23, #0x0]\n"
+      "str s20, [x22, #0x0]\n"
+      "str s26, [x21, #0x0]\n"
       "162:"  // Height 4: Partial direct writeback: Done
       "b 164f\n"
       "163:"  // Height 4: Full writeback
@@ -2765,24 +2762,24 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "str q12, [x14, #0x40]\n"
       "str q13, [x14, #0x50]\n"
       "add x14, x14, #0x60\n"
-      "str q14, [x24, #0x0]\n"
-      "str q15, [x24, #0x10]\n"
-      "str q16, [x24, #0x20]\n"
-      "str q17, [x24, #0x30]\n"
-      "str q18, [x24, #0x40]\n"
-      "str q19, [x24, #0x50]\n"
-      "str q20, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q22, [x23, #0x20]\n"
-      "str q23, [x23, #0x30]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q25, [x23, #0x50]\n"
-      "str q26, [x22, #0x0]\n"
-      "str q27, [x22, #0x10]\n"
-      "str q28, [x22, #0x20]\n"
-      "str q29, [x22, #0x30]\n"
-      "str q30, [x22, #0x40]\n"
-      "str q31, [x22, #0x50]\n"
+      "str q14, [x23, #0x0]\n"
+      "str q15, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "str q17, [x23, #0x30]\n"
+      "str q18, [x23, #0x40]\n"
+      "str q19, [x23, #0x50]\n"
+      "str q20, [x22, #0x0]\n"
+      "str q21, [x22, #0x10]\n"
+      "str q22, [x22, #0x20]\n"
+      "str q23, [x22, #0x30]\n"
+      "str q24, [x22, #0x40]\n"
+      "str q25, [x22, #0x50]\n"
+      "str q26, [x21, #0x0]\n"
+      "str q27, [x21, #0x10]\n"
+      "str q28, [x21, #0x20]\n"
+      "str q29, [x21, #0x30]\n"
+      "str q30, [x21, #0x40]\n"
+      "str q31, [x21, #0x50]\n"
       "164:"  // Height 4: Writeback done
       "subs x16, x16, #0x18\n"
       "bgt 125b\n"
@@ -2798,8 +2795,8 @@ void a64_hybrid_fp32_mla_4x24_a55 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "166:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp
index 38acdd5054..f5863ba348 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp
@@ -47,19 +47,18 @@ void a64_hybrid_fp32_mla_4x24 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void a64_hybrid_fp32_mla_4x24 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -100,10 +98,10 @@ void a64_hybrid_fp32_mla_4x24 (
       "cmp %x[M], #0x2\n"
       "bgt 83f\n"
       "beq 42f\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x10, 3f\n"
       "ldr q8, [x10, #0x0]\n"
@@ -223,8 +221,8 @@ void a64_hybrid_fp32_mla_4x24 (
       "mov x26, #0x0\n"
       "19:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -254,14 +252,10 @@ void a64_hybrid_fp32_mla_4x24 (
       "ldr q17, [x28, #0x60]\n"
       "fmla v11.4s, v7.4s, v0.s[0]\n"
       "ldr q16, [x28, #0x70]\n"
-      "sub x25, x25, #0x4\n"
-      "add x24, x24, #0x10\n"
       "fmla v12.4s, v19.4s, v0.s[0]\n"
       "ldr q19, [x28, #0x80]\n"
       "fmla v13.4s, v18.4s, v0.s[0]\n"
       "ldr q18, [x28, #0x90]\n"
-      "cmp x25, #0x8\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
       "ldr q17, [x28, #0xa0]\n"
       "fmla v9.4s, v16.4s, v0.s[1]\n"
@@ -290,16 +284,20 @@ void a64_hybrid_fp32_mla_4x24 (
       "ldr q17, [x28, #0x160]\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr q16, [x28, #0x170]\n"
-      "add x28, x28, #0x180\n"
+      "sub x25, x25, #0x4\n"
+      "add x24, x24, #0x10\n"
       "fmla v10.4s, v19.4s, v0.s[3]\n"
-      "ldr q4, [x28, #0x0]\n"
       "fmla v11.4s, v18.4s, v0.s[3]\n"
+      "cmp x25, #0x8\n"
+      "add x28, x28, #0x180\n"
+      "ldr q4, [x28, #0x0]\n"
       "ldr q5, [x28, #0x10]\n"
       "fmla v12.4s, v17.4s, v0.s[3]\n"
       "ldr q6, [x28, #0x20]\n"
       "fmla v13.4s, v16.4s, v0.s[3]\n"
       "ldr q0, [x24, #0x0]\n"
       "ldr q7, [x28, #0x30]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "bge 22b\n"
       "23:"  // Height 1: Multiply loop: Single iteration only
       "fmla v8.4s, v4.4s, v0.s[0]\n"
@@ -310,13 +308,10 @@ void a64_hybrid_fp32_mla_4x24 (
       "ldr q17, [x28, #0x60]\n"
       "fmla v11.4s, v7.4s, v0.s[0]\n"
       "ldr q16, [x28, #0x70]\n"
-      "add x24, x24, #0x10\n"
-      "sub x25, x25, #0x4\n"
       "fmla v12.4s, v19.4s, v0.s[0]\n"
       "ldr q19, [x28, #0x80]\n"
       "fmla v13.4s, v18.4s, v0.s[0]\n"
       "ldr q18, [x28, #0x90]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
       "ldr q17, [x28, #0xa0]\n"
       "fmla v9.4s, v16.4s, v0.s[1]\n"
@@ -345,29 +340,32 @@ void a64_hybrid_fp32_mla_4x24 (
       "ldr q17, [x28, #0x160]\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr q16, [x28, #0x170]\n"
-      "add x28, x28, #0x180\n"
+      "add x24, x24, #0x10\n"
+      "sub x25, x25, #0x4\n"
       "fmla v10.4s, v19.4s, v0.s[3]\n"
       "fmla v11.4s, v18.4s, v0.s[3]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "add x28, x28, #0x180\n"
       "fmla v12.4s, v17.4s, v0.s[3]\n"
       "fmla v13.4s, v16.4s, v0.s[3]\n"
       "24:"  // Height 1: Multiply loop: Main loop skip
       "cbz x25, 26f\n"
       "25:"  // Height 1: Multiply loop: Odd block loop
-      "ldr s20, [x24], #0x4\n"
-      "ldr q17, [x28, #0x0]\n"
+      "ldr s18, [x24], #0x4\n"
+      "ldr q16, [x28, #0x0]\n"
+      "fmla v8.4s, v16.4s, v18.s[0]\n"
       "sub x25, x25, #0x1\n"
-      "ldr q16, [x28, #0x10]\n"
-      "ldr q19, [x28, #0x20]\n"
-      "ldr q18, [x28, #0x30]\n"
-      "fmla v8.4s, v17.4s, v20.s[0]\n"
-      "ldr q17, [x28, #0x40]\n"
-      "fmla v9.4s, v16.4s, v20.s[0]\n"
+      "ldr q17, [x28, #0x10]\n"
+      "ldr q16, [x28, #0x20]\n"
+      "fmla v9.4s, v17.4s, v18.s[0]\n"
+      "fmla v10.4s, v16.4s, v18.s[0]\n"
+      "ldr q17, [x28, #0x30]\n"
+      "ldr q16, [x28, #0x40]\n"
+      "fmla v11.4s, v17.4s, v18.s[0]\n"
+      "fmla v12.4s, v16.4s, v18.s[0]\n"
       "ldr q16, [x28, #0x50]\n"
+      "fmla v13.4s, v16.4s, v18.s[0]\n"
       "add x28, x28, #0x60\n"
-      "fmla v10.4s, v19.4s, v20.s[0]\n"
-      "fmla v11.4s, v18.4s, v20.s[0]\n"
-      "fmla v12.4s, v17.4s, v20.s[0]\n"
-      "fmla v13.4s, v16.4s, v20.s[0]\n"
       "cbnz x25, 25b\n"
       "26:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -376,9 +374,9 @@ void a64_hybrid_fp32_mla_4x24 (
       "bne 19b\n"
       "prfm pstl1keep, [x27, #0x0]\n"
       "tbz %x[flags], #1, 27f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v17.4s\n"
       "fmin v9.4s, v9.4s, v17.4s\n"
@@ -481,141 +479,141 @@ void a64_hybrid_fp32_mla_4x24 (
       "bgt 2b\n"
       "b 166f\n"
       "42:"  // Height 2
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "43:"  // Height 2: Column loop
       "cbz x10, 44f\n"
       "ldr q8, [x10, #0x0]\n"
       "ldr q9, [x10, #0x10]\n"
-      "ldr q10, [x10, #0x20]\n"
-      "ldr q11, [x10, #0x30]\n"
-      "ldr q12, [x10, #0x40]\n"
-      "ldr q13, [x10, #0x50]\n"
-      "add x10, x10, #0x60\n"
       "mov v14.16b, v8.16b\n"
       "mov v15.16b, v9.16b\n"
+      "ldr q10, [x10, #0x20]\n"
+      "ldr q11, [x10, #0x30]\n"
       "mov v16.16b, v10.16b\n"
       "mov v17.16b, v11.16b\n"
+      "ldr q12, [x10, #0x40]\n"
+      "ldr q13, [x10, #0x50]\n"
       "mov v18.16b, v12.16b\n"
       "mov v19.16b, v13.16b\n"
+      "add x10, x10, #0x60\n"
       "b 59f\n"
       "44:"  // Height 2: no bias
       "tbz %x[flags], #0, 58f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x9, #0x18\n"
-      "add x24, x27, x20, LSL #2\n"
+      "add x23, x27, x20, LSL #2\n"
       "bge 57f\n"
       "tbz x9, #4, 48f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
       "tbz x9, #2, 46f\n"
       "ld1 { v12.4s }, [x27], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
       "tbz x9, #1, 45f\n"
       "ldr d13, [x27], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
       "tbz x9, #0, 56f\n"
       "ld1 { v13.s }[2], [x27]\n"
-      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v19.s }[2], [x23]\n"
       "b 56f\n"
       "45:"  // Height 2: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x9, #0, 56f\n"
       "ldr s13, [x27, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
       "b 56f\n"
       "46:"  // Height 2: Partial accumulate: partial_2_16
       "tbz x9, #1, 47f\n"
       "ldr d12, [x27], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
       "tbz x9, #0, 56f\n"
       "ld1 { v12.s }[2], [x27]\n"
-      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v18.s }[2], [x23]\n"
       "b 56f\n"
       "47:"  // Height 2: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x9, #0, 56f\n"
       "ldr s12, [x27, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
       "b 56f\n"
       "48:"  // Height 2: Partial accumulate: partial_8_0
       "tbz x9, #3, 52f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
       "tbz x9, #2, 50f\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
       "tbz x9, #1, 49f\n"
       "ldr d11, [x27], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x9, #0, 56f\n"
       "ld1 { v11.s }[2], [x27]\n"
-      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v17.s }[2], [x23]\n"
       "b 56f\n"
       "49:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x9, #0, 56f\n"
       "ldr s11, [x27, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
       "b 56f\n"
       "50:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x9, #1, 51f\n"
       "ldr d10, [x27], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x9, #0, 56f\n"
       "ld1 { v10.s }[2], [x27]\n"
-      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v16.s }[2], [x23]\n"
       "b 56f\n"
       "51:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x9, #0, 56f\n"
       "ldr s10, [x27, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
       "b 56f\n"
       "52:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x9, #2, 54f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "tbz x9, #1, 53f\n"
       "ldr d9, [x27], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x9, #0, 56f\n"
       "ld1 { v9.s }[2], [x27]\n"
-      "ld1 { v15.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x23]\n"
       "b 56f\n"
       "53:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x9, #0, 56f\n"
       "ldr s9, [x27, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
       "b 56f\n"
       "54:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x9, #1, 55f\n"
       "ldr d8, [x27], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x9, #0, 56f\n"
       "ld1 { v8.s }[2], [x27]\n"
-      "ld1 { v14.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x23]\n"
       "b 56f\n"
       "55:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s8, [x27, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
       "56:"  // Height 2: Partial accumulate: Done
       "sub x27, x27, x20\n"
@@ -627,12 +625,12 @@ void a64_hybrid_fp32_mla_4x24 (
       "ldr q11, [x27, #0x30]\n"
       "ldr q12, [x27, #0x40]\n"
       "ldr q13, [x27, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
       "b 59f\n"
       "58:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -651,8 +649,8 @@ void a64_hybrid_fp32_mla_4x24 (
       "mov x26, #0x0\n"
       "60:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 61f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -843,19 +841,19 @@ void a64_hybrid_fp32_mla_4x24 (
       "sub x25, x25, #0x1\n"
       "ldr q21, [x28, #0x0]\n"
       "ldr q20, [x28, #0x10]\n"
-      "ldr q23, [x28, #0x20]\n"
-      "ldr q22, [x28, #0x30]\n"
       "fmla v8.4s, v21.4s, v25.s[0]\n"
       "fmla v14.4s, v21.4s, v24.s[0]\n"
-      "ldr q21, [x28, #0x40]\n"
+      "ldr q23, [x28, #0x20]\n"
+      "ldr q22, [x28, #0x30]\n"
       "fmla v9.4s, v20.4s, v25.s[0]\n"
       "fmla v15.4s, v20.4s, v24.s[0]\n"
+      "ldr q21, [x28, #0x40]\n"
       "ldr q20, [x28, #0x50]\n"
-      "add x28, x28, #0x60\n"
       "fmla v10.4s, v23.4s, v25.s[0]\n"
       "fmla v16.4s, v23.4s, v24.s[0]\n"
       "fmla v11.4s, v22.4s, v25.s[0]\n"
       "fmla v17.4s, v22.4s, v24.s[0]\n"
+      "add x28, x28, #0x60\n"
       "fmla v12.4s, v21.4s, v25.s[0]\n"
       "fmla v18.4s, v21.4s, v24.s[0]\n"
       "fmla v13.4s, v20.4s, v25.s[0]\n"
@@ -867,13 +865,13 @@ void a64_hybrid_fp32_mla_4x24 (
       "cmp x26, x20\n"
       "bne 60b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 68f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.4s }, [x21]\n"
       "ld1r { v20.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v21.4s\n"
       "fmin v9.4s, v9.4s, v21.4s\n"
@@ -907,99 +905,99 @@ void a64_hybrid_fp32_mla_4x24 (
       "st1 { v9.4s }, [x27], #0x10\n"
       "st1 { v10.4s }, [x27], #0x10\n"
       "st1 { v11.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
       "tbz x9, #2, 70f\n"
       "st1 { v12.4s }, [x27], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
       "tbz x9, #1, 69f\n"
       "str d13, [x27], #0x8\n"
-      "str d19, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
       "tbz x9, #0, 80f\n"
       "st1 { v13.s }[2], [x27]\n"
-      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
       "b 80f\n"
       "69:"  // Height 2: Partial direct writeback: partial_1_20
       "tbz x9, #0, 80f\n"
       "str s13, [x27, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
       "b 80f\n"
       "70:"  // Height 2: Partial direct writeback: partial_2_16
       "tbz x9, #1, 71f\n"
       "str d12, [x27], #0x8\n"
-      "str d18, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
       "tbz x9, #0, 80f\n"
       "st1 { v12.s }[2], [x27]\n"
-      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
       "b 80f\n"
       "71:"  // Height 2: Partial direct writeback: partial_1_16
       "tbz x9, #0, 80f\n"
       "str s12, [x27, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
       "b 80f\n"
       "72:"  // Height 2: Partial direct writeback: partial_8_0
       "tbz x9, #3, 76f\n"
       "st1 { v8.4s }, [x27], #0x10\n"
       "st1 { v9.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
       "tbz x9, #2, 74f\n"
       "st1 { v10.4s }, [x27], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
       "tbz x9, #1, 73f\n"
       "str d11, [x27], #0x8\n"
-      "str d17, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
       "tbz x9, #0, 80f\n"
       "st1 { v11.s }[2], [x27]\n"
-      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
       "b 80f\n"
       "73:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x9, #0, 80f\n"
       "str s11, [x27, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
       "b 80f\n"
       "74:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x9, #1, 75f\n"
       "str d10, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
       "tbz x9, #0, 80f\n"
       "st1 { v10.s }[2], [x27]\n"
-      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
       "b 80f\n"
       "75:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x9, #0, 80f\n"
       "str s10, [x27, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
       "b 80f\n"
       "76:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x9, #2, 78f\n"
       "st1 { v8.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
       "tbz x9, #1, 77f\n"
       "str d9, [x27], #0x8\n"
-      "str d15, [x24], #0x8\n"
+      "str d15, [x23], #0x8\n"
       "tbz x9, #0, 80f\n"
       "st1 { v9.s }[2], [x27]\n"
-      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v15.s }[2], [x23]\n"
       "b 80f\n"
       "77:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x9, #0, 80f\n"
       "str s9, [x27, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
+      "str s15, [x23, #0x0]\n"
       "b 80f\n"
       "78:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x9, #1, 79f\n"
       "str d8, [x27], #0x8\n"
-      "str d14, [x24], #0x8\n"
+      "str d14, [x23], #0x8\n"
       "tbz x9, #0, 80f\n"
       "st1 { v8.s }[2], [x27]\n"
-      "st1 { v14.s }[2], [x24]\n"
+      "st1 { v14.s }[2], [x23]\n"
       "b 80f\n"
       "79:"  // Height 2: Partial direct writeback: partial_1_0
       "str s8, [x27, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
+      "str s14, [x23, #0x0]\n"
       "80:"  // Height 2: Partial direct writeback: Done
       "b 82f\n"
       "81:"  // Height 2: Full writeback
@@ -1010,38 +1008,38 @@ void a64_hybrid_fp32_mla_4x24 (
       "str q12, [x27, #0x40]\n"
       "str q13, [x27, #0x50]\n"
       "add x27, x27, #0x60\n"
-      "str q14, [x24, #0x0]\n"
-      "str q15, [x24, #0x10]\n"
-      "str q16, [x24, #0x20]\n"
-      "str q17, [x24, #0x30]\n"
-      "str q18, [x24, #0x40]\n"
-      "str q19, [x24, #0x50]\n"
+      "str q14, [x23, #0x0]\n"
+      "str q15, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "str q17, [x23, #0x30]\n"
+      "str q18, [x23, #0x40]\n"
+      "str q19, [x23, #0x50]\n"
       "82:"  // Height 2: Writeback done
       "subs x9, x9, #0x18\n"
       "bgt 43b\n"
       "b 166f\n"
       "83:"  // Height 3
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "84:"  // Height 3: Column loop
       "cbz x10, 85f\n"
       "ldr q8, [x10, #0x0]\n"
       "ldr q9, [x10, #0x10]\n"
-      "ldr q10, [x10, #0x20]\n"
-      "ldr q11, [x10, #0x30]\n"
-      "ldr q12, [x10, #0x40]\n"
-      "ldr q13, [x10, #0x50]\n"
-      "add x10, x10, #0x60\n"
       "mov v14.16b, v8.16b\n"
       "mov v15.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x10, #0x20]\n"
+      "ldr q11, [x10, #0x30]\n"
       "mov v16.16b, v10.16b\n"
       "mov v17.16b, v11.16b\n"
+      "ldr q12, [x10, #0x40]\n"
+      "ldr q13, [x10, #0x50]\n"
       "mov v18.16b, v12.16b\n"
       "mov v19.16b, v13.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
+      "add x10, x10, #0x60\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
       "mov v24.16b, v12.16b\n"
@@ -1050,147 +1048,147 @@ void a64_hybrid_fp32_mla_4x24 (
       "85:"  // Height 3: no bias
       "tbz %x[flags], #0, 99f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
       "cmp x9, #0x18\n"
-      "add x24, x27, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 98f\n"
       "tbz x9, #4, 89f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
       "tbz x9, #2, 87f\n"
       "ld1 { v12.4s }, [x27], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "tbz x9, #1, 86f\n"
       "ldr d13, [x27], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x9, #0, 97f\n"
       "ld1 { v13.s }[2], [x27]\n"
-      "ld1 { v19.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v19.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 97f\n"
       "86:"  // Height 3: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x9, #0, 97f\n"
       "ldr s13, [x27, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "b 97f\n"
       "87:"  // Height 3: Partial accumulate: partial_2_16
       "tbz x9, #1, 88f\n"
       "ldr d12, [x27], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x9, #0, 97f\n"
       "ld1 { v12.s }[2], [x27]\n"
-      "ld1 { v18.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v18.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
       "b 97f\n"
       "88:"  // Height 3: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x9, #0, 97f\n"
       "ldr s12, [x27, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
       "b 97f\n"
       "89:"  // Height 3: Partial accumulate: partial_8_0
       "tbz x9, #3, 93f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "tbz x9, #2, 91f\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
       "tbz x9, #1, 90f\n"
       "ldr d11, [x27], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d23, [x23], #0x8\n"
+      "ldr d23, [x22], #0x8\n"
       "tbz x9, #0, 97f\n"
       "ld1 { v11.s }[2], [x27]\n"
-      "ld1 { v17.s }[2], [x24]\n"
-      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v17.s }[2], [x23]\n"
+      "ld1 { v23.s }[2], [x22]\n"
       "b 97f\n"
       "90:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x9, #0, 97f\n"
       "ldr s11, [x27, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
-      "ldr s23, [x23, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
+      "ldr s23, [x22, #0x0]\n"
       "b 97f\n"
       "91:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x9, #1, 92f\n"
       "ldr d10, [x27], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d22, [x23], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
       "tbz x9, #0, 97f\n"
       "ld1 { v10.s }[2], [x27]\n"
-      "ld1 { v16.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v16.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
       "b 97f\n"
       "92:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x9, #0, 97f\n"
       "ldr s10, [x27, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
       "b 97f\n"
       "93:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x9, #2, 95f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
       "tbz x9, #1, 94f\n"
       "ldr d9, [x27], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d21, [x23], #0x8\n"
+      "ldr d21, [x22], #0x8\n"
       "tbz x9, #0, 97f\n"
       "ld1 { v9.s }[2], [x27]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x23]\n"
+      "ld1 { v21.s }[2], [x22]\n"
       "b 97f\n"
       "94:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x9, #0, 97f\n"
       "ldr s9, [x27, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s21, [x23, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
+      "ldr s21, [x22, #0x0]\n"
       "b 97f\n"
       "95:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x9, #1, 96f\n"
       "ldr d8, [x27], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d20, [x23], #0x8\n"
+      "ldr d20, [x22], #0x8\n"
       "tbz x9, #0, 97f\n"
       "ld1 { v8.s }[2], [x27]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x23]\n"
+      "ld1 { v20.s }[2], [x22]\n"
       "b 97f\n"
       "96:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s8, [x27, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s20, [x23, #0x0]\n"
+      "ldr s20, [x22, #0x0]\n"
       "97:"  // Height 3: Partial accumulate: Done
       "sub x27, x27, x20\n"
       "b 100f\n"
@@ -1201,18 +1199,18 @@ void a64_hybrid_fp32_mla_4x24 (
       "ldr q11, [x27, #0x30]\n"
       "ldr q12, [x27, #0x40]\n"
       "ldr q13, [x27, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
-      "ldr q20, [x23, #0x0]\n"
-      "ldr q21, [x23, #0x10]\n"
-      "ldr q22, [x23, #0x20]\n"
-      "ldr q23, [x23, #0x30]\n"
-      "ldr q24, [x23, #0x40]\n"
-      "ldr q25, [x23, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
+      "ldr q20, [x22, #0x0]\n"
+      "ldr q21, [x22, #0x10]\n"
+      "ldr q22, [x22, #0x20]\n"
+      "ldr q23, [x22, #0x30]\n"
+      "ldr q24, [x22, #0x40]\n"
+      "ldr q25, [x22, #0x50]\n"
       "b 100f\n"
       "99:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1237,8 +1235,8 @@ void a64_hybrid_fp32_mla_4x24 (
       "mov x26, #0x0\n"
       "101:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 102f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1389,10 +1387,10 @@ void a64_hybrid_fp32_mla_4x24 (
       "fmla v15.4s, v5.4s, v1.s[0]\n"
       "fmla v21.4s, v5.4s, v2.s[0]\n"
       "ldr q28, [x28, #0x50]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
+      "sub x25, x25, #0x4\n"
       "fmla v10.4s, v6.4s, v0.s[0]\n"
       "fmla v16.4s, v6.4s, v1.s[0]\n"
-      "sub x25, x25, #0x4\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "prfm pldl1keep, [x23, #0x80]\n"
       "fmla v22.4s, v6.4s, v2.s[0]\n"
       "ldr q27, [x28, #0x60]\n"
@@ -1486,20 +1484,20 @@ void a64_hybrid_fp32_mla_4x24 (
       "sub x25, x25, #0x1\n"
       "ldr s30, [x22], #0x4\n"
       "ldr q27, [x28, #0x0]\n"
-      "ldr q26, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
       "fmla v8.4s, v27.4s, v0.s[0]\n"
       "fmla v14.4s, v27.4s, v31.s[0]\n"
+      "ldr q26, [x28, #0x10]\n"
+      "ldr q29, [x28, #0x20]\n"
       "fmla v20.4s, v27.4s, v30.s[0]\n"
-      "ldr q27, [x28, #0x40]\n"
       "fmla v9.4s, v26.4s, v0.s[0]\n"
+      "ldr q28, [x28, #0x30]\n"
+      "ldr q27, [x28, #0x40]\n"
       "fmla v15.4s, v26.4s, v31.s[0]\n"
       "fmla v21.4s, v26.4s, v30.s[0]\n"
       "ldr q26, [x28, #0x50]\n"
-      "add x28, x28, #0x60\n"
       "fmla v10.4s, v29.4s, v0.s[0]\n"
       "fmla v16.4s, v29.4s, v31.s[0]\n"
+      "add x28, x28, #0x60\n"
       "fmla v22.4s, v29.4s, v30.s[0]\n"
       "fmla v11.4s, v28.4s, v0.s[0]\n"
       "fmla v17.4s, v28.4s, v31.s[0]\n"
@@ -1517,15 +1515,15 @@ void a64_hybrid_fp32_mla_4x24 (
       "cmp x26, x20\n"
       "bne 101b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 109f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v27.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v27.4s }, [x21]\n"
       "ld1r { v26.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v27.4s\n"
       "fmin v9.4s, v9.4s, v27.4s\n"
@@ -1571,126 +1569,126 @@ void a64_hybrid_fp32_mla_4x24 (
       "st1 { v9.4s }, [x27], #0x10\n"
       "st1 { v10.4s }, [x27], #0x10\n"
       "st1 { v11.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
       "tbz x9, #2, 111f\n"
       "st1 { v12.4s }, [x27], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x9, #1, 110f\n"
       "str d13, [x27], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x9, #0, 121f\n"
       "st1 { v13.s }[2], [x27]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 121f\n"
       "110:"  // Height 3: Partial direct writeback: partial_1_20
       "tbz x9, #0, 121f\n"
       "str s13, [x27, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 121f\n"
       "111:"  // Height 3: Partial direct writeback: partial_2_16
       "tbz x9, #1, 112f\n"
       "str d12, [x27], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x9, #0, 121f\n"
       "st1 { v12.s }[2], [x27]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 121f\n"
       "112:"  // Height 3: Partial direct writeback: partial_1_16
       "tbz x9, #0, 121f\n"
       "str s12, [x27, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "b 121f\n"
       "113:"  // Height 3: Partial direct writeback: partial_8_0
       "tbz x9, #3, 117f\n"
       "st1 { v8.4s }, [x27], #0x10\n"
       "st1 { v9.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
       "tbz x9, #2, 115f\n"
       "st1 { v10.4s }, [x27], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
       "tbz x9, #1, 114f\n"
       "str d11, [x27], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
       "tbz x9, #0, 121f\n"
       "st1 { v11.s }[2], [x27]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
       "b 121f\n"
       "114:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x9, #0, 121f\n"
       "str s11, [x27, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
       "b 121f\n"
       "115:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x9, #1, 116f\n"
       "str d10, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d22, [x22], #0x8\n"
       "tbz x9, #0, 121f\n"
       "st1 { v10.s }[2], [x27]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v22.s }[2], [x22]\n"
       "b 121f\n"
       "116:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x9, #0, 121f\n"
       "str s10, [x27, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s22, [x22, #0x0]\n"
       "b 121f\n"
       "117:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x9, #2, 119f\n"
       "st1 { v8.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
       "tbz x9, #1, 118f\n"
       "str d9, [x27], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
+      "str d15, [x23], #0x8\n"
+      "str d21, [x22], #0x8\n"
       "tbz x9, #0, 121f\n"
       "st1 { v9.s }[2], [x27]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v15.s }[2], [x23]\n"
+      "st1 { v21.s }[2], [x22]\n"
       "b 121f\n"
       "118:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x9, #0, 121f\n"
       "str s9, [x27, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
+      "str s15, [x23, #0x0]\n"
+      "str s21, [x22, #0x0]\n"
       "b 121f\n"
       "119:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x9, #1, 120f\n"
       "str d8, [x27], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
+      "str d14, [x23], #0x8\n"
+      "str d20, [x22], #0x8\n"
       "tbz x9, #0, 121f\n"
       "st1 { v8.s }[2], [x27]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v14.s }[2], [x23]\n"
+      "st1 { v20.s }[2], [x22]\n"
       "b 121f\n"
       "120:"  // Height 3: Partial direct writeback: partial_1_0
       "str s8, [x27, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
+      "str s14, [x23, #0x0]\n"
+      "str s20, [x22, #0x0]\n"
       "121:"  // Height 3: Partial direct writeback: Done
       "b 123f\n"
       "122:"  // Height 3: Full writeback
@@ -1701,48 +1699,47 @@ void a64_hybrid_fp32_mla_4x24 (
       "str q12, [x27, #0x40]\n"
       "str q13, [x27, #0x50]\n"
       "add x27, x27, #0x60\n"
-      "str q14, [x24, #0x0]\n"
-      "str q15, [x24, #0x10]\n"
-      "str q16, [x24, #0x20]\n"
-      "str q17, [x24, #0x30]\n"
-      "str q18, [x24, #0x40]\n"
-      "str q19, [x24, #0x50]\n"
-      "str q20, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q22, [x23, #0x20]\n"
-      "str q23, [x23, #0x30]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q25, [x23, #0x50]\n"
+      "str q14, [x23, #0x0]\n"
+      "str q15, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "str q17, [x23, #0x30]\n"
+      "str q18, [x23, #0x40]\n"
+      "str q19, [x23, #0x50]\n"
+      "str q20, [x22, #0x0]\n"
+      "str q21, [x22, #0x10]\n"
+      "str q22, [x22, #0x20]\n"
+      "str q23, [x22, #0x30]\n"
+      "str q24, [x22, #0x40]\n"
+      "str q25, [x22, #0x50]\n"
       "123:"  // Height 3: Writeback done
       "subs x9, x9, #0x18\n"
       "bgt 84b\n"
       "b 166f\n"
       "124:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x10\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "125:"  // Height 4: Column loop
       "cbz x10, 126f\n"
       "ldr q8, [x10, #0x0]\n"
       "ldr q9, [x10, #0x10]\n"
-      "ldr q10, [x10, #0x20]\n"
-      "ldr q11, [x10, #0x30]\n"
-      "ldr q12, [x10, #0x40]\n"
-      "ldr q13, [x10, #0x50]\n"
-      "add x10, x10, #0x60\n"
       "mov v14.16b, v8.16b\n"
       "mov v15.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x10, #0x20]\n"
+      "ldr q11, [x10, #0x30]\n"
       "mov v16.16b, v10.16b\n"
       "mov v17.16b, v11.16b\n"
+      "ldr q12, [x10, #0x40]\n"
+      "ldr q13, [x10, #0x50]\n"
       "mov v18.16b, v12.16b\n"
       "mov v19.16b, v13.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
+      "add x10, x10, #0x60\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
       "mov v24.16b, v12.16b\n"
@@ -1757,175 +1754,175 @@ void a64_hybrid_fp32_mla_4x24 (
       "126:"  // Height 4: no bias
       "tbz %x[flags], #0, 140f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x9, #0x18\n"
-      "add x24, x27, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x23, x27, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x9, #0x18\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 139f\n"
       "tbz x9, #4, 130f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v27.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x9, #2, 128f\n"
       "ld1 { v12.4s }, [x27], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x9, #1, 127f\n"
       "ldr d13, [x27], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x9, #0, 138f\n"
       "ld1 { v13.s }[2], [x27]\n"
-      "ld1 { v19.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v19.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 138f\n"
       "127:"  // Height 4: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x9, #0, 138f\n"
       "ldr s13, [x27, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 138f\n"
       "128:"  // Height 4: Partial accumulate: partial_2_16
       "tbz x9, #1, 129f\n"
       "ldr d12, [x27], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x9, #0, 138f\n"
       "ld1 { v12.s }[2], [x27]\n"
-      "ld1 { v18.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v18.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 138f\n"
       "129:"  // Height 4: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x9, #0, 138f\n"
       "ldr s12, [x27, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 138f\n"
       "130:"  // Height 4: Partial accumulate: partial_8_0
       "tbz x9, #3, 134f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v27.4s }, [x21], #0x10\n"
       "tbz x9, #2, 132f\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x9, #1, 131f\n"
       "ldr d11, [x27], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d23, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d23, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x9, #0, 138f\n"
       "ld1 { v11.s }[2], [x27]\n"
-      "ld1 { v17.s }[2], [x24]\n"
-      "ld1 { v23.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v17.s }[2], [x23]\n"
+      "ld1 { v23.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 138f\n"
       "131:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x9, #0, 138f\n"
       "ldr s11, [x27, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
-      "ldr s23, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
+      "ldr s23, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 138f\n"
       "132:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x9, #1, 133f\n"
       "ldr d10, [x27], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x9, #0, 138f\n"
       "ld1 { v10.s }[2], [x27]\n"
-      "ld1 { v16.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v16.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 138f\n"
       "133:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x9, #0, 138f\n"
       "ldr s10, [x27, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "b 138f\n"
       "134:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x9, #2, 136f\n"
       "ld1 { v8.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v20.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v20.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "tbz x9, #1, 135f\n"
       "ldr d9, [x27], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d21, [x23], #0x8\n"
-      "ldr d27, [x22], #0x8\n"
+      "ldr d21, [x22], #0x8\n"
+      "ldr d27, [x21], #0x8\n"
       "tbz x9, #0, 138f\n"
       "ld1 { v9.s }[2], [x27]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v21.s }[2], [x23]\n"
-      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x23]\n"
+      "ld1 { v21.s }[2], [x22]\n"
+      "ld1 { v27.s }[2], [x21]\n"
       "b 138f\n"
       "135:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x9, #0, 138f\n"
       "ldr s9, [x27, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s21, [x23, #0x0]\n"
-      "ldr s27, [x22, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
+      "ldr s21, [x22, #0x0]\n"
+      "ldr s27, [x21, #0x0]\n"
       "b 138f\n"
       "136:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x9, #1, 137f\n"
       "ldr d8, [x27], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d20, [x23], #0x8\n"
-      "ldr d26, [x22], #0x8\n"
+      "ldr d20, [x22], #0x8\n"
+      "ldr d26, [x21], #0x8\n"
       "tbz x9, #0, 138f\n"
       "ld1 { v8.s }[2], [x27]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v20.s }[2], [x23]\n"
-      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x23]\n"
+      "ld1 { v20.s }[2], [x22]\n"
+      "ld1 { v26.s }[2], [x21]\n"
       "b 138f\n"
       "137:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s8, [x27, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s20, [x23, #0x0]\n"
-      "ldr s26, [x22, #0x0]\n"
+      "ldr s20, [x22, #0x0]\n"
+      "ldr s26, [x21, #0x0]\n"
       "138:"  // Height 4: Partial accumulate: Done
       "sub x27, x27, x20\n"
       "b 141f\n"
@@ -1936,24 +1933,24 @@ void a64_hybrid_fp32_mla_4x24 (
       "ldr q11, [x27, #0x30]\n"
       "ldr q12, [x27, #0x40]\n"
       "ldr q13, [x27, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
-      "ldr q20, [x23, #0x0]\n"
-      "ldr q21, [x23, #0x10]\n"
-      "ldr q22, [x23, #0x20]\n"
-      "ldr q23, [x23, #0x30]\n"
-      "ldr q24, [x23, #0x40]\n"
-      "ldr q25, [x23, #0x50]\n"
-      "ldr q26, [x22, #0x0]\n"
-      "ldr q27, [x22, #0x10]\n"
-      "ldr q28, [x22, #0x20]\n"
-      "ldr q29, [x22, #0x30]\n"
-      "ldr q30, [x22, #0x40]\n"
-      "ldr q31, [x22, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
+      "ldr q20, [x22, #0x0]\n"
+      "ldr q21, [x22, #0x10]\n"
+      "ldr q22, [x22, #0x20]\n"
+      "ldr q23, [x22, #0x30]\n"
+      "ldr q24, [x22, #0x40]\n"
+      "ldr q25, [x22, #0x50]\n"
+      "ldr q26, [x21, #0x0]\n"
+      "ldr q27, [x21, #0x10]\n"
+      "ldr q28, [x21, #0x20]\n"
+      "ldr q29, [x21, #0x30]\n"
+      "ldr q30, [x21, #0x40]\n"
+      "ldr q31, [x21, #0x50]\n"
       "b 141f\n"
       "140:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1984,8 +1981,8 @@ void a64_hybrid_fp32_mla_4x24 (
       "mov x26, #0x0\n"
       "142:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 143f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2292,10 +2289,10 @@ void a64_hybrid_fp32_mla_4x24 (
       "ldr s4, [x21], #0x4\n"
       "ldr q1, [x28, #0x0]\n"
       "ldr q0, [x28, #0x10]\n"
-      "ldr q3, [x28, #0x20]\n"
-      "ldr q2, [x28, #0x30]\n"
       "fmla v8.4s, v1.4s, v7.s[0]\n"
       "fmla v14.4s, v1.4s, v6.s[0]\n"
+      "ldr q3, [x28, #0x20]\n"
+      "ldr q2, [x28, #0x30]\n"
       "fmla v20.4s, v1.4s, v5.s[0]\n"
       "fmla v26.4s, v1.4s, v4.s[0]\n"
       "ldr q1, [x28, #0x40]\n"
@@ -2328,17 +2325,17 @@ void a64_hybrid_fp32_mla_4x24 (
       "cmp x26, x20\n"
       "bne 142b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "prfm pstl1keep, [x23, #0x0]\n"
-      "add x22, x23, x20, LSL #2\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbz %x[flags], #1, 150f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v1.4s\n"
       "fmin v9.4s, v9.4s, v1.4s\n"
@@ -2396,153 +2393,153 @@ void a64_hybrid_fp32_mla_4x24 (
       "st1 { v9.4s }, [x27], #0x10\n"
       "st1 { v10.4s }, [x27], #0x10\n"
       "st1 { v11.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
-      "st1 { v27.4s }, [x22], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
-      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
+      "st1 { v27.4s }, [x21], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
+      "st1 { v29.4s }, [x21], #0x10\n"
       "tbz x9, #2, 152f\n"
       "st1 { v12.4s }, [x27], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v30.4s }, [x22], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v30.4s }, [x21], #0x10\n"
       "tbz x9, #1, 151f\n"
       "str d13, [x27], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x9, #0, 162f\n"
       "st1 { v13.s }[2], [x27]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
-      "st1 { v31.s }[2], [x22]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v31.s }[2], [x21]\n"
       "b 162f\n"
       "151:"  // Height 4: Partial direct writeback: partial_1_20
       "tbz x9, #0, 162f\n"
       "str s13, [x27, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
-      "str s31, [x22, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
+      "str s31, [x21, #0x0]\n"
       "b 162f\n"
       "152:"  // Height 4: Partial direct writeback: partial_2_16
       "tbz x9, #1, 153f\n"
       "str d12, [x27], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x9, #0, 162f\n"
       "st1 { v12.s }[2], [x27]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
-      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v30.s }[2], [x21]\n"
       "b 162f\n"
       "153:"  // Height 4: Partial direct writeback: partial_1_16
       "tbz x9, #0, 162f\n"
       "str s12, [x27, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
-      "str s30, [x22, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
+      "str s30, [x21, #0x0]\n"
       "b 162f\n"
       "154:"  // Height 4: Partial direct writeback: partial_8_0
       "tbz x9, #3, 158f\n"
       "st1 { v8.4s }, [x27], #0x10\n"
       "st1 { v9.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
-      "st1 { v27.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v15.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
+      "st1 { v27.4s }, [x21], #0x10\n"
       "tbz x9, #2, 156f\n"
       "st1 { v10.4s }, [x27], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
       "tbz x9, #1, 155f\n"
       "str d11, [x27], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x9, #0, 162f\n"
       "st1 { v11.s }[2], [x27]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
-      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
+      "st1 { v29.s }[2], [x21]\n"
       "b 162f\n"
       "155:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x9, #0, 162f\n"
       "str s11, [x27, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
-      "str s29, [x22, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
+      "str s29, [x21, #0x0]\n"
       "b 162f\n"
       "156:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x9, #1, 157f\n"
       "str d10, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d22, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x9, #0, 162f\n"
       "st1 { v10.s }[2], [x27]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
-      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v22.s }[2], [x22]\n"
+      "st1 { v28.s }[2], [x21]\n"
       "b 162f\n"
       "157:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x9, #0, 162f\n"
       "str s10, [x27, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
-      "str s28, [x22, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s22, [x22, #0x0]\n"
+      "str s28, [x21, #0x0]\n"
       "b 162f\n"
       "158:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x9, #2, 160f\n"
       "st1 { v8.4s }, [x27], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
       "tbz x9, #1, 159f\n"
       "str d9, [x27], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
-      "str d27, [x22], #0x8\n"
+      "str d15, [x23], #0x8\n"
+      "str d21, [x22], #0x8\n"
+      "str d27, [x21], #0x8\n"
       "tbz x9, #0, 162f\n"
       "st1 { v9.s }[2], [x27]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
-      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v15.s }[2], [x23]\n"
+      "st1 { v21.s }[2], [x22]\n"
+      "st1 { v27.s }[2], [x21]\n"
       "b 162f\n"
       "159:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x9, #0, 162f\n"
       "str s9, [x27, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
-      "str s27, [x22, #0x0]\n"
+      "str s15, [x23, #0x0]\n"
+      "str s21, [x22, #0x0]\n"
+      "str s27, [x21, #0x0]\n"
       "b 162f\n"
       "160:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x9, #1, 161f\n"
       "str d8, [x27], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
-      "str d26, [x22], #0x8\n"
+      "str d14, [x23], #0x8\n"
+      "str d20, [x22], #0x8\n"
+      "str d26, [x21], #0x8\n"
       "tbz x9, #0, 162f\n"
       "st1 { v8.s }[2], [x27]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
-      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v14.s }[2], [x23]\n"
+      "st1 { v20.s }[2], [x22]\n"
+      "st1 { v26.s }[2], [x21]\n"
       "b 162f\n"
       "161:"  // Height 4: Partial direct writeback: partial_1_0
       "str s8, [x27, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
-      "str s26, [x22, #0x0]\n"
+      "str s14, [x23, #0x0]\n"
+      "str s20, [x22, #0x0]\n"
+      "str s26, [x21, #0x0]\n"
       "162:"  // Height 4: Partial direct writeback: Done
       "b 164f\n"
       "163:"  // Height 4: Full writeback
@@ -2553,24 +2550,24 @@ void a64_hybrid_fp32_mla_4x24 (
       "str q12, [x27, #0x40]\n"
       "str q13, [x27, #0x50]\n"
       "add x27, x27, #0x60\n"
-      "str q14, [x24, #0x0]\n"
-      "str q15, [x24, #0x10]\n"
-      "str q16, [x24, #0x20]\n"
-      "str q17, [x24, #0x30]\n"
-      "str q18, [x24, #0x40]\n"
-      "str q19, [x24, #0x50]\n"
-      "str q20, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q22, [x23, #0x20]\n"
-      "str q23, [x23, #0x30]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q25, [x23, #0x50]\n"
-      "str q26, [x22, #0x0]\n"
-      "str q27, [x22, #0x10]\n"
-      "str q28, [x22, #0x20]\n"
-      "str q29, [x22, #0x30]\n"
-      "str q30, [x22, #0x40]\n"
-      "str q31, [x22, #0x50]\n"
+      "str q14, [x23, #0x0]\n"
+      "str q15, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "str q17, [x23, #0x30]\n"
+      "str q18, [x23, #0x40]\n"
+      "str q19, [x23, #0x50]\n"
+      "str q20, [x22, #0x0]\n"
+      "str q21, [x22, #0x10]\n"
+      "str q22, [x22, #0x20]\n"
+      "str q23, [x22, #0x30]\n"
+      "str q24, [x22, #0x40]\n"
+      "str q25, [x22, #0x50]\n"
+      "str q26, [x21, #0x0]\n"
+      "str q27, [x21, #0x10]\n"
+      "str q28, [x21, #0x20]\n"
+      "str q29, [x21, #0x30]\n"
+      "str q30, [x21, #0x40]\n"
+      "str q31, [x21, #0x50]\n"
       "164:"  // Height 4: Writeback done
       "subs x9, x9, #0x18\n"
       "bgt 125b\n"
@@ -2586,8 +2583,8 @@ void a64_hybrid_fp32_mla_4x24 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "166:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16.hpp
index d47ab34e03..7f85d2dd42 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16.hpp
@@ -71,8 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixedTRB<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 1> transforms = {};
-
+    StdTransformsFixedTRB<rhs_operand_type, result_type, 6, 16, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp
index a315a2fe4f..3428028ac8 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp
@@ -47,19 +47,18 @@ void a64_hybrid_fp32_mla_6x16_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void a64_hybrid_fp32_mla_6x16_a55 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -103,10 +101,10 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "cmp %x[M], #0x2\n"
       "bgt 67f\n"
       "beq 34f\n"
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x7, 3f\n"
       "ldr q8, [x7, #0x0]\n"
@@ -189,8 +187,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "mov x15, #0x0\n"
       "15:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 16f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -212,89 +210,86 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "18:"  // Height 1: Multiply loop: Main loop head
       "fmla v8.4s, v6.4s, v0.s[0]\n"
       "ldr d17, [x17, #0x20]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr x20, [x17, #0x28]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "ldr d16, [x17, #0x30]\n"
-      "add x13, x13, #0x10\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x38]\n"
-      "sub x14, x14, #0x4\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x48]\n"
-      "ldr x22, [x13, #0x8]\n"
-      "cmp x14, #0x8\n"
       "mov v16.d[1], x20\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "ldr d17, [x17, #0x40]\n"
+      "ldr x20, [x17, #0x48]\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
       "ldr d16, [x17, #0x50]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x58]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x68]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
       "ldr d17, [x17, #0x60]\n"
+      "ldr x20, [x17, #0x68]\n"
       "fmla v9.4s, v16.4s, v0.s[1]\n"
       "ldr d16, [x17, #0x70]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x78]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x88]\n"
       "mov v16.d[1], x20\n"
       "fmla v10.4s, v17.4s, v0.s[1]\n"
       "ldr d17, [x17, #0x80]\n"
+      "ldr x20, [x17, #0x88]\n"
       "fmla v11.4s, v16.4s, v0.s[1]\n"
       "ldr d16, [x17, #0x90]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0x98]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xa8]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.4s, v17.4s, v0.s[2]\n"
       "ldr d17, [x17, #0xa0]\n"
+      "ldr x20, [x17, #0xa8]\n"
       "fmla v9.4s, v16.4s, v0.s[2]\n"
       "ldr d16, [x17, #0xb0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0xb8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xc8]\n"
       "mov v16.d[1], x20\n"
       "fmla v10.4s, v17.4s, v0.s[2]\n"
       "ldr d17, [x17, #0xc0]\n"
+      "ldr x20, [x17, #0xc8]\n"
       "fmla v11.4s, v16.4s, v0.s[2]\n"
       "ldr d16, [x17, #0xd0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0xd8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xe8]\n"
       "mov v16.d[1], x20\n"
       "fmla v8.4s, v17.4s, v0.s[3]\n"
       "ldr d17, [x17, #0xe0]\n"
+      "ldr x20, [x17, #0xe8]\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr d16, [x17, #0xf0]\n"
+      "mov v17.d[1], x20\n"
       "ldr x20, [x17, #0xf8]\n"
-      "add x17, x17, #0x100\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x8]\n"
       "mov v16.d[1], x20\n"
+      "add x13, x13, #0x10\n"
+      "add x17, x17, #0x100\n"
       "fmla v10.4s, v17.4s, v0.s[3]\n"
       "ldr d6, [x17, #0x0]\n"
+      "ldr x20, [x17, #0x8]\n"
       "fmla v11.4s, v16.4s, v0.s[3]\n"
       "ldr d0, [x13, #0x0]\n"
+      "sub x14, x14, #0x4\n"
       "ldr d7, [x17, #0x10]\n"
+      "cmp x14, #0x8\n"
+      "ldr x21, [x13, #0x8]\n"
+      "mov v6.d[1], x20\n"
       "ldr x20, [x17, #0x18]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x22\n"
+      "mov v0.d[1], x21\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "bge 18b\n"
       "19:"  // Height 1: Multiply loop: Single iteration only
       "fmla v8.4s, v6.4s, v0.s[0]\n"
       "ldr q17, [x17, #0x20]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "ldr q16, [x17, #0x30]\n"
-      "add x13, x13, #0x10\n"
-      "sub x14, x14, #0x4\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "ldr q17, [x17, #0x40]\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
       "ldr q16, [x17, #0x50]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
       "ldr q17, [x17, #0x60]\n"
       "fmla v9.4s, v16.4s, v0.s[1]\n"
@@ -315,23 +310,26 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr q17, [x17, #0xe0]\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr q16, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "add x13, x13, #0x10\n"
+      "sub x14, x14, #0x4\n"
       "fmla v10.4s, v17.4s, v0.s[3]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v11.4s, v16.4s, v0.s[3]\n"
+      "add x17, x17, #0x100\n"
       "20:"  // Height 1: Multiply loop: Main loop skip
       "cbz x14, 22f\n"
       "21:"  // Height 1: Multiply loop: Odd block loop
-      "ldr s18, [x13], #0x4\n"
+      "ldr s17, [x13], #0x4\n"
       "sub x14, x14, #0x1\n"
-      "ldr q17, [x17, #0x0]\n"
+      "ldr q16, [x17, #0x0]\n"
+      "fmla v8.4s, v16.4s, v17.s[0]\n"
       "ldr q16, [x17, #0x10]\n"
-      "fmla v8.4s, v17.4s, v18.s[0]\n"
-      "ldr q17, [x17, #0x20]\n"
-      "fmla v9.4s, v16.4s, v18.s[0]\n"
+      "fmla v9.4s, v16.4s, v17.s[0]\n"
+      "ldr q16, [x17, #0x20]\n"
+      "fmla v10.4s, v16.4s, v17.s[0]\n"
       "ldr q16, [x17, #0x30]\n"
+      "fmla v11.4s, v16.4s, v17.s[0]\n"
       "add x17, x17, #0x40\n"
-      "fmla v10.4s, v17.4s, v18.s[0]\n"
-      "fmla v11.4s, v16.4s, v18.s[0]\n"
       "cbnz x14, 21b\n"
       "22:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -340,14 +338,14 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "bne 15b\n"
       "prfm pstl1keep, [x16, #0x0]\n"
       "tbz %x[flags], #1, 23f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v16.4s\n"
+      "fmin v9.4s, v9.4s, v16.4s\n"
+      "fmin v10.4s, v10.4s, v16.4s\n"
+      "fmin v11.4s, v11.4s, v16.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v17.4s\n"
-      "fmin v9.4s, v9.4s, v17.4s\n"
-      "fmin v10.4s, v10.4s, v17.4s\n"
-      "fmin v11.4s, v11.4s, v17.4s\n"
       "fmax v8.4s, v8.4s, v16.4s\n"
       "fmax v9.4s, v9.4s, v16.4s\n"
       "fmax v10.4s, v10.4s, v16.4s\n"
@@ -412,96 +410,96 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "bgt 2b\n"
       "b 200f\n"
       "34:"  // Height 2
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "35:"  // Height 2: Column loop
       "cbz x7, 36f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
+      "add x7, x7, #0x40\n"
       "b 47f\n"
       "36:"  // Height 2: no bias
       "tbz %x[flags], #0, 46f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x8, #0x10\n"
-      "add x26, x16, x20, LSL #2\n"
+      "add x25, x16, x20, LSL #2\n"
       "bge 45f\n"
       "tbz x8, #3, 40f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "ld1 { v9.4s }, [x16], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
       "tbz x8, #2, 38f\n"
       "ld1 { v10.4s }, [x16], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
       "tbz x8, #1, 37f\n"
       "ldr d11, [x16], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "tbz x8, #0, 44f\n"
       "ld1 { v11.s }[2], [x16]\n"
-      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v15.s }[2], [x25]\n"
       "b 44f\n"
       "37:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x8, #0, 44f\n"
       "ldr s11, [x16, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
       "b 44f\n"
       "38:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x8, #1, 39f\n"
       "ldr d10, [x16], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "tbz x8, #0, 44f\n"
       "ld1 { v10.s }[2], [x16]\n"
-      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v14.s }[2], [x25]\n"
       "b 44f\n"
       "39:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x8, #0, 44f\n"
       "ldr s10, [x16, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
       "b 44f\n"
       "40:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x8, #2, 42f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "tbz x8, #1, 41f\n"
       "ldr d9, [x16], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "tbz x8, #0, 44f\n"
       "ld1 { v9.s }[2], [x16]\n"
-      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v13.s }[2], [x25]\n"
       "b 44f\n"
       "41:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x8, #0, 44f\n"
       "ldr s9, [x16, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
       "b 44f\n"
       "42:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x8, #1, 43f\n"
       "ldr d8, [x16], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "tbz x8, #0, 44f\n"
       "ld1 { v8.s }[2], [x16]\n"
-      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v12.s }[2], [x25]\n"
       "b 44f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "44:"  // Height 2: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 47f\n"
@@ -510,10 +508,10 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "b 47f\n"
       "46:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -528,8 +526,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "mov x15, #0x0\n"
       "48:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 49f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -554,98 +552,98 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "blt 52f\n"
       "51:"  // Height 2: Multiply loop: Main loop head
       "fmla v8.4s, v6.4s, v0.s[0]\n"
-      "ldr x20, [x17, #0x28]\n"
+      "ldr x21, [x17, #0x28]\n"
       "fmla v12.4s, v6.4s, v1.s[0]\n"
       "ldr d17, [x17, #0x20]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
-      "ldr x21, [x17, #0x38]\n"
+      "ldr x20, [x17, #0x38]\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "ldr d16, [x17, #0x30]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x48]\n"
-      "add x13, x13, #0x10\n"
-      "add x12, x12, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.4s, v17.4s, v1.s[0]\n"
       "ldr d17, [x17, #0x40]\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
-      "ldr x21, [x17, #0x58]\n"
+      "ldr x20, [x17, #0x48]\n"
       "fmla v15.4s, v16.4s, v1.s[0]\n"
       "ldr d16, [x17, #0x50]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x68]\n"
-      "ldr x23, [x13, #0x8]\n"
-      "sub x14, x14, #0x4\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0x58]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v12.4s, v17.4s, v1.s[1]\n"
       "ldr d17, [x17, #0x60]\n"
       "fmla v9.4s, v16.4s, v0.s[1]\n"
-      "ldr x21, [x17, #0x78]\n"
+      "ldr x20, [x17, #0x78]\n"
       "fmla v13.4s, v16.4s, v1.s[1]\n"
       "ldr d16, [x17, #0x70]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x88]\n"
-      "ldr x22, [x12, #0x8]\n"
-      "cmp x14, #0x8\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.4s, v17.4s, v0.s[1]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.4s, v17.4s, v1.s[1]\n"
       "ldr d17, [x17, #0x80]\n"
       "fmla v11.4s, v16.4s, v0.s[1]\n"
-      "ldr x21, [x17, #0x98]\n"
+      "ldr x20, [x17, #0x88]\n"
       "fmla v15.4s, v16.4s, v1.s[1]\n"
       "ldr d16, [x17, #0x90]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xa8]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0x98]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.4s, v17.4s, v0.s[2]\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v12.4s, v17.4s, v1.s[2]\n"
       "ldr d17, [x17, #0xa0]\n"
       "fmla v9.4s, v16.4s, v0.s[2]\n"
-      "ldr x21, [x17, #0xb8]\n"
+      "ldr x20, [x17, #0xb8]\n"
       "fmla v13.4s, v16.4s, v1.s[2]\n"
       "ldr d16, [x17, #0xb0]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xc8]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "mov v17.d[1], x21\n"
       "fmla v10.4s, v17.4s, v0.s[2]\n"
+      "mov v16.d[1], x20\n"
       "fmla v14.4s, v17.4s, v1.s[2]\n"
       "ldr d17, [x17, #0xc0]\n"
       "fmla v11.4s, v16.4s, v0.s[2]\n"
-      "ldr x21, [x17, #0xd8]\n"
+      "ldr x20, [x17, #0xc8]\n"
       "fmla v15.4s, v16.4s, v1.s[2]\n"
       "ldr d16, [x17, #0xd0]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xe8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x17, #0xd8]\n"
+      "mov v16.d[1], x20\n"
       "fmla v8.4s, v17.4s, v0.s[3]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v12.4s, v17.4s, v1.s[3]\n"
       "ldr d17, [x17, #0xe0]\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
-      "ldr x21, [x17, #0xf8]\n"
+      "ldr x20, [x17, #0xf8]\n"
       "fmla v13.4s, v16.4s, v1.s[3]\n"
       "ldr d16, [x17, #0xf0]\n"
-      "mov v17.d[1], x20\n"
+      "mov v17.d[1], x21\n"
+      "add x13, x13, #0x10\n"
+      "mov v16.d[1], x20\n"
+      "add x12, x12, #0x10\n"
       "add x17, x17, #0x100\n"
-      "ldr x20, [x17, #0x8]\n"
-      "mov v16.d[1], x21\n"
       "fmla v10.4s, v17.4s, v0.s[3]\n"
       "fmla v14.4s, v17.4s, v1.s[3]\n"
       "ldr d6, [x17, #0x0]\n"
+      "ldr x21, [x17, #0x8]\n"
       "fmla v11.4s, v16.4s, v0.s[3]\n"
       "ldr d0, [x13, #0x0]\n"
       "fmla v15.4s, v16.4s, v1.s[3]\n"
       "ldr d1, [x12, #0x0]\n"
+      "sub x14, x14, #0x4\n"
       "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x20\n"
+      "cmp x14, #0x8\n"
+      "ldr x20, [x13, #0x8]\n"
+      "mov v6.d[1], x21\n"
+      "ldr x21, [x12, #0x8]\n"
+      "mov v0.d[1], x20\n"
       "ldr x20, [x17, #0x18]\n"
-      "mov v0.d[1], x23\n"
-      "mov v1.d[1], x22\n"
+      "mov v1.d[1], x21\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "bge 51b\n"
       "52:"  // Height 2: Multiply loop: Single iteration only
       "fmla v8.4s, v6.4s, v0.s[0]\n"
@@ -707,8 +705,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "sub x14, x14, #0x1\n"
       "ldr s18, [x12], #0x4\n"
       "ldr q17, [x17, #0x0]\n"
-      "ldr q16, [x17, #0x10]\n"
       "fmla v8.4s, v17.4s, v19.s[0]\n"
+      "ldr q16, [x17, #0x10]\n"
       "fmla v12.4s, v17.4s, v18.s[0]\n"
       "ldr q17, [x17, #0x20]\n"
       "fmla v9.4s, v16.4s, v19.s[0]\n"
@@ -726,22 +724,22 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "cmp x15, x20\n"
       "bne 48b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x16, x20, LSL #2\n"
       "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "tbz %x[flags], #1, 56f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v16.4s\n"
+      "fmin v9.4s, v9.4s, v16.4s\n"
+      "fmin v10.4s, v10.4s, v16.4s\n"
+      "fmin v11.4s, v11.4s, v16.4s\n"
+      "fmin v12.4s, v12.4s, v16.4s\n"
+      "fmin v13.4s, v13.4s, v16.4s\n"
+      "fmin v14.4s, v14.4s, v16.4s\n"
+      "fmin v15.4s, v15.4s, v16.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v17.4s\n"
-      "fmin v9.4s, v9.4s, v17.4s\n"
-      "fmin v10.4s, v10.4s, v17.4s\n"
-      "fmin v11.4s, v11.4s, v17.4s\n"
-      "fmin v12.4s, v12.4s, v17.4s\n"
-      "fmin v13.4s, v13.4s, v17.4s\n"
-      "fmin v14.4s, v14.4s, v17.4s\n"
-      "fmin v15.4s, v15.4s, v17.4s\n"
       "fmax v8.4s, v8.4s, v16.4s\n"
       "fmax v9.4s, v9.4s, v16.4s\n"
       "fmax v10.4s, v10.4s, v16.4s\n"
@@ -756,63 +754,63 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "tbz x8, #3, 60f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
       "st1 { v9.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
       "tbz x8, #2, 58f\n"
       "st1 { v10.4s }, [x16], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
       "tbz x8, #1, 57f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
       "tbz x8, #0, 64f\n"
       "st1 { v11.s }[2], [x16]\n"
-      "st1 { v15.s }[2], [x26]\n"
+      "st1 { v15.s }[2], [x25]\n"
       "b 64f\n"
       "57:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x8, #0, 64f\n"
       "str s11, [x16, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
       "b 64f\n"
       "58:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x8, #1, 59f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
+      "str d14, [x25], #0x8\n"
       "tbz x8, #0, 64f\n"
       "st1 { v10.s }[2], [x16]\n"
-      "st1 { v14.s }[2], [x26]\n"
+      "st1 { v14.s }[2], [x25]\n"
       "b 64f\n"
       "59:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x8, #0, 64f\n"
       "str s10, [x16, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
       "b 64f\n"
       "60:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x8, #2, 62f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
       "tbz x8, #1, 61f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
+      "str d13, [x25], #0x8\n"
       "tbz x8, #0, 64f\n"
       "st1 { v9.s }[2], [x16]\n"
-      "st1 { v13.s }[2], [x26]\n"
+      "st1 { v13.s }[2], [x25]\n"
       "b 64f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x8, #0, 64f\n"
       "str s9, [x16, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
       "b 64f\n"
       "62:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x8, #1, 63f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
+      "str d12, [x25], #0x8\n"
       "tbz x8, #0, 64f\n"
       "st1 { v8.s }[2], [x16]\n"
-      "st1 { v12.s }[2], [x26]\n"
+      "st1 { v12.s }[2], [x25]\n"
       "b 64f\n"
       "63:"  // Height 2: Partial direct writeback: partial_1_0
       "str s8, [x16, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
       "64:"  // Height 2: Partial direct writeback: Done
       "b 66f\n"
       "65:"  // Height 2: Full writeback
@@ -821,31 +819,31 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
       "66:"  // Height 2: Writeback done
       "subs x8, x8, #0x10\n"
       "bgt 35b\n"
       "b 200f\n"
       "67:"  // Height 3
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "68:"  // Height 3: Column loop
       "cbz x7, 69f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
       "mov v16.16b, v8.16b\n"
+      "add x7, x7, #0x40\n"
       "mov v17.16b, v9.16b\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
@@ -853,94 +851,94 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "69:"  // Height 3: no bias
       "tbz %x[flags], #0, 79f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x16, x20, LSL #2\n"
       "cmp x8, #0x10\n"
-      "add x26, x16, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "bge 78f\n"
       "tbz x8, #3, 73f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
       "ld1 { v9.4s }, [x16], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "tbz x8, #2, 71f\n"
       "ld1 { v10.4s }, [x16], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
       "tbz x8, #1, 70f\n"
       "ldr d11, [x16], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x8, #0, 77f\n"
       "ld1 { v11.s }[2], [x16]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
       "b 77f\n"
       "70:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x8, #0, 77f\n"
       "ldr s11, [x16, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
       "b 77f\n"
       "71:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x8, #1, 72f\n"
       "ldr d10, [x16], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x8, #0, 77f\n"
       "ld1 { v10.s }[2], [x16]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
       "b 77f\n"
       "72:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x8, #0, 77f\n"
       "ldr s10, [x16, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
       "b 77f\n"
       "73:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x8, #2, 75f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
       "tbz x8, #1, 74f\n"
       "ldr d9, [x16], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x8, #0, 77f\n"
       "ld1 { v9.s }[2], [x16]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
       "b 77f\n"
       "74:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x8, #0, 77f\n"
       "ldr s9, [x16, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
       "b 77f\n"
       "75:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x8, #1, 76f\n"
       "ldr d8, [x16], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
       "tbz x8, #0, 77f\n"
       "ld1 { v8.s }[2], [x16]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
       "b 77f\n"
       "76:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s12, [x26, #0x0]\n"
-      "ldr s16, [x25, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
       "77:"  // Height 3: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 80f\n"
@@ -949,14 +947,14 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
       "b 80f\n"
       "79:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -975,8 +973,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "mov x15, #0x0\n"
       "81:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 82f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1011,15 +1009,11 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v16.4s, v6.4s, v2.s[0]\n"
       "ldr d21, [x17, #0x20]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
-      "add x13, x13, #0x10\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
-      "add x12, x12, #0x10\n"
+      "ldr x21, [x17, #0x48]\n"
       "fmla v17.4s, v7.4s, v2.s[0]\n"
       "ldr d20, [x17, #0x30]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x48]\n"
-      "add x11, x11, #0x10\n"
-      "ldr x24, [x13, #0x8]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.4s, v21.4s, v0.s[0]\n"
       "fmla v14.4s, v21.4s, v1.s[0]\n"
@@ -1027,15 +1021,11 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v18.4s, v21.4s, v2.s[0]\n"
       "ldr d21, [x17, #0x40]\n"
       "fmla v11.4s, v20.4s, v0.s[0]\n"
-      "ldr x23, [x12, #0x8]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.4s, v20.4s, v1.s[0]\n"
-      "ldr x22, [x11, #0x8]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v19.4s, v20.4s, v2.s[0]\n"
       "ldr d20, [x17, #0x50]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x68]\n"
-      "sub x14, x14, #0x4\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.4s, v21.4s, v0.s[1]\n"
       "fmla v12.4s, v21.4s, v1.s[1]\n"
@@ -1043,14 +1033,11 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v16.4s, v21.4s, v2.s[1]\n"
       "ldr d21, [x17, #0x60]\n"
       "fmla v9.4s, v20.4s, v0.s[1]\n"
-      "cmp x14, #0x8\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.4s, v20.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "ldr x21, [x17, #0x88]\n"
       "fmla v17.4s, v20.4s, v2.s[1]\n"
       "ldr d20, [x17, #0x70]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x88]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.4s, v21.4s, v0.s[1]\n"
       "fmla v14.4s, v21.4s, v1.s[1]\n"
@@ -1058,11 +1045,11 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v18.4s, v21.4s, v2.s[1]\n"
       "ldr d21, [x17, #0x80]\n"
       "fmla v11.4s, v20.4s, v0.s[1]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.4s, v20.4s, v1.s[1]\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v19.4s, v20.4s, v2.s[1]\n"
       "ldr d20, [x17, #0x90]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xa8]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.4s, v21.4s, v0.s[2]\n"
       "fmla v12.4s, v21.4s, v1.s[2]\n"
@@ -1070,11 +1057,11 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v16.4s, v21.4s, v2.s[2]\n"
       "ldr d21, [x17, #0xa0]\n"
       "fmla v9.4s, v20.4s, v0.s[2]\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.4s, v20.4s, v1.s[2]\n"
+      "ldr x21, [x17, #0xc8]\n"
       "fmla v17.4s, v20.4s, v2.s[2]\n"
       "ldr d20, [x17, #0xb0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xc8]\n"
       "mov v20.d[1], x20\n"
       "fmla v10.4s, v21.4s, v0.s[2]\n"
       "fmla v14.4s, v21.4s, v1.s[2]\n"
@@ -1082,11 +1069,11 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v18.4s, v21.4s, v2.s[2]\n"
       "ldr d21, [x17, #0xc0]\n"
       "fmla v11.4s, v20.4s, v0.s[2]\n"
+      "mov v21.d[1], x21\n"
       "fmla v15.4s, v20.4s, v1.s[2]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v19.4s, v20.4s, v2.s[2]\n"
       "ldr d20, [x17, #0xd0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xe8]\n"
       "mov v20.d[1], x20\n"
       "fmla v8.4s, v21.4s, v0.s[3]\n"
       "fmla v12.4s, v21.4s, v1.s[3]\n"
@@ -1094,29 +1081,40 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v16.4s, v21.4s, v2.s[3]\n"
       "ldr d21, [x17, #0xe0]\n"
       "fmla v9.4s, v20.4s, v0.s[3]\n"
+      "mov v21.d[1], x21\n"
       "fmla v13.4s, v20.4s, v1.s[3]\n"
+      "add x13, x13, #0x10\n"
       "fmla v17.4s, v20.4s, v2.s[3]\n"
       "ldr d20, [x17, #0xf0]\n"
-      "mov v21.d[1], x21\n"
-      "add x17, x17, #0x100\n"
-      "ldr x21, [x17, #0x8]\n"
       "mov v20.d[1], x20\n"
+      "add x12, x12, #0x10\n"
+      "add x11, x11, #0x10\n"
+      "add x17, x17, #0x100\n"
       "fmla v10.4s, v21.4s, v0.s[3]\n"
+      "ldr x20, [x17, #0x8]\n"
       "fmla v14.4s, v21.4s, v1.s[3]\n"
-      "ldr x20, [x17, #0x18]\n"
+      "ldr x23, [x13, #0x8]\n"
       "fmla v18.4s, v21.4s, v2.s[3]\n"
       "ldr d6, [x17, #0x0]\n"
       "fmla v11.4s, v20.4s, v0.s[3]\n"
       "ldr d0, [x13, #0x0]\n"
       "fmla v15.4s, v20.4s, v1.s[3]\n"
       "ldr d1, [x12, #0x0]\n"
+      "ldr x22, [x12, #0x8]\n"
       "fmla v19.4s, v20.4s, v2.s[3]\n"
       "ldr d2, [x11, #0x0]\n"
+      "sub x14, x14, #0x4\n"
       "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x24\n"
-      "mov v1.d[1], x23\n"
-      "mov v2.d[1], x22\n"
+      "cmp x14, #0x8\n"
+      "ldr x21, [x11, #0x8]\n"
+      "mov v6.d[1], x20\n"
+      "ldr x20, [x17, #0x18]\n"
+      "mov v0.d[1], x23\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
+      "mov v1.d[1], x22\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v2.d[1], x21\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "mov v7.d[1], x20\n"
       "bge 84b\n"
       "85:"  // Height 3: Multiply loop: Single iteration only
@@ -1198,8 +1196,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr s23, [x12], #0x4\n"
       "ldr s22, [x11], #0x4\n"
       "ldr q21, [x17, #0x0]\n"
-      "ldr q20, [x17, #0x10]\n"
       "fmla v8.4s, v21.4s, v24.s[0]\n"
+      "ldr q20, [x17, #0x10]\n"
       "fmla v12.4s, v21.4s, v23.s[0]\n"
       "fmla v16.4s, v21.4s, v22.s[0]\n"
       "ldr q21, [x17, #0x20]\n"
@@ -1221,28 +1219,28 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "cmp x15, x20\n"
       "bne 81b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x16, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "tbz %x[flags], #1, 89f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v20.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v20.4s\n"
+      "fmin v9.4s, v9.4s, v20.4s\n"
+      "fmin v10.4s, v10.4s, v20.4s\n"
+      "fmin v11.4s, v11.4s, v20.4s\n"
+      "fmin v12.4s, v12.4s, v20.4s\n"
+      "fmin v13.4s, v13.4s, v20.4s\n"
+      "fmin v14.4s, v14.4s, v20.4s\n"
+      "fmin v15.4s, v15.4s, v20.4s\n"
+      "fmin v16.4s, v16.4s, v20.4s\n"
+      "fmin v17.4s, v17.4s, v20.4s\n"
+      "fmin v18.4s, v18.4s, v20.4s\n"
+      "fmin v19.4s, v19.4s, v20.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.4s }, [x21]\n"
       "ld1r { v20.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v21.4s\n"
-      "fmin v9.4s, v9.4s, v21.4s\n"
-      "fmin v10.4s, v10.4s, v21.4s\n"
-      "fmin v11.4s, v11.4s, v21.4s\n"
-      "fmin v12.4s, v12.4s, v21.4s\n"
-      "fmin v13.4s, v13.4s, v21.4s\n"
-      "fmin v14.4s, v14.4s, v21.4s\n"
-      "fmin v15.4s, v15.4s, v21.4s\n"
-      "fmin v16.4s, v16.4s, v21.4s\n"
-      "fmin v17.4s, v17.4s, v21.4s\n"
-      "fmin v18.4s, v18.4s, v21.4s\n"
-      "fmin v19.4s, v19.4s, v21.4s\n"
       "fmax v8.4s, v8.4s, v20.4s\n"
       "fmax v9.4s, v9.4s, v20.4s\n"
       "fmax v10.4s, v10.4s, v20.4s\n"
@@ -1261,79 +1259,79 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "tbz x8, #3, 93f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
       "st1 { v9.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
       "tbz x8, #2, 91f\n"
       "st1 { v10.4s }, [x16], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
       "tbz x8, #1, 90f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x8, #0, 97f\n"
       "st1 { v11.s }[2], [x16]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
       "b 97f\n"
       "90:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x8, #0, 97f\n"
       "str s11, [x16, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
       "b 97f\n"
       "91:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x8, #1, 92f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x8, #0, 97f\n"
       "st1 { v10.s }[2], [x16]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
       "b 97f\n"
       "92:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x8, #0, 97f\n"
       "str s10, [x16, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
       "b 97f\n"
       "93:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x8, #2, 95f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
       "tbz x8, #1, 94f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x8, #0, 97f\n"
       "st1 { v9.s }[2], [x16]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
       "b 97f\n"
       "94:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x8, #0, 97f\n"
       "str s9, [x16, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
       "b 97f\n"
       "95:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x8, #1, 96f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x8, #0, 97f\n"
       "st1 { v8.s }[2], [x16]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
       "b 97f\n"
       "96:"  // Height 3: Partial direct writeback: partial_1_0
       "str s8, [x16, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
       "97:"  // Height 3: Partial direct writeback: Done
       "b 99f\n"
       "98:"  // Height 3: Full writeback
@@ -1342,35 +1340,35 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "99:"  // Height 3: Writeback done
       "subs x8, x8, #0x10\n"
       "bgt 68b\n"
       "b 200f\n"
       "100:"  // Height 4
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "101:"  // Height 4: Column loop
       "cbz x7, 102f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
       "mov v16.16b, v8.16b\n"
+      "add x7, x7, #0x40\n"
       "mov v17.16b, v9.16b\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
@@ -1382,111 +1380,111 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "102:"  // Height 4: no bias
       "tbz %x[flags], #0, 112f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x26, x16, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x16, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "add x23, x24, x20, LSL #2\n"
       "bge 111f\n"
       "tbz x8, #3, 106f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "ld1 { v9.4s }, [x16], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
       "tbz x8, #2, 104f\n"
       "ld1 { v10.4s }, [x16], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
       "tbz x8, #1, 103f\n"
       "ldr d11, [x16], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x8, #0, 110f\n"
       "ld1 { v11.s }[2], [x16]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
       "b 110f\n"
       "103:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x8, #0, 110f\n"
       "ldr s11, [x16, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
       "b 110f\n"
       "104:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x8, #1, 105f\n"
       "ldr d10, [x16], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x8, #0, 110f\n"
       "ld1 { v10.s }[2], [x16]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
       "b 110f\n"
       "105:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x8, #0, 110f\n"
       "ldr s10, [x16, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
       "b 110f\n"
       "106:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x8, #2, 108f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "tbz x8, #1, 107f\n"
       "ldr d9, [x16], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x8, #0, 110f\n"
       "ld1 { v9.s }[2], [x16]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
       "b 110f\n"
       "107:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x8, #0, 110f\n"
       "ldr s9, [x16, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
       "b 110f\n"
       "108:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x8, #1, 109f\n"
       "ldr d8, [x16], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x8, #0, 110f\n"
       "ld1 { v8.s }[2], [x16]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
       "b 110f\n"
       "109:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s12, [x26, #0x0]\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
       "110:"  // Height 4: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 113f\n"
@@ -1495,18 +1493,18 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "b 113f\n"
       "112:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1529,8 +1527,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "mov x15, #0x0\n"
       "114:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 115f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1563,121 +1561,122 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "blt 118f\n"
       "117:"  // Height 4: Multiply loop: Main loop head
       "fmla v8.4s, v6.4s, v0.s[0]\n"
-      "ldr x20, [x17, #0x28]\n"
+      "ldr x21, [x17, #0x28]\n"
       "fmla v12.4s, v6.4s, v1.s[0]\n"
-      "ldr x21, [x17, #0x38]\n"
+      "ldr x20, [x17, #0x38]\n"
       "fmla v16.4s, v6.4s, v2.s[0]\n"
       "add x13, x13, #0x10\n"
       "fmla v20.4s, v6.4s, v3.s[0]\n"
       "ldr d25, [x17, #0x20]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
-      "add x12, x12, #0x10\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
-      "add x11, x11, #0x10\n"
+      "ldr x21, [x17, #0x48]\n"
       "fmla v17.4s, v7.4s, v2.s[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x12, x12, #0x10\n"
       "fmla v21.4s, v7.4s, v3.s[0]\n"
       "ldr d24, [x17, #0x30]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.4s, v25.4s, v0.s[0]\n"
-      "ldr x20, [x17, #0x48]\n"
       "fmla v14.4s, v25.4s, v1.s[0]\n"
-      "add x10, x10, #0x10\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x58]\n"
       "fmla v18.4s, v25.4s, v2.s[0]\n"
+      "add x11, x11, #0x10\n"
       "fmla v22.4s, v25.4s, v3.s[0]\n"
       "ldr d25, [x17, #0x40]\n"
       "fmla v11.4s, v24.4s, v0.s[0]\n"
-      "ldr x21, [x17, #0x58]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.4s, v24.4s, v1.s[0]\n"
-      "ldr x25, [x13, #0x8]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v19.4s, v24.4s, v2.s[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x10, x10, #0x10\n"
       "fmla v23.4s, v24.4s, v3.s[0]\n"
       "ldr d24, [x17, #0x50]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.4s, v25.4s, v0.s[1]\n"
-      "ldr x20, [x17, #0x68]\n"
       "fmla v12.4s, v25.4s, v1.s[1]\n"
-      "ldr x24, [x12, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x78]\n"
       "fmla v16.4s, v25.4s, v2.s[1]\n"
+      "ldr x25, [x13, #0x8]\n"
       "fmla v20.4s, v25.4s, v3.s[1]\n"
       "ldr d25, [x17, #0x60]\n"
       "fmla v9.4s, v24.4s, v0.s[1]\n"
-      "ldr x21, [x17, #0x78]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.4s, v24.4s, v1.s[1]\n"
-      "ldr x23, [x11, #0x8]\n"
+      "ldr x21, [x17, #0x88]\n"
       "fmla v17.4s, v24.4s, v2.s[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x24, [x12, #0x8]\n"
       "fmla v21.4s, v24.4s, v3.s[1]\n"
       "ldr d24, [x17, #0x70]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.4s, v25.4s, v0.s[1]\n"
-      "ldr x20, [x17, #0x88]\n"
       "fmla v14.4s, v25.4s, v1.s[1]\n"
-      "ldr x22, [x10, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x98]\n"
       "fmla v18.4s, v25.4s, v2.s[1]\n"
+      "ldr x23, [x11, #0x8]\n"
       "fmla v22.4s, v25.4s, v3.s[1]\n"
       "ldr d25, [x17, #0x80]\n"
       "fmla v11.4s, v24.4s, v0.s[1]\n"
-      "ldr x21, [x17, #0x98]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.4s, v24.4s, v1.s[1]\n"
-      "sub x14, x14, #0x4\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v19.4s, v24.4s, v2.s[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x22, [x10, #0x8]\n"
       "fmla v23.4s, v24.4s, v3.s[1]\n"
       "ldr d24, [x17, #0x90]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.4s, v25.4s, v0.s[2]\n"
-      "ldr x20, [x17, #0xa8]\n"
       "fmla v12.4s, v25.4s, v1.s[2]\n"
-      "cmp x14, #0x8\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0xb8]\n"
       "fmla v16.4s, v25.4s, v2.s[2]\n"
+      "sub x14, x14, #0x4\n"
       "fmla v20.4s, v25.4s, v3.s[2]\n"
       "ldr d25, [x17, #0xa0]\n"
       "fmla v9.4s, v24.4s, v0.s[2]\n"
-      "ldr x21, [x17, #0xb8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.4s, v24.4s, v1.s[2]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr x21, [x17, #0xc8]\n"
       "fmla v17.4s, v24.4s, v2.s[2]\n"
-      "mov v25.d[1], x20\n"
+      "cmp x14, #0x8\n"
       "fmla v21.4s, v24.4s, v3.s[2]\n"
       "ldr d24, [x17, #0xb0]\n"
+      "mov v24.d[1], x20\n"
       "fmla v10.4s, v25.4s, v0.s[2]\n"
-      "ldr x20, [x17, #0xc8]\n"
       "fmla v14.4s, v25.4s, v1.s[2]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0xd8]\n"
       "fmla v18.4s, v25.4s, v2.s[2]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v22.4s, v25.4s, v3.s[2]\n"
       "ldr d25, [x17, #0xc0]\n"
       "fmla v11.4s, v24.4s, v0.s[2]\n"
-      "ldr x21, [x17, #0xd8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v15.4s, v24.4s, v1.s[2]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v19.4s, v24.4s, v2.s[2]\n"
-      "mov v25.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "fmla v23.4s, v24.4s, v3.s[2]\n"
       "ldr d24, [x17, #0xd0]\n"
+      "mov v24.d[1], x20\n"
       "fmla v8.4s, v25.4s, v0.s[3]\n"
-      "ldr x20, [x17, #0xe8]\n"
       "fmla v12.4s, v25.4s, v1.s[3]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0xf8]\n"
       "fmla v16.4s, v25.4s, v2.s[3]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "fmla v20.4s, v25.4s, v3.s[3]\n"
       "ldr d25, [x17, #0xe0]\n"
       "fmla v9.4s, v24.4s, v0.s[3]\n"
-      "ldr x21, [x17, #0xf8]\n"
+      "mov v25.d[1], x21\n"
       "fmla v13.4s, v24.4s, v1.s[3]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       "fmla v17.4s, v24.4s, v2.s[3]\n"
-      "mov v25.d[1], x20\n"
       "fmla v21.4s, v24.4s, v3.s[3]\n"
       "ldr d24, [x17, #0xf0]\n"
+      "mov v24.d[1], x20\n"
       "add x17, x17, #0x100\n"
       "fmla v10.4s, v25.4s, v0.s[3]\n"
+      "ldr x21, [x17, #0x8]\n"
       "fmla v14.4s, v25.4s, v1.s[3]\n"
-      "ldr x20, [x17, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x17, #0x18]\n"
       "fmla v18.4s, v25.4s, v2.s[3]\n"
       "fmla v22.4s, v25.4s, v3.s[3]\n"
       "ldr d6, [x17, #0x0]\n"
@@ -1690,8 +1689,7 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v23.4s, v24.4s, v3.s[3]\n"
       "ldr d3, [x10, #0x0]\n"
       "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x20\n"
-      "ldr x20, [x17, #0x18]\n"
+      "mov v6.d[1], x21\n"
       "mov v0.d[1], x25\n"
       "mov v1.d[1], x24\n"
       "mov v2.d[1], x23\n"
@@ -1796,8 +1794,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr s27, [x11], #0x4\n"
       "ldr s26, [x10], #0x4\n"
       "ldr q25, [x17, #0x0]\n"
-      "ldr q24, [x17, #0x10]\n"
       "fmla v8.4s, v25.4s, v29.s[0]\n"
+      "ldr q24, [x17, #0x10]\n"
       "fmla v12.4s, v25.4s, v28.s[0]\n"
       "fmla v16.4s, v25.4s, v27.s[0]\n"
       "fmla v20.4s, v25.4s, v26.s[0]\n"
@@ -1823,34 +1821,34 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "cmp x15, x20\n"
       "bne 114b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "add x25, x16, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
+      "prfm pstl1keep, [x16, #0x0]\n"
       "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 122f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v24.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v24.4s\n"
+      "fmin v9.4s, v9.4s, v24.4s\n"
+      "fmin v10.4s, v10.4s, v24.4s\n"
+      "fmin v11.4s, v11.4s, v24.4s\n"
+      "fmin v12.4s, v12.4s, v24.4s\n"
+      "fmin v13.4s, v13.4s, v24.4s\n"
+      "fmin v14.4s, v14.4s, v24.4s\n"
+      "fmin v15.4s, v15.4s, v24.4s\n"
+      "fmin v16.4s, v16.4s, v24.4s\n"
+      "fmin v17.4s, v17.4s, v24.4s\n"
+      "fmin v18.4s, v18.4s, v24.4s\n"
+      "fmin v19.4s, v19.4s, v24.4s\n"
+      "fmin v20.4s, v20.4s, v24.4s\n"
+      "fmin v21.4s, v21.4s, v24.4s\n"
+      "fmin v22.4s, v22.4s, v24.4s\n"
+      "fmin v23.4s, v23.4s, v24.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v25.4s }, [x21]\n"
       "ld1r { v24.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v25.4s\n"
-      "fmin v9.4s, v9.4s, v25.4s\n"
-      "fmin v10.4s, v10.4s, v25.4s\n"
-      "fmin v11.4s, v11.4s, v25.4s\n"
-      "fmin v12.4s, v12.4s, v25.4s\n"
-      "fmin v13.4s, v13.4s, v25.4s\n"
-      "fmin v14.4s, v14.4s, v25.4s\n"
-      "fmin v15.4s, v15.4s, v25.4s\n"
-      "fmin v16.4s, v16.4s, v25.4s\n"
-      "fmin v17.4s, v17.4s, v25.4s\n"
-      "fmin v18.4s, v18.4s, v25.4s\n"
-      "fmin v19.4s, v19.4s, v25.4s\n"
-      "fmin v20.4s, v20.4s, v25.4s\n"
-      "fmin v21.4s, v21.4s, v25.4s\n"
-      "fmin v22.4s, v22.4s, v25.4s\n"
-      "fmin v23.4s, v23.4s, v25.4s\n"
       "fmax v8.4s, v8.4s, v24.4s\n"
       "fmax v9.4s, v9.4s, v24.4s\n"
       "fmax v10.4s, v10.4s, v24.4s\n"
@@ -1873,95 +1871,95 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "tbz x8, #3, 126f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
       "st1 { v9.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
       "tbz x8, #2, 124f\n"
       "st1 { v10.4s }, [x16], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
       "tbz x8, #1, 123f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
       "tbz x8, #0, 130f\n"
       "st1 { v11.s }[2], [x16]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
       "b 130f\n"
       "123:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x8, #0, 130f\n"
       "str s11, [x16, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
       "b 130f\n"
       "124:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x8, #1, 125f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
       "tbz x8, #0, 130f\n"
       "st1 { v10.s }[2], [x16]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
       "b 130f\n"
       "125:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x8, #0, 130f\n"
       "str s10, [x16, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
       "b 130f\n"
       "126:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x8, #2, 128f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
       "tbz x8, #1, 127f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
       "tbz x8, #0, 130f\n"
       "st1 { v9.s }[2], [x16]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
       "b 130f\n"
       "127:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x8, #0, 130f\n"
       "str s9, [x16, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
       "b 130f\n"
       "128:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x8, #1, 129f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x8, #0, 130f\n"
       "st1 { v8.s }[2], [x16]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
       "b 130f\n"
       "129:"  // Height 4: Partial direct writeback: partial_1_0
       "str s8, [x16, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
       "130:"  // Height 4: Partial direct writeback: Done
       "b 132f\n"
       "131:"  // Height 4: Full writeback
@@ -1970,39 +1968,39 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
       "132:"  // Height 4: Writeback done
       "subs x8, x8, #0x10\n"
       "bgt 101b\n"
       "b 200f\n"
       "133:"  // Height 5
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
       "134:"  // Height 5: Column loop
       "cbz x7, 135f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
       "mov v16.16b, v8.16b\n"
+      "add x7, x7, #0x40\n"
       "mov v17.16b, v9.16b\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
@@ -2018,128 +2016,128 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "135:"  // Height 5: no bias
       "tbz %x[flags], #0, 145f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x26, x16, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x16, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 144f\n"
       "tbz x8, #3, 139f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "ld1 { v9.4s }, [x16], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x8, #2, 137f\n"
       "ld1 { v10.4s }, [x16], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
       "tbz x8, #1, 136f\n"
       "ldr d11, [x16], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x8, #0, 143f\n"
       "ld1 { v11.s }[2], [x16]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
       "b 143f\n"
       "136:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x8, #0, 143f\n"
       "ldr s11, [x16, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
       "b 143f\n"
       "137:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x8, #1, 138f\n"
       "ldr d10, [x16], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x8, #0, 143f\n"
       "ld1 { v10.s }[2], [x16]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
       "b 143f\n"
       "138:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x8, #0, 143f\n"
       "ldr s10, [x16, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
       "b 143f\n"
       "139:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x8, #2, 141f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "tbz x8, #1, 140f\n"
       "ldr d9, [x16], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x8, #0, 143f\n"
       "ld1 { v9.s }[2], [x16]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 143f\n"
       "140:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x8, #0, 143f\n"
       "ldr s9, [x16, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "b 143f\n"
       "141:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x8, #1, 142f\n"
       "ldr d8, [x16], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x8, #0, 143f\n"
       "ld1 { v8.s }[2], [x16]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
       "b 143f\n"
       "142:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s12, [x26, #0x0]\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
       "143:"  // Height 5: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 146f\n"
@@ -2148,22 +2146,22 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
       "b 146f\n"
       "145:"  // Height 5: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2190,8 +2188,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "mov x15, #0x0\n"
       "147:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 148f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2238,131 +2236,131 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v24.4s, v6.4s, v4.s[0]\n"
       "ldr d29, [x17, #0x20]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
-      "add x11, x11, #0x10\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
-      "add x10, x10, #0x10\n"
+      "ldr x21, [x17, #0x48]\n"
       "fmla v17.4s, v7.4s, v2.s[0]\n"
-      "mov v29.d[1], x21\n"
+      "add x11, x11, #0x10\n"
       "fmla v21.4s, v7.4s, v3.s[0]\n"
-      "ldr x21, [x17, #0x48]\n"
+      "add x10, x10, #0x10\n"
       "fmla v25.4s, v7.4s, v4.s[0]\n"
       "ldr d28, [x17, #0x30]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.4s, v29.4s, v0.s[0]\n"
-      "add x9, x9, #0x10\n"
       "fmla v14.4s, v29.4s, v1.s[0]\n"
-      "ldr x26, [x13, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x58]\n"
       "fmla v18.4s, v29.4s, v2.s[0]\n"
+      "add x9, x9, #0x10\n"
       "fmla v22.4s, v29.4s, v3.s[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x26, [x13, #0x8]\n"
       "fmla v26.4s, v29.4s, v4.s[0]\n"
       "ldr d29, [x17, #0x40]\n"
       "fmla v11.4s, v28.4s, v0.s[0]\n"
-      "ldr x25, [x12, #0x8]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.4s, v28.4s, v1.s[0]\n"
-      "ldr x24, [x11, #0x8]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v19.4s, v28.4s, v2.s[0]\n"
-      "mov v29.d[1], x21\n"
+      "ldr x25, [x12, #0x8]\n"
       "fmla v23.4s, v28.4s, v3.s[0]\n"
-      "ldr x21, [x17, #0x68]\n"
+      "ldr x24, [x11, #0x8]\n"
       "fmla v27.4s, v28.4s, v4.s[0]\n"
       "ldr d28, [x17, #0x50]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.4s, v29.4s, v0.s[1]\n"
-      "ldr x23, [x10, #0x8]\n"
       "fmla v12.4s, v29.4s, v1.s[1]\n"
-      "ldr x22, [x9, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x78]\n"
       "fmla v16.4s, v29.4s, v2.s[1]\n"
+      "ldr x23, [x10, #0x8]\n"
       "fmla v20.4s, v29.4s, v3.s[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "ldr x22, [x9, #0x8]\n"
       "fmla v24.4s, v29.4s, v4.s[1]\n"
       "ldr d29, [x17, #0x60]\n"
       "fmla v9.4s, v28.4s, v0.s[1]\n"
-      "sub x14, x14, #0x4\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.4s, v28.4s, v1.s[1]\n"
-      "cmp x14, #0x8\n"
+      "ldr x21, [x17, #0x88]\n"
       "fmla v17.4s, v28.4s, v2.s[1]\n"
-      "mov v29.d[1], x21\n"
+      "sub x14, x14, #0x4\n"
       "fmla v21.4s, v28.4s, v3.s[1]\n"
-      "ldr x21, [x17, #0x88]\n"
+      "cmp x14, #0x8\n"
       "fmla v25.4s, v28.4s, v4.s[1]\n"
       "ldr d28, [x17, #0x70]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.4s, v29.4s, v0.s[1]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v14.4s, v29.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0x98]\n"
       "fmla v18.4s, v29.4s, v2.s[1]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v22.4s, v29.4s, v3.s[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "fmla v26.4s, v29.4s, v4.s[1]\n"
       "ldr d29, [x17, #0x80]\n"
       "fmla v11.4s, v28.4s, v0.s[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.4s, v28.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v19.4s, v28.4s, v2.s[1]\n"
-      "mov v29.d[1], x21\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "fmla v23.4s, v28.4s, v3.s[1]\n"
-      "ldr x21, [x17, #0xa8]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       "fmla v27.4s, v28.4s, v4.s[1]\n"
       "ldr d28, [x17, #0x90]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.4s, v29.4s, v0.s[2]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v12.4s, v29.4s, v1.s[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0xb8]\n"
       "fmla v16.4s, v29.4s, v2.s[2]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v20.4s, v29.4s, v3.s[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
       "fmla v24.4s, v29.4s, v4.s[2]\n"
       "ldr d29, [x17, #0xa0]\n"
       "fmla v9.4s, v28.4s, v0.s[2]\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.4s, v28.4s, v1.s[2]\n"
+      "ldr x21, [x17, #0xc8]\n"
       "fmla v17.4s, v28.4s, v2.s[2]\n"
-      "mov v29.d[1], x21\n"
       "fmla v21.4s, v28.4s, v3.s[2]\n"
-      "ldr x21, [x17, #0xc8]\n"
       "fmla v25.4s, v28.4s, v4.s[2]\n"
       "ldr d28, [x17, #0xb0]\n"
+      "mov v28.d[1], x20\n"
       "fmla v10.4s, v29.4s, v0.s[2]\n"
       "fmla v14.4s, v29.4s, v1.s[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0xd8]\n"
       "fmla v18.4s, v29.4s, v2.s[2]\n"
       "fmla v22.4s, v29.4s, v3.s[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
       "fmla v26.4s, v29.4s, v4.s[2]\n"
       "ldr d29, [x17, #0xc0]\n"
       "fmla v11.4s, v28.4s, v0.s[2]\n"
+      "mov v29.d[1], x21\n"
       "fmla v15.4s, v28.4s, v1.s[2]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v19.4s, v28.4s, v2.s[2]\n"
-      "mov v29.d[1], x21\n"
       "fmla v23.4s, v28.4s, v3.s[2]\n"
-      "ldr x21, [x17, #0xe8]\n"
       "fmla v27.4s, v28.4s, v4.s[2]\n"
       "ldr d28, [x17, #0xd0]\n"
+      "mov v28.d[1], x20\n"
       "fmla v8.4s, v29.4s, v0.s[3]\n"
       "fmla v12.4s, v29.4s, v1.s[3]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x17, #0xf8]\n"
       "fmla v16.4s, v29.4s, v2.s[3]\n"
       "fmla v20.4s, v29.4s, v3.s[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
       "fmla v24.4s, v29.4s, v4.s[3]\n"
       "ldr d29, [x17, #0xe0]\n"
       "fmla v9.4s, v28.4s, v0.s[3]\n"
+      "mov v29.d[1], x21\n"
       "fmla v13.4s, v28.4s, v1.s[3]\n"
       "fmla v17.4s, v28.4s, v2.s[3]\n"
-      "mov v29.d[1], x21\n"
       "fmla v21.4s, v28.4s, v3.s[3]\n"
       "fmla v25.4s, v28.4s, v4.s[3]\n"
       "ldr d28, [x17, #0xf0]\n"
+      "mov v28.d[1], x20\n"
       "add x17, x17, #0x100\n"
       "fmla v10.4s, v29.4s, v0.s[3]\n"
-      "fmla v14.4s, v29.4s, v1.s[3]\n"
       "ldr x21, [x17, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "fmla v14.4s, v29.4s, v1.s[3]\n"
+      "ldr x20, [x17, #0x18]\n"
       "fmla v18.4s, v29.4s, v2.s[3]\n"
       "fmla v22.4s, v29.4s, v3.s[3]\n"
-      "ldr x20, [x17, #0x18]\n"
       "fmla v26.4s, v29.4s, v4.s[3]\n"
       "ldr d6, [x17, #0x0]\n"
       "fmla v11.4s, v28.4s, v0.s[3]\n"
@@ -2501,8 +2499,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr s31, [x10], #0x4\n"
       "ldr s30, [x9], #0x4\n"
       "ldr q29, [x17, #0x0]\n"
-      "ldr q28, [x17, #0x10]\n"
       "fmla v8.4s, v29.4s, v2.s[0]\n"
+      "ldr q28, [x17, #0x10]\n"
       "fmla v12.4s, v29.4s, v1.s[0]\n"
       "fmla v16.4s, v29.4s, v0.s[0]\n"
       "fmla v20.4s, v29.4s, v31.s[0]\n"
@@ -2532,40 +2530,40 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "cmp x15, x20\n"
       "bne 147b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "add x25, x16, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
+      "prfm pstl1keep, [x16, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 155f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v28.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v28.4s\n"
+      "fmin v9.4s, v9.4s, v28.4s\n"
+      "fmin v10.4s, v10.4s, v28.4s\n"
+      "fmin v11.4s, v11.4s, v28.4s\n"
+      "fmin v12.4s, v12.4s, v28.4s\n"
+      "fmin v13.4s, v13.4s, v28.4s\n"
+      "fmin v14.4s, v14.4s, v28.4s\n"
+      "fmin v15.4s, v15.4s, v28.4s\n"
+      "fmin v16.4s, v16.4s, v28.4s\n"
+      "fmin v17.4s, v17.4s, v28.4s\n"
+      "fmin v18.4s, v18.4s, v28.4s\n"
+      "fmin v19.4s, v19.4s, v28.4s\n"
+      "fmin v20.4s, v20.4s, v28.4s\n"
+      "fmin v21.4s, v21.4s, v28.4s\n"
+      "fmin v22.4s, v22.4s, v28.4s\n"
+      "fmin v23.4s, v23.4s, v28.4s\n"
+      "fmin v24.4s, v24.4s, v28.4s\n"
+      "fmin v25.4s, v25.4s, v28.4s\n"
+      "fmin v26.4s, v26.4s, v28.4s\n"
+      "fmin v27.4s, v27.4s, v28.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v29.4s }, [x21]\n"
       "ld1r { v28.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v29.4s\n"
-      "fmin v9.4s, v9.4s, v29.4s\n"
-      "fmin v10.4s, v10.4s, v29.4s\n"
-      "fmin v11.4s, v11.4s, v29.4s\n"
-      "fmin v12.4s, v12.4s, v29.4s\n"
-      "fmin v13.4s, v13.4s, v29.4s\n"
-      "fmin v14.4s, v14.4s, v29.4s\n"
-      "fmin v15.4s, v15.4s, v29.4s\n"
-      "fmin v16.4s, v16.4s, v29.4s\n"
-      "fmin v17.4s, v17.4s, v29.4s\n"
-      "fmin v18.4s, v18.4s, v29.4s\n"
-      "fmin v19.4s, v19.4s, v29.4s\n"
-      "fmin v20.4s, v20.4s, v29.4s\n"
-      "fmin v21.4s, v21.4s, v29.4s\n"
-      "fmin v22.4s, v22.4s, v29.4s\n"
-      "fmin v23.4s, v23.4s, v29.4s\n"
-      "fmin v24.4s, v24.4s, v29.4s\n"
-      "fmin v25.4s, v25.4s, v29.4s\n"
-      "fmin v26.4s, v26.4s, v29.4s\n"
-      "fmin v27.4s, v27.4s, v29.4s\n"
       "fmax v8.4s, v8.4s, v28.4s\n"
       "fmax v9.4s, v9.4s, v28.4s\n"
       "fmax v10.4s, v10.4s, v28.4s\n"
@@ -2592,111 +2590,111 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "tbz x8, #3, 159f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
       "st1 { v9.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
       "tbz x8, #2, 157f\n"
       "st1 { v10.4s }, [x16], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
       "tbz x8, #1, 156f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x8, #0, 163f\n"
       "st1 { v11.s }[2], [x16]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
       "b 163f\n"
       "156:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x8, #0, 163f\n"
       "str s11, [x16, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
       "b 163f\n"
       "157:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x8, #1, 158f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x8, #0, 163f\n"
       "st1 { v10.s }[2], [x16]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
       "b 163f\n"
       "158:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x8, #0, 163f\n"
       "str s10, [x16, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
       "b 163f\n"
       "159:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x8, #2, 161f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x8, #1, 160f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x8, #0, 163f\n"
       "st1 { v9.s }[2], [x16]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 163f\n"
       "160:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x8, #0, 163f\n"
       "str s9, [x16, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 163f\n"
       "161:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x8, #1, 162f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x8, #0, 163f\n"
       "st1 { v8.s }[2], [x16]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 163f\n"
       "162:"  // Height 5: Partial direct writeback: partial_1_0
       "str s8, [x16, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "163:"  // Height 5: Partial direct writeback: Done
       "b 165f\n"
       "164:"  // Height 5: Full writeback
@@ -2705,22 +2703,22 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "165:"  // Height 5: Writeback done
       "subs x8, x8, #0x10\n"
       "bgt 134b\n"
@@ -2728,24 +2726,23 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "166:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "mov x20, #0x18\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "ldr x7, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x7, %x[bias]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "madd x20, x21, x20, x16\n"
       "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x16, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "167:"  // Height 6: Column loop
       "cbz x7, 168f\n"
       "ldr q8, [x7, #0x0]\n"
-      "ldr q9, [x7, #0x10]\n"
-      "ldr q10, [x7, #0x20]\n"
-      "ldr q11, [x7, #0x30]\n"
       "mov v12.16b, v8.16b\n"
+      "ldr q9, [x7, #0x10]\n"
       "mov v13.16b, v9.16b\n"
-      "add x7, x7, #0x40\n"
+      "ldr q10, [x7, #0x20]\n"
       "mov v14.16b, v10.16b\n"
+      "ldr q11, [x7, #0x30]\n"
       "mov v15.16b, v11.16b\n"
       "mov v16.16b, v8.16b\n"
+      "add x7, x7, #0x40\n"
       "mov v17.16b, v9.16b\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
@@ -2765,145 +2762,145 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "168:"  // Height 6: no bias
       "tbz %x[flags], #0, 178f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x26, x16, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x16, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 177f\n"
       "tbz x8, #3, 172f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v9.4s }, [x16], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x8, #2, 170f\n"
       "ld1 { v10.4s }, [x16], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x8, #1, 169f\n"
       "ldr d11, [x16], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d15, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x8, #0, 176f\n"
       "ld1 { v11.s }[2], [x16]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 176f\n"
       "169:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x8, #0, 176f\n"
       "ldr s11, [x16, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 176f\n"
       "170:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x8, #1, 171f\n"
       "ldr d10, [x16], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d14, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x8, #0, 176f\n"
       "ld1 { v10.s }[2], [x16]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 176f\n"
       "171:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x8, #0, 176f\n"
       "ldr s10, [x16, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 176f\n"
       "172:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x8, #2, 174f\n"
       "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x8, #1, 173f\n"
       "ldr d9, [x16], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d13, [x26], #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x8, #0, 176f\n"
       "ld1 { v9.s }[2], [x16]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 176f\n"
       "173:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x8, #0, 176f\n"
       "ldr s9, [x16, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 176f\n"
       "174:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x8, #1, 175f\n"
       "ldr d8, [x16], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d12, [x26], #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x8, #0, 176f\n"
       "ld1 { v8.s }[2], [x16]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 176f\n"
       "175:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s8, [x16, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s12, [x26, #0x0]\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "176:"  // Height 6: Partial accumulate: Done
       "sub x16, x16, x20\n"
       "b 179f\n"
@@ -2912,26 +2909,26 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr q9, [x16, #0x10]\n"
       "ldr q10, [x16, #0x20]\n"
       "ldr q11, [x16, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "b 179f\n"
       "178:"  // Height 6: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2962,8 +2959,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "mov x15, #0x0\n"
       "180:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 181f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -3016,146 +3013,146 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "fmla v28.4s, v6.4s, v5.s[0]\n"
       "ldr d6, [x17, #0x20]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
-      "add x10, x10, #0x10\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
-      "add x9, x9, #0x10\n"
+      "ldr x21, [x17, #0x48]\n"
       "fmla v17.4s, v7.4s, v2.s[0]\n"
-      "mov v6.d[1], x21\n"
+      "add x10, x10, #0x10\n"
       "fmla v21.4s, v7.4s, v3.s[0]\n"
-      "ldr x21, [x17, #0x48]\n"
+      "add x9, x9, #0x10\n"
       "fmla v25.4s, v7.4s, v4.s[0]\n"
       "add x28, x28, #0x10\n"
       "fmla v29.4s, v7.4s, v5.s[0]\n"
       "ldr d7, [x17, #0x30]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.4s, v6.4s, v0.s[0]\n"
-      "ldr x27, [x13, #0x8]\n"
       "fmla v14.4s, v6.4s, v1.s[0]\n"
-      "ldr x26, [x12, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x58]\n"
       "fmla v18.4s, v6.4s, v2.s[0]\n"
+      "ldr x27, [x13, #0x8]\n"
       "fmla v22.4s, v6.4s, v3.s[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x26, [x12, #0x8]\n"
       "fmla v26.4s, v6.4s, v4.s[0]\n"
       "ldr x25, [x11, #0x8]\n"
       "fmla v30.4s, v6.4s, v5.s[0]\n"
       "ldr d6, [x17, #0x40]\n"
       "fmla v11.4s, v7.4s, v0.s[0]\n"
-      "ldr x24, [x10, #0x8]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.4s, v7.4s, v1.s[0]\n"
-      "ldr x23, [x9, #0x8]\n"
+      "ldr x21, [x17, #0x68]\n"
       "fmla v19.4s, v7.4s, v2.s[0]\n"
-      "mov v6.d[1], x21\n"
+      "ldr x24, [x10, #0x8]\n"
       "fmla v23.4s, v7.4s, v3.s[0]\n"
-      "ldr x21, [x17, #0x68]\n"
+      "ldr x23, [x9, #0x8]\n"
       "fmla v27.4s, v7.4s, v4.s[0]\n"
       "ldr x22, [x28, #0x8]\n"
       "fmla v31.4s, v7.4s, v5.s[0]\n"
       "ldr d7, [x17, #0x50]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.4s, v6.4s, v0.s[1]\n"
-      "sub x14, x14, #0x4\n"
       "fmla v12.4s, v6.4s, v1.s[1]\n"
-      "cmp x14, #0x8\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x78]\n"
       "fmla v16.4s, v6.4s, v2.s[1]\n"
+      "sub x14, x14, #0x4\n"
       "fmla v20.4s, v6.4s, v3.s[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "cmp x14, #0x8\n"
       "fmla v24.4s, v6.4s, v4.s[1]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v28.4s, v6.4s, v5.s[1]\n"
       "ldr d6, [x17, #0x60]\n"
       "fmla v9.4s, v7.4s, v0.s[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.4s, v7.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x17, #0x88]\n"
       "fmla v17.4s, v7.4s, v2.s[1]\n"
-      "mov v6.d[1], x21\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "fmla v21.4s, v7.4s, v3.s[1]\n"
-      "ldr x21, [x17, #0x88]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "fmla v25.4s, v7.4s, v4.s[1]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       "fmla v29.4s, v7.4s, v5.s[1]\n"
       "ldr d7, [x17, #0x70]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.4s, v6.4s, v0.s[1]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v14.4s, v6.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0x98]\n"
       "fmla v18.4s, v6.4s, v2.s[1]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "fmla v22.4s, v6.4s, v3.s[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v26.4s, v6.4s, v4.s[1]\n"
       "fmla v30.4s, v6.4s, v5.s[1]\n"
       "ldr d6, [x17, #0x80]\n"
       "fmla v11.4s, v7.4s, v0.s[1]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.4s, v7.4s, v1.s[1]\n"
+      "ldr x21, [x17, #0xa8]\n"
       "fmla v19.4s, v7.4s, v2.s[1]\n"
-      "mov v6.d[1], x21\n"
       "fmla v23.4s, v7.4s, v3.s[1]\n"
-      "ldr x21, [x17, #0xa8]\n"
       "fmla v27.4s, v7.4s, v4.s[1]\n"
       "fmla v31.4s, v7.4s, v5.s[1]\n"
       "ldr d7, [x17, #0x90]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.4s, v6.4s, v0.s[2]\n"
       "fmla v12.4s, v6.4s, v1.s[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0xb8]\n"
       "fmla v16.4s, v6.4s, v2.s[2]\n"
       "fmla v20.4s, v6.4s, v3.s[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
       "fmla v24.4s, v6.4s, v4.s[2]\n"
       "fmla v28.4s, v6.4s, v5.s[2]\n"
       "ldr d6, [x17, #0xa0]\n"
       "fmla v9.4s, v7.4s, v0.s[2]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.4s, v7.4s, v1.s[2]\n"
+      "ldr x21, [x17, #0xc8]\n"
       "fmla v17.4s, v7.4s, v2.s[2]\n"
-      "mov v6.d[1], x21\n"
       "fmla v21.4s, v7.4s, v3.s[2]\n"
-      "ldr x21, [x17, #0xc8]\n"
       "fmla v25.4s, v7.4s, v4.s[2]\n"
       "fmla v29.4s, v7.4s, v5.s[2]\n"
       "ldr d7, [x17, #0xb0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v10.4s, v6.4s, v0.s[2]\n"
       "fmla v14.4s, v6.4s, v1.s[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0xd8]\n"
       "fmla v18.4s, v6.4s, v2.s[2]\n"
       "fmla v22.4s, v6.4s, v3.s[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
       "fmla v26.4s, v6.4s, v4.s[2]\n"
       "fmla v30.4s, v6.4s, v5.s[2]\n"
       "ldr d6, [x17, #0xc0]\n"
       "fmla v11.4s, v7.4s, v0.s[2]\n"
+      "mov v6.d[1], x21\n"
       "fmla v15.4s, v7.4s, v1.s[2]\n"
+      "ldr x21, [x17, #0xe8]\n"
       "fmla v19.4s, v7.4s, v2.s[2]\n"
-      "mov v6.d[1], x21\n"
       "fmla v23.4s, v7.4s, v3.s[2]\n"
-      "ldr x21, [x17, #0xe8]\n"
       "fmla v27.4s, v7.4s, v4.s[2]\n"
       "fmla v31.4s, v7.4s, v5.s[2]\n"
       "ldr d7, [x17, #0xd0]\n"
+      "mov v7.d[1], x20\n"
       "fmla v8.4s, v6.4s, v0.s[3]\n"
       "fmla v12.4s, v6.4s, v1.s[3]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x17, #0xf8]\n"
       "fmla v16.4s, v6.4s, v2.s[3]\n"
       "fmla v20.4s, v6.4s, v3.s[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
       "fmla v24.4s, v6.4s, v4.s[3]\n"
       "fmla v28.4s, v6.4s, v5.s[3]\n"
       "ldr d6, [x17, #0xe0]\n"
       "fmla v9.4s, v7.4s, v0.s[3]\n"
+      "mov v6.d[1], x21\n"
       "fmla v13.4s, v7.4s, v1.s[3]\n"
       "fmla v17.4s, v7.4s, v2.s[3]\n"
-      "mov v6.d[1], x21\n"
       "fmla v21.4s, v7.4s, v3.s[3]\n"
       "fmla v25.4s, v7.4s, v4.s[3]\n"
       "fmla v29.4s, v7.4s, v5.s[3]\n"
       "ldr d7, [x17, #0xf0]\n"
+      "mov v7.d[1], x20\n"
       "add x17, x17, #0x100\n"
       "fmla v10.4s, v6.4s, v0.s[3]\n"
-      "fmla v14.4s, v6.4s, v1.s[3]\n"
       "ldr x21, [x17, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "fmla v14.4s, v6.4s, v1.s[3]\n"
+      "ldr x20, [x17, #0x18]\n"
       "fmla v18.4s, v6.4s, v2.s[3]\n"
       "fmla v22.4s, v6.4s, v3.s[3]\n"
-      "ldr x20, [x17, #0x18]\n"
       "fmla v26.4s, v6.4s, v4.s[3]\n"
       "fmla v30.4s, v6.4s, v5.s[3]\n"
       "ldr d6, [x17, #0x0]\n"
@@ -3317,8 +3314,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "ldr s3, [x9], #0x4\n"
       "ldr s2, [x28], #0x4\n"
       "ldr q1, [x17, #0x0]\n"
-      "ldr q0, [x17, #0x10]\n"
       "fmla v8.4s, v1.4s, v7.s[0]\n"
+      "ldr q0, [x17, #0x10]\n"
       "fmla v12.4s, v1.4s, v6.s[0]\n"
       "fmla v16.4s, v1.4s, v5.s[0]\n"
       "fmla v20.4s, v1.4s, v4.s[0]\n"
@@ -3352,46 +3349,46 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "cmp x15, x20\n"
       "bne 180b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x26, x16, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "add x25, x16, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
+      "prfm pstl1keep, [x16, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbz %x[flags], #1, 188f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "fmin v8.4s, v8.4s, v0.4s\n"
+      "fmin v9.4s, v9.4s, v0.4s\n"
+      "fmin v10.4s, v10.4s, v0.4s\n"
+      "fmin v11.4s, v11.4s, v0.4s\n"
+      "fmin v12.4s, v12.4s, v0.4s\n"
+      "fmin v13.4s, v13.4s, v0.4s\n"
+      "fmin v14.4s, v14.4s, v0.4s\n"
+      "fmin v15.4s, v15.4s, v0.4s\n"
+      "fmin v16.4s, v16.4s, v0.4s\n"
+      "fmin v17.4s, v17.4s, v0.4s\n"
+      "fmin v18.4s, v18.4s, v0.4s\n"
+      "fmin v19.4s, v19.4s, v0.4s\n"
+      "fmin v20.4s, v20.4s, v0.4s\n"
+      "fmin v21.4s, v21.4s, v0.4s\n"
+      "fmin v22.4s, v22.4s, v0.4s\n"
+      "fmin v23.4s, v23.4s, v0.4s\n"
+      "fmin v24.4s, v24.4s, v0.4s\n"
+      "fmin v25.4s, v25.4s, v0.4s\n"
+      "fmin v26.4s, v26.4s, v0.4s\n"
+      "fmin v27.4s, v27.4s, v0.4s\n"
+      "fmin v28.4s, v28.4s, v0.4s\n"
+      "fmin v29.4s, v29.4s, v0.4s\n"
+      "fmin v30.4s, v30.4s, v0.4s\n"
+      "fmin v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
-      "fmin v8.4s, v8.4s, v1.4s\n"
-      "fmin v9.4s, v9.4s, v1.4s\n"
-      "fmin v10.4s, v10.4s, v1.4s\n"
-      "fmin v11.4s, v11.4s, v1.4s\n"
-      "fmin v12.4s, v12.4s, v1.4s\n"
-      "fmin v13.4s, v13.4s, v1.4s\n"
-      "fmin v14.4s, v14.4s, v1.4s\n"
-      "fmin v15.4s, v15.4s, v1.4s\n"
-      "fmin v16.4s, v16.4s, v1.4s\n"
-      "fmin v17.4s, v17.4s, v1.4s\n"
-      "fmin v18.4s, v18.4s, v1.4s\n"
-      "fmin v19.4s, v19.4s, v1.4s\n"
-      "fmin v20.4s, v20.4s, v1.4s\n"
-      "fmin v21.4s, v21.4s, v1.4s\n"
-      "fmin v22.4s, v22.4s, v1.4s\n"
-      "fmin v23.4s, v23.4s, v1.4s\n"
-      "fmin v24.4s, v24.4s, v1.4s\n"
-      "fmin v25.4s, v25.4s, v1.4s\n"
-      "fmin v26.4s, v26.4s, v1.4s\n"
-      "fmin v27.4s, v27.4s, v1.4s\n"
-      "fmin v28.4s, v28.4s, v1.4s\n"
-      "fmin v29.4s, v29.4s, v1.4s\n"
-      "fmin v30.4s, v30.4s, v1.4s\n"
-      "fmin v31.4s, v31.4s, v1.4s\n"
       "fmax v8.4s, v8.4s, v0.4s\n"
       "fmax v9.4s, v9.4s, v0.4s\n"
       "fmax v10.4s, v10.4s, v0.4s\n"
@@ -3422,127 +3419,127 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "tbz x8, #3, 192f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
       "st1 { v9.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
-      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
+      "st1 { v29.4s }, [x21], #0x10\n"
       "tbz x8, #2, 190f\n"
       "st1 { v10.4s }, [x16], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
-      "st1 { v30.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v30.4s }, [x21], #0x10\n"
       "tbz x8, #1, 189f\n"
       "str d11, [x16], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x8, #0, 196f\n"
       "st1 { v11.s }[2], [x16]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
-      "st1 { v31.s }[2], [x22]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v31.s }[2], [x21]\n"
       "b 196f\n"
       "189:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x8, #0, 196f\n"
       "str s11, [x16, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
-      "str s31, [x22, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
+      "str s31, [x21, #0x0]\n"
       "b 196f\n"
       "190:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x8, #1, 191f\n"
       "str d10, [x16], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x8, #0, 196f\n"
       "st1 { v10.s }[2], [x16]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
-      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v30.s }[2], [x21]\n"
       "b 196f\n"
       "191:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x8, #0, 196f\n"
       "str s10, [x16, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
-      "str s30, [x22, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
+      "str s30, [x21, #0x0]\n"
       "b 196f\n"
       "192:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x8, #2, 194f\n"
       "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
       "tbz x8, #1, 193f\n"
       "str d9, [x16], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x8, #0, 196f\n"
       "st1 { v9.s }[2], [x16]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
-      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v29.s }[2], [x21]\n"
       "b 196f\n"
       "193:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x8, #0, 196f\n"
       "str s9, [x16, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
-      "str s29, [x22, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
+      "str s29, [x21, #0x0]\n"
       "b 196f\n"
       "194:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x8, #1, 195f\n"
       "str d8, [x16], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x8, #0, 196f\n"
       "st1 { v8.s }[2], [x16]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
-      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v28.s }[2], [x21]\n"
       "b 196f\n"
       "195:"  // Height 6: Partial direct writeback: partial_1_0
       "str s8, [x16, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
-      "str s28, [x22, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
+      "str s28, [x21, #0x0]\n"
       "196:"  // Height 6: Partial direct writeback: Done
       "b 198f\n"
       "197:"  // Height 6: Full writeback
@@ -3551,26 +3548,26 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "str q10, [x16, #0x20]\n"
       "str q11, [x16, #0x30]\n"
       "add x16, x16, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q28, [x22, #0x0]\n"
-      "str q29, [x22, #0x10]\n"
-      "str q30, [x22, #0x20]\n"
-      "str q31, [x22, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
+      "str q28, [x21, #0x0]\n"
+      "str q29, [x21, #0x10]\n"
+      "str q30, [x21, #0x20]\n"
+      "str q31, [x21, #0x30]\n"
       "198:"  // Height 6: Writeback done
       "subs x8, x8, #0x10\n"
       "bgt 167b\n"
@@ -3586,8 +3583,8 @@ void a64_hybrid_fp32_mla_6x16_a55 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "200:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp
index e411da6874..6c51c0ff3c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp
@@ -47,19 +47,18 @@ void a64_hybrid_fp32_mla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void a64_hybrid_fp32_mla_6x16 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -103,10 +101,10 @@ void a64_hybrid_fp32_mla_6x16 (
       "cmp %x[M], #0x2\n"
       "bgt 67f\n"
       "beq 34f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x12, 3f\n"
       "ldr q8, [x12, #0x0]\n"
@@ -189,8 +187,8 @@ void a64_hybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "15:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 16f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -214,10 +212,6 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr q17, [x10, #0x20]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "sub x27, x27, #0x4\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "ldr q17, [x10, #0x40]\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
@@ -242,21 +236,22 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr q17, [x10, #0xe0]\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x4\n"
+      "add x26, x26, #0x10\n"
       "fmla v10.4s, v17.4s, v0.s[3]\n"
-      "ldr q6, [x10, #0x0]\n"
       "fmla v11.4s, v16.4s, v0.s[3]\n"
       "ldr q0, [x26, #0x0]\n"
+      "cmp x27, #0x8\n"
+      "add x10, x10, #0x100\n"
+      "ldr q6, [x10, #0x0]\n"
       "ldr q7, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 18b\n"
       "19:"  // Height 1: Multiply loop: Single iteration only
       "fmla v8.4s, v6.4s, v0.s[0]\n"
       "ldr q17, [x10, #0x20]\n"
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "ldr q17, [x10, #0x40]\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
@@ -281,23 +276,26 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr q17, [x10, #0xe0]\n"
       "fmla v9.4s, v16.4s, v0.s[3]\n"
       "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "fmla v10.4s, v17.4s, v0.s[3]\n"
       "fmla v11.4s, v16.4s, v0.s[3]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "20:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 22f\n"
       "21:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x26], #0x4\n"
-      "ldr q17, [x10, #0x0]\n"
+      "ldr q16, [x10, #0x0]\n"
+      "fmla v8.4s, v16.4s, v18.s[0]\n"
       "sub x27, x27, #0x1\n"
-      "ldr q16, [x10, #0x10]\n"
-      "fmla v8.4s, v17.4s, v18.s[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      "fmla v9.4s, v16.4s, v18.s[0]\n"
+      "ldr q17, [x10, #0x10]\n"
+      "ldr q16, [x10, #0x20]\n"
+      "fmla v9.4s, v17.4s, v18.s[0]\n"
+      "fmla v10.4s, v16.4s, v18.s[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      "fmla v10.4s, v17.4s, v18.s[0]\n"
       "fmla v11.4s, v16.4s, v18.s[0]\n"
+      "add x10, x10, #0x40\n"
       "cbnz x27, 21b\n"
       "22:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -306,9 +304,9 @@ void a64_hybrid_fp32_mla_6x16 (
       "bne 15b\n"
       "prfm pstl1keep, [x9, #0x0]\n"
       "tbz %x[flags], #1, 23f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v17.4s\n"
       "fmin v9.4s, v9.4s, v17.4s\n"
@@ -378,95 +376,95 @@ void a64_hybrid_fp32_mla_6x16 (
       "bgt 2b\n"
       "b 200f\n"
       "34:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "35:"  // Height 2: Column loop
       "cbz x12, 36f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "add x12, x12, #0x40\n"
       "b 47f\n"
       "36:"  // Height 2: no bias
       "tbz %x[flags], #0, 46f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "bge 45f\n"
       "tbz x11, #3, 40f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
       "tbz x11, #2, 38f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
       "tbz x11, #1, 37f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x11, #0, 44f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v15.s }[2], [x25]\n"
       "b 44f\n"
       "37:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 44f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
       "b 44f\n"
       "38:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x11, #1, 39f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x11, #0, 44f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v14.s }[2], [x25]\n"
       "b 44f\n"
       "39:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 44f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
       "b 44f\n"
       "40:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x11, #2, 42f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "tbz x11, #1, 41f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x11, #0, 44f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v13.s }[2], [x25]\n"
       "b 44f\n"
       "41:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 44f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
       "b 44f\n"
       "42:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x11, #1, 43f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x11, #0, 44f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v12.s }[2], [x25]\n"
       "b 44f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
       "44:"  // Height 2: Partial accumulate: Done
       "sub x9, x9, x20\n"
@@ -476,10 +474,10 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "b 47f\n"
       "46:"  // Height 2: no accumulate
       "movi v8.16b, #0x0\n"
@@ -494,8 +492,8 @@ void a64_hybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "48:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 49f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -527,22 +525,22 @@ void a64_hybrid_fp32_mla_6x16 (
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "fmla v14.4s, v17.4s, v1.s[0]\n"
       "ldr q17, [x10, #0x40]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x25, x25, #0x10\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
       "fmla v15.4s, v16.4s, v1.s[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "cmp x27, #0x8\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
       "fmla v12.4s, v17.4s, v1.s[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v9.4s, v16.4s, v0.s[1]\n"
       "fmla v13.4s, v16.4s, v1.s[1]\n"
       "ldr q16, [x10, #0x70]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.4s, v17.4s, v0.s[1]\n"
       "fmla v14.4s, v17.4s, v1.s[1]\n"
       "ldr q17, [x10, #0x80]\n"
@@ -586,18 +584,18 @@ void a64_hybrid_fp32_mla_6x16 (
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.4s, v17.4s, v0.s[0]\n"
       "fmla v14.4s, v17.4s, v1.s[0]\n"
       "ldr q17, [x10, #0x40]\n"
+      "sub x27, x27, #0x4\n"
       "fmla v11.4s, v16.4s, v0.s[0]\n"
       "fmla v15.4s, v16.4s, v1.s[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v8.4s, v17.4s, v0.s[1]\n"
       "fmla v12.4s, v17.4s, v1.s[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v9.4s, v16.4s, v0.s[1]\n"
       "fmla v13.4s, v16.4s, v1.s[1]\n"
       "ldr q16, [x10, #0x70]\n"
@@ -644,9 +642,9 @@ void a64_hybrid_fp32_mla_6x16 (
       "fmla v9.4s, v16.4s, v19.s[0]\n"
       "fmla v13.4s, v16.4s, v18.s[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       "fmla v10.4s, v17.4s, v19.s[0]\n"
       "fmla v14.4s, v17.4s, v18.s[0]\n"
+      "add x10, x10, #0x40\n"
       "fmla v11.4s, v16.4s, v19.s[0]\n"
       "fmla v15.4s, v16.4s, v18.s[0]\n"
       "cbnz x27, 54b\n"
@@ -656,13 +654,13 @@ void a64_hybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 48b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "tbz %x[flags], #1, 56f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v17.4s\n"
       "fmin v9.4s, v9.4s, v17.4s\n"
@@ -686,63 +684,63 @@ void a64_hybrid_fp32_mla_6x16 (
       "tbz x11, #3, 60f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
       "tbz x11, #2, 58f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
       "tbz x11, #1, 57f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
+      "str d15, [x25], #0x8\n"
       "tbz x11, #0, 64f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
+      "st1 { v15.s }[2], [x25]\n"
       "b 64f\n"
       "57:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x11, #0, 64f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
       "b 64f\n"
       "58:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x11, #1, 59f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
+      "str d14, [x25], #0x8\n"
       "tbz x11, #0, 64f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
+      "st1 { v14.s }[2], [x25]\n"
       "b 64f\n"
       "59:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x11, #0, 64f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
       "b 64f\n"
       "60:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x11, #2, 62f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
       "tbz x11, #1, 61f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
+      "str d13, [x25], #0x8\n"
       "tbz x11, #0, 64f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
+      "st1 { v13.s }[2], [x25]\n"
       "b 64f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x11, #0, 64f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
       "b 64f\n"
       "62:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x11, #1, 63f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
+      "str d12, [x25], #0x8\n"
       "tbz x11, #0, 64f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
+      "st1 { v12.s }[2], [x25]\n"
       "b 64f\n"
       "63:"  // Height 2: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
       "64:"  // Height 2: Partial direct writeback: Done
       "b 66f\n"
       "65:"  // Height 2: Full writeback
@@ -751,126 +749,126 @@ void a64_hybrid_fp32_mla_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
       "66:"  // Height 2: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 35b\n"
       "b 200f\n"
       "67:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "68:"  // Height 3: Column loop
       "cbz x12, 69f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
       "b 80f\n"
       "69:"  // Height 3: no bias
       "tbz %x[flags], #0, 79f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "bge 78f\n"
       "tbz x11, #3, 73f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "tbz x11, #2, 71f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
       "tbz x11, #1, 70f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x11, #0, 77f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
       "b 77f\n"
       "70:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 77f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
       "b 77f\n"
       "71:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x11, #1, 72f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x11, #0, 77f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
       "b 77f\n"
       "72:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 77f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
       "b 77f\n"
       "73:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x11, #2, 75f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
       "tbz x11, #1, 74f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x11, #0, 77f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
       "b 77f\n"
       "74:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 77f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
       "b 77f\n"
       "75:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x11, #1, 76f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
       "tbz x11, #0, 77f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
       "b 77f\n"
       "76:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
       "77:"  // Height 3: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 80f\n"
@@ -879,14 +877,14 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
       "b 80f\n"
       "79:"  // Height 3: no accumulate
       "movi v8.16b, #0x0\n"
@@ -905,8 +903,8 @@ void a64_hybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "81:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 82f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -946,18 +944,18 @@ void a64_hybrid_fp32_mla_6x16 (
       "fmla v17.4s, v7.4s, v2.s[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "add x24, x24, #0x10\n"
-      "cmp x27, #0x8\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.4s, v21.4s, v0.s[0]\n"
       "fmla v14.4s, v21.4s, v1.s[0]\n"
+      "cmp x27, #0x8\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v18.4s, v21.4s, v2.s[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v11.4s, v20.4s, v0.s[0]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v15.4s, v20.4s, v1.s[0]\n"
       "fmla v19.4s, v20.4s, v2.s[0]\n"
       "ldr q20, [x10, #0x50]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v8.4s, v21.4s, v0.s[1]\n"
       "fmla v12.4s, v21.4s, v1.s[1]\n"
       "fmla v16.4s, v21.4s, v2.s[1]\n"
@@ -1024,14 +1022,14 @@ void a64_hybrid_fp32_mla_6x16 (
       "fmla v17.4s, v7.4s, v2.s[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.4s, v21.4s, v0.s[0]\n"
       "fmla v14.4s, v21.4s, v1.s[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v18.4s, v21.4s, v2.s[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v11.4s, v20.4s, v0.s[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v15.4s, v20.4s, v1.s[0]\n"
       "fmla v19.4s, v20.4s, v2.s[0]\n"
       "ldr q20, [x10, #0x50]\n"
@@ -1090,9 +1088,9 @@ void a64_hybrid_fp32_mla_6x16 (
       "sub x27, x27, #0x1\n"
       "ldr s22, [x24], #0x4\n"
       "ldr q21, [x10, #0x0]\n"
-      "ldr q20, [x10, #0x10]\n"
       "fmla v8.4s, v21.4s, v24.s[0]\n"
       "fmla v12.4s, v21.4s, v23.s[0]\n"
+      "ldr q20, [x10, #0x10]\n"
       "fmla v16.4s, v21.4s, v22.s[0]\n"
       "ldr q21, [x10, #0x20]\n"
       "fmla v9.4s, v20.4s, v24.s[0]\n"
@@ -1113,15 +1111,15 @@ void a64_hybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 81b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "tbz %x[flags], #1, 89f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v21.4s }, [x21]\n"
       "ld1r { v20.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v21.4s\n"
       "fmin v9.4s, v9.4s, v21.4s\n"
@@ -1153,79 +1151,79 @@ void a64_hybrid_fp32_mla_6x16 (
       "tbz x11, #3, 93f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
       "tbz x11, #2, 91f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
       "tbz x11, #1, 90f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x11, #0, 97f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
       "b 97f\n"
       "90:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x11, #0, 97f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
       "b 97f\n"
       "91:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x11, #1, 92f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x11, #0, 97f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
       "b 97f\n"
       "92:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x11, #0, 97f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
       "b 97f\n"
       "93:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x11, #2, 95f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
       "tbz x11, #1, 94f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x11, #0, 97f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
       "b 97f\n"
       "94:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x11, #0, 97f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
       "b 97f\n"
       "95:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x11, #1, 96f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x11, #0, 97f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
       "b 97f\n"
       "96:"  // Height 3: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
       "97:"  // Height 3: Partial direct writeback: Done
       "b 99f\n"
       "98:"  // Height 3: Full writeback
@@ -1234,39 +1232,39 @@ void a64_hybrid_fp32_mla_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "99:"  // Height 3: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 68b\n"
       "b 200f\n"
       "100:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "101:"  // Height 4: Column loop
       "cbz x12, 102f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1274,111 +1272,111 @@ void a64_hybrid_fp32_mla_6x16 (
       "102:"  // Height 4: no bias
       "tbz %x[flags], #0, 112f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x23, x24, x20, LSL #2\n"
       "bge 111f\n"
       "tbz x11, #3, 106f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
       "tbz x11, #2, 104f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
       "tbz x11, #1, 103f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x11, #0, 110f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
       "b 110f\n"
       "103:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 110f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
       "b 110f\n"
       "104:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x11, #1, 105f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x11, #0, 110f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
       "b 110f\n"
       "105:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 110f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
       "b 110f\n"
       "106:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x11, #2, 108f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "tbz x11, #1, 107f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x11, #0, 110f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
       "b 110f\n"
       "107:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 110f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
       "b 110f\n"
       "108:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x11, #1, 109f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x11, #0, 110f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
       "b 110f\n"
       "109:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
       "110:"  // Height 4: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 113f\n"
@@ -1387,18 +1385,18 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "b 113f\n"
       "112:"  // Height 4: no accumulate
       "movi v8.16b, #0x0\n"
@@ -1421,8 +1419,8 @@ void a64_hybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "114:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 115f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1562,14 +1560,14 @@ void a64_hybrid_fp32_mla_6x16 (
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x4\n"
       "fmla v17.4s, v7.4s, v2.s[0]\n"
       "fmla v21.4s, v7.4s, v3.s[0]\n"
       "ldr q24, [x10, #0x30]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v10.4s, v25.4s, v0.s[0]\n"
       "fmla v14.4s, v25.4s, v1.s[0]\n"
-      "sub x27, x27, #0x4\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v18.4s, v25.4s, v2.s[0]\n"
       "fmla v22.4s, v25.4s, v3.s[0]\n"
@@ -1675,17 +1673,17 @@ void a64_hybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 114b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
-      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 122f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v25.4s }, [x21]\n"
       "ld1r { v24.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v25.4s\n"
       "fmin v9.4s, v9.4s, v25.4s\n"
@@ -1725,95 +1723,95 @@ void a64_hybrid_fp32_mla_6x16 (
       "tbz x11, #3, 126f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
       "tbz x11, #2, 124f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
       "tbz x11, #1, 123f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
       "tbz x11, #0, 130f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
       "b 130f\n"
       "123:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x11, #0, 130f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
       "b 130f\n"
       "124:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x11, #1, 125f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
       "tbz x11, #0, 130f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
       "b 130f\n"
       "125:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x11, #0, 130f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
       "b 130f\n"
       "126:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x11, #2, 128f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
       "tbz x11, #1, 127f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
       "tbz x11, #0, 130f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
       "b 130f\n"
       "127:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x11, #0, 130f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
       "b 130f\n"
       "128:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x11, #1, 129f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x11, #0, 130f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
       "b 130f\n"
       "129:"  // Height 4: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
       "130:"  // Height 4: Partial direct writeback: Done
       "b 132f\n"
       "131:"  // Height 4: Full writeback
@@ -1822,43 +1820,43 @@ void a64_hybrid_fp32_mla_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
       "132:"  // Height 4: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 101b\n"
       "b 200f\n"
       "133:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "134:"  // Height 5: Column loop
       "cbz x12, 135f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -1870,128 +1868,128 @@ void a64_hybrid_fp32_mla_6x16 (
       "135:"  // Height 5: no bias
       "tbz %x[flags], #0, 145f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 144f\n"
       "tbz x11, #3, 139f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x11, #2, 137f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
       "tbz x11, #1, 136f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x11, #0, 143f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
       "b 143f\n"
       "136:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 143f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
       "b 143f\n"
       "137:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x11, #1, 138f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x11, #0, 143f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
       "b 143f\n"
       "138:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 143f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
       "b 143f\n"
       "139:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x11, #2, 141f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "tbz x11, #1, 140f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x11, #0, 143f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 143f\n"
       "140:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 143f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "b 143f\n"
       "141:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x11, #1, 142f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x11, #0, 143f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
       "b 143f\n"
       "142:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
       "143:"  // Height 5: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 146f\n"
@@ -2000,22 +1998,22 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
       "b 146f\n"
       "145:"  // Height 5: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2042,8 +2040,8 @@ void a64_hybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "147:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 148f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2209,12 +2207,12 @@ void a64_hybrid_fp32_mla_6x16 (
       "add x22, x22, #0x10\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "fmla v17.4s, v7.4s, v2.s[0]\n"
+      "sub x27, x27, #0x4\n"
       "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v21.4s, v7.4s, v3.s[0]\n"
       "fmla v25.4s, v7.4s, v4.s[0]\n"
       "ldr q28, [x10, #0x30]\n"
-      "sub x27, x27, #0x4\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v10.4s, v29.4s, v0.s[0]\n"
       "fmla v14.4s, v29.4s, v1.s[0]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
@@ -2311,9 +2309,9 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr s31, [x23], #0x4\n"
       "ldr s30, [x22], #0x4\n"
       "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
       "fmla v8.4s, v29.4s, v2.s[0]\n"
       "fmla v12.4s, v29.4s, v1.s[0]\n"
+      "ldr q28, [x10, #0x10]\n"
       "fmla v16.4s, v29.4s, v0.s[0]\n"
       "fmla v20.4s, v29.4s, v31.s[0]\n"
       "fmla v24.4s, v29.4s, v30.s[0]\n"
@@ -2342,19 +2340,19 @@ void a64_hybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 147b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
-      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 155f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v29.4s }, [x21]\n"
       "ld1r { v28.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v29.4s\n"
       "fmin v9.4s, v9.4s, v29.4s\n"
@@ -2402,111 +2400,111 @@ void a64_hybrid_fp32_mla_6x16 (
       "tbz x11, #3, 159f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
       "tbz x11, #2, 157f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
       "tbz x11, #1, 156f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x11, #0, 163f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
       "b 163f\n"
       "156:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x11, #0, 163f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
       "b 163f\n"
       "157:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x11, #1, 158f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x11, #0, 163f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
       "b 163f\n"
       "158:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x11, #0, 163f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
       "b 163f\n"
       "159:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x11, #2, 161f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x11, #1, 160f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x11, #0, 163f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 163f\n"
       "160:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x11, #0, 163f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 163f\n"
       "161:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x11, #1, 162f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x11, #0, 163f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 163f\n"
       "162:"  // Height 5: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "163:"  // Height 5: Partial direct writeback: Done
       "b 165f\n"
       "164:"  // Height 5: Full writeback
@@ -2515,51 +2513,50 @@ void a64_hybrid_fp32_mla_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "165:"  // Height 5: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 134b\n"
       "b 200f\n"
       "166:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "167:"  // Height 6: Column loop
       "cbz x12, 168f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "mov v12.16b, v8.16b\n"
       "mov v13.16b, v9.16b\n"
-      "mov v16.16b, v8.16b\n"
-      "mov v17.16b, v9.16b\n"
-      "mov v20.16b, v8.16b\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "mov v14.16b, v10.16b\n"
       "mov v15.16b, v11.16b\n"
+      "mov v16.16b, v8.16b\n"
+      "mov v17.16b, v9.16b\n"
+      "add x12, x12, #0x40\n"
       "mov v18.16b, v10.16b\n"
       "mov v19.16b, v11.16b\n"
+      "mov v20.16b, v8.16b\n"
       "mov v21.16b, v9.16b\n"
       "mov v22.16b, v10.16b\n"
       "mov v23.16b, v11.16b\n"
@@ -2575,145 +2572,145 @@ void a64_hybrid_fp32_mla_6x16 (
       "168:"  // Height 6: no bias
       "tbz %x[flags], #0, 178f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 177f\n"
       "tbz x11, #3, 172f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x11, #2, 170f\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x11, #1, 169f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x11, #0, 176f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 176f\n"
       "169:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 176f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 176f\n"
       "170:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x11, #1, 171f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x11, #0, 176f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 176f\n"
       "171:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 176f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 176f\n"
       "172:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x11, #2, 174f\n"
       "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v16.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x11, #1, 173f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x11, #0, 176f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 176f\n"
       "173:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 176f\n"
       "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 176f\n"
       "174:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x11, #1, 175f\n"
       "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d16, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d16, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x11, #0, 176f\n"
       "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v16.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 176f\n"
       "175:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s16, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s16, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "176:"  // Height 6: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 179f\n"
@@ -2722,26 +2719,26 @@ void a64_hybrid_fp32_mla_6x16 (
       "ldr q9, [x9, #0x10]\n"
       "ldr q10, [x9, #0x20]\n"
       "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q16, [x25, #0x0]\n"
-      "ldr q17, [x25, #0x10]\n"
-      "ldr q18, [x25, #0x20]\n"
-      "ldr q19, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q24, [x23, #0x0]\n"
-      "ldr q25, [x23, #0x10]\n"
-      "ldr q26, [x23, #0x20]\n"
-      "ldr q27, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q16, [x24, #0x0]\n"
+      "ldr q17, [x24, #0x10]\n"
+      "ldr q18, [x24, #0x20]\n"
+      "ldr q19, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q24, [x22, #0x0]\n"
+      "ldr q25, [x22, #0x10]\n"
+      "ldr q26, [x22, #0x20]\n"
+      "ldr q27, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "b 179f\n"
       "178:"  // Height 6: no accumulate
       "movi v8.16b, #0x0\n"
@@ -2772,8 +2769,8 @@ void a64_hybrid_fp32_mla_6x16 (
       "mov x28, #0x0\n"
       "180:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 181f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2963,18 +2960,18 @@ void a64_hybrid_fp32_mla_6x16 (
       "fmla v9.4s, v7.4s, v0.s[0]\n"
       "fmla v13.4s, v7.4s, v1.s[0]\n"
       "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x4\n"
       "fmla v17.4s, v7.4s, v2.s[0]\n"
       "fmla v21.4s, v7.4s, v3.s[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v25.4s, v7.4s, v4.s[0]\n"
       "fmla v29.4s, v7.4s, v5.s[0]\n"
       "ldr q7, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v10.4s, v6.4s, v0.s[0]\n"
       "fmla v14.4s, v6.4s, v1.s[0]\n"
-      "sub x27, x27, #0x4\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       "prfm pldl1keep, [x22, #0x80]\n"
       "fmla v18.4s, v6.4s, v2.s[0]\n"
       "fmla v22.4s, v6.4s, v3.s[0]\n"
@@ -3118,21 +3115,21 @@ void a64_hybrid_fp32_mla_6x16 (
       "cmp x28, x20\n"
       "bne 180b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "add x21, x22, x20, LSL #2\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbz %x[flags], #1, 188f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v1.4s\n"
       "fmin v9.4s, v9.4s, v1.4s\n"
@@ -3188,127 +3185,127 @@ void a64_hybrid_fp32_mla_6x16 (
       "tbz x11, #3, 192f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
       "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v13.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
-      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v13.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
+      "st1 { v29.4s }, [x21], #0x10\n"
       "tbz x11, #2, 190f\n"
       "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
-      "st1 { v22.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
-      "st1 { v30.4s }, [x22], #0x10\n"
+      "st1 { v14.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v30.4s }, [x21], #0x10\n"
       "tbz x11, #1, 189f\n"
       "str d11, [x9], #0x8\n"
-      "str d15, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
-      "str d23, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d15, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
+      "str d23, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x11, #0, 196f\n"
       "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
-      "st1 { v23.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
-      "st1 { v31.s }[2], [x22]\n"
+      "st1 { v15.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v31.s }[2], [x21]\n"
       "b 196f\n"
       "189:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x11, #0, 196f\n"
       "str s11, [x9, #0x0]\n"
-      "str s15, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
-      "str s23, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
-      "str s31, [x22, #0x0]\n"
+      "str s15, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
+      "str s23, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
+      "str s31, [x21, #0x0]\n"
       "b 196f\n"
       "190:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x11, #1, 191f\n"
       "str d10, [x9], #0x8\n"
-      "str d14, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
-      "str d22, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d30, [x22], #0x8\n"
+      "str d14, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
+      "str d22, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d30, [x21], #0x8\n"
       "tbz x11, #0, 196f\n"
       "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
-      "st1 { v22.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
-      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v14.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v30.s }[2], [x21]\n"
       "b 196f\n"
       "191:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x11, #0, 196f\n"
       "str s10, [x9, #0x0]\n"
-      "str s14, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
-      "str s22, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
-      "str s30, [x22, #0x0]\n"
+      "str s14, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
+      "str s22, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
+      "str s30, [x21, #0x0]\n"
       "b 196f\n"
       "192:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x11, #2, 194f\n"
       "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x21], #0x10\n"
       "tbz x11, #1, 193f\n"
       "str d9, [x9], #0x8\n"
-      "str d13, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
-      "str d21, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
-      "str d29, [x22], #0x8\n"
+      "str d13, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
+      "str d21, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
+      "str d29, [x21], #0x8\n"
       "tbz x11, #0, 196f\n"
       "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
-      "st1 { v21.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
-      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v13.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v29.s }[2], [x21]\n"
       "b 196f\n"
       "193:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x11, #0, 196f\n"
       "str s9, [x9, #0x0]\n"
-      "str s13, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
-      "str s21, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
-      "str s29, [x22, #0x0]\n"
+      "str s13, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
+      "str s21, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
+      "str s29, [x21, #0x0]\n"
       "b 196f\n"
       "194:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x11, #1, 195f\n"
       "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x11, #0, 196f\n"
       "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
-      "st1 { v20.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
-      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v12.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v28.s }[2], [x21]\n"
       "b 196f\n"
       "195:"  // Height 6: Partial direct writeback: partial_1_0
       "str s8, [x9, #0x0]\n"
-      "str s12, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
-      "str s20, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
-      "str s28, [x22, #0x0]\n"
+      "str s12, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
+      "str s20, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
+      "str s28, [x21, #0x0]\n"
       "196:"  // Height 6: Partial direct writeback: Done
       "b 198f\n"
       "197:"  // Height 6: Full writeback
@@ -3317,26 +3314,26 @@ void a64_hybrid_fp32_mla_6x16 (
       "str q10, [x9, #0x20]\n"
       "str q11, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q12, [x26, #0x0]\n"
-      "str q13, [x26, #0x10]\n"
-      "str q14, [x26, #0x20]\n"
-      "str q15, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q21, [x24, #0x10]\n"
-      "str q22, [x24, #0x20]\n"
-      "str q23, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q28, [x22, #0x0]\n"
-      "str q29, [x22, #0x10]\n"
-      "str q30, [x22, #0x20]\n"
-      "str q31, [x22, #0x30]\n"
+      "str q12, [x25, #0x0]\n"
+      "str q13, [x25, #0x10]\n"
+      "str q14, [x25, #0x20]\n"
+      "str q15, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q21, [x23, #0x10]\n"
+      "str q22, [x23, #0x20]\n"
+      "str q23, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
+      "str q28, [x21, #0x0]\n"
+      "str q29, [x21, #0x10]\n"
+      "str q30, [x21, #0x20]\n"
+      "str q31, [x21, #0x30]\n"
       "198:"  // Height 6: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 167b\n"
@@ -3352,8 +3349,8 @@ void a64_hybrid_fp32_mla_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "200:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4.hpp
index e8f7cdf329..97b6c3f8e0 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4.hpp
@@ -70,7 +70,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 4, 1> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 8, 4, 1> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel=a64_hybrid_fp32_mla_8x4;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp
index f8b117c546..cf1d3070a8 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp
@@ -47,19 +47,18 @@ void a64_hybrid_fp32_mla_8x4_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void a64_hybrid_fp32_mla_8x4_a55 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -106,10 +104,10 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "cmp %x[M], #0x2\n"
       "bgt 43f\n"
       "beq 22f\n"
-      "ldr x3, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x3, %x[bias]\n"
       "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x6, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x6, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x3, 3f\n"
       "ldr q24, [x3, #0x0]\n"
@@ -121,15 +119,15 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bge 6f\n"
       "tbz x4, #1, 4f\n"
       "ldr d24, [x6], #0x8\n"
-      "mov x20, #0x8\n"
+      "mov x26, #0x8\n"
       "tbz x4, #0, 5f\n"
       "ld1 { v24.s }[2], [x6]\n"
       "b 5f\n"
       "4:"  // Height 1: Partial accumulate: partial_1_0
       "ldr s24, [x6, #0x0]\n"
-      "mov x20, #0x0\n"
+      "mov x26, #0x0\n"
       "5:"  // Height 1: Partial accumulate: Done
-      "sub x6, x6, x20\n"
+      "sub x6, x6, x26\n"
       "b 8f\n"
       "6:"  // Height 1: full accumulate
       "ldr q24, [x6, #0x0]\n"
@@ -139,16 +137,16 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "8:"  // Height 1: setup done
       "mov x7, #0x0\n"
       "9:"  // Height 1: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w8, [x20, x7, LSL #0x2]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w8, [x26, x7, LSL #0x2]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 10f\n"
-      "ldr x20, [%x[input_ptr], x7, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x17, [x20, #0x0]\n"
+      "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
+      "add x26, x26, x27, LSL #3\n"
+      "ldr x17, [x26, #0x0]\n"
       "cbnz x7, 11f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x17, x17, x20, LSL #2\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x17, x17, x26, LSL #2\n"
       "b 11f\n"
       "10:"  // Height 1: setup direct input
       "mov x17, %x[input_ptr]\n"
@@ -165,59 +163,59 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "12:"  // Height 1: Multiply loop: Main loop head
       "fmla v24.4s, v8.4s, v0.s[0]\n"
       "add x17, x17, #0x10\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x5, x5, #0x40\n"
-      "ldr x23, [x17, #0x8]\n"
       "ldr d8, [x5, #0x0]\n"
-      "sub x8, x8, #0x4\n"
-      "ldr x20, [x5, #0x8]\n"
-      "cmp x8, #0x8\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
-      "ldr d9, [x5, #0x10]\n"
-      "ldr x22, [x5, #0x18]\n"
-      "ldr x21, [x5, #0x28]\n"
-      "mov v8.d[1], x20\n"
-      "ldr x20, [x5, #0x38]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "ldr d10, [x5, #0x20]\n"
-      "prfm pldl1keep, [x17, #0x80]\n"
-      "mov v9.d[1], x22\n"
-      "mov v10.d[1], x21\n"
+      "ldr d9, [x5, #0x10]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr d0, [x17, #0x0]\n"
+      "sub x8, x8, #0x4\n"
+      "ldr d10, [x5, #0x20]\n"
+      "cmp x8, #0x8\n"
       "ldr d11, [x5, #0x30]\n"
-      "mov v0.d[1], x23\n"
-      "mov v11.d[1], x20\n"
+      "ldr x26, [x5, #0x8]\n"
+      "mov v8.d[1], x26\n"
+      "ldr x26, [x5, #0x18]\n"
+      "mov v9.d[1], x26\n"
+      "ldr x26, [x17, #0x8]\n"
+      "mov v0.d[1], x26\n"
+      "ldr x26, [x5, #0x28]\n"
+      "mov v10.d[1], x26\n"
+      "ldr x26, [x5, #0x38]\n"
+      "mov v11.d[1], x26\n"
+      "prfm pldl1keep, [x17, #0x80]\n"
       "bge 12b\n"
       "13:"  // Height 1: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
       "add x17, x17, #0x10\n"
-      "sub x8, x8, #0x4\n"
-      "prfm pldl1keep, [x17, #0x80]\n"
-      "add x5, x5, #0x40\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
+      "sub x8, x8, #0x4\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
+      "prfm pldl1keep, [x17, #0x80]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
+      "add x5, x5, #0x40\n"
       "14:"  // Height 1: Multiply loop: Main loop skip
       "cbz x8, 16f\n"
       "15:"  // Height 1: Multiply loop: Odd block loop
       "ldr s17, [x17], #0x4\n"
       "sub x8, x8, #0x1\n"
       "ldr q16, [x5, #0x0]\n"
-      "add x5, x5, #0x10\n"
       "fmla v24.4s, v16.4s, v17.s[0]\n"
+      "add x5, x5, #0x10\n"
       "cbnz x8, 15b\n"
       "16:"  // Height 1: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x7, x7, #0x1\n"
-      "cmp x7, x20\n"
+      "cmp x7, x26\n"
       "bne 9b\n"
       "prfm pstl1keep, [x6, #0x0]\n"
       "tbz %x[flags], #1, 17f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
-      "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "fmin v24.4s, v24.4s, v17.4s\n"
+      "add x26, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x26]\n"
+      "fmin v24.4s, v24.4s, v16.4s\n"
+      "add x26, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v16.4s }, [x26]\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
       "17:"  // Height 1: No activation
       "cmp x4, #0x4\n"
@@ -239,40 +237,40 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bgt 2b\n"
       "b 170f\n"
       "22:"  // Height 2
-      "ldr x3, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x3, %x[bias]\n"
       "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x6, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x6, %x[output_ptr]\n"
       "23:"  // Height 2: Column loop
       "cbz x3, 24f\n"
       "ldr q24, [x3, #0x0]\n"
-      "add x3, x3, #0x10\n"
       "mov v25.16b, v24.16b\n"
+      "add x3, x3, #0x10\n"
       "b 29f\n"
       "24:"  // Height 2: no bias
       "tbz %x[flags], #0, 28f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x4, #0x4\n"
-      "add x28, x6, x20, LSL #2\n"
+      "add x13, x6, x26, LSL #2\n"
       "bge 27f\n"
       "tbz x4, #1, 25f\n"
       "ldr d24, [x6], #0x8\n"
-      "mov x20, #0x8\n"
-      "ldr d25, [x28], #0x8\n"
+      "mov x26, #0x8\n"
+      "ldr d25, [x13], #0x8\n"
       "tbz x4, #0, 26f\n"
       "ld1 { v24.s }[2], [x6]\n"
-      "ld1 { v25.s }[2], [x28]\n"
+      "ld1 { v25.s }[2], [x13]\n"
       "b 26f\n"
       "25:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s24, [x6, #0x0]\n"
-      "mov x20, #0x0\n"
-      "ldr s25, [x28, #0x0]\n"
+      "mov x26, #0x0\n"
+      "ldr s25, [x13, #0x0]\n"
       "26:"  // Height 2: Partial accumulate: Done
-      "sub x6, x6, x20\n"
+      "sub x6, x6, x26\n"
       "b 29f\n"
       "27:"  // Height 2: full accumulate
       "ldr q24, [x6, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
+      "ldr q25, [x13, #0x0]\n"
       "b 29f\n"
       "28:"  // Height 2: no accumulate
       "movi v24.16b, #0x0\n"
@@ -280,22 +278,22 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "29:"  // Height 2: setup done
       "mov x7, #0x0\n"
       "30:"  // Height 2: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w8, [x20, x7, LSL #0x2]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w8, [x26, x7, LSL #0x2]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 31f\n"
-      "ldr x20, [%x[input_ptr], x7, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x17, [x20, #0x0]\n"
-      "ldr x16, [x20, #0x8]\n"
+      "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
+      "add x26, x26, x27, LSL #3\n"
+      "ldr x17, [x26, #0x0]\n"
+      "ldr x16, [x26, #0x8]\n"
       "cbnz x7, 32f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x17, x17, x20, LSL #2\n"
-      "add x16, x16, x20, LSL #2\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x17, x17, x26, LSL #2\n"
+      "add x16, x16, x26, LSL #2\n"
       "b 32f\n"
       "31:"  // Height 2: setup direct input
       "mov x17, %x[input_ptr]\n"
-      "add x16, x17, x21, LSL #2\n"
+      "add x16, x17, x27, LSL #2\n"
       "32:"  // Height 2: input setup done
       "cmp x8, #0x4\n"
       "blt 35f\n"
@@ -312,49 +310,49 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "add x17, x17, #0x10\n"
       "fmla v25.4s, v8.4s, v1.s[0]\n"
       "add x16, x16, #0x10\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x5, x5, #0x40\n"
-      "ldr x24, [x17, #0x8]\n"
       "ldr d8, [x5, #0x0]\n"
-      "sub x8, x8, #0x4\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
-      "ldr x20, [x5, #0x8]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
       "ldr d9, [x5, #0x10]\n"
-      "ldr x23, [x5, #0x18]\n"
-      "cmp x8, #0x8\n"
-      "ldr x22, [x5, #0x28]\n"
-      "mov v8.d[1], x20\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "ldr x21, [x16, #0x8]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
       "ldr d10, [x5, #0x20]\n"
-      "ldr x20, [x5, #0x38]\n"
-      "mov v9.d[1], x23\n"
-      "prfm pldl1keep, [x17, #0x80]\n"
+      "ldr x27, [x5, #0x8]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr d0, [x17, #0x0]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
       "ldr d1, [x16, #0x0]\n"
+      "sub x8, x8, #0x4\n"
       "ldr d11, [x5, #0x30]\n"
-      "mov v10.d[1], x22\n"
+      "cmp x8, #0x8\n"
+      "ldr x26, [x5, #0x18]\n"
+      "mov v8.d[1], x27\n"
+      "ldr x27, [x17, #0x8]\n"
+      "mov v9.d[1], x26\n"
+      "ldr x26, [x16, #0x8]\n"
+      "mov v0.d[1], x27\n"
+      "ldr x27, [x5, #0x28]\n"
+      "mov v1.d[1], x26\n"
+      "ldr x26, [x5, #0x38]\n"
+      "mov v10.d[1], x27\n"
+      "mov v11.d[1], x26\n"
+      "prfm pldl1keep, [x17, #0x80]\n"
       "prfm pldl1keep, [x16, #0x80]\n"
-      "mov v0.d[1], x24\n"
-      "mov v1.d[1], x21\n"
-      "mov v11.d[1], x20\n"
       "bge 33b\n"
       "34:"  // Height 2: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
       "add x17, x17, #0x10\n"
       "fmla v25.4s, v8.4s, v1.s[0]\n"
       "add x16, x16, #0x10\n"
-      "sub x8, x8, #0x4\n"
-      "prfm pldl1keep, [x17, #0x80]\n"
-      "add x5, x5, #0x40\n"
-      "prfm pldl1keep, [x16, #0x80]\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
+      "sub x8, x8, #0x4\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "prfm pldl1keep, [x17, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
+      "prfm pldl1keep, [x16, #0x80]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "add x5, x5, #0x40\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
       "35:"  // Height 2: Multiply loop: Main loop skip
@@ -364,26 +362,26 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "sub x8, x8, #0x1\n"
       "ldr s17, [x16], #0x4\n"
       "ldr q16, [x5, #0x0]\n"
-      "add x5, x5, #0x10\n"
       "fmla v24.4s, v16.4s, v18.s[0]\n"
       "fmla v25.4s, v16.4s, v17.s[0]\n"
+      "add x5, x5, #0x10\n"
       "cbnz x8, 36b\n"
       "37:"  // Height 2: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x7, x7, #0x1\n"
-      "cmp x7, x20\n"
+      "cmp x7, x26\n"
       "bne 30b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
       "prfm pstl1keep, [x6, #0x0]\n"
-      "add x28, x6, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
-      "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "fmin v24.4s, v24.4s, v17.4s\n"
-      "fmin v25.4s, v25.4s, v17.4s\n"
+      "add x26, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x26]\n"
+      "fmin v24.4s, v24.4s, v16.4s\n"
+      "fmin v25.4s, v25.4s, v16.4s\n"
+      "add x26, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v16.4s }, [x26]\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
       "fmax v25.4s, v25.4s, v16.4s\n"
       "38:"  // Height 2: No activation
@@ -391,65 +389,65 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bge 41f\n"
       "tbz x4, #1, 39f\n"
       "str d24, [x6], #0x8\n"
-      "str d25, [x28], #0x8\n"
+      "str d25, [x13], #0x8\n"
       "tbz x4, #0, 40f\n"
       "st1 { v24.s }[2], [x6]\n"
-      "st1 { v25.s }[2], [x28]\n"
+      "st1 { v25.s }[2], [x13]\n"
       "b 40f\n"
       "39:"  // Height 2: Partial direct writeback: partial_1_0
       "str s24, [x6, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
+      "str s25, [x13, #0x0]\n"
       "40:"  // Height 2: Partial direct writeback: Done
       "b 42f\n"
       "41:"  // Height 2: Full writeback
       "str q24, [x6, #0x0]\n"
       "add x6, x6, #0x10\n"
-      "str q25, [x28, #0x0]\n"
+      "str q25, [x13, #0x0]\n"
       "42:"  // Height 2: Writeback done
       "subs x4, x4, #0x4\n"
       "bgt 23b\n"
       "b 170f\n"
       "43:"  // Height 3
-      "ldr x3, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x3, %x[bias]\n"
       "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x6, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x6, %x[output_ptr]\n"
       "44:"  // Height 3: Column loop
       "cbz x3, 45f\n"
       "ldr q24, [x3, #0x0]\n"
-      "add x3, x3, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x3, x3, #0x10\n"
       "b 50f\n"
       "45:"  // Height 3: no bias
       "tbz %x[flags], #0, 49f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
       "cmp x4, #0x4\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
       "bge 48f\n"
       "tbz x4, #1, 46f\n"
       "ldr d24, [x6], #0x8\n"
-      "mov x20, #0x8\n"
-      "ldr d25, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
+      "mov x26, #0x8\n"
+      "ldr d25, [x13], #0x8\n"
+      "ldr d26, [x12], #0x8\n"
       "tbz x4, #0, 47f\n"
       "ld1 { v24.s }[2], [x6]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
+      "ld1 { v25.s }[2], [x13]\n"
+      "ld1 { v26.s }[2], [x12]\n"
       "b 47f\n"
       "46:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s24, [x6, #0x0]\n"
-      "mov x20, #0x0\n"
-      "ldr s25, [x28, #0x0]\n"
-      "ldr s26, [x27, #0x0]\n"
+      "mov x26, #0x0\n"
+      "ldr s25, [x13, #0x0]\n"
+      "ldr s26, [x12, #0x0]\n"
       "47:"  // Height 3: Partial accumulate: Done
-      "sub x6, x6, x20\n"
+      "sub x6, x6, x26\n"
       "b 50f\n"
       "48:"  // Height 3: full accumulate
       "ldr q24, [x6, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
+      "ldr q25, [x13, #0x0]\n"
+      "ldr q26, [x12, #0x0]\n"
       "b 50f\n"
       "49:"  // Height 3: no accumulate
       "movi v24.16b, #0x0\n"
@@ -458,25 +456,25 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "50:"  // Height 3: setup done
       "mov x7, #0x0\n"
       "51:"  // Height 3: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w8, [x20, x7, LSL #0x2]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w8, [x26, x7, LSL #0x2]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 52f\n"
-      "ldr x20, [%x[input_ptr], x7, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x17, [x20, #0x0]\n"
-      "ldr x16, [x20, #0x8]\n"
-      "ldr x15, [x20, #0x10]\n"
+      "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
+      "add x26, x26, x27, LSL #3\n"
+      "ldr x17, [x26, #0x0]\n"
+      "ldr x16, [x26, #0x8]\n"
+      "ldr x15, [x26, #0x10]\n"
       "cbnz x7, 53f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x17, x17, x20, LSL #2\n"
-      "add x16, x16, x20, LSL #2\n"
-      "add x15, x15, x20, LSL #2\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x17, x17, x26, LSL #2\n"
+      "add x16, x16, x26, LSL #2\n"
+      "add x15, x15, x26, LSL #2\n"
       "b 53f\n"
       "52:"  // Height 3: setup direct input
       "mov x17, %x[input_ptr]\n"
-      "add x16, x17, x21, LSL #2\n"
-      "add x15, x16, x21, LSL #2\n"
+      "add x16, x17, x27, LSL #2\n"
+      "add x15, x16, x27, LSL #2\n"
       "53:"  // Height 3: input setup done
       "cmp x8, #0x4\n"
       "blt 56f\n"
@@ -496,42 +494,42 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "add x16, x16, #0x10\n"
       "fmla v26.4s, v8.4s, v2.s[0]\n"
       "add x15, x15, #0x10\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x5, x5, #0x40\n"
-      "ldr x25, [x17, #0x8]\n"
       "ldr d8, [x5, #0x0]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
-      "ldr x20, [x5, #0x8]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
       "ldr d9, [x5, #0x10]\n"
-      "ldr x24, [x5, #0x18]\n"
-      "sub x8, x8, #0x4\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "ldr x23, [x5, #0x28]\n"
+      "ldr x28, [x5, #0x8]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
-      "ldr x22, [x16, #0x8]\n"
+      "ldr x27, [x5, #0x18]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "ldr d10, [x5, #0x20]\n"
-      "ldr x21, [x15, #0x8]\n"
-      "cmp x8, #0x8\n"
+      "ldr x26, [x5, #0x28]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr d0, [x17, #0x0]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
       "ldr d1, [x16, #0x0]\n"
       "fmla v26.4s, v11.4s, v2.s[3]\n"
       "ldr d2, [x15, #0x0]\n"
+      "sub x8, x8, #0x4\n"
       "ldr d11, [x5, #0x30]\n"
-      "mov v8.d[1], x20\n"
-      "ldr x20, [x5, #0x38]\n"
-      "mov v9.d[1], x24\n"
+      "cmp x8, #0x8\n"
+      "ldr x9, [x17, #0x8]\n"
+      "mov v8.d[1], x28\n"
+      "ldr x28, [x16, #0x8]\n"
+      "mov v9.d[1], x27\n"
+      "ldr x27, [x15, #0x8]\n"
+      "mov v10.d[1], x26\n"
+      "ldr x26, [x5, #0x38]\n"
+      "mov v0.d[1], x9\n"
+      "mov v1.d[1], x28\n"
       "prfm pldl1keep, [x17, #0x80]\n"
-      "mov v10.d[1], x23\n"
+      "mov v2.d[1], x27\n"
       "prfm pldl1keep, [x16, #0x80]\n"
-      "mov v0.d[1], x25\n"
+      "mov v11.d[1], x26\n"
       "prfm pldl1keep, [x15, #0x80]\n"
-      "mov v1.d[1], x22\n"
-      "mov v2.d[1], x21\n"
-      "mov v11.d[1], x20\n"
       "bge 54b\n"
       "55:"  // Height 3: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
@@ -540,16 +538,16 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "add x16, x16, #0x10\n"
       "fmla v26.4s, v8.4s, v2.s[0]\n"
       "add x15, x15, #0x10\n"
-      "sub x8, x8, #0x4\n"
-      "prfm pldl1keep, [x17, #0x80]\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
-      "prfm pldl1keep, [x16, #0x80]\n"
+      "sub x8, x8, #0x4\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
-      "add x5, x5, #0x40\n"
+      "prfm pldl1keep, [x17, #0x80]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
-      "prfm pldl1keep, [x15, #0x80]\n"
+      "prfm pldl1keep, [x16, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
+      "prfm pldl1keep, [x15, #0x80]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "add x5, x5, #0x40\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
@@ -562,30 +560,30 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr s18, [x16], #0x4\n"
       "ldr s17, [x15], #0x4\n"
       "ldr q16, [x5, #0x0]\n"
-      "add x5, x5, #0x10\n"
       "fmla v24.4s, v16.4s, v19.s[0]\n"
       "fmla v25.4s, v16.4s, v18.s[0]\n"
+      "add x5, x5, #0x10\n"
       "fmla v26.4s, v16.4s, v17.s[0]\n"
       "cbnz x8, 57b\n"
       "58:"  // Height 3: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x7, x7, #0x1\n"
-      "cmp x7, x20\n"
+      "cmp x7, x26\n"
       "bne 51b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
       "prfm pstl1keep, [x6, #0x0]\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x12, #0x0]\n"
       "tbz %x[flags], #1, 59f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
-      "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "fmin v24.4s, v24.4s, v17.4s\n"
-      "fmin v25.4s, v25.4s, v17.4s\n"
-      "fmin v26.4s, v26.4s, v17.4s\n"
+      "add x26, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x26]\n"
+      "fmin v24.4s, v24.4s, v16.4s\n"
+      "fmin v25.4s, v25.4s, v16.4s\n"
+      "fmin v26.4s, v26.4s, v16.4s\n"
+      "add x26, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v16.4s }, [x26]\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
       "fmax v25.4s, v25.4s, v16.4s\n"
       "fmax v26.4s, v26.4s, v16.4s\n"
@@ -594,75 +592,75 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bge 62f\n"
       "tbz x4, #1, 60f\n"
       "str d24, [x6], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
+      "str d25, [x13], #0x8\n"
+      "str d26, [x12], #0x8\n"
       "tbz x4, #0, 61f\n"
       "st1 { v24.s }[2], [x6]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
+      "st1 { v25.s }[2], [x13]\n"
+      "st1 { v26.s }[2], [x12]\n"
       "b 61f\n"
       "60:"  // Height 3: Partial direct writeback: partial_1_0
       "str s24, [x6, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
+      "str s25, [x13, #0x0]\n"
+      "str s26, [x12, #0x0]\n"
       "61:"  // Height 3: Partial direct writeback: Done
       "b 63f\n"
       "62:"  // Height 3: Full writeback
       "str q24, [x6, #0x0]\n"
       "add x6, x6, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
+      "str q25, [x13, #0x0]\n"
+      "str q26, [x12, #0x0]\n"
       "63:"  // Height 3: Writeback done
       "subs x4, x4, #0x4\n"
       "bgt 44b\n"
       "b 170f\n"
       "64:"  // Height 4
-      "ldr x3, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x3, %x[bias]\n"
       "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x6, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x6, %x[output_ptr]\n"
       "65:"  // Height 4: Column loop
       "cbz x3, 66f\n"
       "ldr q24, [x3, #0x0]\n"
-      "add x3, x3, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x3, x3, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "b 71f\n"
       "66:"  // Height 4: no bias
       "tbz %x[flags], #0, 70f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
       "cmp x4, #0x4\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "add x26, x27, x20, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
       "bge 69f\n"
       "tbz x4, #1, 67f\n"
       "ldr d24, [x6], #0x8\n"
-      "mov x20, #0x8\n"
-      "ldr d25, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
+      "mov x26, #0x8\n"
+      "ldr d25, [x13], #0x8\n"
+      "ldr d26, [x12], #0x8\n"
+      "ldr d27, [x11], #0x8\n"
       "tbz x4, #0, 68f\n"
       "ld1 { v24.s }[2], [x6]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
+      "ld1 { v25.s }[2], [x13]\n"
+      "ld1 { v26.s }[2], [x12]\n"
+      "ld1 { v27.s }[2], [x11]\n"
       "b 68f\n"
       "67:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s24, [x6, #0x0]\n"
-      "mov x20, #0x0\n"
-      "ldr s25, [x28, #0x0]\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
+      "mov x26, #0x0\n"
+      "ldr s25, [x13, #0x0]\n"
+      "ldr s26, [x12, #0x0]\n"
+      "ldr s27, [x11, #0x0]\n"
       "68:"  // Height 4: Partial accumulate: Done
-      "sub x6, x6, x20\n"
+      "sub x6, x6, x26\n"
       "b 71f\n"
       "69:"  // Height 4: full accumulate
       "ldr q24, [x6, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
+      "ldr q25, [x13, #0x0]\n"
+      "ldr q26, [x12, #0x0]\n"
+      "ldr q27, [x11, #0x0]\n"
       "b 71f\n"
       "70:"  // Height 4: no accumulate
       "movi v24.16b, #0x0\n"
@@ -672,28 +670,28 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "71:"  // Height 4: setup done
       "mov x7, #0x0\n"
       "72:"  // Height 4: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w8, [x20, x7, LSL #0x2]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w8, [x26, x7, LSL #0x2]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 73f\n"
-      "ldr x20, [%x[input_ptr], x7, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x17, [x20, #0x0]\n"
-      "ldr x16, [x20, #0x8]\n"
-      "ldr x15, [x20, #0x10]\n"
-      "ldr x14, [x20, #0x18]\n"
+      "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
+      "add x26, x26, x27, LSL #3\n"
+      "ldr x17, [x26, #0x0]\n"
+      "ldr x16, [x26, #0x8]\n"
+      "ldr x15, [x26, #0x10]\n"
+      "ldr x14, [x26, #0x18]\n"
       "cbnz x7, 74f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x17, x17, x20, LSL #2\n"
-      "add x16, x16, x20, LSL #2\n"
-      "add x15, x15, x20, LSL #2\n"
-      "add x14, x14, x20, LSL #2\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x17, x17, x26, LSL #2\n"
+      "add x16, x16, x26, LSL #2\n"
+      "add x15, x15, x26, LSL #2\n"
+      "add x14, x14, x26, LSL #2\n"
       "b 74f\n"
       "73:"  // Height 4: setup direct input
       "mov x17, %x[input_ptr]\n"
-      "add x16, x17, x21, LSL #2\n"
-      "add x15, x16, x21, LSL #2\n"
-      "add x14, x15, x21, LSL #2\n"
+      "add x16, x17, x27, LSL #2\n"
+      "add x15, x16, x27, LSL #2\n"
+      "add x14, x15, x27, LSL #2\n"
       "74:"  // Height 4: input setup done
       "cmp x8, #0x4\n"
       "blt 77f\n"
@@ -721,22 +719,22 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr d8, [x5, #0x0]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
-      "ldr x21, [x5, #0x8]\n"
+      "ldr x27, [x5, #0x8]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
       "ldr d9, [x5, #0x10]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "ldr x20, [x5, #0x18]\n"
+      "ldr x26, [x5, #0x18]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
-      "ldr x25, [x5, #0x28]\n"
+      "ldr x11, [x5, #0x28]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
-      "ldr x24, [x17, #0x8]\n"
+      "ldr x10, [x17, #0x8]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
       "ldr d10, [x5, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr d0, [x17, #0x0]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
       "ldr d1, [x16, #0x0]\n"
-      "ldr x23, [x16, #0x8]\n"
+      "ldr x9, [x16, #0x8]\n"
       "fmla v26.4s, v11.4s, v2.s[3]\n"
       "ldr d2, [x15, #0x0]\n"
       "fmla v27.4s, v11.4s, v3.s[3]\n"
@@ -744,21 +742,21 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "sub x8, x8, #0x4\n"
       "ldr d11, [x5, #0x30]\n"
       "cmp x8, #0x8\n"
-      "ldr x22, [x15, #0x8]\n"
-      "mov v8.d[1], x21\n"
-      "ldr x21, [x14, #0x8]\n"
-      "mov v9.d[1], x20\n"
-      "ldr x20, [x5, #0x38]\n"
-      "mov v10.d[1], x25\n"
+      "ldr x28, [x15, #0x8]\n"
+      "mov v8.d[1], x27\n"
+      "ldr x27, [x14, #0x8]\n"
+      "mov v9.d[1], x26\n"
+      "ldr x26, [x5, #0x38]\n"
+      "mov v10.d[1], x11\n"
       "prfm pldl1keep, [x17, #0x80]\n"
-      "mov v0.d[1], x24\n"
+      "mov v0.d[1], x10\n"
       "prfm pldl1keep, [x16, #0x80]\n"
-      "mov v1.d[1], x23\n"
-      "mov v2.d[1], x22\n"
+      "mov v1.d[1], x9\n"
+      "mov v2.d[1], x28\n"
       "prfm pldl1keep, [x15, #0x80]\n"
-      "mov v3.d[1], x21\n"
+      "mov v3.d[1], x27\n"
       "prfm pldl1keep, [x14, #0x80]\n"
-      "mov v11.d[1], x20\n"
+      "mov v11.d[1], x26\n"
       "bge 75b\n"
       "76:"  // Height 4: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
@@ -796,34 +794,34 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr s18, [x15], #0x4\n"
       "ldr s17, [x14], #0x4\n"
       "ldr q16, [x5, #0x0]\n"
-      "add x5, x5, #0x10\n"
       "fmla v24.4s, v16.4s, v20.s[0]\n"
       "fmla v25.4s, v16.4s, v19.s[0]\n"
+      "add x5, x5, #0x10\n"
       "fmla v26.4s, v16.4s, v18.s[0]\n"
       "fmla v27.4s, v16.4s, v17.s[0]\n"
       "cbnz x8, 78b\n"
       "79:"  // Height 4: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x7, x7, #0x1\n"
-      "cmp x7, x20\n"
+      "cmp x7, x26\n"
       "bne 72b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
       "prfm pstl1keep, [x6, #0x0]\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x26, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x12, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "tbz %x[flags], #1, 80f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
-      "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "fmin v24.4s, v24.4s, v17.4s\n"
-      "fmin v25.4s, v25.4s, v17.4s\n"
-      "fmin v26.4s, v26.4s, v17.4s\n"
-      "fmin v27.4s, v27.4s, v17.4s\n"
+      "add x26, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x26]\n"
+      "fmin v24.4s, v24.4s, v16.4s\n"
+      "fmin v25.4s, v25.4s, v16.4s\n"
+      "fmin v26.4s, v26.4s, v16.4s\n"
+      "fmin v27.4s, v27.4s, v16.4s\n"
+      "add x26, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v16.4s }, [x26]\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
       "fmax v25.4s, v25.4s, v16.4s\n"
       "fmax v26.4s, v26.4s, v16.4s\n"
@@ -833,85 +831,85 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bge 83f\n"
       "tbz x4, #1, 81f\n"
       "str d24, [x6], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
+      "str d25, [x13], #0x8\n"
+      "str d26, [x12], #0x8\n"
+      "str d27, [x11], #0x8\n"
       "tbz x4, #0, 82f\n"
       "st1 { v24.s }[2], [x6]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
+      "st1 { v25.s }[2], [x13]\n"
+      "st1 { v26.s }[2], [x12]\n"
+      "st1 { v27.s }[2], [x11]\n"
       "b 82f\n"
       "81:"  // Height 4: Partial direct writeback: partial_1_0
       "str s24, [x6, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
+      "str s25, [x13, #0x0]\n"
+      "str s26, [x12, #0x0]\n"
+      "str s27, [x11, #0x0]\n"
       "82:"  // Height 4: Partial direct writeback: Done
       "b 84f\n"
       "83:"  // Height 4: Full writeback
       "str q24, [x6, #0x0]\n"
       "add x6, x6, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
+      "str q25, [x13, #0x0]\n"
+      "str q26, [x12, #0x0]\n"
+      "str q27, [x11, #0x0]\n"
       "84:"  // Height 4: Writeback done
       "subs x4, x4, #0x4\n"
       "bgt 65b\n"
       "b 170f\n"
       "85:"  // Height 5
-      "ldr x3, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x3, %x[bias]\n"
       "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x6, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x6, %x[output_ptr]\n"
       "86:"  // Height 5: Column loop
       "cbz x3, 87f\n"
       "ldr q24, [x3, #0x0]\n"
-      "add x3, x3, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x3, x3, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "mov v28.16b, v24.16b\n"
       "b 92f\n"
       "87:"  // Height 5: no bias
       "tbz %x[flags], #0, 91f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
       "cmp x4, #0x4\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "add x26, x27, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x10, x11, x26, LSL #2\n"
       "bge 90f\n"
       "tbz x4, #1, 88f\n"
       "ldr d24, [x6], #0x8\n"
-      "mov x20, #0x8\n"
-      "ldr d25, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
+      "mov x26, #0x8\n"
+      "ldr d25, [x13], #0x8\n"
+      "ldr d26, [x12], #0x8\n"
+      "ldr d27, [x11], #0x8\n"
+      "ldr d28, [x10], #0x8\n"
       "tbz x4, #0, 89f\n"
       "ld1 { v24.s }[2], [x6]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v28.s }[2], [x25]\n"
+      "ld1 { v25.s }[2], [x13]\n"
+      "ld1 { v26.s }[2], [x12]\n"
+      "ld1 { v27.s }[2], [x11]\n"
+      "ld1 { v28.s }[2], [x10]\n"
       "b 89f\n"
       "88:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s24, [x6, #0x0]\n"
-      "mov x20, #0x0\n"
-      "ldr s25, [x28, #0x0]\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s28, [x25, #0x0]\n"
+      "mov x26, #0x0\n"
+      "ldr s25, [x13, #0x0]\n"
+      "ldr s26, [x12, #0x0]\n"
+      "ldr s27, [x11, #0x0]\n"
+      "ldr s28, [x10, #0x0]\n"
       "89:"  // Height 5: Partial accumulate: Done
-      "sub x6, x6, x20\n"
+      "sub x6, x6, x26\n"
       "b 92f\n"
       "90:"  // Height 5: full accumulate
       "ldr q24, [x6, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
-      "ldr q28, [x25, #0x0]\n"
+      "ldr q25, [x13, #0x0]\n"
+      "ldr q26, [x12, #0x0]\n"
+      "ldr q27, [x11, #0x0]\n"
+      "ldr q28, [x10, #0x0]\n"
       "b 92f\n"
       "91:"  // Height 5: no accumulate
       "movi v24.16b, #0x0\n"
@@ -922,31 +920,31 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "92:"  // Height 5: setup done
       "mov x7, #0x0\n"
       "93:"  // Height 5: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w8, [x20, x7, LSL #0x2]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w8, [x26, x7, LSL #0x2]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 94f\n"
-      "ldr x20, [%x[input_ptr], x7, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x17, [x20, #0x0]\n"
-      "ldr x16, [x20, #0x8]\n"
-      "ldr x15, [x20, #0x10]\n"
-      "ldr x14, [x20, #0x18]\n"
-      "ldr x13, [x20, #0x20]\n"
+      "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
+      "add x26, x26, x27, LSL #3\n"
+      "ldr x17, [x26, #0x0]\n"
+      "ldr x16, [x26, #0x8]\n"
+      "ldr x15, [x26, #0x10]\n"
+      "ldr x14, [x26, #0x18]\n"
+      "ldr x13, [x26, #0x20]\n"
       "cbnz x7, 95f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x17, x17, x20, LSL #2\n"
-      "add x16, x16, x20, LSL #2\n"
-      "add x15, x15, x20, LSL #2\n"
-      "add x14, x14, x20, LSL #2\n"
-      "add x13, x13, x20, LSL #2\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x17, x17, x26, LSL #2\n"
+      "add x16, x16, x26, LSL #2\n"
+      "add x15, x15, x26, LSL #2\n"
+      "add x14, x14, x26, LSL #2\n"
+      "add x13, x13, x26, LSL #2\n"
       "b 95f\n"
       "94:"  // Height 5: setup direct input
       "mov x17, %x[input_ptr]\n"
-      "add x16, x17, x21, LSL #2\n"
-      "add x15, x16, x21, LSL #2\n"
-      "add x14, x15, x21, LSL #2\n"
-      "add x13, x14, x21, LSL #2\n"
+      "add x16, x17, x27, LSL #2\n"
+      "add x15, x16, x27, LSL #2\n"
+      "add x14, x15, x27, LSL #2\n"
+      "add x13, x14, x27, LSL #2\n"
       "95:"  // Height 5: input setup done
       "cmp x8, #0x4\n"
       "blt 98f\n"
@@ -977,19 +975,19 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr d8, [x5, #0x0]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
-      "ldr x20, [x5, #0x8]\n"
+      "ldr x27, [x5, #0x8]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
-      "ldr x27, [x5, #0x18]\n"
+      "ldr x26, [x5, #0x18]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
       "ldr d9, [x5, #0x10]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "ldr x26, [x5, #0x28]\n"
+      "ldr x12, [x5, #0x28]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
-      "ldr x25, [x17, #0x8]\n"
+      "ldr x11, [x17, #0x8]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
-      "ldr x24, [x16, #0x8]\n"
+      "ldr x10, [x16, #0x8]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
-      "ldr x23, [x15, #0x8]\n"
+      "ldr x9, [x15, #0x8]\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
       "ldr d10, [x5, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
@@ -1000,28 +998,28 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr d2, [x15, #0x0]\n"
       "fmla v27.4s, v11.4s, v3.s[3]\n"
       "ldr d3, [x14, #0x0]\n"
-      "ldr x22, [x14, #0x8]\n"
+      "ldr x28, [x14, #0x8]\n"
       "fmla v28.4s, v11.4s, v4.s[3]\n"
       "ldr d4, [x13, #0x0]\n"
       "sub x8, x8, #0x4\n"
       "ldr d11, [x5, #0x30]\n"
       "cmp x8, #0x8\n"
-      "ldr x21, [x13, #0x8]\n"
-      "mov v8.d[1], x20\n"
-      "ldr x20, [x5, #0x38]\n"
-      "mov v9.d[1], x27\n"
+      "mov v8.d[1], x27\n"
+      "ldr x27, [x13, #0x8]\n"
+      "mov v9.d[1], x26\n"
+      "ldr x26, [x5, #0x38]\n"
       "prfm pldl1keep, [x17, #0x80]\n"
-      "mov v10.d[1], x26\n"
+      "mov v10.d[1], x12\n"
       "prfm pldl1keep, [x16, #0x80]\n"
-      "mov v0.d[1], x25\n"
+      "mov v0.d[1], x11\n"
       "prfm pldl1keep, [x15, #0x80]\n"
-      "mov v1.d[1], x24\n"
+      "mov v1.d[1], x10\n"
       "prfm pldl1keep, [x14, #0x80]\n"
-      "mov v2.d[1], x23\n"
-      "mov v3.d[1], x22\n"
+      "mov v2.d[1], x9\n"
+      "mov v3.d[1], x28\n"
       "prfm pldl1keep, [x13, #0x80]\n"
-      "mov v4.d[1], x21\n"
-      "mov v11.d[1], x20\n"
+      "mov v4.d[1], x27\n"
+      "mov v11.d[1], x26\n"
       "bge 96b\n"
       "97:"  // Height 5: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
@@ -1066,38 +1064,38 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr s18, [x14], #0x4\n"
       "ldr s17, [x13], #0x4\n"
       "ldr q16, [x5, #0x0]\n"
-      "add x5, x5, #0x10\n"
       "fmla v24.4s, v16.4s, v21.s[0]\n"
       "fmla v25.4s, v16.4s, v20.s[0]\n"
+      "add x5, x5, #0x10\n"
       "fmla v26.4s, v16.4s, v19.s[0]\n"
       "fmla v27.4s, v16.4s, v18.s[0]\n"
       "fmla v28.4s, v16.4s, v17.s[0]\n"
       "cbnz x8, 99b\n"
       "100:"  // Height 5: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x7, x7, #0x1\n"
-      "cmp x7, x20\n"
+      "cmp x7, x26\n"
       "bne 93b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
+      "add x10, x11, x26, LSL #2\n"
       "prfm pstl1keep, [x6, #0x0]\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x26, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x12, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
+      "prfm pstl1keep, [x10, #0x0]\n"
       "tbz %x[flags], #1, 101f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
-      "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "fmin v24.4s, v24.4s, v17.4s\n"
-      "fmin v25.4s, v25.4s, v17.4s\n"
-      "fmin v26.4s, v26.4s, v17.4s\n"
-      "fmin v27.4s, v27.4s, v17.4s\n"
-      "fmin v28.4s, v28.4s, v17.4s\n"
+      "add x26, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x26]\n"
+      "fmin v24.4s, v24.4s, v16.4s\n"
+      "fmin v25.4s, v25.4s, v16.4s\n"
+      "fmin v26.4s, v26.4s, v16.4s\n"
+      "fmin v27.4s, v27.4s, v16.4s\n"
+      "fmin v28.4s, v28.4s, v16.4s\n"
+      "add x26, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v16.4s }, [x26]\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
       "fmax v25.4s, v25.4s, v16.4s\n"
       "fmax v26.4s, v26.4s, v16.4s\n"
@@ -1108,95 +1106,95 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bge 104f\n"
       "tbz x4, #1, 102f\n"
       "str d24, [x6], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
-      "str d28, [x25], #0x8\n"
+      "str d25, [x13], #0x8\n"
+      "str d26, [x12], #0x8\n"
+      "str d27, [x11], #0x8\n"
+      "str d28, [x10], #0x8\n"
       "tbz x4, #0, 103f\n"
       "st1 { v24.s }[2], [x6]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
-      "st1 { v28.s }[2], [x25]\n"
+      "st1 { v25.s }[2], [x13]\n"
+      "st1 { v26.s }[2], [x12]\n"
+      "st1 { v27.s }[2], [x11]\n"
+      "st1 { v28.s }[2], [x10]\n"
       "b 103f\n"
       "102:"  // Height 5: Partial direct writeback: partial_1_0
       "str s24, [x6, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
-      "str s28, [x25, #0x0]\n"
+      "str s25, [x13, #0x0]\n"
+      "str s26, [x12, #0x0]\n"
+      "str s27, [x11, #0x0]\n"
+      "str s28, [x10, #0x0]\n"
       "103:"  // Height 5: Partial direct writeback: Done
       "b 105f\n"
       "104:"  // Height 5: Full writeback
       "str q24, [x6, #0x0]\n"
       "add x6, x6, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
-      "str q28, [x25, #0x0]\n"
+      "str q25, [x13, #0x0]\n"
+      "str q26, [x12, #0x0]\n"
+      "str q27, [x11, #0x0]\n"
+      "str q28, [x10, #0x0]\n"
       "105:"  // Height 5: Writeback done
       "subs x4, x4, #0x4\n"
       "bgt 86b\n"
       "b 170f\n"
       "106:"  // Height 6
-      "ldr x3, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x3, %x[bias]\n"
       "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x6, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x6, %x[output_ptr]\n"
       "107:"  // Height 6: Column loop
       "cbz x3, 108f\n"
       "ldr q24, [x3, #0x0]\n"
-      "add x3, x3, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x3, x3, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "mov v28.16b, v24.16b\n"
       "mov v29.16b, v24.16b\n"
       "b 113f\n"
       "108:"  // Height 6: no bias
       "tbz %x[flags], #0, 112f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
+      "add x10, x11, x26, LSL #2\n"
       "cmp x4, #0x4\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "add x26, x27, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
+      "add x9, x10, x26, LSL #2\n"
       "bge 111f\n"
       "tbz x4, #1, 109f\n"
       "ldr d24, [x6], #0x8\n"
-      "mov x20, #0x8\n"
-      "ldr d25, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "ldr d29, [x24], #0x8\n"
+      "mov x26, #0x8\n"
+      "ldr d25, [x13], #0x8\n"
+      "ldr d26, [x12], #0x8\n"
+      "ldr d27, [x11], #0x8\n"
+      "ldr d28, [x10], #0x8\n"
+      "ldr d29, [x9], #0x8\n"
       "tbz x4, #0, 110f\n"
       "ld1 { v24.s }[2], [x6]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v28.s }[2], [x25]\n"
-      "ld1 { v29.s }[2], [x24]\n"
+      "ld1 { v25.s }[2], [x13]\n"
+      "ld1 { v26.s }[2], [x12]\n"
+      "ld1 { v27.s }[2], [x11]\n"
+      "ld1 { v28.s }[2], [x10]\n"
+      "ld1 { v29.s }[2], [x9]\n"
       "b 110f\n"
       "109:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s24, [x6, #0x0]\n"
-      "mov x20, #0x0\n"
-      "ldr s25, [x28, #0x0]\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s28, [x25, #0x0]\n"
-      "ldr s29, [x24, #0x0]\n"
+      "mov x26, #0x0\n"
+      "ldr s25, [x13, #0x0]\n"
+      "ldr s26, [x12, #0x0]\n"
+      "ldr s27, [x11, #0x0]\n"
+      "ldr s28, [x10, #0x0]\n"
+      "ldr s29, [x9, #0x0]\n"
       "110:"  // Height 6: Partial accumulate: Done
-      "sub x6, x6, x20\n"
+      "sub x6, x6, x26\n"
       "b 113f\n"
       "111:"  // Height 6: full accumulate
       "ldr q24, [x6, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
-      "ldr q28, [x25, #0x0]\n"
-      "ldr q29, [x24, #0x0]\n"
+      "ldr q25, [x13, #0x0]\n"
+      "ldr q26, [x12, #0x0]\n"
+      "ldr q27, [x11, #0x0]\n"
+      "ldr q28, [x10, #0x0]\n"
+      "ldr q29, [x9, #0x0]\n"
       "b 113f\n"
       "112:"  // Height 6: no accumulate
       "movi v24.16b, #0x0\n"
@@ -1208,34 +1206,34 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "113:"  // Height 6: setup done
       "mov x7, #0x0\n"
       "114:"  // Height 6: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w8, [x20, x7, LSL #0x2]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w8, [x26, x7, LSL #0x2]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 115f\n"
-      "ldr x20, [%x[input_ptr], x7, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x17, [x20, #0x0]\n"
-      "ldr x16, [x20, #0x8]\n"
-      "ldr x15, [x20, #0x10]\n"
-      "ldr x14, [x20, #0x18]\n"
-      "ldr x13, [x20, #0x20]\n"
-      "ldr x12, [x20, #0x28]\n"
+      "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
+      "add x26, x26, x27, LSL #3\n"
+      "ldr x17, [x26, #0x0]\n"
+      "ldr x16, [x26, #0x8]\n"
+      "ldr x15, [x26, #0x10]\n"
+      "ldr x14, [x26, #0x18]\n"
+      "ldr x13, [x26, #0x20]\n"
+      "ldr x12, [x26, #0x28]\n"
       "cbnz x7, 116f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x17, x17, x20, LSL #2\n"
-      "add x16, x16, x20, LSL #2\n"
-      "add x15, x15, x20, LSL #2\n"
-      "add x14, x14, x20, LSL #2\n"
-      "add x13, x13, x20, LSL #2\n"
-      "add x12, x12, x20, LSL #2\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x17, x17, x26, LSL #2\n"
+      "add x16, x16, x26, LSL #2\n"
+      "add x15, x15, x26, LSL #2\n"
+      "add x14, x14, x26, LSL #2\n"
+      "add x13, x13, x26, LSL #2\n"
+      "add x12, x12, x26, LSL #2\n"
       "b 116f\n"
       "115:"  // Height 6: setup direct input
       "mov x17, %x[input_ptr]\n"
-      "add x16, x17, x21, LSL #2\n"
-      "add x15, x16, x21, LSL #2\n"
-      "add x14, x15, x21, LSL #2\n"
-      "add x13, x14, x21, LSL #2\n"
-      "add x12, x13, x21, LSL #2\n"
+      "add x16, x17, x27, LSL #2\n"
+      "add x15, x16, x27, LSL #2\n"
+      "add x14, x15, x27, LSL #2\n"
+      "add x13, x14, x27, LSL #2\n"
+      "add x12, x13, x27, LSL #2\n"
       "116:"  // Height 6: input setup done
       "cmp x8, #0x4\n"
       "blt 119f\n"
@@ -1279,13 +1277,13 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "ldr x26, [x17, #0x8]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
-      "ldr x25, [x16, #0x8]\n"
+      "ldr x11, [x16, #0x8]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
-      "ldr x24, [x15, #0x8]\n"
+      "ldr x10, [x15, #0x8]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
-      "ldr x23, [x14, #0x8]\n"
+      "sub x8, x8, #0x4\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
-      "ldr x22, [x13, #0x8]\n"
+      "cmp x8, #0x8\n"
       "fmla v29.4s, v10.4s, v5.s[2]\n"
       "ldr d10, [x5, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
@@ -1301,26 +1299,26 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "fmla v29.4s, v11.4s, v5.s[3]\n"
       "ldr d5, [x12, #0x0]\n"
       "ldr d11, [x5, #0x30]\n"
-      "sub x8, x8, #0x4\n"
-      "ldr x21, [x12, #0x8]\n"
-      "cmp x8, #0x8\n"
-      "ldr x20, [x5, #0x38]\n"
       "mov v8.d[1], x9\n"
-      "prfm pldl1keep, [x17, #0x80]\n"
+      "ldr x9, [x14, #0x8]\n"
       "mov v9.d[1], x28\n"
-      "prfm pldl1keep, [x16, #0x80]\n"
+      "ldr x28, [x13, #0x8]\n"
       "mov v10.d[1], x27\n"
-      "prfm pldl1keep, [x15, #0x80]\n"
+      "ldr x27, [x12, #0x8]\n"
       "mov v0.d[1], x26\n"
+      "ldr x26, [x5, #0x38]\n"
+      "mov v1.d[1], x11\n"
+      "prfm pldl1keep, [x17, #0x80]\n"
+      "mov v2.d[1], x10\n"
+      "prfm pldl1keep, [x16, #0x80]\n"
+      "mov v3.d[1], x9\n"
+      "prfm pldl1keep, [x15, #0x80]\n"
+      "mov v4.d[1], x28\n"
       "prfm pldl1keep, [x14, #0x80]\n"
-      "mov v1.d[1], x25\n"
+      "mov v5.d[1], x27\n"
       "prfm pldl1keep, [x13, #0x80]\n"
-      "mov v2.d[1], x24\n"
+      "mov v11.d[1], x26\n"
       "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v3.d[1], x23\n"
-      "mov v4.d[1], x22\n"
-      "mov v5.d[1], x21\n"
-      "mov v11.d[1], x20\n"
       "bge 117b\n"
       "118:"  // Height 6: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
@@ -1372,42 +1370,42 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr s18, [x13], #0x4\n"
       "ldr s17, [x12], #0x4\n"
       "ldr q16, [x5, #0x0]\n"
-      "add x5, x5, #0x10\n"
       "fmla v24.4s, v16.4s, v22.s[0]\n"
       "fmla v25.4s, v16.4s, v21.s[0]\n"
+      "add x5, x5, #0x10\n"
       "fmla v26.4s, v16.4s, v20.s[0]\n"
       "fmla v27.4s, v16.4s, v19.s[0]\n"
       "fmla v28.4s, v16.4s, v18.s[0]\n"
       "fmla v29.4s, v16.4s, v17.s[0]\n"
       "cbnz x8, 120b\n"
       "121:"  // Height 6: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x7, x7, #0x1\n"
-      "cmp x7, x20\n"
+      "cmp x7, x26\n"
       "bne 114b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
+      "add x10, x11, x26, LSL #2\n"
+      "add x9, x10, x26, LSL #2\n"
       "prfm pstl1keep, [x6, #0x0]\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x26, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x24, x25, x20, LSL #2\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x12, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
+      "prfm pstl1keep, [x10, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "tbz %x[flags], #1, 122f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
-      "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "fmin v24.4s, v24.4s, v17.4s\n"
-      "fmin v25.4s, v25.4s, v17.4s\n"
-      "fmin v26.4s, v26.4s, v17.4s\n"
-      "fmin v27.4s, v27.4s, v17.4s\n"
-      "fmin v28.4s, v28.4s, v17.4s\n"
-      "fmin v29.4s, v29.4s, v17.4s\n"
+      "add x26, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x26]\n"
+      "fmin v24.4s, v24.4s, v16.4s\n"
+      "fmin v25.4s, v25.4s, v16.4s\n"
+      "fmin v26.4s, v26.4s, v16.4s\n"
+      "fmin v27.4s, v27.4s, v16.4s\n"
+      "fmin v28.4s, v28.4s, v16.4s\n"
+      "fmin v29.4s, v29.4s, v16.4s\n"
+      "add x26, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v16.4s }, [x26]\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
       "fmax v25.4s, v25.4s, v16.4s\n"
       "fmax v26.4s, v26.4s, v16.4s\n"
@@ -1419,51 +1417,51 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bge 125f\n"
       "tbz x4, #1, 123f\n"
       "str d24, [x6], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
-      "str d28, [x25], #0x8\n"
-      "str d29, [x24], #0x8\n"
+      "str d25, [x13], #0x8\n"
+      "str d26, [x12], #0x8\n"
+      "str d27, [x11], #0x8\n"
+      "str d28, [x10], #0x8\n"
+      "str d29, [x9], #0x8\n"
       "tbz x4, #0, 124f\n"
       "st1 { v24.s }[2], [x6]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
-      "st1 { v28.s }[2], [x25]\n"
-      "st1 { v29.s }[2], [x24]\n"
+      "st1 { v25.s }[2], [x13]\n"
+      "st1 { v26.s }[2], [x12]\n"
+      "st1 { v27.s }[2], [x11]\n"
+      "st1 { v28.s }[2], [x10]\n"
+      "st1 { v29.s }[2], [x9]\n"
       "b 124f\n"
       "123:"  // Height 6: Partial direct writeback: partial_1_0
       "str s24, [x6, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
-      "str s28, [x25, #0x0]\n"
-      "str s29, [x24, #0x0]\n"
+      "str s25, [x13, #0x0]\n"
+      "str s26, [x12, #0x0]\n"
+      "str s27, [x11, #0x0]\n"
+      "str s28, [x10, #0x0]\n"
+      "str s29, [x9, #0x0]\n"
       "124:"  // Height 6: Partial direct writeback: Done
       "b 126f\n"
       "125:"  // Height 6: Full writeback
       "str q24, [x6, #0x0]\n"
       "add x6, x6, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
-      "str q28, [x25, #0x0]\n"
-      "str q29, [x24, #0x0]\n"
+      "str q25, [x13, #0x0]\n"
+      "str q26, [x12, #0x0]\n"
+      "str q27, [x11, #0x0]\n"
+      "str q28, [x10, #0x0]\n"
+      "str q29, [x9, #0x0]\n"
       "126:"  // Height 6: Writeback done
       "subs x4, x4, #0x4\n"
       "bgt 107b\n"
       "b 170f\n"
       "127:"  // Height 7
-      "ldr x3, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x3, %x[bias]\n"
       "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x6, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x6, %x[output_ptr]\n"
       "128:"  // Height 7: Column loop
       "cbz x3, 129f\n"
       "ldr q24, [x3, #0x0]\n"
-      "add x3, x3, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x3, x3, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "mov v28.16b, v24.16b\n"
       "mov v29.16b, v24.16b\n"
@@ -1471,53 +1469,53 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "b 134f\n"
       "129:"  // Height 7: no bias
       "tbz %x[flags], #0, 133f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
+      "add x10, x11, x26, LSL #2\n"
+      "add x9, x10, x26, LSL #2\n"
       "cmp x4, #0x4\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "add x26, x27, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x28, x9, x26, LSL #2\n"
       "bge 132f\n"
       "tbz x4, #1, 130f\n"
       "ldr d24, [x6], #0x8\n"
-      "mov x20, #0x8\n"
-      "ldr d25, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d30, [x23], #0x8\n"
+      "mov x26, #0x8\n"
+      "ldr d25, [x13], #0x8\n"
+      "ldr d26, [x12], #0x8\n"
+      "ldr d27, [x11], #0x8\n"
+      "ldr d28, [x10], #0x8\n"
+      "ldr d29, [x9], #0x8\n"
+      "ldr d30, [x28], #0x8\n"
       "tbz x4, #0, 131f\n"
       "ld1 { v24.s }[2], [x6]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v28.s }[2], [x25]\n"
-      "ld1 { v29.s }[2], [x24]\n"
-      "ld1 { v30.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x13]\n"
+      "ld1 { v26.s }[2], [x12]\n"
+      "ld1 { v27.s }[2], [x11]\n"
+      "ld1 { v28.s }[2], [x10]\n"
+      "ld1 { v29.s }[2], [x9]\n"
+      "ld1 { v30.s }[2], [x28]\n"
       "b 131f\n"
       "130:"  // Height 7: Partial accumulate: partial_1_0
       "ldr s24, [x6, #0x0]\n"
-      "mov x20, #0x0\n"
-      "ldr s25, [x28, #0x0]\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s28, [x25, #0x0]\n"
-      "ldr s29, [x24, #0x0]\n"
-      "ldr s30, [x23, #0x0]\n"
+      "mov x26, #0x0\n"
+      "ldr s25, [x13, #0x0]\n"
+      "ldr s26, [x12, #0x0]\n"
+      "ldr s27, [x11, #0x0]\n"
+      "ldr s28, [x10, #0x0]\n"
+      "ldr s29, [x9, #0x0]\n"
+      "ldr s30, [x28, #0x0]\n"
       "131:"  // Height 7: Partial accumulate: Done
-      "sub x6, x6, x20\n"
+      "sub x6, x6, x26\n"
       "b 134f\n"
       "132:"  // Height 7: full accumulate
       "ldr q24, [x6, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
-      "ldr q28, [x25, #0x0]\n"
-      "ldr q29, [x24, #0x0]\n"
-      "ldr q30, [x23, #0x0]\n"
+      "ldr q25, [x13, #0x0]\n"
+      "ldr q26, [x12, #0x0]\n"
+      "ldr q27, [x11, #0x0]\n"
+      "ldr q28, [x10, #0x0]\n"
+      "ldr q29, [x9, #0x0]\n"
+      "ldr q30, [x28, #0x0]\n"
       "b 134f\n"
       "133:"  // Height 7: no accumulate
       "movi v24.16b, #0x0\n"
@@ -1530,37 +1528,37 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "134:"  // Height 7: setup done
       "mov x7, #0x0\n"
       "135:"  // Height 7: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w8, [x20, x7, LSL #0x2]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w8, [x26, x7, LSL #0x2]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 136f\n"
-      "ldr x20, [%x[input_ptr], x7, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x17, [x20, #0x0]\n"
-      "ldr x16, [x20, #0x8]\n"
-      "ldr x15, [x20, #0x10]\n"
-      "ldr x14, [x20, #0x18]\n"
-      "ldr x13, [x20, #0x20]\n"
-      "ldr x12, [x20, #0x28]\n"
-      "ldr x11, [x20, #0x30]\n"
+      "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
+      "add x26, x26, x27, LSL #3\n"
+      "ldr x17, [x26, #0x0]\n"
+      "ldr x16, [x26, #0x8]\n"
+      "ldr x15, [x26, #0x10]\n"
+      "ldr x14, [x26, #0x18]\n"
+      "ldr x13, [x26, #0x20]\n"
+      "ldr x12, [x26, #0x28]\n"
+      "ldr x11, [x26, #0x30]\n"
       "cbnz x7, 137f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x17, x17, x20, LSL #2\n"
-      "add x16, x16, x20, LSL #2\n"
-      "add x15, x15, x20, LSL #2\n"
-      "add x14, x14, x20, LSL #2\n"
-      "add x13, x13, x20, LSL #2\n"
-      "add x12, x12, x20, LSL #2\n"
-      "add x11, x11, x20, LSL #2\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x17, x17, x26, LSL #2\n"
+      "add x16, x16, x26, LSL #2\n"
+      "add x15, x15, x26, LSL #2\n"
+      "add x14, x14, x26, LSL #2\n"
+      "add x13, x13, x26, LSL #2\n"
+      "add x12, x12, x26, LSL #2\n"
+      "add x11, x11, x26, LSL #2\n"
       "b 137f\n"
       "136:"  // Height 7: setup direct input
       "mov x17, %x[input_ptr]\n"
-      "add x16, x17, x21, LSL #2\n"
-      "add x15, x16, x21, LSL #2\n"
-      "add x14, x15, x21, LSL #2\n"
-      "add x13, x14, x21, LSL #2\n"
-      "add x12, x13, x21, LSL #2\n"
-      "add x11, x12, x21, LSL #2\n"
+      "add x16, x17, x27, LSL #2\n"
+      "add x15, x16, x27, LSL #2\n"
+      "add x14, x15, x27, LSL #2\n"
+      "add x13, x14, x27, LSL #2\n"
+      "add x12, x13, x27, LSL #2\n"
+      "add x11, x12, x27, LSL #2\n"
       "137:"  // Height 7: input setup done
       "cmp x8, #0x4\n"
       "blt 140f\n"
@@ -1597,27 +1595,27 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr d8, [x5, #0x0]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
-      "ldr x10, [x5, #0x8]\n"
+      "ldr x26, [x5, #0x8]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
-      "ldr x9, [x5, #0x18]\n"
+      "ldr x10, [x5, #0x18]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
-      "ldr x28, [x5, #0x28]\n"
+      "ldr x9, [x5, #0x28]\n"
       "fmla v29.4s, v9.4s, v5.s[1]\n"
-      "ldr x27, [x17, #0x8]\n"
+      "ldr x28, [x17, #0x8]\n"
       "fmla v30.4s, v9.4s, v6.s[1]\n"
       "ldr d9, [x5, #0x10]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "ldr x26, [x16, #0x8]\n"
+      "ldr x27, [x16, #0x8]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
-      "ldr x25, [x15, #0x8]\n"
+      "sub x8, x8, #0x4\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
-      "ldr x24, [x14, #0x8]\n"
+      "cmp x8, #0x8\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
-      "ldr x23, [x13, #0x8]\n"
+      "mov v8.d[1], x26\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
-      "ldr x22, [x12, #0x8]\n"
+      "ldr x26, [x15, #0x8]\n"
       "fmla v29.4s, v10.4s, v5.s[2]\n"
-      "ldr x21, [x11, #0x8]\n"
+      "prfm pldl1keep, [x17, #0x80]\n"
       "fmla v30.4s, v10.4s, v6.s[2]\n"
       "ldr d10, [x5, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
@@ -1635,27 +1633,27 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "fmla v30.4s, v11.4s, v6.s[3]\n"
       "ldr d6, [x11, #0x0]\n"
       "ldr d11, [x5, #0x30]\n"
-      "sub x8, x8, #0x4\n"
-      "ldr x20, [x5, #0x38]\n"
-      "cmp x8, #0x8\n"
-      "mov v8.d[1], x10\n"
-      "prfm pldl1keep, [x17, #0x80]\n"
+      "mov v9.d[1], x10\n"
+      "ldr x10, [x14, #0x8]\n"
+      "mov v10.d[1], x9\n"
+      "ldr x9, [x13, #0x8]\n"
+      "mov v0.d[1], x28\n"
+      "ldr x28, [x12, #0x8]\n"
+      "mov v1.d[1], x27\n"
+      "ldr x27, [x11, #0x8]\n"
+      "mov v2.d[1], x26\n"
+      "ldr x26, [x5, #0x38]\n"
+      "mov v3.d[1], x10\n"
       "prfm pldl1keep, [x16, #0x80]\n"
-      "mov v9.d[1], x9\n"
+      "mov v4.d[1], x9\n"
       "prfm pldl1keep, [x15, #0x80]\n"
-      "mov v10.d[1], x28\n"
+      "mov v5.d[1], x28\n"
       "prfm pldl1keep, [x14, #0x80]\n"
-      "mov v0.d[1], x27\n"
+      "mov v6.d[1], x27\n"
       "prfm pldl1keep, [x13, #0x80]\n"
-      "mov v1.d[1], x26\n"
+      "mov v11.d[1], x26\n"
       "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v2.d[1], x25\n"
       "prfm pldl1keep, [x11, #0x80]\n"
-      "mov v3.d[1], x24\n"
-      "mov v4.d[1], x23\n"
-      "mov v5.d[1], x22\n"
-      "mov v6.d[1], x21\n"
-      "mov v11.d[1], x20\n"
       "bge 138b\n"
       "139:"  // Height 7: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
@@ -1714,9 +1712,9 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr s18, [x12], #0x4\n"
       "ldr s17, [x11], #0x4\n"
       "ldr q16, [x5, #0x0]\n"
-      "add x5, x5, #0x10\n"
       "fmla v24.4s, v16.4s, v23.s[0]\n"
       "fmla v25.4s, v16.4s, v22.s[0]\n"
+      "add x5, x5, #0x10\n"
       "fmla v26.4s, v16.4s, v21.s[0]\n"
       "fmla v27.4s, v16.4s, v20.s[0]\n"
       "fmla v28.4s, v16.4s, v19.s[0]\n"
@@ -1724,36 +1722,36 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "fmla v30.4s, v16.4s, v17.s[0]\n"
       "cbnz x8, 141b\n"
       "142:"  // Height 7: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x7, x7, #0x1\n"
-      "cmp x7, x20\n"
+      "cmp x7, x26\n"
       "bne 135b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
+      "add x10, x11, x26, LSL #2\n"
+      "add x9, x10, x26, LSL #2\n"
+      "add x28, x9, x26, LSL #2\n"
       "prfm pstl1keep, [x6, #0x0]\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x12, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
+      "prfm pstl1keep, [x10, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "prfm pstl1keep, [x28, #0x0]\n"
-      "add x26, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x24, x25, x20, LSL #2\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 143f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
-      "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "fmin v24.4s, v24.4s, v17.4s\n"
-      "fmin v25.4s, v25.4s, v17.4s\n"
-      "fmin v26.4s, v26.4s, v17.4s\n"
-      "fmin v27.4s, v27.4s, v17.4s\n"
-      "fmin v28.4s, v28.4s, v17.4s\n"
-      "fmin v29.4s, v29.4s, v17.4s\n"
-      "fmin v30.4s, v30.4s, v17.4s\n"
+      "add x26, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x26]\n"
+      "fmin v24.4s, v24.4s, v16.4s\n"
+      "fmin v25.4s, v25.4s, v16.4s\n"
+      "fmin v26.4s, v26.4s, v16.4s\n"
+      "fmin v27.4s, v27.4s, v16.4s\n"
+      "fmin v28.4s, v28.4s, v16.4s\n"
+      "fmin v29.4s, v29.4s, v16.4s\n"
+      "fmin v30.4s, v30.4s, v16.4s\n"
+      "add x26, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v16.4s }, [x26]\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
       "fmax v25.4s, v25.4s, v16.4s\n"
       "fmax v26.4s, v26.4s, v16.4s\n"
@@ -1766,59 +1764,58 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bge 146f\n"
       "tbz x4, #1, 144f\n"
       "str d24, [x6], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
-      "str d28, [x25], #0x8\n"
-      "str d29, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
+      "str d25, [x13], #0x8\n"
+      "str d26, [x12], #0x8\n"
+      "str d27, [x11], #0x8\n"
+      "str d28, [x10], #0x8\n"
+      "str d29, [x9], #0x8\n"
+      "str d30, [x28], #0x8\n"
       "tbz x4, #0, 145f\n"
       "st1 { v24.s }[2], [x6]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
-      "st1 { v28.s }[2], [x25]\n"
-      "st1 { v29.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x13]\n"
+      "st1 { v26.s }[2], [x12]\n"
+      "st1 { v27.s }[2], [x11]\n"
+      "st1 { v28.s }[2], [x10]\n"
+      "st1 { v29.s }[2], [x9]\n"
+      "st1 { v30.s }[2], [x28]\n"
       "b 145f\n"
       "144:"  // Height 7: Partial direct writeback: partial_1_0
       "str s24, [x6, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
-      "str s28, [x25, #0x0]\n"
-      "str s29, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
+      "str s25, [x13, #0x0]\n"
+      "str s26, [x12, #0x0]\n"
+      "str s27, [x11, #0x0]\n"
+      "str s28, [x10, #0x0]\n"
+      "str s29, [x9, #0x0]\n"
+      "str s30, [x28, #0x0]\n"
       "145:"  // Height 7: Partial direct writeback: Done
       "b 147f\n"
       "146:"  // Height 7: Full writeback
       "str q24, [x6, #0x0]\n"
       "add x6, x6, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
-      "str q28, [x25, #0x0]\n"
-      "str q29, [x24, #0x0]\n"
-      "str q30, [x23, #0x0]\n"
+      "str q25, [x13, #0x0]\n"
+      "str q26, [x12, #0x0]\n"
+      "str q27, [x11, #0x0]\n"
+      "str q28, [x10, #0x0]\n"
+      "str q29, [x9, #0x0]\n"
+      "str q30, [x28, #0x0]\n"
       "147:"  // Height 7: Writeback done
       "subs x4, x4, #0x4\n"
       "bgt 128b\n"
       "b 170f\n"
       "148:"  // Height 8
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "mov x20, #0x20\n"
-      "ldr x6, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "ldr x3, [%x[args_ptr], %[offsetof_bias]]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "mov x26, #0x20\n"
+      "mov x3, %x[bias]\n"
       "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
-      "madd x20, x21, x20, x6\n"
       "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x6, %x[output_ptr]\n"
+      "madd %x[output_ptr], x27, x26, %x[output_ptr]\n"
       "149:"  // Height 8: Column loop
       "cbz x3, 150f\n"
       "ldr q24, [x3, #0x0]\n"
-      "add x3, x3, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x3, x3, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "mov v28.16b, v24.16b\n"
       "mov v29.16b, v24.16b\n"
@@ -1827,58 +1824,58 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "b 155f\n"
       "150:"  // Height 8: no bias
       "tbz %x[flags], #0, 154f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
+      "add x10, x11, x26, LSL #2\n"
+      "add x9, x10, x26, LSL #2\n"
+      "add x28, x9, x26, LSL #2\n"
       "cmp x4, #0x4\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
-      "add x26, x27, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
+      "add x27, x28, x26, LSL #2\n"
       "bge 153f\n"
       "tbz x4, #1, 151f\n"
       "ldr d24, [x6], #0x8\n"
-      "mov x20, #0x8\n"
-      "ldr d25, [x28], #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d30, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "mov x26, #0x8\n"
+      "ldr d25, [x13], #0x8\n"
+      "ldr d26, [x12], #0x8\n"
+      "ldr d27, [x11], #0x8\n"
+      "ldr d28, [x10], #0x8\n"
+      "ldr d29, [x9], #0x8\n"
+      "ldr d30, [x28], #0x8\n"
+      "ldr d31, [x27], #0x8\n"
       "tbz x4, #0, 152f\n"
       "ld1 { v24.s }[2], [x6]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v28.s }[2], [x25]\n"
-      "ld1 { v29.s }[2], [x24]\n"
-      "ld1 { v30.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v25.s }[2], [x13]\n"
+      "ld1 { v26.s }[2], [x12]\n"
+      "ld1 { v27.s }[2], [x11]\n"
+      "ld1 { v28.s }[2], [x10]\n"
+      "ld1 { v29.s }[2], [x9]\n"
+      "ld1 { v30.s }[2], [x28]\n"
+      "ld1 { v31.s }[2], [x27]\n"
       "b 152f\n"
       "151:"  // Height 8: Partial accumulate: partial_1_0
       "ldr s24, [x6, #0x0]\n"
-      "mov x20, #0x0\n"
-      "ldr s25, [x28, #0x0]\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s28, [x25, #0x0]\n"
-      "ldr s29, [x24, #0x0]\n"
-      "ldr s30, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "mov x26, #0x0\n"
+      "ldr s25, [x13, #0x0]\n"
+      "ldr s26, [x12, #0x0]\n"
+      "ldr s27, [x11, #0x0]\n"
+      "ldr s28, [x10, #0x0]\n"
+      "ldr s29, [x9, #0x0]\n"
+      "ldr s30, [x28, #0x0]\n"
+      "ldr s31, [x27, #0x0]\n"
       "152:"  // Height 8: Partial accumulate: Done
-      "sub x6, x6, x20\n"
+      "sub x6, x6, x26\n"
       "b 155f\n"
       "153:"  // Height 8: full accumulate
       "ldr q24, [x6, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
-      "ldr q28, [x25, #0x0]\n"
-      "ldr q29, [x24, #0x0]\n"
-      "ldr q30, [x23, #0x0]\n"
-      "ldr q31, [x22, #0x0]\n"
+      "ldr q25, [x13, #0x0]\n"
+      "ldr q26, [x12, #0x0]\n"
+      "ldr q27, [x11, #0x0]\n"
+      "ldr q28, [x10, #0x0]\n"
+      "ldr q29, [x9, #0x0]\n"
+      "ldr q30, [x28, #0x0]\n"
+      "ldr q31, [x27, #0x0]\n"
       "b 155f\n"
       "154:"  // Height 8: no accumulate
       "movi v24.16b, #0x0\n"
@@ -1892,40 +1889,40 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "155:"  // Height 8: setup done
       "mov x7, #0x0\n"
       "156:"  // Height 8: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w8, [x20, x7, LSL #0x2]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w8, [x26, x7, LSL #0x2]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 157f\n"
-      "ldr x20, [%x[input_ptr], x7, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x17, [x20, #0x0]\n"
-      "ldr x16, [x20, #0x8]\n"
-      "ldr x15, [x20, #0x10]\n"
-      "ldr x14, [x20, #0x18]\n"
-      "ldr x13, [x20, #0x20]\n"
-      "ldr x12, [x20, #0x28]\n"
-      "ldr x11, [x20, #0x30]\n"
-      "ldr x9, [x20, #0x38]\n"
+      "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
+      "add x26, x26, x27, LSL #3\n"
+      "ldr x17, [x26, #0x0]\n"
+      "ldr x16, [x26, #0x8]\n"
+      "ldr x15, [x26, #0x10]\n"
+      "ldr x14, [x26, #0x18]\n"
+      "ldr x13, [x26, #0x20]\n"
+      "ldr x12, [x26, #0x28]\n"
+      "ldr x11, [x26, #0x30]\n"
+      "ldr x27, [x26, #0x38]\n"
       "cbnz x7, 158f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x17, x17, x20, LSL #2\n"
-      "add x16, x16, x20, LSL #2\n"
-      "add x15, x15, x20, LSL #2\n"
-      "add x14, x14, x20, LSL #2\n"
-      "add x13, x13, x20, LSL #2\n"
-      "add x12, x12, x20, LSL #2\n"
-      "add x11, x11, x20, LSL #2\n"
-      "add x9, x9, x20, LSL #2\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+      "add x17, x17, x26, LSL #2\n"
+      "add x16, x16, x26, LSL #2\n"
+      "add x15, x15, x26, LSL #2\n"
+      "add x14, x14, x26, LSL #2\n"
+      "add x13, x13, x26, LSL #2\n"
+      "add x12, x12, x26, LSL #2\n"
+      "add x11, x11, x26, LSL #2\n"
+      "add x27, x27, x26, LSL #2\n"
       "b 158f\n"
       "157:"  // Height 8: setup direct input
       "mov x17, %x[input_ptr]\n"
-      "add x16, x17, x21, LSL #2\n"
-      "add x15, x16, x21, LSL #2\n"
-      "add x14, x15, x21, LSL #2\n"
-      "add x13, x14, x21, LSL #2\n"
-      "add x12, x13, x21, LSL #2\n"
-      "add x11, x12, x21, LSL #2\n"
-      "add x9, x11, x21, LSL #2\n"
+      "add x16, x17, x27, LSL #2\n"
+      "add x15, x16, x27, LSL #2\n"
+      "add x14, x15, x27, LSL #2\n"
+      "add x13, x14, x27, LSL #2\n"
+      "add x12, x13, x27, LSL #2\n"
+      "add x11, x12, x27, LSL #2\n"
+      "add x27, x11, x27, LSL #2\n"
       "158:"  // Height 8: input setup done
       "cmp x8, #0x4\n"
       "blt 161f\n"
@@ -1937,7 +1934,7 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr q4, [x13, #0x0]\n"
       "ldr q5, [x12, #0x0]\n"
       "ldr q6, [x11, #0x0]\n"
-      "ldr q7, [x9, #0x0]\n"
+      "ldr q7, [x27, #0x0]\n"
       "ldr q8, [x5, #0x0]\n"
       "ldr q9, [x5, #0x10]\n"
       "ldr q10, [x5, #0x20]\n"
@@ -1959,37 +1956,37 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "fmla v30.4s, v8.4s, v6.s[0]\n"
       "add x11, x11, #0x10\n"
       "fmla v31.4s, v8.4s, v7.s[0]\n"
-      "add x9, x9, #0x10\n"
+      "add x27, x27, #0x10\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x5, x5, #0x40\n"
       "ldr d8, [x5, #0x0]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
-      "ldr x22, [x5, #0x8]\n"
+      "ldr x26, [x5, #0x8]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
-      "ldr x28, [x5, #0x18]\n"
+      "sub x8, x8, #0x4\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
-      "ldr x21, [x5, #0x28]\n"
+      "cmp x8, #0x8\n"
       "fmla v29.4s, v9.4s, v5.s[1]\n"
-      "ldr x20, [x17, #0x8]\n"
+      "mov v8.d[1], x26\n"
       "fmla v30.4s, v9.4s, v6.s[1]\n"
-      "ldr x27, [x16, #0x8]\n"
+      "ldr x26, [x5, #0x18]\n"
       "fmla v31.4s, v9.4s, v7.s[1]\n"
       "ldr d9, [x5, #0x10]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "ldr x26, [x15, #0x8]\n"
+      "prfm pldl1keep, [x17, #0x80]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
-      "ldr x25, [x14, #0x8]\n"
+      "prfm pldl1keep, [x16, #0x80]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
-      "ldr x24, [x13, #0x8]\n"
+      "prfm pldl1keep, [x15, #0x80]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
-      "ldr x23, [x12, #0x8]\n"
+      "mov v9.d[1], x26\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
-      "sub x8, x8, #0x4\n"
+      "ldr x26, [x5, #0x28]\n"
       "fmla v29.4s, v10.4s, v5.s[2]\n"
-      "cmp x8, #0x8\n"
+      "prfm pldl1keep, [x14, #0x80]\n"
       "fmla v30.4s, v10.4s, v6.s[2]\n"
-      "mov v8.d[1], x22\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "fmla v31.4s, v10.4s, v7.s[2]\n"
       "ldr d10, [x5, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
@@ -2006,31 +2003,31 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr d5, [x12, #0x0]\n"
       "fmla v30.4s, v11.4s, v6.s[3]\n"
       "ldr d6, [x11, #0x0]\n"
-      "ldr x22, [x11, #0x8]\n"
       "fmla v31.4s, v11.4s, v7.s[3]\n"
-      "ldr d7, [x9, #0x0]\n"
-      "mov v9.d[1], x28\n"
+      "ldr d7, [x27, #0x0]\n"
       "ldr d11, [x5, #0x30]\n"
-      "mov v10.d[1], x21\n"
-      "ldr x21, [x9, #0x8]\n"
-      "mov v0.d[1], x20\n"
-      "ldr x20, [x5, #0x38]\n"
-      "mov v1.d[1], x27\n"
-      "prfm pldl1keep, [x17, #0x80]\n"
+      "mov v10.d[1], x26\n"
+      "ldr x26, [x17, #0x8]\n"
+      "mov v0.d[1], x26\n"
+      "ldr x26, [x16, #0x8]\n"
+      "mov v1.d[1], x26\n"
+      "ldr x26, [x15, #0x8]\n"
       "mov v2.d[1], x26\n"
-      "prfm pldl1keep, [x16, #0x80]\n"
-      "mov v3.d[1], x25\n"
-      "prfm pldl1keep, [x15, #0x80]\n"
-      "mov v4.d[1], x24\n"
-      "prfm pldl1keep, [x14, #0x80]\n"
-      "mov v5.d[1], x23\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
-      "mov v6.d[1], x22\n"
+      "ldr x26, [x14, #0x8]\n"
+      "mov v3.d[1], x26\n"
+      "ldr x26, [x13, #0x8]\n"
+      "mov v4.d[1], x26\n"
+      "ldr x26, [x12, #0x8]\n"
+      "mov v5.d[1], x26\n"
+      "ldr x26, [x11, #0x8]\n"
+      "mov v6.d[1], x26\n"
+      "ldr x26, [x27, #0x8]\n"
+      "mov v7.d[1], x26\n"
+      "ldr x26, [x5, #0x38]\n"
+      "mov v11.d[1], x26\n"
       "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v7.d[1], x21\n"
       "prfm pldl1keep, [x11, #0x80]\n"
-      "mov v11.d[1], x20\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "bge 159b\n"
       "160:"  // Height 8: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
@@ -2048,7 +2045,7 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "fmla v30.4s, v8.4s, v6.s[0]\n"
       "add x11, x11, #0x10\n"
       "fmla v31.4s, v8.4s, v7.s[0]\n"
-      "add x9, x9, #0x10\n"
+      "add x27, x27, #0x10\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
       "sub x8, x8, #0x4\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
@@ -2066,7 +2063,7 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "fmla v31.4s, v9.4s, v7.s[1]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
       "add x5, x5, #0x40\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
@@ -2094,11 +2091,11 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "ldr s20, [x13], #0x4\n"
       "ldr s19, [x12], #0x4\n"
       "ldr s18, [x11], #0x4\n"
-      "ldr s17, [x9], #0x4\n"
+      "ldr s17, [x27], #0x4\n"
       "ldr q16, [x5, #0x0]\n"
-      "add x5, x5, #0x10\n"
       "fmla v24.4s, v16.4s, v0.s[0]\n"
       "fmla v25.4s, v16.4s, v23.s[0]\n"
+      "add x5, x5, #0x10\n"
       "fmla v26.4s, v16.4s, v22.s[0]\n"
       "fmla v27.4s, v16.4s, v21.s[0]\n"
       "fmla v28.4s, v16.4s, v20.s[0]\n"
@@ -2107,39 +2104,39 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "fmla v31.4s, v16.4s, v17.s[0]\n"
       "cbnz x8, 162b\n"
       "163:"  // Height 8: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+      "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x7, x7, #0x1\n"
-      "cmp x7, x20\n"
+      "cmp x7, x26\n"
       "bne 156b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x13, x6, x26, LSL #2\n"
+      "add x12, x13, x26, LSL #2\n"
+      "add x11, x12, x26, LSL #2\n"
+      "add x10, x11, x26, LSL #2\n"
+      "add x9, x10, x26, LSL #2\n"
+      "add x28, x9, x26, LSL #2\n"
+      "add x27, x28, x26, LSL #2\n"
       "prfm pstl1keep, [x6, #0x0]\n"
-      "add x28, x6, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
+      "prfm pstl1keep, [x12, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
+      "prfm pstl1keep, [x10, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "prfm pstl1keep, [x28, #0x0]\n"
-      "add x26, x27, x20, LSL #2\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x24, x25, x20, LSL #2\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 164f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
-      "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "fmin v24.4s, v24.4s, v17.4s\n"
-      "fmin v25.4s, v25.4s, v17.4s\n"
-      "fmin v26.4s, v26.4s, v17.4s\n"
-      "fmin v27.4s, v27.4s, v17.4s\n"
-      "fmin v28.4s, v28.4s, v17.4s\n"
-      "fmin v29.4s, v29.4s, v17.4s\n"
-      "fmin v30.4s, v30.4s, v17.4s\n"
-      "fmin v31.4s, v31.4s, v17.4s\n"
+      "add x26, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v16.4s }, [x26]\n"
+      "fmin v24.4s, v24.4s, v16.4s\n"
+      "fmin v25.4s, v25.4s, v16.4s\n"
+      "fmin v26.4s, v26.4s, v16.4s\n"
+      "fmin v27.4s, v27.4s, v16.4s\n"
+      "fmin v28.4s, v28.4s, v16.4s\n"
+      "fmin v29.4s, v29.4s, v16.4s\n"
+      "fmin v30.4s, v30.4s, v16.4s\n"
+      "fmin v31.4s, v31.4s, v16.4s\n"
+      "add x26, %x[args_ptr], %[offset_min]\n"
+      "ld1r { v16.4s }, [x26]\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
       "fmax v25.4s, v25.4s, v16.4s\n"
       "fmax v26.4s, v26.4s, v16.4s\n"
@@ -2153,62 +2150,62 @@ void a64_hybrid_fp32_mla_8x4_a55 (
       "bge 167f\n"
       "tbz x4, #1, 165f\n"
       "str d24, [x6], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
-      "str d28, [x25], #0x8\n"
-      "str d29, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d25, [x13], #0x8\n"
+      "str d26, [x12], #0x8\n"
+      "str d27, [x11], #0x8\n"
+      "str d28, [x10], #0x8\n"
+      "str d29, [x9], #0x8\n"
+      "str d30, [x28], #0x8\n"
+      "str d31, [x27], #0x8\n"
       "tbz x4, #0, 166f\n"
       "st1 { v24.s }[2], [x6]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
-      "st1 { v28.s }[2], [x25]\n"
-      "st1 { v29.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
-      "st1 { v31.s }[2], [x22]\n"
+      "st1 { v25.s }[2], [x13]\n"
+      "st1 { v26.s }[2], [x12]\n"
+      "st1 { v27.s }[2], [x11]\n"
+      "st1 { v28.s }[2], [x10]\n"
+      "st1 { v29.s }[2], [x9]\n"
+      "st1 { v30.s }[2], [x28]\n"
+      "st1 { v31.s }[2], [x27]\n"
       "b 166f\n"
       "165:"  // Height 8: Partial direct writeback: partial_1_0
       "str s24, [x6, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
-      "str s28, [x25, #0x0]\n"
-      "str s29, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
-      "str s31, [x22, #0x0]\n"
+      "str s25, [x13, #0x0]\n"
+      "str s26, [x12, #0x0]\n"
+      "str s27, [x11, #0x0]\n"
+      "str s28, [x10, #0x0]\n"
+      "str s29, [x9, #0x0]\n"
+      "str s30, [x28, #0x0]\n"
+      "str s31, [x27, #0x0]\n"
       "166:"  // Height 8: Partial direct writeback: Done
       "b 168f\n"
       "167:"  // Height 8: Full writeback
       "str q24, [x6, #0x0]\n"
       "add x6, x6, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
-      "str q28, [x25, #0x0]\n"
-      "str q29, [x24, #0x0]\n"
-      "str q30, [x23, #0x0]\n"
-      "str q31, [x22, #0x0]\n"
+      "str q25, [x13, #0x0]\n"
+      "str q26, [x12, #0x0]\n"
+      "str q27, [x11, #0x0]\n"
+      "str q28, [x10, #0x0]\n"
+      "str q29, [x9, #0x0]\n"
+      "str q30, [x28, #0x0]\n"
+      "str q31, [x27, #0x0]\n"
       "168:"  // Height 8: Writeback done
       "subs x4, x4, #0x4\n"
       "bgt 149b\n"
       "subs %x[M], %x[M], #0x8\n"
       "beq 170f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 169f\n"
-      "add x21, x21, #0x8\n"
-      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+      "add x27, x27, #0x8\n"
+      "str x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "b 1b\n"
       "169:"  // Update direct input
-      "mov x20, #0x20\n"
-      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
+      "mov x26, #0x20\n"
+      "madd %x[input_ptr], x26, x27, %x[input_ptr]\n"
       "b 1b\n"
       "170:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp
index 6401b01607..36356dfb35 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp
@@ -47,19 +47,18 @@ void a64_hybrid_fp32_mla_8x4 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void a64_hybrid_fp32_mla_8x4 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -106,10 +104,10 @@ void a64_hybrid_fp32_mla_8x4 (
       "cmp %x[M], #0x2\n"
       "bgt 43f\n"
       "beq 22f\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x14, 3f\n"
       "ldr q24, [x14, #0x0]\n"
@@ -140,8 +138,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "mov x10, #0x0\n"
       "9:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 10f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -164,37 +162,37 @@ void a64_hybrid_fp32_mla_8x4 (
       "blt 13f\n"
       "12:"  // Height 1: Multiply loop: Main loop head
       "fmla v24.4s, v8.4s, v0.s[0]\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "sub x9, x9, #0x4\n"
       "add x28, x28, #0x10\n"
+      "fmla v24.4s, v10.4s, v0.s[2]\n"
       "cmp x9, #0x8\n"
       "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
       "ldr q8, [x12, #0x0]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "ldr q9, [x12, #0x10]\n"
-      "fmla v24.4s, v10.4s, v0.s[2]\n"
       "ldr q10, [x12, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr q0, [x28, #0x0]\n"
       "ldr q11, [x12, #0x30]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "bge 12b\n"
       "13:"  // Height 1: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x28, x28, #0x10\n"
       "sub x9, x9, #0x4\n"
-      "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "add x12, x12, #0x40\n"
       "14:"  // Height 1: Multiply loop: Main loop skip
       "cbz x9, 16f\n"
       "15:"  // Height 1: Multiply loop: Odd block loop
       "ldr s17, [x28], #0x4\n"
       "ldr q16, [x12, #0x0]\n"
       "sub x9, x9, #0x1\n"
-      "add x12, x12, #0x10\n"
       "fmla v24.4s, v16.4s, v17.s[0]\n"
+      "add x12, x12, #0x10\n"
       "cbnz x9, 15b\n"
       "16:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -203,9 +201,9 @@ void a64_hybrid_fp32_mla_8x4 (
       "bne 9b\n"
       "prfm pstl1keep, [x11, #0x0]\n"
       "tbz %x[flags], #1, 17f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v24.4s, v24.4s, v17.4s\n"
       "fmax v24.4s, v24.4s, v16.4s\n"
@@ -229,40 +227,40 @@ void a64_hybrid_fp32_mla_8x4 (
       "bgt 2b\n"
       "b 170f\n"
       "22:"  // Height 2
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "23:"  // Height 2: Column loop
       "cbz x14, 24f\n"
       "ldr q24, [x14, #0x0]\n"
-      "add x14, x14, #0x10\n"
       "mov v25.16b, v24.16b\n"
+      "add x14, x14, #0x10\n"
       "b 29f\n"
       "24:"  // Height 2: no bias
       "tbz %x[flags], #0, 28f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x13, #0x4\n"
-      "add x28, x11, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "bge 27f\n"
       "tbz x13, #1, 25f\n"
       "ldr d24, [x11], #0x8\n"
-      "ldr d25, [x28], #0x8\n"
+      "ldr d25, [x27], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x13, #0, 26f\n"
       "ld1 { v24.s }[2], [x11]\n"
-      "ld1 { v25.s }[2], [x28]\n"
+      "ld1 { v25.s }[2], [x27]\n"
       "b 26f\n"
       "25:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s24, [x11, #0x0]\n"
-      "ldr s25, [x28, #0x0]\n"
+      "ldr s25, [x27, #0x0]\n"
       "mov x20, #0x0\n"
       "26:"  // Height 2: Partial accumulate: Done
       "sub x11, x11, x20\n"
       "b 29f\n"
       "27:"  // Height 2: full accumulate
       "ldr q24, [x11, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
+      "ldr q25, [x27, #0x0]\n"
       "b 29f\n"
       "28:"  // Height 2: no accumulate
       "movi v24.16b, #0x0\n"
@@ -271,8 +269,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "mov x10, #0x0\n"
       "30:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 31f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -302,37 +300,37 @@ void a64_hybrid_fp32_mla_8x4 (
       "fmla v25.4s, v8.4s, v1.s[0]\n"
       "sub x9, x9, #0x4\n"
       "add x28, x28, #0x10\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
+      "fmla v25.4s, v9.4s, v1.s[1]\n"
       "add x27, x27, #0x10\n"
       "cmp x9, #0x8\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
+      "fmla v24.4s, v10.4s, v0.s[2]\n"
+      "fmla v25.4s, v10.4s, v1.s[2]\n"
       "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
       "ldr q8, [x12, #0x0]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
-      "fmla v25.4s, v9.4s, v1.s[1]\n"
       "ldr q9, [x12, #0x10]\n"
-      "fmla v24.4s, v10.4s, v0.s[2]\n"
-      "fmla v25.4s, v10.4s, v1.s[2]\n"
       "ldr q10, [x12, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
-      "ldr q0, [x28, #0x0]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
+      "ldr q0, [x28, #0x0]\n"
       "ldr q1, [x27, #0x0]\n"
       "ldr q11, [x12, #0x30]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "bge 33b\n"
       "34:"  // Height 2: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
       "fmla v25.4s, v8.4s, v1.s[0]\n"
       "add x28, x28, #0x10\n"
       "add x27, x27, #0x10\n"
-      "sub x9, x9, #0x4\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "sub x9, x9, #0x4\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
+      "add x12, x12, #0x40\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
       "35:"  // Height 2: Multiply loop: Main loop skip
@@ -342,9 +340,9 @@ void a64_hybrid_fp32_mla_8x4 (
       "ldr s17, [x27], #0x4\n"
       "sub x9, x9, #0x1\n"
       "ldr q16, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
       "fmla v24.4s, v16.4s, v18.s[0]\n"
       "fmla v25.4s, v16.4s, v17.s[0]\n"
+      "add x12, x12, #0x10\n"
       "cbnz x9, 36b\n"
       "37:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -352,13 +350,13 @@ void a64_hybrid_fp32_mla_8x4 (
       "cmp x10, x20\n"
       "bne 30b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x27, x11, x20, LSL #2\n"
       "prfm pstl1keep, [x11, #0x0]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v24.4s, v24.4s, v17.4s\n"
       "fmin v25.4s, v25.4s, v17.4s\n"
@@ -369,65 +367,65 @@ void a64_hybrid_fp32_mla_8x4 (
       "bge 41f\n"
       "tbz x13, #1, 39f\n"
       "str d24, [x11], #0x8\n"
-      "str d25, [x28], #0x8\n"
+      "str d25, [x27], #0x8\n"
       "tbz x13, #0, 40f\n"
       "st1 { v24.s }[2], [x11]\n"
-      "st1 { v25.s }[2], [x28]\n"
+      "st1 { v25.s }[2], [x27]\n"
       "b 40f\n"
       "39:"  // Height 2: Partial direct writeback: partial_1_0
       "str s24, [x11, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
+      "str s25, [x27, #0x0]\n"
       "40:"  // Height 2: Partial direct writeback: Done
       "b 42f\n"
       "41:"  // Height 2: Full writeback
       "str q24, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "str q25, [x28, #0x0]\n"
+      "str q25, [x27, #0x0]\n"
       "42:"  // Height 2: Writeback done
       "subs x13, x13, #0x4\n"
       "bgt 23b\n"
       "b 170f\n"
       "43:"  // Height 3
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "44:"  // Height 3: Column loop
       "cbz x14, 45f\n"
       "ldr q24, [x14, #0x0]\n"
-      "add x14, x14, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x14, x14, #0x10\n"
       "b 50f\n"
       "45:"  // Height 3: no bias
       "tbz %x[flags], #0, 49f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x27, x11, x20, LSL #2\n"
       "cmp x13, #0x4\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x26, x27, x20, LSL #2\n"
       "bge 48f\n"
       "tbz x13, #1, 46f\n"
       "ldr d24, [x11], #0x8\n"
-      "ldr d25, [x28], #0x8\n"
+      "ldr d25, [x27], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d26, [x27], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
       "tbz x13, #0, 47f\n"
       "ld1 { v24.s }[2], [x11]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
+      "ld1 { v25.s }[2], [x27]\n"
+      "ld1 { v26.s }[2], [x26]\n"
       "b 47f\n"
       "46:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s24, [x11, #0x0]\n"
-      "ldr s25, [x28, #0x0]\n"
+      "ldr s25, [x27, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s26, [x27, #0x0]\n"
+      "ldr s26, [x26, #0x0]\n"
       "47:"  // Height 3: Partial accumulate: Done
       "sub x11, x11, x20\n"
       "b 50f\n"
       "48:"  // Height 3: full accumulate
       "ldr q24, [x11, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
+      "ldr q25, [x27, #0x0]\n"
+      "ldr q26, [x26, #0x0]\n"
       "b 50f\n"
       "49:"  // Height 3: no accumulate
       "movi v24.16b, #0x0\n"
@@ -437,8 +435,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "mov x10, #0x0\n"
       "51:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 52f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -473,17 +471,14 @@ void a64_hybrid_fp32_mla_8x4 (
       "sub x9, x9, #0x4\n"
       "add x28, x28, #0x10\n"
       "fmla v26.4s, v8.4s, v2.s[0]\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x27, x27, #0x10\n"
       "add x26, x26, #0x10\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
+      "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "fmla v26.4s, v9.4s, v2.s[1]\n"
       "cmp x9, #0x8\n"
       "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "ldr q8, [x12, #0x0]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
-      "fmla v25.4s, v9.4s, v1.s[1]\n"
-      "fmla v26.4s, v9.4s, v2.s[1]\n"
       "ldr q9, [x12, #0x10]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
@@ -491,11 +486,14 @@ void a64_hybrid_fp32_mla_8x4 (
       "ldr q10, [x12, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr q0, [x28, #0x0]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
       "ldr q1, [x27, #0x0]\n"
       "fmla v26.4s, v11.4s, v2.s[3]\n"
       "ldr q2, [x26, #0x0]\n"
       "ldr q11, [x12, #0x30]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 54b\n"
       "55:"  // Height 3: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
@@ -503,17 +501,17 @@ void a64_hybrid_fp32_mla_8x4 (
       "add x28, x28, #0x10\n"
       "add x27, x27, #0x10\n"
       "fmla v26.4s, v8.4s, v2.s[0]\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x26, x26, #0x10\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
       "sub x9, x9, #0x4\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
-      "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x12, x12, #0x40\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
@@ -526,10 +524,10 @@ void a64_hybrid_fp32_mla_8x4 (
       "sub x9, x9, #0x1\n"
       "ldr s17, [x26], #0x4\n"
       "ldr q16, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
       "fmla v24.4s, v16.4s, v19.s[0]\n"
       "fmla v25.4s, v16.4s, v18.s[0]\n"
       "fmla v26.4s, v16.4s, v17.s[0]\n"
+      "add x12, x12, #0x10\n"
       "cbnz x9, 57b\n"
       "58:"  // Height 3: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -537,15 +535,15 @@ void a64_hybrid_fp32_mla_8x4 (
       "cmp x10, x20\n"
       "bne 51b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x27, x11, x20, LSL #2\n"
+      "add x26, x27, x20, LSL #2\n"
       "prfm pstl1keep, [x11, #0x0]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x27, x28, x20, LSL #2\n"
       "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x26, #0x0]\n"
       "tbz %x[flags], #1, 59f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v24.4s, v24.4s, v17.4s\n"
       "fmin v25.4s, v25.4s, v17.4s\n"
@@ -558,75 +556,75 @@ void a64_hybrid_fp32_mla_8x4 (
       "bge 62f\n"
       "tbz x13, #1, 60f\n"
       "str d24, [x11], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
+      "str d25, [x27], #0x8\n"
+      "str d26, [x26], #0x8\n"
       "tbz x13, #0, 61f\n"
       "st1 { v24.s }[2], [x11]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
+      "st1 { v25.s }[2], [x27]\n"
+      "st1 { v26.s }[2], [x26]\n"
       "b 61f\n"
       "60:"  // Height 3: Partial direct writeback: partial_1_0
       "str s24, [x11, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
+      "str s25, [x27, #0x0]\n"
+      "str s26, [x26, #0x0]\n"
       "61:"  // Height 3: Partial direct writeback: Done
       "b 63f\n"
       "62:"  // Height 3: Full writeback
       "str q24, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
+      "str q25, [x27, #0x0]\n"
+      "str q26, [x26, #0x0]\n"
       "63:"  // Height 3: Writeback done
       "subs x13, x13, #0x4\n"
       "bgt 44b\n"
       "b 170f\n"
       "64:"  // Height 4
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "65:"  // Height 4: Column loop
       "cbz x14, 66f\n"
       "ldr q24, [x14, #0x0]\n"
-      "add x14, x14, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x14, x14, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "b 71f\n"
       "66:"  // Height 4: no bias
       "tbz %x[flags], #0, 70f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x13, #0x4\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
+      "cmp x13, #0x4\n"
+      "add x25, x26, x20, LSL #2\n"
       "bge 69f\n"
       "tbz x13, #1, 67f\n"
       "ldr d24, [x11], #0x8\n"
-      "ldr d25, [x28], #0x8\n"
+      "ldr d25, [x27], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d27, [x25], #0x8\n"
       "tbz x13, #0, 68f\n"
       "ld1 { v24.s }[2], [x11]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
+      "ld1 { v25.s }[2], [x27]\n"
+      "ld1 { v26.s }[2], [x26]\n"
+      "ld1 { v27.s }[2], [x25]\n"
       "b 68f\n"
       "67:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s24, [x11, #0x0]\n"
-      "ldr s25, [x28, #0x0]\n"
+      "ldr s25, [x27, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
+      "ldr s26, [x26, #0x0]\n"
+      "ldr s27, [x25, #0x0]\n"
       "68:"  // Height 4: Partial accumulate: Done
       "sub x11, x11, x20\n"
       "b 71f\n"
       "69:"  // Height 4: full accumulate
       "ldr q24, [x11, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
+      "ldr q25, [x27, #0x0]\n"
+      "ldr q26, [x26, #0x0]\n"
+      "ldr q27, [x25, #0x0]\n"
       "b 71f\n"
       "70:"  // Height 4: no accumulate
       "movi v24.16b, #0x0\n"
@@ -637,8 +635,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "mov x10, #0x0\n"
       "72:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 73f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -680,24 +678,23 @@ void a64_hybrid_fp32_mla_8x4 (
       "fmla v27.4s, v8.4s, v3.s[0]\n"
       "add x27, x27, #0x10\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x9, #0x8\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
-      "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "ldr q8, [x12, #0x0]\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "add x25, x25, #0x10\n"
+      "cmp x9, #0x8\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
+      "add x12, x12, #0x40\n"
+      "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
       "ldr q10, [x12, #0x20]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr q0, [x28, #0x0]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
@@ -707,6 +704,7 @@ void a64_hybrid_fp32_mla_8x4 (
       "fmla v27.4s, v11.4s, v3.s[3]\n"
       "ldr q3, [x25, #0x0]\n"
       "ldr q11, [x12, #0x30]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "bge 75b\n"
       "76:"  // Height 4: Multiply loop: Single iteration only
       "fmla v24.4s, v8.4s, v0.s[0]\n"
@@ -717,18 +715,18 @@ void a64_hybrid_fp32_mla_8x4 (
       "fmla v27.4s, v8.4s, v3.s[0]\n"
       "add x26, x26, #0x10\n"
       "add x25, x25, #0x10\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
-      "sub x9, x9, #0x4\n"
-      "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "sub x9, x9, #0x4\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x12, x12, #0x40\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
@@ -744,9 +742,9 @@ void a64_hybrid_fp32_mla_8x4 (
       "ldr s18, [x26], #0x4\n"
       "ldr s17, [x25], #0x4\n"
       "ldr q16, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
       "fmla v24.4s, v16.4s, v20.s[0]\n"
       "fmla v25.4s, v16.4s, v19.s[0]\n"
+      "add x12, x12, #0x10\n"
       "fmla v26.4s, v16.4s, v18.s[0]\n"
       "fmla v27.4s, v16.4s, v17.s[0]\n"
       "cbnz x9, 78b\n"
@@ -756,17 +754,17 @@ void a64_hybrid_fp32_mla_8x4 (
       "cmp x10, x20\n"
       "bne 72b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x27, x11, x20, LSL #2\n"
+      "add x26, x27, x20, LSL #2\n"
       "prfm pstl1keep, [x11, #0x0]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x26, x27, x20, LSL #2\n"
       "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "tbz %x[flags], #1, 80f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v24.4s, v24.4s, v17.4s\n"
       "fmin v25.4s, v25.4s, v17.4s\n"
@@ -781,85 +779,85 @@ void a64_hybrid_fp32_mla_8x4 (
       "bge 83f\n"
       "tbz x13, #1, 81f\n"
       "str d24, [x11], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
+      "str d25, [x27], #0x8\n"
+      "str d26, [x26], #0x8\n"
+      "str d27, [x25], #0x8\n"
       "tbz x13, #0, 82f\n"
       "st1 { v24.s }[2], [x11]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
+      "st1 { v25.s }[2], [x27]\n"
+      "st1 { v26.s }[2], [x26]\n"
+      "st1 { v27.s }[2], [x25]\n"
       "b 82f\n"
       "81:"  // Height 4: Partial direct writeback: partial_1_0
       "str s24, [x11, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
+      "str s25, [x27, #0x0]\n"
+      "str s26, [x26, #0x0]\n"
+      "str s27, [x25, #0x0]\n"
       "82:"  // Height 4: Partial direct writeback: Done
       "b 84f\n"
       "83:"  // Height 4: Full writeback
       "str q24, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
+      "str q25, [x27, #0x0]\n"
+      "str q26, [x26, #0x0]\n"
+      "str q27, [x25, #0x0]\n"
       "84:"  // Height 4: Writeback done
       "subs x13, x13, #0x4\n"
       "bgt 65b\n"
       "b 170f\n"
       "85:"  // Height 5
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "86:"  // Height 5: Column loop
       "cbz x14, 87f\n"
       "ldr q24, [x14, #0x0]\n"
-      "add x14, x14, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x14, x14, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "mov v28.16b, v24.16b\n"
       "b 92f\n"
       "87:"  // Height 5: no bias
       "tbz %x[flags], #0, 91f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x13, #0x4\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
+      "cmp x13, #0x4\n"
+      "add x24, x25, x20, LSL #2\n"
       "bge 90f\n"
       "tbz x13, #1, 88f\n"
       "ldr d24, [x11], #0x8\n"
-      "ldr d25, [x28], #0x8\n"
+      "ldr d25, [x27], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d27, [x25], #0x8\n"
+      "ldr d28, [x24], #0x8\n"
       "tbz x13, #0, 89f\n"
       "ld1 { v24.s }[2], [x11]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v28.s }[2], [x25]\n"
+      "ld1 { v25.s }[2], [x27]\n"
+      "ld1 { v26.s }[2], [x26]\n"
+      "ld1 { v27.s }[2], [x25]\n"
+      "ld1 { v28.s }[2], [x24]\n"
       "b 89f\n"
       "88:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s24, [x11, #0x0]\n"
-      "ldr s25, [x28, #0x0]\n"
+      "ldr s25, [x27, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s28, [x25, #0x0]\n"
+      "ldr s26, [x26, #0x0]\n"
+      "ldr s27, [x25, #0x0]\n"
+      "ldr s28, [x24, #0x0]\n"
       "89:"  // Height 5: Partial accumulate: Done
       "sub x11, x11, x20\n"
       "b 92f\n"
       "90:"  // Height 5: full accumulate
       "ldr q24, [x11, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
-      "ldr q28, [x25, #0x0]\n"
+      "ldr q25, [x27, #0x0]\n"
+      "ldr q26, [x26, #0x0]\n"
+      "ldr q27, [x25, #0x0]\n"
+      "ldr q28, [x24, #0x0]\n"
       "b 92f\n"
       "91:"  // Height 5: no accumulate
       "movi v24.16b, #0x0\n"
@@ -871,8 +869,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "mov x10, #0x0\n"
       "93:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 94f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -919,30 +917,30 @@ void a64_hybrid_fp32_mla_8x4 (
       "add x27, x27, #0x10\n"
       "add x26, x26, #0x10\n"
       "fmla v28.4s, v8.4s, v4.s[0]\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x25, x25, #0x10\n"
       "add x24, x24, #0x10\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
+      "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "fmla v26.4s, v9.4s, v2.s[1]\n"
       "cmp x9, #0x8\n"
       "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "ldr q8, [x12, #0x0]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
-      "fmla v25.4s, v9.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "fmla v26.4s, v9.4s, v2.s[1]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
       "ldr q9, [x12, #0x10]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
       "ldr q10, [x12, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr q0, [x28, #0x0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
       "ldr q1, [x27, #0x0]\n"
       "fmla v26.4s, v11.4s, v2.s[3]\n"
@@ -963,21 +961,21 @@ void a64_hybrid_fp32_mla_8x4 (
       "add x26, x26, #0x10\n"
       "add x25, x25, #0x10\n"
       "fmla v28.4s, v8.4s, v4.s[0]\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x24, x24, #0x10\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "sub x9, x9, #0x4\n"
-      "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "add x12, x12, #0x40\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
@@ -996,11 +994,11 @@ void a64_hybrid_fp32_mla_8x4 (
       "ldr s18, [x25], #0x4\n"
       "ldr s17, [x24], #0x4\n"
       "ldr q16, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
       "fmla v24.4s, v16.4s, v21.s[0]\n"
       "fmla v25.4s, v16.4s, v20.s[0]\n"
       "fmla v26.4s, v16.4s, v19.s[0]\n"
       "fmla v27.4s, v16.4s, v18.s[0]\n"
+      "add x12, x12, #0x10\n"
       "fmla v28.4s, v16.4s, v17.s[0]\n"
       "cbnz x9, 99b\n"
       "100:"  // Height 5: Multiply loop: No odd multiplies
@@ -1009,19 +1007,19 @@ void a64_hybrid_fp32_mla_8x4 (
       "cmp x10, x20\n"
       "bne 93b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x27, x11, x20, LSL #2\n"
+      "add x26, x27, x20, LSL #2\n"
       "prfm pstl1keep, [x11, #0x0]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x26, x27, x20, LSL #2\n"
       "prfm pstl1keep, [x26, #0x0]\n"
-      "add x25, x26, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "tbz %x[flags], #1, 101f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v24.4s, v24.4s, v17.4s\n"
       "fmin v25.4s, v25.4s, v17.4s\n"
@@ -1038,47 +1036,47 @@ void a64_hybrid_fp32_mla_8x4 (
       "bge 104f\n"
       "tbz x13, #1, 102f\n"
       "str d24, [x11], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
-      "str d28, [x25], #0x8\n"
+      "str d25, [x27], #0x8\n"
+      "str d26, [x26], #0x8\n"
+      "str d27, [x25], #0x8\n"
+      "str d28, [x24], #0x8\n"
       "tbz x13, #0, 103f\n"
       "st1 { v24.s }[2], [x11]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
-      "st1 { v28.s }[2], [x25]\n"
+      "st1 { v25.s }[2], [x27]\n"
+      "st1 { v26.s }[2], [x26]\n"
+      "st1 { v27.s }[2], [x25]\n"
+      "st1 { v28.s }[2], [x24]\n"
       "b 103f\n"
       "102:"  // Height 5: Partial direct writeback: partial_1_0
       "str s24, [x11, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
-      "str s28, [x25, #0x0]\n"
+      "str s25, [x27, #0x0]\n"
+      "str s26, [x26, #0x0]\n"
+      "str s27, [x25, #0x0]\n"
+      "str s28, [x24, #0x0]\n"
       "103:"  // Height 5: Partial direct writeback: Done
       "b 105f\n"
       "104:"  // Height 5: Full writeback
       "str q24, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
-      "str q28, [x25, #0x0]\n"
+      "str q25, [x27, #0x0]\n"
+      "str q26, [x26, #0x0]\n"
+      "str q27, [x25, #0x0]\n"
+      "str q28, [x24, #0x0]\n"
       "105:"  // Height 5: Writeback done
       "subs x13, x13, #0x4\n"
       "bgt 86b\n"
       "b 170f\n"
       "106:"  // Height 6
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "107:"  // Height 6: Column loop
       "cbz x14, 108f\n"
       "ldr q24, [x14, #0x0]\n"
-      "add x14, x14, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x14, x14, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "mov v28.16b, v24.16b\n"
       "mov v29.16b, v24.16b\n"
@@ -1086,47 +1084,47 @@ void a64_hybrid_fp32_mla_8x4 (
       "108:"  // Height 6: no bias
       "tbz %x[flags], #0, 112f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x13, #0x4\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "cmp x13, #0x4\n"
+      "add x23, x24, x20, LSL #2\n"
       "bge 111f\n"
       "tbz x13, #1, 109f\n"
       "ldr d24, [x11], #0x8\n"
-      "ldr d25, [x28], #0x8\n"
+      "ldr d25, [x27], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "ldr d29, [x24], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d27, [x25], #0x8\n"
+      "ldr d28, [x24], #0x8\n"
+      "ldr d29, [x23], #0x8\n"
       "tbz x13, #0, 110f\n"
       "ld1 { v24.s }[2], [x11]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v28.s }[2], [x25]\n"
-      "ld1 { v29.s }[2], [x24]\n"
+      "ld1 { v25.s }[2], [x27]\n"
+      "ld1 { v26.s }[2], [x26]\n"
+      "ld1 { v27.s }[2], [x25]\n"
+      "ld1 { v28.s }[2], [x24]\n"
+      "ld1 { v29.s }[2], [x23]\n"
       "b 110f\n"
       "109:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s24, [x11, #0x0]\n"
-      "ldr s25, [x28, #0x0]\n"
+      "ldr s25, [x27, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s28, [x25, #0x0]\n"
-      "ldr s29, [x24, #0x0]\n"
+      "ldr s26, [x26, #0x0]\n"
+      "ldr s27, [x25, #0x0]\n"
+      "ldr s28, [x24, #0x0]\n"
+      "ldr s29, [x23, #0x0]\n"
       "110:"  // Height 6: Partial accumulate: Done
       "sub x11, x11, x20\n"
       "b 113f\n"
       "111:"  // Height 6: full accumulate
       "ldr q24, [x11, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
-      "ldr q28, [x25, #0x0]\n"
-      "ldr q29, [x24, #0x0]\n"
+      "ldr q25, [x27, #0x0]\n"
+      "ldr q26, [x26, #0x0]\n"
+      "ldr q27, [x25, #0x0]\n"
+      "ldr q28, [x24, #0x0]\n"
+      "ldr q29, [x23, #0x0]\n"
       "b 113f\n"
       "112:"  // Height 6: no accumulate
       "movi v24.16b, #0x0\n"
@@ -1139,8 +1137,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "mov x10, #0x0\n"
       "114:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 115f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1194,30 +1192,30 @@ void a64_hybrid_fp32_mla_8x4 (
       "fmla v29.4s, v8.4s, v5.s[0]\n"
       "add x25, x25, #0x10\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "cmp x9, #0x8\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
-      "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "ldr q8, [x12, #0x0]\n"
+      "add x23, x23, #0x10\n"
+      "cmp x9, #0x8\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x12, x12, #0x40\n"
+      "ldr q8, [x12, #0x0]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
       "fmla v29.4s, v9.4s, v5.s[1]\n"
       "ldr q9, [x12, #0x10]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
       "fmla v29.4s, v10.4s, v5.s[2]\n"
       "ldr q10, [x12, #0x20]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
       "ldr q0, [x28, #0x0]\n"
       "fmla v25.4s, v11.4s, v1.s[3]\n"
@@ -1245,22 +1243,22 @@ void a64_hybrid_fp32_mla_8x4 (
       "fmla v29.4s, v8.4s, v5.s[0]\n"
       "add x24, x24, #0x10\n"
       "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
-      "sub x9, x9, #0x4\n"
-      "add x12, x12, #0x40\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "sub x9, x9, #0x4\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
       "fmla v29.4s, v9.4s, v5.s[1]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      "add x12, x12, #0x40\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
@@ -1282,9 +1280,9 @@ void a64_hybrid_fp32_mla_8x4 (
       "ldr s18, [x24], #0x4\n"
       "ldr s17, [x23], #0x4\n"
       "ldr q16, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
       "fmla v24.4s, v16.4s, v22.s[0]\n"
       "fmla v25.4s, v16.4s, v21.s[0]\n"
+      "add x12, x12, #0x10\n"
       "fmla v26.4s, v16.4s, v20.s[0]\n"
       "fmla v27.4s, v16.4s, v19.s[0]\n"
       "fmla v28.4s, v16.4s, v18.s[0]\n"
@@ -1296,21 +1294,21 @@ void a64_hybrid_fp32_mla_8x4 (
       "cmp x10, x20\n"
       "bne 114b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x11, #0x0]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x27, x28, x20, LSL #2\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x26, #0x0]\n"
+      "add x23, x24, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbz %x[flags], #1, 122f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v24.4s, v24.4s, v17.4s\n"
       "fmin v25.4s, v25.4s, v17.4s\n"
@@ -1329,51 +1327,51 @@ void a64_hybrid_fp32_mla_8x4 (
       "bge 125f\n"
       "tbz x13, #1, 123f\n"
       "str d24, [x11], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
-      "str d28, [x25], #0x8\n"
-      "str d29, [x24], #0x8\n"
+      "str d25, [x27], #0x8\n"
+      "str d26, [x26], #0x8\n"
+      "str d27, [x25], #0x8\n"
+      "str d28, [x24], #0x8\n"
+      "str d29, [x23], #0x8\n"
       "tbz x13, #0, 124f\n"
       "st1 { v24.s }[2], [x11]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
-      "st1 { v28.s }[2], [x25]\n"
-      "st1 { v29.s }[2], [x24]\n"
+      "st1 { v25.s }[2], [x27]\n"
+      "st1 { v26.s }[2], [x26]\n"
+      "st1 { v27.s }[2], [x25]\n"
+      "st1 { v28.s }[2], [x24]\n"
+      "st1 { v29.s }[2], [x23]\n"
       "b 124f\n"
       "123:"  // Height 6: Partial direct writeback: partial_1_0
       "str s24, [x11, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
-      "str s28, [x25, #0x0]\n"
-      "str s29, [x24, #0x0]\n"
+      "str s25, [x27, #0x0]\n"
+      "str s26, [x26, #0x0]\n"
+      "str s27, [x25, #0x0]\n"
+      "str s28, [x24, #0x0]\n"
+      "str s29, [x23, #0x0]\n"
       "124:"  // Height 6: Partial direct writeback: Done
       "b 126f\n"
       "125:"  // Height 6: Full writeback
       "str q24, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
-      "str q28, [x25, #0x0]\n"
-      "str q29, [x24, #0x0]\n"
+      "str q25, [x27, #0x0]\n"
+      "str q26, [x26, #0x0]\n"
+      "str q27, [x25, #0x0]\n"
+      "str q28, [x24, #0x0]\n"
+      "str q29, [x23, #0x0]\n"
       "126:"  // Height 6: Writeback done
       "subs x13, x13, #0x4\n"
       "bgt 107b\n"
       "b 170f\n"
       "127:"  // Height 7
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "128:"  // Height 7: Column loop
       "cbz x14, 129f\n"
       "ldr q24, [x14, #0x0]\n"
-      "add x14, x14, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x14, x14, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "mov v28.16b, v24.16b\n"
       "mov v29.16b, v24.16b\n"
@@ -1382,52 +1380,52 @@ void a64_hybrid_fp32_mla_8x4 (
       "129:"  // Height 7: no bias
       "tbz %x[flags], #0, 133f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x13, #0x4\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x13, #0x4\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 132f\n"
       "tbz x13, #1, 130f\n"
       "ldr d24, [x11], #0x8\n"
-      "ldr d25, [x28], #0x8\n"
+      "ldr d25, [x27], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d30, [x23], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d27, [x25], #0x8\n"
+      "ldr d28, [x24], #0x8\n"
+      "ldr d29, [x23], #0x8\n"
+      "ldr d30, [x22], #0x8\n"
       "tbz x13, #0, 131f\n"
       "ld1 { v24.s }[2], [x11]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v28.s }[2], [x25]\n"
-      "ld1 { v29.s }[2], [x24]\n"
-      "ld1 { v30.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x27]\n"
+      "ld1 { v26.s }[2], [x26]\n"
+      "ld1 { v27.s }[2], [x25]\n"
+      "ld1 { v28.s }[2], [x24]\n"
+      "ld1 { v29.s }[2], [x23]\n"
+      "ld1 { v30.s }[2], [x22]\n"
       "b 131f\n"
       "130:"  // Height 7: Partial accumulate: partial_1_0
       "ldr s24, [x11, #0x0]\n"
-      "ldr s25, [x28, #0x0]\n"
+      "ldr s25, [x27, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s28, [x25, #0x0]\n"
-      "ldr s29, [x24, #0x0]\n"
-      "ldr s30, [x23, #0x0]\n"
+      "ldr s26, [x26, #0x0]\n"
+      "ldr s27, [x25, #0x0]\n"
+      "ldr s28, [x24, #0x0]\n"
+      "ldr s29, [x23, #0x0]\n"
+      "ldr s30, [x22, #0x0]\n"
       "131:"  // Height 7: Partial accumulate: Done
       "sub x11, x11, x20\n"
       "b 134f\n"
       "132:"  // Height 7: full accumulate
       "ldr q24, [x11, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
-      "ldr q28, [x25, #0x0]\n"
-      "ldr q29, [x24, #0x0]\n"
-      "ldr q30, [x23, #0x0]\n"
+      "ldr q25, [x27, #0x0]\n"
+      "ldr q26, [x26, #0x0]\n"
+      "ldr q27, [x25, #0x0]\n"
+      "ldr q28, [x24, #0x0]\n"
+      "ldr q29, [x23, #0x0]\n"
+      "ldr q30, [x22, #0x0]\n"
       "b 134f\n"
       "133:"  // Height 7: no accumulate
       "movi v24.16b, #0x0\n"
@@ -1441,8 +1439,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "mov x10, #0x0\n"
       "135:"  // Height 7: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 136f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1501,32 +1499,32 @@ void a64_hybrid_fp32_mla_8x4 (
       "add x25, x25, #0x10\n"
       "add x24, x24, #0x10\n"
       "fmla v30.4s, v8.4s, v6.s[0]\n"
+      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "add x23, x23, #0x10\n"
       "add x22, x22, #0x10\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "fmla v26.4s, v9.4s, v2.s[1]\n"
       "cmp x9, #0x8\n"
       "add x12, x12, #0x40\n"
       "ldr q8, [x12, #0x0]\n"
-      "fmla v26.4s, v9.4s, v2.s[1]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "fmla v29.4s, v9.4s, v5.s[1]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v30.4s, v9.4s, v6.s[1]\n"
       "ldr q9, [x12, #0x10]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
       "fmla v29.4s, v10.4s, v5.s[2]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       "fmla v30.4s, v10.4s, v6.s[2]\n"
       "ldr q10, [x12, #0x20]\n"
       "fmla v24.4s, v11.4s, v0.s[3]\n"
@@ -1559,25 +1557,25 @@ void a64_hybrid_fp32_mla_8x4 (
       "add x24, x24, #0x10\n"
       "add x23, x23, #0x10\n"
       "fmla v30.4s, v8.4s, v6.s[0]\n"
-      "add x22, x22, #0x10\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
-      "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "add x22, x22, #0x10\n"
       "sub x9, x9, #0x4\n"
-      "add x12, x12, #0x40\n"
+      "fmla v25.4s, v9.4s, v1.s[1]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
+      "fmla v28.4s, v9.4s, v4.s[1]\n"
       "prfm pldl1keep, [x26, #0x80]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
-      "fmla v28.4s, v9.4s, v4.s[1]\n"
       "fmla v29.4s, v9.4s, v5.s[1]\n"
+      "fmla v30.4s, v9.4s, v6.s[1]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
       "prfm pldl1keep, [x23, #0x80]\n"
-      "fmla v30.4s, v9.4s, v6.s[1]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
+      "add x12, x12, #0x40\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
@@ -1602,11 +1600,11 @@ void a64_hybrid_fp32_mla_8x4 (
       "ldr s18, [x23], #0x4\n"
       "ldr s17, [x22], #0x4\n"
       "ldr q16, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
       "fmla v24.4s, v16.4s, v23.s[0]\n"
       "fmla v25.4s, v16.4s, v22.s[0]\n"
       "fmla v26.4s, v16.4s, v21.s[0]\n"
       "fmla v27.4s, v16.4s, v20.s[0]\n"
+      "add x12, x12, #0x10\n"
       "fmla v28.4s, v16.4s, v19.s[0]\n"
       "fmla v29.4s, v16.4s, v18.s[0]\n"
       "fmla v30.4s, v16.4s, v17.s[0]\n"
@@ -1617,23 +1615,23 @@ void a64_hybrid_fp32_mla_8x4 (
       "cmp x10, x20\n"
       "bne 135b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x11, #0x0]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x27, x28, x20, LSL #2\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x26, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbz %x[flags], #1, 143f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v24.4s, v24.4s, v17.4s\n"
       "fmin v25.4s, v25.4s, v17.4s\n"
@@ -1654,59 +1652,58 @@ void a64_hybrid_fp32_mla_8x4 (
       "bge 146f\n"
       "tbz x13, #1, 144f\n"
       "str d24, [x11], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
-      "str d28, [x25], #0x8\n"
-      "str d29, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
+      "str d25, [x27], #0x8\n"
+      "str d26, [x26], #0x8\n"
+      "str d27, [x25], #0x8\n"
+      "str d28, [x24], #0x8\n"
+      "str d29, [x23], #0x8\n"
+      "str d30, [x22], #0x8\n"
       "tbz x13, #0, 145f\n"
       "st1 { v24.s }[2], [x11]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
-      "st1 { v28.s }[2], [x25]\n"
-      "st1 { v29.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x27]\n"
+      "st1 { v26.s }[2], [x26]\n"
+      "st1 { v27.s }[2], [x25]\n"
+      "st1 { v28.s }[2], [x24]\n"
+      "st1 { v29.s }[2], [x23]\n"
+      "st1 { v30.s }[2], [x22]\n"
       "b 145f\n"
       "144:"  // Height 7: Partial direct writeback: partial_1_0
       "str s24, [x11, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
-      "str s28, [x25, #0x0]\n"
-      "str s29, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
+      "str s25, [x27, #0x0]\n"
+      "str s26, [x26, #0x0]\n"
+      "str s27, [x25, #0x0]\n"
+      "str s28, [x24, #0x0]\n"
+      "str s29, [x23, #0x0]\n"
+      "str s30, [x22, #0x0]\n"
       "145:"  // Height 7: Partial direct writeback: Done
       "b 147f\n"
       "146:"  // Height 7: Full writeback
       "str q24, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
-      "str q28, [x25, #0x0]\n"
-      "str q29, [x24, #0x0]\n"
-      "str q30, [x23, #0x0]\n"
+      "str q25, [x27, #0x0]\n"
+      "str q26, [x26, #0x0]\n"
+      "str q27, [x25, #0x0]\n"
+      "str q28, [x24, #0x0]\n"
+      "str q29, [x23, #0x0]\n"
+      "str q30, [x22, #0x0]\n"
       "147:"  // Height 7: Writeback done
       "subs x13, x13, #0x4\n"
       "bgt 128b\n"
       "b 170f\n"
       "148:"  // Height 8
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x20\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x11\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "149:"  // Height 8: Column loop
       "cbz x14, 150f\n"
       "ldr q24, [x14, #0x0]\n"
-      "add x14, x14, #0x10\n"
       "mov v25.16b, v24.16b\n"
       "mov v26.16b, v24.16b\n"
+      "add x14, x14, #0x10\n"
       "mov v27.16b, v24.16b\n"
       "mov v28.16b, v24.16b\n"
       "mov v29.16b, v24.16b\n"
@@ -1716,57 +1713,57 @@ void a64_hybrid_fp32_mla_8x4 (
       "150:"  // Height 8: no bias
       "tbz %x[flags], #0, 154f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x13, #0x4\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x13, #0x4\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 153f\n"
       "tbz x13, #1, 151f\n"
       "ldr d24, [x11], #0x8\n"
-      "ldr d25, [x28], #0x8\n"
+      "ldr d25, [x27], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d26, [x27], #0x8\n"
-      "ldr d27, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d30, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d27, [x25], #0x8\n"
+      "ldr d28, [x24], #0x8\n"
+      "ldr d29, [x23], #0x8\n"
+      "ldr d30, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x13, #0, 152f\n"
       "ld1 { v24.s }[2], [x11]\n"
-      "ld1 { v25.s }[2], [x28]\n"
-      "ld1 { v26.s }[2], [x27]\n"
-      "ld1 { v27.s }[2], [x26]\n"
-      "ld1 { v28.s }[2], [x25]\n"
-      "ld1 { v29.s }[2], [x24]\n"
-      "ld1 { v30.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v25.s }[2], [x27]\n"
+      "ld1 { v26.s }[2], [x26]\n"
+      "ld1 { v27.s }[2], [x25]\n"
+      "ld1 { v28.s }[2], [x24]\n"
+      "ld1 { v29.s }[2], [x23]\n"
+      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 152f\n"
       "151:"  // Height 8: Partial accumulate: partial_1_0
       "ldr s24, [x11, #0x0]\n"
-      "ldr s25, [x28, #0x0]\n"
+      "ldr s25, [x27, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s26, [x27, #0x0]\n"
-      "ldr s27, [x26, #0x0]\n"
-      "ldr s28, [x25, #0x0]\n"
-      "ldr s29, [x24, #0x0]\n"
-      "ldr s30, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s26, [x26, #0x0]\n"
+      "ldr s27, [x25, #0x0]\n"
+      "ldr s28, [x24, #0x0]\n"
+      "ldr s29, [x23, #0x0]\n"
+      "ldr s30, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "152:"  // Height 8: Partial accumulate: Done
       "sub x11, x11, x20\n"
       "b 155f\n"
       "153:"  // Height 8: full accumulate
       "ldr q24, [x11, #0x0]\n"
-      "ldr q25, [x28, #0x0]\n"
-      "ldr q26, [x27, #0x0]\n"
-      "ldr q27, [x26, #0x0]\n"
-      "ldr q28, [x25, #0x0]\n"
-      "ldr q29, [x24, #0x0]\n"
-      "ldr q30, [x23, #0x0]\n"
-      "ldr q31, [x22, #0x0]\n"
+      "ldr q25, [x27, #0x0]\n"
+      "ldr q26, [x26, #0x0]\n"
+      "ldr q27, [x25, #0x0]\n"
+      "ldr q28, [x24, #0x0]\n"
+      "ldr q29, [x23, #0x0]\n"
+      "ldr q30, [x22, #0x0]\n"
+      "ldr q31, [x21, #0x0]\n"
       "b 155f\n"
       "154:"  // Height 8: no accumulate
       "movi v24.16b, #0x0\n"
@@ -1781,8 +1778,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "mov x10, #0x0\n"
       "156:"  // Height 8: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 157f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1855,10 +1852,10 @@ void a64_hybrid_fp32_mla_8x4 (
       "fmla v26.4s, v9.4s, v2.s[1]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
       "add x12, x12, #0x40\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
       "ldr q8, [x12, #0x0]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
       "fmla v29.4s, v9.4s, v5.s[1]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v30.4s, v9.4s, v6.s[1]\n"
       "fmla v31.4s, v9.4s, v7.s[1]\n"
@@ -1915,24 +1912,24 @@ void a64_hybrid_fp32_mla_8x4 (
       "add x21, x21, #0x10\n"
       "fmla v24.4s, v9.4s, v0.s[1]\n"
       "fmla v25.4s, v9.4s, v1.s[1]\n"
+      "sub x9, x9, #0x4\n"
       "prfm pldl1keep, [x28, #0x80]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
       "fmla v26.4s, v9.4s, v2.s[1]\n"
       "fmla v27.4s, v9.4s, v3.s[1]\n"
-      "sub x9, x9, #0x4\n"
-      "add x12, x12, #0x40\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "fmla v28.4s, v9.4s, v4.s[1]\n"
       "fmla v29.4s, v9.4s, v5.s[1]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       "fmla v30.4s, v9.4s, v6.s[1]\n"
       "fmla v31.4s, v9.4s, v7.s[1]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "prfm pldl1keep, [x23, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       "fmla v24.4s, v10.4s, v0.s[2]\n"
       "fmla v25.4s, v10.4s, v1.s[2]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       "prfm pldl1keep, [x21, #0x80]\n"
+      "add x12, x12, #0x40\n"
       "fmla v26.4s, v10.4s, v2.s[2]\n"
       "fmla v27.4s, v10.4s, v3.s[2]\n"
       "fmla v28.4s, v10.4s, v4.s[2]\n"
@@ -1960,9 +1957,9 @@ void a64_hybrid_fp32_mla_8x4 (
       "ldr s18, [x22], #0x4\n"
       "ldr s17, [x21], #0x4\n"
       "ldr q16, [x12, #0x0]\n"
-      "add x12, x12, #0x10\n"
       "fmla v24.4s, v16.4s, v0.s[0]\n"
       "fmla v25.4s, v16.4s, v23.s[0]\n"
+      "add x12, x12, #0x10\n"
       "fmla v26.4s, v16.4s, v22.s[0]\n"
       "fmla v27.4s, v16.4s, v21.s[0]\n"
       "fmla v28.4s, v16.4s, v20.s[0]\n"
@@ -1976,25 +1973,25 @@ void a64_hybrid_fp32_mla_8x4 (
       "cmp x10, x20\n"
       "bne 156b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x11, #0x0]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "prfm pstl1keep, [x28, #0x0]\n"
-      "add x27, x28, x20, LSL #2\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x26, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
-      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbz %x[flags], #1, 164f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v17.4s }, [x21]\n"
       "ld1r { v16.4s }, [x20]\n"
       "fmin v24.4s, v24.4s, v17.4s\n"
       "fmin v25.4s, v25.4s, v17.4s\n"
@@ -2017,44 +2014,44 @@ void a64_hybrid_fp32_mla_8x4 (
       "bge 167f\n"
       "tbz x13, #1, 165f\n"
       "str d24, [x11], #0x8\n"
-      "str d25, [x28], #0x8\n"
-      "str d26, [x27], #0x8\n"
-      "str d27, [x26], #0x8\n"
-      "str d28, [x25], #0x8\n"
-      "str d29, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
-      "str d31, [x22], #0x8\n"
+      "str d25, [x27], #0x8\n"
+      "str d26, [x26], #0x8\n"
+      "str d27, [x25], #0x8\n"
+      "str d28, [x24], #0x8\n"
+      "str d29, [x23], #0x8\n"
+      "str d30, [x22], #0x8\n"
+      "str d31, [x21], #0x8\n"
       "tbz x13, #0, 166f\n"
       "st1 { v24.s }[2], [x11]\n"
-      "st1 { v25.s }[2], [x28]\n"
-      "st1 { v26.s }[2], [x27]\n"
-      "st1 { v27.s }[2], [x26]\n"
-      "st1 { v28.s }[2], [x25]\n"
-      "st1 { v29.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
-      "st1 { v31.s }[2], [x22]\n"
+      "st1 { v25.s }[2], [x27]\n"
+      "st1 { v26.s }[2], [x26]\n"
+      "st1 { v27.s }[2], [x25]\n"
+      "st1 { v28.s }[2], [x24]\n"
+      "st1 { v29.s }[2], [x23]\n"
+      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v31.s }[2], [x21]\n"
       "b 166f\n"
       "165:"  // Height 8: Partial direct writeback: partial_1_0
       "str s24, [x11, #0x0]\n"
-      "str s25, [x28, #0x0]\n"
-      "str s26, [x27, #0x0]\n"
-      "str s27, [x26, #0x0]\n"
-      "str s28, [x25, #0x0]\n"
-      "str s29, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
-      "str s31, [x22, #0x0]\n"
+      "str s25, [x27, #0x0]\n"
+      "str s26, [x26, #0x0]\n"
+      "str s27, [x25, #0x0]\n"
+      "str s28, [x24, #0x0]\n"
+      "str s29, [x23, #0x0]\n"
+      "str s30, [x22, #0x0]\n"
+      "str s31, [x21, #0x0]\n"
       "166:"  // Height 8: Partial direct writeback: Done
       "b 168f\n"
       "167:"  // Height 8: Full writeback
       "str q24, [x11, #0x0]\n"
       "add x11, x11, #0x10\n"
-      "str q25, [x28, #0x0]\n"
-      "str q26, [x27, #0x0]\n"
-      "str q27, [x26, #0x0]\n"
-      "str q28, [x25, #0x0]\n"
-      "str q29, [x24, #0x0]\n"
-      "str q30, [x23, #0x0]\n"
-      "str q31, [x22, #0x0]\n"
+      "str q25, [x27, #0x0]\n"
+      "str q26, [x26, #0x0]\n"
+      "str q27, [x25, #0x0]\n"
+      "str q28, [x24, #0x0]\n"
+      "str q29, [x23, #0x0]\n"
+      "str q30, [x22, #0x0]\n"
+      "str q31, [x21, #0x0]\n"
       "168:"  // Height 8: Writeback done
       "subs x13, x13, #0x4\n"
       "bgt 149b\n"
@@ -2070,8 +2067,8 @@ void a64_hybrid_fp32_mla_8x4 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "170:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24.hpp
index 453ef4888f..e04820cc10 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 24, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 4, 24, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp
index 39a34898d7..d36059d6a1 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp
@@ -48,19 +48,18 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -81,7 +80,6 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -101,27 +99,27 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "cmp %x[M], #0x2\n"
       "bgt 87f\n"
       "beq 44f\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x10, 3f\n"
       "ldr q8, [x10, #0x0]\n"
       "ldr q9, [x10, #0x10]\n"
-      "ldr q10, [x10, #0x20]\n"
-      "ldr q11, [x10, #0x30]\n"
-      "ldr q12, [x10, #0x40]\n"
-      "ldr q13, [x10, #0x50]\n"
-      "add x10, x10, #0x60\n"
       "zip2 v14.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x10, #0x20]\n"
+      "ldr q11, [x10, #0x30]\n"
       "zip2 v15.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
+      "ldr q12, [x10, #0x40]\n"
+      "ldr q13, [x10, #0x50]\n"
       "zip2 v16.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
       "zip2 v17.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
+      "add x10, x10, #0x60\n"
       "zip2 v18.2d, v12.2d, v12.2d\n"
       "zip1 v12.2d, v12.2d, v12.2d\n"
       "zip2 v19.2d, v13.2d, v13.2d\n"
@@ -255,8 +253,8 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "mov x26, #0x0\n"
       "20:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 21f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -279,9 +277,6 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "blt 24f\n"
       "23:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x25, x25, #0x4\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "cmp x25, #0x8\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
       "ldr q24, [x28, #0x40]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
@@ -298,42 +293,45 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q22, [x28, #0xa0]\n"
       ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
       "ldr q21, [x28, #0xb0]\n"
-      "add x28, x28, #0xc0\n"
+      "sub x25, x25, #0x4\n"
+      "cmp x25, #0x8\n"
       ".inst 0x6e58ec0c  // bfmmla v12.4s, v0.8h, v24.8h\n"
+      "add x28, x28, #0xc0\n"
       "ldr q4, [x28, #0x0]\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
       "ldr q5, [x28, #0x10]\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       "ldr q6, [x28, #0x20]\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
-      "ld1 { v0.4s }, [x24], #0x10\n"
       "ldr q7, [x28, #0x30]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "ld1 { v0.4s }, [x24], #0x10\n"
       "bge 23b\n"
       "24:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x25, x25, #0x4\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
       "ldr q23, [x28, #0x40]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
-      "ldr q22, [x28, #0x50]\n"
+      "ldr q25, [x28, #0x50]\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       "ldr q21, [x28, #0x60]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       "ldr q24, [x28, #0x70]\n"
       ".inst 0x6e57ec0a  // bfmmla v10.4s, v0.8h, v23.8h\n"
       "ldr q23, [x28, #0x80]\n"
-      ".inst 0x6e56ec10  // bfmmla v16.4s, v0.8h, v22.8h\n"
+      ".inst 0x6e59ec10  // bfmmla v16.4s, v0.8h, v25.8h\n"
       "ldr q22, [x28, #0x90]\n"
       ".inst 0x6e55ec0b  // bfmmla v11.4s, v0.8h, v21.8h\n"
       "ldr q21, [x28, #0xa0]\n"
       ".inst 0x6e58ec11  // bfmmla v17.4s, v0.8h, v24.8h\n"
-      "ldr q3, [x28, #0xb0]\n"
-      "add x28, x28, #0xc0\n"
+      "ldr q5, [x28, #0xb0]\n"
+      "sub x25, x25, #0x4\n"
       ".inst 0x6e57ec0c  // bfmmla v12.4s, v0.8h, v23.8h\n"
       ".inst 0x6e56ec12  // bfmmla v18.4s, v0.8h, v22.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "add x28, x28, #0xc0\n"
       ".inst 0x6e55ec0d  // bfmmla v13.4s, v0.8h, v21.8h\n"
-      ".inst 0x6e43ec13  // bfmmla v19.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e45ec13  // bfmmla v19.4s, v0.8h, v5.8h\n"
       "25:"  // Height 1: Multiply loop: Main loop skip
       "cbz x25, 28f\n"
       "cbz x25, 28f\n"
@@ -345,32 +343,32 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "26:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr s0, [x24, #0x0]\n"
       "27:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q23, [x28, #0x0]\n"
-      "ldr q29, [x28, #0x10]\n"
+      "ldr q21, [x28, #0x0]\n"
+      "ldr q1, [x28, #0x10]\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "ldr q22, [x28, #0x20]\n"
-      "ldr q21, [x28, #0x30]\n"
-      ".inst 0x6e57ec08  // bfmmla v8.4s, v0.8h, v23.8h\n"
-      "ldr q24, [x28, #0x40]\n"
-      ".inst 0x6e5dec0e  // bfmmla v14.4s, v0.8h, v29.8h\n"
+      ".inst 0x6e55ec08  // bfmmla v8.4s, v0.8h, v21.8h\n"
+      "ldr q21, [x28, #0x20]\n"
+      "ldr q22, [x28, #0x30]\n"
+      ".inst 0x6e41ec0e  // bfmmla v14.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e55ec09  // bfmmla v9.4s, v0.8h, v21.8h\n"
+      "ldr q21, [x28, #0x40]\n"
       "ldr q23, [x28, #0x50]\n"
-      ".inst 0x6e56ec09  // bfmmla v9.4s, v0.8h, v22.8h\n"
-      "ldr q22, [x28, #0x60]\n"
-      ".inst 0x6e55ec0f  // bfmmla v15.4s, v0.8h, v21.8h\n"
-      "ldr q21, [x28, #0x70]\n"
-      ".inst 0x6e58ec0a  // bfmmla v10.4s, v0.8h, v24.8h\n"
-      "ldr q24, [x28, #0x80]\n"
+      ".inst 0x6e56ec0f  // bfmmla v15.4s, v0.8h, v22.8h\n"
+      ".inst 0x6e55ec0a  // bfmmla v10.4s, v0.8h, v21.8h\n"
+      "ldr q21, [x28, #0x60]\n"
+      "ldr q22, [x28, #0x70]\n"
       ".inst 0x6e57ec10  // bfmmla v16.4s, v0.8h, v23.8h\n"
+      ".inst 0x6e55ec0b  // bfmmla v11.4s, v0.8h, v21.8h\n"
+      "ldr q21, [x28, #0x80]\n"
       "ldr q23, [x28, #0x90]\n"
-      ".inst 0x6e56ec0b  // bfmmla v11.4s, v0.8h, v22.8h\n"
+      ".inst 0x6e56ec11  // bfmmla v17.4s, v0.8h, v22.8h\n"
+      ".inst 0x6e55ec0c  // bfmmla v12.4s, v0.8h, v21.8h\n"
       "ldr q22, [x28, #0xa0]\n"
-      ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
       "ldr q21, [x28, #0xb0]\n"
-      "add x28, x28, #0xc0\n"
-      ".inst 0x6e58ec0c  // bfmmla v12.4s, v0.8h, v24.8h\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
+      "add x28, x28, #0xc0\n"
       "28:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -384,9 +382,9 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "uzp1 v12.2d, v12.2d, v18.2d\n"
       "uzp1 v13.2d, v13.2d, v19.2d\n"
       "tbz %x[flags], #1, 29f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v22.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v22.4s }, [x21]\n"
       "ld1r { v21.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v22.4s\n"
       "fmin v9.4s, v9.4s, v22.4s\n"
@@ -489,27 +487,27 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "bgt 2b\n"
       "b 174f\n"
       "44:"  // Height 2
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "45:"  // Height 2: Column loop
       "cbz x10, 46f\n"
       "ldr q8, [x10, #0x0]\n"
       "ldr q9, [x10, #0x10]\n"
-      "ldr q10, [x10, #0x20]\n"
-      "ldr q11, [x10, #0x30]\n"
-      "ldr q12, [x10, #0x40]\n"
-      "ldr q13, [x10, #0x50]\n"
-      "add x10, x10, #0x60\n"
       "zip2 v14.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x10, #0x20]\n"
+      "ldr q11, [x10, #0x30]\n"
       "zip2 v15.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
+      "ldr q12, [x10, #0x40]\n"
+      "ldr q13, [x10, #0x50]\n"
       "zip2 v16.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
       "zip2 v17.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
+      "add x10, x10, #0x60\n"
       "zip2 v18.2d, v12.2d, v12.2d\n"
       "zip1 v12.2d, v12.2d, v12.2d\n"
       "zip2 v19.2d, v13.2d, v13.2d\n"
@@ -519,117 +517,117 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "tbz %x[flags], #0, 61f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x9, #0x18\n"
-      "add x24, x27, x20, LSL #2\n"
+      "add x23, x27, x20, LSL #2\n"
       "bge 59f\n"
       "tbz x9, #4, 50f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v12.4s }, [x27], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
       "tbz x9, #2, 48f\n"
       "ld1 { v13.4s }, [x27], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
       "tbz x9, #1, 47f\n"
       "ldr d20, [x27], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
       "tbz x9, #0, 58f\n"
       "ld1 { v20.s }[2], [x27]\n"
-      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v19.s }[2], [x23]\n"
       "b 58f\n"
       "47:"  // Height 2: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x9, #0, 58f\n"
       "ldr s20, [x27, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
       "b 58f\n"
       "48:"  // Height 2: Partial accumulate: partial_2_16
       "tbz x9, #1, 49f\n"
       "ldr d13, [x27], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
       "tbz x9, #0, 58f\n"
       "ld1 { v13.s }[2], [x27]\n"
-      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v18.s }[2], [x23]\n"
       "b 58f\n"
       "49:"  // Height 2: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x9, #0, 58f\n"
       "ldr s13, [x27, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
       "b 58f\n"
       "50:"  // Height 2: Partial accumulate: partial_8_0
       "tbz x9, #3, 54f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
       "tbz x9, #2, 52f\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
       "tbz x9, #1, 51f\n"
       "ldr d12, [x27], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x9, #0, 58f\n"
       "ld1 { v12.s }[2], [x27]\n"
-      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v17.s }[2], [x23]\n"
       "b 58f\n"
       "51:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x9, #0, 58f\n"
       "ldr s12, [x27, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
       "b 58f\n"
       "52:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x9, #1, 53f\n"
       "ldr d11, [x27], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x9, #0, 58f\n"
       "ld1 { v11.s }[2], [x27]\n"
-      "ld1 { v16.s }[2], [x24]\n"
+      "ld1 { v16.s }[2], [x23]\n"
       "b 58f\n"
       "53:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x9, #0, 58f\n"
       "ldr s11, [x27, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
       "b 58f\n"
       "54:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x9, #2, 56f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
       "tbz x9, #1, 55f\n"
       "ldr d10, [x27], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x9, #0, 58f\n"
       "ld1 { v10.s }[2], [x27]\n"
-      "ld1 { v15.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x23]\n"
       "b 58f\n"
       "55:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x9, #0, 58f\n"
       "ldr s10, [x27, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
       "b 58f\n"
       "56:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x9, #1, 57f\n"
       "ldr d9, [x27], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x9, #0, 58f\n"
       "ld1 { v9.s }[2], [x27]\n"
-      "ld1 { v14.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x23]\n"
       "b 58f\n"
       "57:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s9, [x27, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
       "58:"  // Height 2: Partial accumulate: Done
       "sub x27, x27, x20\n"
@@ -641,12 +639,12 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q12, [x27, #0x30]\n"
       "ldr q13, [x27, #0x40]\n"
       "ldr q20, [x27, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
       "60:"  // Height 2: MMLA fixup
       "zip1 v8.2d, v9.2d, v14.2d\n"
       "zip2 v14.2d, v9.2d, v14.2d\n"
@@ -678,8 +676,8 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "mov x26, #0x0\n"
       "63:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 64f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -706,44 +704,41 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "blt 67f\n"
       "66:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x25, x25, #0x4\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "cmp x25, #0x8\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      "ld1 { v1.4s }, [x23], #0x10\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
-      "ldr q29, [x28, #0x40]\n"
+      "ldr q3, [x28, #0x40]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
       "ldr q23, [x28, #0x50]\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       "ldr q22, [x28, #0x60]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       "ldr q21, [x28, #0x70]\n"
-      ".inst 0x6e5dec0a  // bfmmla v10.4s, v0.8h, v29.8h\n"
-      "ldr q30, [x28, #0x80]\n"
+      ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
+      "ldr q1, [x28, #0x80]\n"
       ".inst 0x6e57ec10  // bfmmla v16.4s, v0.8h, v23.8h\n"
       "ldr q23, [x28, #0x90]\n"
       ".inst 0x6e56ec0b  // bfmmla v11.4s, v0.8h, v22.8h\n"
       "ldr q22, [x28, #0xa0]\n"
       ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
       "ldr q21, [x28, #0xb0]\n"
+      "sub x25, x25, #0x4\n"
+      "cmp x25, #0x8\n"
       "add x28, x28, #0xc0\n"
-      ".inst 0x6e5eec0c  // bfmmla v12.4s, v0.8h, v30.8h\n"
+      ".inst 0x6e41ec0c  // bfmmla v12.4s, v0.8h, v1.8h\n"
       "ldr q4, [x28, #0x0]\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
       "ldr q5, [x28, #0x10]\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       "ldr q6, [x28, #0x20]\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
-      "ld1 { v0.4s }, [x24], #0x10\n"
       "ldr q7, [x28, #0x30]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "ld1 { v0.4s }, [x24], #0x10\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      "ld1 { v1.4s }, [x23], #0x10\n"
       "bge 66b\n"
       "67:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x25, x25, #0x4\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
       "ldr q24, [x28, #0x40]\n"
@@ -761,11 +756,14 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q22, [x28, #0xa0]\n"
       ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
       "ldr q21, [x28, #0xb0]\n"
-      "add x28, x28, #0xc0\n"
+      "sub x25, x25, #0x4\n"
       ".inst 0x6e58ec0c  // bfmmla v12.4s, v0.8h, v24.8h\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
+      "add x28, x28, #0xc0\n"
       "68:"  // Height 2: Multiply loop: Main loop skip
       "cbz x25, 71f\n"
       "cbz x25, 71f\n"
@@ -783,55 +781,55 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q24, [x28, #0x0]\n"
       "ldr q23, [x28, #0x10]\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       "ldr q22, [x28, #0x20]\n"
       "ldr q21, [x28, #0x30]\n"
-      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x6e58ec08  // bfmmla v8.4s, v0.8h, v24.8h\n"
-      "ldr q24, [x28, #0x40]\n"
       ".inst 0x6e57ec0e  // bfmmla v14.4s, v0.8h, v23.8h\n"
+      "ldr q24, [x28, #0x40]\n"
       "ldr q23, [x28, #0x50]\n"
       ".inst 0x6e56ec09  // bfmmla v9.4s, v0.8h, v22.8h\n"
-      "ldr q22, [x28, #0x60]\n"
       ".inst 0x6e55ec0f  // bfmmla v15.4s, v0.8h, v21.8h\n"
+      "ldr q22, [x28, #0x60]\n"
       "ldr q21, [x28, #0x70]\n"
       ".inst 0x6e58ec0a  // bfmmla v10.4s, v0.8h, v24.8h\n"
-      "ldr q24, [x28, #0x80]\n"
       ".inst 0x6e57ec10  // bfmmla v16.4s, v0.8h, v23.8h\n"
+      "ldr q24, [x28, #0x80]\n"
       "ldr q23, [x28, #0x90]\n"
       ".inst 0x6e56ec0b  // bfmmla v11.4s, v0.8h, v22.8h\n"
-      "ldr q22, [x28, #0xa0]\n"
       ".inst 0x6e55ec11  // bfmmla v17.4s, v0.8h, v21.8h\n"
+      "ldr q22, [x28, #0xa0]\n"
       "ldr q21, [x28, #0xb0]\n"
-      "add x28, x28, #0xc0\n"
       ".inst 0x6e58ec0c  // bfmmla v12.4s, v0.8h, v24.8h\n"
       ".inst 0x6e57ec12  // bfmmla v18.4s, v0.8h, v23.8h\n"
       ".inst 0x6e56ec0d  // bfmmla v13.4s, v0.8h, v22.8h\n"
       ".inst 0x6e55ec13  // bfmmla v19.4s, v0.8h, v21.8h\n"
+      "add x28, x28, #0xc0\n"
       "71:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
       "cmp x26, x20\n"
       "bne 63b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
       "uzp1 v4.2d, v8.2d, v14.2d\n"
       "uzp2 v8.2d, v8.2d, v14.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v14.2d, v9.2d, v15.2d\n"
       "uzp2 v9.2d, v9.2d, v15.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v15.2d, v10.2d, v16.2d\n"
       "uzp2 v10.2d, v10.2d, v16.2d\n"
-      "add x24, x27, x20, LSL #2\n"
       "uzp1 v16.2d, v11.2d, v17.2d\n"
       "uzp2 v11.2d, v11.2d, v17.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v17.2d, v12.2d, v18.2d\n"
       "uzp2 v12.2d, v12.2d, v18.2d\n"
       "uzp1 v18.2d, v13.2d, v19.2d\n"
       "uzp2 v13.2d, v13.2d, v19.2d\n"
       "tbz %x[flags], #1, 72f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v22.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v22.4s }, [x21]\n"
       "ld1r { v21.4s }, [x20]\n"
       "fmin v4.4s, v4.4s, v22.4s\n"
       "fmin v14.4s, v14.4s, v22.4s\n"
@@ -865,99 +863,99 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "st1 { v14.4s }, [x27], #0x10\n"
       "st1 { v15.4s }, [x27], #0x10\n"
       "st1 { v16.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v11.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v11.4s }, [x23], #0x10\n"
       "tbz x9, #2, 74f\n"
       "st1 { v17.4s }, [x27], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
+      "st1 { v12.4s }, [x23], #0x10\n"
       "tbz x9, #1, 73f\n"
       "str d18, [x27], #0x8\n"
-      "str d13, [x24], #0x8\n"
+      "str d13, [x23], #0x8\n"
       "tbz x9, #0, 84f\n"
       "st1 { v18.s }[2], [x27]\n"
-      "st1 { v13.s }[2], [x24]\n"
+      "st1 { v13.s }[2], [x23]\n"
       "b 84f\n"
       "73:"  // Height 2: Partial direct writeback: partial_1_20
       "tbz x9, #0, 84f\n"
       "str s18, [x27, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
+      "str s13, [x23, #0x0]\n"
       "b 84f\n"
       "74:"  // Height 2: Partial direct writeback: partial_2_16
       "tbz x9, #1, 75f\n"
       "str d17, [x27], #0x8\n"
-      "str d12, [x24], #0x8\n"
+      "str d12, [x23], #0x8\n"
       "tbz x9, #0, 84f\n"
       "st1 { v17.s }[2], [x27]\n"
-      "st1 { v12.s }[2], [x24]\n"
+      "st1 { v12.s }[2], [x23]\n"
       "b 84f\n"
       "75:"  // Height 2: Partial direct writeback: partial_1_16
       "tbz x9, #0, 84f\n"
       "str s17, [x27, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
+      "str s12, [x23, #0x0]\n"
       "b 84f\n"
       "76:"  // Height 2: Partial direct writeback: partial_8_0
       "tbz x9, #3, 80f\n"
       "st1 { v4.4s }, [x27], #0x10\n"
       "st1 { v14.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
       "tbz x9, #2, 78f\n"
       "st1 { v15.4s }, [x27], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
       "tbz x9, #1, 77f\n"
       "str d16, [x27], #0x8\n"
-      "str d11, [x24], #0x8\n"
+      "str d11, [x23], #0x8\n"
       "tbz x9, #0, 84f\n"
       "st1 { v16.s }[2], [x27]\n"
-      "st1 { v11.s }[2], [x24]\n"
+      "st1 { v11.s }[2], [x23]\n"
       "b 84f\n"
       "77:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x9, #0, 84f\n"
       "str s16, [x27, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
+      "str s11, [x23, #0x0]\n"
       "b 84f\n"
       "78:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x9, #1, 79f\n"
       "str d15, [x27], #0x8\n"
-      "str d10, [x24], #0x8\n"
+      "str d10, [x23], #0x8\n"
       "tbz x9, #0, 84f\n"
       "st1 { v15.s }[2], [x27]\n"
-      "st1 { v10.s }[2], [x24]\n"
+      "st1 { v10.s }[2], [x23]\n"
       "b 84f\n"
       "79:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x9, #0, 84f\n"
       "str s15, [x27, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
+      "str s10, [x23, #0x0]\n"
       "b 84f\n"
       "80:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x9, #2, 82f\n"
       "st1 { v4.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
       "tbz x9, #1, 81f\n"
       "str d14, [x27], #0x8\n"
-      "str d9, [x24], #0x8\n"
+      "str d9, [x23], #0x8\n"
       "tbz x9, #0, 84f\n"
       "st1 { v14.s }[2], [x27]\n"
-      "st1 { v9.s }[2], [x24]\n"
+      "st1 { v9.s }[2], [x23]\n"
       "b 84f\n"
       "81:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x9, #0, 84f\n"
       "str s14, [x27, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
+      "str s9, [x23, #0x0]\n"
       "b 84f\n"
       "82:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x9, #1, 83f\n"
       "str d4, [x27], #0x8\n"
-      "str d8, [x24], #0x8\n"
+      "str d8, [x23], #0x8\n"
       "tbz x9, #0, 84f\n"
       "st1 { v4.s }[2], [x27]\n"
-      "st1 { v8.s }[2], [x24]\n"
+      "st1 { v8.s }[2], [x23]\n"
       "b 84f\n"
       "83:"  // Height 2: Partial direct writeback: partial_1_0
       "str s4, [x27, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
+      "str s8, [x23, #0x0]\n"
       "84:"  // Height 2: Partial direct writeback: Done
       "b 86f\n"
       "85:"  // Height 2: Full writeback
@@ -968,38 +966,38 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "str q17, [x27, #0x40]\n"
       "str q18, [x27, #0x50]\n"
       "add x27, x27, #0x60\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q12, [x24, #0x40]\n"
-      "str q13, [x24, #0x50]\n"
+      "str q8, [x23, #0x0]\n"
+      "str q9, [x23, #0x10]\n"
+      "str q10, [x23, #0x20]\n"
+      "str q11, [x23, #0x30]\n"
+      "str q12, [x23, #0x40]\n"
+      "str q13, [x23, #0x50]\n"
       "86:"  // Height 2: Writeback done
       "subs x9, x9, #0x18\n"
       "bgt 45b\n"
       "b 174f\n"
       "87:"  // Height 3
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "88:"  // Height 3: Column loop
       "cbz x10, 89f\n"
       "ldr q8, [x10, #0x0]\n"
       "ldr q9, [x10, #0x10]\n"
-      "ldr q10, [x10, #0x20]\n"
-      "ldr q11, [x10, #0x30]\n"
-      "ldr q12, [x10, #0x40]\n"
-      "ldr q13, [x10, #0x50]\n"
-      "add x10, x10, #0x60\n"
       "zip2 v14.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x10, #0x20]\n"
+      "ldr q11, [x10, #0x30]\n"
       "zip2 v15.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
+      "ldr q12, [x10, #0x40]\n"
+      "ldr q13, [x10, #0x50]\n"
       "zip2 v16.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
       "zip2 v17.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
+      "add x10, x10, #0x60\n"
       "zip2 v18.2d, v12.2d, v12.2d\n"
       "zip1 v12.2d, v12.2d, v12.2d\n"
       "zip2 v19.2d, v13.2d, v13.2d\n"
@@ -1020,147 +1018,147 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "89:"  // Height 3: no bias
       "tbz %x[flags], #0, 104f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
       "cmp x9, #0x18\n"
-      "add x24, x27, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 102f\n"
       "tbz x9, #4, 93f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
       "ld1 { v12.4s }, [x27], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
       "tbz x9, #2, 91f\n"
       "ld1 { v13.4s }, [x27], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x9, #1, 90f\n"
       "ldr d20, [x27], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d4, [x23], #0x8\n"
+      "ldr d4, [x22], #0x8\n"
       "tbz x9, #0, 101f\n"
       "ld1 { v20.s }[2], [x27]\n"
-      "ld1 { v19.s }[2], [x24]\n"
-      "ld1 { v4.s }[2], [x23]\n"
+      "ld1 { v19.s }[2], [x23]\n"
+      "ld1 { v4.s }[2], [x22]\n"
       "b 101f\n"
       "90:"  // Height 3: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x9, #0, 101f\n"
       "ldr s20, [x27, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
-      "ldr s4, [x23, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
+      "ldr s4, [x22, #0x0]\n"
       "b 101f\n"
       "91:"  // Height 3: Partial accumulate: partial_2_16
       "tbz x9, #1, 92f\n"
       "ldr d13, [x27], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x9, #0, 101f\n"
       "ld1 { v13.s }[2], [x27]\n"
-      "ld1 { v18.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v18.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 101f\n"
       "92:"  // Height 3: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x9, #0, 101f\n"
       "ldr s13, [x27, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "b 101f\n"
       "93:"  // Height 3: Partial accumulate: partial_8_0
       "tbz x9, #3, 97f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
       "tbz x9, #2, 95f\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
       "tbz x9, #1, 94f\n"
       "ldr d12, [x27], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x23], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
       "tbz x9, #0, 101f\n"
       "ld1 { v12.s }[2], [x27]\n"
-      "ld1 { v17.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
+      "ld1 { v17.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
       "b 101f\n"
       "94:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x9, #0, 101f\n"
       "ldr s12, [x27, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
       "b 101f\n"
       "95:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x9, #1, 96f\n"
       "ldr d11, [x27], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d23, [x23], #0x8\n"
+      "ldr d23, [x22], #0x8\n"
       "tbz x9, #0, 101f\n"
       "ld1 { v11.s }[2], [x27]\n"
-      "ld1 { v16.s }[2], [x24]\n"
-      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v16.s }[2], [x23]\n"
+      "ld1 { v23.s }[2], [x22]\n"
       "b 101f\n"
       "96:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x9, #0, 101f\n"
       "ldr s11, [x27, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
-      "ldr s23, [x23, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
+      "ldr s23, [x22, #0x0]\n"
       "b 101f\n"
       "97:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x9, #2, 99f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
       "tbz x9, #1, 98f\n"
       "ldr d10, [x27], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d22, [x23], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
       "tbz x9, #0, 101f\n"
       "ld1 { v10.s }[2], [x27]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
       "b 101f\n"
       "98:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x9, #0, 101f\n"
       "ldr s10, [x27, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
       "b 101f\n"
       "99:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x9, #1, 100f\n"
       "ldr d9, [x27], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d21, [x23], #0x8\n"
+      "ldr d21, [x22], #0x8\n"
       "tbz x9, #0, 101f\n"
       "ld1 { v9.s }[2], [x27]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x23]\n"
+      "ld1 { v21.s }[2], [x22]\n"
       "b 101f\n"
       "100:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s9, [x27, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s21, [x23, #0x0]\n"
+      "ldr s21, [x22, #0x0]\n"
       "101:"  // Height 3: Partial accumulate: Done
       "sub x27, x27, x20\n"
       "b 103f\n"
@@ -1171,18 +1169,18 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q12, [x27, #0x30]\n"
       "ldr q13, [x27, #0x40]\n"
       "ldr q20, [x27, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
-      "ldr q21, [x23, #0x0]\n"
-      "ldr q22, [x23, #0x10]\n"
-      "ldr q23, [x23, #0x20]\n"
-      "ldr q24, [x23, #0x30]\n"
-      "ldr q25, [x23, #0x40]\n"
-      "ldr q4, [x23, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
+      "ldr q21, [x22, #0x0]\n"
+      "ldr q22, [x22, #0x10]\n"
+      "ldr q23, [x22, #0x20]\n"
+      "ldr q24, [x22, #0x30]\n"
+      "ldr q25, [x22, #0x40]\n"
+      "ldr q4, [x22, #0x50]\n"
       "103:"  // Height 3: MMLA fixup
       "zip1 v8.2d, v9.2d, v14.2d\n"
       "zip2 v14.2d, v9.2d, v14.2d\n"
@@ -1238,8 +1236,8 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "mov x26, #0x0\n"
       "106:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 107f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1270,33 +1268,33 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "blt 110f\n"
       "109:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "sub x25, x25, #0x4\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "cmp x25, #0x8\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      "ld1 { v1.4s }, [x23], #0x10\n"
-      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
-      ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
-      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
+      "sub x25, x25, #0x4\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x40]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
       "ldr q5, [x28, #0x50]\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      "cmp x25, #0x8\n"
+      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
       "ldr q6, [x28, #0x60]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       "ldr q3, [x28, #0x70]\n"
       ".inst 0x6e44ec0a  // bfmmla v10.4s, v0.8h, v4.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      "ld1 { v1.4s }, [x23], #0x10\n"
       ".inst 0x6e44ec56  // bfmmla v22.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x80]\n"
       ".inst 0x6e45ec10  // bfmmla v16.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec5c  // bfmmla v28.4s, v2.8h, v5.8h\n"
       "ldr q5, [x28, #0x90]\n"
       ".inst 0x6e46ec0b  // bfmmla v11.4s, v0.8h, v6.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e46ec57  // bfmmla v23.4s, v2.8h, v6.8h\n"
       "ldr q6, [x28, #0xa0]\n"
       ".inst 0x6e43ec11  // bfmmla v17.4s, v0.8h, v3.8h\n"
@@ -1320,25 +1318,25 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "bge 109b\n"
       "110:"  // Height 3: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "sub x25, x25, #0x4\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
-      ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
-      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
+      "sub x25, x25, #0x4\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
       "ldr q3, [x28, #0x40]\n"
       ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
       "ldr q4, [x28, #0x50]\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
       "ldr q6, [x28, #0x60]\n"
       ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       "ldr q1, [x28, #0x70]\n"
       ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e43ec56  // bfmmla v22.4s, v2.8h, v3.8h\n"
       "ldr q5, [x28, #0x80]\n"
       ".inst 0x6e44ec10  // bfmmla v16.4s, v0.8h, v4.8h\n"
@@ -1379,21 +1377,21 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q5, [x28, #0x0]\n"
       "ldr q4, [x28, #0x10]\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "ldr q3, [x28, #0x20]\n"
-      "ldr q6, [x28, #0x30]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      ".inst 0x6e45ec54  // bfmmla v20.4s, v2.8h, v5.8h\n"
-      ".inst 0x6e44ec5a  // bfmmla v26.4s, v2.8h, v4.8h\n"
+      "ldr q3, [x28, #0x20]\n"
+      "ldr q1, [x28, #0x30]\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       ".inst 0x6e45ec08  // bfmmla v8.4s, v0.8h, v5.8h\n"
+      ".inst 0x6e45ec54  // bfmmla v20.4s, v2.8h, v5.8h\n"
       "ldr q5, [x28, #0x40]\n"
       ".inst 0x6e44ec0e  // bfmmla v14.4s, v0.8h, v4.8h\n"
+      ".inst 0x6e44ec5a  // bfmmla v26.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x50]\n"
       ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
       ".inst 0x6e43ec55  // bfmmla v21.4s, v2.8h, v3.8h\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x6e46ec0f  // bfmmla v15.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e46ec5b  // bfmmla v27.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e41ec0f  // bfmmla v15.4s, v0.8h, v1.8h\n"
+      ".inst 0x6e41ec5b  // bfmmla v27.4s, v2.8h, v1.8h\n"
       "ldr q1, [x28, #0x70]\n"
       ".inst 0x6e45ec0a  // bfmmla v10.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec56  // bfmmla v22.4s, v2.8h, v5.8h\n"
@@ -1422,19 +1420,19 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "cmp x26, x20\n"
       "bne 106b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp1 v4.2d, v8.2d, v14.2d\n"
       "uzp2 v8.2d, v8.2d, v14.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v14.2d, v9.2d, v15.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v9.2d, v9.2d, v15.2d\n"
       "uzp1 v15.2d, v10.2d, v16.2d\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v16.2d\n"
-      "add x24, x27, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v16.2d, v11.2d, v17.2d\n"
       "uzp2 v11.2d, v11.2d, v17.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v17.2d, v12.2d, v18.2d\n"
       "uzp2 v12.2d, v12.2d, v18.2d\n"
       "uzp1 v18.2d, v13.2d, v19.2d\n"
@@ -1446,9 +1444,9 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "uzp1 v24.2d, v24.2d, v30.2d\n"
       "uzp1 v25.2d, v25.2d, v31.2d\n"
       "tbz %x[flags], #1, 115f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v4.4s, v4.4s, v1.4s\n"
       "fmin v14.4s, v14.4s, v1.4s\n"
@@ -1494,126 +1492,126 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "st1 { v14.4s }, [x27], #0x10\n"
       "st1 { v15.4s }, [x27], #0x10\n"
       "st1 { v16.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v11.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v11.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
       "tbz x9, #2, 117f\n"
       "st1 { v17.4s }, [x27], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v12.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x9, #1, 116f\n"
       "str d18, [x27], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d13, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x9, #0, 127f\n"
       "st1 { v18.s }[2], [x27]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v13.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 127f\n"
       "116:"  // Height 3: Partial direct writeback: partial_1_20
       "tbz x9, #0, 127f\n"
       "str s18, [x27, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s13, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 127f\n"
       "117:"  // Height 3: Partial direct writeback: partial_2_16
       "tbz x9, #1, 118f\n"
       "str d17, [x27], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d12, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x9, #0, 127f\n"
       "st1 { v17.s }[2], [x27]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v12.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 127f\n"
       "118:"  // Height 3: Partial direct writeback: partial_1_16
       "tbz x9, #0, 127f\n"
       "str s17, [x27, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s12, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "b 127f\n"
       "119:"  // Height 3: Partial direct writeback: partial_8_0
       "tbz x9, #3, 123f\n"
       "st1 { v4.4s }, [x27], #0x10\n"
       "st1 { v14.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v21.4s }, [x22], #0x10\n"
       "tbz x9, #2, 121f\n"
       "st1 { v15.4s }, [x27], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v22.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v22.4s }, [x22], #0x10\n"
       "tbz x9, #1, 120f\n"
       "str d16, [x27], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
+      "str d11, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
       "tbz x9, #0, 127f\n"
       "st1 { v16.s }[2], [x27]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
+      "st1 { v11.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
       "b 127f\n"
       "120:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x9, #0, 127f\n"
       "str s16, [x27, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
+      "str s11, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
       "b 127f\n"
       "121:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x9, #1, 122f\n"
       "str d15, [x27], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
+      "str d10, [x23], #0x8\n"
+      "str d22, [x22], #0x8\n"
       "tbz x9, #0, 127f\n"
       "st1 { v15.s }[2], [x27]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
+      "st1 { v10.s }[2], [x23]\n"
+      "st1 { v22.s }[2], [x22]\n"
       "b 127f\n"
       "122:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x9, #0, 127f\n"
       "str s15, [x27, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
+      "str s10, [x23, #0x0]\n"
+      "str s22, [x22, #0x0]\n"
       "b 127f\n"
       "123:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x9, #2, 125f\n"
       "st1 { v4.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v20.4s }, [x22], #0x10\n"
       "tbz x9, #1, 124f\n"
       "str d14, [x27], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
+      "str d9, [x23], #0x8\n"
+      "str d21, [x22], #0x8\n"
       "tbz x9, #0, 127f\n"
       "st1 { v14.s }[2], [x27]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
+      "st1 { v9.s }[2], [x23]\n"
+      "st1 { v21.s }[2], [x22]\n"
       "b 127f\n"
       "124:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x9, #0, 127f\n"
       "str s14, [x27, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
+      "str s9, [x23, #0x0]\n"
+      "str s21, [x22, #0x0]\n"
       "b 127f\n"
       "125:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x9, #1, 126f\n"
       "str d4, [x27], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
+      "str d8, [x23], #0x8\n"
+      "str d20, [x22], #0x8\n"
       "tbz x9, #0, 127f\n"
       "st1 { v4.s }[2], [x27]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
+      "st1 { v8.s }[2], [x23]\n"
+      "st1 { v20.s }[2], [x22]\n"
       "b 127f\n"
       "126:"  // Height 3: Partial direct writeback: partial_1_0
       "str s4, [x27, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
+      "str s8, [x23, #0x0]\n"
+      "str s20, [x22, #0x0]\n"
       "127:"  // Height 3: Partial direct writeback: Done
       "b 129f\n"
       "128:"  // Height 3: Full writeback
@@ -1624,48 +1622,47 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "str q17, [x27, #0x40]\n"
       "str q18, [x27, #0x50]\n"
       "add x27, x27, #0x60\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q12, [x24, #0x40]\n"
-      "str q13, [x24, #0x50]\n"
-      "str q20, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q22, [x23, #0x20]\n"
-      "str q23, [x23, #0x30]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q25, [x23, #0x50]\n"
+      "str q8, [x23, #0x0]\n"
+      "str q9, [x23, #0x10]\n"
+      "str q10, [x23, #0x20]\n"
+      "str q11, [x23, #0x30]\n"
+      "str q12, [x23, #0x40]\n"
+      "str q13, [x23, #0x50]\n"
+      "str q20, [x22, #0x0]\n"
+      "str q21, [x22, #0x10]\n"
+      "str q22, [x22, #0x20]\n"
+      "str q23, [x22, #0x30]\n"
+      "str q24, [x22, #0x40]\n"
+      "str q25, [x22, #0x50]\n"
       "129:"  // Height 3: Writeback done
       "subs x9, x9, #0x18\n"
       "bgt 88b\n"
       "b 174f\n"
       "130:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x10\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "131:"  // Height 4: Column loop
       "cbz x10, 132f\n"
       "ldr q8, [x10, #0x0]\n"
       "ldr q9, [x10, #0x10]\n"
-      "ldr q10, [x10, #0x20]\n"
-      "ldr q11, [x10, #0x30]\n"
-      "ldr q12, [x10, #0x40]\n"
-      "ldr q13, [x10, #0x50]\n"
-      "add x10, x10, #0x60\n"
       "zip2 v14.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x10, #0x20]\n"
+      "ldr q11, [x10, #0x30]\n"
       "zip2 v15.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
+      "ldr q12, [x10, #0x40]\n"
+      "ldr q13, [x10, #0x50]\n"
       "zip2 v16.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
       "zip2 v17.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
+      "add x10, x10, #0x60\n"
       "zip2 v18.2d, v12.2d, v12.2d\n"
       "zip1 v12.2d, v12.2d, v12.2d\n"
       "zip2 v19.2d, v13.2d, v13.2d\n"
@@ -1686,175 +1683,175 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "132:"  // Height 4: no bias
       "tbz %x[flags], #0, 147f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x9, #0x18\n"
-      "add x24, x27, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
+      "add x23, x27, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x9, #0x18\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 145f\n"
       "tbz x9, #4, 136f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
-      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
+      "ld1 { v27.4s }, [x21], #0x10\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v12.4s }, [x27], #0x10\n"
-      "ld1 { v17.4s }, [x24], #0x10\n"
-      "ld1 { v24.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v17.4s }, [x23], #0x10\n"
+      "ld1 { v24.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x9, #2, 134f\n"
       "ld1 { v13.4s }, [x27], #0x10\n"
-      "ld1 { v18.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v18.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x9, #1, 133f\n"
       "ldr d20, [x27], #0x8\n"
-      "ldr d19, [x24], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "mov x20, #0x58\n"
-      "ldr d4, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d4, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x9, #0, 144f\n"
       "ld1 { v20.s }[2], [x27]\n"
-      "ld1 { v19.s }[2], [x24]\n"
-      "ld1 { v4.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v19.s }[2], [x23]\n"
+      "ld1 { v4.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 144f\n"
       "133:"  // Height 4: Partial accumulate: partial_1_20
       "mov x20, #0x50\n"
       "tbz x9, #0, 144f\n"
       "ldr s20, [x27, #0x0]\n"
-      "ldr s19, [x24, #0x0]\n"
-      "ldr s4, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s19, [x23, #0x0]\n"
+      "ldr s4, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 144f\n"
       "134:"  // Height 4: Partial accumulate: partial_2_16
       "tbz x9, #1, 135f\n"
       "ldr d13, [x27], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "mov x20, #0x48\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x9, #0, 144f\n"
       "ld1 { v13.s }[2], [x27]\n"
-      "ld1 { v18.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v18.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 144f\n"
       "135:"  // Height 4: Partial accumulate: partial_1_16
       "mov x20, #0x40\n"
       "tbz x9, #0, 144f\n"
       "ldr s13, [x27, #0x0]\n"
-      "ldr s18, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s18, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 144f\n"
       "136:"  // Height 4: Partial accumulate: partial_8_0
       "tbz x9, #3, 140f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x27], #0x10\n"
-      "ld1 { v15.4s }, [x24], #0x10\n"
-      "ld1 { v22.4s }, [x23], #0x10\n"
-      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v15.4s }, [x23], #0x10\n"
+      "ld1 { v22.4s }, [x22], #0x10\n"
+      "ld1 { v27.4s }, [x21], #0x10\n"
       "tbz x9, #2, 138f\n"
       "ld1 { v11.4s }, [x27], #0x10\n"
-      "ld1 { v16.4s }, [x24], #0x10\n"
-      "ld1 { v23.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v16.4s }, [x23], #0x10\n"
+      "ld1 { v23.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x9, #1, 137f\n"
       "ldr d12, [x27], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d24, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x9, #0, 144f\n"
       "ld1 { v12.s }[2], [x27]\n"
-      "ld1 { v17.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v17.s }[2], [x23]\n"
+      "ld1 { v24.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 144f\n"
       "137:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x9, #0, 144f\n"
       "ldr s12, [x27, #0x0]\n"
-      "ldr s17, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s17, [x23, #0x0]\n"
+      "ldr s24, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 144f\n"
       "138:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x9, #1, 139f\n"
       "ldr d11, [x27], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d23, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d23, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x9, #0, 144f\n"
       "ld1 { v11.s }[2], [x27]\n"
-      "ld1 { v16.s }[2], [x24]\n"
-      "ld1 { v23.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v16.s }[2], [x23]\n"
+      "ld1 { v23.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 144f\n"
       "139:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x9, #0, 144f\n"
       "ldr s11, [x27, #0x0]\n"
-      "ldr s16, [x24, #0x0]\n"
-      "ldr s23, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s16, [x23, #0x0]\n"
+      "ldr s23, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "b 144f\n"
       "140:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x9, #2, 142f\n"
       "ld1 { v9.4s }, [x27], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v21.4s }, [x23], #0x10\n"
-      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x23], #0x10\n"
+      "ld1 { v21.4s }, [x22], #0x10\n"
+      "ld1 { v26.4s }, [x21], #0x10\n"
       "tbz x9, #1, 141f\n"
       "ldr d10, [x27], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
+      "ldr d15, [x23], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d27, [x22], #0x8\n"
+      "ldr d22, [x22], #0x8\n"
+      "ldr d27, [x21], #0x8\n"
       "tbz x9, #0, 144f\n"
       "ld1 { v10.s }[2], [x27]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v22.s }[2], [x23]\n"
-      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x23]\n"
+      "ld1 { v22.s }[2], [x22]\n"
+      "ld1 { v27.s }[2], [x21]\n"
       "b 144f\n"
       "141:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x9, #0, 144f\n"
       "ldr s10, [x27, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s22, [x23, #0x0]\n"
-      "ldr s27, [x22, #0x0]\n"
+      "ldr s15, [x23, #0x0]\n"
+      "ldr s22, [x22, #0x0]\n"
+      "ldr s27, [x21, #0x0]\n"
       "b 144f\n"
       "142:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x9, #1, 143f\n"
       "ldr d9, [x27], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
+      "ldr d14, [x23], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d21, [x23], #0x8\n"
-      "ldr d26, [x22], #0x8\n"
+      "ldr d21, [x22], #0x8\n"
+      "ldr d26, [x21], #0x8\n"
       "tbz x9, #0, 144f\n"
       "ld1 { v9.s }[2], [x27]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v21.s }[2], [x23]\n"
-      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x23]\n"
+      "ld1 { v21.s }[2], [x22]\n"
+      "ld1 { v26.s }[2], [x21]\n"
       "b 144f\n"
       "143:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s9, [x27, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
+      "ldr s14, [x23, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s21, [x23, #0x0]\n"
-      "ldr s26, [x22, #0x0]\n"
+      "ldr s21, [x22, #0x0]\n"
+      "ldr s26, [x21, #0x0]\n"
       "144:"  // Height 4: Partial accumulate: Done
       "sub x27, x27, x20\n"
       "b 146f\n"
@@ -1865,24 +1862,24 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "ldr q12, [x27, #0x30]\n"
       "ldr q13, [x27, #0x40]\n"
       "ldr q20, [x27, #0x50]\n"
-      "ldr q14, [x24, #0x0]\n"
-      "ldr q15, [x24, #0x10]\n"
-      "ldr q16, [x24, #0x20]\n"
-      "ldr q17, [x24, #0x30]\n"
-      "ldr q18, [x24, #0x40]\n"
-      "ldr q19, [x24, #0x50]\n"
-      "ldr q21, [x23, #0x0]\n"
-      "ldr q22, [x23, #0x10]\n"
-      "ldr q23, [x23, #0x20]\n"
-      "ldr q24, [x23, #0x30]\n"
-      "ldr q25, [x23, #0x40]\n"
-      "ldr q4, [x23, #0x50]\n"
-      "ldr q26, [x22, #0x0]\n"
-      "ldr q27, [x22, #0x10]\n"
-      "ldr q28, [x22, #0x20]\n"
-      "ldr q29, [x22, #0x30]\n"
-      "ldr q30, [x22, #0x40]\n"
-      "ldr q31, [x22, #0x50]\n"
+      "ldr q14, [x23, #0x0]\n"
+      "ldr q15, [x23, #0x10]\n"
+      "ldr q16, [x23, #0x20]\n"
+      "ldr q17, [x23, #0x30]\n"
+      "ldr q18, [x23, #0x40]\n"
+      "ldr q19, [x23, #0x50]\n"
+      "ldr q21, [x22, #0x0]\n"
+      "ldr q22, [x22, #0x10]\n"
+      "ldr q23, [x22, #0x20]\n"
+      "ldr q24, [x22, #0x30]\n"
+      "ldr q25, [x22, #0x40]\n"
+      "ldr q4, [x22, #0x50]\n"
+      "ldr q26, [x21, #0x0]\n"
+      "ldr q27, [x21, #0x10]\n"
+      "ldr q28, [x21, #0x20]\n"
+      "ldr q29, [x21, #0x30]\n"
+      "ldr q30, [x21, #0x40]\n"
+      "ldr q31, [x21, #0x50]\n"
       "146:"  // Height 4: MMLA fixup
       "zip1 v8.2d, v9.2d, v14.2d\n"
       "zip2 v14.2d, v9.2d, v14.2d\n"
@@ -1938,8 +1935,8 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "mov x26, #0x0\n"
       "149:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 150f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1976,28 +1973,28 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x25, x25, #0x4\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       "cmp x25, #0x8\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      "ld1 { v1.4s }, [x23], #0x10\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
-      "ld1 { v3.4s }, [x21], #0x10\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
-      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
-      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x40]\n"
+      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      "ld1 { v1.4s }, [x23], #0x10\n"
       ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
       "ldr q5, [x28, #0x50]\n"
+      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
       "ldr q6, [x28, #0x60]\n"
+      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       "ldr q7, [x28, #0x70]\n"
       ".inst 0x6e44ec0a  // bfmmla v10.4s, v0.8h, v4.8h\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
+      "ld1 { v3.4s }, [x21], #0x10\n"
       ".inst 0x6e44ec56  // bfmmla v22.4s, v2.8h, v4.8h\n"
       "ldr q4, [x28, #0x80]\n"
       ".inst 0x6e45ec10  // bfmmla v16.4s, v0.8h, v5.8h\n"
@@ -2030,21 +2027,21 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x25, x25, #0x4\n"
       "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
       ".inst 0x6e44ec08  // bfmmla v8.4s, v0.8h, v4.8h\n"
-      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
-      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e44ec54  // bfmmla v20.4s, v2.8h, v4.8h\n"
       "ldr q3, [x28, #0x40]\n"
+      ".inst 0x6e45ec0e  // bfmmla v14.4s, v0.8h, v5.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e45ec5a  // bfmmla v26.4s, v2.8h, v5.8h\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e46ec55  // bfmmla v21.4s, v2.8h, v6.8h\n"
       "ldr q6, [x28, #0x60]\n"
+      ".inst 0x6e47ec0f  // bfmmla v15.4s, v0.8h, v7.8h\n"
       ".inst 0x6e47ec5b  // bfmmla v27.4s, v2.8h, v7.8h\n"
       "ldr q1, [x28, #0x70]\n"
       ".inst 0x6e43ec0a  // bfmmla v10.4s, v0.8h, v3.8h\n"
@@ -2135,23 +2132,23 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "cmp x26, x20\n"
       "bne 149b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp1 v4.2d, v8.2d, v14.2d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v14.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v14.2d, v9.2d, v15.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp2 v9.2d, v9.2d, v15.2d\n"
       "uzp1 v15.2d, v10.2d, v16.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v16.2d\n"
-      "add x24, x27, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v16.2d, v11.2d, v17.2d\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v17.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v17.2d, v12.2d, v18.2d\n"
       "uzp2 v12.2d, v12.2d, v18.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v18.2d, v13.2d, v19.2d\n"
       "uzp2 v13.2d, v13.2d, v19.2d\n"
       "uzp1 v19.2d, v20.2d, v26.2d\n"
@@ -2167,9 +2164,9 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "uzp1 v30.2d, v25.2d, v31.2d\n"
       "uzp2 v25.2d, v25.2d, v31.2d\n"
       "tbz %x[flags], #1, 158f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v4.4s, v4.4s, v1.4s\n"
       "fmin v14.4s, v14.4s, v1.4s\n"
@@ -2227,153 +2224,153 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "st1 { v14.4s }, [x27], #0x10\n"
       "st1 { v15.4s }, [x27], #0x10\n"
       "st1 { v16.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v11.4s }, [x24], #0x10\n"
-      "st1 { v19.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
-      "st1 { v27.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v21.4s }, [x22], #0x10\n"
-      "st1 { v22.4s }, [x22], #0x10\n"
-      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v11.4s }, [x23], #0x10\n"
+      "st1 { v19.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v27.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v20.4s }, [x21], #0x10\n"
+      "st1 { v21.4s }, [x21], #0x10\n"
+      "st1 { v22.4s }, [x21], #0x10\n"
+      "st1 { v23.4s }, [x21], #0x10\n"
       "tbz x9, #2, 160f\n"
       "st1 { v17.4s }, [x27], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v29.4s }, [x23], #0x10\n"
-      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v12.4s }, [x23], #0x10\n"
+      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v24.4s }, [x21], #0x10\n"
       "tbz x9, #1, 159f\n"
       "str d18, [x27], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
-      "str d25, [x22], #0x8\n"
+      "str d13, [x23], #0x8\n"
+      "str d30, [x22], #0x8\n"
+      "str d25, [x21], #0x8\n"
       "tbz x9, #0, 170f\n"
       "st1 { v18.s }[2], [x27]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
-      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v13.s }[2], [x23]\n"
+      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v25.s }[2], [x21]\n"
       "b 170f\n"
       "159:"  // Height 4: Partial direct writeback: partial_1_20
       "tbz x9, #0, 170f\n"
       "str s18, [x27, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
-      "str s25, [x22, #0x0]\n"
+      "str s13, [x23, #0x0]\n"
+      "str s30, [x22, #0x0]\n"
+      "str s25, [x21, #0x0]\n"
       "b 170f\n"
       "160:"  // Height 4: Partial direct writeback: partial_2_16
       "tbz x9, #1, 161f\n"
       "str d17, [x27], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "str d29, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
+      "str d12, [x23], #0x8\n"
+      "str d29, [x22], #0x8\n"
+      "str d24, [x21], #0x8\n"
       "tbz x9, #0, 170f\n"
       "st1 { v17.s }[2], [x27]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "st1 { v29.s }[2], [x23]\n"
-      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v12.s }[2], [x23]\n"
+      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v24.s }[2], [x21]\n"
       "b 170f\n"
       "161:"  // Height 4: Partial direct writeback: partial_1_16
       "tbz x9, #0, 170f\n"
       "str s17, [x27, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "str s29, [x23, #0x0]\n"
-      "str s24, [x22, #0x0]\n"
+      "str s12, [x23, #0x0]\n"
+      "str s29, [x22, #0x0]\n"
+      "str s24, [x21, #0x0]\n"
       "b 170f\n"
       "162:"  // Height 4: Partial direct writeback: partial_8_0
       "tbz x9, #3, 166f\n"
       "st1 { v4.4s }, [x27], #0x10\n"
       "st1 { v14.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v19.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v21.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v9.4s }, [x23], #0x10\n"
+      "st1 { v19.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v20.4s }, [x21], #0x10\n"
+      "st1 { v21.4s }, [x21], #0x10\n"
       "tbz x9, #2, 164f\n"
       "st1 { v15.4s }, [x27], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v27.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x22], #0x10\n"
+      "st1 { v10.4s }, [x23], #0x10\n"
+      "st1 { v27.4s }, [x22], #0x10\n"
+      "st1 { v22.4s }, [x21], #0x10\n"
       "tbz x9, #1, 163f\n"
       "str d16, [x27], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "str d28, [x23], #0x8\n"
-      "str d23, [x22], #0x8\n"
+      "str d11, [x23], #0x8\n"
+      "str d28, [x22], #0x8\n"
+      "str d23, [x21], #0x8\n"
       "tbz x9, #0, 170f\n"
       "st1 { v16.s }[2], [x27]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "st1 { v28.s }[2], [x23]\n"
-      "st1 { v23.s }[2], [x22]\n"
+      "st1 { v11.s }[2], [x23]\n"
+      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v23.s }[2], [x21]\n"
       "b 170f\n"
       "163:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x9, #0, 170f\n"
       "str s16, [x27, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "str s28, [x23, #0x0]\n"
-      "str s23, [x22, #0x0]\n"
+      "str s11, [x23, #0x0]\n"
+      "str s28, [x22, #0x0]\n"
+      "str s23, [x21, #0x0]\n"
       "b 170f\n"
       "164:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x9, #1, 165f\n"
       "str d15, [x27], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
-      "str d22, [x22], #0x8\n"
+      "str d10, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
+      "str d22, [x21], #0x8\n"
       "tbz x9, #0, 170f\n"
       "st1 { v15.s }[2], [x27]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
-      "st1 { v22.s }[2], [x22]\n"
+      "st1 { v10.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v22.s }[2], [x21]\n"
       "b 170f\n"
       "165:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x9, #0, 170f\n"
       "str s15, [x27, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
-      "str s22, [x22, #0x0]\n"
+      "str s10, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
+      "str s22, [x21, #0x0]\n"
       "b 170f\n"
       "166:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x9, #2, 168f\n"
       "st1 { v4.4s }, [x27], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v19.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x23], #0x10\n"
+      "st1 { v19.4s }, [x22], #0x10\n"
+      "st1 { v20.4s }, [x21], #0x10\n"
       "tbz x9, #1, 167f\n"
       "str d14, [x27], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
-      "str d21, [x22], #0x8\n"
+      "str d9, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
+      "str d21, [x21], #0x8\n"
       "tbz x9, #0, 170f\n"
       "st1 { v14.s }[2], [x27]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
-      "st1 { v21.s }[2], [x22]\n"
+      "st1 { v9.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v21.s }[2], [x21]\n"
       "b 170f\n"
       "167:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x9, #0, 170f\n"
       "str s14, [x27, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
-      "str s21, [x22, #0x0]\n"
+      "str s9, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
+      "str s21, [x21, #0x0]\n"
       "b 170f\n"
       "168:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x9, #1, 169f\n"
       "str d4, [x27], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "str d19, [x23], #0x8\n"
-      "str d20, [x22], #0x8\n"
+      "str d8, [x23], #0x8\n"
+      "str d19, [x22], #0x8\n"
+      "str d20, [x21], #0x8\n"
       "tbz x9, #0, 170f\n"
       "st1 { v4.s }[2], [x27]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "st1 { v19.s }[2], [x23]\n"
-      "st1 { v20.s }[2], [x22]\n"
+      "st1 { v8.s }[2], [x23]\n"
+      "st1 { v19.s }[2], [x22]\n"
+      "st1 { v20.s }[2], [x21]\n"
       "b 170f\n"
       "169:"  // Height 4: Partial direct writeback: partial_1_0
       "str s4, [x27, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "str s19, [x23, #0x0]\n"
-      "str s20, [x22, #0x0]\n"
+      "str s8, [x23, #0x0]\n"
+      "str s19, [x22, #0x0]\n"
+      "str s20, [x21, #0x0]\n"
       "170:"  // Height 4: Partial direct writeback: Done
       "b 172f\n"
       "171:"  // Height 4: Full writeback
@@ -2384,24 +2381,24 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "str q17, [x27, #0x40]\n"
       "str q18, [x27, #0x50]\n"
       "add x27, x27, #0x60\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q12, [x24, #0x40]\n"
-      "str q13, [x24, #0x50]\n"
-      "str q19, [x23, #0x0]\n"
-      "str q26, [x23, #0x10]\n"
-      "str q27, [x23, #0x20]\n"
-      "str q28, [x23, #0x30]\n"
-      "str q29, [x23, #0x40]\n"
-      "str q30, [x23, #0x50]\n"
-      "str q20, [x22, #0x0]\n"
-      "str q21, [x22, #0x10]\n"
-      "str q22, [x22, #0x20]\n"
-      "str q23, [x22, #0x30]\n"
-      "str q24, [x22, #0x40]\n"
-      "str q25, [x22, #0x50]\n"
+      "str q8, [x23, #0x0]\n"
+      "str q9, [x23, #0x10]\n"
+      "str q10, [x23, #0x20]\n"
+      "str q11, [x23, #0x30]\n"
+      "str q12, [x23, #0x40]\n"
+      "str q13, [x23, #0x50]\n"
+      "str q19, [x22, #0x0]\n"
+      "str q26, [x22, #0x10]\n"
+      "str q27, [x22, #0x20]\n"
+      "str q28, [x22, #0x30]\n"
+      "str q29, [x22, #0x40]\n"
+      "str q30, [x22, #0x50]\n"
+      "str q20, [x21, #0x0]\n"
+      "str q21, [x21, #0x10]\n"
+      "str q22, [x21, #0x20]\n"
+      "str q23, [x21, #0x30]\n"
+      "str q24, [x21, #0x40]\n"
+      "str q25, [x21, #0x50]\n"
       "172:"  // Height 4: Writeback done
       "subs x9, x9, #0x18\n"
       "bgt 131b\n"
@@ -2417,8 +2414,8 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "174:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16.hpp
index 6780b76a3a..191528d7f9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp
index d4a4f18d2b..9dde5b0b92 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp
@@ -48,19 +48,18 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -81,7 +80,6 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -104,23 +102,23 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "cmp %x[M], #0x2\n"
       "bgt 71f\n"
       "beq 36f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "cbz x12, 3f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "b 15f\n"
@@ -211,8 +209,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "16:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 17f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -233,9 +231,6 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "blt 20f\n"
       "19:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "cmp x27, #0x8\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
@@ -248,17 +243,18 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
+      "add x10, x10, #0x80\n"
       "ldr q6, [x10, #0x0]\n"
       ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
-      "ld1 { v0.4s }, [x26], #0x10\n"
       "ldr q7, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
       "bge 19b\n"
       "20:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
@@ -271,9 +267,11 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "sub x27, x27, #0x4\n"
       ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
       ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x80\n"
       "21:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 24f\n"
       "cbz x27, 24f\n"
@@ -300,9 +298,9 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
       ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "add x10, x10, #0x80\n"
       "24:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -314,9 +312,9 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "uzp1 v10.2d, v10.2d, v14.2d\n"
       "uzp1 v11.2d, v11.2d, v15.2d\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v18.4s }, [x21]\n"
       "ld1r { v17.4s }, [x20]\n"
       "fmin v8.4s, v8.4s, v18.4s\n"
       "fmin v9.4s, v9.4s, v18.4s\n"
@@ -386,23 +384,23 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "bgt 2b\n"
       "b 212f\n"
       "36:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "37:"  // Height 2: Column loop
       "cbz x12, 38f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "b 50f\n"
@@ -410,75 +408,75 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "tbz %x[flags], #0, 49f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "bge 47f\n"
       "tbz x11, #3, 42f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
       "tbz x11, #2, 40f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
       "tbz x11, #1, 39f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
       "tbz x11, #0, 46f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
+      "ld1 { v15.s }[2], [x25]\n"
       "b 46f\n"
       "39:"  // Height 2: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 46f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
       "b 46f\n"
       "40:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x11, #1, 41f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
       "tbz x11, #0, 46f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
+      "ld1 { v14.s }[2], [x25]\n"
       "b 46f\n"
       "41:"  // Height 2: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 46f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
       "b 46f\n"
       "42:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x11, #2, 44f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
       "tbz x11, #1, 43f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
       "tbz x11, #0, 46f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
+      "ld1 { v13.s }[2], [x25]\n"
       "b 46f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 46f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
       "b 46f\n"
       "44:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x11, #1, 45f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
       "tbz x11, #0, 46f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
+      "ld1 { v12.s }[2], [x25]\n"
       "b 46f\n"
       "45:"  // Height 2: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
       "46:"  // Height 2: Partial accumulate: Done
       "sub x9, x9, x20\n"
@@ -488,10 +486,10 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
       "48:"  // Height 2: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -515,8 +513,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "51:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 52f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -541,12 +539,7 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "blt 55f\n"
       "54:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "cmp x27, #0x8\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      "ld1 { v1.4s }, [x25], #0x10\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
@@ -559,18 +552,20 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x8\n"
       "add x10, x10, #0x80\n"
       ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
       "ldr q6, [x10, #0x0]\n"
       ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
-      "ld1 { v0.4s }, [x26], #0x10\n"
       "ldr q7, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "ld1 { v0.4s }, [x26], #0x10\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
       "bge 54b\n"
       "55:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
       "ldr q18, [x10, #0x20]\n"
@@ -584,9 +579,12 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "sub x27, x27, #0x4\n"
       ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
       ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x10, x10, #0x80\n"
       "56:"  // Height 2: Multiply loop: Main loop skip
       "cbz x27, 59f\n"
       "cbz x27, 59f\n"
@@ -617,30 +615,30 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e51ec0e  // bfmmla v14.4s, v0.8h, v17.8h\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e52ec0b  // bfmmla v11.4s, v0.8h, v18.8h\n"
       ".inst 0x6e51ec0f  // bfmmla v15.4s, v0.8h, v17.8h\n"
+      "add x10, x10, #0x80\n"
       "59:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 51b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "uzp1 v6.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "tbz %x[flags], #1, 60f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v18.4s }, [x21]\n"
       "ld1r { v17.4s }, [x20]\n"
       "fmin v6.4s, v6.4s, v18.4s\n"
       "fmin v12.4s, v12.4s, v18.4s\n"
@@ -664,63 +662,63 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "tbz x11, #3, 64f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
       "tbz x11, #2, 62f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
       "tbz x11, #1, 61f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
+      "str d11, [x25], #0x8\n"
       "tbz x11, #0, 68f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
+      "st1 { v11.s }[2], [x25]\n"
       "b 68f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x11, #0, 68f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
       "b 68f\n"
       "62:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x11, #1, 63f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
+      "str d10, [x25], #0x8\n"
       "tbz x11, #0, 68f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
+      "st1 { v10.s }[2], [x25]\n"
       "b 68f\n"
       "63:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x11, #0, 68f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
       "b 68f\n"
       "64:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x11, #2, 66f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
       "tbz x11, #1, 65f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
+      "str d9, [x25], #0x8\n"
       "tbz x11, #0, 68f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
+      "st1 { v9.s }[2], [x25]\n"
       "b 68f\n"
       "65:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x11, #0, 68f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
       "b 68f\n"
       "66:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x11, #1, 67f\n"
       "str d6, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
+      "str d8, [x25], #0x8\n"
       "tbz x11, #0, 68f\n"
       "st1 { v6.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
+      "st1 { v8.s }[2], [x25]\n"
       "b 68f\n"
       "67:"  // Height 2: Partial direct writeback: partial_1_0
       "str s6, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
       "68:"  // Height 2: Partial direct writeback: Done
       "b 70f\n"
       "69:"  // Height 2: Full writeback
@@ -729,32 +727,32 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
       "70:"  // Height 2: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 37b\n"
       "b 212f\n"
       "71:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "72:"  // Height 3: Column loop
       "cbz x12, 73f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -769,94 +767,94 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "73:"  // Height 3: no bias
       "tbz %x[flags], #0, 84f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "bge 82f\n"
       "tbz x11, #3, 77f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
       "tbz x11, #2, 75f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
       "tbz x11, #1, 74f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
       "tbz x11, #0, 81f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
       "b 81f\n"
       "74:"  // Height 3: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 81f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
       "b 81f\n"
       "75:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x11, #1, 76f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
       "tbz x11, #0, 81f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
       "b 81f\n"
       "76:"  // Height 3: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 81f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
       "b 81f\n"
       "77:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x11, #2, 79f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
       "tbz x11, #1, 78f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
       "tbz x11, #0, 81f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
       "b 81f\n"
       "78:"  // Height 3: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 81f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
       "b 81f\n"
       "79:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x11, #1, 80f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
       "tbz x11, #0, 81f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
       "b 81f\n"
       "80:"  // Height 3: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
       "81:"  // Height 3: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 83f\n"
@@ -865,14 +863,14 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
       "83:"  // Height 3: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -912,8 +910,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "86:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 87f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -942,33 +940,33 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "blt 90f\n"
       "89:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "cmp x27, #0x8\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      "ld1 { v1.4s }, [x25], #0x10\n"
-      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "sub x27, x27, #0x4\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
       ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
       ".inst 0x6e59ec56  // bfmmla v22.4s, v2.8h, v25.8h\n"
       "ldr q25, [x10, #0x70]\n"
       "add x10, x10, #0x80\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e5aec0b  // bfmmla v11.4s, v0.8h, v26.8h\n"
       ".inst 0x6e5aec53  // bfmmla v19.4s, v2.8h, v26.8h\n"
       "ldr q6, [x10, #0x0]\n"
@@ -980,25 +978,25 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "bge 89b\n"
       "90:"  // Height 3: Multiply loop: Single iteration only
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "sub x27, x27, #0x4\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
@@ -1029,13 +1027,13 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q26, [x10, #0x0]\n"
       "ldr q25, [x10, #0x10]\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      ".inst 0x6e5aec50  // bfmmla v16.4s, v2.8h, v26.8h\n"
-      ".inst 0x6e59ec54  // bfmmla v20.4s, v2.8h, v25.8h\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       ".inst 0x6e5aec08  // bfmmla v8.4s, v0.8h, v26.8h\n"
+      ".inst 0x6e5aec50  // bfmmla v16.4s, v2.8h, v26.8h\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e59ec0c  // bfmmla v12.4s, v0.8h, v25.8h\n"
+      ".inst 0x6e59ec54  // bfmmla v20.4s, v2.8h, v25.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
       ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
@@ -1060,27 +1058,27 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 86b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v6.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp1 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v18.2d, v18.2d, v22.2d\n"
       "uzp1 v19.2d, v19.2d, v23.2d\n"
       "tbz %x[flags], #1, 95f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v26.4s }, [x21]\n"
       "ld1r { v25.4s }, [x20]\n"
       "fmin v6.4s, v6.4s, v26.4s\n"
       "fmin v12.4s, v12.4s, v26.4s\n"
@@ -1112,79 +1110,79 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "tbz x11, #3, 99f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
-      "st1 { v17.4s }, [x25], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v17.4s }, [x24], #0x10\n"
       "tbz x11, #2, 97f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v18.4s }, [x25], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v18.4s }, [x24], #0x10\n"
       "tbz x11, #1, 96f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d19, [x25], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d19, [x24], #0x8\n"
       "tbz x11, #0, 103f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v19.s }[2], [x25]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v19.s }[2], [x24]\n"
       "b 103f\n"
       "96:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x11, #0, 103f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s19, [x25, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s19, [x24, #0x0]\n"
       "b 103f\n"
       "97:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x11, #1, 98f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d18, [x25], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d18, [x24], #0x8\n"
       "tbz x11, #0, 103f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v18.s }[2], [x25]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v18.s }[2], [x24]\n"
       "b 103f\n"
       "98:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x11, #0, 103f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s18, [x25, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s18, [x24, #0x0]\n"
       "b 103f\n"
       "99:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x11, #2, 101f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v16.4s }, [x25], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v16.4s }, [x24], #0x10\n"
       "tbz x11, #1, 100f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d17, [x25], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d17, [x24], #0x8\n"
       "tbz x11, #0, 103f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v17.s }[2], [x25]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v17.s }[2], [x24]\n"
       "b 103f\n"
       "100:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x11, #0, 103f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s17, [x25, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s17, [x24, #0x0]\n"
       "b 103f\n"
       "101:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x11, #1, 102f\n"
       "str d6, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
       "tbz x11, #0, 103f\n"
       "st1 { v6.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v16.s }[2], [x25]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v16.s }[2], [x24]\n"
       "b 103f\n"
       "102:"  // Height 3: Partial direct writeback: partial_1_0
       "str s6, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s16, [x25, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s16, [x24, #0x0]\n"
       "103:"  // Height 3: Partial direct writeback: Done
       "b 105f\n"
       "104:"  // Height 3: Full writeback
@@ -1193,36 +1191,36 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q17, [x25, #0x10]\n"
-      "str q18, [x25, #0x20]\n"
-      "str q19, [x25, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q17, [x24, #0x10]\n"
+      "str q18, [x24, #0x20]\n"
+      "str q19, [x24, #0x30]\n"
       "105:"  // Height 3: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 72b\n"
       "b 212f\n"
       "106:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "107:"  // Height 4: Column loop
       "cbz x12, 108f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -1237,111 +1235,111 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "108:"  // Height 4: no bias
       "tbz %x[flags], #0, 119f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x23, x24, x20, LSL #2\n"
       "bge 117f\n"
       "tbz x11, #3, 112f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
       "tbz x11, #2, 110f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
       "tbz x11, #1, 109f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
       "tbz x11, #0, 116f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
       "b 116f\n"
       "109:"  // Height 4: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 116f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
       "b 116f\n"
       "110:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x11, #1, 111f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
       "tbz x11, #0, 116f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
       "b 116f\n"
       "111:"  // Height 4: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 116f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
       "b 116f\n"
       "112:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x11, #2, 114f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
       "tbz x11, #1, 113f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
       "tbz x11, #0, 116f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
       "b 116f\n"
       "113:"  // Height 4: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 116f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
       "b 116f\n"
       "114:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x11, #1, 115f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
       "tbz x11, #0, 116f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
       "b 116f\n"
       "115:"  // Height 4: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
       "116:"  // Height 4: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 118f\n"
@@ -1350,18 +1348,18 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
       "118:"  // Height 4: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -1401,8 +1399,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "121:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 122f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1437,28 +1435,28 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "cmp x27, #0x8\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      "ld1 { v1.4s }, [x25], #0x10\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
-      "ld1 { v3.4s }, [x23], #0x10\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
       ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
       ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e59ec55  // bfmmla v21.4s, v2.8h, v25.8h\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e5aec0a  // bfmmla v10.4s, v0.8h, v26.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
       ".inst 0x6e5aec52  // bfmmla v18.4s, v2.8h, v26.8h\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
@@ -1479,18 +1477,18 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x27, x27, #0x4\n"
       "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e5aec09  // bfmmla v9.4s, v0.8h, v26.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e5aec51  // bfmmla v17.4s, v2.8h, v26.8h\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e59ec0d  // bfmmla v13.4s, v0.8h, v25.8h\n"
@@ -1551,8 +1549,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       ".inst 0x6e59ec0e  // bfmmla v14.4s, v0.8h, v25.8h\n"
       ".inst 0x6e59ec56  // bfmmla v22.4s, v2.8h, v25.8h\n"
       "ldr q25, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e5aec0b  // bfmmla v11.4s, v0.8h, v26.8h\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x6e5aec53  // bfmmla v19.4s, v2.8h, v26.8h\n"
       ".inst 0x6e59ec0f  // bfmmla v15.4s, v0.8h, v25.8h\n"
       ".inst 0x6e59ec57  // bfmmla v23.4s, v2.8h, v25.8h\n"
@@ -1562,23 +1560,23 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 121b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v6.2d, v8.2d, v12.2d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
@@ -1586,9 +1584,9 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
       "tbz %x[flags], #1, 130f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v26.4s }, [x21]\n"
       "ld1r { v25.4s }, [x20]\n"
       "fmin v6.4s, v6.4s, v26.4s\n"
       "fmin v12.4s, v12.4s, v26.4s\n"
@@ -1628,95 +1626,95 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "tbz x11, #3, 134f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
       "tbz x11, #2, 132f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
       "tbz x11, #1, 131f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
       "tbz x11, #0, 138f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
       "b 138f\n"
       "131:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x11, #0, 138f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
       "b 138f\n"
       "132:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x11, #1, 133f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
       "tbz x11, #0, 138f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
       "b 138f\n"
       "133:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x11, #0, 138f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
       "b 138f\n"
       "134:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x11, #2, 136f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
       "tbz x11, #1, 135f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
       "tbz x11, #0, 138f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
       "b 138f\n"
       "135:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x11, #0, 138f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
       "b 138f\n"
       "136:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x11, #1, 137f\n"
       "str d6, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
       "tbz x11, #0, 138f\n"
       "st1 { v6.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
       "b 138f\n"
       "137:"  // Height 4: Partial direct writeback: partial_1_0
       "str s6, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
       "138:"  // Height 4: Partial direct writeback: Done
       "b 140f\n"
       "139:"  // Height 4: Full writeback
@@ -1725,40 +1723,40 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
       "140:"  // Height 4: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 107b\n"
       "b 212f\n"
       "141:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "142:"  // Height 5: Column loop
       "cbz x12, 143f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -1781,128 +1779,128 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "143:"  // Height 5: no bias
       "tbz %x[flags], #0, 154f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x22, x23, x20, LSL #2\n"
       "bge 152f\n"
       "tbz x11, #3, 147f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
       "tbz x11, #2, 145f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v27.4s }, [x23], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v27.4s }, [x22], #0x10\n"
       "tbz x11, #1, 144f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d6, [x23], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d6, [x22], #0x8\n"
       "tbz x11, #0, 151f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v6.s }[2], [x23]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v6.s }[2], [x22]\n"
       "b 151f\n"
       "144:"  // Height 5: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 151f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s6, [x23, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s6, [x22, #0x0]\n"
       "b 151f\n"
       "145:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x11, #1, 146f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
       "tbz x11, #0, 151f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
       "b 151f\n"
       "146:"  // Height 5: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 151f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
       "b 151f\n"
       "147:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x11, #2, 149f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
       "tbz x11, #1, 148f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
       "tbz x11, #0, 151f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
       "b 151f\n"
       "148:"  // Height 5: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 151f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
       "b 151f\n"
       "149:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x11, #1, 150f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
       "tbz x11, #0, 151f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
       "b 151f\n"
       "150:"  // Height 5: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
       "151:"  // Height 5: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 153f\n"
@@ -1911,22 +1909,22 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q25, [x23, #0x0]\n"
-      "ldr q26, [x23, #0x10]\n"
-      "ldr q27, [x23, #0x20]\n"
-      "ldr q6, [x23, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q25, [x22, #0x0]\n"
+      "ldr q26, [x22, #0x10]\n"
+      "ldr q27, [x22, #0x20]\n"
+      "ldr q6, [x22, #0x30]\n"
       "153:"  // Height 5: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -1982,8 +1980,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "156:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 157f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2022,31 +2020,31 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
       "cmp x27, #0x8\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      "ld1 { v1.4s }, [x25], #0x10\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
-      "ld1 { v3.4s }, [x23], #0x10\n"
-      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
-      ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
       ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
-      "ldr q6, [x10, #0x20]\n"
+      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
+      "ldr q3, [x10, #0x20]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
       ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
+      ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
       "ldr q5, [x10, #0x30]\n"
-      ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e46ec51  // bfmmla v17.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e46ec99  // bfmmla v25.4s, v4.8h, v6.8h\n"
+      ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      ".inst 0x6e43ec51  // bfmmla v17.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e43ec99  // bfmmla v25.4s, v4.8h, v3.8h\n"
       "ldr q6, [x10, #0x40]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
       ".inst 0x6e45ec0d  // bfmmla v13.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec55  // bfmmla v21.4s, v2.8h, v5.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e45ec9d  // bfmmla v29.4s, v4.8h, v5.8h\n"
       "ldr q5, [x10, #0x50]\n"
       ".inst 0x6e46ec0a  // bfmmla v10.4s, v0.8h, v6.8h\n"
@@ -2075,22 +2073,22 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x27, x27, #0x4\n"
       "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
-      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
-      ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
       ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
       "ldr q3, [x10, #0x20]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
       ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
       "ldr q1, [x10, #0x30]\n"
       ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e43ec51  // bfmmla v17.4s, v2.8h, v3.8h\n"
       ".inst 0x6e43ec99  // bfmmla v25.4s, v4.8h, v3.8h\n"
       "ldr q3, [x10, #0x40]\n"
@@ -2140,16 +2138,16 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q5, [x10, #0x10]\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
       ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
-      ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
-      ".inst 0x6e45ec9c  // bfmmla v28.4s, v4.8h, v5.8h\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
       ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
+      ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
       "ldr q3, [x10, #0x20]\n"
+      ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
       ".inst 0x6e45ec54  // bfmmla v20.4s, v2.8h, v5.8h\n"
+      ".inst 0x6e45ec9c  // bfmmla v28.4s, v4.8h, v5.8h\n"
       "ldr q1, [x10, #0x30]\n"
       ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
       ".inst 0x6e43ec51  // bfmmla v17.4s, v2.8h, v3.8h\n"
@@ -2167,8 +2165,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       ".inst 0x6e41ec56  // bfmmla v22.4s, v2.8h, v1.8h\n"
       ".inst 0x6e41ec9e  // bfmmla v30.4s, v4.8h, v1.8h\n"
       "ldr q1, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x6e43ec53  // bfmmla v19.4s, v2.8h, v3.8h\n"
       ".inst 0x6e43ec9b  // bfmmla v27.4s, v4.8h, v3.8h\n"
       ".inst 0x6e41ec0f  // bfmmla v15.4s, v0.8h, v1.8h\n"
@@ -2180,27 +2178,27 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 156b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v6.2d, v8.2d, v12.2d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
@@ -2210,9 +2208,9 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "uzp1 v26.2d, v26.2d, v30.2d\n"
       "uzp1 v27.2d, v27.2d, v31.2d\n"
       "tbz %x[flags], #1, 165f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v6.4s, v6.4s, v1.4s\n"
       "fmin v12.4s, v12.4s, v1.4s\n"
@@ -2260,111 +2258,111 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "tbz x11, #3, 169f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
-      "st1 { v25.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v25.4s }, [x22], #0x10\n"
       "tbz x11, #2, 167f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v26.4s }, [x23], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v26.4s }, [x22], #0x10\n"
       "tbz x11, #1, 166f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d27, [x23], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d27, [x22], #0x8\n"
       "tbz x11, #0, 173f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v27.s }[2], [x23]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v27.s }[2], [x22]\n"
       "b 173f\n"
       "166:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x11, #0, 173f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s27, [x23, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s27, [x22, #0x0]\n"
       "b 173f\n"
       "167:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x11, #1, 168f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d26, [x23], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d26, [x22], #0x8\n"
       "tbz x11, #0, 173f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v26.s }[2], [x23]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v26.s }[2], [x22]\n"
       "b 173f\n"
       "168:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x11, #0, 173f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s26, [x23, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s26, [x22, #0x0]\n"
       "b 173f\n"
       "169:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x11, #2, 171f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v24.4s }, [x23], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v24.4s }, [x22], #0x10\n"
       "tbz x11, #1, 170f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d25, [x23], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d25, [x22], #0x8\n"
       "tbz x11, #0, 173f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v25.s }[2], [x23]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v25.s }[2], [x22]\n"
       "b 173f\n"
       "170:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x11, #0, 173f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s25, [x23, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s25, [x22, #0x0]\n"
       "b 173f\n"
       "171:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x11, #1, 172f\n"
       "str d6, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x11, #0, 173f\n"
       "st1 { v6.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v24.s }[2], [x23]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v24.s }[2], [x22]\n"
       "b 173f\n"
       "172:"  // Height 5: Partial direct writeback: partial_1_0
       "str s6, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s24, [x23, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s24, [x22, #0x0]\n"
       "173:"  // Height 5: Partial direct writeback: Done
       "b 175f\n"
       "174:"  // Height 5: Full writeback
@@ -2373,48 +2371,47 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q25, [x23, #0x10]\n"
-      "str q26, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q25, [x22, #0x10]\n"
+      "str q26, [x22, #0x20]\n"
+      "str q27, [x22, #0x30]\n"
       "175:"  // Height 5: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 142b\n"
       "b 212f\n"
       "176:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "177:"  // Height 6: Column loop
       "cbz x12, 178f\n"
       "ldr q8, [x12, #0x0]\n"
       "ldr q9, [x12, #0x10]\n"
-      "ldr q10, [x12, #0x20]\n"
-      "ldr q11, [x12, #0x30]\n"
-      "add x12, x12, #0x40\n"
       "zip2 v12.2d, v8.2d, v8.2d\n"
       "zip1 v8.2d, v8.2d, v8.2d\n"
+      "ldr q10, [x12, #0x20]\n"
+      "ldr q11, [x12, #0x30]\n"
       "zip2 v13.2d, v9.2d, v9.2d\n"
       "zip1 v9.2d, v9.2d, v9.2d\n"
       "zip2 v14.2d, v10.2d, v10.2d\n"
       "zip1 v10.2d, v10.2d, v10.2d\n"
+      "add x12, x12, #0x40\n"
       "zip2 v15.2d, v11.2d, v11.2d\n"
       "zip1 v11.2d, v11.2d, v11.2d\n"
       "mov v16.16b, v8.16b\n"
@@ -2437,145 +2434,145 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "178:"  // Height 6: no bias
       "tbz %x[flags], #0, 189f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "add x21, x22, x20, LSL #2\n"
       "bge 187f\n"
       "tbz x11, #3, 182f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x26], #0x10\n"
-      "ld1 { v18.4s }, [x25], #0x10\n"
-      "ld1 { v21.4s }, [x24], #0x10\n"
-      "ld1 { v26.4s }, [x23], #0x10\n"
-      "ld1 { v29.4s }, [x22], #0x10\n"
+      "ld1 { v13.4s }, [x25], #0x10\n"
+      "ld1 { v18.4s }, [x24], #0x10\n"
+      "ld1 { v21.4s }, [x23], #0x10\n"
+      "ld1 { v26.4s }, [x22], #0x10\n"
+      "ld1 { v29.4s }, [x21], #0x10\n"
       "tbz x11, #2, 180f\n"
       "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x26], #0x10\n"
-      "ld1 { v19.4s }, [x25], #0x10\n"
-      "ld1 { v22.4s }, [x24], #0x10\n"
-      "ld1 { v27.4s }, [x23], #0x10\n"
-      "ld1 { v30.4s }, [x22], #0x10\n"
+      "ld1 { v14.4s }, [x25], #0x10\n"
+      "ld1 { v19.4s }, [x24], #0x10\n"
+      "ld1 { v22.4s }, [x23], #0x10\n"
+      "ld1 { v27.4s }, [x22], #0x10\n"
+      "ld1 { v30.4s }, [x21], #0x10\n"
       "tbz x11, #1, 179f\n"
       "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x26], #0x8\n"
+      "ldr d15, [x25], #0x8\n"
       "mov x20, #0x38\n"
-      "ldr d24, [x25], #0x8\n"
-      "ldr d23, [x24], #0x8\n"
-      "ldr d6, [x23], #0x8\n"
-      "ldr d31, [x22], #0x8\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr d23, [x23], #0x8\n"
+      "ldr d6, [x22], #0x8\n"
+      "ldr d31, [x21], #0x8\n"
       "tbz x11, #0, 186f\n"
       "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x26]\n"
-      "ld1 { v24.s }[2], [x25]\n"
-      "ld1 { v23.s }[2], [x24]\n"
-      "ld1 { v6.s }[2], [x23]\n"
-      "ld1 { v31.s }[2], [x22]\n"
+      "ld1 { v15.s }[2], [x25]\n"
+      "ld1 { v24.s }[2], [x24]\n"
+      "ld1 { v23.s }[2], [x23]\n"
+      "ld1 { v6.s }[2], [x22]\n"
+      "ld1 { v31.s }[2], [x21]\n"
       "b 186f\n"
       "179:"  // Height 6: Partial accumulate: partial_1_12
       "mov x20, #0x30\n"
       "tbz x11, #0, 186f\n"
       "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x26, #0x0]\n"
-      "ldr s24, [x25, #0x0]\n"
-      "ldr s23, [x24, #0x0]\n"
-      "ldr s6, [x23, #0x0]\n"
-      "ldr s31, [x22, #0x0]\n"
+      "ldr s15, [x25, #0x0]\n"
+      "ldr s24, [x24, #0x0]\n"
+      "ldr s23, [x23, #0x0]\n"
+      "ldr s6, [x22, #0x0]\n"
+      "ldr s31, [x21, #0x0]\n"
       "b 186f\n"
       "180:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x11, #1, 181f\n"
       "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x26], #0x8\n"
+      "ldr d14, [x25], #0x8\n"
       "mov x20, #0x28\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d22, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "ldr d30, [x22], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d22, [x23], #0x8\n"
+      "ldr d27, [x22], #0x8\n"
+      "ldr d30, [x21], #0x8\n"
       "tbz x11, #0, 186f\n"
       "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x26]\n"
-      "ld1 { v19.s }[2], [x25]\n"
-      "ld1 { v22.s }[2], [x24]\n"
-      "ld1 { v27.s }[2], [x23]\n"
-      "ld1 { v30.s }[2], [x22]\n"
+      "ld1 { v14.s }[2], [x25]\n"
+      "ld1 { v19.s }[2], [x24]\n"
+      "ld1 { v22.s }[2], [x23]\n"
+      "ld1 { v27.s }[2], [x22]\n"
+      "ld1 { v30.s }[2], [x21]\n"
       "b 186f\n"
       "181:"  // Height 6: Partial accumulate: partial_1_8
       "mov x20, #0x20\n"
       "tbz x11, #0, 186f\n"
       "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x26, #0x0]\n"
-      "ldr s19, [x25, #0x0]\n"
-      "ldr s22, [x24, #0x0]\n"
-      "ldr s27, [x23, #0x0]\n"
-      "ldr s30, [x22, #0x0]\n"
+      "ldr s14, [x25, #0x0]\n"
+      "ldr s19, [x24, #0x0]\n"
+      "ldr s22, [x23, #0x0]\n"
+      "ldr s27, [x22, #0x0]\n"
+      "ldr s30, [x21, #0x0]\n"
       "b 186f\n"
       "182:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x11, #2, 184f\n"
       "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x26], #0x10\n"
-      "ld1 { v17.4s }, [x25], #0x10\n"
-      "ld1 { v20.4s }, [x24], #0x10\n"
-      "ld1 { v25.4s }, [x23], #0x10\n"
-      "ld1 { v28.4s }, [x22], #0x10\n"
+      "ld1 { v12.4s }, [x25], #0x10\n"
+      "ld1 { v17.4s }, [x24], #0x10\n"
+      "ld1 { v20.4s }, [x23], #0x10\n"
+      "ld1 { v25.4s }, [x22], #0x10\n"
+      "ld1 { v28.4s }, [x21], #0x10\n"
       "tbz x11, #1, 183f\n"
       "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x26], #0x8\n"
+      "ldr d13, [x25], #0x8\n"
       "mov x20, #0x18\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d21, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
-      "ldr d29, [x22], #0x8\n"
+      "ldr d18, [x24], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "ldr d26, [x22], #0x8\n"
+      "ldr d29, [x21], #0x8\n"
       "tbz x11, #0, 186f\n"
       "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x26]\n"
-      "ld1 { v18.s }[2], [x25]\n"
-      "ld1 { v21.s }[2], [x24]\n"
-      "ld1 { v26.s }[2], [x23]\n"
-      "ld1 { v29.s }[2], [x22]\n"
+      "ld1 { v13.s }[2], [x25]\n"
+      "ld1 { v18.s }[2], [x24]\n"
+      "ld1 { v21.s }[2], [x23]\n"
+      "ld1 { v26.s }[2], [x22]\n"
+      "ld1 { v29.s }[2], [x21]\n"
       "b 186f\n"
       "183:"  // Height 6: Partial accumulate: partial_1_4
       "mov x20, #0x10\n"
       "tbz x11, #0, 186f\n"
       "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x26, #0x0]\n"
-      "ldr s18, [x25, #0x0]\n"
-      "ldr s21, [x24, #0x0]\n"
-      "ldr s26, [x23, #0x0]\n"
-      "ldr s29, [x22, #0x0]\n"
+      "ldr s13, [x25, #0x0]\n"
+      "ldr s18, [x24, #0x0]\n"
+      "ldr s21, [x23, #0x0]\n"
+      "ldr s26, [x22, #0x0]\n"
+      "ldr s29, [x21, #0x0]\n"
       "b 186f\n"
       "184:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x11, #1, 185f\n"
       "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x26], #0x8\n"
+      "ldr d12, [x25], #0x8\n"
       "mov x20, #0x8\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "ldr d25, [x23], #0x8\n"
-      "ldr d28, [x22], #0x8\n"
+      "ldr d17, [x24], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "ldr d25, [x22], #0x8\n"
+      "ldr d28, [x21], #0x8\n"
       "tbz x11, #0, 186f\n"
       "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x26]\n"
-      "ld1 { v17.s }[2], [x25]\n"
-      "ld1 { v20.s }[2], [x24]\n"
-      "ld1 { v25.s }[2], [x23]\n"
-      "ld1 { v28.s }[2], [x22]\n"
+      "ld1 { v12.s }[2], [x25]\n"
+      "ld1 { v17.s }[2], [x24]\n"
+      "ld1 { v20.s }[2], [x23]\n"
+      "ld1 { v25.s }[2], [x22]\n"
+      "ld1 { v28.s }[2], [x21]\n"
       "b 186f\n"
       "185:"  // Height 6: Partial accumulate: partial_1_0
       "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x26, #0x0]\n"
+      "ldr s12, [x25, #0x0]\n"
       "mov x20, #0x0\n"
-      "ldr s17, [x25, #0x0]\n"
-      "ldr s20, [x24, #0x0]\n"
-      "ldr s25, [x23, #0x0]\n"
-      "ldr s28, [x22, #0x0]\n"
+      "ldr s17, [x24, #0x0]\n"
+      "ldr s20, [x23, #0x0]\n"
+      "ldr s25, [x22, #0x0]\n"
+      "ldr s28, [x21, #0x0]\n"
       "186:"  // Height 6: Partial accumulate: Done
       "sub x9, x9, x20\n"
       "b 188f\n"
@@ -2584,26 +2581,26 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "ldr q10, [x9, #0x10]\n"
       "ldr q11, [x9, #0x20]\n"
       "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x26, #0x0]\n"
-      "ldr q13, [x26, #0x10]\n"
-      "ldr q14, [x26, #0x20]\n"
-      "ldr q15, [x26, #0x30]\n"
-      "ldr q17, [x25, #0x0]\n"
-      "ldr q18, [x25, #0x10]\n"
-      "ldr q19, [x25, #0x20]\n"
-      "ldr q24, [x25, #0x30]\n"
-      "ldr q20, [x24, #0x0]\n"
-      "ldr q21, [x24, #0x10]\n"
-      "ldr q22, [x24, #0x20]\n"
-      "ldr q23, [x24, #0x30]\n"
-      "ldr q25, [x23, #0x0]\n"
-      "ldr q26, [x23, #0x10]\n"
-      "ldr q27, [x23, #0x20]\n"
-      "ldr q6, [x23, #0x30]\n"
-      "ldr q28, [x22, #0x0]\n"
-      "ldr q29, [x22, #0x10]\n"
-      "ldr q30, [x22, #0x20]\n"
-      "ldr q31, [x22, #0x30]\n"
+      "ldr q12, [x25, #0x0]\n"
+      "ldr q13, [x25, #0x10]\n"
+      "ldr q14, [x25, #0x20]\n"
+      "ldr q15, [x25, #0x30]\n"
+      "ldr q17, [x24, #0x0]\n"
+      "ldr q18, [x24, #0x10]\n"
+      "ldr q19, [x24, #0x20]\n"
+      "ldr q24, [x24, #0x30]\n"
+      "ldr q20, [x23, #0x0]\n"
+      "ldr q21, [x23, #0x10]\n"
+      "ldr q22, [x23, #0x20]\n"
+      "ldr q23, [x23, #0x30]\n"
+      "ldr q25, [x22, #0x0]\n"
+      "ldr q26, [x22, #0x10]\n"
+      "ldr q27, [x22, #0x20]\n"
+      "ldr q6, [x22, #0x30]\n"
+      "ldr q28, [x21, #0x0]\n"
+      "ldr q29, [x21, #0x10]\n"
+      "ldr q30, [x21, #0x20]\n"
+      "ldr q31, [x21, #0x30]\n"
       "188:"  // Height 6: MMLA fixup
       "zip1 v8.2d, v9.2d, v12.2d\n"
       "zip2 v12.2d, v9.2d, v12.2d\n"
@@ -2659,8 +2656,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "mov x28, #0x0\n"
       "191:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 192f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2703,36 +2700,36 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       "sub x27, x27, #0x4\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
       "cmp x27, #0x8\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      "ld1 { v1.4s }, [x25], #0x10\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
-      "ld1 { v3.4s }, [x23], #0x10\n"
       ".inst 0x4ea168a4  // bfcvtn2 v4.8h, v5.4s\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      "ld1 { v5.4s }, [x21], #0x10\n"
-      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
       ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "ld1 { v1.4s }, [x25], #0x10\n"
       ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x20]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
-      "ldr q7, [x10, #0x30]\n"
+      "ldr q5, [x10, #0x30]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e46ec09  // bfmmla v9.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec51  // bfmmla v17.4s, v2.8h, v6.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      "ld1 { v3.4s }, [x23], #0x10\n"
       ".inst 0x6e46ec99  // bfmmla v25.4s, v4.8h, v6.8h\n"
       "ldr q6, [x10, #0x40]\n"
-      ".inst 0x6e47ec0d  // bfmmla v13.4s, v0.8h, v7.8h\n"
-      ".inst 0x6e47ec55  // bfmmla v21.4s, v2.8h, v7.8h\n"
-      ".inst 0x6e47ec9d  // bfmmla v29.4s, v4.8h, v7.8h\n"
+      ".inst 0x6e45ec0d  // bfmmla v13.4s, v0.8h, v5.8h\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
+      ".inst 0x6e45ec55  // bfmmla v21.4s, v2.8h, v5.8h\n"
+      ".inst 0x6e45ec9d  // bfmmla v29.4s, v4.8h, v5.8h\n"
       "ldr q7, [x10, #0x50]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
+      "ld1 { v5.4s }, [x21], #0x10\n"
       ".inst 0x6e46ec0a  // bfmmla v10.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec52  // bfmmla v18.4s, v2.8h, v6.8h\n"
       ".inst 0x6e46ec9a  // bfmmla v26.4s, v4.8h, v6.8h\n"
@@ -2760,24 +2757,24 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "sub x27, x27, #0x4\n"
       "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4ea168a4  // bfcvtn2 v4.8h, v5.4s\n"
-      ".inst 0x6e46ec08  // bfmmla v8.4s, v0.8h, v6.8h\n"
-      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
       ".inst 0x6e46ec50  // bfmmla v16.4s, v2.8h, v6.8h\n"
-      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e46ec98  // bfmmla v24.4s, v4.8h, v6.8h\n"
       "ldr q3, [x10, #0x20]\n"
+      ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec9c  // bfmmla v28.4s, v4.8h, v7.8h\n"
       "ldr q1, [x10, #0x30]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
       ".inst 0x6e43ec51  // bfmmla v17.4s, v2.8h, v3.8h\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e43ec99  // bfmmla v25.4s, v4.8h, v3.8h\n"
       "ldr q3, [x10, #0x40]\n"
       ".inst 0x6e41ec0d  // bfmmla v13.4s, v0.8h, v1.8h\n"
@@ -2831,13 +2828,13 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
       ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
       ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
+      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
       ".inst 0x4ea16862  // bfcvtn2 v2.8h, v3.4s\n"
       ".inst 0x4ea168a4  // bfcvtn2 v4.8h, v5.4s\n"
-      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
-      ".inst 0x6e46ec0c  // bfmmla v12.4s, v0.8h, v6.8h\n"
       ".inst 0x6e47ec50  // bfmmla v16.4s, v2.8h, v7.8h\n"
       ".inst 0x6e47ec98  // bfmmla v24.4s, v4.8h, v7.8h\n"
       "ldr q3, [x10, #0x20]\n"
+      ".inst 0x6e46ec0c  // bfmmla v12.4s, v0.8h, v6.8h\n"
       ".inst 0x6e46ec54  // bfmmla v20.4s, v2.8h, v6.8h\n"
       ".inst 0x6e46ec9c  // bfmmla v28.4s, v4.8h, v6.8h\n"
       "ldr q1, [x10, #0x30]\n"
@@ -2870,31 +2867,31 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 191b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 v6.2d, v8.2d, v12.2d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
       "uzp1 v23.2d, v24.2d, v28.2d\n"
@@ -2906,9 +2903,9 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "uzp1 v30.2d, v27.2d, v31.2d\n"
       "uzp2 v27.2d, v27.2d, v31.2d\n"
       "tbz %x[flags], #1, 200f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1r { v1.4s }, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1r { v1.4s }, [x21]\n"
       "ld1r { v0.4s }, [x20]\n"
       "fmin v6.4s, v6.4s, v1.4s\n"
       "fmin v12.4s, v12.4s, v1.4s\n"
@@ -2964,127 +2961,127 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "tbz x11, #3, 204f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
       "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v9.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v20.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v17.4s }, [x24], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
-      "st1 { v28.4s }, [x23], #0x10\n"
-      "st1 { v24.4s }, [x22], #0x10\n"
-      "st1 { v25.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v9.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v20.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v17.4s }, [x23], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v28.4s }, [x22], #0x10\n"
+      "st1 { v24.4s }, [x21], #0x10\n"
+      "st1 { v25.4s }, [x21], #0x10\n"
       "tbz x11, #2, 202f\n"
       "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x26], #0x10\n"
-      "st1 { v21.4s }, [x25], #0x10\n"
-      "st1 { v18.4s }, [x24], #0x10\n"
-      "st1 { v29.4s }, [x23], #0x10\n"
-      "st1 { v26.4s }, [x22], #0x10\n"
+      "st1 { v10.4s }, [x25], #0x10\n"
+      "st1 { v21.4s }, [x24], #0x10\n"
+      "st1 { v18.4s }, [x23], #0x10\n"
+      "st1 { v29.4s }, [x22], #0x10\n"
+      "st1 { v26.4s }, [x21], #0x10\n"
       "tbz x11, #1, 201f\n"
       "str d14, [x9], #0x8\n"
-      "str d11, [x26], #0x8\n"
-      "str d22, [x25], #0x8\n"
-      "str d19, [x24], #0x8\n"
-      "str d30, [x23], #0x8\n"
-      "str d27, [x22], #0x8\n"
+      "str d11, [x25], #0x8\n"
+      "str d22, [x24], #0x8\n"
+      "str d19, [x23], #0x8\n"
+      "str d30, [x22], #0x8\n"
+      "str d27, [x21], #0x8\n"
       "tbz x11, #0, 208f\n"
       "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x26]\n"
-      "st1 { v22.s }[2], [x25]\n"
-      "st1 { v19.s }[2], [x24]\n"
-      "st1 { v30.s }[2], [x23]\n"
-      "st1 { v27.s }[2], [x22]\n"
+      "st1 { v11.s }[2], [x25]\n"
+      "st1 { v22.s }[2], [x24]\n"
+      "st1 { v19.s }[2], [x23]\n"
+      "st1 { v30.s }[2], [x22]\n"
+      "st1 { v27.s }[2], [x21]\n"
       "b 208f\n"
       "201:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x11, #0, 208f\n"
       "str s14, [x9, #0x0]\n"
-      "str s11, [x26, #0x0]\n"
-      "str s22, [x25, #0x0]\n"
-      "str s19, [x24, #0x0]\n"
-      "str s30, [x23, #0x0]\n"
-      "str s27, [x22, #0x0]\n"
+      "str s11, [x25, #0x0]\n"
+      "str s22, [x24, #0x0]\n"
+      "str s19, [x23, #0x0]\n"
+      "str s30, [x22, #0x0]\n"
+      "str s27, [x21, #0x0]\n"
       "b 208f\n"
       "202:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x11, #1, 203f\n"
       "str d13, [x9], #0x8\n"
-      "str d10, [x26], #0x8\n"
-      "str d21, [x25], #0x8\n"
-      "str d18, [x24], #0x8\n"
-      "str d29, [x23], #0x8\n"
-      "str d26, [x22], #0x8\n"
+      "str d10, [x25], #0x8\n"
+      "str d21, [x24], #0x8\n"
+      "str d18, [x23], #0x8\n"
+      "str d29, [x22], #0x8\n"
+      "str d26, [x21], #0x8\n"
       "tbz x11, #0, 208f\n"
       "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x26]\n"
-      "st1 { v21.s }[2], [x25]\n"
-      "st1 { v18.s }[2], [x24]\n"
-      "st1 { v29.s }[2], [x23]\n"
-      "st1 { v26.s }[2], [x22]\n"
+      "st1 { v10.s }[2], [x25]\n"
+      "st1 { v21.s }[2], [x24]\n"
+      "st1 { v18.s }[2], [x23]\n"
+      "st1 { v29.s }[2], [x22]\n"
+      "st1 { v26.s }[2], [x21]\n"
       "b 208f\n"
       "203:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x11, #0, 208f\n"
       "str s13, [x9, #0x0]\n"
-      "str s10, [x26, #0x0]\n"
-      "str s21, [x25, #0x0]\n"
-      "str s18, [x24, #0x0]\n"
-      "str s29, [x23, #0x0]\n"
-      "str s26, [x22, #0x0]\n"
+      "str s10, [x25, #0x0]\n"
+      "str s21, [x24, #0x0]\n"
+      "str s18, [x23, #0x0]\n"
+      "str s29, [x22, #0x0]\n"
+      "str s26, [x21, #0x0]\n"
       "b 208f\n"
       "204:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x11, #2, 206f\n"
       "st1 { v6.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x26], #0x10\n"
-      "st1 { v15.4s }, [x25], #0x10\n"
-      "st1 { v16.4s }, [x24], #0x10\n"
-      "st1 { v23.4s }, [x23], #0x10\n"
-      "st1 { v24.4s }, [x22], #0x10\n"
+      "st1 { v8.4s }, [x25], #0x10\n"
+      "st1 { v15.4s }, [x24], #0x10\n"
+      "st1 { v16.4s }, [x23], #0x10\n"
+      "st1 { v23.4s }, [x22], #0x10\n"
+      "st1 { v24.4s }, [x21], #0x10\n"
       "tbz x11, #1, 205f\n"
       "str d12, [x9], #0x8\n"
-      "str d9, [x26], #0x8\n"
-      "str d20, [x25], #0x8\n"
-      "str d17, [x24], #0x8\n"
-      "str d28, [x23], #0x8\n"
-      "str d25, [x22], #0x8\n"
+      "str d9, [x25], #0x8\n"
+      "str d20, [x24], #0x8\n"
+      "str d17, [x23], #0x8\n"
+      "str d28, [x22], #0x8\n"
+      "str d25, [x21], #0x8\n"
       "tbz x11, #0, 208f\n"
       "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x26]\n"
-      "st1 { v20.s }[2], [x25]\n"
-      "st1 { v17.s }[2], [x24]\n"
-      "st1 { v28.s }[2], [x23]\n"
-      "st1 { v25.s }[2], [x22]\n"
+      "st1 { v9.s }[2], [x25]\n"
+      "st1 { v20.s }[2], [x24]\n"
+      "st1 { v17.s }[2], [x23]\n"
+      "st1 { v28.s }[2], [x22]\n"
+      "st1 { v25.s }[2], [x21]\n"
       "b 208f\n"
       "205:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x11, #0, 208f\n"
       "str s12, [x9, #0x0]\n"
-      "str s9, [x26, #0x0]\n"
-      "str s20, [x25, #0x0]\n"
-      "str s17, [x24, #0x0]\n"
-      "str s28, [x23, #0x0]\n"
-      "str s25, [x22, #0x0]\n"
+      "str s9, [x25, #0x0]\n"
+      "str s20, [x24, #0x0]\n"
+      "str s17, [x23, #0x0]\n"
+      "str s28, [x22, #0x0]\n"
+      "str s25, [x21, #0x0]\n"
       "b 208f\n"
       "206:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x11, #1, 207f\n"
       "str d6, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
+      "str d24, [x21], #0x8\n"
       "tbz x11, #0, 208f\n"
       "st1 { v6.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x26]\n"
-      "st1 { v15.s }[2], [x25]\n"
-      "st1 { v16.s }[2], [x24]\n"
-      "st1 { v23.s }[2], [x23]\n"
-      "st1 { v24.s }[2], [x22]\n"
+      "st1 { v8.s }[2], [x25]\n"
+      "st1 { v15.s }[2], [x24]\n"
+      "st1 { v16.s }[2], [x23]\n"
+      "st1 { v23.s }[2], [x22]\n"
+      "st1 { v24.s }[2], [x21]\n"
       "b 208f\n"
       "207:"  // Height 6: Partial direct writeback: partial_1_0
       "str s6, [x9, #0x0]\n"
-      "str s8, [x26, #0x0]\n"
-      "str s15, [x25, #0x0]\n"
-      "str s16, [x24, #0x0]\n"
-      "str s23, [x23, #0x0]\n"
-      "str s24, [x22, #0x0]\n"
+      "str s8, [x25, #0x0]\n"
+      "str s15, [x24, #0x0]\n"
+      "str s16, [x23, #0x0]\n"
+      "str s23, [x22, #0x0]\n"
+      "str s24, [x21, #0x0]\n"
       "208:"  // Height 6: Partial direct writeback: Done
       "b 210f\n"
       "209:"  // Height 6: Full writeback
@@ -3093,26 +3090,26 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "str q13, [x9, #0x20]\n"
       "str q14, [x9, #0x30]\n"
       "add x9, x9, #0x40\n"
-      "str q8, [x26, #0x0]\n"
-      "str q9, [x26, #0x10]\n"
-      "str q10, [x26, #0x20]\n"
-      "str q11, [x26, #0x30]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q20, [x25, #0x10]\n"
-      "str q21, [x25, #0x20]\n"
-      "str q22, [x25, #0x30]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q17, [x24, #0x10]\n"
-      "str q18, [x24, #0x20]\n"
-      "str q19, [x24, #0x30]\n"
-      "str q23, [x23, #0x0]\n"
-      "str q28, [x23, #0x10]\n"
-      "str q29, [x23, #0x20]\n"
-      "str q30, [x23, #0x30]\n"
-      "str q24, [x22, #0x0]\n"
-      "str q25, [x22, #0x10]\n"
-      "str q26, [x22, #0x20]\n"
-      "str q27, [x22, #0x30]\n"
+      "str q8, [x25, #0x0]\n"
+      "str q9, [x25, #0x10]\n"
+      "str q10, [x25, #0x20]\n"
+      "str q11, [x25, #0x30]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q20, [x24, #0x10]\n"
+      "str q21, [x24, #0x20]\n"
+      "str q22, [x24, #0x30]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q18, [x23, #0x20]\n"
+      "str q19, [x23, #0x30]\n"
+      "str q23, [x22, #0x0]\n"
+      "str q28, [x22, #0x10]\n"
+      "str q29, [x22, #0x20]\n"
+      "str q30, [x22, #0x30]\n"
+      "str q24, [x21, #0x0]\n"
+      "str q25, [x21, #0x10]\n"
+      "str q26, [x21, #0x20]\n"
+      "str q27, [x21, #0x30]\n"
       "210:"  // Height 6: Writeback done
       "subs x11, x11, #0x10\n"
       "bgt 177b\n"
@@ -3128,8 +3125,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "212:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16.hpp
index becd43516f..857de3c6d0 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16.hpp
@@ -71,7 +71,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 4, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp
index 4d23660942..6ef2fd5d34 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp
@@ -45,18 +45,18 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -84,133 +84,133 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "cmp %x[M], #0x2\n"
       "bgt 61f\n"
       "beq 31f\n"
-      "mov x16, %x[col_bias]\n"
+      "mov x15, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v15.16b, #0x1\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x13, %x[output_ptr]\n"
+      "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
       "movi v18.4s, #0x0\n"
       "movi v19.4s, #0x0\n"
       "3:"  // Height 1: setup done
-      "mov x12, #0x0\n"
+      "mov x11, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w10, [x20, x11, LSL #0x2]\n"
       "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w11, [x20, x12, LSL #0x2]\n"
       "tbz %x[flags], #3, 5f\n"
-      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
-      "ldr x10, [x20, #0x0]\n"
-      "cbnz x12, 6f\n"
+      "ldr x9, [x20, #0x0]\n"
+      "cbnz x11, 6f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x10, x10, x20\n"
+      "add x9, x9, x20\n"
       "b 6f\n"
       "5:"  // Height 1: setup direct input
-      "mov x10, %x[input_ptr]\n"
+      "mov x9, %x[input_ptr]\n"
       "6:"  // Height 1: input setup done
-      "cmp x11, #0x10\n"
+      "cmp x10, #0x10\n"
       "blt 11f\n"
-      "ldr q0, [x10, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "blt 9f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x4f80e090  // sdot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr d21, [x14, #0x70]\n"
-      "ldr x20, [x14, #0x78]\n"
+      "ldr d21, [x12, #0x70]\n"
+      "ldr x20, [x12, #0x78]\n"
       ".inst 0x4f80e0b1  // sdot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr d20, [x14, #0x80]\n"
+      "ldr d20, [x12, #0x80]\n"
       ".inst 0x4f80e0d2  // sdot v18.4s, v6.16b, v0.4b[0]\n"
-      "ldr d26, [x14, #0x90]\n"
+      "ldr d26, [x12, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr d25, [x14, #0xa0]\n"
+      "ldr d25, [x12, #0xa0]\n"
       "mov v21.d[1], x20\n"
-      "ldr x20, [x14, #0x88]\n"
+      "ldr x20, [x12, #0x88]\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
-      "ldr d24, [x14, #0xb0]\n"
+      "ldr d24, [x12, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
-      "ldr d23, [x14, #0xc0]\n"
+      "ldr d23, [x12, #0xc0]\n"
       ".inst 0x4fa0e152  // sdot v18.4s, v10.16b, v0.4b[1]\n"
-      "ldr d22, [x14, #0xd0]\n"
+      "ldr d22, [x12, #0xd0]\n"
       ".inst 0x4fa0e2b3  // sdot v19.4s, v21.16b, v0.4b[1]\n"
-      "ldr d21, [x14, #0xe0]\n"
+      "ldr d21, [x12, #0xe0]\n"
       "mov v20.d[1], x20\n"
-      "ldr x22, [x14, #0x98]\n"
-      "add x10, x10, #0x10\n"
-      "ldr x21, [x14, #0xa8]\n"
+      "ldr x20, [x12, #0x98]\n"
+      "mov v26.d[1], x20\n"
+      "ldr x20, [x12, #0xa8]\n"
+      "mov v25.d[1], x20\n"
+      "ldr x20, [x12, #0xb8]\n"
+      "mov v24.d[1], x20\n"
+      "ldr x23, [x12, #0xc8]\n"
       ".inst 0x4f80ea90  // sdot v16.4s, v20.16b, v0.4b[2]\n"
-      "ldr d20, [x14, #0xf0]\n"
-      "ldr x20, [x14, #0xb8]\n"
-      "mov v26.d[1], x22\n"
-      "mov v25.d[1], x21\n"
-      "ldr x23, [x14, #0xc8]\n"
-      "ldr x22, [x14, #0xd8]\n"
+      "ldr d20, [x12, #0xf0]\n"
       ".inst 0x4f80eb51  // sdot v17.4s, v26.16b, v0.4b[2]\n"
-      "mov v24.d[1], x20\n"
-      "ldr x21, [x14, #0xe8]\n"
-      "ldr x20, [x14, #0xf8]\n"
+      "ldr x22, [x12, #0xd8]\n"
       ".inst 0x4f80eb32  // sdot v18.4s, v25.16b, v0.4b[2]\n"
+      "ldr x21, [x12, #0xe8]\n"
       ".inst 0x4f80eb13  // sdot v19.4s, v24.16b, v0.4b[2]\n"
+      "ldr x20, [x12, #0xf8]\n"
       "mov v23.d[1], x23\n"
       "mov v22.d[1], x22\n"
-      "add x14, x14, #0x100\n"
+      "add x9, x9, #0x10\n"
       "mov v21.d[1], x21\n"
-      ".inst 0x4fa0eaf0  // sdot v16.4s, v23.16b, v0.4b[3]\n"
+      "add x12, x12, #0x100\n"
       "mov v20.d[1], x20\n"
+      ".inst 0x4fa0eaf0  // sdot v16.4s, v23.16b, v0.4b[3]\n"
       ".inst 0x4fa0ead1  // sdot v17.4s, v22.16b, v0.4b[3]\n"
       ".inst 0x4fa0eab2  // sdot v18.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa0ea93  // sdot v19.4s, v20.16b, v0.4b[3]\n"
       "tbnz %x[flags], #31, 8f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "8:"  // Height 1: Multiply loop: unique 1: skip row sum
-      "ldr q0, [x10, #0x0]\n"
-      "sub x11, x11, #0x10\n"
-      "ldr q4, [x14, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "sub x10, x10, #0x10\n"
+      "ldr q4, [x12, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "bge 7b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x4f80e090  // sdot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr q21, [x14, #0x70]\n"
+      "ldr q21, [x12, #0x70]\n"
       ".inst 0x4f80e0b1  // sdot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr q20, [x14, #0x80]\n"
+      "ldr q20, [x12, #0x80]\n"
       ".inst 0x4f80e0d2  // sdot v18.4s, v6.16b, v0.4b[0]\n"
-      "ldr q26, [x14, #0x90]\n"
+      "ldr q26, [x12, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr q25, [x14, #0xa0]\n"
+      "ldr q25, [x12, #0xa0]\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
-      "ldr q24, [x14, #0xb0]\n"
+      "ldr q24, [x12, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
-      "ldr q23, [x14, #0xc0]\n"
+      "ldr q23, [x12, #0xc0]\n"
       ".inst 0x4fa0e152  // sdot v18.4s, v10.16b, v0.4b[1]\n"
-      "ldr q22, [x14, #0xd0]\n"
+      "ldr q22, [x12, #0xd0]\n"
       ".inst 0x4fa0e2b3  // sdot v19.4s, v21.16b, v0.4b[1]\n"
-      "ldr q21, [x14, #0xe0]\n"
+      "ldr q21, [x12, #0xe0]\n"
       ".inst 0x4f80ea90  // sdot v16.4s, v20.16b, v0.4b[2]\n"
-      "ldr q20, [x14, #0xf0]\n"
+      "ldr q20, [x12, #0xf0]\n"
       ".inst 0x4f80eb51  // sdot v17.4s, v26.16b, v0.4b[2]\n"
-      "sub x11, x11, #0x10\n"
+      "sub x10, x10, #0x10\n"
       ".inst 0x4f80eb32  // sdot v18.4s, v25.16b, v0.4b[2]\n"
-      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x4f80eb13  // sdot v19.4s, v24.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x4fa0eaf0  // sdot v16.4s, v23.16b, v0.4b[3]\n"
       ".inst 0x4fa0ead1  // sdot v17.4s, v22.16b, v0.4b[3]\n"
       ".inst 0x4fa0eab2  // sdot v18.4s, v21.16b, v0.4b[3]\n"
@@ -218,54 +218,54 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "tbnz %x[flags], #31, 10f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "10:"  // Height 1: Multiply loop: unique 2: skip row sum
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "11:"  // Height 1: Multiply loop: Main loop skip
-      "cbz x11, 18f\n"
-      "cmp x11, #0x4\n"
+      "cbz x10, 18f\n"
+      "cmp x10, #0x4\n"
       "blt 14f\n"
       "12:"  // Height 1: Multiply loop: Odd block loop
-      "ldr s0, [x10], #0x4\n"
+      "ldr s0, [x9], #0x4\n"
       "tbnz %x[flags], #31, 13f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "13:"  // Height 1: Multiply loop: unique 3: skip row sum
-      "ldr q23, [x14, #0x0]\n"
-      "sub x11, x11, #0x4\n"
-      "ldr q22, [x14, #0x10]\n"
-      "cmp x11, #0x4\n"
-      "ldr q21, [x14, #0x20]\n"
-      "ldr q20, [x14, #0x30]\n"
-      ".inst 0x4f80e2f0  // sdot v16.4s, v23.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x0]\n"
+      "sub x10, x10, #0x4\n"
+      "ldr q22, [x12, #0x10]\n"
+      "cmp x10, #0x4\n"
+      "ldr q21, [x12, #0x20]\n"
+      ".inst 0x4f80e290  // sdot v16.4s, v20.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x30]\n"
       ".inst 0x4f80e2d1  // sdot v17.4s, v22.16b, v0.4b[0]\n"
-      "add x14, x14, #0x40\n"
       ".inst 0x4f80e2b2  // sdot v18.4s, v21.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x4f80e293  // sdot v19.4s, v20.16b, v0.4b[0]\n"
       "bge 12b\n"
       "14:"  // Height 1: Multiply loop: Skip odd blocks
-      "cbz x11, 18f\n"
-      "tbz x11, #1, 15f\n"
-      "ldr h0, [x10], #0x2\n"
-      "tbz x11, #0, 16f\n"
-      "ld1 { v0.b }[2], [x10]\n"
+      "cbz x10, 18f\n"
+      "tbz x10, #1, 15f\n"
+      "ldr h0, [x9], #0x2\n"
+      "tbz x10, #0, 16f\n"
+      "ld1 { v0.b }[2], [x9]\n"
       "b 16f\n"
       "15:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x10, #0x0]\n"
+      "ldr b0, [x9, #0x0]\n"
       "16:"  // Height 1: Multiply loop: Ragged operand read: Done
       "tbnz %x[flags], #31, 17f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "17:"  // Height 1: Multiply loop: unique 4: skip row sum
-      "ldr q23, [x14, #0x0]\n"
-      "ldr q22, [x14, #0x10]\n"
-      "ldr q21, [x14, #0x20]\n"
-      "ldr q20, [x14, #0x30]\n"
-      ".inst 0x4f80e2f0  // sdot v16.4s, v23.16b, v0.4b[0]\n"
-      ".inst 0x4f80e2d1  // sdot v17.4s, v22.16b, v0.4b[0]\n"
-      "add x14, x14, #0x40\n"
-      ".inst 0x4f80e2b2  // sdot v18.4s, v21.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x0]\n"
+      ".inst 0x4f80e290  // sdot v16.4s, v20.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x10]\n"
+      ".inst 0x4f80e291  // sdot v17.4s, v20.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x20]\n"
+      ".inst 0x4f80e292  // sdot v18.4s, v20.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x30]\n"
       ".inst 0x4f80e293  // sdot v19.4s, v20.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       "18:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x12, x12, #0x1\n"
-      "cmp x12, x20\n"
+      "add x11, x11, #0x1\n"
+      "cmp x11, x20\n"
       "bne 4b\n"
       "prfm pstl1keep, [x13, #0x0]\n"
       "tbnz %x[flags], #31, 19f\n"
@@ -276,28 +276,28 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "addp v11.4s, v11.4s, v11.4s\n"
       "mul v11.4s, v11.4s, v20.4s\n"
       "19:"  // Height 1: skip row sum fixup
-      "ldr q24, [x16, #0x0]\n"
+      "ldr q23, [x15, #0x0]\n"
       "add v16.4s, v16.4s, v11.4s\n"
-      "ldr q23, [x16, #0x10]\n"
+      "ldr q22, [x15, #0x10]\n"
       "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q22, [x16, #0x20]\n"
+      "ldr q21, [x15, #0x20]\n"
       "add v18.4s, v18.4s, v11.4s\n"
-      "ldr q21, [x16, #0x30]\n"
+      "ldr q20, [x15, #0x30]\n"
       "add v19.4s, v19.4s, v11.4s\n"
+      "add v16.4s, v16.4s, v23.4s\n"
+      "add v17.4s, v17.4s, v22.4s\n"
+      "add v18.4s, v18.4s, v21.4s\n"
+      "add v19.4s, v19.4s, v20.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v20.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v24.4s\n"
-      "add v17.4s, v17.4s, v23.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v22.4s\n"
-      "add v19.4s, v19.4s, v21.4s\n"
-      "add x16, x16, #0x40\n"
       "sqrdmulh v16.4s, v16.4s, v20.4s\n"
       "sqrdmulh v17.4s, v17.4s, v20.4s\n"
       "sqrdmulh v18.4s, v18.4s, v20.4s\n"
       "sqrdmulh v19.4s, v19.4s, v20.4s\n"
+      "add x15, x15, #0x40\n"
       "tbz %x[flags], #5, 20f\n"
       "and v23.16b, v16.16b, v0.16b\n"
       "and v22.16b, v17.16b, v0.16b\n"
@@ -317,67 +317,67 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v22.4s }, [x20]\n"
+      "ld1r { v20.4s }, [x20]\n"
+      "add v16.4s, v16.4s, v20.4s\n"
+      "add v17.4s, v17.4s, v20.4s\n"
+      "add v18.4s, v18.4s, v20.4s\n"
+      "add v19.4s, v19.4s, v20.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v20.4s }, [x20]\n"
+      "smin v16.4s, v16.4s, v20.4s\n"
+      "smin v17.4s, v17.4s, v20.4s\n"
+      "smin v18.4s, v18.4s, v20.4s\n"
+      "smin v19.4s, v19.4s, v20.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v21.4s }, [x21]\n"
-      "cmp x15, #0x10\n"
       "ld1r { v20.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v22.4s\n"
-      "add v17.4s, v17.4s, v22.4s\n"
-      "add v18.4s, v18.4s, v22.4s\n"
-      "add v19.4s, v19.4s, v22.4s\n"
-      "smin v16.4s, v16.4s, v21.4s\n"
-      "smin v17.4s, v17.4s, v21.4s\n"
-      "smin v18.4s, v18.4s, v21.4s\n"
-      "smin v19.4s, v19.4s, v21.4s\n"
       "smax v16.4s, v16.4s, v20.4s\n"
       "smax v17.4s, v17.4s, v20.4s\n"
       "smax v18.4s, v18.4s, v20.4s\n"
       "smax v19.4s, v19.4s, v20.4s\n"
       "uzp1 v16.8h, v16.8h, v17.8h\n"
       "uzp1 v17.8h, v18.8h, v19.8h\n"
+      "cmp x14, #0x10\n"
       "uzp1 v16.16b, v16.16b, v17.16b\n"
       "bge 29f\n"
-      "tbz x15, #3, 24f\n"
+      "tbz x14, #3, 24f\n"
       "str d16, [x13], #0x8\n"
-      "tbz x15, #2, 22f\n"
+      "tbz x14, #2, 22f\n"
       "st1 { v16.s }[2], [x13], #0x4\n"
-      "tbz x15, #1, 21f\n"
+      "tbz x14, #1, 21f\n"
       "st1 { v16.h }[6], [x13], #0x2\n"
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[14], [x13]\n"
       "b 28f\n"
       "21:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[12], [x13]\n"
       "b 28f\n"
       "22:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x15, #1, 23f\n"
+      "tbz x14, #1, 23f\n"
       "st1 { v16.h }[4], [x13], #0x2\n"
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[10], [x13]\n"
       "b 28f\n"
       "23:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[8], [x13]\n"
       "b 28f\n"
       "24:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x15, #2, 26f\n"
+      "tbz x14, #2, 26f\n"
       "str s16, [x13], #0x4\n"
-      "tbz x15, #1, 25f\n"
+      "tbz x14, #1, 25f\n"
       "st1 { v16.h }[2], [x13], #0x2\n"
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[6], [x13]\n"
       "b 28f\n"
       "25:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[4], [x13]\n"
       "b 28f\n"
       "26:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x15, #1, 27f\n"
+      "tbz x14, #1, 27f\n"
       "str h16, [x13], #0x2\n"
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[2], [x13]\n"
       "b 28f\n"
       "27:"  // Height 1: Partial direct writeback: partial_1_0
@@ -388,18 +388,18 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "str q16, [x13, #0x0]\n"
       "add x13, x13, #0x10\n"
       "30:"  // Height 1: Writeback done
-      "subs x15, x15, #0x10\n"
+      "subs x14, x14, #0x10\n"
       "bgt 2b\n"
       "b 122f\n"
       "31:"  // Height 2
-      "mov x16, %x[col_bias]\n"
+      "mov x15, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "movi v15.16b, #0x1\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "32:"  // Height 2: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -410,80 +410,80 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "movi v22.4s, #0x0\n"
       "movi v23.4s, #0x0\n"
       "33:"  // Height 2: setup done
-      "mov x12, #0x0\n"
+      "mov x11, #0x0\n"
       "34:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w10, [x20, x11, LSL #0x2]\n"
       "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w11, [x20, x12, LSL #0x2]\n"
       "tbz %x[flags], #3, 35f\n"
-      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
-      "ldr x10, [x20, #0x0]\n"
-      "ldr x9, [x20, #0x8]\n"
-      "cbnz x12, 36f\n"
+      "ldr x9, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "cbnz x11, 36f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x10, x10, x20\n"
       "add x9, x9, x20\n"
+      "add x28, x28, x20\n"
       "b 36f\n"
       "35:"  // Height 2: setup direct input
-      "mov x10, %x[input_ptr]\n"
-      "add x9, x10, x21\n"
+      "mov x9, %x[input_ptr]\n"
+      "add x28, x9, x21\n"
       "36:"  // Height 2: input setup done
-      "cmp x11, #0x10\n"
+      "cmp x10, #0x10\n"
       "blt 41f\n"
-      "ldr q0, [x10, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q1, [x9, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "blt 39f\n"
       "37:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x4f80e090  // sdot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr x20, [x14, #0x78]\n"
+      "ldr x20, [x12, #0x78]\n"
       ".inst 0x4f81e094  // sdot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr d25, [x14, #0x70]\n"
+      "ldr d25, [x12, #0x70]\n"
       ".inst 0x4f80e0b1  // sdot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr x23, [x14, #0x88]\n"
+      "mov v25.d[1], x20\n"
       ".inst 0x4f81e0b5  // sdot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr d24, [x14, #0x80]\n"
+      "ldr d24, [x12, #0x80]\n"
       ".inst 0x4f80e0d2  // sdot v18.4s, v6.16b, v0.4b[0]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x23, [x12, #0x88]\n"
       ".inst 0x4f81e0d6  // sdot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr d30, [x14, #0x90]\n"
+      "ldr d30, [x12, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr x22, [x14, #0x98]\n"
+      "ldr x22, [x12, #0x98]\n"
       ".inst 0x4f81e0f7  // sdot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr d29, [x14, #0xa0]\n"
-      "ldr x21, [x14, #0xa8]\n"
+      "ldr d29, [x12, #0xa0]\n"
+      "ldr x21, [x12, #0xa8]\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
       ".inst 0x4fa1e114  // sdot v20.4s, v8.16b, v1.4b[1]\n"
-      "ldr d28, [x14, #0xb0]\n"
-      "ldr x20, [x14, #0xb8]\n"
+      "ldr d28, [x12, #0xb0]\n"
+      "ldr x20, [x12, #0xb8]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
       ".inst 0x4fa1e135  // sdot v21.4s, v9.16b, v1.4b[1]\n"
-      "ldr d27, [x14, #0xc0]\n"
+      "ldr d27, [x12, #0xc0]\n"
       ".inst 0x4fa0e152  // sdot v18.4s, v10.16b, v0.4b[1]\n"
       "mov v24.d[1], x23\n"
       ".inst 0x4fa1e156  // sdot v22.4s, v10.16b, v1.4b[1]\n"
-      "ldr d26, [x14, #0xd0]\n"
+      "ldr d26, [x12, #0xd0]\n"
       ".inst 0x4fa0e333  // sdot v19.4s, v25.16b, v0.4b[1]\n"
       "mov v30.d[1], x22\n"
       ".inst 0x4fa1e337  // sdot v23.4s, v25.16b, v1.4b[1]\n"
-      "ldr d25, [x14, #0xe0]\n"
+      "ldr d25, [x12, #0xe0]\n"
       "mov v29.d[1], x21\n"
-      "ldr x23, [x14, #0xc8]\n"
+      "ldr x23, [x12, #0xc8]\n"
       "mov v28.d[1], x20\n"
-      "ldr x22, [x14, #0xd8]\n"
-      "ldr x21, [x14, #0xe8]\n"
+      "ldr x22, [x12, #0xd8]\n"
+      "ldr x21, [x12, #0xe8]\n"
       ".inst 0x4f80eb10  // sdot v16.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb14  // sdot v20.4s, v24.16b, v1.4b[2]\n"
-      "ldr d24, [x14, #0xf0]\n"
-      "ldr x20, [x14, #0xf8]\n"
+      "ldr d24, [x12, #0xf0]\n"
+      "ldr x20, [x12, #0xf8]\n"
       ".inst 0x4f80ebd1  // sdot v17.4s, v30.16b, v0.4b[2]\n"
       ".inst 0x4f81ebd5  // sdot v21.4s, v30.16b, v1.4b[2]\n"
       "mov v27.d[1], x23\n"
@@ -494,9 +494,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4f80eb93  // sdot v19.4s, v28.16b, v0.4b[2]\n"
       "mov v24.d[1], x20\n"
       ".inst 0x4f81eb97  // sdot v23.4s, v28.16b, v1.4b[2]\n"
-      "add x10, x10, #0x10\n"
       "add x9, x9, #0x10\n"
-      "add x14, x14, #0x100\n"
+      "add x28, x28, #0x10\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x4fa0eb70  // sdot v16.4s, v27.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb74  // sdot v20.4s, v27.16b, v1.4b[3]\n"
       ".inst 0x4fa0eb51  // sdot v17.4s, v26.16b, v0.4b[3]\n"
@@ -509,53 +509,53 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       "38:"  // Height 2: Multiply loop: unique 5: skip row sum
-      "ldr q0, [x10, #0x0]\n"
-      "sub x11, x11, #0x10\n"
-      "ldr q1, [x9, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "sub x10, x10, #0x10\n"
+      "ldr q1, [x28, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "bge 37b\n"
       "39:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x4f80e090  // sdot v16.4s, v4.16b, v0.4b[0]\n"
-      "sub x11, x11, #0x10\n"
+      "sub x10, x10, #0x10\n"
       ".inst 0x4f81e094  // sdot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr q25, [x14, #0x70]\n"
+      "ldr q25, [x12, #0x70]\n"
       ".inst 0x4f80e0b1  // sdot v17.4s, v5.16b, v0.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x4f81e0b5  // sdot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr q24, [x14, #0x80]\n"
+      "ldr q24, [x12, #0x80]\n"
       ".inst 0x4f80e0d2  // sdot v18.4s, v6.16b, v0.4b[0]\n"
-      "add x9, x9, #0x10\n"
+      "add x28, x28, #0x10\n"
       ".inst 0x4f81e0d6  // sdot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr q30, [x14, #0x90]\n"
+      "ldr q30, [x12, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0f7  // sdot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr q29, [x14, #0xa0]\n"
+      "ldr q29, [x12, #0xa0]\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
       ".inst 0x4fa1e114  // sdot v20.4s, v8.16b, v1.4b[1]\n"
-      "ldr q28, [x14, #0xb0]\n"
+      "ldr q28, [x12, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
       ".inst 0x4fa1e135  // sdot v21.4s, v9.16b, v1.4b[1]\n"
-      "ldr q27, [x14, #0xc0]\n"
+      "ldr q27, [x12, #0xc0]\n"
       ".inst 0x4fa0e152  // sdot v18.4s, v10.16b, v0.4b[1]\n"
       ".inst 0x4fa1e156  // sdot v22.4s, v10.16b, v1.4b[1]\n"
-      "ldr q26, [x14, #0xd0]\n"
+      "ldr q26, [x12, #0xd0]\n"
       ".inst 0x4fa0e333  // sdot v19.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e337  // sdot v23.4s, v25.16b, v1.4b[1]\n"
-      "ldr q25, [x14, #0xe0]\n"
+      "ldr q25, [x12, #0xe0]\n"
       ".inst 0x4f80eb10  // sdot v16.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb14  // sdot v20.4s, v24.16b, v1.4b[2]\n"
-      "ldr q24, [x14, #0xf0]\n"
+      "ldr q24, [x12, #0xf0]\n"
       ".inst 0x4f80ebd1  // sdot v17.4s, v30.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x4f81ebd5  // sdot v21.4s, v30.16b, v1.4b[2]\n"
       ".inst 0x4f80ebb2  // sdot v18.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebb6  // sdot v22.4s, v29.16b, v1.4b[2]\n"
@@ -573,29 +573,29 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       "40:"  // Height 2: Multiply loop: unique 6: skip row sum
-      "prfm pldl1keep, [x10, #0x80]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "41:"  // Height 2: Multiply loop: Main loop skip
-      "cbz x11, 48f\n"
-      "cmp x11, #0x4\n"
+      "cbz x10, 48f\n"
+      "cmp x10, #0x4\n"
       "blt 44f\n"
       "42:"  // Height 2: Multiply loop: Odd block loop
-      "ldr s0, [x10], #0x4\n"
-      "ldr s1, [x9], #0x4\n"
+      "ldr s0, [x9], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
       "tbnz %x[flags], #31, 43f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       "43:"  // Height 2: Multiply loop: unique 7: skip row sum
-      "ldr q27, [x14, #0x0]\n"
-      "sub x11, x11, #0x4\n"
-      "ldr q26, [x14, #0x10]\n"
-      "cmp x11, #0x4\n"
-      "ldr q25, [x14, #0x20]\n"
-      "ldr q24, [x14, #0x30]\n"
+      "ldr q27, [x12, #0x0]\n"
+      "sub x10, x10, #0x4\n"
+      "ldr q26, [x12, #0x10]\n"
+      "cmp x10, #0x4\n"
+      "ldr q25, [x12, #0x20]\n"
       ".inst 0x4f80e370  // sdot v16.4s, v27.16b, v0.4b[0]\n"
+      "ldr q24, [x12, #0x30]\n"
       ".inst 0x4f81e374  // sdot v20.4s, v27.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
       ".inst 0x4f80e351  // sdot v17.4s, v26.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x4f81e355  // sdot v21.4s, v26.16b, v1.4b[0]\n"
       ".inst 0x4f80e332  // sdot v18.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x4f81e336  // sdot v22.4s, v25.16b, v1.4b[0]\n"
@@ -603,44 +603,44 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4f81e317  // sdot v23.4s, v24.16b, v1.4b[0]\n"
       "bge 42b\n"
       "44:"  // Height 2: Multiply loop: Skip odd blocks
-      "cbz x11, 48f\n"
-      "tbz x11, #1, 45f\n"
-      "ldr h0, [x10], #0x2\n"
-      "ldr h1, [x9], #0x2\n"
-      "tbz x11, #0, 46f\n"
-      "ld1 { v0.b }[2], [x10]\n"
-      "ld1 { v1.b }[2], [x9]\n"
+      "cbz x10, 48f\n"
+      "tbz x10, #1, 45f\n"
+      "ldr h0, [x9], #0x2\n"
+      "ldr h1, [x28], #0x2\n"
+      "tbz x10, #0, 46f\n"
+      "ld1 { v0.b }[2], [x9]\n"
+      "ld1 { v1.b }[2], [x28]\n"
       "b 46f\n"
       "45:"  // Height 2: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x10, #0x0]\n"
-      "ldr b1, [x9, #0x0]\n"
+      "ldr b0, [x9, #0x0]\n"
+      "ldr b1, [x28, #0x0]\n"
       "46:"  // Height 2: Multiply loop: Ragged operand read: Done
       "tbnz %x[flags], #31, 47f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       "47:"  // Height 2: Multiply loop: unique 8: skip row sum
-      "ldr q27, [x14, #0x0]\n"
-      "ldr q26, [x14, #0x10]\n"
-      "ldr q25, [x14, #0x20]\n"
-      "ldr q24, [x14, #0x30]\n"
-      ".inst 0x4f80e370  // sdot v16.4s, v27.16b, v0.4b[0]\n"
-      ".inst 0x4f81e374  // sdot v20.4s, v27.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
+      "ldr q24, [x12, #0x0]\n"
+      ".inst 0x4f80e310  // sdot v16.4s, v24.16b, v0.4b[0]\n"
+      "ldr q26, [x12, #0x10]\n"
+      ".inst 0x4f81e314  // sdot v20.4s, v24.16b, v1.4b[0]\n"
+      "ldr q25, [x12, #0x20]\n"
       ".inst 0x4f80e351  // sdot v17.4s, v26.16b, v0.4b[0]\n"
+      "ldr q24, [x12, #0x30]\n"
       ".inst 0x4f81e355  // sdot v21.4s, v26.16b, v1.4b[0]\n"
       ".inst 0x4f80e332  // sdot v18.4s, v25.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x4f81e336  // sdot v22.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x4f80e313  // sdot v19.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e317  // sdot v23.4s, v24.16b, v1.4b[0]\n"
       "48:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x12, x12, #0x1\n"
-      "cmp x12, x20\n"
+      "add x11, x11, #0x1\n"
+      "cmp x11, x20\n"
       "bne 34b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20\n"
       "prfm pstl1keep, [x13, #0x0]\n"
-      "add x24, x13, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbnz %x[flags], #31, 49f\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
@@ -652,28 +652,28 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "mul v11.4s, v11.4s, v24.4s\n"
       "mul v12.4s, v12.4s, v24.4s\n"
       "49:"  // Height 2: skip row sum fixup
-      "ldr q28, [x16, #0x0]\n"
+      "ldr q27, [x15, #0x0]\n"
       "add v16.4s, v16.4s, v11.4s\n"
-      "ldr q27, [x16, #0x10]\n"
+      "ldr q26, [x15, #0x10]\n"
       "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q26, [x16, #0x20]\n"
+      "ldr q25, [x15, #0x20]\n"
       "add v18.4s, v18.4s, v11.4s\n"
-      "ldr q25, [x16, #0x30]\n"
+      "ldr q24, [x15, #0x30]\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
       "add v22.4s, v22.4s, v12.4s\n"
       "add v23.4s, v23.4s, v12.4s\n"
-      "add v16.4s, v16.4s, v28.4s\n"
-      "add v17.4s, v17.4s, v27.4s\n"
-      "add v18.4s, v18.4s, v26.4s\n"
+      "add v16.4s, v16.4s, v27.4s\n"
+      "add v17.4s, v17.4s, v26.4s\n"
+      "add v18.4s, v18.4s, v25.4s\n"
+      "add v19.4s, v19.4s, v24.4s\n"
+      "add v20.4s, v20.4s, v27.4s\n"
+      "add v21.4s, v21.4s, v26.4s\n"
+      "add v22.4s, v22.4s, v25.4s\n"
+      "add v23.4s, v23.4s, v24.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
       "ld1r { v24.4s }, [x20]\n"
-      "add v19.4s, v19.4s, v25.4s\n"
-      "add v20.4s, v20.4s, v28.4s\n"
-      "add v21.4s, v21.4s, v27.4s\n"
-      "add v22.4s, v22.4s, v26.4s\n"
-      "add v23.4s, v23.4s, v25.4s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "ld1r { v0.4s }, [x20]\n"
@@ -685,31 +685,31 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "sqrdmulh v21.4s, v21.4s, v24.4s\n"
       "sqrdmulh v22.4s, v22.4s, v24.4s\n"
       "sqrdmulh v23.4s, v23.4s, v24.4s\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       "tbz %x[flags], #5, 50f\n"
       "and v24.16b, v16.16b, v0.16b\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
+      "sqadd v16.4s, v16.4s, v24.4s\n"
       "and v30.16b, v17.16b, v0.16b\n"
       "and v29.16b, v18.16b, v0.16b\n"
       "and v28.16b, v19.16b, v0.16b\n"
       "and v27.16b, v20.16b, v0.16b\n"
       "and v26.16b, v21.16b, v0.16b\n"
       "and v25.16b, v22.16b, v0.16b\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
+      "and v24.16b, v23.16b, v0.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
       "sshr v26.4s, v26.4s, #0x1f\n"
       "sshr v25.4s, v25.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v24.4s\n"
-      "and v24.16b, v23.16b, v0.16b\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v17.4s, v17.4s, v30.4s\n"
       "sqadd v18.4s, v18.4s, v29.4s\n"
       "sqadd v19.4s, v19.4s, v28.4s\n"
       "sqadd v20.4s, v20.4s, v27.4s\n"
       "sqadd v21.4s, v21.4s, v26.4s\n"
       "sqadd v22.4s, v22.4s, v25.4s\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v23.4s, v23.4s, v24.4s\n"
       "50:"  // Height 2: no shift correction
       "srshl v16.4s, v16.4s, v0.4s\n"
@@ -721,28 +721,27 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "srshl v22.4s, v22.4s, v0.4s\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x20]\n"
+      "ld1r { v24.4s }, [x20]\n"
+      "add v16.4s, v16.4s, v24.4s\n"
+      "add v17.4s, v17.4s, v24.4s\n"
+      "add v18.4s, v18.4s, v24.4s\n"
+      "add v19.4s, v19.4s, v24.4s\n"
+      "add v20.4s, v20.4s, v24.4s\n"
+      "add v21.4s, v21.4s, v24.4s\n"
+      "add v22.4s, v22.4s, v24.4s\n"
+      "add v23.4s, v23.4s, v24.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v24.4s }, [x20]\n"
+      "smin v16.4s, v16.4s, v24.4s\n"
+      "smin v17.4s, v17.4s, v24.4s\n"
+      "smin v18.4s, v18.4s, v24.4s\n"
+      "smin v19.4s, v19.4s, v24.4s\n"
+      "smin v20.4s, v20.4s, v24.4s\n"
+      "smin v21.4s, v21.4s, v24.4s\n"
+      "smin v22.4s, v22.4s, v24.4s\n"
+      "smin v23.4s, v23.4s, v24.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v25.4s }, [x21]\n"
-      "cmp x15, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v26.4s\n"
-      "add v17.4s, v17.4s, v26.4s\n"
-      "add v18.4s, v18.4s, v26.4s\n"
-      "add v19.4s, v19.4s, v26.4s\n"
-      "add v20.4s, v20.4s, v26.4s\n"
-      "add v21.4s, v21.4s, v26.4s\n"
-      "add v22.4s, v22.4s, v26.4s\n"
-      "add v23.4s, v23.4s, v26.4s\n"
-      "smin v16.4s, v16.4s, v25.4s\n"
-      "smin v17.4s, v17.4s, v25.4s\n"
-      "smin v18.4s, v18.4s, v25.4s\n"
-      "smin v19.4s, v19.4s, v25.4s\n"
-      "smin v20.4s, v20.4s, v25.4s\n"
-      "smin v21.4s, v21.4s, v25.4s\n"
-      "smin v22.4s, v22.4s, v25.4s\n"
-      "smin v23.4s, v23.4s, v25.4s\n"
       "smax v16.4s, v16.4s, v24.4s\n"
       "smax v17.4s, v17.4s, v24.4s\n"
       "smax v18.4s, v18.4s, v24.4s\n"
@@ -755,87 +754,88 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "uzp1 v18.8h, v18.8h, v19.8h\n"
       "uzp1 v20.8h, v20.8h, v21.8h\n"
       "uzp1 v17.8h, v22.8h, v23.8h\n"
+      "cmp x14, #0x10\n"
       "uzp1 v16.16b, v16.16b, v18.16b\n"
       "uzp1 v20.16b, v20.16b, v17.16b\n"
       "bge 59f\n"
-      "tbz x15, #3, 54f\n"
+      "tbz x14, #3, 54f\n"
       "str d16, [x13], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "tbz x15, #2, 52f\n"
+      "str d20, [x23], #0x8\n"
+      "tbz x14, #2, 52f\n"
       "st1 { v16.s }[2], [x13], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "tbz x15, #1, 51f\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "tbz x14, #1, 51f\n"
       "st1 { v16.h }[6], [x13], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "tbz x15, #0, 58f\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[14], [x13]\n"
-      "st1 { v20.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
       "b 58f\n"
       "51:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x15, #0, 58f\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[12], [x13]\n"
-      "st1 { v20.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
       "b 58f\n"
       "52:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x15, #1, 53f\n"
+      "tbz x14, #1, 53f\n"
       "st1 { v16.h }[4], [x13], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "tbz x15, #0, 58f\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[10], [x13]\n"
-      "st1 { v20.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
       "b 58f\n"
       "53:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x15, #0, 58f\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[8], [x13]\n"
-      "st1 { v20.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
       "b 58f\n"
       "54:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x15, #2, 56f\n"
+      "tbz x14, #2, 56f\n"
       "str s16, [x13], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "tbz x15, #1, 55f\n"
+      "str s20, [x23], #0x4\n"
+      "tbz x14, #1, 55f\n"
       "st1 { v16.h }[2], [x13], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "tbz x15, #0, 58f\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[6], [x13]\n"
-      "st1 { v20.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
       "b 58f\n"
       "55:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x15, #0, 58f\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[4], [x13]\n"
-      "st1 { v20.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
       "b 58f\n"
       "56:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x15, #1, 57f\n"
+      "tbz x14, #1, 57f\n"
       "str h16, [x13], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "tbz x15, #0, 58f\n"
+      "str h20, [x23], #0x2\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[2], [x13]\n"
-      "st1 { v20.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
       "b 58f\n"
       "57:"  // Height 2: Partial direct writeback: partial_1_0
       "str b16, [x13, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
       "58:"  // Height 2: Partial direct writeback: Done
       "b 60f\n"
       "59:"  // Height 2: Full writeback
       "str q16, [x13, #0x0]\n"
       "add x13, x13, #0x10\n"
-      "str q20, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
       "60:"  // Height 2: Writeback done
-      "subs x15, x15, #0x10\n"
+      "subs x14, x14, #0x10\n"
       "bgt 32b\n"
       "b 122f\n"
       "61:"  // Height 3
-      "mov x16, %x[col_bias]\n"
+      "mov x15, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "movi v13.4s, #0x0\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "movi v15.16b, #0x1\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "62:"  // Height 3: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -850,105 +850,105 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "movi v26.4s, #0x0\n"
       "movi v27.4s, #0x0\n"
       "63:"  // Height 3: setup done
-      "mov x12, #0x0\n"
+      "mov x11, #0x0\n"
       "64:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w10, [x20, x11, LSL #0x2]\n"
       "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w11, [x20, x12, LSL #0x2]\n"
       "tbz %x[flags], #3, 65f\n"
-      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
-      "ldr x10, [x20, #0x0]\n"
-      "ldr x9, [x20, #0x8]\n"
-      "ldr x28, [x20, #0x10]\n"
-      "cbnz x12, 66f\n"
+      "ldr x9, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "ldr x27, [x20, #0x10]\n"
+      "cbnz x11, 66f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x10, x10, x20\n"
       "add x9, x9, x20\n"
       "add x28, x28, x20\n"
+      "add x27, x27, x20\n"
       "b 66f\n"
       "65:"  // Height 3: setup direct input
-      "mov x10, %x[input_ptr]\n"
-      "add x9, x10, x21\n"
+      "mov x9, %x[input_ptr]\n"
       "add x28, x9, x21\n"
+      "add x27, x28, x21\n"
       "66:"  // Height 3: input setup done
-      "cmp x11, #0x10\n"
+      "cmp x10, #0x10\n"
       "blt 71f\n"
-      "ldr q0, [x10, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q1, [x9, #0x0]\n"
-      "ldr q2, [x28, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x27, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "blt 69f\n"
       "67:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x4f80e090  // sdot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr x20, [x14, #0x78]\n"
+      "ldr x20, [x12, #0x78]\n"
       ".inst 0x4f81e094  // sdot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr x23, [x14, #0x88]\n"
+      "ldr x23, [x12, #0x88]\n"
       ".inst 0x4f82e098  // sdot v24.4s, v4.16b, v2.4b[0]\n"
-      "ldr d29, [x14, #0x70]\n"
+      "ldr d29, [x12, #0x70]\n"
       ".inst 0x4f80e0b1  // sdot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr x22, [x14, #0x98]\n"
+      "mov v29.d[1], x20\n"
       ".inst 0x4f81e0b5  // sdot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr x21, [x14, #0xa8]\n"
+      "ldr x22, [x12, #0x98]\n"
       ".inst 0x4f82e0b9  // sdot v25.4s, v5.16b, v2.4b[0]\n"
-      "ldr d28, [x14, #0x80]\n"
+      "ldr d28, [x12, #0x80]\n"
       ".inst 0x4f80e0d2  // sdot v18.4s, v6.16b, v0.4b[0]\n"
-      "mov v29.d[1], x20\n"
+      "ldr x21, [x12, #0xa8]\n"
       ".inst 0x4f81e0d6  // sdot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x14, #0xb8]\n"
+      "ldr x20, [x12, #0xb8]\n"
       ".inst 0x4f82e0da  // sdot v26.4s, v6.16b, v2.4b[0]\n"
-      "ldr d5, [x14, #0x90]\n"
+      "ldr d5, [x12, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
       "mov v28.d[1], x23\n"
       ".inst 0x4f81e0f7  // sdot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr x23, [x14, #0xc8]\n"
+      "mov v5.d[1], x22\n"
       ".inst 0x4f82e0fb  // sdot v27.4s, v7.16b, v2.4b[0]\n"
-      "ldr d4, [x14, #0xa0]\n"
+      "ldr d4, [x12, #0xa0]\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
-      "mov v5.d[1], x22\n"
+      "mov v4.d[1], x21\n"
       ".inst 0x4fa1e114  // sdot v20.4s, v8.16b, v1.4b[1]\n"
-      "ldr x22, [x14, #0xd8]\n"
+      "ldr x23, [x12, #0xc8]\n"
       ".inst 0x4fa2e118  // sdot v24.4s, v8.16b, v2.4b[1]\n"
-      "ldr d3, [x14, #0xb0]\n"
+      "ldr d3, [x12, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
-      "mov v4.d[1], x21\n"
+      "mov v3.d[1], x20\n"
       ".inst 0x4fa1e135  // sdot v21.4s, v9.16b, v1.4b[1]\n"
-      "ldr x21, [x14, #0xe8]\n"
+      "ldr x22, [x12, #0xd8]\n"
       ".inst 0x4fa2e139  // sdot v25.4s, v9.16b, v2.4b[1]\n"
-      "ldr d31, [x14, #0xc0]\n"
+      "ldr d31, [x12, #0xc0]\n"
       ".inst 0x4fa0e152  // sdot v18.4s, v10.16b, v0.4b[1]\n"
-      "mov v3.d[1], x20\n"
+      "ldr x21, [x12, #0xe8]\n"
       ".inst 0x4fa1e156  // sdot v22.4s, v10.16b, v1.4b[1]\n"
-      "ldr x20, [x14, #0xf8]\n"
+      "ldr x20, [x12, #0xf8]\n"
       ".inst 0x4fa2e15a  // sdot v26.4s, v10.16b, v2.4b[1]\n"
-      "ldr d30, [x14, #0xd0]\n"
+      "ldr d30, [x12, #0xd0]\n"
       ".inst 0x4fa0e3b3  // sdot v19.4s, v29.16b, v0.4b[1]\n"
       "mov v31.d[1], x23\n"
       ".inst 0x4fa1e3b7  // sdot v23.4s, v29.16b, v1.4b[1]\n"
-      "add x10, x10, #0x10\n"
+      "mov v30.d[1], x22\n"
       ".inst 0x4fa2e3bb  // sdot v27.4s, v29.16b, v2.4b[1]\n"
-      "ldr d29, [x14, #0xe0]\n"
+      "ldr d29, [x12, #0xe0]\n"
       ".inst 0x4f80eb90  // sdot v16.4s, v28.16b, v0.4b[2]\n"
-      "mov v30.d[1], x22\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81eb94  // sdot v20.4s, v28.16b, v1.4b[2]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x4f82eb98  // sdot v24.4s, v28.16b, v2.4b[2]\n"
-      "ldr d28, [x14, #0xf0]\n"
+      "ldr d28, [x12, #0xf0]\n"
       ".inst 0x4f80e8b1  // sdot v17.4s, v5.16b, v0.4b[2]\n"
-      "mov v29.d[1], x21\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4f81e8b5  // sdot v21.4s, v5.16b, v1.4b[2]\n"
       "add x28, x28, #0x10\n"
       ".inst 0x4f82e8b9  // sdot v25.4s, v5.16b, v2.4b[2]\n"
-      "mov v28.d[1], x20\n"
+      "add x27, x27, #0x10\n"
       ".inst 0x4f80e892  // sdot v18.4s, v4.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x4f81e896  // sdot v22.4s, v4.16b, v1.4b[2]\n"
       ".inst 0x4f82e89a  // sdot v26.4s, v4.16b, v2.4b[2]\n"
       ".inst 0x4f80e873  // sdot v19.4s, v3.16b, v0.4b[2]\n"
@@ -971,65 +971,65 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       "68:"  // Height 3: Multiply loop: unique 9: skip row sum
-      "ldr q0, [x10, #0x0]\n"
-      "sub x11, x11, #0x10\n"
-      "ldr q1, [x9, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q2, [x28, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "sub x10, x10, #0x10\n"
+      "ldr q1, [x28, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q2, [x27, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "bge 67b\n"
       "69:"  // Height 3: Multiply loop: Single iteration only
       ".inst 0x4f80e090  // sdot v16.4s, v4.16b, v0.4b[0]\n"
-      "sub x11, x11, #0x10\n"
+      "sub x10, x10, #0x10\n"
       ".inst 0x4f81e094  // sdot v20.4s, v4.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x4f82e098  // sdot v24.4s, v4.16b, v2.4b[0]\n"
-      "ldr q29, [x14, #0x70]\n"
+      "ldr q29, [x12, #0x70]\n"
       ".inst 0x4f80e0b1  // sdot v17.4s, v5.16b, v0.4b[0]\n"
-      "add x9, x9, #0x10\n"
-      ".inst 0x4f81e0b5  // sdot v21.4s, v5.16b, v1.4b[0]\n"
       "add x28, x28, #0x10\n"
+      ".inst 0x4f81e0b5  // sdot v21.4s, v5.16b, v1.4b[0]\n"
+      "add x27, x27, #0x10\n"
       ".inst 0x4f82e0b9  // sdot v25.4s, v5.16b, v2.4b[0]\n"
-      "ldr q28, [x14, #0x80]\n"
+      "ldr q28, [x12, #0x80]\n"
       ".inst 0x4f80e0d2  // sdot v18.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0d6  // sdot v22.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x4f82e0da  // sdot v26.4s, v6.16b, v2.4b[0]\n"
-      "ldr q5, [x14, #0x90]\n"
+      "ldr q5, [x12, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0f7  // sdot v23.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0fb  // sdot v27.4s, v7.16b, v2.4b[0]\n"
-      "ldr q4, [x14, #0xa0]\n"
+      "ldr q4, [x12, #0xa0]\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
       ".inst 0x4fa1e114  // sdot v20.4s, v8.16b, v1.4b[1]\n"
       ".inst 0x4fa2e118  // sdot v24.4s, v8.16b, v2.4b[1]\n"
-      "ldr q3, [x14, #0xb0]\n"
+      "ldr q3, [x12, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
       ".inst 0x4fa1e135  // sdot v21.4s, v9.16b, v1.4b[1]\n"
       ".inst 0x4fa2e139  // sdot v25.4s, v9.16b, v2.4b[1]\n"
-      "ldr q31, [x14, #0xc0]\n"
+      "ldr q31, [x12, #0xc0]\n"
       ".inst 0x4fa0e152  // sdot v18.4s, v10.16b, v0.4b[1]\n"
       ".inst 0x4fa1e156  // sdot v22.4s, v10.16b, v1.4b[1]\n"
       ".inst 0x4fa2e15a  // sdot v26.4s, v10.16b, v2.4b[1]\n"
-      "ldr q30, [x14, #0xd0]\n"
+      "ldr q30, [x12, #0xd0]\n"
       ".inst 0x4fa0e3b3  // sdot v19.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3b7  // sdot v23.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3bb  // sdot v27.4s, v29.16b, v2.4b[1]\n"
-      "ldr q29, [x14, #0xe0]\n"
+      "ldr q29, [x12, #0xe0]\n"
       ".inst 0x4f80eb90  // sdot v16.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb94  // sdot v20.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb98  // sdot v24.4s, v28.16b, v2.4b[2]\n"
-      "ldr q28, [x14, #0xf0]\n"
+      "ldr q28, [x12, #0xf0]\n"
       ".inst 0x4f80e8b1  // sdot v17.4s, v5.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x4f81e8b5  // sdot v21.4s, v5.16b, v1.4b[2]\n"
       ".inst 0x4f82e8b9  // sdot v25.4s, v5.16b, v2.4b[2]\n"
       ".inst 0x4f80e892  // sdot v18.4s, v4.16b, v0.4b[2]\n"
@@ -1055,32 +1055,32 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       "70:"  // Height 3: Multiply loop: unique 10: skip row sum
-      "prfm pldl1keep, [x10, #0x80]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "71:"  // Height 3: Multiply loop: Main loop skip
-      "cbz x11, 78f\n"
-      "cmp x11, #0x4\n"
+      "cbz x10, 78f\n"
+      "cmp x10, #0x4\n"
       "blt 74f\n"
       "72:"  // Height 3: Multiply loop: Odd block loop
-      "ldr s0, [x10], #0x4\n"
-      "ldr s1, [x9], #0x4\n"
-      "ldr s2, [x28], #0x4\n"
+      "ldr s0, [x9], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr s2, [x27], #0x4\n"
       "tbnz %x[flags], #31, 73f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       "73:"  // Height 3: Multiply loop: unique 11: skip row sum
-      "ldr q31, [x14, #0x0]\n"
-      "sub x11, x11, #0x4\n"
-      "ldr q30, [x14, #0x10]\n"
-      "cmp x11, #0x4\n"
-      "ldr q29, [x14, #0x20]\n"
-      "ldr q28, [x14, #0x30]\n"
+      "ldr q31, [x12, #0x0]\n"
+      "sub x10, x10, #0x4\n"
+      "ldr q30, [x12, #0x10]\n"
+      "cmp x10, #0x4\n"
+      "ldr q29, [x12, #0x20]\n"
       ".inst 0x4f80e3f0  // sdot v16.4s, v31.16b, v0.4b[0]\n"
+      "ldr q28, [x12, #0x30]\n"
       ".inst 0x4f81e3f4  // sdot v20.4s, v31.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
       ".inst 0x4f82e3f8  // sdot v24.4s, v31.16b, v2.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x4f80e3d1  // sdot v17.4s, v30.16b, v0.4b[0]\n"
       ".inst 0x4f81e3d5  // sdot v21.4s, v30.16b, v1.4b[0]\n"
       ".inst 0x4f82e3d9  // sdot v25.4s, v30.16b, v2.4b[0]\n"
@@ -1092,36 +1092,36 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4f82e39b  // sdot v27.4s, v28.16b, v2.4b[0]\n"
       "bge 72b\n"
       "74:"  // Height 3: Multiply loop: Skip odd blocks
-      "cbz x11, 78f\n"
-      "tbz x11, #1, 75f\n"
-      "ldr h0, [x10], #0x2\n"
-      "ldr h1, [x9], #0x2\n"
-      "ldr h2, [x28], #0x2\n"
-      "tbz x11, #0, 76f\n"
-      "ld1 { v0.b }[2], [x10]\n"
-      "ld1 { v1.b }[2], [x9]\n"
-      "ld1 { v2.b }[2], [x28]\n"
+      "cbz x10, 78f\n"
+      "tbz x10, #1, 75f\n"
+      "ldr h0, [x9], #0x2\n"
+      "ldr h1, [x28], #0x2\n"
+      "ldr h2, [x27], #0x2\n"
+      "tbz x10, #0, 76f\n"
+      "ld1 { v0.b }[2], [x9]\n"
+      "ld1 { v1.b }[2], [x28]\n"
+      "ld1 { v2.b }[2], [x27]\n"
       "b 76f\n"
       "75:"  // Height 3: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x10, #0x0]\n"
-      "ldr b1, [x9, #0x0]\n"
-      "ldr b2, [x28, #0x0]\n"
+      "ldr b0, [x9, #0x0]\n"
+      "ldr b1, [x28, #0x0]\n"
+      "ldr b2, [x27, #0x0]\n"
       "76:"  // Height 3: Multiply loop: Ragged operand read: Done
       "tbnz %x[flags], #31, 77f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       "77:"  // Height 3: Multiply loop: unique 12: skip row sum
-      "ldr q31, [x14, #0x0]\n"
-      "ldr q30, [x14, #0x10]\n"
-      "ldr q29, [x14, #0x20]\n"
-      "ldr q28, [x14, #0x30]\n"
-      ".inst 0x4f80e3f0  // sdot v16.4s, v31.16b, v0.4b[0]\n"
-      ".inst 0x4f81e3f4  // sdot v20.4s, v31.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
-      ".inst 0x4f82e3f8  // sdot v24.4s, v31.16b, v2.4b[0]\n"
+      "ldr q28, [x12, #0x0]\n"
+      ".inst 0x4f80e390  // sdot v16.4s, v28.16b, v0.4b[0]\n"
+      "ldr q30, [x12, #0x10]\n"
+      ".inst 0x4f81e394  // sdot v20.4s, v28.16b, v1.4b[0]\n"
+      "ldr q29, [x12, #0x20]\n"
+      ".inst 0x4f82e398  // sdot v24.4s, v28.16b, v2.4b[0]\n"
+      "ldr q28, [x12, #0x30]\n"
       ".inst 0x4f80e3d1  // sdot v17.4s, v30.16b, v0.4b[0]\n"
       ".inst 0x4f81e3d5  // sdot v21.4s, v30.16b, v1.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x4f82e3d9  // sdot v25.4s, v30.16b, v2.4b[0]\n"
       ".inst 0x4f80e3b2  // sdot v18.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f81e3b6  // sdot v22.4s, v29.16b, v1.4b[0]\n"
@@ -1131,15 +1131,15 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4f82e39b  // sdot v27.4s, v28.16b, v2.4b[0]\n"
       "78:"  // Height 3: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x12, x12, #0x1\n"
-      "cmp x12, x20\n"
+      "add x11, x11, #0x1\n"
+      "cmp x11, x20\n"
       "bne 64b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20\n"
+      "add x22, x23, x20\n"
       "prfm pstl1keep, [x13, #0x0]\n"
-      "add x24, x13, x20\n"
-      "add x23, x24, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbnz %x[flags], #31, 79f\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
@@ -1154,13 +1154,13 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "mul v12.4s, v12.4s, v28.4s\n"
       "mul v13.4s, v13.4s, v28.4s\n"
       "79:"  // Height 3: skip row sum fixup
-      "ldr q31, [x16, #0x0]\n"
+      "ldr q31, [x15, #0x0]\n"
       "add v16.4s, v16.4s, v11.4s\n"
-      "ldr q30, [x16, #0x10]\n"
+      "ldr q30, [x15, #0x10]\n"
       "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q29, [x16, #0x20]\n"
+      "ldr q29, [x15, #0x20]\n"
       "add v18.4s, v18.4s, v11.4s\n"
-      "ldr q28, [x16, #0x30]\n"
+      "ldr q28, [x15, #0x30]\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
@@ -1183,11 +1183,10 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "add v26.4s, v26.4s, v29.4s\n"
       "add v27.4s, v27.4s, v28.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v28.4s }, [x20]\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add x16, x16, #0x40\n"
       "sqrdmulh v16.4s, v16.4s, v28.4s\n"
       "sqrdmulh v17.4s, v17.4s, v28.4s\n"
       "sqrdmulh v18.4s, v18.4s, v28.4s\n"
@@ -1200,38 +1199,39 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "sqrdmulh v25.4s, v25.4s, v28.4s\n"
       "sqrdmulh v26.4s, v26.4s, v28.4s\n"
       "sqrdmulh v27.4s, v27.4s, v28.4s\n"
+      "add x15, x15, #0x40\n"
       "tbz %x[flags], #5, 80f\n"
       "and v1.16b, v16.16b, v0.16b\n"
       "and v31.16b, v17.16b, v0.16b\n"
       "and v30.16b, v18.16b, v0.16b\n"
       "and v29.16b, v19.16b, v0.16b\n"
       "and v28.16b, v20.16b, v0.16b\n"
-      "and v3.16b, v21.16b, v0.16b\n"
-      "and v2.16b, v22.16b, v0.16b\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
-      "sshr v3.4s, v3.4s, #0x1f\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
       "sqadd v16.4s, v16.4s, v1.4s\n"
       "sqadd v17.4s, v17.4s, v31.4s\n"
       "sqadd v18.4s, v18.4s, v30.4s\n"
       "sqadd v19.4s, v19.4s, v29.4s\n"
       "sqadd v20.4s, v20.4s, v28.4s\n"
+      "and v3.16b, v21.16b, v0.16b\n"
+      "and v2.16b, v22.16b, v0.16b\n"
       "and v1.16b, v23.16b, v0.16b\n"
       "and v31.16b, v24.16b, v0.16b\n"
       "and v30.16b, v25.16b, v0.16b\n"
       "and v29.16b, v26.16b, v0.16b\n"
       "and v28.16b, v27.16b, v0.16b\n"
-      "sqadd v21.4s, v21.4s, v3.4s\n"
-      "sqadd v22.4s, v22.4s, v2.4s\n"
+      "sshr v3.4s, v3.4s, #0x1f\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
+      "sqadd v21.4s, v21.4s, v3.4s\n"
+      "sqadd v22.4s, v22.4s, v2.4s\n"
       "sqadd v23.4s, v23.4s, v1.4s\n"
       "sqadd v24.4s, v24.4s, v31.4s\n"
       "sqadd v25.4s, v25.4s, v30.4s\n"
@@ -1251,36 +1251,35 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "srshl v26.4s, v26.4s, v0.4s\n"
       "srshl v27.4s, v27.4s, v0.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v30.4s }, [x20]\n"
+      "ld1r { v28.4s }, [x20]\n"
+      "add v16.4s, v16.4s, v28.4s\n"
+      "add v17.4s, v17.4s, v28.4s\n"
+      "add v18.4s, v18.4s, v28.4s\n"
+      "add v19.4s, v19.4s, v28.4s\n"
+      "add v20.4s, v20.4s, v28.4s\n"
+      "add v21.4s, v21.4s, v28.4s\n"
+      "add v22.4s, v22.4s, v28.4s\n"
+      "add v23.4s, v23.4s, v28.4s\n"
+      "add v24.4s, v24.4s, v28.4s\n"
+      "add v25.4s, v25.4s, v28.4s\n"
+      "add v26.4s, v26.4s, v28.4s\n"
+      "add v27.4s, v27.4s, v28.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v28.4s }, [x20]\n"
+      "smin v16.4s, v16.4s, v28.4s\n"
+      "smin v17.4s, v17.4s, v28.4s\n"
+      "smin v18.4s, v18.4s, v28.4s\n"
+      "smin v19.4s, v19.4s, v28.4s\n"
+      "smin v20.4s, v20.4s, v28.4s\n"
+      "smin v21.4s, v21.4s, v28.4s\n"
+      "smin v22.4s, v22.4s, v28.4s\n"
+      "smin v23.4s, v23.4s, v28.4s\n"
+      "smin v24.4s, v24.4s, v28.4s\n"
+      "smin v25.4s, v25.4s, v28.4s\n"
+      "smin v26.4s, v26.4s, v28.4s\n"
+      "smin v27.4s, v27.4s, v28.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v29.4s }, [x21]\n"
-      "cmp x15, #0x10\n"
       "ld1r { v28.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v30.4s\n"
-      "add v17.4s, v17.4s, v30.4s\n"
-      "add v18.4s, v18.4s, v30.4s\n"
-      "add v19.4s, v19.4s, v30.4s\n"
-      "add v20.4s, v20.4s, v30.4s\n"
-      "add v21.4s, v21.4s, v30.4s\n"
-      "add v22.4s, v22.4s, v30.4s\n"
-      "add v23.4s, v23.4s, v30.4s\n"
-      "add v24.4s, v24.4s, v30.4s\n"
-      "add v25.4s, v25.4s, v30.4s\n"
-      "add v26.4s, v26.4s, v30.4s\n"
-      "add v27.4s, v27.4s, v30.4s\n"
-      "smin v16.4s, v16.4s, v29.4s\n"
-      "smin v17.4s, v17.4s, v29.4s\n"
-      "smin v18.4s, v18.4s, v29.4s\n"
-      "smin v19.4s, v19.4s, v29.4s\n"
-      "smin v20.4s, v20.4s, v29.4s\n"
-      "smin v21.4s, v21.4s, v29.4s\n"
-      "smin v22.4s, v22.4s, v29.4s\n"
-      "smin v23.4s, v23.4s, v29.4s\n"
-      "smin v24.4s, v24.4s, v29.4s\n"
-      "smin v25.4s, v25.4s, v29.4s\n"
-      "smin v26.4s, v26.4s, v29.4s\n"
-      "smin v27.4s, v27.4s, v29.4s\n"
       "smax v16.4s, v16.4s, v28.4s\n"
       "smax v17.4s, v17.4s, v28.4s\n"
       "smax v18.4s, v18.4s, v28.4s\n"
@@ -1299,109 +1298,109 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "uzp1 v18.8h, v22.8h, v23.8h\n"
       "uzp1 v24.8h, v24.8h, v25.8h\n"
       "uzp1 v17.8h, v26.8h, v27.8h\n"
+      "cmp x14, #0x10\n"
       "uzp1 v16.16b, v16.16b, v19.16b\n"
       "uzp1 v20.16b, v20.16b, v18.16b\n"
       "uzp1 v24.16b, v24.16b, v17.16b\n"
       "bge 89f\n"
-      "tbz x15, #3, 84f\n"
+      "tbz x14, #3, 84f\n"
       "str d16, [x13], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "tbz x15, #2, 82f\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "tbz x14, #2, 82f\n"
       "st1 { v16.s }[2], [x13], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "tbz x15, #1, 81f\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "tbz x14, #1, 81f\n"
       "st1 { v16.h }[6], [x13], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "tbz x15, #0, 88f\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[14], [x13]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 88f\n"
       "81:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x15, #0, 88f\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[12], [x13]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 88f\n"
       "82:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x15, #1, 83f\n"
+      "tbz x14, #1, 83f\n"
       "st1 { v16.h }[4], [x13], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "tbz x15, #0, 88f\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[10], [x13]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 88f\n"
       "83:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x15, #0, 88f\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[8], [x13]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 88f\n"
       "84:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x15, #2, 86f\n"
+      "tbz x14, #2, 86f\n"
       "str s16, [x13], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "tbz x15, #1, 85f\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "tbz x14, #1, 85f\n"
       "st1 { v16.h }[2], [x13], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "tbz x15, #0, 88f\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[6], [x13]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 88f\n"
       "85:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x15, #0, 88f\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[4], [x13]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 88f\n"
       "86:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x15, #1, 87f\n"
+      "tbz x14, #1, 87f\n"
       "str h16, [x13], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "tbz x15, #0, 88f\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[2], [x13]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 88f\n"
       "87:"  // Height 3: Partial direct writeback: partial_1_0
       "str b16, [x13, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "88:"  // Height 3: Partial direct writeback: Done
       "b 90f\n"
       "89:"  // Height 3: Full writeback
       "str q16, [x13, #0x0]\n"
       "add x13, x13, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "90:"  // Height 3: Writeback done
-      "subs x15, x15, #0x10\n"
+      "subs x14, x14, #0x10\n"
       "bgt 62b\n"
       "b 122f\n"
       "91:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "mov x20, #0x4\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x16, %x[col_bias]\n"
+      "mov x15, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
       "movi v12.4s, #0x0\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_N]]\n"
-      "madd x20, x21, x20, x13\n"
+      "bic %x[flags], %x[flags], #0x80000000\n"
       "movi v13.4s, #0x0\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "movi v14.4s, #0x0\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "movi v15.16b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "92:"  // Height 4: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -1420,118 +1419,118 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "movi v30.4s, #0x0\n"
       "movi v31.4s, #0x0\n"
       "93:"  // Height 4: setup done
-      "mov x12, #0x0\n"
+      "mov x11, #0x0\n"
       "94:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w10, [x20, x11, LSL #0x2]\n"
       "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w11, [x20, x12, LSL #0x2]\n"
       "tbz %x[flags], #3, 95f\n"
-      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
-      "ldr x10, [x20, #0x0]\n"
-      "ldr x9, [x20, #0x8]\n"
-      "ldr x28, [x20, #0x10]\n"
-      "ldr x27, [x20, #0x18]\n"
-      "cbnz x12, 96f\n"
+      "ldr x9, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "ldr x27, [x20, #0x10]\n"
+      "ldr x26, [x20, #0x18]\n"
+      "cbnz x11, 96f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x10, x10, x20\n"
       "add x9, x9, x20\n"
       "add x28, x28, x20\n"
       "add x27, x27, x20\n"
+      "add x26, x26, x20\n"
       "b 96f\n"
       "95:"  // Height 4: setup direct input
-      "mov x10, %x[input_ptr]\n"
-      "add x9, x10, x21\n"
+      "mov x9, %x[input_ptr]\n"
       "add x28, x9, x21\n"
       "add x27, x28, x21\n"
+      "add x26, x27, x21\n"
       "96:"  // Height 4: input setup done
-      "cmp x11, #0x10\n"
+      "cmp x10, #0x10\n"
       "blt 101f\n"
-      "ldr q0, [x10, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q1, [x9, #0x0]\n"
-      "ldr q2, [x28, #0x0]\n"
-      "ldr q3, [x27, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x27, #0x0]\n"
+      "ldr q3, [x26, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "blt 99f\n"
       "97:"  // Height 4: Multiply loop: Main loop head
       ".inst 0x4f80e090  // sdot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr x21, [x14, #0x78]\n"
+      "ldr x22, [x12, #0x78]\n"
       ".inst 0x4f81e094  // sdot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr x20, [x14, #0x88]\n"
+      "ldr x21, [x12, #0x88]\n"
       ".inst 0x4f82e098  // sdot v24.4s, v4.16b, v2.4b[0]\n"
-      "ldr x26, [x14, #0x98]\n"
+      "ldr x20, [x12, #0x98]\n"
       ".inst 0x4f83e09c  // sdot v28.4s, v4.16b, v3.4b[0]\n"
-      "ldr d4, [x14, #0x70]\n"
+      "ldr d4, [x12, #0x70]\n"
       ".inst 0x4f80e0b1  // sdot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr x25, [x14, #0xa8]\n"
+      "mov v4.d[1], x22\n"
       ".inst 0x4f81e0b5  // sdot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr x24, [x14, #0xb8]\n"
+      "ldr x25, [x12, #0xa8]\n"
       ".inst 0x4f82e0b9  // sdot v25.4s, v5.16b, v2.4b[0]\n"
-      "mov v4.d[1], x21\n"
+      "ldr x24, [x12, #0xb8]\n"
       ".inst 0x4f83e0bd  // sdot v29.4s, v5.16b, v3.4b[0]\n"
-      "ldr d5, [x14, #0x80]\n"
+      "ldr d5, [x12, #0x80]\n"
       ".inst 0x4f80e0d2  // sdot v18.4s, v6.16b, v0.4b[0]\n"
-      "ldr x23, [x14, #0xc8]\n"
+      "mov v5.d[1], x21\n"
       ".inst 0x4f81e0d6  // sdot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr x22, [x14, #0xd8]\n"
+      "ldr x23, [x12, #0xc8]\n"
       ".inst 0x4f82e0da  // sdot v26.4s, v6.16b, v2.4b[0]\n"
-      "mov v5.d[1], x20\n"
+      "ldr x22, [x12, #0xd8]\n"
       ".inst 0x4f83e0de  // sdot v30.4s, v6.16b, v3.4b[0]\n"
-      "ldr d6, [x14, #0x90]\n"
+      "ldr d6, [x12, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr x21, [x14, #0xe8]\n"
+      "mov v6.d[1], x20\n"
       ".inst 0x4f81e0f7  // sdot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr x20, [x14, #0xf8]\n"
+      "ldr x21, [x12, #0xe8]\n"
       ".inst 0x4f82e0fb  // sdot v27.4s, v7.16b, v2.4b[0]\n"
-      "mov v6.d[1], x26\n"
+      "ldr x20, [x12, #0xf8]\n"
       ".inst 0x4f83e0ff  // sdot v31.4s, v7.16b, v3.4b[0]\n"
-      "ldr d7, [x14, #0xa0]\n"
+      "ldr d7, [x12, #0xa0]\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
-      "add x10, x10, #0x10\n"
+      "mov v7.d[1], x25\n"
       ".inst 0x4fa1e114  // sdot v20.4s, v8.16b, v1.4b[1]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x4fa2e118  // sdot v24.4s, v8.16b, v2.4b[1]\n"
-      "mov v7.d[1], x25\n"
+      "add x28, x28, #0x10\n"
       ".inst 0x4fa3e11c  // sdot v28.4s, v8.16b, v3.4b[1]\n"
-      "ldr d8, [x14, #0xb0]\n"
+      "ldr d8, [x12, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
-      "add x28, x28, #0x10\n"
+      "mov v8.d[1], x24\n"
       ".inst 0x4fa1e135  // sdot v21.4s, v9.16b, v1.4b[1]\n"
       "add x27, x27, #0x10\n"
       ".inst 0x4fa2e139  // sdot v25.4s, v9.16b, v2.4b[1]\n"
-      "mov v8.d[1], x24\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4fa3e13d  // sdot v29.4s, v9.16b, v3.4b[1]\n"
-      "ldr d9, [x14, #0xc0]\n"
+      "ldr d9, [x12, #0xc0]\n"
       ".inst 0x4fa0e152  // sdot v18.4s, v10.16b, v0.4b[1]\n"
+      "mov v9.d[1], x23\n"
       ".inst 0x4fa1e156  // sdot v22.4s, v10.16b, v1.4b[1]\n"
       ".inst 0x4fa2e15a  // sdot v26.4s, v10.16b, v2.4b[1]\n"
-      "mov v9.d[1], x23\n"
       ".inst 0x4fa3e15e  // sdot v30.4s, v10.16b, v3.4b[1]\n"
-      "ldr d10, [x14, #0xd0]\n"
+      "ldr d10, [x12, #0xd0]\n"
       ".inst 0x4fa0e093  // sdot v19.4s, v4.16b, v0.4b[1]\n"
+      "mov v10.d[1], x22\n"
       ".inst 0x4fa1e097  // sdot v23.4s, v4.16b, v1.4b[1]\n"
       ".inst 0x4fa2e09b  // sdot v27.4s, v4.16b, v2.4b[1]\n"
-      "mov v10.d[1], x22\n"
       ".inst 0x4fa3e09f  // sdot v31.4s, v4.16b, v3.4b[1]\n"
-      "ldr d4, [x14, #0xe0]\n"
+      "ldr d4, [x12, #0xe0]\n"
       ".inst 0x4f80e8b0  // sdot v16.4s, v5.16b, v0.4b[2]\n"
+      "mov v4.d[1], x21\n"
       ".inst 0x4f81e8b4  // sdot v20.4s, v5.16b, v1.4b[2]\n"
       ".inst 0x4f82e8b8  // sdot v24.4s, v5.16b, v2.4b[2]\n"
-      "mov v4.d[1], x21\n"
       ".inst 0x4f83e8bc  // sdot v28.4s, v5.16b, v3.4b[2]\n"
-      "ldr d5, [x14, #0xf0]\n"
+      "ldr d5, [x12, #0xf0]\n"
       ".inst 0x4f80e8d1  // sdot v17.4s, v6.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "mov v5.d[1], x20\n"
       ".inst 0x4f81e8d5  // sdot v21.4s, v6.16b, v1.4b[2]\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x4f82e8d9  // sdot v25.4s, v6.16b, v2.4b[2]\n"
-      "mov v5.d[1], x20\n"
       ".inst 0x4f83e8dd  // sdot v29.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f80e8f2  // sdot v18.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8f6  // sdot v22.4s, v7.16b, v1.4b[2]\n"
@@ -1563,77 +1562,77 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       ".inst 0x4e8f946e  // sdot v14.4s, v3.16b, v15.16b\n"
       "98:"  // Height 4: Multiply loop: unique 13: skip row sum
-      "ldr q0, [x10, #0x0]\n"
-      "sub x11, x11, #0x10\n"
-      "ldr q1, [x9, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q2, [x28, #0x0]\n"
-      "ldr q3, [x27, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "sub x10, x10, #0x10\n"
+      "ldr q1, [x28, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q2, [x27, #0x0]\n"
+      "ldr q3, [x26, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
       "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 97b\n"
       "99:"  // Height 4: Multiply loop: Single iteration only
       ".inst 0x4f80e090  // sdot v16.4s, v4.16b, v0.4b[0]\n"
-      "sub x11, x11, #0x10\n"
+      "sub x10, x10, #0x10\n"
       ".inst 0x4f81e094  // sdot v20.4s, v4.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
-      ".inst 0x4f82e098  // sdot v24.4s, v4.16b, v2.4b[0]\n"
       "add x9, x9, #0x10\n"
+      ".inst 0x4f82e098  // sdot v24.4s, v4.16b, v2.4b[0]\n"
+      "add x28, x28, #0x10\n"
       ".inst 0x4f83e09c  // sdot v28.4s, v4.16b, v3.4b[0]\n"
-      "ldr q4, [x14, #0x70]\n"
+      "ldr q4, [x12, #0x70]\n"
       ".inst 0x4f80e0b1  // sdot v17.4s, v5.16b, v0.4b[0]\n"
-      "add x28, x28, #0x10\n"
-      ".inst 0x4f81e0b5  // sdot v21.4s, v5.16b, v1.4b[0]\n"
       "add x27, x27, #0x10\n"
+      ".inst 0x4f81e0b5  // sdot v21.4s, v5.16b, v1.4b[0]\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4f82e0b9  // sdot v25.4s, v5.16b, v2.4b[0]\n"
       ".inst 0x4f83e0bd  // sdot v29.4s, v5.16b, v3.4b[0]\n"
-      "ldr q5, [x14, #0x80]\n"
+      "ldr q5, [x12, #0x80]\n"
       ".inst 0x4f80e0d2  // sdot v18.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0d6  // sdot v22.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x4f82e0da  // sdot v26.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0de  // sdot v30.4s, v6.16b, v3.4b[0]\n"
-      "ldr q6, [x14, #0x90]\n"
+      "ldr q6, [x12, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0f7  // sdot v23.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0fb  // sdot v27.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0ff  // sdot v31.4s, v7.16b, v3.4b[0]\n"
-      "ldr q7, [x14, #0xa0]\n"
+      "ldr q7, [x12, #0xa0]\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
       ".inst 0x4fa1e114  // sdot v20.4s, v8.16b, v1.4b[1]\n"
       ".inst 0x4fa2e118  // sdot v24.4s, v8.16b, v2.4b[1]\n"
       ".inst 0x4fa3e11c  // sdot v28.4s, v8.16b, v3.4b[1]\n"
-      "ldr q8, [x14, #0xb0]\n"
+      "ldr q8, [x12, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
       ".inst 0x4fa1e135  // sdot v21.4s, v9.16b, v1.4b[1]\n"
       ".inst 0x4fa2e139  // sdot v25.4s, v9.16b, v2.4b[1]\n"
       ".inst 0x4fa3e13d  // sdot v29.4s, v9.16b, v3.4b[1]\n"
-      "ldr q9, [x14, #0xc0]\n"
+      "ldr q9, [x12, #0xc0]\n"
       ".inst 0x4fa0e152  // sdot v18.4s, v10.16b, v0.4b[1]\n"
       ".inst 0x4fa1e156  // sdot v22.4s, v10.16b, v1.4b[1]\n"
       ".inst 0x4fa2e15a  // sdot v26.4s, v10.16b, v2.4b[1]\n"
       ".inst 0x4fa3e15e  // sdot v30.4s, v10.16b, v3.4b[1]\n"
-      "ldr q10, [x14, #0xd0]\n"
+      "ldr q10, [x12, #0xd0]\n"
       ".inst 0x4fa0e093  // sdot v19.4s, v4.16b, v0.4b[1]\n"
       ".inst 0x4fa1e097  // sdot v23.4s, v4.16b, v1.4b[1]\n"
       ".inst 0x4fa2e09b  // sdot v27.4s, v4.16b, v2.4b[1]\n"
       ".inst 0x4fa3e09f  // sdot v31.4s, v4.16b, v3.4b[1]\n"
-      "ldr q4, [x14, #0xe0]\n"
+      "ldr q4, [x12, #0xe0]\n"
       ".inst 0x4f80e8b0  // sdot v16.4s, v5.16b, v0.4b[2]\n"
       ".inst 0x4f81e8b4  // sdot v20.4s, v5.16b, v1.4b[2]\n"
       ".inst 0x4f82e8b8  // sdot v24.4s, v5.16b, v2.4b[2]\n"
       ".inst 0x4f83e8bc  // sdot v28.4s, v5.16b, v3.4b[2]\n"
-      "ldr q5, [x14, #0xf0]\n"
+      "ldr q5, [x12, #0xf0]\n"
       ".inst 0x4f80e8d1  // sdot v17.4s, v6.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x4f81e8d5  // sdot v21.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d9  // sdot v25.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8dd  // sdot v29.4s, v6.16b, v3.4b[2]\n"
@@ -1667,35 +1666,35 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       ".inst 0x4e8f946e  // sdot v14.4s, v3.16b, v15.16b\n"
       "100:"  // Height 4: Multiply loop: unique 14: skip row sum
-      "prfm pldl1keep, [x10, #0x80]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
       "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "101:"  // Height 4: Multiply loop: Main loop skip
-      "cbz x11, 108f\n"
-      "cmp x11, #0x4\n"
+      "cbz x10, 108f\n"
+      "cmp x10, #0x4\n"
       "blt 104f\n"
       "102:"  // Height 4: Multiply loop: Odd block loop
-      "ldr s0, [x10], #0x4\n"
-      "ldr s1, [x9], #0x4\n"
-      "ldr s2, [x28], #0x4\n"
-      "ldr s3, [x27], #0x4\n"
+      "ldr s0, [x9], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr s2, [x27], #0x4\n"
+      "ldr s3, [x26], #0x4\n"
       "tbnz %x[flags], #31, 103f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       ".inst 0x4e8f946e  // sdot v14.4s, v3.16b, v15.16b\n"
       "103:"  // Height 4: Multiply loop: unique 15: skip row sum
-      "ldr q7, [x14, #0x0]\n"
-      "sub x11, x11, #0x4\n"
-      "ldr q6, [x14, #0x10]\n"
-      "cmp x11, #0x4\n"
-      "ldr q5, [x14, #0x20]\n"
-      "ldr q4, [x14, #0x30]\n"
+      "ldr q7, [x12, #0x0]\n"
+      "sub x10, x10, #0x4\n"
+      "ldr q6, [x12, #0x10]\n"
+      "cmp x10, #0x4\n"
+      "ldr q5, [x12, #0x20]\n"
       ".inst 0x4f80e0f0  // sdot v16.4s, v7.16b, v0.4b[0]\n"
+      "ldr q4, [x12, #0x30]\n"
       ".inst 0x4f81e0f4  // sdot v20.4s, v7.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
       ".inst 0x4f82e0f8  // sdot v24.4s, v7.16b, v2.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x4f83e0fc  // sdot v28.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f80e0d1  // sdot v17.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0d5  // sdot v21.4s, v6.16b, v1.4b[0]\n"
@@ -1711,23 +1710,23 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4f83e09f  // sdot v31.4s, v4.16b, v3.4b[0]\n"
       "bge 102b\n"
       "104:"  // Height 4: Multiply loop: Skip odd blocks
-      "cbz x11, 108f\n"
-      "tbz x11, #1, 105f\n"
-      "ldr h0, [x10], #0x2\n"
-      "ldr h1, [x9], #0x2\n"
-      "ldr h2, [x28], #0x2\n"
-      "ldr h3, [x27], #0x2\n"
-      "tbz x11, #0, 106f\n"
-      "ld1 { v0.b }[2], [x10]\n"
-      "ld1 { v1.b }[2], [x9]\n"
-      "ld1 { v2.b }[2], [x28]\n"
-      "ld1 { v3.b }[2], [x27]\n"
+      "cbz x10, 108f\n"
+      "tbz x10, #1, 105f\n"
+      "ldr h0, [x9], #0x2\n"
+      "ldr h1, [x28], #0x2\n"
+      "ldr h2, [x27], #0x2\n"
+      "ldr h3, [x26], #0x2\n"
+      "tbz x10, #0, 106f\n"
+      "ld1 { v0.b }[2], [x9]\n"
+      "ld1 { v1.b }[2], [x28]\n"
+      "ld1 { v2.b }[2], [x27]\n"
+      "ld1 { v3.b }[2], [x26]\n"
       "b 106f\n"
       "105:"  // Height 4: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x10, #0x0]\n"
-      "ldr b1, [x9, #0x0]\n"
-      "ldr b2, [x28, #0x0]\n"
-      "ldr b3, [x27, #0x0]\n"
+      "ldr b0, [x9, #0x0]\n"
+      "ldr b1, [x28, #0x0]\n"
+      "ldr b2, [x27, #0x0]\n"
+      "ldr b3, [x26, #0x0]\n"
       "106:"  // Height 4: Multiply loop: Ragged operand read: Done
       "tbnz %x[flags], #31, 107f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
@@ -1735,16 +1734,16 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       ".inst 0x4e8f946e  // sdot v14.4s, v3.16b, v15.16b\n"
       "107:"  // Height 4: Multiply loop: unique 16: skip row sum
-      "ldr q7, [x14, #0x0]\n"
-      "ldr q6, [x14, #0x10]\n"
-      "ldr q5, [x14, #0x20]\n"
-      "ldr q4, [x14, #0x30]\n"
+      "ldr q7, [x12, #0x0]\n"
       ".inst 0x4f80e0f0  // sdot v16.4s, v7.16b, v0.4b[0]\n"
+      "ldr q6, [x12, #0x10]\n"
       ".inst 0x4f81e0f4  // sdot v20.4s, v7.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
+      "ldr q5, [x12, #0x20]\n"
       ".inst 0x4f82e0f8  // sdot v24.4s, v7.16b, v2.4b[0]\n"
+      "ldr q4, [x12, #0x30]\n"
       ".inst 0x4f83e0fc  // sdot v28.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f80e0d1  // sdot v17.4s, v6.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x4f81e0d5  // sdot v21.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x4f82e0d9  // sdot v25.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0dd  // sdot v29.4s, v6.16b, v3.4b[0]\n"
@@ -1758,17 +1757,17 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       ".inst 0x4f83e09f  // sdot v31.4s, v4.16b, v3.4b[0]\n"
       "108:"  // Height 4: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x12, x12, #0x1\n"
-      "cmp x12, x20\n"
+      "add x11, x11, #0x1\n"
+      "cmp x11, x20\n"
       "bne 94b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x13, #0x0]\n"
-      "add x24, x13, x20\n"
-      "add x23, x24, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "add x23, x13, x20\n"
       "add x22, x23, x20\n"
+      "add x21, x22, x20\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbnz %x[flags], #31, 109f\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
@@ -1776,9 +1775,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "addp v14.4s, v14.4s, v14.4s\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1r { v0.4s }, [x20]\n"
+      "neg v0.4s, v0.4s\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
-      "neg v0.4s, v0.4s\n"
       "addp v13.4s, v13.4s, v13.4s\n"
       "addp v14.4s, v14.4s, v14.4s\n"
       "mul v11.4s, v11.4s, v0.4s\n"
@@ -1786,13 +1785,13 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "mul v13.4s, v13.4s, v0.4s\n"
       "mul v14.4s, v14.4s, v0.4s\n"
       "109:"  // Height 4: skip row sum fixup
-      "ldr q3, [x16, #0x0]\n"
+      "ldr q3, [x15, #0x0]\n"
       "add v16.4s, v16.4s, v11.4s\n"
-      "ldr q2, [x16, #0x10]\n"
+      "ldr q2, [x15, #0x10]\n"
       "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q1, [x16, #0x20]\n"
+      "ldr q1, [x15, #0x20]\n"
       "add v18.4s, v18.4s, v11.4s\n"
-      "ldr q0, [x16, #0x30]\n"
+      "ldr q0, [x15, #0x30]\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
@@ -1823,11 +1822,10 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "add v30.4s, v30.4s, v1.4s\n"
       "add v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v1.4s }, [x20]\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add x16, x16, #0x40\n"
       "sqrdmulh v16.4s, v16.4s, v1.4s\n"
       "sqrdmulh v17.4s, v17.4s, v1.4s\n"
       "sqrdmulh v18.4s, v18.4s, v1.4s\n"
@@ -1844,51 +1842,52 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "sqrdmulh v29.4s, v29.4s, v1.4s\n"
       "sqrdmulh v30.4s, v30.4s, v1.4s\n"
       "sqrdmulh v31.4s, v31.4s, v1.4s\n"
+      "add x15, x15, #0x40\n"
       "tbz %x[flags], #5, 110f\n"
       "and v2.16b, v16.16b, v0.16b\n"
       "and v1.16b, v17.16b, v0.16b\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v16.4s, v16.4s, v2.4s\n"
+      "sqadd v17.4s, v17.4s, v1.4s\n"
       "and v7.16b, v18.16b, v0.16b\n"
       "and v6.16b, v19.16b, v0.16b\n"
       "and v5.16b, v20.16b, v0.16b\n"
       "and v4.16b, v21.16b, v0.16b\n"
       "and v3.16b, v22.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
+      "and v2.16b, v23.16b, v0.16b\n"
+      "and v1.16b, v24.16b, v0.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
       "sshr v6.4s, v6.4s, #0x1f\n"
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sshr v3.4s, v3.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v2.4s\n"
-      "sqadd v17.4s, v17.4s, v1.4s\n"
-      "and v2.16b, v23.16b, v0.16b\n"
-      "and v1.16b, v24.16b, v0.16b\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
       "sqadd v18.4s, v18.4s, v7.4s\n"
       "sqadd v19.4s, v19.4s, v6.4s\n"
       "sqadd v20.4s, v20.4s, v5.4s\n"
       "sqadd v21.4s, v21.4s, v4.4s\n"
       "sqadd v22.4s, v22.4s, v3.4s\n"
+      "sqadd v23.4s, v23.4s, v2.4s\n"
+      "sqadd v24.4s, v24.4s, v1.4s\n"
       "and v7.16b, v25.16b, v0.16b\n"
       "and v6.16b, v26.16b, v0.16b\n"
       "and v5.16b, v27.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
       "and v4.16b, v28.16b, v0.16b\n"
       "and v3.16b, v29.16b, v0.16b\n"
+      "and v2.16b, v30.16b, v0.16b\n"
+      "and v1.16b, v31.16b, v0.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
       "sshr v6.4s, v6.4s, #0x1f\n"
       "sshr v5.4s, v5.4s, #0x1f\n"
-      "sqadd v23.4s, v23.4s, v2.4s\n"
-      "sqadd v24.4s, v24.4s, v1.4s\n"
-      "and v2.16b, v30.16b, v0.16b\n"
-      "and v1.16b, v31.16b, v0.16b\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sshr v3.4s, v3.4s, #0x1f\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
       "sqadd v25.4s, v25.4s, v7.4s\n"
       "sqadd v26.4s, v26.4s, v6.4s\n"
       "sqadd v27.4s, v27.4s, v5.4s\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
       "sqadd v28.4s, v28.4s, v4.4s\n"
       "sqadd v29.4s, v29.4s, v3.4s\n"
       "sqadd v30.4s, v30.4s, v2.4s\n"
@@ -1911,44 +1910,43 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "srshl v30.4s, v30.4s, v0.4s\n"
       "srshl v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v2.4s }, [x20]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "add v16.4s, v16.4s, v0.4s\n"
+      "add v17.4s, v17.4s, v0.4s\n"
+      "add v18.4s, v18.4s, v0.4s\n"
+      "add v19.4s, v19.4s, v0.4s\n"
+      "add v20.4s, v20.4s, v0.4s\n"
+      "add v21.4s, v21.4s, v0.4s\n"
+      "add v22.4s, v22.4s, v0.4s\n"
+      "add v23.4s, v23.4s, v0.4s\n"
+      "add v24.4s, v24.4s, v0.4s\n"
+      "add v25.4s, v25.4s, v0.4s\n"
+      "add v26.4s, v26.4s, v0.4s\n"
+      "add v27.4s, v27.4s, v0.4s\n"
+      "add v28.4s, v28.4s, v0.4s\n"
+      "add v29.4s, v29.4s, v0.4s\n"
+      "add v30.4s, v30.4s, v0.4s\n"
+      "add v31.4s, v31.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "smin v16.4s, v16.4s, v0.4s\n"
+      "smin v17.4s, v17.4s, v0.4s\n"
+      "smin v18.4s, v18.4s, v0.4s\n"
+      "smin v19.4s, v19.4s, v0.4s\n"
+      "smin v20.4s, v20.4s, v0.4s\n"
+      "smin v21.4s, v21.4s, v0.4s\n"
+      "smin v22.4s, v22.4s, v0.4s\n"
+      "smin v23.4s, v23.4s, v0.4s\n"
+      "smin v24.4s, v24.4s, v0.4s\n"
+      "smin v25.4s, v25.4s, v0.4s\n"
+      "smin v26.4s, v26.4s, v0.4s\n"
+      "smin v27.4s, v27.4s, v0.4s\n"
+      "smin v28.4s, v28.4s, v0.4s\n"
+      "smin v29.4s, v29.4s, v0.4s\n"
+      "smin v30.4s, v30.4s, v0.4s\n"
+      "smin v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v1.4s }, [x21]\n"
-      "cmp x15, #0x10\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v2.4s\n"
-      "add v17.4s, v17.4s, v2.4s\n"
-      "add v18.4s, v18.4s, v2.4s\n"
-      "add v19.4s, v19.4s, v2.4s\n"
-      "add v20.4s, v20.4s, v2.4s\n"
-      "add v21.4s, v21.4s, v2.4s\n"
-      "add v22.4s, v22.4s, v2.4s\n"
-      "add v23.4s, v23.4s, v2.4s\n"
-      "add v24.4s, v24.4s, v2.4s\n"
-      "add v25.4s, v25.4s, v2.4s\n"
-      "add v26.4s, v26.4s, v2.4s\n"
-      "add v27.4s, v27.4s, v2.4s\n"
-      "add v28.4s, v28.4s, v2.4s\n"
-      "add v29.4s, v29.4s, v2.4s\n"
-      "add v30.4s, v30.4s, v2.4s\n"
-      "add v31.4s, v31.4s, v2.4s\n"
-      "smin v16.4s, v16.4s, v1.4s\n"
-      "smin v17.4s, v17.4s, v1.4s\n"
-      "smin v18.4s, v18.4s, v1.4s\n"
-      "smin v19.4s, v19.4s, v1.4s\n"
-      "smin v20.4s, v20.4s, v1.4s\n"
-      "smin v21.4s, v21.4s, v1.4s\n"
-      "smin v22.4s, v22.4s, v1.4s\n"
-      "smin v23.4s, v23.4s, v1.4s\n"
-      "smin v24.4s, v24.4s, v1.4s\n"
-      "smin v25.4s, v25.4s, v1.4s\n"
-      "smin v26.4s, v26.4s, v1.4s\n"
-      "smin v27.4s, v27.4s, v1.4s\n"
-      "smin v28.4s, v28.4s, v1.4s\n"
-      "smin v29.4s, v29.4s, v1.4s\n"
-      "smin v30.4s, v30.4s, v1.4s\n"
-      "smin v31.4s, v31.4s, v1.4s\n"
       "smax v16.4s, v16.4s, v0.4s\n"
       "smax v17.4s, v17.4s, v0.4s\n"
       "smax v18.4s, v18.4s, v0.4s\n"
@@ -1973,109 +1971,110 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "uzp1 v18.8h, v26.8h, v27.8h\n"
       "uzp1 v28.8h, v28.8h, v29.8h\n"
       "uzp1 v17.8h, v30.8h, v31.8h\n"
+      "cmp x14, #0x10\n"
       "uzp1 v16.16b, v16.16b, v0.16b\n"
       "uzp1 v20.16b, v20.16b, v19.16b\n"
       "uzp1 v24.16b, v24.16b, v18.16b\n"
       "uzp1 v28.16b, v28.16b, v17.16b\n"
       "bge 119f\n"
-      "tbz x15, #3, 114f\n"
+      "tbz x14, #3, 114f\n"
       "str d16, [x13], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
-      "tbz x15, #2, 112f\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
+      "tbz x14, #2, 112f\n"
       "st1 { v16.s }[2], [x13], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
-      "tbz x15, #1, 111f\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
+      "tbz x14, #1, 111f\n"
       "st1 { v16.h }[6], [x13], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "st1 { v28.h }[6], [x22], #0x2\n"
-      "tbz x15, #0, 118f\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "st1 { v28.h }[6], [x21], #0x2\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[14], [x13]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "st1 { v28.b }[14], [x22]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
+      "st1 { v28.b }[14], [x21]\n"
       "b 118f\n"
       "111:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x15, #0, 118f\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[12], [x13]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "st1 { v28.b }[12], [x22]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
+      "st1 { v28.b }[12], [x21]\n"
       "b 118f\n"
       "112:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x15, #1, 113f\n"
+      "tbz x14, #1, 113f\n"
       "st1 { v16.h }[4], [x13], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "st1 { v28.h }[4], [x22], #0x2\n"
-      "tbz x15, #0, 118f\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "st1 { v28.h }[4], [x21], #0x2\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[10], [x13]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "st1 { v28.b }[10], [x22]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
+      "st1 { v28.b }[10], [x21]\n"
       "b 118f\n"
       "113:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x15, #0, 118f\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[8], [x13]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "st1 { v28.b }[8], [x22]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
+      "st1 { v28.b }[8], [x21]\n"
       "b 118f\n"
       "114:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x15, #2, 116f\n"
+      "tbz x14, #2, 116f\n"
       "str s16, [x13], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
-      "tbz x15, #1, 115f\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
+      "tbz x14, #1, 115f\n"
       "st1 { v16.h }[2], [x13], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "st1 { v28.h }[2], [x22], #0x2\n"
-      "tbz x15, #0, 118f\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "st1 { v28.h }[2], [x21], #0x2\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[6], [x13]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "st1 { v28.b }[6], [x22]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
+      "st1 { v28.b }[6], [x21]\n"
       "b 118f\n"
       "115:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x15, #0, 118f\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[4], [x13]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "st1 { v28.b }[4], [x22]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
+      "st1 { v28.b }[4], [x21]\n"
       "b 118f\n"
       "116:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x15, #1, 117f\n"
+      "tbz x14, #1, 117f\n"
       "str h16, [x13], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "str h28, [x22], #0x2\n"
-      "tbz x15, #0, 118f\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "str h28, [x21], #0x2\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[2], [x13]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "st1 { v28.b }[2], [x22]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
+      "st1 { v28.b }[2], [x21]\n"
       "b 118f\n"
       "117:"  // Height 4: Partial direct writeback: partial_1_0
       "str b16, [x13, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "str b28, [x22, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
+      "str b28, [x21, #0x0]\n"
       "118:"  // Height 4: Partial direct writeback: Done
       "b 120f\n"
       "119:"  // Height 4: Full writeback
       "str q16, [x13, #0x0]\n"
       "add x13, x13, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q28, [x22, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q28, [x21, #0x0]\n"
       "120:"  // Height 4: Writeback done
-      "subs x15, x15, #0x10\n"
+      "subs x14, x14, #0x10\n"
       "bgt 92b\n"
       "subs %x[M], %x[M], #0x4\n"
       "beq 122f\n"
@@ -2089,9 +2088,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "122:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp
index 75e35a3e98..0c3470ec16 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp
@@ -45,18 +45,18 @@ void a64_hybrid_s8qa_dot_4x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -90,7 +90,7 @@ void a64_hybrid_s8qa_dot_4x16 (
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -100,8 +100,8 @@ void a64_hybrid_s8qa_dot_4x16 (
       "mov x26, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -134,7 +134,6 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q26, [x28, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
       "ldr q25, [x28, #0xa0]\n"
-      "add x24, x24, #0x10\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
       "ldr q24, [x28, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
@@ -145,10 +144,11 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q21, [x28, #0xe0]\n"
       ".inst 0x4f80ea90  // sdot v16.4s, v20.16b, v0.4b[2]\n"
       "ldr q20, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
       ".inst 0x4f80eb51  // sdot v17.4s, v26.16b, v0.4b[2]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4f80eb32  // sdot v18.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f80eb13  // sdot v19.4s, v24.16b, v0.4b[2]\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x4fa0eaf0  // sdot v16.4s, v23.16b, v0.4b[3]\n"
       ".inst 0x4fa0ead1  // sdot v17.4s, v22.16b, v0.4b[3]\n"
       ".inst 0x4fa0eab2  // sdot v18.4s, v21.16b, v0.4b[3]\n"
@@ -159,9 +159,9 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q0, [x24, #0x0]\n"
       "ldr q4, [x28, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q5, [x28, #0x10]\n"
       "ldr q6, [x28, #0x20]\n"
-      "cmp x25, #0x20\n"
       "ldr q7, [x28, #0x30]\n"
       "ldr q8, [x28, #0x40]\n"
       "ldr q9, [x28, #0x50]\n"
@@ -177,8 +177,6 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q26, [x28, #0x90]\n"
       ".inst 0x4f80e0f3  // sdot v19.4s, v7.16b, v0.4b[0]\n"
       "ldr q25, [x28, #0xa0]\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
       ".inst 0x4fa0e110  // sdot v16.4s, v8.16b, v0.4b[1]\n"
       "ldr q24, [x28, #0xb0]\n"
       ".inst 0x4fa0e131  // sdot v17.4s, v9.16b, v0.4b[1]\n"
@@ -189,10 +187,12 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q21, [x28, #0xe0]\n"
       ".inst 0x4f80ea90  // sdot v16.4s, v20.16b, v0.4b[2]\n"
       "ldr q20, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
       ".inst 0x4f80eb51  // sdot v17.4s, v26.16b, v0.4b[2]\n"
+      "sub x25, x25, #0x10\n"
       ".inst 0x4f80eb32  // sdot v18.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f80eb13  // sdot v19.4s, v24.16b, v0.4b[2]\n"
+      "add x24, x24, #0x10\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x4fa0eaf0  // sdot v16.4s, v23.16b, v0.4b[3]\n"
       ".inst 0x4fa0ead1  // sdot v17.4s, v22.16b, v0.4b[3]\n"
       ".inst 0x4fa0eab2  // sdot v18.4s, v21.16b, v0.4b[3]\n"
@@ -213,14 +213,14 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q23, [x28, #0x0]\n"
       "ldr q22, [x28, #0x10]\n"
       "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ldr q21, [x28, #0x20]\n"
       "ldr q20, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x4f80e2f0  // sdot v16.4s, v23.16b, v0.4b[0]\n"
       ".inst 0x4f80e2d1  // sdot v17.4s, v22.16b, v0.4b[0]\n"
       ".inst 0x4f80e2b2  // sdot v18.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f80e293  // sdot v19.4s, v20.16b, v0.4b[0]\n"
+      "add x28, x28, #0x40\n"
       "bge 12b\n"
       "14:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x25, 18f\n"
@@ -235,15 +235,15 @@ void a64_hybrid_s8qa_dot_4x16 (
       "tbnz %x[flags], #31, 17f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "17:"  // Height 1: Multiply loop: unique 4: skip row sum
-      "ldr q23, [x28, #0x0]\n"
-      "ldr q22, [x28, #0x10]\n"
+      "ldr q21, [x28, #0x0]\n"
+      "ldr q20, [x28, #0x10]\n"
+      ".inst 0x4f80e2b0  // sdot v16.4s, v21.16b, v0.4b[0]\n"
+      ".inst 0x4f80e291  // sdot v17.4s, v20.16b, v0.4b[0]\n"
       "ldr q21, [x28, #0x20]\n"
       "ldr q20, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f80e2f0  // sdot v16.4s, v23.16b, v0.4b[0]\n"
-      ".inst 0x4f80e2d1  // sdot v17.4s, v22.16b, v0.4b[0]\n"
       ".inst 0x4f80e2b2  // sdot v18.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f80e293  // sdot v19.4s, v20.16b, v0.4b[0]\n"
+      "add x28, x28, #0x40\n"
       "18:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -252,8 +252,8 @@ void a64_hybrid_s8qa_dot_4x16 (
       "prfm pstl1keep, [x27, #0x0]\n"
       "tbnz %x[flags], #31, 19f\n"
       "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
       "ld1r { v20.4s }, [x20]\n"
+      "addp v11.4s, v11.4s, v11.4s\n"
       "neg v20.4s, v20.4s\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "mul v11.4s, v11.4s, v20.4s\n"
@@ -267,16 +267,16 @@ void a64_hybrid_s8qa_dot_4x16 (
       "add v18.4s, v18.4s, v11.4s\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v20.4s }, [x20]\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add v16.4s, v16.4s, v24.4s\n"
       "add v17.4s, v17.4s, v23.4s\n"
+      "add v18.4s, v18.4s, v22.4s\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add x10, x10, #0x40\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v22.4s\n"
       "add v19.4s, v19.4s, v21.4s\n"
       "sqrdmulh v16.4s, v16.4s, v20.4s\n"
+      "add x10, x10, #0x40\n"
       "sqrdmulh v17.4s, v17.4s, v20.4s\n"
       "sqrdmulh v18.4s, v18.4s, v20.4s\n"
       "sqrdmulh v19.4s, v19.4s, v20.4s\n"
@@ -294,21 +294,21 @@ void a64_hybrid_s8qa_dot_4x16 (
       "sqadd v18.4s, v18.4s, v21.4s\n"
       "sqadd v19.4s, v19.4s, v20.4s\n"
       "20:"  // Height 1: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v22.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v22.4s }, [x21]\n"
-      "ld1r { v21.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v20.4s }, [x20]\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "add v16.4s, v16.4s, v22.4s\n"
       "add v17.4s, v17.4s, v22.4s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1r { v20.4s }, [x20]\n"
       "add v18.4s, v18.4s, v22.4s\n"
       "add v19.4s, v19.4s, v22.4s\n"
+      "cmp x9, #0x10\n"
       "smin v16.4s, v16.4s, v21.4s\n"
       "smin v17.4s, v17.4s, v21.4s\n"
       "smin v18.4s, v18.4s, v21.4s\n"
@@ -381,7 +381,7 @@ void a64_hybrid_s8qa_dot_4x16 (
       "movi v15.16b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "32:"  // Height 2: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -395,8 +395,8 @@ void a64_hybrid_s8qa_dot_4x16 (
       "mov x26, #0x0\n"
       "34:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 35f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -476,9 +476,9 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q0, [x24, #0x0]\n"
       "ldr q1, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q4, [x28, #0x0]\n"
       "ldr q5, [x28, #0x10]\n"
-      "cmp x25, #0x20\n"
       "ldr q6, [x28, #0x20]\n"
       "ldr q7, [x28, #0x30]\n"
       "ldr q8, [x28, #0x40]\n"
@@ -553,14 +553,14 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q27, [x28, #0x0]\n"
       "ldr q26, [x28, #0x10]\n"
       "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ldr q25, [x28, #0x20]\n"
       "ldr q24, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x4f80e370  // sdot v16.4s, v27.16b, v0.4b[0]\n"
       ".inst 0x4f81e374  // sdot v20.4s, v27.16b, v1.4b[0]\n"
       ".inst 0x4f80e351  // sdot v17.4s, v26.16b, v0.4b[0]\n"
       ".inst 0x4f81e355  // sdot v21.4s, v26.16b, v1.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x4f80e332  // sdot v18.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x4f81e336  // sdot v22.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x4f80e313  // sdot v19.4s, v24.16b, v0.4b[0]\n"
@@ -583,17 +583,17 @@ void a64_hybrid_s8qa_dot_4x16 (
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f942c  // sdot v12.4s, v1.16b, v15.16b\n"
       "47:"  // Height 2: Multiply loop: unique 8: skip row sum
-      "ldr q27, [x28, #0x0]\n"
+      "ldr q24, [x28, #0x0]\n"
       "ldr q26, [x28, #0x10]\n"
+      ".inst 0x4f80e310  // sdot v16.4s, v24.16b, v0.4b[0]\n"
+      ".inst 0x4f81e314  // sdot v20.4s, v24.16b, v1.4b[0]\n"
       "ldr q25, [x28, #0x20]\n"
       "ldr q24, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f80e370  // sdot v16.4s, v27.16b, v0.4b[0]\n"
-      ".inst 0x4f81e374  // sdot v20.4s, v27.16b, v1.4b[0]\n"
       ".inst 0x4f80e351  // sdot v17.4s, v26.16b, v0.4b[0]\n"
       ".inst 0x4f81e355  // sdot v21.4s, v26.16b, v1.4b[0]\n"
       ".inst 0x4f80e332  // sdot v18.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x4f81e336  // sdot v22.4s, v25.16b, v1.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x4f80e313  // sdot v19.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e317  // sdot v23.4s, v24.16b, v1.4b[0]\n"
       "48:"  // Height 2: Multiply loop: No odd multiplies
@@ -602,14 +602,14 @@ void a64_hybrid_s8qa_dot_4x16 (
       "cmp x26, x20\n"
       "bne 34b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbnz %x[flags], #31, 49f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v24.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v24.4s }, [x20]\n"
       "neg v24.4s, v24.4s\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
@@ -627,10 +627,10 @@ void a64_hybrid_s8qa_dot_4x16 (
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v24.4s }, [x20]\n"
       "add v22.4s, v22.4s, v12.4s\n"
       "add v23.4s, v23.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v16.4s, v16.4s, v28.4s\n"
       "add v17.4s, v17.4s, v27.4s\n"
@@ -652,45 +652,45 @@ void a64_hybrid_s8qa_dot_4x16 (
       "sqrdmulh v23.4s, v23.4s, v24.4s\n"
       "tbz %x[flags], #5, 50f\n"
       "and v24.16b, v16.16b, v0.16b\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
+      "sqadd v16.4s, v16.4s, v24.4s\n"
       "and v30.16b, v17.16b, v0.16b\n"
       "and v29.16b, v18.16b, v0.16b\n"
       "and v28.16b, v19.16b, v0.16b\n"
       "and v27.16b, v20.16b, v0.16b\n"
       "and v26.16b, v21.16b, v0.16b\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "and v25.16b, v22.16b, v0.16b\n"
+      "and v24.16b, v23.16b, v0.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v24.4s\n"
-      "and v24.16b, v23.16b, v0.16b\n"
       "sshr v26.4s, v26.4s, #0x1f\n"
       "sshr v25.4s, v25.4s, #0x1f\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v17.4s, v17.4s, v30.4s\n"
       "sqadd v18.4s, v18.4s, v29.4s\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v19.4s, v19.4s, v28.4s\n"
       "sqadd v20.4s, v20.4s, v27.4s\n"
       "sqadd v21.4s, v21.4s, v26.4s\n"
       "sqadd v22.4s, v22.4s, v25.4s\n"
       "sqadd v23.4s, v23.4s, v24.4s\n"
       "50:"  // Height 2: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x21]\n"
-      "ld1r { v25.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "add v16.4s, v16.4s, v26.4s\n"
       "add v17.4s, v17.4s, v26.4s\n"
       "add v18.4s, v18.4s, v26.4s\n"
@@ -724,68 +724,68 @@ void a64_hybrid_s8qa_dot_4x16 (
       "bge 59f\n"
       "tbz x9, #3, 54f\n"
       "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x9, #2, 52f\n"
       "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
       "tbz x9, #1, 51f\n"
       "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
       "b 58f\n"
       "51:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
       "b 58f\n"
       "52:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x9, #1, 53f\n"
       "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
       "b 58f\n"
       "53:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
       "b 58f\n"
       "54:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x9, #2, 56f\n"
       "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
       "tbz x9, #1, 55f\n"
       "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
       "b 58f\n"
       "55:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
       "b 58f\n"
       "56:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x9, #1, 57f\n"
       "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
+      "str h20, [x23], #0x2\n"
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
       "b 58f\n"
       "57:"  // Height 2: Partial direct writeback: partial_1_0
       "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
       "58:"  // Height 2: Partial direct writeback: Done
       "b 60f\n"
       "59:"  // Height 2: Full writeback
       "str q16, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
       "60:"  // Height 2: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 32b\n"
@@ -799,7 +799,7 @@ void a64_hybrid_s8qa_dot_4x16 (
       "movi v15.16b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "62:"  // Height 3: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -817,8 +817,8 @@ void a64_hybrid_s8qa_dot_4x16 (
       "mov x26, #0x0\n"
       "64:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 65f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -920,9 +920,9 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q0, [x24, #0x0]\n"
       "ldr q1, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q2, [x22, #0x0]\n"
       "ldr q4, [x28, #0x0]\n"
-      "cmp x25, #0x20\n"
       "ldr q5, [x28, #0x10]\n"
       "ldr q6, [x28, #0x20]\n"
       "ldr q7, [x28, #0x30]\n"
@@ -1020,14 +1020,14 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q31, [x28, #0x0]\n"
       "ldr q30, [x28, #0x10]\n"
       "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ldr q29, [x28, #0x20]\n"
       "ldr q28, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x4f80e3f0  // sdot v16.4s, v31.16b, v0.4b[0]\n"
       ".inst 0x4f81e3f4  // sdot v20.4s, v31.16b, v1.4b[0]\n"
       ".inst 0x4f82e3f8  // sdot v24.4s, v31.16b, v2.4b[0]\n"
       ".inst 0x4f80e3d1  // sdot v17.4s, v30.16b, v0.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x4f81e3d5  // sdot v21.4s, v30.16b, v1.4b[0]\n"
       ".inst 0x4f82e3d9  // sdot v25.4s, v30.16b, v2.4b[0]\n"
       ".inst 0x4f80e3b2  // sdot v18.4s, v29.16b, v0.4b[0]\n"
@@ -1060,15 +1060,15 @@ void a64_hybrid_s8qa_dot_4x16 (
       "77:"  // Height 3: Multiply loop: unique 12: skip row sum
       "ldr q31, [x28, #0x0]\n"
       "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x4f80e3f0  // sdot v16.4s, v31.16b, v0.4b[0]\n"
       ".inst 0x4f81e3f4  // sdot v20.4s, v31.16b, v1.4b[0]\n"
+      "ldr q29, [x28, #0x20]\n"
+      "ldr q28, [x28, #0x30]\n"
       ".inst 0x4f82e3f8  // sdot v24.4s, v31.16b, v2.4b[0]\n"
       ".inst 0x4f80e3d1  // sdot v17.4s, v30.16b, v0.4b[0]\n"
       ".inst 0x4f81e3d5  // sdot v21.4s, v30.16b, v1.4b[0]\n"
       ".inst 0x4f82e3d9  // sdot v25.4s, v30.16b, v2.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x4f80e3b2  // sdot v18.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f81e3b6  // sdot v22.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f82e3ba  // sdot v26.4s, v29.16b, v2.4b[0]\n"
@@ -1081,16 +1081,16 @@ void a64_hybrid_s8qa_dot_4x16 (
       "cmp x26, x20\n"
       "bne 64b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbnz %x[flags], #31, 79f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v28.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v28.4s }, [x20]\n"
       "addp v13.4s, v13.4s, v13.4s\n"
       "neg v28.4s, v28.4s\n"
       "addp v11.4s, v11.4s, v11.4s\n"
@@ -1111,10 +1111,10 @@ void a64_hybrid_s8qa_dot_4x16 (
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v28.4s }, [x20]\n"
       "add v22.4s, v22.4s, v12.4s\n"
       "add v23.4s, v23.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v24.4s, v24.4s, v13.4s\n"
       "add v25.4s, v25.4s, v13.4s\n"
@@ -1152,18 +1152,18 @@ void a64_hybrid_s8qa_dot_4x16 (
       "and v30.16b, v18.16b, v0.16b\n"
       "and v29.16b, v19.16b, v0.16b\n"
       "and v28.16b, v20.16b, v0.16b\n"
-      "and v3.16b, v21.16b, v0.16b\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
-      "and v2.16b, v22.16b, v0.16b\n"
       "sqadd v16.4s, v16.4s, v1.4s\n"
       "sqadd v17.4s, v17.4s, v31.4s\n"
       "sqadd v18.4s, v18.4s, v30.4s\n"
       "sqadd v19.4s, v19.4s, v29.4s\n"
       "sqadd v20.4s, v20.4s, v28.4s\n"
+      "and v3.16b, v21.16b, v0.16b\n"
+      "and v2.16b, v22.16b, v0.16b\n"
       "and v1.16b, v23.16b, v0.16b\n"
       "and v31.16b, v24.16b, v0.16b\n"
       "and v30.16b, v25.16b, v0.16b\n"
@@ -1184,21 +1184,21 @@ void a64_hybrid_s8qa_dot_4x16 (
       "sqadd v26.4s, v26.4s, v29.4s\n"
       "sqadd v27.4s, v27.4s, v28.4s\n"
       "80:"  // Height 3: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v30.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v30.4s }, [x21]\n"
-      "ld1r { v29.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v28.4s }, [x20]\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "srshl v24.4s, v24.4s, v0.4s\n"
       "srshl v25.4s, v25.4s, v0.4s\n"
       "srshl v26.4s, v26.4s, v0.4s\n"
@@ -1251,103 +1251,102 @@ void a64_hybrid_s8qa_dot_4x16 (
       "bge 89f\n"
       "tbz x9, #3, 84f\n"
       "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x9, #2, 82f\n"
       "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
       "tbz x9, #1, 81f\n"
       "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 88f\n"
       "81:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 88f\n"
       "82:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x9, #1, 83f\n"
       "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 88f\n"
       "83:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 88f\n"
       "84:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x9, #2, 86f\n"
       "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
       "tbz x9, #1, 85f\n"
       "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 88f\n"
       "85:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 88f\n"
       "86:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x9, #1, 87f\n"
       "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 88f\n"
       "87:"  // Height 3: Partial direct writeback: partial_1_0
       "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "88:"  // Height 3: Partial direct writeback: Done
       "b 90f\n"
       "89:"  // Height 3: Full writeback
       "str q16, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "90:"  // Height 3: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 62b\n"
       "b 122f\n"
       "91:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x4\n"
       "mov x10, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
+      "movi v13.4s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "movi v13.4s, #0x0\n"
       "movi v14.4s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
       "movi v15.16b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "92:"  // Height 4: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -1369,8 +1368,8 @@ void a64_hybrid_s8qa_dot_4x16 (
       "mov x26, #0x0\n"
       "94:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 95f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1494,9 +1493,9 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q0, [x24, #0x0]\n"
       "ldr q1, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q2, [x22, #0x0]\n"
       "ldr q3, [x21, #0x0]\n"
-      "cmp x25, #0x20\n"
       "ldr q4, [x28, #0x0]\n"
       "ldr q5, [x28, #0x10]\n"
       "ldr q6, [x28, #0x20]\n"
@@ -1617,14 +1616,14 @@ void a64_hybrid_s8qa_dot_4x16 (
       "ldr q7, [x28, #0x0]\n"
       "ldr q6, [x28, #0x10]\n"
       "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ldr q5, [x28, #0x20]\n"
       "ldr q4, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x4f80e0f0  // sdot v16.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0f4  // sdot v20.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f8  // sdot v24.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0fc  // sdot v28.4s, v7.16b, v3.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x4f80e0d1  // sdot v17.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0d5  // sdot v21.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x4f82e0d9  // sdot v25.4s, v6.16b, v2.4b[0]\n"
@@ -1665,15 +1664,15 @@ void a64_hybrid_s8qa_dot_4x16 (
       "107:"  // Height 4: Multiply loop: unique 16: skip row sum
       "ldr q7, [x28, #0x0]\n"
       "ldr q6, [x28, #0x10]\n"
-      "ldr q5, [x28, #0x20]\n"
-      "ldr q4, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x4f80e0f0  // sdot v16.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0f4  // sdot v20.4s, v7.16b, v1.4b[0]\n"
+      "ldr q5, [x28, #0x20]\n"
+      "ldr q4, [x28, #0x30]\n"
       ".inst 0x4f82e0f8  // sdot v24.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0fc  // sdot v28.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f80e0d1  // sdot v17.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0d5  // sdot v21.4s, v6.16b, v1.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x4f82e0d9  // sdot v25.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0dd  // sdot v29.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x4f80e0b2  // sdot v18.4s, v5.16b, v0.4b[0]\n"
@@ -1690,18 +1689,18 @@ void a64_hybrid_s8qa_dot_4x16 (
       "cmp x26, x20\n"
       "bne 94b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20\n"
+      "add x21, x22, x20\n"
       "prfm pstl1keep, [x23, #0x0]\n"
-      "add x22, x23, x20\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbnz %x[flags], #31, 109f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
       "addp v13.4s, v13.4s, v13.4s\n"
       "addp v14.4s, v14.4s, v14.4s\n"
       "neg v0.4s, v0.4s\n"
@@ -1725,10 +1724,10 @@ void a64_hybrid_s8qa_dot_4x16 (
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v1.4s }, [x20]\n"
       "add v22.4s, v22.4s, v12.4s\n"
       "add v23.4s, v23.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v24.4s, v24.4s, v13.4s\n"
       "add v25.4s, v25.4s, v13.4s\n"
@@ -1775,32 +1774,32 @@ void a64_hybrid_s8qa_dot_4x16 (
       "tbz %x[flags], #5, 110f\n"
       "and v2.16b, v16.16b, v0.16b\n"
       "and v1.16b, v17.16b, v0.16b\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v16.4s, v16.4s, v2.4s\n"
+      "sqadd v17.4s, v17.4s, v1.4s\n"
       "and v7.16b, v18.16b, v0.16b\n"
       "and v6.16b, v19.16b, v0.16b\n"
       "and v5.16b, v20.16b, v0.16b\n"
       "and v4.16b, v21.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
       "and v3.16b, v22.16b, v0.16b\n"
+      "and v2.16b, v23.16b, v0.16b\n"
+      "and v1.16b, v24.16b, v0.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
       "sshr v6.4s, v6.4s, #0x1f\n"
       "sshr v5.4s, v5.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v2.4s\n"
-      "sqadd v17.4s, v17.4s, v1.4s\n"
-      "and v2.16b, v23.16b, v0.16b\n"
-      "and v1.16b, v24.16b, v0.16b\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sshr v3.4s, v3.4s, #0x1f\n"
-      "sqadd v18.4s, v18.4s, v7.4s\n"
-      "sqadd v19.4s, v19.4s, v6.4s\n"
       "sshr v2.4s, v2.4s, #0x1f\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v18.4s, v18.4s, v7.4s\n"
+      "sqadd v19.4s, v19.4s, v6.4s\n"
       "sqadd v20.4s, v20.4s, v5.4s\n"
       "sqadd v21.4s, v21.4s, v4.4s\n"
       "sqadd v22.4s, v22.4s, v3.4s\n"
-      "and v7.16b, v25.16b, v0.16b\n"
       "sqadd v23.4s, v23.4s, v2.4s\n"
       "sqadd v24.4s, v24.4s, v1.4s\n"
+      "and v7.16b, v25.16b, v0.16b\n"
       "and v6.16b, v26.16b, v0.16b\n"
       "and v5.16b, v27.16b, v0.16b\n"
       "and v4.16b, v28.16b, v0.16b\n"
@@ -1822,21 +1821,21 @@ void a64_hybrid_s8qa_dot_4x16 (
       "sqadd v30.4s, v30.4s, v2.4s\n"
       "sqadd v31.4s, v31.4s, v1.4s\n"
       "110:"  // Height 4: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v3.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v3.4s }, [x21]\n"
-      "ld1r { v2.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v2.4s }, [x20]\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v1.4s }, [x20]\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "srshl v24.4s, v24.4s, v0.4s\n"
       "srshl v25.4s, v25.4s, v0.4s\n"
       "srshl v26.4s, v26.4s, v0.4s\n"
@@ -1908,100 +1907,100 @@ void a64_hybrid_s8qa_dot_4x16 (
       "bge 119f\n"
       "tbz x9, #3, 114f\n"
       "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x9, #2, 112f\n"
       "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
       "tbz x9, #1, 111f\n"
       "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "st1 { v28.h }[6], [x22], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "st1 { v28.h }[6], [x21], #0x2\n"
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "st1 { v28.b }[14], [x22]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
+      "st1 { v28.b }[14], [x21]\n"
       "b 118f\n"
       "111:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "st1 { v28.b }[12], [x22]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
+      "st1 { v28.b }[12], [x21]\n"
       "b 118f\n"
       "112:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x9, #1, 113f\n"
       "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "st1 { v28.h }[4], [x22], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "st1 { v28.h }[4], [x21], #0x2\n"
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "st1 { v28.b }[10], [x22]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
+      "st1 { v28.b }[10], [x21]\n"
       "b 118f\n"
       "113:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "st1 { v28.b }[8], [x22]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
+      "st1 { v28.b }[8], [x21]\n"
       "b 118f\n"
       "114:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x9, #2, 116f\n"
       "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
       "tbz x9, #1, 115f\n"
       "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "st1 { v28.h }[2], [x22], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "st1 { v28.h }[2], [x21], #0x2\n"
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "st1 { v28.b }[6], [x22]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
+      "st1 { v28.b }[6], [x21]\n"
       "b 118f\n"
       "115:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "st1 { v28.b }[4], [x22]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
+      "st1 { v28.b }[4], [x21]\n"
       "b 118f\n"
       "116:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x9, #1, 117f\n"
       "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "str h28, [x22], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "str h28, [x21], #0x2\n"
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "st1 { v28.b }[2], [x22]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
+      "st1 { v28.b }[2], [x21]\n"
       "b 118f\n"
       "117:"  // Height 4: Partial direct writeback: partial_1_0
       "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "str b28, [x22, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
+      "str b28, [x21, #0x0]\n"
       "118:"  // Height 4: Partial direct writeback: Done
       "b 120f\n"
       "119:"  // Height 4: Full writeback
       "str q16, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q28, [x22, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q28, [x21, #0x0]\n"
       "120:"  // Height 4: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 92b\n"
@@ -2017,8 +2016,8 @@ void a64_hybrid_s8qa_dot_4x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "122:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16.hpp
index 55290826d1..7052d1cc41 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 16, 8> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 4, 16, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp
index 8f70b3dc26..722f9af535 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp
@@ -45,18 +45,18 @@ void a64_hybrid_s8qa_mmla_4x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -90,7 +90,7 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -104,8 +104,8 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "mov x26, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -130,7 +130,6 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "ldr q4, [x28, #0x60]\n"
       "blt 9f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
-      "add x24, x24, #0x10\n"
       "trn1 v0.2d, v1.2d, v27.2d\n"
       ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
       "ldr q25, [x28, #0x70]\n"
@@ -152,8 +151,9 @@ void a64_hybrid_s8qa_mmla_4x16 (
       ".inst 0x4e98a430  // smmla v16.4s, v1.16b, v24.16b\n"
       "ldr q24, [x28, #0xf0]\n"
       ".inst 0x4e9ea434  // smmla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e9da431  // smmla v17.4s, v1.16b, v29.16b\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x4e9ca435  // smmla v21.4s, v1.16b, v28.16b\n"
       ".inst 0x4e9ba432  // smmla v18.4s, v1.16b, v27.16b\n"
       ".inst 0x4e9aa436  // smmla v22.4s, v1.16b, v26.16b\n"
@@ -166,9 +166,9 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "ldr q1, [x24, #0x0]\n"
       "ldr q5, [x28, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q6, [x28, #0x10]\n"
       "ldr q7, [x28, #0x20]\n"
-      "cmp x25, #0x20\n"
       "ldr q8, [x28, #0x30]\n"
       "ldr q9, [x28, #0x40]\n"
       "ldr q10, [x28, #0x50]\n"
@@ -176,12 +176,10 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "prfm pldl1keep, [x24, #0x80]\n"
       "bge 7b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
       "trn1 v0.2d, v1.2d, v24.2d\n"
-      "trn2 v1.2d, v1.2d, v24.2d\n"
       ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
       "ldr q25, [x28, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v24.2d\n"
       ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
       "ldr q24, [x28, #0x80]\n"
       ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
@@ -198,9 +196,11 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "ldr q25, [x28, #0xe0]\n"
       ".inst 0x4e98a430  // smmla v16.4s, v1.16b, v24.16b\n"
       "ldr q24, [x28, #0xf0]\n"
+      "sub x25, x25, #0x10\n"
       ".inst 0x4e9ea434  // smmla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
       ".inst 0x4e9da431  // smmla v17.4s, v1.16b, v29.16b\n"
+      "add x24, x24, #0x10\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x4e9ca435  // smmla v21.4s, v1.16b, v28.16b\n"
       ".inst 0x4e9ba432  // smmla v18.4s, v1.16b, v27.16b\n"
       ".inst 0x4e9aa436  // smmla v22.4s, v1.16b, v26.16b\n"
@@ -222,24 +222,24 @@ void a64_hybrid_s8qa_mmla_4x16 (
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "13:"  // Height 1: Multiply loop: unique 3: skip row sum
       "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
+      "ldr q26, [x28, #0x10]\n"
+      ".inst 0x4e98a410  // smmla v16.4s, v0.16b, v24.16b\n"
       "sub x25, x25, #0x8\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
+      "ldr q25, [x28, #0x20]\n"
+      "ldr q24, [x28, #0x30]\n"
       "cmp x25, #0x8\n"
+      ".inst 0x4e9aa414  // smmla v20.4s, v0.16b, v26.16b\n"
       "ldr q27, [x28, #0x40]\n"
       "ldr q26, [x28, #0x50]\n"
+      ".inst 0x4e99a411  // smmla v17.4s, v0.16b, v25.16b\n"
+      ".inst 0x4e98a415  // smmla v21.4s, v0.16b, v24.16b\n"
       "ldr q25, [x28, #0x60]\n"
-      ".inst 0x4e98a410  // smmla v16.4s, v0.16b, v24.16b\n"
       "ldr q24, [x28, #0x70]\n"
-      ".inst 0x4e9ea414  // smmla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x4e9da411  // smmla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x4e9ca415  // smmla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
       ".inst 0x4e9ba412  // smmla v18.4s, v0.16b, v27.16b\n"
       ".inst 0x4e9aa416  // smmla v22.4s, v0.16b, v26.16b\n"
       ".inst 0x4e99a413  // smmla v19.4s, v0.16b, v25.16b\n"
       ".inst 0x4e98a417  // smmla v23.4s, v0.16b, v24.16b\n"
+      "add x28, x28, #0x80\n"
       "bge 12b\n"
       "14:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x25, 20f\n"
@@ -267,23 +267,23 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "tbnz %x[flags], #31, 19f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "19:"  // Height 1: Multiply loop: unique 4: skip row sum
-      "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "ldr q27, [x28, #0x40]\n"
-      "ldr q26, [x28, #0x50]\n"
+      "ldr q25, [x28, #0x0]\n"
+      "ldr q24, [x28, #0x10]\n"
+      ".inst 0x4e99a410  // smmla v16.4s, v0.16b, v25.16b\n"
+      ".inst 0x4e98a414  // smmla v20.4s, v0.16b, v24.16b\n"
+      "ldr q25, [x28, #0x20]\n"
+      "ldr q24, [x28, #0x30]\n"
+      ".inst 0x4e99a411  // smmla v17.4s, v0.16b, v25.16b\n"
+      ".inst 0x4e98a415  // smmla v21.4s, v0.16b, v24.16b\n"
+      "ldr q25, [x28, #0x40]\n"
+      "ldr q24, [x28, #0x50]\n"
+      ".inst 0x4e99a412  // smmla v18.4s, v0.16b, v25.16b\n"
+      ".inst 0x4e98a416  // smmla v22.4s, v0.16b, v24.16b\n"
       "ldr q25, [x28, #0x60]\n"
-      ".inst 0x4e98a410  // smmla v16.4s, v0.16b, v24.16b\n"
       "ldr q24, [x28, #0x70]\n"
-      ".inst 0x4e9ea414  // smmla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x4e9da411  // smmla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x4e9ca415  // smmla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e9ba412  // smmla v18.4s, v0.16b, v27.16b\n"
-      ".inst 0x4e9aa416  // smmla v22.4s, v0.16b, v26.16b\n"
       ".inst 0x4e99a413  // smmla v19.4s, v0.16b, v25.16b\n"
       ".inst 0x4e98a417  // smmla v23.4s, v0.16b, v24.16b\n"
+      "add x28, x28, #0x80\n"
       "20:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -297,8 +297,8 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "mov v23.16b, v16.16b\n"
       "tbnz %x[flags], #31, 21f\n"
       "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
       "ld1r { v16.4s }, [x20]\n"
+      "addp v11.4s, v11.4s, v11.4s\n"
       "neg v16.4s, v16.4s\n"
       "dup v11.4s, v11.s[0]\n"
       "mul v11.4s, v11.4s, v16.4s\n"
@@ -312,16 +312,16 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "add v18.4s, v18.4s, v11.4s\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v16.4s }, [x20]\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add v23.4s, v23.4s, v24.4s\n"
       "add v17.4s, v17.4s, v22.4s\n"
+      "add v18.4s, v18.4s, v21.4s\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add x10, x10, #0x40\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v21.4s\n"
       "add v19.4s, v19.4s, v20.4s\n"
       "sqrdmulh v23.4s, v23.4s, v16.4s\n"
+      "add x10, x10, #0x40\n"
       "sqrdmulh v17.4s, v17.4s, v16.4s\n"
       "sqrdmulh v18.4s, v18.4s, v16.4s\n"
       "sqrdmulh v19.4s, v19.4s, v16.4s\n"
@@ -339,21 +339,21 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "sqadd v18.4s, v18.4s, v20.4s\n"
       "sqadd v19.4s, v19.4s, v16.4s\n"
       "22:"  // Height 1: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v21.4s }, [x21]\n"
-      "ld1r { v20.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v16.4s }, [x20]\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v20.4s }, [x20]\n"
       "add v23.4s, v23.4s, v21.4s\n"
       "add v17.4s, v17.4s, v21.4s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1r { v16.4s }, [x20]\n"
       "add v18.4s, v18.4s, v21.4s\n"
       "add v19.4s, v19.4s, v21.4s\n"
+      "cmp x9, #0x10\n"
       "smin v23.4s, v23.4s, v20.4s\n"
       "smin v17.4s, v17.4s, v20.4s\n"
       "smin v18.4s, v18.4s, v20.4s\n"
@@ -426,7 +426,7 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "movi v15.16b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "34:"  // Height 2: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -440,8 +440,8 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "mov x26, #0x0\n"
       "36:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 37f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -471,13 +471,11 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "blt 41f\n"
       "39:"  // Height 2: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
       "ldr q25, [x28, #0x70]\n"
       ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
       "ldr q24, [x28, #0x80]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
       "ldr q30, [x28, #0x90]\n"
       ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
@@ -493,9 +491,11 @@ void a64_hybrid_s8qa_mmla_4x16 (
       ".inst 0x4e98a430  // smmla v16.4s, v1.16b, v24.16b\n"
       "ldr q24, [x28, #0xf0]\n"
       ".inst 0x4e9ea434  // smmla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e9da431  // smmla v17.4s, v1.16b, v29.16b\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e9ca435  // smmla v21.4s, v1.16b, v28.16b\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x4e9ba432  // smmla v18.4s, v1.16b, v27.16b\n"
       ".inst 0x4e9aa436  // smmla v22.4s, v1.16b, v26.16b\n"
       ".inst 0x4e99a433  // smmla v19.4s, v1.16b, v25.16b\n"
@@ -507,9 +507,9 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "ldr q1, [x24, #0x0]\n"
       "ldr q2, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q5, [x28, #0x0]\n"
       "ldr q6, [x28, #0x10]\n"
-      "cmp x25, #0x20\n"
       "ldr q7, [x28, #0x20]\n"
       "ldr q8, [x28, #0x30]\n"
       "ldr q9, [x28, #0x40]\n"
@@ -520,14 +520,11 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "bge 39b\n"
       "41:"  // Height 2: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
       "ldr q25, [x28, #0x70]\n"
       ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
       "ldr q24, [x28, #0x80]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
       "ldr q30, [x28, #0x90]\n"
       ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
@@ -542,11 +539,14 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "ldr q25, [x28, #0xe0]\n"
       ".inst 0x4e98a430  // smmla v16.4s, v1.16b, v24.16b\n"
       "ldr q24, [x28, #0xf0]\n"
+      "sub x25, x25, #0x10\n"
       ".inst 0x4e9ea434  // smmla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
       ".inst 0x4e9da431  // smmla v17.4s, v1.16b, v29.16b\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e9ca435  // smmla v21.4s, v1.16b, v28.16b\n"
       ".inst 0x4e9ba432  // smmla v18.4s, v1.16b, v27.16b\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x4e9aa436  // smmla v22.4s, v1.16b, v26.16b\n"
       ".inst 0x4e99a433  // smmla v19.4s, v1.16b, v25.16b\n"
       ".inst 0x4e98a437  // smmla v23.4s, v1.16b, v24.16b\n"
@@ -568,24 +568,24 @@ void a64_hybrid_s8qa_mmla_4x16 (
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "45:"  // Height 2: Multiply loop: unique 7: skip row sum
       "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
+      "ldr q26, [x28, #0x10]\n"
+      ".inst 0x4e98a410  // smmla v16.4s, v0.16b, v24.16b\n"
       "sub x25, x25, #0x8\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
+      "ldr q25, [x28, #0x20]\n"
+      "ldr q24, [x28, #0x30]\n"
       "cmp x25, #0x8\n"
+      ".inst 0x4e9aa414  // smmla v20.4s, v0.16b, v26.16b\n"
       "ldr q27, [x28, #0x40]\n"
       "ldr q26, [x28, #0x50]\n"
+      ".inst 0x4e99a411  // smmla v17.4s, v0.16b, v25.16b\n"
+      ".inst 0x4e98a415  // smmla v21.4s, v0.16b, v24.16b\n"
       "ldr q25, [x28, #0x60]\n"
-      ".inst 0x4e98a410  // smmla v16.4s, v0.16b, v24.16b\n"
       "ldr q24, [x28, #0x70]\n"
-      ".inst 0x4e9ea414  // smmla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x4e9da411  // smmla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x4e9ca415  // smmla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
       ".inst 0x4e9ba412  // smmla v18.4s, v0.16b, v27.16b\n"
       ".inst 0x4e9aa416  // smmla v22.4s, v0.16b, v26.16b\n"
       ".inst 0x4e99a413  // smmla v19.4s, v0.16b, v25.16b\n"
       ".inst 0x4e98a417  // smmla v23.4s, v0.16b, v24.16b\n"
+      "add x28, x28, #0x80\n"
       "bge 44b\n"
       "46:"  // Height 2: Multiply loop: Skip odd blocks
       "cbz x25, 52f\n"
@@ -620,23 +620,23 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "tbnz %x[flags], #31, 51f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       "51:"  // Height 2: Multiply loop: unique 8: skip row sum
-      "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "ldr q27, [x28, #0x40]\n"
-      "ldr q26, [x28, #0x50]\n"
+      "ldr q25, [x28, #0x0]\n"
+      "ldr q24, [x28, #0x10]\n"
+      ".inst 0x4e99a410  // smmla v16.4s, v0.16b, v25.16b\n"
+      ".inst 0x4e98a414  // smmla v20.4s, v0.16b, v24.16b\n"
+      "ldr q25, [x28, #0x20]\n"
+      "ldr q24, [x28, #0x30]\n"
+      ".inst 0x4e99a411  // smmla v17.4s, v0.16b, v25.16b\n"
+      ".inst 0x4e98a415  // smmla v21.4s, v0.16b, v24.16b\n"
+      "ldr q25, [x28, #0x40]\n"
+      "ldr q24, [x28, #0x50]\n"
+      ".inst 0x4e99a412  // smmla v18.4s, v0.16b, v25.16b\n"
+      ".inst 0x4e98a416  // smmla v22.4s, v0.16b, v24.16b\n"
       "ldr q25, [x28, #0x60]\n"
-      ".inst 0x4e98a410  // smmla v16.4s, v0.16b, v24.16b\n"
       "ldr q24, [x28, #0x70]\n"
-      ".inst 0x4e9ea414  // smmla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x4e9da411  // smmla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x4e9ca415  // smmla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e9ba412  // smmla v18.4s, v0.16b, v27.16b\n"
-      ".inst 0x4e9aa416  // smmla v22.4s, v0.16b, v26.16b\n"
       ".inst 0x4e99a413  // smmla v19.4s, v0.16b, v25.16b\n"
       ".inst 0x4e98a417  // smmla v23.4s, v0.16b, v24.16b\n"
+      "add x28, x28, #0x80\n"
       "52:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -644,21 +644,21 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "bne 36b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v24.2d, v16.2d, v20.2d\n"
+      "add x23, x27, x20\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "mov v23.16b, v24.16b\n"
       "tbnz %x[flags], #31, 53f\n"
       "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
       "ld1r { v24.4s }, [x20]\n"
+      "addp v11.4s, v11.4s, v11.4s\n"
       "neg v24.4s, v24.4s\n"
       "dup v12.4s, v11.s[3]\n"
       "dup v11.4s, v11.s[0]\n"
@@ -676,10 +676,10 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "add v16.4s, v16.4s, v12.4s\n"
       "add v17.4s, v17.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v24.4s }, [x20]\n"
       "add v18.4s, v18.4s, v12.4s\n"
       "add v19.4s, v19.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v23.4s, v23.4s, v28.4s\n"
       "add v20.4s, v20.4s, v27.4s\n"
@@ -701,45 +701,45 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "sqrdmulh v19.4s, v19.4s, v24.4s\n"
       "tbz %x[flags], #5, 54f\n"
       "and v24.16b, v23.16b, v0.16b\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
+      "sqadd v23.4s, v23.4s, v24.4s\n"
       "and v30.16b, v20.16b, v0.16b\n"
       "and v29.16b, v21.16b, v0.16b\n"
       "and v28.16b, v22.16b, v0.16b\n"
       "and v27.16b, v16.16b, v0.16b\n"
       "and v26.16b, v17.16b, v0.16b\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "and v25.16b, v18.16b, v0.16b\n"
+      "and v24.16b, v19.16b, v0.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
-      "sqadd v23.4s, v23.4s, v24.4s\n"
-      "and v24.16b, v19.16b, v0.16b\n"
       "sshr v26.4s, v26.4s, #0x1f\n"
       "sshr v25.4s, v25.4s, #0x1f\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v20.4s, v20.4s, v30.4s\n"
       "sqadd v21.4s, v21.4s, v29.4s\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v22.4s, v22.4s, v28.4s\n"
       "sqadd v16.4s, v16.4s, v27.4s\n"
       "sqadd v17.4s, v17.4s, v26.4s\n"
       "sqadd v18.4s, v18.4s, v25.4s\n"
       "sqadd v19.4s, v19.4s, v24.4s\n"
       "54:"  // Height 2: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x21]\n"
-      "ld1r { v25.4s }, [x20]\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "add v23.4s, v23.4s, v26.4s\n"
       "add v20.4s, v20.4s, v26.4s\n"
       "add v21.4s, v21.4s, v26.4s\n"
@@ -773,68 +773,68 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "bge 63f\n"
       "tbz x9, #3, 58f\n"
       "str d23, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
       "tbz x9, #2, 56f\n"
       "st1 { v23.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
       "tbz x9, #1, 55f\n"
       "st1 { v23.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
+      "st1 { v16.b }[14], [x23]\n"
       "b 62f\n"
       "55:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
+      "st1 { v16.b }[12], [x23]\n"
       "b 62f\n"
       "56:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x9, #1, 57f\n"
       "st1 { v23.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
+      "st1 { v16.b }[10], [x23]\n"
       "b 62f\n"
       "57:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
+      "st1 { v16.b }[8], [x23]\n"
       "b 62f\n"
       "58:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x9, #2, 60f\n"
       "str s23, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
+      "str s16, [x23], #0x4\n"
       "tbz x9, #1, 59f\n"
       "st1 { v23.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
+      "st1 { v16.b }[6], [x23]\n"
       "b 62f\n"
       "59:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
+      "st1 { v16.b }[4], [x23]\n"
       "b 62f\n"
       "60:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x9, #1, 61f\n"
       "str h23, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
+      "str h16, [x23], #0x2\n"
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
+      "st1 { v16.b }[2], [x23]\n"
       "b 62f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_0
       "str b23, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
       "62:"  // Height 2: Partial direct writeback: Done
       "b 64f\n"
       "63:"  // Height 2: Full writeback
       "str q23, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
       "64:"  // Height 2: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 34b\n"
@@ -848,7 +848,7 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "movi v15.16b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "66:"  // Height 3: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -870,8 +870,8 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "mov x26, #0x0\n"
       "68:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 69f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -905,35 +905,35 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "71:"  // Height 3: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "ldr q14, [x28, #0x60]\n"
       ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
+      "trn1 v2.2d, v3.2d, v4.2d\n"
       ".inst 0x4e85a458  // smmla v24.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0x70]\n"
+      "ldr q14, [x28, #0x70]\n"
+      ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
+      "ldr q5, [x28, #0x60]\n"
       ".inst 0x4e86a45c  // smmla v28.4s, v2.16b, v6.16b\n"
       "ldr q4, [x28, #0x80]\n"
-      ".inst 0x4e89a412  // smmla v18.4s, v0.16b, v9.16b\n"
+      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x4e87a459  // smmla v25.4s, v2.16b, v7.16b\n"
       "ldr q7, [x28, #0x90]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x4e88a45d  // smmla v29.4s, v2.16b, v8.16b\n"
       "ldr q8, [x28, #0xa0]\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x4e89a412  // smmla v18.4s, v0.16b, v9.16b\n"
       ".inst 0x4e89a45a  // smmla v26.4s, v2.16b, v9.16b\n"
       "ldr q9, [x28, #0xb0]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4e8aa416  // smmla v22.4s, v0.16b, v10.16b\n"
       ".inst 0x4e8aa45e  // smmla v30.4s, v2.16b, v10.16b\n"
       "ldr q10, [x28, #0xc0]\n"
-      ".inst 0x4e8ea413  // smmla v19.4s, v0.16b, v14.16b\n"
-      ".inst 0x4e8ea45b  // smmla v27.4s, v2.16b, v14.16b\n"
+      ".inst 0x4e85a413  // smmla v19.4s, v0.16b, v5.16b\n"
+      ".inst 0x4e85a45b  // smmla v27.4s, v2.16b, v5.16b\n"
       "ldr q6, [x28, #0xd0]\n"
-      ".inst 0x4e85a417  // smmla v23.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85a45f  // smmla v31.4s, v2.16b, v5.16b\n"
+      ".inst 0x4e8ea417  // smmla v23.4s, v0.16b, v14.16b\n"
+      ".inst 0x4e8ea45f  // smmla v31.4s, v2.16b, v14.16b\n"
       "ldr q5, [x28, #0xe0]\n"
       ".inst 0x4e84a430  // smmla v16.4s, v1.16b, v4.16b\n"
       ".inst 0x4e84a478  // smmla v24.4s, v3.16b, v4.16b\n"
@@ -962,9 +962,9 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "ldr q1, [x24, #0x0]\n"
       "ldr q2, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q3, [x22, #0x0]\n"
       "ldr q5, [x28, #0x0]\n"
-      "cmp x25, #0x20\n"
       "ldr q6, [x28, #0x10]\n"
       "ldr q7, [x28, #0x20]\n"
       "ldr q8, [x28, #0x30]\n"
@@ -977,36 +977,36 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "73:"  // Height 3: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "ldr q14, [x28, #0x60]\n"
       ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
+      "trn1 v2.2d, v3.2d, v4.2d\n"
       ".inst 0x4e85a458  // smmla v24.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0x70]\n"
+      "ldr q14, [x28, #0x70]\n"
+      ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
+      "ldr q5, [x28, #0x60]\n"
       ".inst 0x4e86a45c  // smmla v28.4s, v2.16b, v6.16b\n"
       "ldr q4, [x28, #0x80]\n"
-      ".inst 0x4e89a412  // smmla v18.4s, v0.16b, v9.16b\n"
+      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x4e87a459  // smmla v25.4s, v2.16b, v7.16b\n"
       "ldr q7, [x28, #0x90]\n"
+      "sub x25, x25, #0x10\n"
+      ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x4e88a45d  // smmla v29.4s, v2.16b, v8.16b\n"
       "ldr q8, [x28, #0xa0]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x4e89a412  // smmla v18.4s, v0.16b, v9.16b\n"
       ".inst 0x4e89a45a  // smmla v26.4s, v2.16b, v9.16b\n"
       "ldr q9, [x28, #0xb0]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e8aa416  // smmla v22.4s, v0.16b, v10.16b\n"
       ".inst 0x4e8aa45e  // smmla v30.4s, v2.16b, v10.16b\n"
       "ldr q10, [x28, #0xc0]\n"
-      ".inst 0x4e8ea413  // smmla v19.4s, v0.16b, v14.16b\n"
-      ".inst 0x4e8ea45b  // smmla v27.4s, v2.16b, v14.16b\n"
+      "add x22, x22, #0x10\n"
+      ".inst 0x4e85a413  // smmla v19.4s, v0.16b, v5.16b\n"
+      ".inst 0x4e85a45b  // smmla v27.4s, v2.16b, v5.16b\n"
       "ldr q6, [x28, #0xd0]\n"
-      ".inst 0x4e85a417  // smmla v23.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85a45f  // smmla v31.4s, v2.16b, v5.16b\n"
+      ".inst 0x4e8ea417  // smmla v23.4s, v0.16b, v14.16b\n"
+      ".inst 0x4e8ea45f  // smmla v31.4s, v2.16b, v14.16b\n"
       "ldr q5, [x28, #0xe0]\n"
       ".inst 0x4e84a430  // smmla v16.4s, v1.16b, v4.16b\n"
       ".inst 0x4e84a478  // smmla v24.4s, v3.16b, v4.16b\n"
@@ -1040,34 +1040,34 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "cmp x25, #0x8\n"
       "blt 78f\n"
       "76:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d3, [x24], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
       "ldr d0, [x23], #0x8\n"
+      "trn1 v0.2d, v1.2d, v0.2d\n"
       "ldr d1, [x22], #0x8\n"
-      "trn1 v0.2d, v3.2d, v0.2d\n"
       "trn1 v2.2d, v1.2d, v2.2d\n"
       "tbnz %x[flags], #31, 77f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       "77:"  // Height 3: Multiply loop: unique 11: skip row sum
-      "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "sub x25, x25, #0x8\n"
+      "ldr q3, [x28, #0x0]\n"
+      "ldr q1, [x28, #0x10]\n"
+      ".inst 0x4e83a410  // smmla v16.4s, v0.16b, v3.16b\n"
+      ".inst 0x4e83a458  // smmla v24.4s, v2.16b, v3.16b\n"
       "ldr q7, [x28, #0x20]\n"
       "ldr q6, [x28, #0x30]\n"
+      "sub x25, x25, #0x8\n"
       "cmp x25, #0x8\n"
       "ldr q5, [x28, #0x40]\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x4e81a414  // smmla v20.4s, v0.16b, v1.16b\n"
+      ".inst 0x4e81a45c  // smmla v28.4s, v2.16b, v1.16b\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x4e81a410  // smmla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x28, #0x70]\n"
-      ".inst 0x4e88a414  // smmla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e88a45c  // smmla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
       ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x4e87a459  // smmla v25.4s, v2.16b, v7.16b\n"
       ".inst 0x4e86a415  // smmla v21.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a45d  // smmla v29.4s, v2.16b, v6.16b\n"
+      "add x28, x28, #0x80\n"
       ".inst 0x4e85a412  // smmla v18.4s, v0.16b, v5.16b\n"
       ".inst 0x4e85a45a  // smmla v26.4s, v2.16b, v5.16b\n"
       ".inst 0x4e84a416  // smmla v22.4s, v0.16b, v4.16b\n"
@@ -1120,24 +1120,24 @@ void a64_hybrid_s8qa_mmla_4x16 (
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       "83:"  // Height 3: Multiply loop: unique 12: skip row sum
       "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
+      "ldr q3, [x28, #0x10]\n"
+      ".inst 0x4e81a410  // smmla v16.4s, v0.16b, v1.16b\n"
+      ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
+      "ldr q1, [x28, #0x20]\n"
       "ldr q6, [x28, #0x30]\n"
+      ".inst 0x4e83a414  // smmla v20.4s, v0.16b, v3.16b\n"
+      ".inst 0x4e83a45c  // smmla v28.4s, v2.16b, v3.16b\n"
       "ldr q5, [x28, #0x40]\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x4e81a411  // smmla v17.4s, v0.16b, v1.16b\n"
+      ".inst 0x4e81a459  // smmla v25.4s, v2.16b, v1.16b\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x4e81a410  // smmla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x28, #0x70]\n"
-      ".inst 0x4e88a414  // smmla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e88a45c  // smmla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87a459  // smmla v25.4s, v2.16b, v7.16b\n"
       ".inst 0x4e86a415  // smmla v21.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a45d  // smmla v29.4s, v2.16b, v6.16b\n"
       ".inst 0x4e85a412  // smmla v18.4s, v0.16b, v5.16b\n"
       ".inst 0x4e85a45a  // smmla v26.4s, v2.16b, v5.16b\n"
+      "add x28, x28, #0x80\n"
       ".inst 0x4e84a416  // smmla v22.4s, v0.16b, v4.16b\n"
       ".inst 0x4e84a45e  // smmla v30.4s, v2.16b, v4.16b\n"
       ".inst 0x4e83a413  // smmla v19.4s, v0.16b, v3.16b\n"
@@ -1151,18 +1151,18 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "bne 68b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v0.2d, v16.2d, v20.2d\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v24.2d, v24.2d, v28.2d\n"
       "uzp1 v25.2d, v25.2d, v29.2d\n"
       "uzp1 v26.2d, v26.2d, v30.2d\n"
@@ -1170,9 +1170,9 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "mov v31.16b, v0.16b\n"
       "tbnz %x[flags], #31, 85f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v23.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v13.4s, v13.4s, v13.4s\n"
-      "ld1r { v23.4s }, [x20]\n"
       "neg v23.4s, v23.4s\n"
       "dup v12.4s, v11.s[3]\n"
       "dup v11.4s, v11.s[0]\n"
@@ -1192,10 +1192,10 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "add v16.4s, v16.4s, v12.4s\n"
       "add v17.4s, v17.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v23.4s }, [x20]\n"
       "add v18.4s, v18.4s, v12.4s\n"
       "add v19.4s, v19.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v24.4s, v24.4s, v13.4s\n"
       "add v25.4s, v25.4s, v13.4s\n"
@@ -1233,18 +1233,18 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "and v29.16b, v21.16b, v0.16b\n"
       "and v28.16b, v22.16b, v0.16b\n"
       "and v23.16b, v16.16b, v0.16b\n"
-      "and v3.16b, v17.16b, v0.16b\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v23.4s, v23.4s, #0x1f\n"
-      "and v2.16b, v18.16b, v0.16b\n"
       "sqadd v31.4s, v31.4s, v1.4s\n"
       "sqadd v20.4s, v20.4s, v30.4s\n"
       "sqadd v21.4s, v21.4s, v29.4s\n"
       "sqadd v22.4s, v22.4s, v28.4s\n"
       "sqadd v16.4s, v16.4s, v23.4s\n"
+      "and v3.16b, v17.16b, v0.16b\n"
+      "and v2.16b, v18.16b, v0.16b\n"
       "and v1.16b, v19.16b, v0.16b\n"
       "and v30.16b, v24.16b, v0.16b\n"
       "and v29.16b, v25.16b, v0.16b\n"
@@ -1265,21 +1265,21 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "sqadd v26.4s, v26.4s, v28.4s\n"
       "sqadd v27.4s, v27.4s, v23.4s\n"
       "86:"  // Height 3: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "srshl v31.4s, v31.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v29.4s }, [x21]\n"
-      "ld1r { v28.4s }, [x20]\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v28.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v23.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "srshl v24.4s, v24.4s, v0.4s\n"
       "srshl v25.4s, v25.4s, v0.4s\n"
       "srshl v26.4s, v26.4s, v0.4s\n"
@@ -1332,103 +1332,102 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "bge 95f\n"
       "tbz x9, #3, 90f\n"
       "str d31, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x9, #2, 88f\n"
       "st1 { v31.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
       "tbz x9, #1, 87f\n"
       "st1 { v31.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "st1 { v16.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 94f\n"
       "87:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "st1 { v16.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 94f\n"
       "88:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x9, #1, 89f\n"
       "st1 { v31.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "st1 { v16.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 94f\n"
       "89:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "st1 { v16.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 94f\n"
       "90:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x9, #2, 92f\n"
       "str s31, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
+      "str s16, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
       "tbz x9, #1, 91f\n"
       "st1 { v31.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "st1 { v16.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 94f\n"
       "91:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "st1 { v16.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 94f\n"
       "92:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x9, #1, 93f\n"
       "str h31, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
+      "str h16, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "st1 { v16.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 94f\n"
       "93:"  // Height 3: Partial direct writeback: partial_1_0
       "str b31, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "94:"  // Height 3: Partial direct writeback: Done
       "b 96f\n"
       "95:"  // Height 3: Full writeback
       "str q31, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "96:"  // Height 3: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 66b\n"
       "b 130f\n"
       "97:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x4\n"
       "mov x10, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
+      "movi v13.4s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "movi v13.4s, #0x0\n"
       "movi v14.4s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
       "movi v15.16b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "98:"  // Height 4: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -1450,8 +1449,8 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "mov x26, #0x0\n"
       "100:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 101f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1489,28 +1488,28 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "103:"  // Height 4: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
       "ldr q4, [x28, #0x60]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x4e85a458  // smmla v24.4s, v2.16b, v5.16b\n"
       "ldr q5, [x28, #0x70]\n"
+      ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a45c  // smmla v28.4s, v2.16b, v6.16b\n"
       "ldr q6, [x28, #0x80]\n"
+      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x4e87a459  // smmla v25.4s, v2.16b, v7.16b\n"
       "ldr q7, [x28, #0x90]\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x4e88a45d  // smmla v29.4s, v2.16b, v8.16b\n"
       "ldr q8, [x28, #0xa0]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4e89a412  // smmla v18.4s, v0.16b, v9.16b\n"
       ".inst 0x4e89a45a  // smmla v26.4s, v2.16b, v9.16b\n"
       "ldr q9, [x28, #0xb0]\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x4e8aa416  // smmla v22.4s, v0.16b, v10.16b\n"
       ".inst 0x4e8aa45e  // smmla v30.4s, v2.16b, v10.16b\n"
       "ldr q10, [x28, #0xc0]\n"
@@ -1547,9 +1546,9 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "ldr q1, [x24, #0x0]\n"
       "ldr q2, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q3, [x22, #0x0]\n"
       "ldr q4, [x21, #0x0]\n"
-      "cmp x25, #0x20\n"
       "ldr q5, [x28, #0x0]\n"
       "ldr q6, [x28, #0x10]\n"
       "ldr q7, [x28, #0x20]\n"
@@ -1564,32 +1563,32 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "105:"  // Height 4: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
       "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
       "ldr q4, [x28, #0x60]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x4e85a410  // smmla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x4e85a458  // smmla v24.4s, v2.16b, v5.16b\n"
       "ldr q5, [x28, #0x70]\n"
+      ".inst 0x4e86a414  // smmla v20.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a45c  // smmla v28.4s, v2.16b, v6.16b\n"
       "ldr q6, [x28, #0x80]\n"
+      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x4e87a459  // smmla v25.4s, v2.16b, v7.16b\n"
       "ldr q7, [x28, #0x90]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x4e88a415  // smmla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x4e88a45d  // smmla v29.4s, v2.16b, v8.16b\n"
       "ldr q8, [x28, #0xa0]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e89a412  // smmla v18.4s, v0.16b, v9.16b\n"
       ".inst 0x4e89a45a  // smmla v26.4s, v2.16b, v9.16b\n"
       "ldr q9, [x28, #0xb0]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4e8aa416  // smmla v22.4s, v0.16b, v10.16b\n"
       ".inst 0x4e8aa45e  // smmla v30.4s, v2.16b, v10.16b\n"
       "ldr q10, [x28, #0xc0]\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x4e84a413  // smmla v19.4s, v0.16b, v4.16b\n"
       ".inst 0x4e84a45b  // smmla v27.4s, v2.16b, v4.16b\n"
       "ldr q4, [x28, #0xd0]\n"
@@ -1629,35 +1628,35 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "cmp x25, #0x8\n"
       "blt 110f\n"
       "108:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d3, [x24], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
       "ldr d0, [x23], #0x8\n"
+      "trn1 v0.2d, v1.2d, v0.2d\n"
       "ldr d2, [x22], #0x8\n"
       "ldr d1, [x21], #0x8\n"
-      "trn1 v0.2d, v3.2d, v0.2d\n"
       "trn1 v2.2d, v2.2d, v1.2d\n"
       "tbnz %x[flags], #31, 109f\n"
       ".inst 0x4e8f940b  // sdot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       "109:"  // Height 4: Multiply loop: unique 15: skip row sum
-      "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "sub x25, x25, #0x8\n"
+      "ldr q3, [x28, #0x0]\n"
+      "ldr q1, [x28, #0x10]\n"
+      ".inst 0x4e83a410  // smmla v16.4s, v0.16b, v3.16b\n"
+      ".inst 0x4e83a458  // smmla v24.4s, v2.16b, v3.16b\n"
       "ldr q7, [x28, #0x20]\n"
       "ldr q6, [x28, #0x30]\n"
+      "sub x25, x25, #0x8\n"
       "cmp x25, #0x8\n"
       "ldr q5, [x28, #0x40]\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x4e81a414  // smmla v20.4s, v0.16b, v1.16b\n"
+      ".inst 0x4e81a45c  // smmla v28.4s, v2.16b, v1.16b\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x4e81a410  // smmla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x28, #0x70]\n"
-      ".inst 0x4e88a414  // smmla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e88a45c  // smmla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
       ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x4e87a459  // smmla v25.4s, v2.16b, v7.16b\n"
       ".inst 0x4e86a415  // smmla v21.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a45d  // smmla v29.4s, v2.16b, v6.16b\n"
+      "add x28, x28, #0x80\n"
       ".inst 0x4e85a412  // smmla v18.4s, v0.16b, v5.16b\n"
       ".inst 0x4e85a45a  // smmla v26.4s, v2.16b, v5.16b\n"
       ".inst 0x4e84a416  // smmla v22.4s, v0.16b, v4.16b\n"
@@ -1717,24 +1716,24 @@ void a64_hybrid_s8qa_mmla_4x16 (
       ".inst 0x4e8f944d  // sdot v13.4s, v2.16b, v15.16b\n"
       "115:"  // Height 4: Multiply loop: unique 16: skip row sum
       "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
+      "ldr q3, [x28, #0x10]\n"
+      ".inst 0x4e81a410  // smmla v16.4s, v0.16b, v1.16b\n"
+      ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
+      "ldr q1, [x28, #0x20]\n"
       "ldr q6, [x28, #0x30]\n"
+      ".inst 0x4e83a414  // smmla v20.4s, v0.16b, v3.16b\n"
+      ".inst 0x4e83a45c  // smmla v28.4s, v2.16b, v3.16b\n"
       "ldr q5, [x28, #0x40]\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x4e81a411  // smmla v17.4s, v0.16b, v1.16b\n"
+      ".inst 0x4e81a459  // smmla v25.4s, v2.16b, v1.16b\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x4e81a410  // smmla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x28, #0x70]\n"
-      ".inst 0x4e88a414  // smmla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e88a45c  // smmla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e87a411  // smmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87a459  // smmla v25.4s, v2.16b, v7.16b\n"
       ".inst 0x4e86a415  // smmla v21.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a45d  // smmla v29.4s, v2.16b, v6.16b\n"
       ".inst 0x4e85a412  // smmla v18.4s, v0.16b, v5.16b\n"
       ".inst 0x4e85a45a  // smmla v26.4s, v2.16b, v5.16b\n"
+      "add x28, x28, #0x80\n"
       ".inst 0x4e84a416  // smmla v22.4s, v0.16b, v4.16b\n"
       ".inst 0x4e84a45e  // smmla v30.4s, v2.16b, v4.16b\n"
       ".inst 0x4e83a413  // smmla v19.4s, v0.16b, v3.16b\n"
@@ -1748,22 +1747,22 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "bne 100b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v0.2d, v16.2d, v20.2d\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
+      "add x21, x22, x20\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
-      "add x22, x23, x20\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v23.2d, v24.2d, v28.2d\n"
       "uzp2 v24.2d, v24.2d, v28.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v28.2d, v25.2d, v29.2d\n"
       "uzp2 v25.2d, v25.2d, v29.2d\n"
       "uzp1 v29.2d, v26.2d, v30.2d\n"
@@ -1773,9 +1772,9 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "mov v31.16b, v0.16b\n"
       "tbnz %x[flags], #31, 117f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v13.4s, v13.4s, v13.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
       "neg v0.4s, v0.4s\n"
       "dup v12.4s, v11.s[3]\n"
       "dup v11.4s, v11.s[0]\n"
@@ -1783,8 +1782,8 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "dup v13.4s, v13.s[0]\n"
       "mul v11.4s, v11.4s, v0.4s\n"
       "mul v12.4s, v12.4s, v0.4s\n"
-      "mul v14.4s, v14.4s, v0.4s\n"
       "mul v13.4s, v13.4s, v0.4s\n"
+      "mul v14.4s, v14.4s, v0.4s\n"
       "117:"  // Height 4: skip row sum fixup
       "ldr q0, [x10, #0x0]\n"
       "ldr q4, [x10, #0x10]\n"
@@ -1797,10 +1796,10 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "add v16.4s, v16.4s, v12.4s\n"
       "add v17.4s, v17.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v1.4s }, [x20]\n"
       "add v18.4s, v18.4s, v12.4s\n"
       "add v19.4s, v19.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v23.4s, v23.4s, v13.4s\n"
       "add v28.4s, v28.4s, v13.4s\n"
@@ -1847,32 +1846,32 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "tbz %x[flags], #5, 118f\n"
       "and v2.16b, v31.16b, v0.16b\n"
       "and v1.16b, v20.16b, v0.16b\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v31.4s, v31.4s, v2.4s\n"
+      "sqadd v20.4s, v20.4s, v1.4s\n"
       "and v7.16b, v21.16b, v0.16b\n"
       "and v6.16b, v22.16b, v0.16b\n"
       "and v5.16b, v16.16b, v0.16b\n"
       "and v4.16b, v17.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
       "and v3.16b, v18.16b, v0.16b\n"
+      "and v2.16b, v19.16b, v0.16b\n"
+      "and v1.16b, v23.16b, v0.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
       "sshr v6.4s, v6.4s, #0x1f\n"
       "sshr v5.4s, v5.4s, #0x1f\n"
-      "sqadd v31.4s, v31.4s, v2.4s\n"
-      "sqadd v20.4s, v20.4s, v1.4s\n"
-      "and v2.16b, v19.16b, v0.16b\n"
-      "and v1.16b, v23.16b, v0.16b\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sshr v3.4s, v3.4s, #0x1f\n"
-      "sqadd v21.4s, v21.4s, v7.4s\n"
-      "sqadd v22.4s, v22.4s, v6.4s\n"
       "sshr v2.4s, v2.4s, #0x1f\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v21.4s, v21.4s, v7.4s\n"
+      "sqadd v22.4s, v22.4s, v6.4s\n"
       "sqadd v16.4s, v16.4s, v5.4s\n"
       "sqadd v17.4s, v17.4s, v4.4s\n"
       "sqadd v18.4s, v18.4s, v3.4s\n"
-      "and v7.16b, v28.16b, v0.16b\n"
       "sqadd v19.4s, v19.4s, v2.4s\n"
       "sqadd v23.4s, v23.4s, v1.4s\n"
+      "and v7.16b, v28.16b, v0.16b\n"
       "and v6.16b, v29.16b, v0.16b\n"
       "and v5.16b, v30.16b, v0.16b\n"
       "and v4.16b, v24.16b, v0.16b\n"
@@ -1894,21 +1893,21 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "sqadd v26.4s, v26.4s, v2.4s\n"
       "sqadd v27.4s, v27.4s, v1.4s\n"
       "118:"  // Height 4: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v3.4s }, [x20]\n"
       "srshl v31.4s, v31.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v3.4s }, [x21]\n"
-      "ld1r { v2.4s }, [x20]\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v2.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v1.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "srshl v28.4s, v28.4s, v0.4s\n"
       "srshl v29.4s, v29.4s, v0.4s\n"
@@ -1980,100 +1979,100 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "bge 127f\n"
       "tbz x9, #3, 122f\n"
       "str d31, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
+      "str d24, [x21], #0x8\n"
       "tbz x9, #2, 120f\n"
       "st1 { v31.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "st1 { v23.s }[2], [x23], #0x4\n"
-      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
+      "st1 { v23.s }[2], [x22], #0x4\n"
+      "st1 { v24.s }[2], [x21], #0x4\n"
       "tbz x9, #1, 119f\n"
       "st1 { v31.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "st1 { v23.h }[6], [x23], #0x2\n"
-      "st1 { v24.h }[6], [x22], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
+      "st1 { v23.h }[6], [x22], #0x2\n"
+      "st1 { v24.h }[6], [x21], #0x2\n"
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "st1 { v23.b }[14], [x23]\n"
-      "st1 { v24.b }[14], [x22]\n"
+      "st1 { v16.b }[14], [x23]\n"
+      "st1 { v23.b }[14], [x22]\n"
+      "st1 { v24.b }[14], [x21]\n"
       "b 126f\n"
       "119:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "st1 { v23.b }[12], [x23]\n"
-      "st1 { v24.b }[12], [x22]\n"
+      "st1 { v16.b }[12], [x23]\n"
+      "st1 { v23.b }[12], [x22]\n"
+      "st1 { v24.b }[12], [x21]\n"
       "b 126f\n"
       "120:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x9, #1, 121f\n"
       "st1 { v31.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "st1 { v23.h }[4], [x23], #0x2\n"
-      "st1 { v24.h }[4], [x22], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
+      "st1 { v23.h }[4], [x22], #0x2\n"
+      "st1 { v24.h }[4], [x21], #0x2\n"
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "st1 { v23.b }[10], [x23]\n"
-      "st1 { v24.b }[10], [x22]\n"
+      "st1 { v16.b }[10], [x23]\n"
+      "st1 { v23.b }[10], [x22]\n"
+      "st1 { v24.b }[10], [x21]\n"
       "b 126f\n"
       "121:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "st1 { v23.b }[8], [x23]\n"
-      "st1 { v24.b }[8], [x22]\n"
+      "st1 { v16.b }[8], [x23]\n"
+      "st1 { v23.b }[8], [x22]\n"
+      "st1 { v24.b }[8], [x21]\n"
       "b 126f\n"
       "122:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x9, #2, 124f\n"
       "str s31, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "str s23, [x23], #0x4\n"
-      "str s24, [x22], #0x4\n"
+      "str s16, [x23], #0x4\n"
+      "str s23, [x22], #0x4\n"
+      "str s24, [x21], #0x4\n"
       "tbz x9, #1, 123f\n"
       "st1 { v31.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "st1 { v23.h }[2], [x23], #0x2\n"
-      "st1 { v24.h }[2], [x22], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
+      "st1 { v23.h }[2], [x22], #0x2\n"
+      "st1 { v24.h }[2], [x21], #0x2\n"
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "st1 { v23.b }[6], [x23]\n"
-      "st1 { v24.b }[6], [x22]\n"
+      "st1 { v16.b }[6], [x23]\n"
+      "st1 { v23.b }[6], [x22]\n"
+      "st1 { v24.b }[6], [x21]\n"
       "b 126f\n"
       "123:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "st1 { v23.b }[4], [x23]\n"
-      "st1 { v24.b }[4], [x22]\n"
+      "st1 { v16.b }[4], [x23]\n"
+      "st1 { v23.b }[4], [x22]\n"
+      "st1 { v24.b }[4], [x21]\n"
       "b 126f\n"
       "124:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x9, #1, 125f\n"
       "str h31, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "str h23, [x23], #0x2\n"
-      "str h24, [x22], #0x2\n"
+      "str h16, [x23], #0x2\n"
+      "str h23, [x22], #0x2\n"
+      "str h24, [x21], #0x2\n"
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "st1 { v23.b }[2], [x23]\n"
-      "st1 { v24.b }[2], [x22]\n"
+      "st1 { v16.b }[2], [x23]\n"
+      "st1 { v23.b }[2], [x22]\n"
+      "st1 { v24.b }[2], [x21]\n"
       "b 126f\n"
       "125:"  // Height 4: Partial direct writeback: partial_1_0
       "str b31, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "str b23, [x23, #0x0]\n"
-      "str b24, [x22, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
+      "str b23, [x22, #0x0]\n"
+      "str b24, [x21, #0x0]\n"
       "126:"  // Height 4: Partial direct writeback: Done
       "b 128f\n"
       "127:"  // Height 4: Full writeback
       "str q31, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
-      "str q23, [x23, #0x0]\n"
-      "str q24, [x22, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q23, [x22, #0x0]\n"
+      "str q24, [x21, #0x0]\n"
       "128:"  // Height 4: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 98b\n"
@@ -2089,8 +2088,8 @@ void a64_hybrid_s8qa_mmla_4x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "130:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16.hpp
index 51057a6ffc..0ec35f7f76 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16.hpp
@@ -71,7 +71,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp
index 559b492871..f8f6579d8c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp
@@ -47,18 +47,18 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -97,9 +97,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "mov x6, %x[col_bias]\n"
       "ldr x7, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -109,8 +109,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "mov x14, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w13, [x20, x14, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -126,118 +126,118 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "blt 9f\n"
       "ldr q0, [x12, #0x0]\n"
       "cmp x13, #0x20\n"
-      "ldr q6, [x16, #0x0]\n"
-      "ldr q7, [x16, #0x10]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "ldr q7, [x15, #0x10]\n"
       "blt 8f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr d17, [x16, #0x20]\n"
-      "ldr x21, [x16, #0x28]\n"
+      "ldr d17, [x15, #0x20]\n"
+      "ldr x20, [x15, #0x28]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr d16, [x16, #0x30]\n"
-      "add x12, x12, #0x10\n"
-      "ldr x20, [x16, #0x38]\n"
-      "sub x13, x13, #0x10\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x16, #0x48]\n"
-      "ldr x22, [x12, #0x8]\n"
-      "cmp x13, #0x20\n"
+      "ldr d16, [x15, #0x30]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0x38]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr d17, [x16, #0x40]\n"
+      "ldr d17, [x15, #0x40]\n"
+      "ldr x20, [x15, #0x48]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr d16, [x16, #0x50]\n"
-      "ldr x20, [x16, #0x58]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x16, #0x68]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "ldr d16, [x15, #0x50]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0x58]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr d17, [x16, #0x60]\n"
+      "ldr d17, [x15, #0x60]\n"
+      "ldr x20, [x15, #0x68]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr d16, [x16, #0x70]\n"
-      "ldr x20, [x16, #0x78]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x16, #0x88]\n"
+      "ldr d16, [x15, #0x70]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0x78]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr d17, [x16, #0x80]\n"
+      "ldr d17, [x15, #0x80]\n"
+      "ldr x20, [x15, #0x88]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr d16, [x16, #0x90]\n"
-      "ldr x20, [x16, #0x98]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x16, #0xa8]\n"
+      "ldr d16, [x15, #0x90]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0x98]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr d17, [x16, #0xa0]\n"
+      "ldr d17, [x15, #0xa0]\n"
+      "ldr x20, [x15, #0xa8]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr d16, [x16, #0xb0]\n"
-      "ldr x20, [x16, #0xb8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x16, #0xc8]\n"
+      "ldr d16, [x15, #0xb0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0xb8]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr d17, [x16, #0xc0]\n"
+      "ldr d17, [x15, #0xc0]\n"
+      "ldr x20, [x15, #0xc8]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr d16, [x16, #0xd0]\n"
-      "ldr x20, [x16, #0xd8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x16, #0xe8]\n"
+      "ldr d16, [x15, #0xd0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0xd8]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr d17, [x16, #0xe0]\n"
+      "ldr d17, [x15, #0xe0]\n"
+      "ldr x20, [x15, #0xe8]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr d16, [x16, #0xf0]\n"
-      "ldr x20, [x16, #0xf8]\n"
-      "add x16, x16, #0x100\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x16, #0x8]\n"
+      "ldr d16, [x15, #0xf0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x15, #0xf8]\n"
       "mov v16.d[1], x20\n"
+      "add x12, x12, #0x10\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
-      "ldr d6, [x16, #0x0]\n"
+      "ldr d6, [x15, #0x0]\n"
+      "ldr x20, [x15, #0x8]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr d0, [x12, #0x0]\n"
-      "ldr d7, [x16, #0x10]\n"
-      "ldr x20, [x16, #0x18]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x22\n"
+      "sub x13, x13, #0x10\n"
+      "ldr d7, [x15, #0x10]\n"
+      "cmp x13, #0x20\n"
+      "ldr x21, [x12, #0x8]\n"
+      "mov v6.d[1], x20\n"
+      "ldr x20, [x15, #0x18]\n"
+      "mov v0.d[1], x21\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "bge 7b\n"
       "8:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr q17, [x16, #0x20]\n"
+      "ldr q17, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr q16, [x16, #0x30]\n"
-      "add x12, x12, #0x10\n"
-      "sub x13, x13, #0x10\n"
+      "ldr q16, [x15, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x16, #0x40]\n"
+      "ldr q17, [x15, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x16, #0x50]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "ldr q16, [x15, #0x50]\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x16, #0x60]\n"
+      "ldr q17, [x15, #0x60]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x16, #0x70]\n"
+      "ldr q16, [x15, #0x70]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x16, #0x80]\n"
+      "ldr q17, [x15, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x16, #0x90]\n"
+      "ldr q16, [x15, #0x90]\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x16, #0xa0]\n"
+      "ldr q17, [x15, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x16, #0xb0]\n"
+      "ldr q16, [x15, #0xb0]\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x16, #0xc0]\n"
+      "ldr q17, [x15, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x16, #0xd0]\n"
+      "ldr q16, [x15, #0xd0]\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr q17, [x16, #0xe0]\n"
+      "ldr q17, [x15, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr q16, [x16, #0xf0]\n"
-      "add x16, x16, #0x100\n"
+      "ldr q16, [x15, #0xf0]\n"
+      "add x12, x12, #0x10\n"
+      "sub x13, x13, #0x10\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
+      "add x15, x15, #0x100\n"
       "9:"  // Height 1: Multiply loop: Main loop skip
       "cbz x13, 14f\n"
       "cmp x13, #0x4\n"
@@ -245,16 +245,16 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "10:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x12], #0x4\n"
       "sub x13, x13, #0x4\n"
-      "ldr q17, [x16, #0x0]\n"
-      "cmp x13, #0x4\n"
-      "ldr q16, [x16, #0x10]\n"
-      ".inst 0x4f92e228  // sdot v8.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x16, #0x20]\n"
+      "ldr q16, [x15, #0x0]\n"
+      ".inst 0x4f92e208  // sdot v8.4s, v16.16b, v18.4b[0]\n"
+      "ldr q16, [x15, #0x10]\n"
       ".inst 0x4f92e209  // sdot v9.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "ldr q17, [x15, #0x20]\n"
+      "cmp x13, #0x4\n"
+      "ldr q16, [x15, #0x30]\n"
       ".inst 0x4f92e22a  // sdot v10.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x4f92e20b  // sdot v11.4s, v16.16b, v18.4b[0]\n"
+      "add x15, x15, #0x40\n"
       "bge 10b\n"
       "11:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x13, 14f\n"
@@ -266,30 +266,30 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "12:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr b0, [x12, #0x0]\n"
       "13:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x16, #0x0]\n"
-      "ldr q16, [x16, #0x10]\n"
-      ".inst 0x4f80e228  // sdot v8.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x16, #0x20]\n"
+      "ldr q16, [x15, #0x0]\n"
+      ".inst 0x4f80e208  // sdot v8.4s, v16.16b, v0.4b[0]\n"
+      "ldr q16, [x15, #0x10]\n"
       ".inst 0x4f80e209  // sdot v9.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
-      ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
+      "ldr q16, [x15, #0x20]\n"
+      ".inst 0x4f80e20a  // sdot v10.4s, v16.16b, v0.4b[0]\n"
+      "ldr q16, [x15, #0x30]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
+      "add x15, x15, #0x40\n"
       "14:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x14, x14, #0x1\n"
       "cmp x14, x20\n"
       "bne 4b\n"
-      "ldr q19, [x6, #0x0]\n"
-      "ldr q18, [x6, #0x10]\n"
-      "ldr q17, [x6, #0x20]\n"
+      "ldr q16, [x6, #0x0]\n"
+      "add v8.4s, v8.4s, v16.4s\n"
+      "ldr q16, [x6, #0x10]\n"
+      "add v9.4s, v9.4s, v16.4s\n"
+      "ldr q16, [x6, #0x20]\n"
+      "add v10.4s, v10.4s, v16.4s\n"
       "ldr q16, [x6, #0x30]\n"
-      "add v8.4s, v8.4s, v19.4s\n"
-      "prfm pstl1keep, [x15, #0x0]\n"
-      "add v9.4s, v9.4s, v18.4s\n"
-      "add v10.4s, v10.4s, v17.4s\n"
-      "add x6, x6, #0x40\n"
       "add v11.4s, v11.4s, v16.4s\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "add x6, x6, #0x40\n"
       "tbz %x[flags], #4, 15f\n"
       "ldr q0, [x8, #0x0]\n"
       "ldr q4, [x7, #0x0]\n"
@@ -303,9 +303,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "add x7, x7, #0x40\n"
       "b 16f\n"
       "15:"  // Height 1: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -337,87 +337,87 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v18.4s }, [x20]\n"
+      "ld1r { v16.4s }, [x20]\n"
+      "add v8.4s, v8.4s, v16.4s\n"
+      "add v9.4s, v9.4s, v16.4s\n"
+      "add v10.4s, v10.4s, v16.4s\n"
+      "add v11.4s, v11.4s, v16.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v16.4s }, [x20]\n"
+      "smin v8.4s, v8.4s, v16.4s\n"
+      "smin v9.4s, v9.4s, v16.4s\n"
+      "smin v10.4s, v10.4s, v16.4s\n"
+      "smin v11.4s, v11.4s, v16.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "cmp x17, #0x10\n"
       "ld1r { v16.4s }, [x20]\n"
-      "add v8.4s, v8.4s, v18.4s\n"
-      "add v9.4s, v9.4s, v18.4s\n"
-      "add v10.4s, v10.4s, v18.4s\n"
-      "add v11.4s, v11.4s, v18.4s\n"
-      "smin v8.4s, v8.4s, v17.4s\n"
-      "smin v9.4s, v9.4s, v17.4s\n"
-      "smin v10.4s, v10.4s, v17.4s\n"
-      "smin v11.4s, v11.4s, v17.4s\n"
       "smax v8.4s, v8.4s, v16.4s\n"
       "smax v9.4s, v9.4s, v16.4s\n"
       "smax v10.4s, v10.4s, v16.4s\n"
       "smax v11.4s, v11.4s, v16.4s\n"
       "uzp1 v8.8h, v8.8h, v9.8h\n"
       "uzp1 v16.8h, v10.8h, v11.8h\n"
+      "cmp x16, #0x10\n"
       "uzp1 v8.16b, v8.16b, v16.16b\n"
       "bge 26f\n"
-      "tbz x17, #3, 21f\n"
-      "str d8, [x15], #0x8\n"
-      "tbz x17, #2, 19f\n"
-      "st1 { v8.s }[2], [x15], #0x4\n"
-      "tbz x17, #1, 18f\n"
-      "st1 { v8.h }[6], [x15], #0x2\n"
-      "tbz x17, #0, 25f\n"
-      "st1 { v8.b }[14], [x15]\n"
+      "tbz x16, #3, 21f\n"
+      "str d8, [x17], #0x8\n"
+      "tbz x16, #2, 19f\n"
+      "st1 { v8.s }[2], [x17], #0x4\n"
+      "tbz x16, #1, 18f\n"
+      "st1 { v8.h }[6], [x17], #0x2\n"
+      "tbz x16, #0, 25f\n"
+      "st1 { v8.b }[14], [x17]\n"
       "b 25f\n"
       "18:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x17, #0, 25f\n"
-      "st1 { v8.b }[12], [x15]\n"
+      "tbz x16, #0, 25f\n"
+      "st1 { v8.b }[12], [x17]\n"
       "b 25f\n"
       "19:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x17, #1, 20f\n"
-      "st1 { v8.h }[4], [x15], #0x2\n"
-      "tbz x17, #0, 25f\n"
-      "st1 { v8.b }[10], [x15]\n"
+      "tbz x16, #1, 20f\n"
+      "st1 { v8.h }[4], [x17], #0x2\n"
+      "tbz x16, #0, 25f\n"
+      "st1 { v8.b }[10], [x17]\n"
       "b 25f\n"
       "20:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x17, #0, 25f\n"
-      "st1 { v8.b }[8], [x15]\n"
+      "tbz x16, #0, 25f\n"
+      "st1 { v8.b }[8], [x17]\n"
       "b 25f\n"
       "21:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x17, #2, 23f\n"
-      "str s8, [x15], #0x4\n"
-      "tbz x17, #1, 22f\n"
-      "st1 { v8.h }[2], [x15], #0x2\n"
-      "tbz x17, #0, 25f\n"
-      "st1 { v8.b }[6], [x15]\n"
+      "tbz x16, #2, 23f\n"
+      "str s8, [x17], #0x4\n"
+      "tbz x16, #1, 22f\n"
+      "st1 { v8.h }[2], [x17], #0x2\n"
+      "tbz x16, #0, 25f\n"
+      "st1 { v8.b }[6], [x17]\n"
       "b 25f\n"
       "22:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x17, #0, 25f\n"
-      "st1 { v8.b }[4], [x15]\n"
+      "tbz x16, #0, 25f\n"
+      "st1 { v8.b }[4], [x17]\n"
       "b 25f\n"
       "23:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x17, #1, 24f\n"
-      "str h8, [x15], #0x2\n"
-      "tbz x17, #0, 25f\n"
-      "st1 { v8.b }[2], [x15]\n"
+      "tbz x16, #1, 24f\n"
+      "str h8, [x17], #0x2\n"
+      "tbz x16, #0, 25f\n"
+      "st1 { v8.b }[2], [x17]\n"
       "b 25f\n"
       "24:"  // Height 1: Partial direct writeback: partial_1_0
-      "str b8, [x15, #0x0]\n"
+      "str b8, [x17, #0x0]\n"
       "25:"  // Height 1: Partial direct writeback: Done
       "b 27f\n"
       "26:"  // Height 1: Full writeback
-      "str q8, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
+      "str q8, [x17, #0x0]\n"
+      "add x17, x17, #0x10\n"
       "27:"  // Height 1: Writeback done
-      "subs x17, x17, #0x10\n"
+      "subs x16, x16, #0x10\n"
       "bgt 2b\n"
       "b 164f\n"
       "28:"  // Height 2
       "mov x6, %x[col_bias]\n"
       "ldr x7, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "29:"  // Height 2: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -431,8 +431,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "mov x14, #0x0\n"
       "31:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w13, [x20, x14, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 32f\n"
       "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -452,154 +452,154 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr q0, [x12, #0x0]\n"
       "cmp x13, #0x20\n"
       "ldr q1, [x11, #0x0]\n"
-      "ldr q6, [x16, #0x0]\n"
-      "ldr q7, [x16, #0x10]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "ldr q7, [x15, #0x10]\n"
       "blt 35f\n"
       "34:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x20, [x16, #0x28]\n"
+      "ldr x21, [x15, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr d17, [x16, #0x20]\n"
+      "ldr d17, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr x21, [x16, #0x38]\n"
+      "ldr x20, [x15, #0x38]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr d16, [x16, #0x30]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x16, #0x48]\n"
-      "add x12, x12, #0x10\n"
-      "add x11, x11, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x15, #0x30]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr d17, [x16, #0x40]\n"
+      "ldr d17, [x15, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr x21, [x16, #0x58]\n"
+      "ldr x20, [x15, #0x48]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr d16, [x16, #0x50]\n"
+      "ldr d16, [x15, #0x50]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x16, #0x68]\n"
-      "ldr x23, [x12, #0x8]\n"
-      "sub x13, x13, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x15, #0x58]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
+      "ldr x21, [x15, #0x68]\n"
       ".inst 0x4fa1e22c  // sdot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr d17, [x16, #0x60]\n"
+      "ldr d17, [x15, #0x60]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr x21, [x16, #0x78]\n"
+      "ldr x20, [x15, #0x78]\n"
       ".inst 0x4fa1e20d  // sdot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr d16, [x16, #0x70]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x16, #0x88]\n"
-      "ldr x22, [x11, #0x8]\n"
-      "cmp x13, #0x20\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x15, #0x70]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4fa1e22e  // sdot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr d17, [x16, #0x80]\n"
+      "ldr d17, [x15, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr x21, [x16, #0x98]\n"
+      "ldr x20, [x15, #0x88]\n"
       ".inst 0x4fa1e20f  // sdot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr d16, [x16, #0x90]\n"
+      "ldr d16, [x15, #0x90]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x16, #0xa8]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x15, #0x98]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
+      "ldr x21, [x15, #0xa8]\n"
       ".inst 0x4f81ea2c  // sdot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr d17, [x16, #0xa0]\n"
+      "ldr d17, [x15, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr x21, [x16, #0xb8]\n"
+      "ldr x20, [x15, #0xb8]\n"
       ".inst 0x4f81ea0d  // sdot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr d16, [x16, #0xb0]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x16, #0xc8]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x15, #0xb0]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4f81ea2e  // sdot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr d17, [x16, #0xc0]\n"
+      "ldr d17, [x15, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr x21, [x16, #0xd8]\n"
+      "ldr x20, [x15, #0xc8]\n"
       ".inst 0x4f81ea0f  // sdot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr d16, [x16, #0xd0]\n"
+      "ldr d16, [x15, #0xd0]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x16, #0xe8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x15, #0xd8]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
+      "ldr x21, [x15, #0xe8]\n"
       ".inst 0x4fa1ea2c  // sdot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr d17, [x16, #0xe0]\n"
+      "ldr d17, [x15, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr x21, [x16, #0xf8]\n"
+      "ldr x20, [x15, #0xf8]\n"
       ".inst 0x4fa1ea0d  // sdot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr d16, [x16, #0xf0]\n"
-      "mov v17.d[1], x20\n"
-      "add x16, x16, #0x100\n"
-      "ldr x20, [x16, #0x8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x15, #0xf0]\n"
+      "mov v17.d[1], x21\n"
+      "add x12, x12, #0x10\n"
+      "mov v16.d[1], x20\n"
+      "add x11, x11, #0x10\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea2e  // sdot v14.4s, v17.16b, v1.4b[3]\n"
-      "ldr d6, [x16, #0x0]\n"
+      "ldr d6, [x15, #0x0]\n"
+      "ldr x21, [x15, #0x8]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr d0, [x12, #0x0]\n"
       ".inst 0x4fa1ea0f  // sdot v15.4s, v16.16b, v1.4b[3]\n"
       "ldr d1, [x11, #0x0]\n"
-      "ldr d7, [x16, #0x10]\n"
-      "mov v6.d[1], x20\n"
-      "ldr x20, [x16, #0x18]\n"
-      "mov v0.d[1], x23\n"
-      "mov v1.d[1], x22\n"
+      "sub x13, x13, #0x10\n"
+      "ldr d7, [x15, #0x10]\n"
+      "cmp x13, #0x20\n"
+      "ldr x20, [x12, #0x8]\n"
+      "mov v6.d[1], x21\n"
+      "ldr x21, [x11, #0x8]\n"
+      "mov v0.d[1], x20\n"
+      "ldr x20, [x15, #0x18]\n"
+      "mov v1.d[1], x21\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "bge 34b\n"
       "35:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr q17, [x16, #0x20]\n"
+      "ldr q17, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr q16, [x16, #0x30]\n"
+      "ldr q16, [x15, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       "sub x13, x13, #0x10\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x16, #0x40]\n"
+      "ldr q17, [x15, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x16, #0x50]\n"
+      "ldr q16, [x15, #0x50]\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4fa1e22c  // sdot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x16, #0x60]\n"
+      "ldr q17, [x15, #0x60]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20d  // sdot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x16, #0x70]\n"
+      "ldr q16, [x15, #0x70]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22e  // sdot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x16, #0x80]\n"
+      "ldr q17, [x15, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20f  // sdot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x16, #0x90]\n"
+      "ldr q16, [x15, #0x90]\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x4f81ea2c  // sdot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x16, #0xa0]\n"
+      "ldr q17, [x15, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x4f81ea0d  // sdot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x16, #0xb0]\n"
+      "ldr q16, [x15, #0xb0]\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x4f81ea2e  // sdot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x16, #0xc0]\n"
+      "ldr q17, [x15, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x4f81ea0f  // sdot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x16, #0xd0]\n"
+      "ldr q16, [x15, #0xd0]\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea2c  // sdot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr q17, [x16, #0xe0]\n"
+      "ldr q17, [x15, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea0d  // sdot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr q16, [x16, #0xf0]\n"
+      "ldr q16, [x15, #0xf0]\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
-      "add x16, x16, #0x100\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa1ea2e  // sdot v14.4s, v17.16b, v1.4b[3]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea0f  // sdot v15.4s, v16.16b, v1.4b[3]\n"
@@ -612,16 +612,16 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "sub x13, x13, #0x4\n"
       "ldr s18, [x11], #0x4\n"
       "cmp x13, #0x4\n"
-      "ldr q17, [x16, #0x0]\n"
-      "ldr q16, [x16, #0x10]\n"
+      "ldr q17, [x15, #0x0]\n"
       ".inst 0x4f93e228  // sdot v8.4s, v17.16b, v19.4b[0]\n"
+      "ldr q16, [x15, #0x10]\n"
       ".inst 0x4f92e22c  // sdot v12.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x16, #0x20]\n"
+      "ldr q17, [x15, #0x20]\n"
       ".inst 0x4f93e209  // sdot v9.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x4f92e20d  // sdot v13.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x16, #0x30]\n"
+      "ldr q16, [x15, #0x30]\n"
       ".inst 0x4f93e22a  // sdot v10.4s, v17.16b, v19.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f92e22e  // sdot v14.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x4f93e20b  // sdot v11.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x4f92e20f  // sdot v15.4s, v16.16b, v18.4b[0]\n"
@@ -639,16 +639,16 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr b0, [x12, #0x0]\n"
       "ldr b1, [x11, #0x0]\n"
       "40:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x16, #0x0]\n"
-      "ldr q16, [x16, #0x10]\n"
+      "ldr q17, [x15, #0x0]\n"
       ".inst 0x4f80e228  // sdot v8.4s, v17.16b, v0.4b[0]\n"
+      "ldr q16, [x15, #0x10]\n"
       ".inst 0x4f81e22c  // sdot v12.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x16, #0x20]\n"
+      "ldr q17, [x15, #0x20]\n"
       ".inst 0x4f80e209  // sdot v9.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20d  // sdot v13.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x16, #0x30]\n"
+      "ldr q16, [x15, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
@@ -658,19 +658,19 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "cmp x14, x20\n"
       "bne 31b\n"
       "ldr q19, [x6, #0x0]\n"
-      "ldr q18, [x6, #0x10]\n"
-      "ldr q17, [x6, #0x20]\n"
-      "ldr q16, [x6, #0x30]\n"
       "add v8.4s, v8.4s, v19.4s\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr q18, [x6, #0x10]\n"
       "add v9.4s, v9.4s, v18.4s\n"
-      "prfm pstl1keep, [x15, #0x0]\n"
+      "ldr q17, [x6, #0x20]\n"
       "add v10.4s, v10.4s, v17.4s\n"
+      "ldr q16, [x6, #0x30]\n"
       "add v11.4s, v11.4s, v16.4s\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x17, x20\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
       "add v12.4s, v12.4s, v19.4s\n"
-      "add x26, x15, x20\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "add v13.4s, v13.4s, v18.4s\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "add v14.4s, v14.4s, v17.4s\n"
       "add v15.4s, v15.4s, v16.4s\n"
       "add x6, x6, #0x40\n"
@@ -687,9 +687,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "add x7, x7, #0x40\n"
       "b 43f\n"
       "42:"  // Height 2: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -741,28 +741,27 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "srshl v14.4s, v14.4s, v2.4s\n"
       "srshl v15.4s, v15.4s, v3.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v18.4s }, [x20]\n"
+      "ld1r { v16.4s }, [x20]\n"
+      "add v8.4s, v8.4s, v16.4s\n"
+      "add v9.4s, v9.4s, v16.4s\n"
+      "add v10.4s, v10.4s, v16.4s\n"
+      "add v11.4s, v11.4s, v16.4s\n"
+      "add v12.4s, v12.4s, v16.4s\n"
+      "add v13.4s, v13.4s, v16.4s\n"
+      "add v14.4s, v14.4s, v16.4s\n"
+      "add v15.4s, v15.4s, v16.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v16.4s }, [x20]\n"
+      "smin v8.4s, v8.4s, v16.4s\n"
+      "smin v9.4s, v9.4s, v16.4s\n"
+      "smin v10.4s, v10.4s, v16.4s\n"
+      "smin v11.4s, v11.4s, v16.4s\n"
+      "smin v12.4s, v12.4s, v16.4s\n"
+      "smin v13.4s, v13.4s, v16.4s\n"
+      "smin v14.4s, v14.4s, v16.4s\n"
+      "smin v15.4s, v15.4s, v16.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v17.4s }, [x21]\n"
-      "cmp x17, #0x10\n"
       "ld1r { v16.4s }, [x20]\n"
-      "add v8.4s, v8.4s, v18.4s\n"
-      "add v9.4s, v9.4s, v18.4s\n"
-      "add v10.4s, v10.4s, v18.4s\n"
-      "add v11.4s, v11.4s, v18.4s\n"
-      "add v12.4s, v12.4s, v18.4s\n"
-      "add v13.4s, v13.4s, v18.4s\n"
-      "add v14.4s, v14.4s, v18.4s\n"
-      "add v15.4s, v15.4s, v18.4s\n"
-      "smin v8.4s, v8.4s, v17.4s\n"
-      "smin v9.4s, v9.4s, v17.4s\n"
-      "smin v10.4s, v10.4s, v17.4s\n"
-      "smin v11.4s, v11.4s, v17.4s\n"
-      "smin v12.4s, v12.4s, v17.4s\n"
-      "smin v13.4s, v13.4s, v17.4s\n"
-      "smin v14.4s, v14.4s, v17.4s\n"
-      "smin v15.4s, v15.4s, v17.4s\n"
       "smax v8.4s, v8.4s, v16.4s\n"
       "smax v9.4s, v9.4s, v16.4s\n"
       "smax v10.4s, v10.4s, v16.4s\n"
@@ -775,84 +774,85 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "uzp1 v17.8h, v10.8h, v11.8h\n"
       "uzp1 v12.8h, v12.8h, v13.8h\n"
       "uzp1 v16.8h, v14.8h, v15.8h\n"
+      "cmp x16, #0x10\n"
       "uzp1 v8.16b, v8.16b, v17.16b\n"
       "uzp1 v12.16b, v12.16b, v16.16b\n"
       "bge 53f\n"
-      "tbz x17, #3, 48f\n"
-      "str d8, [x15], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "tbz x17, #2, 46f\n"
-      "st1 { v8.s }[2], [x15], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "tbz x17, #1, 45f\n"
-      "st1 { v8.h }[6], [x15], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "tbz x17, #0, 52f\n"
-      "st1 { v8.b }[14], [x15]\n"
-      "st1 { v12.b }[14], [x26]\n"
+      "tbz x16, #3, 48f\n"
+      "str d8, [x17], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "tbz x16, #2, 46f\n"
+      "st1 { v8.s }[2], [x17], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "tbz x16, #1, 45f\n"
+      "st1 { v8.h }[6], [x17], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "tbz x16, #0, 52f\n"
+      "st1 { v8.b }[14], [x17]\n"
+      "st1 { v12.b }[14], [x25]\n"
       "b 52f\n"
       "45:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x17, #0, 52f\n"
-      "st1 { v8.b }[12], [x15]\n"
-      "st1 { v12.b }[12], [x26]\n"
+      "tbz x16, #0, 52f\n"
+      "st1 { v8.b }[12], [x17]\n"
+      "st1 { v12.b }[12], [x25]\n"
       "b 52f\n"
       "46:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x17, #1, 47f\n"
-      "st1 { v8.h }[4], [x15], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "tbz x17, #0, 52f\n"
-      "st1 { v8.b }[10], [x15]\n"
-      "st1 { v12.b }[10], [x26]\n"
+      "tbz x16, #1, 47f\n"
+      "st1 { v8.h }[4], [x17], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "tbz x16, #0, 52f\n"
+      "st1 { v8.b }[10], [x17]\n"
+      "st1 { v12.b }[10], [x25]\n"
       "b 52f\n"
       "47:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x17, #0, 52f\n"
-      "st1 { v8.b }[8], [x15]\n"
-      "st1 { v12.b }[8], [x26]\n"
+      "tbz x16, #0, 52f\n"
+      "st1 { v8.b }[8], [x17]\n"
+      "st1 { v12.b }[8], [x25]\n"
       "b 52f\n"
       "48:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x17, #2, 50f\n"
-      "str s8, [x15], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "tbz x17, #1, 49f\n"
-      "st1 { v8.h }[2], [x15], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "tbz x17, #0, 52f\n"
-      "st1 { v8.b }[6], [x15]\n"
-      "st1 { v12.b }[6], [x26]\n"
+      "tbz x16, #2, 50f\n"
+      "str s8, [x17], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "tbz x16, #1, 49f\n"
+      "st1 { v8.h }[2], [x17], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "tbz x16, #0, 52f\n"
+      "st1 { v8.b }[6], [x17]\n"
+      "st1 { v12.b }[6], [x25]\n"
       "b 52f\n"
       "49:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x17, #0, 52f\n"
-      "st1 { v8.b }[4], [x15]\n"
-      "st1 { v12.b }[4], [x26]\n"
+      "tbz x16, #0, 52f\n"
+      "st1 { v8.b }[4], [x17]\n"
+      "st1 { v12.b }[4], [x25]\n"
       "b 52f\n"
       "50:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x17, #1, 51f\n"
-      "str h8, [x15], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "tbz x17, #0, 52f\n"
-      "st1 { v8.b }[2], [x15]\n"
-      "st1 { v12.b }[2], [x26]\n"
+      "tbz x16, #1, 51f\n"
+      "str h8, [x17], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "tbz x16, #0, 52f\n"
+      "st1 { v8.b }[2], [x17]\n"
+      "st1 { v12.b }[2], [x25]\n"
       "b 52f\n"
       "51:"  // Height 2: Partial direct writeback: partial_1_0
-      "str b8, [x15, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
+      "str b8, [x17, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
       "52:"  // Height 2: Partial direct writeback: Done
       "b 54f\n"
       "53:"  // Height 2: Full writeback
-      "str q8, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
-      "str q12, [x26, #0x0]\n"
+      "str q8, [x17, #0x0]\n"
+      "add x17, x17, #0x10\n"
+      "str q12, [x25, #0x0]\n"
       "54:"  // Height 2: Writeback done
-      "subs x17, x17, #0x10\n"
+      "subs x16, x16, #0x10\n"
       "bgt 29b\n"
       "b 164f\n"
       "55:"  // Height 3
       "mov x6, %x[col_bias]\n"
       "ldr x7, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "56:"  // Height 3: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -870,8 +870,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "mov x14, #0x0\n"
       "58:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w13, [x20, x14, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 59f\n"
       "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -895,123 +895,123 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "cmp x13, #0x20\n"
       "ldr q1, [x11, #0x0]\n"
       "ldr q2, [x10, #0x0]\n"
-      "ldr q6, [x16, #0x0]\n"
-      "ldr q7, [x16, #0x10]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "ldr q7, [x15, #0x10]\n"
       "blt 62f\n"
       "61:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x16, #0x28]\n"
+      "ldr x21, [x15, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x16, #0x38]\n"
+      "ldr x20, [x15, #0x38]\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr d21, [x16, #0x20]\n"
+      "ldr d21, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x12, x12, #0x10\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x11, x11, #0x10\n"
+      "ldr x21, [x15, #0x48]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr d20, [x16, #0x30]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x16, #0x48]\n"
-      "add x10, x10, #0x10\n"
-      "ldr x24, [x12, #0x8]\n"
+      "ldr d20, [x15, #0x30]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
-      "ldr x20, [x16, #0x58]\n"
+      "ldr x20, [x15, #0x58]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr d21, [x16, #0x40]\n"
+      "ldr d21, [x15, #0x40]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
-      "ldr x23, [x11, #0x8]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4f81e28f  // sdot v15.4s, v20.16b, v1.4b[0]\n"
-      "ldr x22, [x10, #0x8]\n"
+      "ldr x21, [x15, #0x68]\n"
       ".inst 0x4f82e293  // sdot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr d20, [x16, #0x50]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x16, #0x68]\n"
-      "sub x13, x13, #0x10\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "ldr d20, [x15, #0x50]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4fa0e2a8  // sdot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ac  // sdot v12.4s, v21.16b, v1.4b[1]\n"
-      "ldr x20, [x16, #0x78]\n"
+      "ldr x20, [x15, #0x78]\n"
       ".inst 0x4fa2e2b0  // sdot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr d21, [x16, #0x60]\n"
+      "ldr d21, [x15, #0x60]\n"
       ".inst 0x4fa0e289  // sdot v9.4s, v20.16b, v0.4b[1]\n"
-      "cmp x13, #0x20\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4fa1e28d  // sdot v13.4s, v20.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x15, #0x88]\n"
       ".inst 0x4fa2e291  // sdot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr d20, [x16, #0x70]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x16, #0x88]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr d20, [x15, #0x70]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4fa0e2aa  // sdot v10.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ae  // sdot v14.4s, v21.16b, v1.4b[1]\n"
-      "ldr x20, [x16, #0x98]\n"
+      "ldr x20, [x15, #0x98]\n"
       ".inst 0x4fa2e2b2  // sdot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr d21, [x16, #0x80]\n"
+      "ldr d21, [x15, #0x80]\n"
       ".inst 0x4fa0e28b  // sdot v11.4s, v20.16b, v0.4b[1]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4fa1e28f  // sdot v15.4s, v20.16b, v1.4b[1]\n"
+      "ldr x21, [x15, #0xa8]\n"
       ".inst 0x4fa2e293  // sdot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr d20, [x16, #0x90]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x16, #0xa8]\n"
+      "ldr d20, [x15, #0x90]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4f80eaa8  // sdot v8.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaac  // sdot v12.4s, v21.16b, v1.4b[2]\n"
-      "ldr x20, [x16, #0xb8]\n"
+      "ldr x20, [x15, #0xb8]\n"
       ".inst 0x4f82eab0  // sdot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr d21, [x16, #0xa0]\n"
+      "ldr d21, [x15, #0xa0]\n"
       ".inst 0x4f80ea89  // sdot v9.4s, v20.16b, v0.4b[2]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4f81ea8d  // sdot v13.4s, v20.16b, v1.4b[2]\n"
+      "ldr x21, [x15, #0xc8]\n"
       ".inst 0x4f82ea91  // sdot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr d20, [x16, #0xb0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x16, #0xc8]\n"
+      "ldr d20, [x15, #0xb0]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4f80eaaa  // sdot v10.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaae  // sdot v14.4s, v21.16b, v1.4b[2]\n"
-      "ldr x20, [x16, #0xd8]\n"
+      "ldr x20, [x15, #0xd8]\n"
       ".inst 0x4f82eab2  // sdot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr d21, [x16, #0xc0]\n"
+      "ldr d21, [x15, #0xc0]\n"
       ".inst 0x4f80ea8b  // sdot v11.4s, v20.16b, v0.4b[2]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4f81ea8f  // sdot v15.4s, v20.16b, v1.4b[2]\n"
+      "ldr x21, [x15, #0xe8]\n"
       ".inst 0x4f82ea93  // sdot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr d20, [x16, #0xd0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x16, #0xe8]\n"
+      "ldr d20, [x15, #0xd0]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4fa0eaa8  // sdot v8.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa1eaac  // sdot v12.4s, v21.16b, v1.4b[3]\n"
-      "ldr x20, [x16, #0xf8]\n"
+      "ldr x20, [x15, #0xf8]\n"
       ".inst 0x4fa2eab0  // sdot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr d21, [x16, #0xe0]\n"
+      "ldr d21, [x15, #0xe0]\n"
       ".inst 0x4fa0ea89  // sdot v9.4s, v20.16b, v0.4b[3]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4fa1ea8d  // sdot v13.4s, v20.16b, v1.4b[3]\n"
+      "add x12, x12, #0x10\n"
       ".inst 0x4fa2ea91  // sdot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr d20, [x16, #0xf0]\n"
-      "mov v21.d[1], x21\n"
-      "add x16, x16, #0x100\n"
-      "ldr x21, [x16, #0x8]\n"
+      "ldr d20, [x15, #0xf0]\n"
       "mov v20.d[1], x20\n"
+      "add x11, x11, #0x10\n"
+      "add x10, x10, #0x10\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa0eaaa  // sdot v10.4s, v21.16b, v0.4b[3]\n"
+      "ldr x20, [x15, #0x8]\n"
       ".inst 0x4fa1eaae  // sdot v14.4s, v21.16b, v1.4b[3]\n"
-      "ldr x20, [x16, #0x18]\n"
+      "ldr x23, [x12, #0x8]\n"
       ".inst 0x4fa2eab2  // sdot v18.4s, v21.16b, v2.4b[3]\n"
-      "ldr d6, [x16, #0x0]\n"
+      "ldr d6, [x15, #0x0]\n"
       ".inst 0x4fa0ea8b  // sdot v11.4s, v20.16b, v0.4b[3]\n"
       "ldr d0, [x12, #0x0]\n"
       ".inst 0x4fa1ea8f  // sdot v15.4s, v20.16b, v1.4b[3]\n"
       "ldr d1, [x11, #0x0]\n"
+      "ldr x22, [x11, #0x8]\n"
       ".inst 0x4fa2ea93  // sdot v19.4s, v20.16b, v2.4b[3]\n"
       "ldr d2, [x10, #0x0]\n"
-      "ldr d7, [x16, #0x10]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x24\n"
-      "mov v1.d[1], x23\n"
-      "mov v2.d[1], x22\n"
+      "sub x13, x13, #0x10\n"
+      "ldr d7, [x15, #0x10]\n"
+      "cmp x13, #0x20\n"
+      "ldr x21, [x10, #0x8]\n"
+      "mov v6.d[1], x20\n"
+      "ldr x20, [x15, #0x18]\n"
+      "mov v0.d[1], x23\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v1.d[1], x22\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
+      "mov v2.d[1], x21\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       "mov v7.d[1], x20\n"
       "bge 61b\n"
       "62:"  // Height 3: Multiply loop: Single iteration only
@@ -1020,66 +1020,66 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr q21, [x16, #0x20]\n"
+      "ldr q21, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x10, x10, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       "sub x13, x13, #0x10\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr q20, [x16, #0x30]\n"
+      "ldr q20, [x15, #0x30]\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x16, #0x40]\n"
+      "ldr q21, [x15, #0x40]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4f81e28f  // sdot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e293  // sdot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x16, #0x50]\n"
+      "ldr q20, [x15, #0x50]\n"
       ".inst 0x4fa0e2a8  // sdot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ac  // sdot v12.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b0  // sdot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x16, #0x60]\n"
+      "ldr q21, [x15, #0x60]\n"
       ".inst 0x4fa0e289  // sdot v9.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x4fa1e28d  // sdot v13.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x4fa2e291  // sdot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x16, #0x70]\n"
+      "ldr q20, [x15, #0x70]\n"
       ".inst 0x4fa0e2aa  // sdot v10.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ae  // sdot v14.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b2  // sdot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x16, #0x80]\n"
+      "ldr q21, [x15, #0x80]\n"
       ".inst 0x4fa0e28b  // sdot v11.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x4fa1e28f  // sdot v15.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x4fa2e293  // sdot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x16, #0x90]\n"
+      "ldr q20, [x15, #0x90]\n"
       ".inst 0x4f80eaa8  // sdot v8.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaac  // sdot v12.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x4f82eab0  // sdot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x16, #0xa0]\n"
+      "ldr q21, [x15, #0xa0]\n"
       ".inst 0x4f80ea89  // sdot v9.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x4f81ea8d  // sdot v13.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x4f82ea91  // sdot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x16, #0xb0]\n"
+      "ldr q20, [x15, #0xb0]\n"
       ".inst 0x4f80eaaa  // sdot v10.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaae  // sdot v14.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x4f82eab2  // sdot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x16, #0xc0]\n"
+      "ldr q21, [x15, #0xc0]\n"
       ".inst 0x4f80ea8b  // sdot v11.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x4f81ea8f  // sdot v15.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x4f82ea93  // sdot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x16, #0xd0]\n"
+      "ldr q20, [x15, #0xd0]\n"
       ".inst 0x4fa0eaa8  // sdot v8.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa1eaac  // sdot v12.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x4fa2eab0  // sdot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr q21, [x16, #0xe0]\n"
+      "ldr q21, [x15, #0xe0]\n"
       ".inst 0x4fa0ea89  // sdot v9.4s, v20.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea8d  // sdot v13.4s, v20.16b, v1.4b[3]\n"
       ".inst 0x4fa2ea91  // sdot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr q20, [x16, #0xf0]\n"
+      "ldr q20, [x15, #0xf0]\n"
       ".inst 0x4fa0eaaa  // sdot v10.4s, v21.16b, v0.4b[3]\n"
-      "add x16, x16, #0x100\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa1eaae  // sdot v14.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x4fa2eab2  // sdot v18.4s, v21.16b, v2.4b[3]\n"
       ".inst 0x4fa0ea8b  // sdot v11.4s, v20.16b, v0.4b[3]\n"
@@ -1095,18 +1095,18 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr s23, [x11], #0x4\n"
       "cmp x13, #0x4\n"
       "ldr s22, [x10], #0x4\n"
-      "ldr q21, [x16, #0x0]\n"
-      "ldr q20, [x16, #0x10]\n"
+      "ldr q21, [x15, #0x0]\n"
       ".inst 0x4f98e2a8  // sdot v8.4s, v21.16b, v24.4b[0]\n"
+      "ldr q20, [x15, #0x10]\n"
       ".inst 0x4f97e2ac  // sdot v12.4s, v21.16b, v23.4b[0]\n"
       ".inst 0x4f96e2b0  // sdot v16.4s, v21.16b, v22.4b[0]\n"
-      "ldr q21, [x16, #0x20]\n"
+      "ldr q21, [x15, #0x20]\n"
       ".inst 0x4f98e289  // sdot v9.4s, v20.16b, v24.4b[0]\n"
       ".inst 0x4f97e28d  // sdot v13.4s, v20.16b, v23.4b[0]\n"
       ".inst 0x4f96e291  // sdot v17.4s, v20.16b, v22.4b[0]\n"
-      "ldr q20, [x16, #0x30]\n"
+      "ldr q20, [x15, #0x30]\n"
       ".inst 0x4f98e2aa  // sdot v10.4s, v21.16b, v24.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f97e2ae  // sdot v14.4s, v21.16b, v23.4b[0]\n"
       ".inst 0x4f96e2b2  // sdot v18.4s, v21.16b, v22.4b[0]\n"
       ".inst 0x4f98e28b  // sdot v11.4s, v20.16b, v24.4b[0]\n"
@@ -1129,18 +1129,18 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr b1, [x11, #0x0]\n"
       "ldr b2, [x10, #0x0]\n"
       "67:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "ldr q21, [x16, #0x0]\n"
-      "ldr q20, [x16, #0x10]\n"
+      "ldr q21, [x15, #0x0]\n"
       ".inst 0x4f80e2a8  // sdot v8.4s, v21.16b, v0.4b[0]\n"
+      "ldr q20, [x15, #0x10]\n"
       ".inst 0x4f81e2ac  // sdot v12.4s, v21.16b, v1.4b[0]\n"
       ".inst 0x4f82e2b0  // sdot v16.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x16, #0x20]\n"
+      "ldr q21, [x15, #0x20]\n"
       ".inst 0x4f80e289  // sdot v9.4s, v20.16b, v0.4b[0]\n"
       ".inst 0x4f81e28d  // sdot v13.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e291  // sdot v17.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x16, #0x30]\n"
+      "ldr q20, [x15, #0x30]\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
@@ -1152,21 +1152,21 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "cmp x14, x20\n"
       "bne 58b\n"
       "ldr q23, [x6, #0x0]\n"
-      "ldr q22, [x6, #0x10]\n"
-      "ldr q21, [x6, #0x20]\n"
-      "ldr q20, [x6, #0x30]\n"
       "add v8.4s, v8.4s, v23.4s\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr q22, [x6, #0x10]\n"
       "add v9.4s, v9.4s, v22.4s\n"
-      "prfm pstl1keep, [x15, #0x0]\n"
+      "ldr q21, [x6, #0x20]\n"
       "add v10.4s, v10.4s, v21.4s\n"
+      "ldr q20, [x6, #0x30]\n"
       "add v11.4s, v11.4s, v20.4s\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x17, x20\n"
+      "add x24, x25, x20\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "add v12.4s, v12.4s, v23.4s\n"
-      "add x26, x15, x20\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "add v13.4s, v13.4s, v22.4s\n"
-      "add x25, x26, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "add v14.4s, v14.4s, v21.4s\n"
       "add v15.4s, v15.4s, v20.4s\n"
       "add v16.4s, v16.4s, v23.4s\n"
@@ -1187,9 +1187,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "add x7, x7, #0x40\n"
       "b 70f\n"
       "69:"  // Height 3: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -1261,36 +1261,35 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "srshl v18.4s, v18.4s, v2.4s\n"
       "srshl v19.4s, v19.4s, v3.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v22.4s }, [x20]\n"
+      "ld1r { v20.4s }, [x20]\n"
+      "add v8.4s, v8.4s, v20.4s\n"
+      "add v9.4s, v9.4s, v20.4s\n"
+      "add v10.4s, v10.4s, v20.4s\n"
+      "add v11.4s, v11.4s, v20.4s\n"
+      "add v12.4s, v12.4s, v20.4s\n"
+      "add v13.4s, v13.4s, v20.4s\n"
+      "add v14.4s, v14.4s, v20.4s\n"
+      "add v15.4s, v15.4s, v20.4s\n"
+      "add v16.4s, v16.4s, v20.4s\n"
+      "add v17.4s, v17.4s, v20.4s\n"
+      "add v18.4s, v18.4s, v20.4s\n"
+      "add v19.4s, v19.4s, v20.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v20.4s }, [x20]\n"
+      "smin v8.4s, v8.4s, v20.4s\n"
+      "smin v9.4s, v9.4s, v20.4s\n"
+      "smin v10.4s, v10.4s, v20.4s\n"
+      "smin v11.4s, v11.4s, v20.4s\n"
+      "smin v12.4s, v12.4s, v20.4s\n"
+      "smin v13.4s, v13.4s, v20.4s\n"
+      "smin v14.4s, v14.4s, v20.4s\n"
+      "smin v15.4s, v15.4s, v20.4s\n"
+      "smin v16.4s, v16.4s, v20.4s\n"
+      "smin v17.4s, v17.4s, v20.4s\n"
+      "smin v18.4s, v18.4s, v20.4s\n"
+      "smin v19.4s, v19.4s, v20.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v21.4s }, [x21]\n"
-      "cmp x17, #0x10\n"
       "ld1r { v20.4s }, [x20]\n"
-      "add v8.4s, v8.4s, v22.4s\n"
-      "add v9.4s, v9.4s, v22.4s\n"
-      "add v10.4s, v10.4s, v22.4s\n"
-      "add v11.4s, v11.4s, v22.4s\n"
-      "add v12.4s, v12.4s, v22.4s\n"
-      "add v13.4s, v13.4s, v22.4s\n"
-      "add v14.4s, v14.4s, v22.4s\n"
-      "add v15.4s, v15.4s, v22.4s\n"
-      "add v16.4s, v16.4s, v22.4s\n"
-      "add v17.4s, v17.4s, v22.4s\n"
-      "add v18.4s, v18.4s, v22.4s\n"
-      "add v19.4s, v19.4s, v22.4s\n"
-      "smin v8.4s, v8.4s, v21.4s\n"
-      "smin v9.4s, v9.4s, v21.4s\n"
-      "smin v10.4s, v10.4s, v21.4s\n"
-      "smin v11.4s, v11.4s, v21.4s\n"
-      "smin v12.4s, v12.4s, v21.4s\n"
-      "smin v13.4s, v13.4s, v21.4s\n"
-      "smin v14.4s, v14.4s, v21.4s\n"
-      "smin v15.4s, v15.4s, v21.4s\n"
-      "smin v16.4s, v16.4s, v21.4s\n"
-      "smin v17.4s, v17.4s, v21.4s\n"
-      "smin v18.4s, v18.4s, v21.4s\n"
-      "smin v19.4s, v19.4s, v21.4s\n"
       "smax v8.4s, v8.4s, v20.4s\n"
       "smax v9.4s, v9.4s, v20.4s\n"
       "smax v10.4s, v10.4s, v20.4s\n"
@@ -1309,101 +1308,102 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "uzp1 v20.8h, v14.8h, v15.8h\n"
       "uzp1 v16.8h, v16.8h, v17.8h\n"
       "uzp1 v17.8h, v18.8h, v19.8h\n"
+      "cmp x16, #0x10\n"
       "uzp1 v8.16b, v8.16b, v21.16b\n"
       "uzp1 v12.16b, v12.16b, v20.16b\n"
       "uzp1 v16.16b, v16.16b, v17.16b\n"
       "bge 80f\n"
-      "tbz x17, #3, 75f\n"
-      "str d8, [x15], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "tbz x17, #2, 73f\n"
-      "st1 { v8.s }[2], [x15], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "tbz x17, #1, 72f\n"
-      "st1 { v8.h }[6], [x15], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "tbz x17, #0, 79f\n"
-      "st1 { v8.b }[14], [x15]\n"
-      "st1 { v12.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
+      "tbz x16, #3, 75f\n"
+      "str d8, [x17], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "tbz x16, #2, 73f\n"
+      "st1 { v8.s }[2], [x17], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "tbz x16, #1, 72f\n"
+      "st1 { v8.h }[6], [x17], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "tbz x16, #0, 79f\n"
+      "st1 { v8.b }[14], [x17]\n"
+      "st1 { v12.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
       "b 79f\n"
       "72:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x17, #0, 79f\n"
-      "st1 { v8.b }[12], [x15]\n"
-      "st1 { v12.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
+      "tbz x16, #0, 79f\n"
+      "st1 { v8.b }[12], [x17]\n"
+      "st1 { v12.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
       "b 79f\n"
       "73:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x17, #1, 74f\n"
-      "st1 { v8.h }[4], [x15], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "tbz x17, #0, 79f\n"
-      "st1 { v8.b }[10], [x15]\n"
-      "st1 { v12.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
+      "tbz x16, #1, 74f\n"
+      "st1 { v8.h }[4], [x17], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "tbz x16, #0, 79f\n"
+      "st1 { v8.b }[10], [x17]\n"
+      "st1 { v12.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
       "b 79f\n"
       "74:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x17, #0, 79f\n"
-      "st1 { v8.b }[8], [x15]\n"
-      "st1 { v12.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
+      "tbz x16, #0, 79f\n"
+      "st1 { v8.b }[8], [x17]\n"
+      "st1 { v12.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
       "b 79f\n"
       "75:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x17, #2, 77f\n"
-      "str s8, [x15], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "tbz x17, #1, 76f\n"
-      "st1 { v8.h }[2], [x15], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "tbz x17, #0, 79f\n"
-      "st1 { v8.b }[6], [x15]\n"
-      "st1 { v12.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
+      "tbz x16, #2, 77f\n"
+      "str s8, [x17], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "tbz x16, #1, 76f\n"
+      "st1 { v8.h }[2], [x17], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "tbz x16, #0, 79f\n"
+      "st1 { v8.b }[6], [x17]\n"
+      "st1 { v12.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
       "b 79f\n"
       "76:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x17, #0, 79f\n"
-      "st1 { v8.b }[4], [x15]\n"
-      "st1 { v12.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
+      "tbz x16, #0, 79f\n"
+      "st1 { v8.b }[4], [x17]\n"
+      "st1 { v12.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
       "b 79f\n"
       "77:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x17, #1, 78f\n"
-      "str h8, [x15], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "tbz x17, #0, 79f\n"
-      "st1 { v8.b }[2], [x15]\n"
-      "st1 { v12.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
+      "tbz x16, #1, 78f\n"
+      "str h8, [x17], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "tbz x16, #0, 79f\n"
+      "st1 { v8.b }[2], [x17]\n"
+      "st1 { v12.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
       "b 79f\n"
       "78:"  // Height 3: Partial direct writeback: partial_1_0
-      "str b8, [x15, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
+      "str b8, [x17, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
       "79:"  // Height 3: Partial direct writeback: Done
       "b 81f\n"
       "80:"  // Height 3: Full writeback
-      "str q8, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
-      "str q12, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
+      "str q8, [x17, #0x0]\n"
+      "add x17, x17, #0x10\n"
+      "str q12, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
       "81:"  // Height 3: Writeback done
-      "subs x17, x17, #0x10\n"
+      "subs x16, x16, #0x10\n"
       "bgt 56b\n"
       "b 164f\n"
       "82:"  // Height 4
       "mov x6, %x[col_bias]\n"
       "ldr x7, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "83:"  // Height 4: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -1425,8 +1425,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "mov x14, #0x0\n"
       "85:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w13, [x20, x14, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 86f\n"
       "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1454,129 +1454,130 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr q1, [x11, #0x0]\n"
       "ldr q2, [x10, #0x0]\n"
       "ldr q3, [x9, #0x0]\n"
-      "ldr q6, [x16, #0x0]\n"
-      "ldr q7, [x16, #0x10]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "ldr q7, [x15, #0x10]\n"
       "blt 89f\n"
       "88:"  // Height 4: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x20, [x16, #0x28]\n"
+      "ldr x21, [x15, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x21, [x16, #0x38]\n"
+      "ldr x20, [x15, #0x38]\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr d25, [x16, #0x20]\n"
+      "ldr d25, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x11, x11, #0x10\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "ldr x21, [x15, #0x48]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x11, x11, #0x10\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr d24, [x16, #0x30]\n"
+      "ldr d24, [x15, #0x30]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
-      "ldr x20, [x16, #0x48]\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
-      "add x9, x9, #0x10\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x15, #0x58]\n"
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
+      "add x10, x10, #0x10\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr d25, [x16, #0x40]\n"
+      "ldr d25, [x15, #0x40]\n"
       ".inst 0x4f80e30b  // sdot v11.4s, v24.16b, v0.4b[0]\n"
-      "ldr x21, [x16, #0x58]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4f81e30f  // sdot v15.4s, v24.16b, v1.4b[0]\n"
-      "ldr x25, [x12, #0x8]\n"
+      "ldr x21, [x15, #0x68]\n"
       ".inst 0x4f82e313  // sdot v19.4s, v24.16b, v2.4b[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x4f83e317  // sdot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr d24, [x16, #0x50]\n"
+      "ldr d24, [x15, #0x50]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4fa0e328  // sdot v8.4s, v25.16b, v0.4b[1]\n"
-      "ldr x20, [x16, #0x68]\n"
       ".inst 0x4fa1e32c  // sdot v12.4s, v25.16b, v1.4b[1]\n"
-      "ldr x24, [x11, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x15, #0x78]\n"
       ".inst 0x4fa2e330  // sdot v16.4s, v25.16b, v2.4b[1]\n"
+      "ldr x25, [x12, #0x8]\n"
       ".inst 0x4fa3e334  // sdot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr d25, [x16, #0x60]\n"
+      "ldr d25, [x15, #0x60]\n"
       ".inst 0x4fa0e309  // sdot v9.4s, v24.16b, v0.4b[1]\n"
-      "ldr x21, [x16, #0x78]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4fa1e30d  // sdot v13.4s, v24.16b, v1.4b[1]\n"
-      "ldr x23, [x10, #0x8]\n"
+      "ldr x21, [x15, #0x88]\n"
       ".inst 0x4fa2e311  // sdot v17.4s, v24.16b, v2.4b[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x24, [x11, #0x8]\n"
       ".inst 0x4fa3e315  // sdot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr d24, [x16, #0x70]\n"
+      "ldr d24, [x15, #0x70]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4fa0e32a  // sdot v10.4s, v25.16b, v0.4b[1]\n"
-      "ldr x20, [x16, #0x88]\n"
       ".inst 0x4fa1e32e  // sdot v14.4s, v25.16b, v1.4b[1]\n"
-      "ldr x22, [x9, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x15, #0x98]\n"
       ".inst 0x4fa2e332  // sdot v18.4s, v25.16b, v2.4b[1]\n"
+      "ldr x23, [x10, #0x8]\n"
       ".inst 0x4fa3e336  // sdot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr d25, [x16, #0x80]\n"
+      "ldr d25, [x15, #0x80]\n"
       ".inst 0x4fa0e30b  // sdot v11.4s, v24.16b, v0.4b[1]\n"
-      "ldr x21, [x16, #0x98]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4fa1e30f  // sdot v15.4s, v24.16b, v1.4b[1]\n"
-      "sub x13, x13, #0x10\n"
+      "ldr x21, [x15, #0xa8]\n"
       ".inst 0x4fa2e313  // sdot v19.4s, v24.16b, v2.4b[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x22, [x9, #0x8]\n"
       ".inst 0x4fa3e317  // sdot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr d24, [x16, #0x90]\n"
+      "ldr d24, [x15, #0x90]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4f80eb28  // sdot v8.4s, v25.16b, v0.4b[2]\n"
-      "ldr x20, [x16, #0xa8]\n"
       ".inst 0x4f81eb2c  // sdot v12.4s, v25.16b, v1.4b[2]\n"
-      "cmp x13, #0x20\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x15, #0xb8]\n"
       ".inst 0x4f82eb30  // sdot v16.4s, v25.16b, v2.4b[2]\n"
+      "sub x13, x13, #0x10\n"
       ".inst 0x4f83eb34  // sdot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr d25, [x16, #0xa0]\n"
+      "ldr d25, [x15, #0xa0]\n"
       ".inst 0x4f80eb09  // sdot v9.4s, v24.16b, v0.4b[2]\n"
-      "ldr x21, [x16, #0xb8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4f81eb0d  // sdot v13.4s, v24.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "ldr x21, [x15, #0xc8]\n"
       ".inst 0x4f82eb11  // sdot v17.4s, v24.16b, v2.4b[2]\n"
-      "mov v25.d[1], x20\n"
+      "cmp x13, #0x20\n"
       ".inst 0x4f83eb15  // sdot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr d24, [x16, #0xb0]\n"
+      "ldr d24, [x15, #0xb0]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4f80eb2a  // sdot v10.4s, v25.16b, v0.4b[2]\n"
-      "ldr x20, [x16, #0xc8]\n"
       ".inst 0x4f81eb2e  // sdot v14.4s, v25.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x15, #0xd8]\n"
       ".inst 0x4f82eb32  // sdot v18.4s, v25.16b, v2.4b[2]\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4f83eb36  // sdot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr d25, [x16, #0xc0]\n"
+      "ldr d25, [x15, #0xc0]\n"
       ".inst 0x4f80eb0b  // sdot v11.4s, v24.16b, v0.4b[2]\n"
-      "ldr x21, [x16, #0xd8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4f81eb0f  // sdot v15.4s, v24.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr x21, [x15, #0xe8]\n"
       ".inst 0x4f82eb13  // sdot v19.4s, v24.16b, v2.4b[2]\n"
-      "mov v25.d[1], x20\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4f83eb17  // sdot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr d24, [x16, #0xd0]\n"
+      "ldr d24, [x15, #0xd0]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4fa0eb28  // sdot v8.4s, v25.16b, v0.4b[3]\n"
-      "ldr x20, [x16, #0xe8]\n"
       ".inst 0x4fa1eb2c  // sdot v12.4s, v25.16b, v1.4b[3]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x15, #0xf8]\n"
       ".inst 0x4fa2eb30  // sdot v16.4s, v25.16b, v2.4b[3]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4fa3eb34  // sdot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr d25, [x16, #0xe0]\n"
+      "ldr d25, [x15, #0xe0]\n"
       ".inst 0x4fa0eb09  // sdot v9.4s, v24.16b, v0.4b[3]\n"
-      "ldr x21, [x16, #0xf8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4fa1eb0d  // sdot v13.4s, v24.16b, v1.4b[3]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4fa2eb11  // sdot v17.4s, v24.16b, v2.4b[3]\n"
-      "mov v25.d[1], x20\n"
       ".inst 0x4fa3eb15  // sdot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr d24, [x16, #0xf0]\n"
-      "add x16, x16, #0x100\n"
+      "ldr d24, [x15, #0xf0]\n"
+      "mov v24.d[1], x20\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa0eb2a  // sdot v10.4s, v25.16b, v0.4b[3]\n"
+      "ldr x21, [x15, #0x8]\n"
       ".inst 0x4fa1eb2e  // sdot v14.4s, v25.16b, v1.4b[3]\n"
-      "ldr x20, [x16, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x15, #0x18]\n"
       ".inst 0x4fa2eb32  // sdot v18.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb36  // sdot v22.4s, v25.16b, v3.4b[3]\n"
-      "ldr d6, [x16, #0x0]\n"
+      "ldr d6, [x15, #0x0]\n"
       ".inst 0x4fa0eb0b  // sdot v11.4s, v24.16b, v0.4b[3]\n"
       "ldr d0, [x12, #0x0]\n"
       ".inst 0x4fa1eb0f  // sdot v15.4s, v24.16b, v1.4b[3]\n"
@@ -1585,9 +1586,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr d2, [x10, #0x0]\n"
       ".inst 0x4fa3eb17  // sdot v23.4s, v24.16b, v3.4b[3]\n"
       "ldr d3, [x9, #0x0]\n"
-      "ldr d7, [x16, #0x10]\n"
-      "mov v6.d[1], x20\n"
-      "ldr x20, [x16, #0x18]\n"
+      "ldr d7, [x15, #0x10]\n"
+      "mov v6.d[1], x21\n"
       "mov v0.d[1], x25\n"
       "mov v1.d[1], x24\n"
       "mov v2.d[1], x23\n"
@@ -1602,7 +1602,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       "add x10, x10, #0x10\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr q25, [x16, #0x20]\n"
+      "ldr q25, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -1610,7 +1610,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr q24, [x16, #0x30]\n"
+      "ldr q24, [x15, #0x30]\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
@@ -1618,64 +1618,64 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x16, #0x40]\n"
+      "ldr q25, [x15, #0x40]\n"
       ".inst 0x4f80e30b  // sdot v11.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e30f  // sdot v15.4s, v24.16b, v1.4b[0]\n"
       ".inst 0x4f82e313  // sdot v19.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x4f83e317  // sdot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x16, #0x50]\n"
+      "ldr q24, [x15, #0x50]\n"
       ".inst 0x4fa0e328  // sdot v8.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e32c  // sdot v12.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x4fa2e330  // sdot v16.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x4fa3e334  // sdot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x16, #0x60]\n"
+      "ldr q25, [x15, #0x60]\n"
       ".inst 0x4fa0e309  // sdot v9.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x4fa1e30d  // sdot v13.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x4fa2e311  // sdot v17.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x4fa3e315  // sdot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x16, #0x70]\n"
+      "ldr q24, [x15, #0x70]\n"
       ".inst 0x4fa0e32a  // sdot v10.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e32e  // sdot v14.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x4fa2e332  // sdot v18.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x4fa3e336  // sdot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x16, #0x80]\n"
+      "ldr q25, [x15, #0x80]\n"
       ".inst 0x4fa0e30b  // sdot v11.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x4fa1e30f  // sdot v15.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x4fa2e313  // sdot v19.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x4fa3e317  // sdot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x16, #0x90]\n"
+      "ldr q24, [x15, #0x90]\n"
       ".inst 0x4f80eb28  // sdot v8.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f81eb2c  // sdot v12.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x4f82eb30  // sdot v16.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x4f83eb34  // sdot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x16, #0xa0]\n"
+      "ldr q25, [x15, #0xa0]\n"
       ".inst 0x4f80eb09  // sdot v9.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb0d  // sdot v13.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x4f82eb11  // sdot v17.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x4f83eb15  // sdot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x16, #0xb0]\n"
+      "ldr q24, [x15, #0xb0]\n"
       ".inst 0x4f80eb2a  // sdot v10.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f81eb2e  // sdot v14.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x4f82eb32  // sdot v18.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x4f83eb36  // sdot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x16, #0xc0]\n"
+      "ldr q25, [x15, #0xc0]\n"
       ".inst 0x4f80eb0b  // sdot v11.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb0f  // sdot v15.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x4f82eb13  // sdot v19.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x4f83eb17  // sdot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x16, #0xd0]\n"
+      "ldr q24, [x15, #0xd0]\n"
       ".inst 0x4fa0eb28  // sdot v8.4s, v25.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb2c  // sdot v12.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb30  // sdot v16.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb34  // sdot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr q25, [x16, #0xe0]\n"
+      "ldr q25, [x15, #0xe0]\n"
       ".inst 0x4fa0eb09  // sdot v9.4s, v24.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb0d  // sdot v13.4s, v24.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb11  // sdot v17.4s, v24.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb15  // sdot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr q24, [x16, #0xf0]\n"
+      "ldr q24, [x15, #0xf0]\n"
       ".inst 0x4fa0eb2a  // sdot v10.4s, v25.16b, v0.4b[3]\n"
-      "add x16, x16, #0x100\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa1eb2e  // sdot v14.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb32  // sdot v18.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb36  // sdot v22.4s, v25.16b, v3.4b[3]\n"
@@ -1694,20 +1694,20 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "cmp x13, #0x4\n"
       "ldr s27, [x10], #0x4\n"
       "ldr s26, [x9], #0x4\n"
-      "ldr q25, [x16, #0x0]\n"
-      "ldr q24, [x16, #0x10]\n"
+      "ldr q25, [x15, #0x0]\n"
       ".inst 0x4f9de328  // sdot v8.4s, v25.16b, v29.4b[0]\n"
+      "ldr q24, [x15, #0x10]\n"
       ".inst 0x4f9ce32c  // sdot v12.4s, v25.16b, v28.4b[0]\n"
       ".inst 0x4f9be330  // sdot v16.4s, v25.16b, v27.4b[0]\n"
       ".inst 0x4f9ae334  // sdot v20.4s, v25.16b, v26.4b[0]\n"
-      "ldr q25, [x16, #0x20]\n"
+      "ldr q25, [x15, #0x20]\n"
       ".inst 0x4f9de309  // sdot v9.4s, v24.16b, v29.4b[0]\n"
       ".inst 0x4f9ce30d  // sdot v13.4s, v24.16b, v28.4b[0]\n"
       ".inst 0x4f9be311  // sdot v17.4s, v24.16b, v27.4b[0]\n"
       ".inst 0x4f9ae315  // sdot v21.4s, v24.16b, v26.4b[0]\n"
-      "ldr q24, [x16, #0x30]\n"
+      "ldr q24, [x15, #0x30]\n"
       ".inst 0x4f9de32a  // sdot v10.4s, v25.16b, v29.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f9ce32e  // sdot v14.4s, v25.16b, v28.4b[0]\n"
       ".inst 0x4f9be332  // sdot v18.4s, v25.16b, v27.4b[0]\n"
       ".inst 0x4f9ae336  // sdot v22.4s, v25.16b, v26.4b[0]\n"
@@ -1735,20 +1735,20 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr b2, [x10, #0x0]\n"
       "ldr b3, [x9, #0x0]\n"
       "94:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "ldr q25, [x16, #0x0]\n"
-      "ldr q24, [x16, #0x10]\n"
+      "ldr q25, [x15, #0x0]\n"
       ".inst 0x4f80e328  // sdot v8.4s, v25.16b, v0.4b[0]\n"
+      "ldr q24, [x15, #0x10]\n"
       ".inst 0x4f81e32c  // sdot v12.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x4f82e330  // sdot v16.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x4f83e334  // sdot v20.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x16, #0x20]\n"
+      "ldr q25, [x15, #0x20]\n"
       ".inst 0x4f80e309  // sdot v9.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e30d  // sdot v13.4s, v24.16b, v1.4b[0]\n"
       ".inst 0x4f82e311  // sdot v17.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x4f83e315  // sdot v21.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x16, #0x30]\n"
+      "ldr q24, [x15, #0x30]\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
@@ -1762,24 +1762,24 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "cmp x14, x20\n"
       "bne 85b\n"
       "ldr q27, [x6, #0x0]\n"
-      "ldr q26, [x6, #0x10]\n"
-      "ldr q25, [x6, #0x20]\n"
-      "ldr q24, [x6, #0x30]\n"
       "add v8.4s, v8.4s, v27.4s\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr q26, [x6, #0x10]\n"
       "add v9.4s, v9.4s, v26.4s\n"
-      "prfm pstl1keep, [x15, #0x0]\n"
+      "ldr q25, [x6, #0x20]\n"
       "add v10.4s, v10.4s, v25.4s\n"
+      "ldr q24, [x6, #0x30]\n"
       "add v11.4s, v11.4s, v24.4s\n"
-      "add v12.4s, v12.4s, v27.4s\n"
-      "add x26, x15, x20\n"
-      "add v13.4s, v13.4s, v26.4s\n"
-      "add x25, x26, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x17, x20\n"
       "add x24, x25, x20\n"
+      "add x23, x24, x20\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "add v12.4s, v12.4s, v27.4s\n"
       "prfm pstl1keep, [x25, #0x0]\n"
+      "add v13.4s, v13.4s, v26.4s\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "add v14.4s, v14.4s, v25.4s\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "add v15.4s, v15.4s, v24.4s\n"
       "add v16.4s, v16.4s, v27.4s\n"
       "add v17.4s, v17.4s, v26.4s\n"
@@ -1803,9 +1803,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "add x7, x7, #0x40\n"
       "b 97f\n"
       "96:"  // Height 4: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -1897,44 +1897,43 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "srshl v22.4s, v22.4s, v2.4s\n"
       "srshl v23.4s, v23.4s, v3.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x20]\n"
+      "ld1r { v24.4s }, [x20]\n"
+      "add v8.4s, v8.4s, v24.4s\n"
+      "add v9.4s, v9.4s, v24.4s\n"
+      "add v10.4s, v10.4s, v24.4s\n"
+      "add v11.4s, v11.4s, v24.4s\n"
+      "add v12.4s, v12.4s, v24.4s\n"
+      "add v13.4s, v13.4s, v24.4s\n"
+      "add v14.4s, v14.4s, v24.4s\n"
+      "add v15.4s, v15.4s, v24.4s\n"
+      "add v16.4s, v16.4s, v24.4s\n"
+      "add v17.4s, v17.4s, v24.4s\n"
+      "add v18.4s, v18.4s, v24.4s\n"
+      "add v19.4s, v19.4s, v24.4s\n"
+      "add v20.4s, v20.4s, v24.4s\n"
+      "add v21.4s, v21.4s, v24.4s\n"
+      "add v22.4s, v22.4s, v24.4s\n"
+      "add v23.4s, v23.4s, v24.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v24.4s }, [x20]\n"
+      "smin v8.4s, v8.4s, v24.4s\n"
+      "smin v9.4s, v9.4s, v24.4s\n"
+      "smin v10.4s, v10.4s, v24.4s\n"
+      "smin v11.4s, v11.4s, v24.4s\n"
+      "smin v12.4s, v12.4s, v24.4s\n"
+      "smin v13.4s, v13.4s, v24.4s\n"
+      "smin v14.4s, v14.4s, v24.4s\n"
+      "smin v15.4s, v15.4s, v24.4s\n"
+      "smin v16.4s, v16.4s, v24.4s\n"
+      "smin v17.4s, v17.4s, v24.4s\n"
+      "smin v18.4s, v18.4s, v24.4s\n"
+      "smin v19.4s, v19.4s, v24.4s\n"
+      "smin v20.4s, v20.4s, v24.4s\n"
+      "smin v21.4s, v21.4s, v24.4s\n"
+      "smin v22.4s, v22.4s, v24.4s\n"
+      "smin v23.4s, v23.4s, v24.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v25.4s }, [x21]\n"
-      "cmp x17, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
-      "add v8.4s, v8.4s, v26.4s\n"
-      "add v9.4s, v9.4s, v26.4s\n"
-      "add v10.4s, v10.4s, v26.4s\n"
-      "add v11.4s, v11.4s, v26.4s\n"
-      "add v12.4s, v12.4s, v26.4s\n"
-      "add v13.4s, v13.4s, v26.4s\n"
-      "add v14.4s, v14.4s, v26.4s\n"
-      "add v15.4s, v15.4s, v26.4s\n"
-      "add v16.4s, v16.4s, v26.4s\n"
-      "add v17.4s, v17.4s, v26.4s\n"
-      "add v18.4s, v18.4s, v26.4s\n"
-      "add v19.4s, v19.4s, v26.4s\n"
-      "add v20.4s, v20.4s, v26.4s\n"
-      "add v21.4s, v21.4s, v26.4s\n"
-      "add v22.4s, v22.4s, v26.4s\n"
-      "add v23.4s, v23.4s, v26.4s\n"
-      "smin v8.4s, v8.4s, v25.4s\n"
-      "smin v9.4s, v9.4s, v25.4s\n"
-      "smin v10.4s, v10.4s, v25.4s\n"
-      "smin v11.4s, v11.4s, v25.4s\n"
-      "smin v12.4s, v12.4s, v25.4s\n"
-      "smin v13.4s, v13.4s, v25.4s\n"
-      "smin v14.4s, v14.4s, v25.4s\n"
-      "smin v15.4s, v15.4s, v25.4s\n"
-      "smin v16.4s, v16.4s, v25.4s\n"
-      "smin v17.4s, v17.4s, v25.4s\n"
-      "smin v18.4s, v18.4s, v25.4s\n"
-      "smin v19.4s, v19.4s, v25.4s\n"
-      "smin v20.4s, v20.4s, v25.4s\n"
-      "smin v21.4s, v21.4s, v25.4s\n"
-      "smin v22.4s, v22.4s, v25.4s\n"
-      "smin v23.4s, v23.4s, v25.4s\n"
       "smax v8.4s, v8.4s, v24.4s\n"
       "smax v9.4s, v9.4s, v24.4s\n"
       "smax v10.4s, v10.4s, v24.4s\n"
@@ -1959,118 +1958,119 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "uzp1 v18.8h, v18.8h, v19.8h\n"
       "uzp1 v20.8h, v20.8h, v21.8h\n"
       "uzp1 v17.8h, v22.8h, v23.8h\n"
+      "cmp x16, #0x10\n"
       "uzp1 v8.16b, v8.16b, v25.16b\n"
       "uzp1 v12.16b, v12.16b, v24.16b\n"
       "uzp1 v16.16b, v16.16b, v18.16b\n"
       "uzp1 v20.16b, v20.16b, v17.16b\n"
       "bge 107f\n"
-      "tbz x17, #3, 102f\n"
-      "str d8, [x15], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "tbz x17, #2, 100f\n"
-      "st1 { v8.s }[2], [x15], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "tbz x17, #1, 99f\n"
-      "st1 { v8.h }[6], [x15], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "tbz x17, #0, 106f\n"
-      "st1 { v8.b }[14], [x15]\n"
-      "st1 { v12.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
-      "st1 { v20.b }[14], [x24]\n"
+      "tbz x16, #3, 102f\n"
+      "str d8, [x17], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "tbz x16, #2, 100f\n"
+      "st1 { v8.s }[2], [x17], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "tbz x16, #1, 99f\n"
+      "st1 { v8.h }[6], [x17], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "tbz x16, #0, 106f\n"
+      "st1 { v8.b }[14], [x17]\n"
+      "st1 { v12.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
       "b 106f\n"
       "99:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x17, #0, 106f\n"
-      "st1 { v8.b }[12], [x15]\n"
-      "st1 { v12.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
-      "st1 { v20.b }[12], [x24]\n"
+      "tbz x16, #0, 106f\n"
+      "st1 { v8.b }[12], [x17]\n"
+      "st1 { v12.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
       "b 106f\n"
       "100:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x17, #1, 101f\n"
-      "st1 { v8.h }[4], [x15], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "tbz x17, #0, 106f\n"
-      "st1 { v8.b }[10], [x15]\n"
-      "st1 { v12.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
-      "st1 { v20.b }[10], [x24]\n"
+      "tbz x16, #1, 101f\n"
+      "st1 { v8.h }[4], [x17], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "tbz x16, #0, 106f\n"
+      "st1 { v8.b }[10], [x17]\n"
+      "st1 { v12.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
       "b 106f\n"
       "101:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x17, #0, 106f\n"
-      "st1 { v8.b }[8], [x15]\n"
-      "st1 { v12.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
-      "st1 { v20.b }[8], [x24]\n"
+      "tbz x16, #0, 106f\n"
+      "st1 { v8.b }[8], [x17]\n"
+      "st1 { v12.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
       "b 106f\n"
       "102:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x17, #2, 104f\n"
-      "str s8, [x15], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "tbz x17, #1, 103f\n"
-      "st1 { v8.h }[2], [x15], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "tbz x17, #0, 106f\n"
-      "st1 { v8.b }[6], [x15]\n"
-      "st1 { v12.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
-      "st1 { v20.b }[6], [x24]\n"
+      "tbz x16, #2, 104f\n"
+      "str s8, [x17], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "tbz x16, #1, 103f\n"
+      "st1 { v8.h }[2], [x17], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "tbz x16, #0, 106f\n"
+      "st1 { v8.b }[6], [x17]\n"
+      "st1 { v12.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
       "b 106f\n"
       "103:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x17, #0, 106f\n"
-      "st1 { v8.b }[4], [x15]\n"
-      "st1 { v12.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
-      "st1 { v20.b }[4], [x24]\n"
+      "tbz x16, #0, 106f\n"
+      "st1 { v8.b }[4], [x17]\n"
+      "st1 { v12.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
       "b 106f\n"
       "104:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x17, #1, 105f\n"
-      "str h8, [x15], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "tbz x17, #0, 106f\n"
-      "st1 { v8.b }[2], [x15]\n"
-      "st1 { v12.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
-      "st1 { v20.b }[2], [x24]\n"
+      "tbz x16, #1, 105f\n"
+      "str h8, [x17], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "tbz x16, #0, 106f\n"
+      "st1 { v8.b }[2], [x17]\n"
+      "st1 { v12.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
       "b 106f\n"
       "105:"  // Height 4: Partial direct writeback: partial_1_0
-      "str b8, [x15, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
+      "str b8, [x17, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
       "106:"  // Height 4: Partial direct writeback: Done
       "b 108f\n"
       "107:"  // Height 4: Full writeback
-      "str q8, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
-      "str q12, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q20, [x24, #0x0]\n"
+      "str q8, [x17, #0x0]\n"
+      "add x17, x17, #0x10\n"
+      "str q12, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
       "108:"  // Height 4: Writeback done
-      "subs x17, x17, #0x10\n"
+      "subs x16, x16, #0x10\n"
       "bgt 83b\n"
       "b 164f\n"
       "109:"  // Height 5
       "mov x6, %x[col_bias]\n"
       "ldr x7, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "110:"  // Height 5: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -2096,8 +2096,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "mov x14, #0x0\n"
       "112:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w13, [x20, x14, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 113f\n"
       "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2129,148 +2129,148 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr q2, [x10, #0x0]\n"
       "ldr q3, [x9, #0x0]\n"
       "ldr q4, [x28, #0x0]\n"
-      "ldr q6, [x16, #0x0]\n"
-      "ldr q7, [x16, #0x10]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "ldr q7, [x15, #0x10]\n"
       "blt 116f\n"
       "115:"  // Height 5: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x16, #0x28]\n"
+      "ldr x21, [x15, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x16, #0x38]\n"
+      "ldr x20, [x15, #0x38]\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr d29, [x16, #0x20]\n"
+      "ldr d29, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x9, x9, #0x10\n"
+      "ldr x21, [x15, #0x48]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v29.d[1], x21\n"
+      "add x10, x10, #0x10\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x16, #0x48]\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr d28, [x16, #0x30]\n"
+      "ldr d28, [x15, #0x30]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
-      "add x28, x28, #0x10\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
-      "ldr x26, [x12, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x15, #0x58]\n"
       ".inst 0x4f82e3b2  // sdot v18.4s, v29.16b, v2.4b[0]\n"
+      "add x28, x28, #0x10\n"
       ".inst 0x4f83e3b6  // sdot v22.4s, v29.16b, v3.4b[0]\n"
-      "ldr x20, [x16, #0x58]\n"
+      "ldr x26, [x12, #0x8]\n"
       ".inst 0x4f84e3ba  // sdot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr d29, [x16, #0x40]\n"
+      "ldr d29, [x15, #0x40]\n"
       ".inst 0x4f80e38b  // sdot v11.4s, v28.16b, v0.4b[0]\n"
-      "ldr x25, [x11, #0x8]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81e38f  // sdot v15.4s, v28.16b, v1.4b[0]\n"
-      "ldr x24, [x10, #0x8]\n"
+      "ldr x21, [x15, #0x68]\n"
       ".inst 0x4f82e393  // sdot v19.4s, v28.16b, v2.4b[0]\n"
-      "mov v29.d[1], x21\n"
+      "ldr x25, [x11, #0x8]\n"
       ".inst 0x4f83e397  // sdot v23.4s, v28.16b, v3.4b[0]\n"
-      "ldr x21, [x16, #0x68]\n"
+      "ldr x24, [x10, #0x8]\n"
       ".inst 0x4f84e39b  // sdot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr d28, [x16, #0x50]\n"
+      "ldr d28, [x15, #0x50]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4fa0e3a8  // sdot v8.4s, v29.16b, v0.4b[1]\n"
-      "ldr x23, [x9, #0x8]\n"
       ".inst 0x4fa1e3ac  // sdot v12.4s, v29.16b, v1.4b[1]\n"
-      "ldr x22, [x28, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x15, #0x78]\n"
       ".inst 0x4fa2e3b0  // sdot v16.4s, v29.16b, v2.4b[1]\n"
+      "ldr x23, [x9, #0x8]\n"
       ".inst 0x4fa3e3b4  // sdot v20.4s, v29.16b, v3.4b[1]\n"
-      "ldr x20, [x16, #0x78]\n"
+      "ldr x22, [x28, #0x8]\n"
       ".inst 0x4fa4e3b8  // sdot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr d29, [x16, #0x60]\n"
+      "ldr d29, [x15, #0x60]\n"
       ".inst 0x4fa0e389  // sdot v9.4s, v28.16b, v0.4b[1]\n"
-      "sub x13, x13, #0x10\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4fa1e38d  // sdot v13.4s, v28.16b, v1.4b[1]\n"
-      "cmp x13, #0x20\n"
+      "ldr x21, [x15, #0x88]\n"
       ".inst 0x4fa2e391  // sdot v17.4s, v28.16b, v2.4b[1]\n"
-      "mov v29.d[1], x21\n"
+      "sub x13, x13, #0x10\n"
       ".inst 0x4fa3e395  // sdot v21.4s, v28.16b, v3.4b[1]\n"
-      "ldr x21, [x16, #0x88]\n"
+      "cmp x13, #0x20\n"
       ".inst 0x4fa4e399  // sdot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr d28, [x16, #0x70]\n"
+      "ldr d28, [x15, #0x70]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4fa0e3aa  // sdot v10.4s, v29.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4fa1e3ae  // sdot v14.4s, v29.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x15, #0x98]\n"
       ".inst 0x4fa2e3b2  // sdot v18.4s, v29.16b, v2.4b[1]\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4fa3e3b6  // sdot v22.4s, v29.16b, v3.4b[1]\n"
-      "ldr x20, [x16, #0x98]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4fa4e3ba  // sdot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr d29, [x16, #0x80]\n"
+      "ldr d29, [x15, #0x80]\n"
       ".inst 0x4fa0e38b  // sdot v11.4s, v28.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4fa1e38f  // sdot v15.4s, v28.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
+      "ldr x21, [x15, #0xa8]\n"
       ".inst 0x4fa2e393  // sdot v19.4s, v28.16b, v2.4b[1]\n"
-      "mov v29.d[1], x21\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4fa3e397  // sdot v23.4s, v28.16b, v3.4b[1]\n"
-      "ldr x21, [x16, #0xa8]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4fa4e39b  // sdot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr d28, [x16, #0x90]\n"
+      "ldr d28, [x15, #0x90]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4f80eba8  // sdot v8.4s, v29.16b, v0.4b[2]\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
       ".inst 0x4f81ebac  // sdot v12.4s, v29.16b, v1.4b[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x15, #0xb8]\n"
       ".inst 0x4f82ebb0  // sdot v16.4s, v29.16b, v2.4b[2]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       ".inst 0x4f83ebb4  // sdot v20.4s, v29.16b, v3.4b[2]\n"
-      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x4f84ebb8  // sdot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr d29, [x16, #0xa0]\n"
+      "ldr d29, [x15, #0xa0]\n"
       ".inst 0x4f80eb89  // sdot v9.4s, v28.16b, v0.4b[2]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81eb8d  // sdot v13.4s, v28.16b, v1.4b[2]\n"
+      "ldr x21, [x15, #0xc8]\n"
       ".inst 0x4f82eb91  // sdot v17.4s, v28.16b, v2.4b[2]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x4f83eb95  // sdot v21.4s, v28.16b, v3.4b[2]\n"
-      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x4f84eb99  // sdot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr d28, [x16, #0xb0]\n"
+      "ldr d28, [x15, #0xb0]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4f80ebaa  // sdot v10.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebae  // sdot v14.4s, v29.16b, v1.4b[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x15, #0xd8]\n"
       ".inst 0x4f82ebb2  // sdot v18.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb6  // sdot v22.4s, v29.16b, v3.4b[2]\n"
-      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x4f84ebba  // sdot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr d29, [x16, #0xc0]\n"
+      "ldr d29, [x15, #0xc0]\n"
       ".inst 0x4f80eb8b  // sdot v11.4s, v28.16b, v0.4b[2]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81eb8f  // sdot v15.4s, v28.16b, v1.4b[2]\n"
+      "ldr x21, [x15, #0xe8]\n"
       ".inst 0x4f82eb93  // sdot v19.4s, v28.16b, v2.4b[2]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x4f83eb97  // sdot v23.4s, v28.16b, v3.4b[2]\n"
-      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x4f84eb9b  // sdot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr d28, [x16, #0xd0]\n"
+      "ldr d28, [x15, #0xd0]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4fa0eba8  // sdot v8.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x4fa1ebac  // sdot v12.4s, v29.16b, v1.4b[3]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x15, #0xf8]\n"
       ".inst 0x4fa2ebb0  // sdot v16.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb4  // sdot v20.4s, v29.16b, v3.4b[3]\n"
-      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x4fa4ebb8  // sdot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr d29, [x16, #0xe0]\n"
+      "ldr d29, [x15, #0xe0]\n"
       ".inst 0x4fa0eb89  // sdot v9.4s, v28.16b, v0.4b[3]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4fa1eb8d  // sdot v13.4s, v28.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb91  // sdot v17.4s, v28.16b, v2.4b[3]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x4fa3eb95  // sdot v21.4s, v28.16b, v3.4b[3]\n"
       ".inst 0x4fa4eb99  // sdot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr d28, [x16, #0xf0]\n"
-      "add x16, x16, #0x100\n"
+      "ldr d28, [x15, #0xf0]\n"
+      "mov v28.d[1], x20\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa0ebaa  // sdot v10.4s, v29.16b, v0.4b[3]\n"
+      "ldr x21, [x15, #0x8]\n"
       ".inst 0x4fa1ebae  // sdot v14.4s, v29.16b, v1.4b[3]\n"
-      "ldr x21, [x16, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x15, #0x18]\n"
       ".inst 0x4fa2ebb2  // sdot v18.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb6  // sdot v22.4s, v29.16b, v3.4b[3]\n"
-      "ldr x20, [x16, #0x18]\n"
       ".inst 0x4fa4ebba  // sdot v26.4s, v29.16b, v4.4b[3]\n"
-      "ldr d6, [x16, #0x0]\n"
+      "ldr d6, [x15, #0x0]\n"
       ".inst 0x4fa0eb8b  // sdot v11.4s, v28.16b, v0.4b[3]\n"
       "ldr d0, [x12, #0x0]\n"
       ".inst 0x4fa1eb8f  // sdot v15.4s, v28.16b, v1.4b[3]\n"
@@ -2281,7 +2281,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr d3, [x9, #0x0]\n"
       ".inst 0x4fa4eb9b  // sdot v27.4s, v28.16b, v4.4b[3]\n"
       "ldr d4, [x28, #0x0]\n"
-      "ldr d7, [x16, #0x10]\n"
+      "ldr d7, [x15, #0x10]\n"
       "mov v6.d[1], x21\n"
       "mov v0.d[1], x26\n"
       "mov v1.d[1], x25\n"
@@ -2300,7 +2300,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr q29, [x16, #0x20]\n"
+      "ldr q29, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x28, x28, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2310,7 +2310,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr q28, [x16, #0x30]\n"
+      "ldr q28, [x15, #0x30]\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
@@ -2319,75 +2319,75 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "prfm pldl1keep, [x28, #0x80]\n"
       ".inst 0x4f83e3b6  // sdot v22.4s, v29.16b, v3.4b[0]\n"
       ".inst 0x4f84e3ba  // sdot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x16, #0x40]\n"
+      "ldr q29, [x15, #0x40]\n"
       ".inst 0x4f80e38b  // sdot v11.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f81e38f  // sdot v15.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f82e393  // sdot v19.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f83e397  // sdot v23.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x4f84e39b  // sdot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x16, #0x50]\n"
+      "ldr q28, [x15, #0x50]\n"
       ".inst 0x4fa0e3a8  // sdot v8.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3ac  // sdot v12.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3b0  // sdot v16.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x4fa3e3b4  // sdot v20.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x4fa4e3b8  // sdot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x16, #0x60]\n"
+      "ldr q29, [x15, #0x60]\n"
       ".inst 0x4fa0e389  // sdot v9.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x4fa1e38d  // sdot v13.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x4fa2e391  // sdot v17.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x4fa3e395  // sdot v21.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x4fa4e399  // sdot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x16, #0x70]\n"
+      "ldr q28, [x15, #0x70]\n"
       ".inst 0x4fa0e3aa  // sdot v10.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3ae  // sdot v14.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3b2  // sdot v18.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x4fa3e3b6  // sdot v22.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x4fa4e3ba  // sdot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x16, #0x80]\n"
+      "ldr q29, [x15, #0x80]\n"
       ".inst 0x4fa0e38b  // sdot v11.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x4fa1e38f  // sdot v15.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x4fa2e393  // sdot v19.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x4fa3e397  // sdot v23.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x4fa4e39b  // sdot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x16, #0x90]\n"
+      "ldr q28, [x15, #0x90]\n"
       ".inst 0x4f80eba8  // sdot v8.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebac  // sdot v12.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x4f82ebb0  // sdot v16.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb4  // sdot v20.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x4f84ebb8  // sdot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x16, #0xa0]\n"
+      "ldr q29, [x15, #0xa0]\n"
       ".inst 0x4f80eb89  // sdot v9.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb8d  // sdot v13.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb91  // sdot v17.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x4f83eb95  // sdot v21.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x4f84eb99  // sdot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x16, #0xb0]\n"
+      "ldr q28, [x15, #0xb0]\n"
       ".inst 0x4f80ebaa  // sdot v10.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebae  // sdot v14.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x4f82ebb2  // sdot v18.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb6  // sdot v22.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x4f84ebba  // sdot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x16, #0xc0]\n"
+      "ldr q29, [x15, #0xc0]\n"
       ".inst 0x4f80eb8b  // sdot v11.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb8f  // sdot v15.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb93  // sdot v19.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x4f83eb97  // sdot v23.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x4f84eb9b  // sdot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x16, #0xd0]\n"
+      "ldr q28, [x15, #0xd0]\n"
       ".inst 0x4fa0eba8  // sdot v8.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x4fa1ebac  // sdot v12.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x4fa2ebb0  // sdot v16.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb4  // sdot v20.4s, v29.16b, v3.4b[3]\n"
       ".inst 0x4fa4ebb8  // sdot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr q29, [x16, #0xe0]\n"
+      "ldr q29, [x15, #0xe0]\n"
       ".inst 0x4fa0eb89  // sdot v9.4s, v28.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb8d  // sdot v13.4s, v28.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb91  // sdot v17.4s, v28.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb95  // sdot v21.4s, v28.16b, v3.4b[3]\n"
       ".inst 0x4fa4eb99  // sdot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr q28, [x16, #0xf0]\n"
+      "ldr q28, [x15, #0xf0]\n"
       ".inst 0x4fa0ebaa  // sdot v10.4s, v29.16b, v0.4b[3]\n"
-      "add x16, x16, #0x100\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa1ebae  // sdot v14.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x4fa2ebb2  // sdot v18.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb6  // sdot v22.4s, v29.16b, v3.4b[3]\n"
@@ -2409,22 +2409,22 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr s0, [x10], #0x4\n"
       "ldr s31, [x9], #0x4\n"
       "ldr s30, [x28], #0x4\n"
-      "ldr q29, [x16, #0x0]\n"
-      "ldr q28, [x16, #0x10]\n"
+      "ldr q29, [x15, #0x0]\n"
       ".inst 0x4f82e3a8  // sdot v8.4s, v29.16b, v2.4b[0]\n"
+      "ldr q28, [x15, #0x10]\n"
       ".inst 0x4f81e3ac  // sdot v12.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f80e3b0  // sdot v16.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f9fe3b4  // sdot v20.4s, v29.16b, v31.4b[0]\n"
       ".inst 0x4f9ee3b8  // sdot v24.4s, v29.16b, v30.4b[0]\n"
-      "ldr q29, [x16, #0x20]\n"
+      "ldr q29, [x15, #0x20]\n"
       ".inst 0x4f82e389  // sdot v9.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f81e38d  // sdot v13.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f80e391  // sdot v17.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f9fe395  // sdot v21.4s, v28.16b, v31.4b[0]\n"
       ".inst 0x4f9ee399  // sdot v25.4s, v28.16b, v30.4b[0]\n"
-      "ldr q28, [x16, #0x30]\n"
+      "ldr q28, [x15, #0x30]\n"
       ".inst 0x4f82e3aa  // sdot v10.4s, v29.16b, v2.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f80e3b2  // sdot v18.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f9fe3b6  // sdot v22.4s, v29.16b, v31.4b[0]\n"
@@ -2457,22 +2457,22 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr b3, [x9, #0x0]\n"
       "ldr b4, [x28, #0x0]\n"
       "121:"  // Height 5: Multiply loop: Ragged operand read: Done
-      "ldr q29, [x16, #0x0]\n"
-      "ldr q28, [x16, #0x10]\n"
+      "ldr q29, [x15, #0x0]\n"
       ".inst 0x4f80e3a8  // sdot v8.4s, v29.16b, v0.4b[0]\n"
+      "ldr q28, [x15, #0x10]\n"
       ".inst 0x4f81e3ac  // sdot v12.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f82e3b0  // sdot v16.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x4f83e3b4  // sdot v20.4s, v29.16b, v3.4b[0]\n"
       ".inst 0x4f84e3b8  // sdot v24.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x16, #0x20]\n"
+      "ldr q29, [x15, #0x20]\n"
       ".inst 0x4f80e389  // sdot v9.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f81e38d  // sdot v13.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f82e391  // sdot v17.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f83e395  // sdot v21.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x4f84e399  // sdot v25.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x16, #0x30]\n"
+      "ldr q28, [x15, #0x30]\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f82e3b2  // sdot v18.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x4f83e3b6  // sdot v22.4s, v29.16b, v3.4b[0]\n"
@@ -2488,26 +2488,26 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "cmp x14, x20\n"
       "bne 112b\n"
       "ldr q31, [x6, #0x0]\n"
-      "ldr q30, [x6, #0x10]\n"
-      "ldr q29, [x6, #0x20]\n"
-      "ldr q28, [x6, #0x30]\n"
       "add v8.4s, v8.4s, v31.4s\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr q30, [x6, #0x10]\n"
       "add v9.4s, v9.4s, v30.4s\n"
-      "prfm pstl1keep, [x15, #0x0]\n"
+      "ldr q29, [x6, #0x20]\n"
       "add v10.4s, v10.4s, v29.4s\n"
+      "ldr q28, [x6, #0x30]\n"
       "add v11.4s, v11.4s, v28.4s\n"
-      "add v12.4s, v12.4s, v31.4s\n"
-      "add x26, x15, x20\n"
-      "add v13.4s, v13.4s, v30.4s\n"
-      "add x25, x26, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x17, x20\n"
       "add x24, x25, x20\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "add x23, x24, x20\n"
+      "add x22, x23, x20\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "add v12.4s, v12.4s, v31.4s\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "add v13.4s, v13.4s, v30.4s\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "add v14.4s, v14.4s, v29.4s\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "add v15.4s, v15.4s, v28.4s\n"
       "add v16.4s, v16.4s, v31.4s\n"
       "add v17.4s, v17.4s, v30.4s\n"
@@ -2535,9 +2535,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "add x7, x7, #0x40\n"
       "b 124f\n"
       "123:"  // Height 5: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -2649,52 +2649,51 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "srshl v26.4s, v26.4s, v2.4s\n"
       "srshl v27.4s, v27.4s, v3.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v30.4s }, [x20]\n"
+      "ld1r { v28.4s }, [x20]\n"
+      "add v8.4s, v8.4s, v28.4s\n"
+      "add v9.4s, v9.4s, v28.4s\n"
+      "add v10.4s, v10.4s, v28.4s\n"
+      "add v11.4s, v11.4s, v28.4s\n"
+      "add v12.4s, v12.4s, v28.4s\n"
+      "add v13.4s, v13.4s, v28.4s\n"
+      "add v14.4s, v14.4s, v28.4s\n"
+      "add v15.4s, v15.4s, v28.4s\n"
+      "add v16.4s, v16.4s, v28.4s\n"
+      "add v17.4s, v17.4s, v28.4s\n"
+      "add v18.4s, v18.4s, v28.4s\n"
+      "add v19.4s, v19.4s, v28.4s\n"
+      "add v20.4s, v20.4s, v28.4s\n"
+      "add v21.4s, v21.4s, v28.4s\n"
+      "add v22.4s, v22.4s, v28.4s\n"
+      "add v23.4s, v23.4s, v28.4s\n"
+      "add v24.4s, v24.4s, v28.4s\n"
+      "add v25.4s, v25.4s, v28.4s\n"
+      "add v26.4s, v26.4s, v28.4s\n"
+      "add v27.4s, v27.4s, v28.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v28.4s }, [x20]\n"
+      "smin v8.4s, v8.4s, v28.4s\n"
+      "smin v9.4s, v9.4s, v28.4s\n"
+      "smin v10.4s, v10.4s, v28.4s\n"
+      "smin v11.4s, v11.4s, v28.4s\n"
+      "smin v12.4s, v12.4s, v28.4s\n"
+      "smin v13.4s, v13.4s, v28.4s\n"
+      "smin v14.4s, v14.4s, v28.4s\n"
+      "smin v15.4s, v15.4s, v28.4s\n"
+      "smin v16.4s, v16.4s, v28.4s\n"
+      "smin v17.4s, v17.4s, v28.4s\n"
+      "smin v18.4s, v18.4s, v28.4s\n"
+      "smin v19.4s, v19.4s, v28.4s\n"
+      "smin v20.4s, v20.4s, v28.4s\n"
+      "smin v21.4s, v21.4s, v28.4s\n"
+      "smin v22.4s, v22.4s, v28.4s\n"
+      "smin v23.4s, v23.4s, v28.4s\n"
+      "smin v24.4s, v24.4s, v28.4s\n"
+      "smin v25.4s, v25.4s, v28.4s\n"
+      "smin v26.4s, v26.4s, v28.4s\n"
+      "smin v27.4s, v27.4s, v28.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v29.4s }, [x21]\n"
-      "cmp x17, #0x10\n"
       "ld1r { v28.4s }, [x20]\n"
-      "add v8.4s, v8.4s, v30.4s\n"
-      "add v9.4s, v9.4s, v30.4s\n"
-      "add v10.4s, v10.4s, v30.4s\n"
-      "add v11.4s, v11.4s, v30.4s\n"
-      "add v12.4s, v12.4s, v30.4s\n"
-      "add v13.4s, v13.4s, v30.4s\n"
-      "add v14.4s, v14.4s, v30.4s\n"
-      "add v15.4s, v15.4s, v30.4s\n"
-      "add v16.4s, v16.4s, v30.4s\n"
-      "add v17.4s, v17.4s, v30.4s\n"
-      "add v18.4s, v18.4s, v30.4s\n"
-      "add v19.4s, v19.4s, v30.4s\n"
-      "add v20.4s, v20.4s, v30.4s\n"
-      "add v21.4s, v21.4s, v30.4s\n"
-      "add v22.4s, v22.4s, v30.4s\n"
-      "add v23.4s, v23.4s, v30.4s\n"
-      "add v24.4s, v24.4s, v30.4s\n"
-      "add v25.4s, v25.4s, v30.4s\n"
-      "add v26.4s, v26.4s, v30.4s\n"
-      "add v27.4s, v27.4s, v30.4s\n"
-      "smin v8.4s, v8.4s, v29.4s\n"
-      "smin v9.4s, v9.4s, v29.4s\n"
-      "smin v10.4s, v10.4s, v29.4s\n"
-      "smin v11.4s, v11.4s, v29.4s\n"
-      "smin v12.4s, v12.4s, v29.4s\n"
-      "smin v13.4s, v13.4s, v29.4s\n"
-      "smin v14.4s, v14.4s, v29.4s\n"
-      "smin v15.4s, v15.4s, v29.4s\n"
-      "smin v16.4s, v16.4s, v29.4s\n"
-      "smin v17.4s, v17.4s, v29.4s\n"
-      "smin v18.4s, v18.4s, v29.4s\n"
-      "smin v19.4s, v19.4s, v29.4s\n"
-      "smin v20.4s, v20.4s, v29.4s\n"
-      "smin v21.4s, v21.4s, v29.4s\n"
-      "smin v22.4s, v22.4s, v29.4s\n"
-      "smin v23.4s, v23.4s, v29.4s\n"
-      "smin v24.4s, v24.4s, v29.4s\n"
-      "smin v25.4s, v25.4s, v29.4s\n"
-      "smin v26.4s, v26.4s, v29.4s\n"
-      "smin v27.4s, v27.4s, v29.4s\n"
       "smax v8.4s, v8.4s, v28.4s\n"
       "smax v9.4s, v9.4s, v28.4s\n"
       "smax v10.4s, v10.4s, v28.4s\n"
@@ -2725,139 +2724,139 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "uzp1 v18.8h, v22.8h, v23.8h\n"
       "uzp1 v24.8h, v24.8h, v25.8h\n"
       "uzp1 v17.8h, v26.8h, v27.8h\n"
+      "cmp x16, #0x10\n"
       "uzp1 v8.16b, v8.16b, v29.16b\n"
       "uzp1 v12.16b, v12.16b, v28.16b\n"
       "uzp1 v16.16b, v16.16b, v19.16b\n"
       "uzp1 v20.16b, v20.16b, v18.16b\n"
       "uzp1 v24.16b, v24.16b, v17.16b\n"
       "bge 134f\n"
-      "tbz x17, #3, 129f\n"
-      "str d8, [x15], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "tbz x17, #2, 127f\n"
-      "st1 { v8.s }[2], [x15], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "tbz x17, #1, 126f\n"
-      "st1 { v8.h }[6], [x15], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "tbz x17, #0, 133f\n"
-      "st1 { v8.b }[14], [x15]\n"
-      "st1 { v12.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "tbz x16, #3, 129f\n"
+      "str d8, [x17], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "tbz x16, #2, 127f\n"
+      "st1 { v8.s }[2], [x17], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "tbz x16, #1, 126f\n"
+      "st1 { v8.h }[6], [x17], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "tbz x16, #0, 133f\n"
+      "st1 { v8.b }[14], [x17]\n"
+      "st1 { v12.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 133f\n"
       "126:"  // Height 5: Partial direct writeback: partial_1_12
-      "tbz x17, #0, 133f\n"
-      "st1 { v8.b }[12], [x15]\n"
-      "st1 { v12.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "tbz x16, #0, 133f\n"
+      "st1 { v8.b }[12], [x17]\n"
+      "st1 { v12.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 133f\n"
       "127:"  // Height 5: Partial direct writeback: partial_2_8
-      "tbz x17, #1, 128f\n"
-      "st1 { v8.h }[4], [x15], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "tbz x17, #0, 133f\n"
-      "st1 { v8.b }[10], [x15]\n"
-      "st1 { v12.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "tbz x16, #1, 128f\n"
+      "st1 { v8.h }[4], [x17], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "tbz x16, #0, 133f\n"
+      "st1 { v8.b }[10], [x17]\n"
+      "st1 { v12.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 133f\n"
       "128:"  // Height 5: Partial direct writeback: partial_1_8
-      "tbz x17, #0, 133f\n"
-      "st1 { v8.b }[8], [x15]\n"
-      "st1 { v12.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "tbz x16, #0, 133f\n"
+      "st1 { v8.b }[8], [x17]\n"
+      "st1 { v12.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 133f\n"
       "129:"  // Height 5: Partial direct writeback: partial_4_0
-      "tbz x17, #2, 131f\n"
-      "str s8, [x15], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "tbz x17, #1, 130f\n"
-      "st1 { v8.h }[2], [x15], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "tbz x17, #0, 133f\n"
-      "st1 { v8.b }[6], [x15]\n"
-      "st1 { v12.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "tbz x16, #2, 131f\n"
+      "str s8, [x17], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "tbz x16, #1, 130f\n"
+      "st1 { v8.h }[2], [x17], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "tbz x16, #0, 133f\n"
+      "st1 { v8.b }[6], [x17]\n"
+      "st1 { v12.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 133f\n"
       "130:"  // Height 5: Partial direct writeback: partial_1_4
-      "tbz x17, #0, 133f\n"
-      "st1 { v8.b }[4], [x15]\n"
-      "st1 { v12.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "tbz x16, #0, 133f\n"
+      "st1 { v8.b }[4], [x17]\n"
+      "st1 { v12.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 133f\n"
       "131:"  // Height 5: Partial direct writeback: partial_2_0
-      "tbz x17, #1, 132f\n"
-      "str h8, [x15], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "tbz x17, #0, 133f\n"
-      "st1 { v8.b }[2], [x15]\n"
-      "st1 { v12.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "tbz x16, #1, 132f\n"
+      "str h8, [x17], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "tbz x16, #0, 133f\n"
+      "st1 { v8.b }[2], [x17]\n"
+      "st1 { v12.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 133f\n"
       "132:"  // Height 5: Partial direct writeback: partial_1_0
-      "str b8, [x15, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b8, [x17, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "133:"  // Height 5: Partial direct writeback: Done
       "b 135f\n"
       "134:"  // Height 5: Full writeback
-      "str q8, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
-      "str q12, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q8, [x17, #0x0]\n"
+      "add x17, x17, #0x10\n"
+      "str q12, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "135:"  // Height 5: Writeback done
-      "subs x17, x17, #0x10\n"
+      "subs x16, x16, #0x10\n"
       "bgt 110b\n"
       "b 164f\n"
       "136:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "mov x20, #0x6\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x6, %x[col_bias]\n"
       "ldr x7, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "madd x20, x21, x20, x15\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_N]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
+      "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "137:"  // Height 6: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -2887,8 +2886,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "mov x14, #0x0\n"
       "139:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w13, [x20, x14, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 140f\n"
       "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2924,14 +2923,14 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr q3, [x9, #0x0]\n"
       "ldr q4, [x28, #0x0]\n"
       "ldr q5, [x27, #0x0]\n"
-      "ldr q6, [x16, #0x0]\n"
-      "ldr q7, [x16, #0x10]\n"
+      "ldr q6, [x15, #0x0]\n"
+      "ldr q7, [x15, #0x10]\n"
       "blt 143f\n"
       "142:"  // Height 6: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x16, #0x28]\n"
+      "ldr x21, [x15, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x16, #0x38]\n"
+      "ldr x20, [x15, #0x38]\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
@@ -2939,150 +2938,149 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
       "add x10, x10, #0x10\n"
       ".inst 0x4f85e0dc  // sdot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr d6, [x16, #0x20]\n"
+      "ldr d6, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x9, x9, #0x10\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x28, x28, #0x10\n"
+      "ldr x21, [x15, #0x48]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v6.d[1], x21\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x16, #0x48]\n"
+      "add x28, x28, #0x10\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
       "add x27, x27, #0x10\n"
       ".inst 0x4f85e0fd  // sdot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr d7, [x16, #0x30]\n"
+      "ldr d7, [x15, #0x30]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4f80e0ca  // sdot v10.4s, v6.16b, v0.4b[0]\n"
-      "ldr x26, [x12, #0x8]\n"
       ".inst 0x4f81e0ce  // sdot v14.4s, v6.16b, v1.4b[0]\n"
-      "ldr x25, [x11, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0x58]\n"
       ".inst 0x4f82e0d2  // sdot v18.4s, v6.16b, v2.4b[0]\n"
+      "ldr x26, [x12, #0x8]\n"
       ".inst 0x4f83e0d6  // sdot v22.4s, v6.16b, v3.4b[0]\n"
-      "ldr x20, [x16, #0x58]\n"
+      "ldr x25, [x11, #0x8]\n"
       ".inst 0x4f84e0da  // sdot v26.4s, v6.16b, v4.4b[0]\n"
       "ldr x24, [x10, #0x8]\n"
       ".inst 0x4f85e0de  // sdot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr d6, [x16, #0x40]\n"
+      "ldr d6, [x15, #0x40]\n"
       ".inst 0x4f80e0eb  // sdot v11.4s, v7.16b, v0.4b[0]\n"
-      "ldr x23, [x9, #0x8]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4f81e0ef  // sdot v15.4s, v7.16b, v1.4b[0]\n"
-      "ldr x22, [x28, #0x8]\n"
+      "ldr x21, [x15, #0x68]\n"
       ".inst 0x4f82e0f3  // sdot v19.4s, v7.16b, v2.4b[0]\n"
-      "mov v6.d[1], x21\n"
+      "ldr x23, [x9, #0x8]\n"
       ".inst 0x4f83e0f7  // sdot v23.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x16, #0x68]\n"
-      ".inst 0x4f84e0fb  // sdot v27.4s, v7.16b, v4.4b[0]\n"
       "sub x13, x13, #0x10\n"
+      ".inst 0x4f84e0fb  // sdot v27.4s, v7.16b, v4.4b[0]\n"
+      "cmp x13, #0x20\n"
       ".inst 0x4f85e0ff  // sdot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr d7, [x16, #0x50]\n"
+      "ldr d7, [x15, #0x50]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4fa0e0c8  // sdot v8.4s, v6.16b, v0.4b[1]\n"
-      "cmp x13, #0x20\n"
       ".inst 0x4fa1e0cc  // sdot v12.4s, v6.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0x78]\n"
       ".inst 0x4fa2e0d0  // sdot v16.4s, v6.16b, v2.4b[1]\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4fa3e0d4  // sdot v20.4s, v6.16b, v3.4b[1]\n"
-      "ldr x20, [x16, #0x78]\n"
-      ".inst 0x4fa4e0d8  // sdot v24.4s, v6.16b, v4.4b[1]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
+      ".inst 0x4fa4e0d8  // sdot v24.4s, v6.16b, v4.4b[1]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4fa5e0dc  // sdot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr d6, [x16, #0x60]\n"
+      "ldr d6, [x15, #0x60]\n"
       ".inst 0x4fa0e0e9  // sdot v9.4s, v7.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4fa1e0ed  // sdot v13.4s, v7.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
+      "ldr x21, [x15, #0x88]\n"
       ".inst 0x4fa2e0f1  // sdot v17.4s, v7.16b, v2.4b[1]\n"
-      "mov v6.d[1], x21\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4fa3e0f5  // sdot v21.4s, v7.16b, v3.4b[1]\n"
-      "ldr x21, [x16, #0x88]\n"
-      ".inst 0x4fa4e0f9  // sdot v25.4s, v7.16b, v4.4b[1]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
+      ".inst 0x4fa4e0f9  // sdot v25.4s, v7.16b, v4.4b[1]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       ".inst 0x4fa5e0fd  // sdot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr d7, [x16, #0x70]\n"
+      "ldr d7, [x15, #0x70]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4fa0e0ca  // sdot v10.4s, v6.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x27, #0x80]\n"
       ".inst 0x4fa1e0ce  // sdot v14.4s, v6.16b, v1.4b[1]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0x98]\n"
       ".inst 0x4fa2e0d2  // sdot v18.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d6  // sdot v22.4s, v6.16b, v3.4b[1]\n"
-      "ldr x20, [x16, #0x98]\n"
       ".inst 0x4fa4e0da  // sdot v26.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0de  // sdot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr d6, [x16, #0x80]\n"
+      "ldr d6, [x15, #0x80]\n"
       ".inst 0x4fa0e0eb  // sdot v11.4s, v7.16b, v0.4b[1]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4fa1e0ef  // sdot v15.4s, v7.16b, v1.4b[1]\n"
+      "ldr x21, [x15, #0xa8]\n"
       ".inst 0x4fa2e0f3  // sdot v19.4s, v7.16b, v2.4b[1]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x4fa3e0f7  // sdot v23.4s, v7.16b, v3.4b[1]\n"
-      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x4fa4e0fb  // sdot v27.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0ff  // sdot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr d7, [x16, #0x90]\n"
+      "ldr d7, [x15, #0x90]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4f80e8c8  // sdot v8.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8cc  // sdot v12.4s, v6.16b, v1.4b[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0xb8]\n"
       ".inst 0x4f82e8d0  // sdot v16.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d4  // sdot v20.4s, v6.16b, v3.4b[2]\n"
-      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x4f84e8d8  // sdot v24.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8dc  // sdot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr d6, [x16, #0xa0]\n"
+      "ldr d6, [x15, #0xa0]\n"
       ".inst 0x4f80e8e9  // sdot v9.4s, v7.16b, v0.4b[2]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4f81e8ed  // sdot v13.4s, v7.16b, v1.4b[2]\n"
+      "ldr x21, [x15, #0xc8]\n"
       ".inst 0x4f82e8f1  // sdot v17.4s, v7.16b, v2.4b[2]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x4f83e8f5  // sdot v21.4s, v7.16b, v3.4b[2]\n"
-      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x4f84e8f9  // sdot v25.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8fd  // sdot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr d7, [x16, #0xb0]\n"
+      "ldr d7, [x15, #0xb0]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4f80e8ca  // sdot v10.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ce  // sdot v14.4s, v6.16b, v1.4b[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0xd8]\n"
       ".inst 0x4f82e8d2  // sdot v18.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d6  // sdot v22.4s, v6.16b, v3.4b[2]\n"
-      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x4f84e8da  // sdot v26.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8de  // sdot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr d6, [x16, #0xc0]\n"
+      "ldr d6, [x15, #0xc0]\n"
       ".inst 0x4f80e8eb  // sdot v11.4s, v7.16b, v0.4b[2]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4f81e8ef  // sdot v15.4s, v7.16b, v1.4b[2]\n"
+      "ldr x21, [x15, #0xe8]\n"
       ".inst 0x4f82e8f3  // sdot v19.4s, v7.16b, v2.4b[2]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x4f83e8f7  // sdot v23.4s, v7.16b, v3.4b[2]\n"
-      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x4f84e8fb  // sdot v27.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8ff  // sdot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr d7, [x16, #0xd0]\n"
+      "ldr d7, [x15, #0xd0]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4fa0e8c8  // sdot v8.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8cc  // sdot v12.4s, v6.16b, v1.4b[3]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x15, #0xf8]\n"
       ".inst 0x4fa2e8d0  // sdot v16.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d4  // sdot v20.4s, v6.16b, v3.4b[3]\n"
-      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x4fa4e8d8  // sdot v24.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8dc  // sdot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr d6, [x16, #0xe0]\n"
+      "ldr d6, [x15, #0xe0]\n"
       ".inst 0x4fa0e8e9  // sdot v9.4s, v7.16b, v0.4b[3]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4fa1e8ed  // sdot v13.4s, v7.16b, v1.4b[3]\n"
+      "ldr x22, [x28, #0x8]\n"
       ".inst 0x4fa2e8f1  // sdot v17.4s, v7.16b, v2.4b[3]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x4fa3e8f5  // sdot v21.4s, v7.16b, v3.4b[3]\n"
-      "ldr x21, [x27, #0x8]\n"
       ".inst 0x4fa4e8f9  // sdot v25.4s, v7.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8fd  // sdot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr d7, [x16, #0xf0]\n"
-      "add x16, x16, #0x100\n"
+      "ldr d7, [x15, #0xf0]\n"
+      "mov v7.d[1], x20\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa0e8ca  // sdot v10.4s, v6.16b, v0.4b[3]\n"
+      "ldr x20, [x15, #0x8]\n"
       ".inst 0x4fa1e8ce  // sdot v14.4s, v6.16b, v1.4b[3]\n"
-      "mov v7.d[1], x20\n"
       ".inst 0x4fa2e8d2  // sdot v18.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d6  // sdot v22.4s, v6.16b, v3.4b[3]\n"
-      "ldr x20, [x16, #0x8]\n"
       ".inst 0x4fa4e8da  // sdot v26.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8de  // sdot v30.4s, v6.16b, v5.4b[3]\n"
-      "ldr d6, [x16, #0x0]\n"
+      "ldr d6, [x15, #0x0]\n"
       ".inst 0x4fa0e8eb  // sdot v11.4s, v7.16b, v0.4b[3]\n"
       "ldr d0, [x12, #0x0]\n"
       ".inst 0x4fa1e8ef  // sdot v15.4s, v7.16b, v1.4b[3]\n"
@@ -3095,10 +3093,11 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr d4, [x28, #0x0]\n"
       ".inst 0x4fa5e8ff  // sdot v31.4s, v7.16b, v5.4b[3]\n"
       "ldr d5, [x27, #0x0]\n"
-      "ldr d7, [x16, #0x10]\n"
+      "ldr d7, [x15, #0x10]\n"
       "mov v6.d[1], x20\n"
-      "ldr x20, [x16, #0x18]\n"
+      "ldr x21, [x27, #0x8]\n"
       "mov v0.d[1], x26\n"
+      "ldr x20, [x15, #0x18]\n"
       "mov v1.d[1], x25\n"
       "mov v2.d[1], x24\n"
       "mov v3.d[1], x23\n"
@@ -3118,7 +3117,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
       "add x28, x28, #0x10\n"
       ".inst 0x4f85e0dc  // sdot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x16, #0x20]\n"
+      "ldr q6, [x15, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x27, x27, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -3130,7 +3129,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4f85e0fd  // sdot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x16, #0x30]\n"
+      "ldr q7, [x15, #0x30]\n"
       ".inst 0x4f80e0ca  // sdot v10.4s, v6.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4f81e0ce  // sdot v14.4s, v6.16b, v1.4b[0]\n"
@@ -3140,86 +3139,86 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       ".inst 0x4f83e0d6  // sdot v22.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x4f84e0da  // sdot v26.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0de  // sdot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x16, #0x40]\n"
+      "ldr q6, [x15, #0x40]\n"
       ".inst 0x4f80e0eb  // sdot v11.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ef  // sdot v15.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f3  // sdot v19.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f7  // sdot v23.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0fb  // sdot v27.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0ff  // sdot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x16, #0x50]\n"
+      "ldr q7, [x15, #0x50]\n"
       ".inst 0x4fa0e0c8  // sdot v8.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0cc  // sdot v12.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0d0  // sdot v16.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d4  // sdot v20.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0d8  // sdot v24.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0dc  // sdot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x16, #0x60]\n"
+      "ldr q6, [x15, #0x60]\n"
       ".inst 0x4fa0e0e9  // sdot v9.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ed  // sdot v13.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0f1  // sdot v17.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0f5  // sdot v21.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0f9  // sdot v25.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0fd  // sdot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x16, #0x70]\n"
+      "ldr q7, [x15, #0x70]\n"
       ".inst 0x4fa0e0ca  // sdot v10.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ce  // sdot v14.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0d2  // sdot v18.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d6  // sdot v22.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0da  // sdot v26.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0de  // sdot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x16, #0x80]\n"
+      "ldr q6, [x15, #0x80]\n"
       ".inst 0x4fa0e0eb  // sdot v11.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ef  // sdot v15.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0f3  // sdot v19.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0f7  // sdot v23.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0fb  // sdot v27.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0ff  // sdot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x16, #0x90]\n"
+      "ldr q7, [x15, #0x90]\n"
       ".inst 0x4f80e8c8  // sdot v8.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8cc  // sdot v12.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d0  // sdot v16.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d4  // sdot v20.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f84e8d8  // sdot v24.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8dc  // sdot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x16, #0xa0]\n"
+      "ldr q6, [x15, #0xa0]\n"
       ".inst 0x4f80e8e9  // sdot v9.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ed  // sdot v13.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x4f82e8f1  // sdot v17.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x4f83e8f5  // sdot v21.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x4f84e8f9  // sdot v25.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8fd  // sdot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x16, #0xb0]\n"
+      "ldr q7, [x15, #0xb0]\n"
       ".inst 0x4f80e8ca  // sdot v10.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ce  // sdot v14.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d2  // sdot v18.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d6  // sdot v22.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f84e8da  // sdot v26.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8de  // sdot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x16, #0xc0]\n"
+      "ldr q6, [x15, #0xc0]\n"
       ".inst 0x4f80e8eb  // sdot v11.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ef  // sdot v15.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x4f82e8f3  // sdot v19.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x4f83e8f7  // sdot v23.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x4f84e8fb  // sdot v27.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8ff  // sdot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x16, #0xd0]\n"
+      "ldr q7, [x15, #0xd0]\n"
       ".inst 0x4fa0e8c8  // sdot v8.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8cc  // sdot v12.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8d0  // sdot v16.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d4  // sdot v20.4s, v6.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8d8  // sdot v24.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8dc  // sdot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x16, #0xe0]\n"
+      "ldr q6, [x15, #0xe0]\n"
       ".inst 0x4fa0e8e9  // sdot v9.4s, v7.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8ed  // sdot v13.4s, v7.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8f1  // sdot v17.4s, v7.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8f5  // sdot v21.4s, v7.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8f9  // sdot v25.4s, v7.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8fd  // sdot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr q7, [x16, #0xf0]\n"
+      "ldr q7, [x15, #0xf0]\n"
       ".inst 0x4fa0e8ca  // sdot v10.4s, v6.16b, v0.4b[3]\n"
-      "add x16, x16, #0x100\n"
+      "add x15, x15, #0x100\n"
       ".inst 0x4fa1e8ce  // sdot v14.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8d2  // sdot v18.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d6  // sdot v22.4s, v6.16b, v3.4b[3]\n"
@@ -3244,24 +3243,24 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr s4, [x9], #0x4\n"
       "ldr s3, [x28], #0x4\n"
       "ldr s2, [x27], #0x4\n"
-      "ldr q1, [x16, #0x0]\n"
-      "ldr q0, [x16, #0x10]\n"
+      "ldr q1, [x15, #0x0]\n"
       ".inst 0x4f87e028  // sdot v8.4s, v1.16b, v7.4b[0]\n"
+      "ldr q0, [x15, #0x10]\n"
       ".inst 0x4f86e02c  // sdot v12.4s, v1.16b, v6.4b[0]\n"
       ".inst 0x4f85e030  // sdot v16.4s, v1.16b, v5.4b[0]\n"
       ".inst 0x4f84e034  // sdot v20.4s, v1.16b, v4.4b[0]\n"
       ".inst 0x4f83e038  // sdot v24.4s, v1.16b, v3.4b[0]\n"
       ".inst 0x4f82e03c  // sdot v28.4s, v1.16b, v2.4b[0]\n"
-      "ldr q1, [x16, #0x20]\n"
+      "ldr q1, [x15, #0x20]\n"
       ".inst 0x4f87e009  // sdot v9.4s, v0.16b, v7.4b[0]\n"
       ".inst 0x4f86e00d  // sdot v13.4s, v0.16b, v6.4b[0]\n"
       ".inst 0x4f85e011  // sdot v17.4s, v0.16b, v5.4b[0]\n"
       ".inst 0x4f84e015  // sdot v21.4s, v0.16b, v4.4b[0]\n"
       ".inst 0x4f83e019  // sdot v25.4s, v0.16b, v3.4b[0]\n"
       ".inst 0x4f82e01d  // sdot v29.4s, v0.16b, v2.4b[0]\n"
-      "ldr q0, [x16, #0x30]\n"
+      "ldr q0, [x15, #0x30]\n"
       ".inst 0x4f87e02a  // sdot v10.4s, v1.16b, v7.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f86e02e  // sdot v14.4s, v1.16b, v6.4b[0]\n"
       ".inst 0x4f85e032  // sdot v18.4s, v1.16b, v5.4b[0]\n"
       ".inst 0x4f84e036  // sdot v22.4s, v1.16b, v4.4b[0]\n"
@@ -3299,24 +3298,24 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "ldr b4, [x28, #0x0]\n"
       "ldr b5, [x27, #0x0]\n"
       "148:"  // Height 6: Multiply loop: Ragged operand read: Done
-      "ldr q7, [x16, #0x0]\n"
-      "ldr q6, [x16, #0x10]\n"
+      "ldr q7, [x15, #0x0]\n"
       ".inst 0x4f80e0e8  // sdot v8.4s, v7.16b, v0.4b[0]\n"
+      "ldr q6, [x15, #0x10]\n"
       ".inst 0x4f81e0ec  // sdot v12.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f0  // sdot v16.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f4  // sdot v20.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0f8  // sdot v24.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0fc  // sdot v28.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x16, #0x20]\n"
+      "ldr q7, [x15, #0x20]\n"
       ".inst 0x4f80e0c9  // sdot v9.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0cd  // sdot v13.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x4f82e0d1  // sdot v17.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0d5  // sdot v21.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x4f84e0d9  // sdot v25.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0dd  // sdot v29.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x16, #0x30]\n"
+      "ldr q6, [x15, #0x30]\n"
       ".inst 0x4f80e0ea  // sdot v10.4s, v7.16b, v0.4b[0]\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       ".inst 0x4f81e0ee  // sdot v14.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f2  // sdot v18.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f6  // sdot v22.4s, v7.16b, v3.4b[0]\n"
@@ -3334,30 +3333,30 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "cmp x14, x20\n"
       "bne 139b\n"
       "ldr q3, [x6, #0x0]\n"
-      "ldr q2, [x6, #0x10]\n"
-      "ldr q1, [x6, #0x20]\n"
-      "ldr q0, [x6, #0x30]\n"
       "add v8.4s, v8.4s, v3.4s\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr q2, [x6, #0x10]\n"
       "add v9.4s, v9.4s, v2.4s\n"
-      "prfm pstl1keep, [x15, #0x0]\n"
+      "ldr q1, [x6, #0x20]\n"
       "add v10.4s, v10.4s, v1.4s\n"
+      "ldr q0, [x6, #0x30]\n"
       "add v11.4s, v11.4s, v0.4s\n"
-      "add v12.4s, v12.4s, v3.4s\n"
-      "add x26, x15, x20\n"
-      "add v13.4s, v13.4s, v2.4s\n"
-      "add x25, x26, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x17, x20\n"
       "add x24, x25, x20\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "add x23, x24, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
+      "add x21, x22, x20\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "add v12.4s, v12.4s, v3.4s\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "add v13.4s, v13.4s, v2.4s\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "add v14.4s, v14.4s, v1.4s\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "add v15.4s, v15.4s, v0.4s\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "add v16.4s, v16.4s, v3.4s\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "add v17.4s, v17.4s, v2.4s\n"
       "add v18.4s, v18.4s, v1.4s\n"
       "add v19.4s, v19.4s, v0.4s\n"
@@ -3387,9 +3386,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "add x7, x7, #0x40\n"
       "b 151f\n"
       "150:"  // Height 6: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -3521,60 +3520,59 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "srshl v30.4s, v30.4s, v2.4s\n"
       "srshl v31.4s, v31.4s, v3.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v2.4s }, [x20]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "add v8.4s, v8.4s, v0.4s\n"
+      "add v9.4s, v9.4s, v0.4s\n"
+      "add v10.4s, v10.4s, v0.4s\n"
+      "add v11.4s, v11.4s, v0.4s\n"
+      "add v12.4s, v12.4s, v0.4s\n"
+      "add v13.4s, v13.4s, v0.4s\n"
+      "add v14.4s, v14.4s, v0.4s\n"
+      "add v15.4s, v15.4s, v0.4s\n"
+      "add v16.4s, v16.4s, v0.4s\n"
+      "add v17.4s, v17.4s, v0.4s\n"
+      "add v18.4s, v18.4s, v0.4s\n"
+      "add v19.4s, v19.4s, v0.4s\n"
+      "add v20.4s, v20.4s, v0.4s\n"
+      "add v21.4s, v21.4s, v0.4s\n"
+      "add v22.4s, v22.4s, v0.4s\n"
+      "add v23.4s, v23.4s, v0.4s\n"
+      "add v24.4s, v24.4s, v0.4s\n"
+      "add v25.4s, v25.4s, v0.4s\n"
+      "add v26.4s, v26.4s, v0.4s\n"
+      "add v27.4s, v27.4s, v0.4s\n"
+      "add v28.4s, v28.4s, v0.4s\n"
+      "add v29.4s, v29.4s, v0.4s\n"
+      "add v30.4s, v30.4s, v0.4s\n"
+      "add v31.4s, v31.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "smin v8.4s, v8.4s, v0.4s\n"
+      "smin v9.4s, v9.4s, v0.4s\n"
+      "smin v10.4s, v10.4s, v0.4s\n"
+      "smin v11.4s, v11.4s, v0.4s\n"
+      "smin v12.4s, v12.4s, v0.4s\n"
+      "smin v13.4s, v13.4s, v0.4s\n"
+      "smin v14.4s, v14.4s, v0.4s\n"
+      "smin v15.4s, v15.4s, v0.4s\n"
+      "smin v16.4s, v16.4s, v0.4s\n"
+      "smin v17.4s, v17.4s, v0.4s\n"
+      "smin v18.4s, v18.4s, v0.4s\n"
+      "smin v19.4s, v19.4s, v0.4s\n"
+      "smin v20.4s, v20.4s, v0.4s\n"
+      "smin v21.4s, v21.4s, v0.4s\n"
+      "smin v22.4s, v22.4s, v0.4s\n"
+      "smin v23.4s, v23.4s, v0.4s\n"
+      "smin v24.4s, v24.4s, v0.4s\n"
+      "smin v25.4s, v25.4s, v0.4s\n"
+      "smin v26.4s, v26.4s, v0.4s\n"
+      "smin v27.4s, v27.4s, v0.4s\n"
+      "smin v28.4s, v28.4s, v0.4s\n"
+      "smin v29.4s, v29.4s, v0.4s\n"
+      "smin v30.4s, v30.4s, v0.4s\n"
+      "smin v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v1.4s }, [x21]\n"
-      "cmp x17, #0x10\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v8.4s, v8.4s, v2.4s\n"
-      "add v9.4s, v9.4s, v2.4s\n"
-      "add v10.4s, v10.4s, v2.4s\n"
-      "add v11.4s, v11.4s, v2.4s\n"
-      "add v12.4s, v12.4s, v2.4s\n"
-      "add v13.4s, v13.4s, v2.4s\n"
-      "add v14.4s, v14.4s, v2.4s\n"
-      "add v15.4s, v15.4s, v2.4s\n"
-      "add v16.4s, v16.4s, v2.4s\n"
-      "add v17.4s, v17.4s, v2.4s\n"
-      "add v18.4s, v18.4s, v2.4s\n"
-      "add v19.4s, v19.4s, v2.4s\n"
-      "add v20.4s, v20.4s, v2.4s\n"
-      "add v21.4s, v21.4s, v2.4s\n"
-      "add v22.4s, v22.4s, v2.4s\n"
-      "add v23.4s, v23.4s, v2.4s\n"
-      "add v24.4s, v24.4s, v2.4s\n"
-      "add v25.4s, v25.4s, v2.4s\n"
-      "add v26.4s, v26.4s, v2.4s\n"
-      "add v27.4s, v27.4s, v2.4s\n"
-      "add v28.4s, v28.4s, v2.4s\n"
-      "add v29.4s, v29.4s, v2.4s\n"
-      "add v30.4s, v30.4s, v2.4s\n"
-      "add v31.4s, v31.4s, v2.4s\n"
-      "smin v8.4s, v8.4s, v1.4s\n"
-      "smin v9.4s, v9.4s, v1.4s\n"
-      "smin v10.4s, v10.4s, v1.4s\n"
-      "smin v11.4s, v11.4s, v1.4s\n"
-      "smin v12.4s, v12.4s, v1.4s\n"
-      "smin v13.4s, v13.4s, v1.4s\n"
-      "smin v14.4s, v14.4s, v1.4s\n"
-      "smin v15.4s, v15.4s, v1.4s\n"
-      "smin v16.4s, v16.4s, v1.4s\n"
-      "smin v17.4s, v17.4s, v1.4s\n"
-      "smin v18.4s, v18.4s, v1.4s\n"
-      "smin v19.4s, v19.4s, v1.4s\n"
-      "smin v20.4s, v20.4s, v1.4s\n"
-      "smin v21.4s, v21.4s, v1.4s\n"
-      "smin v22.4s, v22.4s, v1.4s\n"
-      "smin v23.4s, v23.4s, v1.4s\n"
-      "smin v24.4s, v24.4s, v1.4s\n"
-      "smin v25.4s, v25.4s, v1.4s\n"
-      "smin v26.4s, v26.4s, v1.4s\n"
-      "smin v27.4s, v27.4s, v1.4s\n"
-      "smin v28.4s, v28.4s, v1.4s\n"
-      "smin v29.4s, v29.4s, v1.4s\n"
-      "smin v30.4s, v30.4s, v1.4s\n"
-      "smin v31.4s, v31.4s, v1.4s\n"
       "smax v8.4s, v8.4s, v0.4s\n"
       "smax v9.4s, v9.4s, v0.4s\n"
       "smax v10.4s, v10.4s, v0.4s\n"
@@ -3611,6 +3609,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "uzp1 v18.8h, v26.8h, v27.8h\n"
       "uzp1 v28.8h, v28.8h, v29.8h\n"
       "uzp1 v17.8h, v30.8h, v31.8h\n"
+      "cmp x16, #0x10\n"
       "uzp1 v8.16b, v8.16b, v2.16b\n"
       "uzp1 v12.16b, v12.16b, v1.16b\n"
       "uzp1 v16.16b, v16.16b, v0.16b\n"
@@ -3618,136 +3617,136 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "uzp1 v24.16b, v24.16b, v18.16b\n"
       "uzp1 v28.16b, v28.16b, v17.16b\n"
       "bge 161f\n"
-      "tbz x17, #3, 156f\n"
-      "str d8, [x15], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
-      "tbz x17, #2, 154f\n"
-      "st1 { v8.s }[2], [x15], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
-      "tbz x17, #1, 153f\n"
-      "st1 { v8.h }[6], [x15], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "st1 { v28.h }[6], [x22], #0x2\n"
-      "tbz x17, #0, 160f\n"
-      "st1 { v8.b }[14], [x15]\n"
-      "st1 { v12.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "st1 { v28.b }[14], [x22]\n"
+      "tbz x16, #3, 156f\n"
+      "str d8, [x17], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
+      "tbz x16, #2, 154f\n"
+      "st1 { v8.s }[2], [x17], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
+      "tbz x16, #1, 153f\n"
+      "st1 { v8.h }[6], [x17], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "st1 { v28.h }[6], [x21], #0x2\n"
+      "tbz x16, #0, 160f\n"
+      "st1 { v8.b }[14], [x17]\n"
+      "st1 { v12.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
+      "st1 { v28.b }[14], [x21]\n"
       "b 160f\n"
       "153:"  // Height 6: Partial direct writeback: partial_1_12
-      "tbz x17, #0, 160f\n"
-      "st1 { v8.b }[12], [x15]\n"
-      "st1 { v12.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "st1 { v28.b }[12], [x22]\n"
+      "tbz x16, #0, 160f\n"
+      "st1 { v8.b }[12], [x17]\n"
+      "st1 { v12.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
+      "st1 { v28.b }[12], [x21]\n"
       "b 160f\n"
       "154:"  // Height 6: Partial direct writeback: partial_2_8
-      "tbz x17, #1, 155f\n"
-      "st1 { v8.h }[4], [x15], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "st1 { v28.h }[4], [x22], #0x2\n"
-      "tbz x17, #0, 160f\n"
-      "st1 { v8.b }[10], [x15]\n"
-      "st1 { v12.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "st1 { v28.b }[10], [x22]\n"
+      "tbz x16, #1, 155f\n"
+      "st1 { v8.h }[4], [x17], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "st1 { v28.h }[4], [x21], #0x2\n"
+      "tbz x16, #0, 160f\n"
+      "st1 { v8.b }[10], [x17]\n"
+      "st1 { v12.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
+      "st1 { v28.b }[10], [x21]\n"
       "b 160f\n"
       "155:"  // Height 6: Partial direct writeback: partial_1_8
-      "tbz x17, #0, 160f\n"
-      "st1 { v8.b }[8], [x15]\n"
-      "st1 { v12.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "st1 { v28.b }[8], [x22]\n"
+      "tbz x16, #0, 160f\n"
+      "st1 { v8.b }[8], [x17]\n"
+      "st1 { v12.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
+      "st1 { v28.b }[8], [x21]\n"
       "b 160f\n"
       "156:"  // Height 6: Partial direct writeback: partial_4_0
-      "tbz x17, #2, 158f\n"
-      "str s8, [x15], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
-      "tbz x17, #1, 157f\n"
-      "st1 { v8.h }[2], [x15], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "st1 { v28.h }[2], [x22], #0x2\n"
-      "tbz x17, #0, 160f\n"
-      "st1 { v8.b }[6], [x15]\n"
-      "st1 { v12.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "st1 { v28.b }[6], [x22]\n"
+      "tbz x16, #2, 158f\n"
+      "str s8, [x17], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
+      "tbz x16, #1, 157f\n"
+      "st1 { v8.h }[2], [x17], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "st1 { v28.h }[2], [x21], #0x2\n"
+      "tbz x16, #0, 160f\n"
+      "st1 { v8.b }[6], [x17]\n"
+      "st1 { v12.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
+      "st1 { v28.b }[6], [x21]\n"
       "b 160f\n"
       "157:"  // Height 6: Partial direct writeback: partial_1_4
-      "tbz x17, #0, 160f\n"
-      "st1 { v8.b }[4], [x15]\n"
-      "st1 { v12.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "st1 { v28.b }[4], [x22]\n"
+      "tbz x16, #0, 160f\n"
+      "st1 { v8.b }[4], [x17]\n"
+      "st1 { v12.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
+      "st1 { v28.b }[4], [x21]\n"
       "b 160f\n"
       "158:"  // Height 6: Partial direct writeback: partial_2_0
-      "tbz x17, #1, 159f\n"
-      "str h8, [x15], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "str h28, [x22], #0x2\n"
-      "tbz x17, #0, 160f\n"
-      "st1 { v8.b }[2], [x15]\n"
-      "st1 { v12.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "st1 { v28.b }[2], [x22]\n"
+      "tbz x16, #1, 159f\n"
+      "str h8, [x17], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "str h28, [x21], #0x2\n"
+      "tbz x16, #0, 160f\n"
+      "st1 { v8.b }[2], [x17]\n"
+      "st1 { v12.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
+      "st1 { v28.b }[2], [x21]\n"
       "b 160f\n"
       "159:"  // Height 6: Partial direct writeback: partial_1_0
-      "str b8, [x15, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "str b28, [x22, #0x0]\n"
+      "str b8, [x17, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
+      "str b28, [x21, #0x0]\n"
       "160:"  // Height 6: Partial direct writeback: Done
       "b 162f\n"
       "161:"  // Height 6: Full writeback
-      "str q8, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
-      "str q12, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q28, [x22, #0x0]\n"
+      "str q8, [x17, #0x0]\n"
+      "add x17, x17, #0x10\n"
+      "str q12, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q28, [x21, #0x0]\n"
       "162:"  // Height 6: Writeback done
-      "subs x17, x17, #0x10\n"
+      "subs x16, x16, #0x10\n"
       "bgt 137b\n"
       "subs %x[M], %x[M], #0x6\n"
       "beq 164f\n"
@@ -3761,8 +3760,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "164:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp
index dbff7baee7..6cdca85bd2 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp
@@ -47,18 +47,18 @@ void a64_hybrid_s8qs_dot_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -97,9 +97,9 @@ void a64_hybrid_s8qs_dot_6x16 (
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -109,8 +109,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "mov x28, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -125,102 +125,102 @@ void a64_hybrid_s8qs_dot_6x16 (
       "cmp x27, #0x10\n"
       "blt 9f\n"
       "ldr q0, [x26, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
+      "ldr q6, [x9, #0x0]\n"
       "cmp x27, #0x20\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x10]\n"
       "blt 8f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x10, #0x70]\n"
+      "ldr q16, [x9, #0x70]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x10, #0x80]\n"
+      "ldr q17, [x9, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x10, #0x90]\n"
+      "ldr q16, [x9, #0x90]\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x10, #0xa0]\n"
+      "ldr q17, [x9, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x10, #0xb0]\n"
+      "ldr q16, [x9, #0xb0]\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x10, #0xc0]\n"
+      "ldr q17, [x9, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x10, #0xd0]\n"
+      "ldr q16, [x9, #0xd0]\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr q17, [x10, #0xe0]\n"
+      "ldr q17, [x9, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q16, [x9, #0xf0]\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr q0, [x26, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "cmp x27, #0x20\n"
+      "add x9, x9, #0x100\n"
+      "ldr q6, [x9, #0x0]\n"
+      "ldr q7, [x9, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 7b\n"
       "8:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x10, #0x70]\n"
+      "ldr q16, [x9, #0x70]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x10, #0x80]\n"
+      "ldr q17, [x9, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x10, #0x90]\n"
+      "ldr q16, [x9, #0x90]\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x10, #0xa0]\n"
+      "ldr q17, [x9, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x10, #0xb0]\n"
+      "ldr q16, [x9, #0xb0]\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x10, #0xc0]\n"
+      "ldr q17, [x9, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x10, #0xd0]\n"
+      "ldr q16, [x9, #0xd0]\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr q17, [x10, #0xe0]\n"
+      "ldr q17, [x9, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q16, [x9, #0xf0]\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x9, x9, #0x100\n"
       "9:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 14f\n"
       "cmp x27, #0x4\n"
       "blt 11f\n"
       "10:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x26], #0x4\n"
-      "ldr q17, [x10, #0x0]\n"
+      "ldr q16, [x9, #0x0]\n"
+      ".inst 0x4f92e208  // sdot v8.4s, v16.16b, v18.4b[0]\n"
       "sub x27, x27, #0x4\n"
-      "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
-      ".inst 0x4f92e228  // sdot v8.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q16, [x9, #0x10]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4f92e209  // sdot v9.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "cmp x27, #0x4\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4f92e22a  // sdot v10.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x4f92e20b  // sdot v11.4s, v16.16b, v18.4b[0]\n"
+      "add x9, x9, #0x40\n"
       "bge 10b\n"
       "11:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 14f\n"
@@ -232,30 +232,30 @@ void a64_hybrid_s8qs_dot_6x16 (
       "12:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr b0, [x26, #0x0]\n"
       "13:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q16, [x10, #0x10]\n"
+      "ldr q17, [x9, #0x0]\n"
+      "ldr q16, [x9, #0x10]\n"
       ".inst 0x4f80e228  // sdot v8.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f80e209  // sdot v9.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q17, [x9, #0x20]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
+      "add x9, x9, #0x40\n"
       "14:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 4b\n"
-      "ldr q19, [x14, #0x0]\n"
-      "ldr q18, [x14, #0x10]\n"
+      "ldr q17, [x14, #0x0]\n"
+      "ldr q16, [x14, #0x10]\n"
+      "add v8.4s, v8.4s, v17.4s\n"
+      "add v9.4s, v9.4s, v16.4s\n"
       "ldr q17, [x14, #0x20]\n"
       "ldr q16, [x14, #0x30]\n"
-      "add x14, x14, #0x40\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add v8.4s, v8.4s, v19.4s\n"
-      "add v9.4s, v9.4s, v18.4s\n"
       "add v10.4s, v10.4s, v17.4s\n"
       "add v11.4s, v11.4s, v16.4s\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
+      "add x14, x14, #0x40\n"
       "tbz %x[flags], #4, 15f\n"
       "ldr q0, [x12, #0x0]\n"
       "ldr q4, [x13, #0x0]\n"
@@ -269,9 +269,9 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x13, x13, #0x40\n"
       "b 16f\n"
       "15:"  // Height 1: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -298,21 +298,21 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sqadd v10.4s, v10.4s, v17.4s\n"
       "sqadd v11.4s, v11.4s, v16.4s\n"
       "17:"  // Height 1: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v18.4s }, [x21]\n"
-      "ld1r { v17.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
-      "ld1r { v16.4s }, [x20]\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add v8.4s, v8.4s, v18.4s\n"
       "add v9.4s, v9.4s, v18.4s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1r { v16.4s }, [x20]\n"
       "add v10.4s, v10.4s, v18.4s\n"
       "add v11.4s, v11.4s, v18.4s\n"
+      "cmp x10, #0x10\n"
       "smin v8.4s, v8.4s, v17.4s\n"
       "smin v9.4s, v9.4s, v17.4s\n"
       "smin v10.4s, v10.4s, v17.4s\n"
@@ -325,65 +325,65 @@ void a64_hybrid_s8qs_dot_6x16 (
       "uzp1 v16.8h, v10.8h, v11.8h\n"
       "uzp1 v8.16b, v8.16b, v16.16b\n"
       "bge 26f\n"
-      "tbz x11, #3, 21f\n"
-      "str d8, [x9], #0x8\n"
-      "tbz x11, #2, 19f\n"
-      "st1 { v8.s }[2], [x9], #0x4\n"
-      "tbz x11, #1, 18f\n"
-      "st1 { v8.h }[6], [x9], #0x2\n"
-      "tbz x11, #0, 25f\n"
-      "st1 { v8.b }[14], [x9]\n"
+      "tbz x10, #3, 21f\n"
+      "str d8, [x11], #0x8\n"
+      "tbz x10, #2, 19f\n"
+      "st1 { v8.s }[2], [x11], #0x4\n"
+      "tbz x10, #1, 18f\n"
+      "st1 { v8.h }[6], [x11], #0x2\n"
+      "tbz x10, #0, 25f\n"
+      "st1 { v8.b }[14], [x11]\n"
       "b 25f\n"
       "18:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 25f\n"
-      "st1 { v8.b }[12], [x9]\n"
+      "tbz x10, #0, 25f\n"
+      "st1 { v8.b }[12], [x11]\n"
       "b 25f\n"
       "19:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 20f\n"
-      "st1 { v8.h }[4], [x9], #0x2\n"
-      "tbz x11, #0, 25f\n"
-      "st1 { v8.b }[10], [x9]\n"
+      "tbz x10, #1, 20f\n"
+      "st1 { v8.h }[4], [x11], #0x2\n"
+      "tbz x10, #0, 25f\n"
+      "st1 { v8.b }[10], [x11]\n"
       "b 25f\n"
       "20:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 25f\n"
-      "st1 { v8.b }[8], [x9]\n"
+      "tbz x10, #0, 25f\n"
+      "st1 { v8.b }[8], [x11]\n"
       "b 25f\n"
       "21:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 23f\n"
-      "str s8, [x9], #0x4\n"
-      "tbz x11, #1, 22f\n"
-      "st1 { v8.h }[2], [x9], #0x2\n"
-      "tbz x11, #0, 25f\n"
-      "st1 { v8.b }[6], [x9]\n"
+      "tbz x10, #2, 23f\n"
+      "str s8, [x11], #0x4\n"
+      "tbz x10, #1, 22f\n"
+      "st1 { v8.h }[2], [x11], #0x2\n"
+      "tbz x10, #0, 25f\n"
+      "st1 { v8.b }[6], [x11]\n"
       "b 25f\n"
       "22:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 25f\n"
-      "st1 { v8.b }[4], [x9]\n"
+      "tbz x10, #0, 25f\n"
+      "st1 { v8.b }[4], [x11]\n"
       "b 25f\n"
       "23:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 24f\n"
-      "str h8, [x9], #0x2\n"
-      "tbz x11, #0, 25f\n"
-      "st1 { v8.b }[2], [x9]\n"
+      "tbz x10, #1, 24f\n"
+      "str h8, [x11], #0x2\n"
+      "tbz x10, #0, 25f\n"
+      "st1 { v8.b }[2], [x11]\n"
       "b 25f\n"
       "24:"  // Height 1: Partial direct writeback: partial_1_0
-      "str b8, [x9, #0x0]\n"
+      "str b8, [x11, #0x0]\n"
       "25:"  // Height 1: Partial direct writeback: Done
       "b 27f\n"
       "26:"  // Height 1: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
+      "str q8, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
       "27:"  // Height 1: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 2b\n"
       "b 164f\n"
       "28:"  // Height 2
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "29:"  // Height 2: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -397,8 +397,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "mov x28, #0x0\n"
       "31:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 32f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -418,117 +418,117 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr q0, [x26, #0x0]\n"
       "ldr q1, [x25, #0x0]\n"
       "cmp x27, #0x20\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x0]\n"
+      "ldr q7, [x9, #0x10]\n"
       "blt 35f\n"
       "34:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       "sub x27, x27, #0x10\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q16, [x9, #0x30]\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x40]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "ldr q17, [x9, #0x40]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22c  // sdot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20d  // sdot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x10, #0x70]\n"
+      "ldr q16, [x9, #0x70]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22e  // sdot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x10, #0x80]\n"
+      "ldr q17, [x9, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20f  // sdot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x10, #0x90]\n"
+      "ldr q16, [x9, #0x90]\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x4f81ea2c  // sdot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x10, #0xa0]\n"
+      "ldr q17, [x9, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x4f81ea0d  // sdot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x10, #0xb0]\n"
+      "ldr q16, [x9, #0xb0]\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x4f81ea2e  // sdot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x10, #0xc0]\n"
+      "ldr q17, [x9, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x4f81ea0f  // sdot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x10, #0xd0]\n"
+      "ldr q16, [x9, #0xd0]\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea2c  // sdot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr q17, [x10, #0xe0]\n"
+      "ldr q17, [x9, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea0d  // sdot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q16, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea2e  // sdot v14.4s, v17.16b, v1.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
+      "ldr q6, [x9, #0x0]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr q0, [x26, #0x0]\n"
       ".inst 0x4fa1ea0f  // sdot v15.4s, v16.16b, v1.4b[3]\n"
       "ldr q1, [x25, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x10]\n"
       "bge 34b\n"
       "35:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       "add x26, x26, #0x10\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q16, [x9, #0x30]\n"
       "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22c  // sdot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20d  // sdot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x10, #0x70]\n"
+      "ldr q16, [x9, #0x70]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22e  // sdot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x10, #0x80]\n"
+      "ldr q17, [x9, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20f  // sdot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x10, #0x90]\n"
+      "ldr q16, [x9, #0x90]\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x4f81ea2c  // sdot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x10, #0xa0]\n"
+      "ldr q17, [x9, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x4f81ea0d  // sdot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x10, #0xb0]\n"
+      "ldr q16, [x9, #0xb0]\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x4f81ea2e  // sdot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x10, #0xc0]\n"
+      "ldr q17, [x9, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x4f81ea0f  // sdot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x10, #0xd0]\n"
+      "ldr q16, [x9, #0xd0]\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea2c  // sdot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr q17, [x10, #0xe0]\n"
+      "ldr q17, [x9, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea0d  // sdot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q16, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea2e  // sdot v14.4s, v17.16b, v1.4b[3]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
@@ -541,18 +541,18 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr s19, [x26], #0x4\n"
       "ldr s18, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q16, [x10, #0x10]\n"
       "cmp x27, #0x4\n"
+      "ldr q17, [x9, #0x0]\n"
+      "ldr q16, [x9, #0x10]\n"
       ".inst 0x4f93e228  // sdot v8.4s, v17.16b, v19.4b[0]\n"
       ".inst 0x4f92e22c  // sdot v12.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4f93e209  // sdot v9.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x4f92e20d  // sdot v13.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4f93e22a  // sdot v10.4s, v17.16b, v19.4b[0]\n"
       ".inst 0x4f92e22e  // sdot v14.4s, v17.16b, v18.4b[0]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f93e20b  // sdot v11.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x4f92e20f  // sdot v15.4s, v16.16b, v18.4b[0]\n"
       "bge 37b\n"
@@ -569,17 +569,17 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr b0, [x26, #0x0]\n"
       "ldr b1, [x25, #0x0]\n"
       "40:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q16, [x10, #0x10]\n"
+      "ldr q17, [x9, #0x0]\n"
+      "ldr q16, [x9, #0x10]\n"
       ".inst 0x4f80e228  // sdot v8.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f81e22c  // sdot v12.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4f80e209  // sdot v9.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20d  // sdot v13.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
       "41:"  // Height 2: Multiply loop: No odd multiplies
@@ -589,20 +589,20 @@ void a64_hybrid_s8qs_dot_6x16 (
       "bne 31b\n"
       "ldr q19, [x14, #0x0]\n"
       "ldr q18, [x14, #0x10]\n"
+      "add v8.4s, v8.4s, v19.4s\n"
+      "add v9.4s, v9.4s, v18.4s\n"
       "ldr q17, [x14, #0x20]\n"
       "ldr q16, [x14, #0x30]\n"
-      "add x14, x14, #0x40\n"
+      "add v10.4s, v10.4s, v17.4s\n"
+      "add v11.4s, v11.4s, v16.4s\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add v8.4s, v8.4s, v19.4s\n"
-      "add v9.4s, v9.4s, v18.4s\n"
+      "add x25, x11, x20\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "add v12.4s, v12.4s, v19.4s\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "add v13.4s, v13.4s, v18.4s\n"
-      "add v10.4s, v10.4s, v17.4s\n"
-      "add v11.4s, v11.4s, v16.4s\n"
-      "add x26, x9, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "add v14.4s, v14.4s, v17.4s\n"
+      "add x14, x14, #0x40\n"
       "add v15.4s, v15.4s, v16.4s\n"
       "tbz %x[flags], #4, 42f\n"
       "ldr q0, [x12, #0x0]\n"
@@ -617,9 +617,9 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x13, x13, #0x40\n"
       "b 43f\n"
       "42:"  // Height 2: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -646,11 +646,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v17.4s, v17.4s, #0x1f\n"
       "sshr v16.4s, v16.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v19.4s\n"
-      "and v19.16b, v12.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v18.4s\n"
-      "and v18.16b, v13.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v17.4s\n"
       "sqadd v11.4s, v11.4s, v16.4s\n"
+      "and v19.16b, v12.16b, v0.16b\n"
+      "and v18.16b, v13.16b, v1.16b\n"
       "and v17.16b, v14.16b, v2.16b\n"
       "and v16.16b, v15.16b, v3.16b\n"
       "sshr v19.4s, v19.4s, #0x1f\n"
@@ -662,21 +662,21 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sqadd v14.4s, v14.4s, v17.4s\n"
       "sqadd v15.4s, v15.4s, v16.4s\n"
       "44:"  // Height 2: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v18.4s }, [x21]\n"
-      "ld1r { v17.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "srshl v12.4s, v12.4s, v0.4s\n"
       "srshl v13.4s, v13.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v16.4s }, [x20]\n"
       "srshl v14.4s, v14.4s, v2.4s\n"
       "srshl v15.4s, v15.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "add v8.4s, v8.4s, v18.4s\n"
       "add v9.4s, v9.4s, v18.4s\n"
       "add v10.4s, v10.4s, v18.4s\n"
@@ -708,81 +708,81 @@ void a64_hybrid_s8qs_dot_6x16 (
       "uzp1 v8.16b, v8.16b, v17.16b\n"
       "uzp1 v12.16b, v12.16b, v16.16b\n"
       "bge 53f\n"
-      "tbz x11, #3, 48f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "tbz x11, #2, 46f\n"
-      "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "tbz x11, #1, 45f\n"
-      "st1 { v8.h }[6], [x9], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "tbz x11, #0, 52f\n"
-      "st1 { v8.b }[14], [x9]\n"
-      "st1 { v12.b }[14], [x26]\n"
+      "tbz x10, #3, 48f\n"
+      "str d8, [x11], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "tbz x10, #2, 46f\n"
+      "st1 { v8.s }[2], [x11], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "tbz x10, #1, 45f\n"
+      "st1 { v8.h }[6], [x11], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "tbz x10, #0, 52f\n"
+      "st1 { v8.b }[14], [x11]\n"
+      "st1 { v12.b }[14], [x25]\n"
       "b 52f\n"
       "45:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 52f\n"
-      "st1 { v8.b }[12], [x9]\n"
-      "st1 { v12.b }[12], [x26]\n"
+      "tbz x10, #0, 52f\n"
+      "st1 { v8.b }[12], [x11]\n"
+      "st1 { v12.b }[12], [x25]\n"
       "b 52f\n"
       "46:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 47f\n"
-      "st1 { v8.h }[4], [x9], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "tbz x11, #0, 52f\n"
-      "st1 { v8.b }[10], [x9]\n"
-      "st1 { v12.b }[10], [x26]\n"
+      "tbz x10, #1, 47f\n"
+      "st1 { v8.h }[4], [x11], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "tbz x10, #0, 52f\n"
+      "st1 { v8.b }[10], [x11]\n"
+      "st1 { v12.b }[10], [x25]\n"
       "b 52f\n"
       "47:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 52f\n"
-      "st1 { v8.b }[8], [x9]\n"
-      "st1 { v12.b }[8], [x26]\n"
+      "tbz x10, #0, 52f\n"
+      "st1 { v8.b }[8], [x11]\n"
+      "st1 { v12.b }[8], [x25]\n"
       "b 52f\n"
       "48:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 50f\n"
-      "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "tbz x11, #1, 49f\n"
-      "st1 { v8.h }[2], [x9], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "tbz x11, #0, 52f\n"
-      "st1 { v8.b }[6], [x9]\n"
-      "st1 { v12.b }[6], [x26]\n"
+      "tbz x10, #2, 50f\n"
+      "str s8, [x11], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "tbz x10, #1, 49f\n"
+      "st1 { v8.h }[2], [x11], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "tbz x10, #0, 52f\n"
+      "st1 { v8.b }[6], [x11]\n"
+      "st1 { v12.b }[6], [x25]\n"
       "b 52f\n"
       "49:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 52f\n"
-      "st1 { v8.b }[4], [x9]\n"
-      "st1 { v12.b }[4], [x26]\n"
+      "tbz x10, #0, 52f\n"
+      "st1 { v8.b }[4], [x11]\n"
+      "st1 { v12.b }[4], [x25]\n"
       "b 52f\n"
       "50:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 51f\n"
-      "str h8, [x9], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "tbz x11, #0, 52f\n"
-      "st1 { v8.b }[2], [x9]\n"
-      "st1 { v12.b }[2], [x26]\n"
+      "tbz x10, #1, 51f\n"
+      "str h8, [x11], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "tbz x10, #0, 52f\n"
+      "st1 { v8.b }[2], [x11]\n"
+      "st1 { v12.b }[2], [x25]\n"
       "b 52f\n"
       "51:"  // Height 2: Partial direct writeback: partial_1_0
-      "str b8, [x9, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
+      "str b8, [x11, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
       "52:"  // Height 2: Partial direct writeback: Done
       "b 54f\n"
       "53:"  // Height 2: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q12, [x26, #0x0]\n"
+      "str q8, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q12, [x25, #0x0]\n"
       "54:"  // Height 2: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 29b\n"
       "b 164f\n"
       "55:"  // Height 3
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "56:"  // Height 3: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -800,8 +800,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "mov x28, #0x0\n"
       "58:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 59f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -825,8 +825,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr q1, [x25, #0x0]\n"
       "cmp x27, #0x20\n"
       "ldr q2, [x24, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x0]\n"
+      "ldr q7, [x9, #0x10]\n"
       "blt 62f\n"
       "61:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
@@ -834,77 +834,77 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sub x27, x27, #0x10\n"
       "add x26, x26, #0x10\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x20]\n"
+      "ldr q21, [x9, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x25, x25, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x30]\n"
+      "ldr q20, [x9, #0x30]\n"
       "add x24, x24, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
+      "cmp x27, #0x20\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
+      "ldr q21, [x9, #0x40]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f81e28f  // sdot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e293  // sdot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x50]\n"
+      "ldr q20, [x9, #0x50]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4fa0e2a8  // sdot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ac  // sdot v12.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b0  // sdot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x10, #0x60]\n"
+      "ldr q21, [x9, #0x60]\n"
       ".inst 0x4fa0e289  // sdot v9.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x4fa1e28d  // sdot v13.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x4fa2e291  // sdot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x10, #0x70]\n"
+      "ldr q20, [x9, #0x70]\n"
       ".inst 0x4fa0e2aa  // sdot v10.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ae  // sdot v14.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b2  // sdot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x10, #0x80]\n"
+      "ldr q21, [x9, #0x80]\n"
       ".inst 0x4fa0e28b  // sdot v11.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x4fa1e28f  // sdot v15.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x4fa2e293  // sdot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x10, #0x90]\n"
+      "ldr q20, [x9, #0x90]\n"
       ".inst 0x4f80eaa8  // sdot v8.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaac  // sdot v12.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x4f82eab0  // sdot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x10, #0xa0]\n"
+      "ldr q21, [x9, #0xa0]\n"
       ".inst 0x4f80ea89  // sdot v9.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x4f81ea8d  // sdot v13.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x4f82ea91  // sdot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x10, #0xb0]\n"
+      "ldr q20, [x9, #0xb0]\n"
       ".inst 0x4f80eaaa  // sdot v10.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaae  // sdot v14.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x4f82eab2  // sdot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x10, #0xc0]\n"
+      "ldr q21, [x9, #0xc0]\n"
       ".inst 0x4f80ea8b  // sdot v11.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x4f81ea8f  // sdot v15.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x4f82ea93  // sdot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x10, #0xd0]\n"
+      "ldr q20, [x9, #0xd0]\n"
       ".inst 0x4fa0eaa8  // sdot v8.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa1eaac  // sdot v12.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x4fa2eab0  // sdot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr q21, [x10, #0xe0]\n"
+      "ldr q21, [x9, #0xe0]\n"
       ".inst 0x4fa0ea89  // sdot v9.4s, v20.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea8d  // sdot v13.4s, v20.16b, v1.4b[3]\n"
       ".inst 0x4fa2ea91  // sdot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr q20, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q20, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0eaaa  // sdot v10.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa1eaae  // sdot v14.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x4fa2eab2  // sdot v18.4s, v21.16b, v2.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
+      "ldr q6, [x9, #0x0]\n"
       ".inst 0x4fa0ea8b  // sdot v11.4s, v20.16b, v0.4b[3]\n"
       "ldr q0, [x26, #0x0]\n"
       ".inst 0x4fa1ea8f  // sdot v15.4s, v20.16b, v1.4b[3]\n"
       "ldr q1, [x25, #0x0]\n"
       ".inst 0x4fa2ea93  // sdot v19.4s, v20.16b, v2.4b[3]\n"
       "ldr q2, [x24, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x10]\n"
       "bge 61b\n"
       "62:"  // Height 3: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
@@ -912,65 +912,65 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x26, x26, #0x10\n"
       "add x25, x25, #0x10\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x20]\n"
+      "ldr q21, [x9, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x24, x24, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x30]\n"
+      "ldr q20, [x9, #0x30]\n"
       "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
+      "ldr q21, [x9, #0x40]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f81e28f  // sdot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e293  // sdot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x50]\n"
+      "ldr q20, [x9, #0x50]\n"
       ".inst 0x4fa0e2a8  // sdot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ac  // sdot v12.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b0  // sdot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x10, #0x60]\n"
+      "ldr q21, [x9, #0x60]\n"
       ".inst 0x4fa0e289  // sdot v9.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x4fa1e28d  // sdot v13.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x4fa2e291  // sdot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x10, #0x70]\n"
+      "ldr q20, [x9, #0x70]\n"
       ".inst 0x4fa0e2aa  // sdot v10.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ae  // sdot v14.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b2  // sdot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x10, #0x80]\n"
+      "ldr q21, [x9, #0x80]\n"
       ".inst 0x4fa0e28b  // sdot v11.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x4fa1e28f  // sdot v15.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x4fa2e293  // sdot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x10, #0x90]\n"
+      "ldr q20, [x9, #0x90]\n"
       ".inst 0x4f80eaa8  // sdot v8.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaac  // sdot v12.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x4f82eab0  // sdot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x10, #0xa0]\n"
+      "ldr q21, [x9, #0xa0]\n"
       ".inst 0x4f80ea89  // sdot v9.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x4f81ea8d  // sdot v13.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x4f82ea91  // sdot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x10, #0xb0]\n"
+      "ldr q20, [x9, #0xb0]\n"
       ".inst 0x4f80eaaa  // sdot v10.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaae  // sdot v14.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x4f82eab2  // sdot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x10, #0xc0]\n"
+      "ldr q21, [x9, #0xc0]\n"
       ".inst 0x4f80ea8b  // sdot v11.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x4f81ea8f  // sdot v15.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x4f82ea93  // sdot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x10, #0xd0]\n"
+      "ldr q20, [x9, #0xd0]\n"
       ".inst 0x4fa0eaa8  // sdot v8.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa1eaac  // sdot v12.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x4fa2eab0  // sdot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr q21, [x10, #0xe0]\n"
+      "ldr q21, [x9, #0xe0]\n"
       ".inst 0x4fa0ea89  // sdot v9.4s, v20.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea8d  // sdot v13.4s, v20.16b, v1.4b[3]\n"
       ".inst 0x4fa2ea91  // sdot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr q20, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q20, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0eaaa  // sdot v10.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa1eaae  // sdot v14.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x4fa2eab2  // sdot v18.4s, v21.16b, v2.4b[3]\n"
@@ -985,19 +985,19 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr s24, [x26], #0x4\n"
       "ldr s23, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr q21, [x10, #0x0]\n"
       "cmp x27, #0x4\n"
-      "ldr q20, [x10, #0x10]\n"
+      "ldr s22, [x24], #0x4\n"
+      "ldr q21, [x9, #0x0]\n"
       ".inst 0x4f98e2a8  // sdot v8.4s, v21.16b, v24.4b[0]\n"
       ".inst 0x4f97e2ac  // sdot v12.4s, v21.16b, v23.4b[0]\n"
+      "ldr q20, [x9, #0x10]\n"
       ".inst 0x4f96e2b0  // sdot v16.4s, v21.16b, v22.4b[0]\n"
-      "ldr q21, [x10, #0x20]\n"
+      "ldr q21, [x9, #0x20]\n"
       ".inst 0x4f98e289  // sdot v9.4s, v20.16b, v24.4b[0]\n"
       ".inst 0x4f97e28d  // sdot v13.4s, v20.16b, v23.4b[0]\n"
       ".inst 0x4f96e291  // sdot v17.4s, v20.16b, v22.4b[0]\n"
-      "ldr q20, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q20, [x9, #0x30]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f98e2aa  // sdot v10.4s, v21.16b, v24.4b[0]\n"
       ".inst 0x4f97e2ae  // sdot v14.4s, v21.16b, v23.4b[0]\n"
       ".inst 0x4f96e2b2  // sdot v18.4s, v21.16b, v22.4b[0]\n"
@@ -1021,17 +1021,17 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr b1, [x25, #0x0]\n"
       "ldr b2, [x24, #0x0]\n"
       "67:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "ldr q21, [x10, #0x0]\n"
-      "ldr q20, [x10, #0x10]\n"
+      "ldr q21, [x9, #0x0]\n"
+      "ldr q20, [x9, #0x10]\n"
       ".inst 0x4f80e2a8  // sdot v8.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f81e2ac  // sdot v12.4s, v21.16b, v1.4b[0]\n"
       ".inst 0x4f82e2b0  // sdot v16.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x20]\n"
+      "ldr q21, [x9, #0x20]\n"
       ".inst 0x4f80e289  // sdot v9.4s, v20.16b, v0.4b[0]\n"
       ".inst 0x4f81e28d  // sdot v13.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e291  // sdot v17.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q20, [x9, #0x30]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
@@ -1045,23 +1045,23 @@ void a64_hybrid_s8qs_dot_6x16 (
       "bne 58b\n"
       "ldr q23, [x14, #0x0]\n"
       "ldr q22, [x14, #0x10]\n"
-      "ldr q21, [x14, #0x20]\n"
-      "ldr q20, [x14, #0x30]\n"
-      "add x14, x14, #0x40\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add v8.4s, v8.4s, v23.4s\n"
       "add v9.4s, v9.4s, v22.4s\n"
-      "add v12.4s, v12.4s, v23.4s\n"
-      "add v13.4s, v13.4s, v22.4s\n"
+      "ldr q21, [x14, #0x20]\n"
+      "ldr q20, [x14, #0x30]\n"
       "add v10.4s, v10.4s, v21.4s\n"
       "add v11.4s, v11.4s, v20.4s\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x11, x20\n"
+      "add x24, x25, x20\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "add v12.4s, v12.4s, v23.4s\n"
+      "add v13.4s, v13.4s, v22.4s\n"
       "add v14.4s, v14.4s, v21.4s\n"
       "add v15.4s, v15.4s, v20.4s\n"
+      "add x14, x14, #0x40\n"
       "add v16.4s, v16.4s, v23.4s\n"
       "add v17.4s, v17.4s, v22.4s\n"
       "add v18.4s, v18.4s, v21.4s\n"
@@ -1079,9 +1079,9 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x13, x13, #0x40\n"
       "b 70f\n"
       "69:"  // Height 3: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -1112,11 +1112,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v21.4s, v21.4s, #0x1f\n"
       "sshr v20.4s, v20.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v23.4s\n"
-      "and v23.16b, v12.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v22.4s\n"
-      "and v22.16b, v13.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v21.4s\n"
       "sqadd v11.4s, v11.4s, v20.4s\n"
+      "and v23.16b, v12.16b, v0.16b\n"
+      "and v22.16b, v13.16b, v1.16b\n"
       "and v21.16b, v14.16b, v2.16b\n"
       "and v20.16b, v15.16b, v3.16b\n"
       "sshr v23.4s, v23.4s, #0x1f\n"
@@ -1124,11 +1124,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v21.4s, v21.4s, #0x1f\n"
       "sshr v20.4s, v20.4s, #0x1f\n"
       "sqadd v12.4s, v12.4s, v23.4s\n"
-      "and v23.16b, v16.16b, v0.16b\n"
       "sqadd v13.4s, v13.4s, v22.4s\n"
-      "and v22.16b, v17.16b, v1.16b\n"
       "sqadd v14.4s, v14.4s, v21.4s\n"
       "sqadd v15.4s, v15.4s, v20.4s\n"
+      "and v23.16b, v16.16b, v0.16b\n"
+      "and v22.16b, v17.16b, v1.16b\n"
       "and v21.16b, v18.16b, v2.16b\n"
       "and v20.16b, v19.16b, v3.16b\n"
       "sshr v23.4s, v23.4s, #0x1f\n"
@@ -1140,21 +1140,21 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sqadd v18.4s, v18.4s, v21.4s\n"
       "sqadd v19.4s, v19.4s, v20.4s\n"
       "71:"  // Height 3: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v22.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v22.4s }, [x21]\n"
-      "ld1r { v21.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "srshl v12.4s, v12.4s, v0.4s\n"
       "srshl v13.4s, v13.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v20.4s }, [x20]\n"
       "srshl v14.4s, v14.4s, v2.4s\n"
       "srshl v15.4s, v15.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v1.4s\n"
       "srshl v18.4s, v18.4s, v2.4s\n"
@@ -1205,97 +1205,97 @@ void a64_hybrid_s8qs_dot_6x16 (
       "uzp1 v12.16b, v12.16b, v20.16b\n"
       "uzp1 v16.16b, v16.16b, v17.16b\n"
       "bge 80f\n"
-      "tbz x11, #3, 75f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "tbz x11, #2, 73f\n"
-      "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "tbz x11, #1, 72f\n"
-      "st1 { v8.h }[6], [x9], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "tbz x11, #0, 79f\n"
-      "st1 { v8.b }[14], [x9]\n"
-      "st1 { v12.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
+      "tbz x10, #3, 75f\n"
+      "str d8, [x11], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "tbz x10, #2, 73f\n"
+      "st1 { v8.s }[2], [x11], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "tbz x10, #1, 72f\n"
+      "st1 { v8.h }[6], [x11], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "tbz x10, #0, 79f\n"
+      "st1 { v8.b }[14], [x11]\n"
+      "st1 { v12.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
       "b 79f\n"
       "72:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 79f\n"
-      "st1 { v8.b }[12], [x9]\n"
-      "st1 { v12.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
+      "tbz x10, #0, 79f\n"
+      "st1 { v8.b }[12], [x11]\n"
+      "st1 { v12.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
       "b 79f\n"
       "73:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 74f\n"
-      "st1 { v8.h }[4], [x9], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "tbz x11, #0, 79f\n"
-      "st1 { v8.b }[10], [x9]\n"
-      "st1 { v12.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
+      "tbz x10, #1, 74f\n"
+      "st1 { v8.h }[4], [x11], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "tbz x10, #0, 79f\n"
+      "st1 { v8.b }[10], [x11]\n"
+      "st1 { v12.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
       "b 79f\n"
       "74:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 79f\n"
-      "st1 { v8.b }[8], [x9]\n"
-      "st1 { v12.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
+      "tbz x10, #0, 79f\n"
+      "st1 { v8.b }[8], [x11]\n"
+      "st1 { v12.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
       "b 79f\n"
       "75:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 77f\n"
-      "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "tbz x11, #1, 76f\n"
-      "st1 { v8.h }[2], [x9], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "tbz x11, #0, 79f\n"
-      "st1 { v8.b }[6], [x9]\n"
-      "st1 { v12.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
+      "tbz x10, #2, 77f\n"
+      "str s8, [x11], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "tbz x10, #1, 76f\n"
+      "st1 { v8.h }[2], [x11], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "tbz x10, #0, 79f\n"
+      "st1 { v8.b }[6], [x11]\n"
+      "st1 { v12.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
       "b 79f\n"
       "76:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 79f\n"
-      "st1 { v8.b }[4], [x9]\n"
-      "st1 { v12.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
+      "tbz x10, #0, 79f\n"
+      "st1 { v8.b }[4], [x11]\n"
+      "st1 { v12.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
       "b 79f\n"
       "77:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 78f\n"
-      "str h8, [x9], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "tbz x11, #0, 79f\n"
-      "st1 { v8.b }[2], [x9]\n"
-      "st1 { v12.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
+      "tbz x10, #1, 78f\n"
+      "str h8, [x11], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "tbz x10, #0, 79f\n"
+      "st1 { v8.b }[2], [x11]\n"
+      "st1 { v12.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
       "b 79f\n"
       "78:"  // Height 3: Partial direct writeback: partial_1_0
-      "str b8, [x9, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
+      "str b8, [x11, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
       "79:"  // Height 3: Partial direct writeback: Done
       "b 81f\n"
       "80:"  // Height 3: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q12, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
+      "str q8, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q12, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
       "81:"  // Height 3: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 56b\n"
       "b 164f\n"
       "82:"  // Height 4
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "83:"  // Height 4: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -1317,8 +1317,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "mov x28, #0x0\n"
       "85:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 86f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1346,8 +1346,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "cmp x27, #0x20\n"
       "ldr q2, [x24, #0x0]\n"
       "ldr q3, [x23, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x0]\n"
+      "ldr q7, [x9, #0x10]\n"
       "blt 89f\n"
       "88:"  // Height 4: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
@@ -1356,7 +1356,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x26, x26, #0x10\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
       "add x25, x25, #0x10\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -1364,7 +1364,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x23, x23, #0x10\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x30]\n"
+      "ldr q24, [x9, #0x30]\n"
       "cmp x27, #0x20\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
@@ -1372,70 +1372,70 @@ void a64_hybrid_s8qs_dot_6x16 (
       "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f80e30b  // sdot v11.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e30f  // sdot v15.4s, v24.16b, v1.4b[0]\n"
       "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4f82e313  // sdot v19.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x4f83e317  // sdot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4fa0e328  // sdot v8.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e32c  // sdot v12.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x4fa2e330  // sdot v16.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x4fa3e334  // sdot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4fa0e309  // sdot v9.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x4fa1e30d  // sdot v13.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x4fa2e311  // sdot v17.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x4fa3e315  // sdot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x10, #0x70]\n"
+      "ldr q24, [x9, #0x70]\n"
       ".inst 0x4fa0e32a  // sdot v10.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e32e  // sdot v14.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x4fa2e332  // sdot v18.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x4fa3e336  // sdot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x10, #0x80]\n"
+      "ldr q25, [x9, #0x80]\n"
       ".inst 0x4fa0e30b  // sdot v11.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x4fa1e30f  // sdot v15.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x4fa2e313  // sdot v19.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x4fa3e317  // sdot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x10, #0x90]\n"
+      "ldr q24, [x9, #0x90]\n"
       ".inst 0x4f80eb28  // sdot v8.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f81eb2c  // sdot v12.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x4f82eb30  // sdot v16.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x4f83eb34  // sdot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x10, #0xa0]\n"
+      "ldr q25, [x9, #0xa0]\n"
       ".inst 0x4f80eb09  // sdot v9.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb0d  // sdot v13.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x4f82eb11  // sdot v17.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x4f83eb15  // sdot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x10, #0xb0]\n"
+      "ldr q24, [x9, #0xb0]\n"
       ".inst 0x4f80eb2a  // sdot v10.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f81eb2e  // sdot v14.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x4f82eb32  // sdot v18.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x4f83eb36  // sdot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x10, #0xc0]\n"
+      "ldr q25, [x9, #0xc0]\n"
       ".inst 0x4f80eb0b  // sdot v11.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb0f  // sdot v15.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x4f82eb13  // sdot v19.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x4f83eb17  // sdot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x10, #0xd0]\n"
+      "ldr q24, [x9, #0xd0]\n"
       ".inst 0x4fa0eb28  // sdot v8.4s, v25.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb2c  // sdot v12.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb30  // sdot v16.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb34  // sdot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr q25, [x10, #0xe0]\n"
+      "ldr q25, [x9, #0xe0]\n"
       ".inst 0x4fa0eb09  // sdot v9.4s, v24.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb0d  // sdot v13.4s, v24.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb11  // sdot v17.4s, v24.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb15  // sdot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr q24, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q24, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0eb2a  // sdot v10.4s, v25.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb2e  // sdot v14.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb32  // sdot v18.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb36  // sdot v22.4s, v25.16b, v3.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
+      "ldr q6, [x9, #0x0]\n"
       ".inst 0x4fa0eb0b  // sdot v11.4s, v24.16b, v0.4b[3]\n"
       "ldr q0, [x26, #0x0]\n"
       ".inst 0x4fa1eb0f  // sdot v15.4s, v24.16b, v1.4b[3]\n"
@@ -1444,7 +1444,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr q2, [x24, #0x0]\n"
       ".inst 0x4fa3eb17  // sdot v23.4s, v24.16b, v3.4b[3]\n"
       "ldr q3, [x23, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x10]\n"
       "bge 88b\n"
       "89:"  // Height 4: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
@@ -1453,80 +1453,80 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x25, x25, #0x10\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
       "add x24, x24, #0x10\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x30]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "ldr q24, [x9, #0x30]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4f80e30b  // sdot v11.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e30f  // sdot v15.4s, v24.16b, v1.4b[0]\n"
       ".inst 0x4f82e313  // sdot v19.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x4f83e317  // sdot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4fa0e328  // sdot v8.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e32c  // sdot v12.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x4fa2e330  // sdot v16.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x4fa3e334  // sdot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4fa0e309  // sdot v9.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x4fa1e30d  // sdot v13.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x4fa2e311  // sdot v17.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x4fa3e315  // sdot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x10, #0x70]\n"
+      "ldr q24, [x9, #0x70]\n"
       ".inst 0x4fa0e32a  // sdot v10.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e32e  // sdot v14.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x4fa2e332  // sdot v18.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x4fa3e336  // sdot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x10, #0x80]\n"
+      "ldr q25, [x9, #0x80]\n"
       ".inst 0x4fa0e30b  // sdot v11.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x4fa1e30f  // sdot v15.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x4fa2e313  // sdot v19.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x4fa3e317  // sdot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x10, #0x90]\n"
+      "ldr q24, [x9, #0x90]\n"
       ".inst 0x4f80eb28  // sdot v8.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f81eb2c  // sdot v12.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x4f82eb30  // sdot v16.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x4f83eb34  // sdot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x10, #0xa0]\n"
+      "ldr q25, [x9, #0xa0]\n"
       ".inst 0x4f80eb09  // sdot v9.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb0d  // sdot v13.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x4f82eb11  // sdot v17.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x4f83eb15  // sdot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x10, #0xb0]\n"
+      "ldr q24, [x9, #0xb0]\n"
       ".inst 0x4f80eb2a  // sdot v10.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f81eb2e  // sdot v14.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x4f82eb32  // sdot v18.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x4f83eb36  // sdot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x10, #0xc0]\n"
+      "ldr q25, [x9, #0xc0]\n"
       ".inst 0x4f80eb0b  // sdot v11.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb0f  // sdot v15.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x4f82eb13  // sdot v19.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x4f83eb17  // sdot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x10, #0xd0]\n"
+      "ldr q24, [x9, #0xd0]\n"
       ".inst 0x4fa0eb28  // sdot v8.4s, v25.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb2c  // sdot v12.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb30  // sdot v16.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb34  // sdot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr q25, [x10, #0xe0]\n"
+      "ldr q25, [x9, #0xe0]\n"
       ".inst 0x4fa0eb09  // sdot v9.4s, v24.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb0d  // sdot v13.4s, v24.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb11  // sdot v17.4s, v24.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb15  // sdot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr q24, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q24, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0eb2a  // sdot v10.4s, v25.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb2e  // sdot v14.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb32  // sdot v18.4s, v25.16b, v2.4b[3]\n"
@@ -1543,22 +1543,22 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr s29, [x26], #0x4\n"
       "ldr s28, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s27, [x24], #0x4\n"
       "ldr s26, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
-      "ldr q25, [x10, #0x0]\n"
-      "ldr q24, [x10, #0x10]\n"
+      "ldr q25, [x9, #0x0]\n"
+      "ldr q24, [x9, #0x10]\n"
       ".inst 0x4f9de328  // sdot v8.4s, v25.16b, v29.4b[0]\n"
       ".inst 0x4f9ce32c  // sdot v12.4s, v25.16b, v28.4b[0]\n"
       ".inst 0x4f9be330  // sdot v16.4s, v25.16b, v27.4b[0]\n"
       ".inst 0x4f9ae334  // sdot v20.4s, v25.16b, v26.4b[0]\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
       ".inst 0x4f9de309  // sdot v9.4s, v24.16b, v29.4b[0]\n"
       ".inst 0x4f9ce30d  // sdot v13.4s, v24.16b, v28.4b[0]\n"
       ".inst 0x4f9be311  // sdot v17.4s, v24.16b, v27.4b[0]\n"
       ".inst 0x4f9ae315  // sdot v21.4s, v24.16b, v26.4b[0]\n"
-      "ldr q24, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q24, [x9, #0x30]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f9de32a  // sdot v10.4s, v25.16b, v29.4b[0]\n"
       ".inst 0x4f9ce32e  // sdot v14.4s, v25.16b, v28.4b[0]\n"
       ".inst 0x4f9be332  // sdot v18.4s, v25.16b, v27.4b[0]\n"
@@ -1587,19 +1587,19 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr b2, [x24, #0x0]\n"
       "ldr b3, [x23, #0x0]\n"
       "94:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "ldr q25, [x10, #0x0]\n"
-      "ldr q24, [x10, #0x10]\n"
+      "ldr q25, [x9, #0x0]\n"
+      "ldr q24, [x9, #0x10]\n"
       ".inst 0x4f80e328  // sdot v8.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x4f81e32c  // sdot v12.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x4f82e330  // sdot v16.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x4f83e334  // sdot v20.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
       ".inst 0x4f80e309  // sdot v9.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e30d  // sdot v13.4s, v24.16b, v1.4b[0]\n"
       ".inst 0x4f82e311  // sdot v17.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x4f83e315  // sdot v21.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q24, [x9, #0x30]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
@@ -1615,24 +1615,24 @@ void a64_hybrid_s8qs_dot_6x16 (
       "bne 85b\n"
       "ldr q27, [x14, #0x0]\n"
       "ldr q26, [x14, #0x10]\n"
-      "ldr q25, [x14, #0x20]\n"
-      "ldr q24, [x14, #0x30]\n"
-      "add x14, x14, #0x40\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add v8.4s, v8.4s, v27.4s\n"
       "add v9.4s, v9.4s, v26.4s\n"
-      "add v12.4s, v12.4s, v27.4s\n"
-      "add v13.4s, v13.4s, v26.4s\n"
+      "ldr q25, [x14, #0x20]\n"
+      "ldr q24, [x14, #0x30]\n"
       "add v10.4s, v10.4s, v25.4s\n"
       "add v11.4s, v11.4s, v24.4s\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x11, x20\n"
       "add x24, x25, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
+      "add x23, x24, x20\n"
       "prfm pstl1keep, [x25, #0x0]\n"
-      "add v14.4s, v14.4s, v25.4s\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "add v12.4s, v12.4s, v27.4s\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "add v13.4s, v13.4s, v26.4s\n"
+      "add v14.4s, v14.4s, v25.4s\n"
+      "add x14, x14, #0x40\n"
       "add v15.4s, v15.4s, v24.4s\n"
       "add v16.4s, v16.4s, v27.4s\n"
       "add v17.4s, v17.4s, v26.4s\n"
@@ -1655,9 +1655,9 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x13, x13, #0x40\n"
       "b 97f\n"
       "96:"  // Height 4: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -1692,11 +1692,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v25.4s, v25.4s, #0x1f\n"
       "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v27.4s\n"
-      "and v27.16b, v12.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v26.4s\n"
-      "and v26.16b, v13.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v25.4s\n"
       "sqadd v11.4s, v11.4s, v24.4s\n"
+      "and v27.16b, v12.16b, v0.16b\n"
+      "and v26.16b, v13.16b, v1.16b\n"
       "and v25.16b, v14.16b, v2.16b\n"
       "and v24.16b, v15.16b, v3.16b\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
@@ -1704,11 +1704,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v25.4s, v25.4s, #0x1f\n"
       "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v12.4s, v12.4s, v27.4s\n"
-      "and v27.16b, v16.16b, v0.16b\n"
       "sqadd v13.4s, v13.4s, v26.4s\n"
-      "and v26.16b, v17.16b, v1.16b\n"
       "sqadd v14.4s, v14.4s, v25.4s\n"
       "sqadd v15.4s, v15.4s, v24.4s\n"
+      "and v27.16b, v16.16b, v0.16b\n"
+      "and v26.16b, v17.16b, v1.16b\n"
       "and v25.16b, v18.16b, v2.16b\n"
       "and v24.16b, v19.16b, v3.16b\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
@@ -1716,11 +1716,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v25.4s, v25.4s, #0x1f\n"
       "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v16.4s, v16.4s, v27.4s\n"
-      "and v27.16b, v20.16b, v0.16b\n"
       "sqadd v17.4s, v17.4s, v26.4s\n"
-      "and v26.16b, v21.16b, v1.16b\n"
       "sqadd v18.4s, v18.4s, v25.4s\n"
       "sqadd v19.4s, v19.4s, v24.4s\n"
+      "and v27.16b, v20.16b, v0.16b\n"
+      "and v26.16b, v21.16b, v1.16b\n"
       "and v25.16b, v22.16b, v2.16b\n"
       "and v24.16b, v23.16b, v3.16b\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
@@ -1732,21 +1732,21 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sqadd v22.4s, v22.4s, v25.4s\n"
       "sqadd v23.4s, v23.4s, v24.4s\n"
       "98:"  // Height 4: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x21]\n"
-      "ld1r { v25.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "srshl v12.4s, v12.4s, v0.4s\n"
       "srshl v13.4s, v13.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
       "srshl v14.4s, v14.4s, v2.4s\n"
       "srshl v15.4s, v15.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v1.4s\n"
       "srshl v18.4s, v18.4s, v2.4s\n"
@@ -1816,113 +1816,113 @@ void a64_hybrid_s8qs_dot_6x16 (
       "uzp1 v16.16b, v16.16b, v18.16b\n"
       "uzp1 v20.16b, v20.16b, v17.16b\n"
       "bge 107f\n"
-      "tbz x11, #3, 102f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "tbz x11, #2, 100f\n"
-      "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "tbz x11, #1, 99f\n"
-      "st1 { v8.h }[6], [x9], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "tbz x11, #0, 106f\n"
-      "st1 { v8.b }[14], [x9]\n"
-      "st1 { v12.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
-      "st1 { v20.b }[14], [x24]\n"
+      "tbz x10, #3, 102f\n"
+      "str d8, [x11], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "tbz x10, #2, 100f\n"
+      "st1 { v8.s }[2], [x11], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "tbz x10, #1, 99f\n"
+      "st1 { v8.h }[6], [x11], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "tbz x10, #0, 106f\n"
+      "st1 { v8.b }[14], [x11]\n"
+      "st1 { v12.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
       "b 106f\n"
       "99:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 106f\n"
-      "st1 { v8.b }[12], [x9]\n"
-      "st1 { v12.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
-      "st1 { v20.b }[12], [x24]\n"
+      "tbz x10, #0, 106f\n"
+      "st1 { v8.b }[12], [x11]\n"
+      "st1 { v12.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
       "b 106f\n"
       "100:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 101f\n"
-      "st1 { v8.h }[4], [x9], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "tbz x11, #0, 106f\n"
-      "st1 { v8.b }[10], [x9]\n"
-      "st1 { v12.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
-      "st1 { v20.b }[10], [x24]\n"
+      "tbz x10, #1, 101f\n"
+      "st1 { v8.h }[4], [x11], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "tbz x10, #0, 106f\n"
+      "st1 { v8.b }[10], [x11]\n"
+      "st1 { v12.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
       "b 106f\n"
       "101:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 106f\n"
-      "st1 { v8.b }[8], [x9]\n"
-      "st1 { v12.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
-      "st1 { v20.b }[8], [x24]\n"
+      "tbz x10, #0, 106f\n"
+      "st1 { v8.b }[8], [x11]\n"
+      "st1 { v12.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
       "b 106f\n"
       "102:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 104f\n"
-      "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "tbz x11, #1, 103f\n"
-      "st1 { v8.h }[2], [x9], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "tbz x11, #0, 106f\n"
-      "st1 { v8.b }[6], [x9]\n"
-      "st1 { v12.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
-      "st1 { v20.b }[6], [x24]\n"
+      "tbz x10, #2, 104f\n"
+      "str s8, [x11], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "tbz x10, #1, 103f\n"
+      "st1 { v8.h }[2], [x11], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "tbz x10, #0, 106f\n"
+      "st1 { v8.b }[6], [x11]\n"
+      "st1 { v12.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
       "b 106f\n"
       "103:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 106f\n"
-      "st1 { v8.b }[4], [x9]\n"
-      "st1 { v12.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
-      "st1 { v20.b }[4], [x24]\n"
+      "tbz x10, #0, 106f\n"
+      "st1 { v8.b }[4], [x11]\n"
+      "st1 { v12.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
       "b 106f\n"
       "104:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 105f\n"
-      "str h8, [x9], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "tbz x11, #0, 106f\n"
-      "st1 { v8.b }[2], [x9]\n"
-      "st1 { v12.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
-      "st1 { v20.b }[2], [x24]\n"
+      "tbz x10, #1, 105f\n"
+      "str h8, [x11], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "tbz x10, #0, 106f\n"
+      "st1 { v8.b }[2], [x11]\n"
+      "st1 { v12.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
       "b 106f\n"
       "105:"  // Height 4: Partial direct writeback: partial_1_0
-      "str b8, [x9, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
+      "str b8, [x11, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
       "106:"  // Height 4: Partial direct writeback: Done
       "b 108f\n"
       "107:"  // Height 4: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q12, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q20, [x24, #0x0]\n"
+      "str q8, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q12, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
       "108:"  // Height 4: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 83b\n"
       "b 164f\n"
       "109:"  // Height 5
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "110:"  // Height 5: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -1948,8 +1948,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "mov x28, #0x0\n"
       "112:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 113f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1981,8 +1981,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr q2, [x24, #0x0]\n"
       "ldr q3, [x23, #0x0]\n"
       "ldr q4, [x22, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x0]\n"
+      "ldr q7, [x9, #0x10]\n"
       "blt 116f\n"
       "115:"  // Height 5: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
@@ -1994,7 +1994,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x25, x25, #0x10\n"
       "add x24, x24, #0x10\n"
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x20]\n"
+      "ldr q29, [x9, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x23, x23, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2003,7 +2003,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "cmp x27, #0x20\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x30]\n"
+      "ldr q28, [x9, #0x30]\n"
       "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
@@ -2014,80 +2014,80 @@ void a64_hybrid_s8qs_dot_6x16 (
       "prfm pldl1keep, [x23, #0x80]\n"
       "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4f84e3ba  // sdot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x40]\n"
+      "ldr q29, [x9, #0x40]\n"
       ".inst 0x4f80e38b  // sdot v11.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f81e38f  // sdot v15.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f82e393  // sdot v19.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f83e397  // sdot v23.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x4f84e39b  // sdot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x50]\n"
+      "ldr q28, [x9, #0x50]\n"
       ".inst 0x4fa0e3a8  // sdot v8.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3ac  // sdot v12.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3b0  // sdot v16.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x4fa3e3b4  // sdot v20.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x4fa4e3b8  // sdot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x10, #0x60]\n"
+      "ldr q29, [x9, #0x60]\n"
       ".inst 0x4fa0e389  // sdot v9.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x4fa1e38d  // sdot v13.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x4fa2e391  // sdot v17.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x4fa3e395  // sdot v21.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x4fa4e399  // sdot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x10, #0x70]\n"
+      "ldr q28, [x9, #0x70]\n"
       ".inst 0x4fa0e3aa  // sdot v10.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3ae  // sdot v14.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3b2  // sdot v18.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x4fa3e3b6  // sdot v22.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x4fa4e3ba  // sdot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x10, #0x80]\n"
+      "ldr q29, [x9, #0x80]\n"
       ".inst 0x4fa0e38b  // sdot v11.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x4fa1e38f  // sdot v15.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x4fa2e393  // sdot v19.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x4fa3e397  // sdot v23.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x4fa4e39b  // sdot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x10, #0x90]\n"
+      "ldr q28, [x9, #0x90]\n"
       ".inst 0x4f80eba8  // sdot v8.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebac  // sdot v12.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x4f82ebb0  // sdot v16.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb4  // sdot v20.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x4f84ebb8  // sdot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x10, #0xa0]\n"
+      "ldr q29, [x9, #0xa0]\n"
       ".inst 0x4f80eb89  // sdot v9.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb8d  // sdot v13.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb91  // sdot v17.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x4f83eb95  // sdot v21.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x4f84eb99  // sdot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x10, #0xb0]\n"
+      "ldr q28, [x9, #0xb0]\n"
       ".inst 0x4f80ebaa  // sdot v10.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebae  // sdot v14.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x4f82ebb2  // sdot v18.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb6  // sdot v22.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x4f84ebba  // sdot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x10, #0xc0]\n"
+      "ldr q29, [x9, #0xc0]\n"
       ".inst 0x4f80eb8b  // sdot v11.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb8f  // sdot v15.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb93  // sdot v19.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x4f83eb97  // sdot v23.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x4f84eb9b  // sdot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x10, #0xd0]\n"
+      "ldr q28, [x9, #0xd0]\n"
       ".inst 0x4fa0eba8  // sdot v8.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x4fa1ebac  // sdot v12.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x4fa2ebb0  // sdot v16.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb4  // sdot v20.4s, v29.16b, v3.4b[3]\n"
       ".inst 0x4fa4ebb8  // sdot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr q29, [x10, #0xe0]\n"
+      "ldr q29, [x9, #0xe0]\n"
       ".inst 0x4fa0eb89  // sdot v9.4s, v28.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb8d  // sdot v13.4s, v28.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb91  // sdot v17.4s, v28.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb95  // sdot v21.4s, v28.16b, v3.4b[3]\n"
       ".inst 0x4fa4eb99  // sdot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr q28, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q28, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0ebaa  // sdot v10.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x4fa1ebae  // sdot v14.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x4fa2ebb2  // sdot v18.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb6  // sdot v22.4s, v29.16b, v3.4b[3]\n"
       ".inst 0x4fa4ebba  // sdot v26.4s, v29.16b, v4.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
+      "ldr q6, [x9, #0x0]\n"
       ".inst 0x4fa0eb8b  // sdot v11.4s, v28.16b, v0.4b[3]\n"
       "ldr q0, [x26, #0x0]\n"
       ".inst 0x4fa1eb8f  // sdot v15.4s, v28.16b, v1.4b[3]\n"
@@ -2098,7 +2098,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr q3, [x23, #0x0]\n"
       ".inst 0x4fa4eb9b  // sdot v27.4s, v28.16b, v4.4b[3]\n"
       "ldr q4, [x22, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x10]\n"
       "bge 115b\n"
       "116:"  // Height 5: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
@@ -2110,17 +2110,17 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x24, x24, #0x10\n"
       "add x23, x23, #0x10\n"
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x20]\n"
+      "ldr q29, [x9, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x22, x22, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
+      "sub x27, x27, #0x10\n"
       "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
+      "ldr q28, [x9, #0x30]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
@@ -2129,74 +2129,74 @@ void a64_hybrid_s8qs_dot_6x16 (
       ".inst 0x4f83e3b6  // sdot v22.4s, v29.16b, v3.4b[0]\n"
       "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4f84e3ba  // sdot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x40]\n"
+      "ldr q29, [x9, #0x40]\n"
       ".inst 0x4f80e38b  // sdot v11.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f81e38f  // sdot v15.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f82e393  // sdot v19.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f83e397  // sdot v23.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x4f84e39b  // sdot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x50]\n"
+      "ldr q28, [x9, #0x50]\n"
       ".inst 0x4fa0e3a8  // sdot v8.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3ac  // sdot v12.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3b0  // sdot v16.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x4fa3e3b4  // sdot v20.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x4fa4e3b8  // sdot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x10, #0x60]\n"
+      "ldr q29, [x9, #0x60]\n"
       ".inst 0x4fa0e389  // sdot v9.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x4fa1e38d  // sdot v13.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x4fa2e391  // sdot v17.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x4fa3e395  // sdot v21.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x4fa4e399  // sdot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x10, #0x70]\n"
+      "ldr q28, [x9, #0x70]\n"
       ".inst 0x4fa0e3aa  // sdot v10.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3ae  // sdot v14.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3b2  // sdot v18.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x4fa3e3b6  // sdot v22.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x4fa4e3ba  // sdot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x10, #0x80]\n"
+      "ldr q29, [x9, #0x80]\n"
       ".inst 0x4fa0e38b  // sdot v11.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x4fa1e38f  // sdot v15.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x4fa2e393  // sdot v19.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x4fa3e397  // sdot v23.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x4fa4e39b  // sdot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x10, #0x90]\n"
+      "ldr q28, [x9, #0x90]\n"
       ".inst 0x4f80eba8  // sdot v8.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebac  // sdot v12.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x4f82ebb0  // sdot v16.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb4  // sdot v20.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x4f84ebb8  // sdot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x10, #0xa0]\n"
+      "ldr q29, [x9, #0xa0]\n"
       ".inst 0x4f80eb89  // sdot v9.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb8d  // sdot v13.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb91  // sdot v17.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x4f83eb95  // sdot v21.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x4f84eb99  // sdot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x10, #0xb0]\n"
+      "ldr q28, [x9, #0xb0]\n"
       ".inst 0x4f80ebaa  // sdot v10.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebae  // sdot v14.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x4f82ebb2  // sdot v18.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb6  // sdot v22.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x4f84ebba  // sdot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x10, #0xc0]\n"
+      "ldr q29, [x9, #0xc0]\n"
       ".inst 0x4f80eb8b  // sdot v11.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb8f  // sdot v15.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb93  // sdot v19.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x4f83eb97  // sdot v23.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x4f84eb9b  // sdot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x10, #0xd0]\n"
+      "ldr q28, [x9, #0xd0]\n"
       ".inst 0x4fa0eba8  // sdot v8.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x4fa1ebac  // sdot v12.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x4fa2ebb0  // sdot v16.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb4  // sdot v20.4s, v29.16b, v3.4b[3]\n"
       ".inst 0x4fa4ebb8  // sdot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr q29, [x10, #0xe0]\n"
+      "ldr q29, [x9, #0xe0]\n"
       ".inst 0x4fa0eb89  // sdot v9.4s, v28.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb8d  // sdot v13.4s, v28.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb91  // sdot v17.4s, v28.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb95  // sdot v21.4s, v28.16b, v3.4b[3]\n"
       ".inst 0x4fa4eb99  // sdot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr q28, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q28, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0ebaa  // sdot v10.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x4fa1ebae  // sdot v14.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x4fa2ebb2  // sdot v18.4s, v29.16b, v2.4b[3]\n"
@@ -2215,25 +2215,25 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr s2, [x26], #0x4\n"
       "ldr s1, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s0, [x24], #0x4\n"
       "ldr s31, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
       "ldr s30, [x22], #0x4\n"
-      "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
+      "ldr q29, [x9, #0x0]\n"
       ".inst 0x4f82e3a8  // sdot v8.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x4f81e3ac  // sdot v12.4s, v29.16b, v1.4b[0]\n"
+      "ldr q28, [x9, #0x10]\n"
       ".inst 0x4f80e3b0  // sdot v16.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f9fe3b4  // sdot v20.4s, v29.16b, v31.4b[0]\n"
       ".inst 0x4f9ee3b8  // sdot v24.4s, v29.16b, v30.4b[0]\n"
-      "ldr q29, [x10, #0x20]\n"
+      "ldr q29, [x9, #0x20]\n"
       ".inst 0x4f82e389  // sdot v9.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f81e38d  // sdot v13.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f80e391  // sdot v17.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f9fe395  // sdot v21.4s, v28.16b, v31.4b[0]\n"
       ".inst 0x4f9ee399  // sdot v25.4s, v28.16b, v30.4b[0]\n"
-      "ldr q28, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q28, [x9, #0x30]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f82e3aa  // sdot v10.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f80e3b2  // sdot v18.4s, v29.16b, v0.4b[0]\n"
@@ -2267,21 +2267,21 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr b3, [x23, #0x0]\n"
       "ldr b4, [x22, #0x0]\n"
       "121:"  // Height 5: Multiply loop: Ragged operand read: Done
-      "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
+      "ldr q29, [x9, #0x0]\n"
+      "ldr q28, [x9, #0x10]\n"
       ".inst 0x4f80e3a8  // sdot v8.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f81e3ac  // sdot v12.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f82e3b0  // sdot v16.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x4f83e3b4  // sdot v20.4s, v29.16b, v3.4b[0]\n"
       ".inst 0x4f84e3b8  // sdot v24.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x20]\n"
+      "ldr q29, [x9, #0x20]\n"
       ".inst 0x4f80e389  // sdot v9.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f81e38d  // sdot v13.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f82e391  // sdot v17.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f83e395  // sdot v21.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x4f84e399  // sdot v25.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q28, [x9, #0x30]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f82e3b2  // sdot v18.4s, v29.16b, v2.4b[0]\n"
@@ -2299,28 +2299,28 @@ void a64_hybrid_s8qs_dot_6x16 (
       "bne 112b\n"
       "ldr q31, [x14, #0x0]\n"
       "ldr q30, [x14, #0x10]\n"
-      "ldr q29, [x14, #0x20]\n"
-      "ldr q28, [x14, #0x30]\n"
-      "add x14, x14, #0x40\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add v8.4s, v8.4s, v31.4s\n"
       "add v9.4s, v9.4s, v30.4s\n"
-      "add v12.4s, v12.4s, v31.4s\n"
-      "add v13.4s, v13.4s, v30.4s\n"
+      "ldr q29, [x14, #0x20]\n"
+      "ldr q28, [x14, #0x30]\n"
       "add v10.4s, v10.4s, v29.4s\n"
       "add v11.4s, v11.4s, v28.4s\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x11, x20\n"
       "add x24, x25, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "add v14.4s, v14.4s, v29.4s\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "add x23, x24, x20\n"
+      "add x22, x23, x20\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
+      "add v12.4s, v12.4s, v31.4s\n"
+      "add v13.4s, v13.4s, v30.4s\n"
+      "add v14.4s, v14.4s, v29.4s\n"
       "add v15.4s, v15.4s, v28.4s\n"
+      "add x14, x14, #0x40\n"
       "add v16.4s, v16.4s, v31.4s\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "add v17.4s, v17.4s, v30.4s\n"
       "add v18.4s, v18.4s, v29.4s\n"
       "add v19.4s, v19.4s, v28.4s\n"
@@ -2345,9 +2345,9 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x13, x13, #0x40\n"
       "b 124f\n"
       "123:"  // Height 5: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -2386,11 +2386,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v31.4s\n"
-      "and v31.16b, v12.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v30.4s\n"
-      "and v30.16b, v13.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v29.4s\n"
       "sqadd v11.4s, v11.4s, v28.4s\n"
+      "and v31.16b, v12.16b, v0.16b\n"
+      "and v30.16b, v13.16b, v1.16b\n"
       "and v29.16b, v14.16b, v2.16b\n"
       "and v28.16b, v15.16b, v3.16b\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
@@ -2398,11 +2398,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sqadd v12.4s, v12.4s, v31.4s\n"
-      "and v31.16b, v16.16b, v0.16b\n"
       "sqadd v13.4s, v13.4s, v30.4s\n"
-      "and v30.16b, v17.16b, v1.16b\n"
       "sqadd v14.4s, v14.4s, v29.4s\n"
       "sqadd v15.4s, v15.4s, v28.4s\n"
+      "and v31.16b, v16.16b, v0.16b\n"
+      "and v30.16b, v17.16b, v1.16b\n"
       "and v29.16b, v18.16b, v2.16b\n"
       "and v28.16b, v19.16b, v3.16b\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
@@ -2410,11 +2410,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sqadd v16.4s, v16.4s, v31.4s\n"
-      "and v31.16b, v20.16b, v0.16b\n"
       "sqadd v17.4s, v17.4s, v30.4s\n"
-      "and v30.16b, v21.16b, v1.16b\n"
       "sqadd v18.4s, v18.4s, v29.4s\n"
       "sqadd v19.4s, v19.4s, v28.4s\n"
+      "and v31.16b, v20.16b, v0.16b\n"
+      "and v30.16b, v21.16b, v1.16b\n"
       "and v29.16b, v22.16b, v2.16b\n"
       "and v28.16b, v23.16b, v3.16b\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
@@ -2422,11 +2422,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sqadd v20.4s, v20.4s, v31.4s\n"
-      "and v31.16b, v24.16b, v0.16b\n"
       "sqadd v21.4s, v21.4s, v30.4s\n"
-      "and v30.16b, v25.16b, v1.16b\n"
       "sqadd v22.4s, v22.4s, v29.4s\n"
       "sqadd v23.4s, v23.4s, v28.4s\n"
+      "and v31.16b, v24.16b, v0.16b\n"
+      "and v30.16b, v25.16b, v1.16b\n"
       "and v29.16b, v26.16b, v2.16b\n"
       "and v28.16b, v27.16b, v3.16b\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
@@ -2438,21 +2438,21 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sqadd v26.4s, v26.4s, v29.4s\n"
       "sqadd v27.4s, v27.4s, v28.4s\n"
       "125:"  // Height 5: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v30.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v30.4s }, [x21]\n"
-      "ld1r { v29.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "srshl v12.4s, v12.4s, v0.4s\n"
       "srshl v13.4s, v13.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v28.4s }, [x20]\n"
       "srshl v14.4s, v14.4s, v2.4s\n"
       "srshl v15.4s, v15.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v1.4s\n"
       "srshl v18.4s, v18.4s, v2.4s\n"
@@ -2541,133 +2541,132 @@ void a64_hybrid_s8qs_dot_6x16 (
       "uzp1 v20.16b, v20.16b, v18.16b\n"
       "uzp1 v24.16b, v24.16b, v17.16b\n"
       "bge 134f\n"
-      "tbz x11, #3, 129f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "tbz x11, #2, 127f\n"
-      "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "tbz x11, #1, 126f\n"
-      "st1 { v8.h }[6], [x9], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "tbz x11, #0, 133f\n"
-      "st1 { v8.b }[14], [x9]\n"
-      "st1 { v12.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "tbz x10, #3, 129f\n"
+      "str d8, [x11], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "tbz x10, #2, 127f\n"
+      "st1 { v8.s }[2], [x11], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "tbz x10, #1, 126f\n"
+      "st1 { v8.h }[6], [x11], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "tbz x10, #0, 133f\n"
+      "st1 { v8.b }[14], [x11]\n"
+      "st1 { v12.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 133f\n"
       "126:"  // Height 5: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 133f\n"
-      "st1 { v8.b }[12], [x9]\n"
-      "st1 { v12.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "tbz x10, #0, 133f\n"
+      "st1 { v8.b }[12], [x11]\n"
+      "st1 { v12.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 133f\n"
       "127:"  // Height 5: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 128f\n"
-      "st1 { v8.h }[4], [x9], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "tbz x11, #0, 133f\n"
-      "st1 { v8.b }[10], [x9]\n"
-      "st1 { v12.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "tbz x10, #1, 128f\n"
+      "st1 { v8.h }[4], [x11], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "tbz x10, #0, 133f\n"
+      "st1 { v8.b }[10], [x11]\n"
+      "st1 { v12.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 133f\n"
       "128:"  // Height 5: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 133f\n"
-      "st1 { v8.b }[8], [x9]\n"
-      "st1 { v12.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "tbz x10, #0, 133f\n"
+      "st1 { v8.b }[8], [x11]\n"
+      "st1 { v12.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 133f\n"
       "129:"  // Height 5: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 131f\n"
-      "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "tbz x11, #1, 130f\n"
-      "st1 { v8.h }[2], [x9], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "tbz x11, #0, 133f\n"
-      "st1 { v8.b }[6], [x9]\n"
-      "st1 { v12.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "tbz x10, #2, 131f\n"
+      "str s8, [x11], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "tbz x10, #1, 130f\n"
+      "st1 { v8.h }[2], [x11], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "tbz x10, #0, 133f\n"
+      "st1 { v8.b }[6], [x11]\n"
+      "st1 { v12.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 133f\n"
       "130:"  // Height 5: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 133f\n"
-      "st1 { v8.b }[4], [x9]\n"
-      "st1 { v12.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "tbz x10, #0, 133f\n"
+      "st1 { v8.b }[4], [x11]\n"
+      "st1 { v12.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 133f\n"
       "131:"  // Height 5: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 132f\n"
-      "str h8, [x9], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "tbz x11, #0, 133f\n"
-      "st1 { v8.b }[2], [x9]\n"
-      "st1 { v12.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "tbz x10, #1, 132f\n"
+      "str h8, [x11], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "tbz x10, #0, 133f\n"
+      "st1 { v8.b }[2], [x11]\n"
+      "st1 { v12.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 133f\n"
       "132:"  // Height 5: Partial direct writeback: partial_1_0
-      "str b8, [x9, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b8, [x11, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "133:"  // Height 5: Partial direct writeback: Done
       "b 135f\n"
       "134:"  // Height 5: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q12, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q8, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q12, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "135:"  // Height 5: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 110b\n"
       "b 164f\n"
       "136:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x6\n"
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "137:"  // Height 6: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -2697,8 +2696,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "mov x28, #0x0\n"
       "139:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 140f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2734,8 +2733,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr q3, [x23, #0x0]\n"
       "ldr q4, [x22, #0x0]\n"
       "ldr q5, [x21, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x0]\n"
+      "ldr q7, [x9, #0x10]\n"
       "blt 143f\n"
       "142:"  // Height 6: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
@@ -2748,7 +2747,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x24, x24, #0x10\n"
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0dc  // sdot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x20]\n"
+      "ldr q6, [x9, #0x20]\n"
       "add x23, x23, #0x10\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2760,7 +2759,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0fd  // sdot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x30]\n"
+      "ldr q7, [x9, #0x30]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e0ca  // sdot v10.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ce  // sdot v14.4s, v6.16b, v1.4b[0]\n"
@@ -2772,92 +2771,92 @@ void a64_hybrid_s8qs_dot_6x16 (
       "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4f84e0da  // sdot v26.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0de  // sdot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x40]\n"
+      "ldr q6, [x9, #0x40]\n"
       ".inst 0x4f80e0eb  // sdot v11.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ef  // sdot v15.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f3  // sdot v19.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f7  // sdot v23.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0fb  // sdot v27.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0ff  // sdot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x50]\n"
+      "ldr q7, [x9, #0x50]\n"
       ".inst 0x4fa0e0c8  // sdot v8.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0cc  // sdot v12.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0d0  // sdot v16.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d4  // sdot v20.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0d8  // sdot v24.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0dc  // sdot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x10, #0x60]\n"
+      "ldr q6, [x9, #0x60]\n"
       ".inst 0x4fa0e0e9  // sdot v9.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ed  // sdot v13.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0f1  // sdot v17.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0f5  // sdot v21.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0f9  // sdot v25.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0fd  // sdot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x10, #0x70]\n"
+      "ldr q7, [x9, #0x70]\n"
       ".inst 0x4fa0e0ca  // sdot v10.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ce  // sdot v14.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0d2  // sdot v18.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d6  // sdot v22.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0da  // sdot v26.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0de  // sdot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x10, #0x80]\n"
+      "ldr q6, [x9, #0x80]\n"
       ".inst 0x4fa0e0eb  // sdot v11.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ef  // sdot v15.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0f3  // sdot v19.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0f7  // sdot v23.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0fb  // sdot v27.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0ff  // sdot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x10, #0x90]\n"
+      "ldr q7, [x9, #0x90]\n"
       ".inst 0x4f80e8c8  // sdot v8.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8cc  // sdot v12.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d0  // sdot v16.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d4  // sdot v20.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f84e8d8  // sdot v24.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8dc  // sdot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x10, #0xa0]\n"
+      "ldr q6, [x9, #0xa0]\n"
       ".inst 0x4f80e8e9  // sdot v9.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ed  // sdot v13.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x4f82e8f1  // sdot v17.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x4f83e8f5  // sdot v21.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x4f84e8f9  // sdot v25.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8fd  // sdot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x10, #0xb0]\n"
+      "ldr q7, [x9, #0xb0]\n"
       ".inst 0x4f80e8ca  // sdot v10.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ce  // sdot v14.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d2  // sdot v18.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d6  // sdot v22.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f84e8da  // sdot v26.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8de  // sdot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x10, #0xc0]\n"
+      "ldr q6, [x9, #0xc0]\n"
       ".inst 0x4f80e8eb  // sdot v11.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ef  // sdot v15.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x4f82e8f3  // sdot v19.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x4f83e8f7  // sdot v23.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x4f84e8fb  // sdot v27.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8ff  // sdot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x10, #0xd0]\n"
+      "ldr q7, [x9, #0xd0]\n"
       ".inst 0x4fa0e8c8  // sdot v8.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8cc  // sdot v12.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8d0  // sdot v16.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d4  // sdot v20.4s, v6.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8d8  // sdot v24.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8dc  // sdot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x10, #0xe0]\n"
+      "ldr q6, [x9, #0xe0]\n"
       ".inst 0x4fa0e8e9  // sdot v9.4s, v7.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8ed  // sdot v13.4s, v7.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8f1  // sdot v17.4s, v7.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8f5  // sdot v21.4s, v7.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8f9  // sdot v25.4s, v7.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8fd  // sdot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr q7, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q7, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0e8ca  // sdot v10.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8ce  // sdot v14.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8d2  // sdot v18.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d6  // sdot v22.4s, v6.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8da  // sdot v26.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8de  // sdot v30.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
+      "ldr q6, [x9, #0x0]\n"
       ".inst 0x4fa0e8eb  // sdot v11.4s, v7.16b, v0.4b[3]\n"
       "ldr q0, [x26, #0x0]\n"
       ".inst 0x4fa1e8ef  // sdot v15.4s, v7.16b, v1.4b[3]\n"
@@ -2870,7 +2869,7 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr q4, [x22, #0x0]\n"
       ".inst 0x4fa5e8ff  // sdot v31.4s, v7.16b, v5.4b[3]\n"
       "ldr q5, [x21, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x10]\n"
       "bge 142b\n"
       "143:"  // Height 6: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
@@ -2883,108 +2882,108 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x23, x23, #0x10\n"
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0dc  // sdot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x20]\n"
+      "ldr q6, [x9, #0x20]\n"
       "add x22, x22, #0x10\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0fd  // sdot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "ldr q7, [x9, #0x30]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f80e0ca  // sdot v10.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ce  // sdot v14.4s, v6.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4f82e0d2  // sdot v18.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0d6  // sdot v22.4s, v6.16b, v3.4b[0]\n"
       "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4f84e0da  // sdot v26.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0de  // sdot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x40]\n"
+      "ldr q6, [x9, #0x40]\n"
       ".inst 0x4f80e0eb  // sdot v11.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ef  // sdot v15.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f3  // sdot v19.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f7  // sdot v23.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0fb  // sdot v27.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0ff  // sdot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x50]\n"
+      "ldr q7, [x9, #0x50]\n"
       ".inst 0x4fa0e0c8  // sdot v8.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0cc  // sdot v12.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0d0  // sdot v16.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d4  // sdot v20.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0d8  // sdot v24.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0dc  // sdot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x10, #0x60]\n"
+      "ldr q6, [x9, #0x60]\n"
       ".inst 0x4fa0e0e9  // sdot v9.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ed  // sdot v13.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0f1  // sdot v17.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0f5  // sdot v21.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0f9  // sdot v25.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0fd  // sdot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x10, #0x70]\n"
+      "ldr q7, [x9, #0x70]\n"
       ".inst 0x4fa0e0ca  // sdot v10.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ce  // sdot v14.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0d2  // sdot v18.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d6  // sdot v22.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0da  // sdot v26.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0de  // sdot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x10, #0x80]\n"
+      "ldr q6, [x9, #0x80]\n"
       ".inst 0x4fa0e0eb  // sdot v11.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ef  // sdot v15.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0f3  // sdot v19.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0f7  // sdot v23.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0fb  // sdot v27.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0ff  // sdot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x10, #0x90]\n"
+      "ldr q7, [x9, #0x90]\n"
       ".inst 0x4f80e8c8  // sdot v8.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8cc  // sdot v12.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d0  // sdot v16.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d4  // sdot v20.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f84e8d8  // sdot v24.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8dc  // sdot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x10, #0xa0]\n"
+      "ldr q6, [x9, #0xa0]\n"
       ".inst 0x4f80e8e9  // sdot v9.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ed  // sdot v13.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x4f82e8f1  // sdot v17.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x4f83e8f5  // sdot v21.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x4f84e8f9  // sdot v25.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8fd  // sdot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x10, #0xb0]\n"
+      "ldr q7, [x9, #0xb0]\n"
       ".inst 0x4f80e8ca  // sdot v10.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ce  // sdot v14.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d2  // sdot v18.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d6  // sdot v22.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f84e8da  // sdot v26.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8de  // sdot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x10, #0xc0]\n"
+      "ldr q6, [x9, #0xc0]\n"
       ".inst 0x4f80e8eb  // sdot v11.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ef  // sdot v15.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x4f82e8f3  // sdot v19.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x4f83e8f7  // sdot v23.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x4f84e8fb  // sdot v27.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8ff  // sdot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x10, #0xd0]\n"
+      "ldr q7, [x9, #0xd0]\n"
       ".inst 0x4fa0e8c8  // sdot v8.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8cc  // sdot v12.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8d0  // sdot v16.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d4  // sdot v20.4s, v6.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8d8  // sdot v24.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8dc  // sdot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x10, #0xe0]\n"
+      "ldr q6, [x9, #0xe0]\n"
       ".inst 0x4fa0e8e9  // sdot v9.4s, v7.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8ed  // sdot v13.4s, v7.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8f1  // sdot v17.4s, v7.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8f5  // sdot v21.4s, v7.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8f9  // sdot v25.4s, v7.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8fd  // sdot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr q7, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q7, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4fa0e8ca  // sdot v10.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8ce  // sdot v14.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8d2  // sdot v18.4s, v6.16b, v2.4b[3]\n"
@@ -3005,28 +3004,28 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr s7, [x26], #0x4\n"
       "ldr s6, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s5, [x24], #0x4\n"
       "ldr s4, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
       "ldr s3, [x22], #0x4\n"
       "ldr s2, [x21], #0x4\n"
-      "ldr q1, [x10, #0x0]\n"
-      "ldr q0, [x10, #0x10]\n"
+      "ldr q1, [x9, #0x0]\n"
+      "ldr q0, [x9, #0x10]\n"
       ".inst 0x4f87e028  // sdot v8.4s, v1.16b, v7.4b[0]\n"
       ".inst 0x4f86e02c  // sdot v12.4s, v1.16b, v6.4b[0]\n"
       ".inst 0x4f85e030  // sdot v16.4s, v1.16b, v5.4b[0]\n"
       ".inst 0x4f84e034  // sdot v20.4s, v1.16b, v4.4b[0]\n"
       ".inst 0x4f83e038  // sdot v24.4s, v1.16b, v3.4b[0]\n"
       ".inst 0x4f82e03c  // sdot v28.4s, v1.16b, v2.4b[0]\n"
-      "ldr q1, [x10, #0x20]\n"
+      "ldr q1, [x9, #0x20]\n"
       ".inst 0x4f87e009  // sdot v9.4s, v0.16b, v7.4b[0]\n"
       ".inst 0x4f86e00d  // sdot v13.4s, v0.16b, v6.4b[0]\n"
       ".inst 0x4f85e011  // sdot v17.4s, v0.16b, v5.4b[0]\n"
       ".inst 0x4f84e015  // sdot v21.4s, v0.16b, v4.4b[0]\n"
       ".inst 0x4f83e019  // sdot v25.4s, v0.16b, v3.4b[0]\n"
       ".inst 0x4f82e01d  // sdot v29.4s, v0.16b, v2.4b[0]\n"
-      "ldr q0, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q0, [x9, #0x30]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f87e02a  // sdot v10.4s, v1.16b, v7.4b[0]\n"
       ".inst 0x4f86e02e  // sdot v14.4s, v1.16b, v6.4b[0]\n"
       ".inst 0x4f85e032  // sdot v18.4s, v1.16b, v5.4b[0]\n"
@@ -3065,23 +3064,23 @@ void a64_hybrid_s8qs_dot_6x16 (
       "ldr b4, [x22, #0x0]\n"
       "ldr b5, [x21, #0x0]\n"
       "148:"  // Height 6: Multiply loop: Ragged operand read: Done
-      "ldr q7, [x10, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x0]\n"
+      "ldr q6, [x9, #0x10]\n"
       ".inst 0x4f80e0e8  // sdot v8.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ec  // sdot v12.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f0  // sdot v16.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f4  // sdot v20.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0f8  // sdot v24.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0fc  // sdot v28.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x20]\n"
+      "ldr q7, [x9, #0x20]\n"
       ".inst 0x4f80e0c9  // sdot v9.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0cd  // sdot v13.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x4f82e0d1  // sdot v17.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0d5  // sdot v21.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x4f84e0d9  // sdot v25.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0dd  // sdot v29.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
+      "ldr q6, [x9, #0x30]\n"
+      "add x9, x9, #0x40\n"
       ".inst 0x4f80e0ea  // sdot v10.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ee  // sdot v14.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f2  // sdot v18.4s, v7.16b, v2.4b[0]\n"
@@ -3101,32 +3100,32 @@ void a64_hybrid_s8qs_dot_6x16 (
       "bne 139b\n"
       "ldr q3, [x14, #0x0]\n"
       "ldr q2, [x14, #0x10]\n"
-      "ldr q1, [x14, #0x20]\n"
-      "ldr q0, [x14, #0x30]\n"
-      "add x14, x14, #0x40\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add v8.4s, v8.4s, v3.4s\n"
       "add v9.4s, v9.4s, v2.4s\n"
-      "add v12.4s, v12.4s, v3.4s\n"
-      "add v13.4s, v13.4s, v2.4s\n"
+      "ldr q1, [x14, #0x20]\n"
+      "ldr q0, [x14, #0x30]\n"
       "add v10.4s, v10.4s, v1.4s\n"
       "add v11.4s, v11.4s, v0.4s\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x11, x20\n"
       "add x24, x25, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
-      "add v14.4s, v14.4s, v1.4s\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "add x23, x24, x20\n"
+      "add x22, x23, x20\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
+      "add x21, x22, x20\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
+      "add v12.4s, v12.4s, v3.4s\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
+      "add v13.4s, v13.4s, v2.4s\n"
+      "add v14.4s, v14.4s, v1.4s\n"
+      "add x14, x14, #0x40\n"
       "add v15.4s, v15.4s, v0.4s\n"
       "add v16.4s, v16.4s, v3.4s\n"
-      "add x22, x23, x20\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "add v17.4s, v17.4s, v2.4s\n"
       "add v18.4s, v18.4s, v1.4s\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "add v19.4s, v19.4s, v0.4s\n"
       "add v20.4s, v20.4s, v3.4s\n"
       "add v21.4s, v21.4s, v2.4s\n"
@@ -3153,9 +3152,9 @@ void a64_hybrid_s8qs_dot_6x16 (
       "add x13, x13, #0x40\n"
       "b 151f\n"
       "150:"  // Height 6: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -3198,11 +3197,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v7.4s\n"
-      "and v7.16b, v12.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v6.4s\n"
-      "and v6.16b, v13.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v5.4s\n"
       "sqadd v11.4s, v11.4s, v4.4s\n"
+      "and v7.16b, v12.16b, v0.16b\n"
+      "and v6.16b, v13.16b, v1.16b\n"
       "and v5.16b, v14.16b, v2.16b\n"
       "and v4.16b, v15.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3210,11 +3209,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v12.4s, v12.4s, v7.4s\n"
-      "and v7.16b, v16.16b, v0.16b\n"
       "sqadd v13.4s, v13.4s, v6.4s\n"
-      "and v6.16b, v17.16b, v1.16b\n"
       "sqadd v14.4s, v14.4s, v5.4s\n"
       "sqadd v15.4s, v15.4s, v4.4s\n"
+      "and v7.16b, v16.16b, v0.16b\n"
+      "and v6.16b, v17.16b, v1.16b\n"
       "and v5.16b, v18.16b, v2.16b\n"
       "and v4.16b, v19.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3222,11 +3221,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v16.4s, v16.4s, v7.4s\n"
-      "and v7.16b, v20.16b, v0.16b\n"
       "sqadd v17.4s, v17.4s, v6.4s\n"
-      "and v6.16b, v21.16b, v1.16b\n"
       "sqadd v18.4s, v18.4s, v5.4s\n"
       "sqadd v19.4s, v19.4s, v4.4s\n"
+      "and v7.16b, v20.16b, v0.16b\n"
+      "and v6.16b, v21.16b, v1.16b\n"
       "and v5.16b, v22.16b, v2.16b\n"
       "and v4.16b, v23.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3234,11 +3233,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v20.4s, v20.4s, v7.4s\n"
-      "and v7.16b, v24.16b, v0.16b\n"
       "sqadd v21.4s, v21.4s, v6.4s\n"
-      "and v6.16b, v25.16b, v1.16b\n"
       "sqadd v22.4s, v22.4s, v5.4s\n"
       "sqadd v23.4s, v23.4s, v4.4s\n"
+      "and v7.16b, v24.16b, v0.16b\n"
+      "and v6.16b, v25.16b, v1.16b\n"
       "and v5.16b, v26.16b, v2.16b\n"
       "and v4.16b, v27.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3246,11 +3245,11 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v24.4s, v24.4s, v7.4s\n"
-      "and v7.16b, v28.16b, v0.16b\n"
       "sqadd v25.4s, v25.4s, v6.4s\n"
-      "and v6.16b, v29.16b, v1.16b\n"
       "sqadd v26.4s, v26.4s, v5.4s\n"
       "sqadd v27.4s, v27.4s, v4.4s\n"
+      "and v7.16b, v28.16b, v0.16b\n"
+      "and v6.16b, v29.16b, v1.16b\n"
       "and v5.16b, v30.16b, v2.16b\n"
       "and v4.16b, v31.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3262,21 +3261,21 @@ void a64_hybrid_s8qs_dot_6x16 (
       "sqadd v30.4s, v30.4s, v5.4s\n"
       "sqadd v31.4s, v31.4s, v4.4s\n"
       "152:"  // Height 6: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v6.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v6.4s }, [x21]\n"
-      "ld1r { v5.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v5.4s }, [x20]\n"
       "srshl v12.4s, v12.4s, v0.4s\n"
       "srshl v13.4s, v13.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v4.4s }, [x20]\n"
       "srshl v14.4s, v14.4s, v2.4s\n"
       "srshl v15.4s, v15.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v1.4s\n"
       "srshl v18.4s, v18.4s, v2.4s\n"
@@ -3384,136 +3383,136 @@ void a64_hybrid_s8qs_dot_6x16 (
       "uzp1 v24.16b, v24.16b, v18.16b\n"
       "uzp1 v28.16b, v28.16b, v17.16b\n"
       "bge 161f\n"
-      "tbz x11, #3, 156f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
-      "tbz x11, #2, 154f\n"
-      "st1 { v8.s }[2], [x9], #0x4\n"
-      "st1 { v12.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
-      "tbz x11, #1, 153f\n"
-      "st1 { v8.h }[6], [x9], #0x2\n"
-      "st1 { v12.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "st1 { v28.h }[6], [x22], #0x2\n"
-      "tbz x11, #0, 160f\n"
-      "st1 { v8.b }[14], [x9]\n"
-      "st1 { v12.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "st1 { v28.b }[14], [x22]\n"
+      "tbz x10, #3, 156f\n"
+      "str d8, [x11], #0x8\n"
+      "str d12, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
+      "tbz x10, #2, 154f\n"
+      "st1 { v8.s }[2], [x11], #0x4\n"
+      "st1 { v12.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
+      "tbz x10, #1, 153f\n"
+      "st1 { v8.h }[6], [x11], #0x2\n"
+      "st1 { v12.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "st1 { v28.h }[6], [x21], #0x2\n"
+      "tbz x10, #0, 160f\n"
+      "st1 { v8.b }[14], [x11]\n"
+      "st1 { v12.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
+      "st1 { v28.b }[14], [x21]\n"
       "b 160f\n"
       "153:"  // Height 6: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 160f\n"
-      "st1 { v8.b }[12], [x9]\n"
-      "st1 { v12.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "st1 { v28.b }[12], [x22]\n"
+      "tbz x10, #0, 160f\n"
+      "st1 { v8.b }[12], [x11]\n"
+      "st1 { v12.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
+      "st1 { v28.b }[12], [x21]\n"
       "b 160f\n"
       "154:"  // Height 6: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 155f\n"
-      "st1 { v8.h }[4], [x9], #0x2\n"
-      "st1 { v12.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "st1 { v28.h }[4], [x22], #0x2\n"
-      "tbz x11, #0, 160f\n"
-      "st1 { v8.b }[10], [x9]\n"
-      "st1 { v12.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "st1 { v28.b }[10], [x22]\n"
+      "tbz x10, #1, 155f\n"
+      "st1 { v8.h }[4], [x11], #0x2\n"
+      "st1 { v12.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "st1 { v28.h }[4], [x21], #0x2\n"
+      "tbz x10, #0, 160f\n"
+      "st1 { v8.b }[10], [x11]\n"
+      "st1 { v12.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
+      "st1 { v28.b }[10], [x21]\n"
       "b 160f\n"
       "155:"  // Height 6: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 160f\n"
-      "st1 { v8.b }[8], [x9]\n"
-      "st1 { v12.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "st1 { v28.b }[8], [x22]\n"
+      "tbz x10, #0, 160f\n"
+      "st1 { v8.b }[8], [x11]\n"
+      "st1 { v12.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
+      "st1 { v28.b }[8], [x21]\n"
       "b 160f\n"
       "156:"  // Height 6: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 158f\n"
-      "str s8, [x9], #0x4\n"
-      "str s12, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
-      "tbz x11, #1, 157f\n"
-      "st1 { v8.h }[2], [x9], #0x2\n"
-      "st1 { v12.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "st1 { v28.h }[2], [x22], #0x2\n"
-      "tbz x11, #0, 160f\n"
-      "st1 { v8.b }[6], [x9]\n"
-      "st1 { v12.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "st1 { v28.b }[6], [x22]\n"
+      "tbz x10, #2, 158f\n"
+      "str s8, [x11], #0x4\n"
+      "str s12, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
+      "tbz x10, #1, 157f\n"
+      "st1 { v8.h }[2], [x11], #0x2\n"
+      "st1 { v12.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "st1 { v28.h }[2], [x21], #0x2\n"
+      "tbz x10, #0, 160f\n"
+      "st1 { v8.b }[6], [x11]\n"
+      "st1 { v12.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
+      "st1 { v28.b }[6], [x21]\n"
       "b 160f\n"
       "157:"  // Height 6: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 160f\n"
-      "st1 { v8.b }[4], [x9]\n"
-      "st1 { v12.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "st1 { v28.b }[4], [x22]\n"
+      "tbz x10, #0, 160f\n"
+      "st1 { v8.b }[4], [x11]\n"
+      "st1 { v12.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
+      "st1 { v28.b }[4], [x21]\n"
       "b 160f\n"
       "158:"  // Height 6: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 159f\n"
-      "str h8, [x9], #0x2\n"
-      "str h12, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "str h28, [x22], #0x2\n"
-      "tbz x11, #0, 160f\n"
-      "st1 { v8.b }[2], [x9]\n"
-      "st1 { v12.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "st1 { v28.b }[2], [x22]\n"
+      "tbz x10, #1, 159f\n"
+      "str h8, [x11], #0x2\n"
+      "str h12, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "str h28, [x21], #0x2\n"
+      "tbz x10, #0, 160f\n"
+      "st1 { v8.b }[2], [x11]\n"
+      "st1 { v12.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
+      "st1 { v28.b }[2], [x21]\n"
       "b 160f\n"
       "159:"  // Height 6: Partial direct writeback: partial_1_0
-      "str b8, [x9, #0x0]\n"
-      "str b12, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "str b28, [x22, #0x0]\n"
+      "str b8, [x11, #0x0]\n"
+      "str b12, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
+      "str b28, [x21, #0x0]\n"
       "160:"  // Height 6: Partial direct writeback: Done
       "b 162f\n"
       "161:"  // Height 6: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q12, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q28, [x22, #0x0]\n"
+      "str q8, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q12, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q28, [x21, #0x0]\n"
       "162:"  // Height 6: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 137b\n"
       "subs %x[M], %x[M], #0x6\n"
       "beq 164f\n"
@@ -3527,8 +3526,8 @@ void a64_hybrid_s8qs_dot_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "164:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16.hpp
index 494370ade7..65f654012e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 8> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp
index 867bcded1f..5d34a5e9d2 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp
@@ -47,18 +47,18 @@ void a64_hybrid_s8qs_mmla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -97,9 +97,9 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -113,8 +113,8 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "mov x28, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -129,115 +129,115 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "cmp x27, #0x10\n"
       "blt 9f\n"
       "ldr q1, [x26, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       "cmp x27, #0x20\n"
-      "ldr q6, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x10]\n"
       "blt 8f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "trn1 v18.2d, v1.2d, v21.2d\n"
-      "trn2 v1.2d, v1.2d, v21.2d\n"
       ".inst 0x4e87a648  // smmla v8.4s, v18.16b, v7.16b\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4e86a64c  // smmla v12.4s, v18.16b, v6.16b\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4e91a649  // smmla v9.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
       ".inst 0x4e90a64d  // smmla v13.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4e91a64a  // smmla v10.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
       ".inst 0x4e90a64e  // smmla v14.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x70]\n"
+      "ldr q16, [x9, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v21.2d\n"
       ".inst 0x4e91a64b  // smmla v11.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x80]\n"
+      "ldr q17, [x9, #0x80]\n"
       ".inst 0x4e90a64f  // smmla v15.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x90]\n"
+      "ldr q16, [x9, #0x90]\n"
       ".inst 0x4e91a428  // smmla v8.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xa0]\n"
+      "ldr q17, [x9, #0xa0]\n"
       ".inst 0x4e90a42c  // smmla v12.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xb0]\n"
+      "ldr q16, [x9, #0xb0]\n"
       ".inst 0x4e91a429  // smmla v9.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xc0]\n"
+      "ldr q17, [x9, #0xc0]\n"
       ".inst 0x4e90a42d  // smmla v13.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xd0]\n"
+      "ldr q16, [x9, #0xd0]\n"
       ".inst 0x4e91a42a  // smmla v10.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xe0]\n"
+      "ldr q17, [x9, #0xe0]\n"
       ".inst 0x4e90a42e  // smmla v14.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q16, [x9, #0xf0]\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e91a42b  // smmla v11.4s, v1.16b, v17.16b\n"
-      "ldr q7, [x10, #0x0]\n"
       ".inst 0x4e90a42f  // smmla v15.4s, v1.16b, v16.16b\n"
       "ldr q1, [x26, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
+      "add x9, x9, #0x100\n"
+      "ldr q7, [x9, #0x0]\n"
+      "ldr q6, [x9, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 7b\n"
       "8:"  // Height 1: Multiply loop: Single iteration only
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "trn1 v18.2d, v1.2d, v16.2d\n"
-      "trn2 v1.2d, v1.2d, v16.2d\n"
+      "trn1 v18.2d, v1.2d, v19.2d\n"
       ".inst 0x4e87a648  // smmla v8.4s, v18.16b, v7.16b\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4e86a64c  // smmla v12.4s, v18.16b, v6.16b\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4e91a649  // smmla v9.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
       ".inst 0x4e90a64d  // smmla v13.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4e91a64a  // smmla v10.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
       ".inst 0x4e90a64e  // smmla v14.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x70]\n"
+      "ldr q16, [x9, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v19.2d\n"
       ".inst 0x4e91a64b  // smmla v11.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x80]\n"
+      "ldr q17, [x9, #0x80]\n"
       ".inst 0x4e90a64f  // smmla v15.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x90]\n"
+      "ldr q16, [x9, #0x90]\n"
       ".inst 0x4e91a428  // smmla v8.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xa0]\n"
+      "ldr q17, [x9, #0xa0]\n"
       ".inst 0x4e90a42c  // smmla v12.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xb0]\n"
+      "ldr q16, [x9, #0xb0]\n"
       ".inst 0x4e91a429  // smmla v9.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xc0]\n"
+      "ldr q17, [x9, #0xc0]\n"
       ".inst 0x4e90a42d  // smmla v13.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xd0]\n"
+      "ldr q16, [x9, #0xd0]\n"
       ".inst 0x4e91a42a  // smmla v10.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xe0]\n"
+      "ldr q17, [x9, #0xe0]\n"
       ".inst 0x4e90a42e  // smmla v14.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q16, [x9, #0xf0]\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e91a42b  // smmla v11.4s, v1.16b, v17.16b\n"
       ".inst 0x4e90a42f  // smmla v15.4s, v1.16b, v16.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x9, x9, #0x100\n"
       "9:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 16f\n"
       "cmp x27, #0x8\n"
       "blt 11f\n"
       "10:"  // Height 1: Multiply loop: Odd block loop
       "ldr d18, [x26], #0x8\n"
-      "ldr q19, [x10, #0x0]\n"
-      "sub x27, x27, #0x8\n"
-      "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x8\n"
-      "trn1 v18.2d, v18.2d, v17.2d\n"
-      ".inst 0x4e93a648  // smmla v8.4s, v18.16b, v19.16b\n"
-      "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4e90a64c  // smmla v12.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q17, [x9, #0x0]\n"
+      "trn1 v18.2d, v18.2d, v16.2d\n"
+      "ldr q31, [x9, #0x10]\n"
+      ".inst 0x4e91a648  // smmla v8.4s, v18.16b, v17.16b\n"
+      "ldr q17, [x9, #0x20]\n"
+      ".inst 0x4e9fa64c  // smmla v12.4s, v18.16b, v31.16b\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4e91a649  // smmla v9.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
       ".inst 0x4e90a64d  // smmla v13.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4e91a64a  // smmla v10.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
       ".inst 0x4e90a64e  // smmla v14.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q16, [x9, #0x70]\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
       ".inst 0x4e91a64b  // smmla v11.4s, v18.16b, v17.16b\n"
       ".inst 0x4e90a64f  // smmla v15.4s, v18.16b, v16.16b\n"
+      "add x9, x9, #0x80\n"
       "bge 10b\n"
       "11:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 16f\n"
@@ -261,24 +261,24 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "14:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr b1, [x26, #0x0]\n"
       "15:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q19, [x10, #0x10]\n"
+      "ldr q17, [x9, #0x0]\n"
+      "ldr q19, [x9, #0x10]\n"
       "trn1 v18.2d, v1.2d, v16.2d\n"
       ".inst 0x4e91a648  // smmla v8.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4e93a64c  // smmla v12.4s, v18.16b, v19.16b\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4e91a649  // smmla v9.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
       ".inst 0x4e90a64d  // smmla v13.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4e91a64a  // smmla v10.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
       ".inst 0x4e90a64e  // smmla v14.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q16, [x9, #0x70]\n"
       ".inst 0x4e91a64b  // smmla v11.4s, v18.16b, v17.16b\n"
       ".inst 0x4e90a64f  // smmla v15.4s, v18.16b, v16.16b\n"
+      "add x9, x9, #0x80\n"
       "16:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -292,11 +292,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr q16, [x14, #0x30]\n"
       "uzp1 v10.2d, v10.2d, v14.2d\n"
       "uzp1 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x14, x14, #0x40\n"
       "mov v15.16b, v8.16b\n"
-      "add v9.4s, v9.4s, v18.4s\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "add v15.4s, v15.4s, v19.4s\n"
+      "add x14, x14, #0x40\n"
+      "add v9.4s, v9.4s, v18.4s\n"
       "add v10.4s, v10.4s, v17.4s\n"
       "add v11.4s, v11.4s, v16.4s\n"
       "tbz %x[flags], #4, 17f\n"
@@ -312,9 +312,9 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "add x13, x13, #0x40\n"
       "b 18f\n"
       "17:"  // Height 1: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -328,34 +328,34 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sqrdmulh v10.4s, v10.4s, v6.4s\n"
       "sqrdmulh v11.4s, v11.4s, v7.4s\n"
       "tbz %x[flags], #5, 19f\n"
-      "and v18.16b, v15.16b, v0.16b\n"
-      "and v17.16b, v9.16b, v1.16b\n"
-      "and v26.16b, v10.16b, v2.16b\n"
-      "and v16.16b, v11.16b, v3.16b\n"
-      "sshr v18.4s, v18.4s, #0x1f\n"
+      "and v17.16b, v15.16b, v0.16b\n"
+      "and v16.16b, v9.16b, v1.16b\n"
+      "and v25.16b, v10.16b, v2.16b\n"
+      "and v18.16b, v11.16b, v3.16b\n"
       "sshr v17.4s, v17.4s, #0x1f\n"
-      "sshr v26.4s, v26.4s, #0x1f\n"
       "sshr v16.4s, v16.4s, #0x1f\n"
-      "sqadd v15.4s, v15.4s, v18.4s\n"
-      "sqadd v9.4s, v9.4s, v17.4s\n"
-      "sqadd v10.4s, v10.4s, v26.4s\n"
-      "sqadd v11.4s, v11.4s, v16.4s\n"
+      "sshr v25.4s, v25.4s, #0x1f\n"
+      "sshr v18.4s, v18.4s, #0x1f\n"
+      "sqadd v15.4s, v15.4s, v17.4s\n"
+      "sqadd v9.4s, v9.4s, v16.4s\n"
+      "sqadd v10.4s, v10.4s, v25.4s\n"
+      "sqadd v11.4s, v11.4s, v18.4s\n"
       "19:"  // Height 1: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "srshl v15.4s, v15.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v18.4s }, [x21]\n"
-      "ld1r { v17.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
-      "ld1r { v16.4s }, [x20]\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "add v15.4s, v15.4s, v18.4s\n"
       "add v9.4s, v9.4s, v18.4s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1r { v16.4s }, [x20]\n"
       "add v10.4s, v10.4s, v18.4s\n"
       "add v11.4s, v11.4s, v18.4s\n"
+      "cmp x10, #0x10\n"
       "smin v15.4s, v15.4s, v17.4s\n"
       "smin v9.4s, v9.4s, v17.4s\n"
       "smin v10.4s, v10.4s, v17.4s\n"
@@ -368,65 +368,65 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v16.8h, v10.8h, v11.8h\n"
       "uzp1 v15.16b, v15.16b, v16.16b\n"
       "bge 28f\n"
-      "tbz x11, #3, 23f\n"
-      "str d15, [x9], #0x8\n"
-      "tbz x11, #2, 21f\n"
-      "st1 { v15.s }[2], [x9], #0x4\n"
-      "tbz x11, #1, 20f\n"
-      "st1 { v15.h }[6], [x9], #0x2\n"
-      "tbz x11, #0, 27f\n"
-      "st1 { v15.b }[14], [x9]\n"
+      "tbz x10, #3, 23f\n"
+      "str d15, [x11], #0x8\n"
+      "tbz x10, #2, 21f\n"
+      "st1 { v15.s }[2], [x11], #0x4\n"
+      "tbz x10, #1, 20f\n"
+      "st1 { v15.h }[6], [x11], #0x2\n"
+      "tbz x10, #0, 27f\n"
+      "st1 { v15.b }[14], [x11]\n"
       "b 27f\n"
       "20:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 27f\n"
-      "st1 { v15.b }[12], [x9]\n"
+      "tbz x10, #0, 27f\n"
+      "st1 { v15.b }[12], [x11]\n"
       "b 27f\n"
       "21:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 22f\n"
-      "st1 { v15.h }[4], [x9], #0x2\n"
-      "tbz x11, #0, 27f\n"
-      "st1 { v15.b }[10], [x9]\n"
+      "tbz x10, #1, 22f\n"
+      "st1 { v15.h }[4], [x11], #0x2\n"
+      "tbz x10, #0, 27f\n"
+      "st1 { v15.b }[10], [x11]\n"
       "b 27f\n"
       "22:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 27f\n"
-      "st1 { v15.b }[8], [x9]\n"
+      "tbz x10, #0, 27f\n"
+      "st1 { v15.b }[8], [x11]\n"
       "b 27f\n"
       "23:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 25f\n"
-      "str s15, [x9], #0x4\n"
-      "tbz x11, #1, 24f\n"
-      "st1 { v15.h }[2], [x9], #0x2\n"
-      "tbz x11, #0, 27f\n"
-      "st1 { v15.b }[6], [x9]\n"
+      "tbz x10, #2, 25f\n"
+      "str s15, [x11], #0x4\n"
+      "tbz x10, #1, 24f\n"
+      "st1 { v15.h }[2], [x11], #0x2\n"
+      "tbz x10, #0, 27f\n"
+      "st1 { v15.b }[6], [x11]\n"
       "b 27f\n"
       "24:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 27f\n"
-      "st1 { v15.b }[4], [x9]\n"
+      "tbz x10, #0, 27f\n"
+      "st1 { v15.b }[4], [x11]\n"
       "b 27f\n"
       "25:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 26f\n"
-      "str h15, [x9], #0x2\n"
-      "tbz x11, #0, 27f\n"
-      "st1 { v15.b }[2], [x9]\n"
+      "tbz x10, #1, 26f\n"
+      "str h15, [x11], #0x2\n"
+      "tbz x10, #0, 27f\n"
+      "st1 { v15.b }[2], [x11]\n"
       "b 27f\n"
       "26:"  // Height 1: Partial direct writeback: partial_1_0
-      "str b15, [x9, #0x0]\n"
+      "str b15, [x11, #0x0]\n"
       "27:"  // Height 1: Partial direct writeback: Done
       "b 29f\n"
       "28:"  // Height 1: Full writeback
-      "str q15, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
+      "str q15, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
       "29:"  // Height 1: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 2b\n"
       "b 176f\n"
       "30:"  // Height 2
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "31:"  // Height 2: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -440,8 +440,8 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "mov x28, #0x0\n"
       "33:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 34f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -461,120 +461,120 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr q1, [x26, #0x0]\n"
       "ldr q2, [x25, #0x0]\n"
       "cmp x27, #0x20\n"
-      "ldr q7, [x10, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x0]\n"
+      "ldr q6, [x9, #0x10]\n"
       "blt 37f\n"
       "36:"  // Height 2: Multiply loop: Main loop head
       "trn1 v18.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e87a648  // smmla v8.4s, v18.16b, v7.16b\n"
-      "ldr q16, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4e86a64c  // smmla v12.4s, v18.16b, v6.16b\n"
-      "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e90a649  // smmla v9.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x40]\n"
-      ".inst 0x4e91a64d  // smmla v13.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e90a64a  // smmla v10.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x60]\n"
-      ".inst 0x4e91a64e  // smmla v14.4s, v18.16b, v17.16b\n"
-      "ldr q31, [x10, #0x70]\n"
-      ".inst 0x4e90a64b  // smmla v11.4s, v18.16b, v16.16b\n"
-      "ldr q17, [x10, #0x80]\n"
-      ".inst 0x4e9fa64f  // smmla v15.4s, v18.16b, v31.16b\n"
-      "ldr q16, [x10, #0x90]\n"
+      "ldr q16, [x9, #0x30]\n"
+      ".inst 0x4e91a649  // smmla v9.4s, v18.16b, v17.16b\n"
+      "ldr q17, [x9, #0x40]\n"
+      ".inst 0x4e90a64d  // smmla v13.4s, v18.16b, v16.16b\n"
+      "ldr q16, [x9, #0x50]\n"
+      ".inst 0x4e91a64a  // smmla v10.4s, v18.16b, v17.16b\n"
+      "ldr q17, [x9, #0x60]\n"
+      ".inst 0x4e90a64e  // smmla v14.4s, v18.16b, v16.16b\n"
+      "ldr q16, [x9, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e91a64b  // smmla v11.4s, v18.16b, v17.16b\n"
+      "ldr q17, [x9, #0x80]\n"
+      ".inst 0x4e90a64f  // smmla v15.4s, v18.16b, v16.16b\n"
+      "ldr q16, [x9, #0x90]\n"
       ".inst 0x4e91a428  // smmla v8.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xa0]\n"
+      "ldr q17, [x9, #0xa0]\n"
       ".inst 0x4e90a42c  // smmla v12.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xb0]\n"
+      "ldr q16, [x9, #0xb0]\n"
       ".inst 0x4e91a429  // smmla v9.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xc0]\n"
+      "ldr q17, [x9, #0xc0]\n"
       ".inst 0x4e90a42d  // smmla v13.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xd0]\n"
+      "ldr q16, [x9, #0xd0]\n"
       ".inst 0x4e91a42a  // smmla v10.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xe0]\n"
+      "ldr q17, [x9, #0xe0]\n"
       ".inst 0x4e90a42e  // smmla v14.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q16, [x9, #0xf0]\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "ldr q2, [x25, #0x0]\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e91a42b  // smmla v11.4s, v1.16b, v17.16b\n"
-      "ldr q7, [x10, #0x0]\n"
+      "add x9, x9, #0x100\n"
+      "ldr q7, [x9, #0x0]\n"
       ".inst 0x4e90a42f  // smmla v15.4s, v1.16b, v16.16b\n"
       "ldr q1, [x26, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "bge 36b\n"
       "37:"  // Height 2: Multiply loop: Single iteration only
       "trn1 v18.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e87a648  // smmla v8.4s, v18.16b, v7.16b\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4e86a64c  // smmla v12.4s, v18.16b, v6.16b\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4e91a649  // smmla v9.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
       ".inst 0x4e90a64d  // smmla v13.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4e91a64a  // smmla v10.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
       ".inst 0x4e90a64e  // smmla v14.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x70]\n"
+      "ldr q16, [x9, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x4e91a64b  // smmla v11.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x80]\n"
+      "ldr q17, [x9, #0x80]\n"
       ".inst 0x4e90a64f  // smmla v15.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x90]\n"
+      "ldr q16, [x9, #0x90]\n"
       ".inst 0x4e91a428  // smmla v8.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xa0]\n"
+      "ldr q17, [x9, #0xa0]\n"
       ".inst 0x4e90a42c  // smmla v12.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xb0]\n"
+      "ldr q16, [x9, #0xb0]\n"
       ".inst 0x4e91a429  // smmla v9.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xc0]\n"
+      "ldr q17, [x9, #0xc0]\n"
       ".inst 0x4e90a42d  // smmla v13.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xd0]\n"
+      "ldr q16, [x9, #0xd0]\n"
       ".inst 0x4e91a42a  // smmla v10.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xe0]\n"
+      "ldr q17, [x9, #0xe0]\n"
       ".inst 0x4e90a42e  // smmla v14.4s, v1.16b, v16.16b\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q16, [x9, #0xf0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e91a42b  // smmla v11.4s, v1.16b, v17.16b\n"
       ".inst 0x4e90a42f  // smmla v15.4s, v1.16b, v16.16b\n"
+      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x9, x9, #0x100\n"
       "38:"  // Height 2: Multiply loop: Main loop skip
       "cbz x27, 45f\n"
       "cmp x27, #0x8\n"
       "blt 40f\n"
       "39:"  // Height 2: Multiply loop: Odd block loop
-      "ldr d19, [x26], #0x8\n"
-      "ldr d18, [x25], #0x8\n"
+      "ldr d17, [x26], #0x8\n"
+      "ldr d16, [x25], #0x8\n"
+      "trn1 v18.2d, v17.2d, v16.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x8\n"
-      "trn1 v18.2d, v19.2d, v18.2d\n"
+      "ldr q17, [x9, #0x0]\n"
+      "ldr q16, [x9, #0x10]\n"
       ".inst 0x4e91a648  // smmla v8.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x20]\n"
       ".inst 0x4e90a64c  // smmla v12.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q17, [x9, #0x20]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4e91a649  // smmla v9.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x40]\n"
       ".inst 0x4e90a64d  // smmla v13.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q17, [x9, #0x40]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4e91a64a  // smmla v10.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x60]\n"
       ".inst 0x4e90a64e  // smmla v14.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q17, [x9, #0x60]\n"
+      "ldr q16, [x9, #0x70]\n"
+      "cmp x27, #0x8\n"
       ".inst 0x4e91a64b  // smmla v11.4s, v18.16b, v17.16b\n"
       ".inst 0x4e90a64f  // smmla v15.4s, v18.16b, v16.16b\n"
+      "add x9, x9, #0x80\n"
       "bge 39b\n"
       "40:"  // Height 2: Multiply loop: Skip odd blocks
       "cbz x27, 45f\n"
@@ -605,24 +605,24 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr b1, [x26, #0x0]\n"
       "ldr b2, [x25, #0x0]\n"
       "44:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q16, [x10, #0x10]\n"
+      "ldr q17, [x9, #0x0]\n"
+      "ldr q16, [x9, #0x10]\n"
       "trn1 v18.2d, v1.2d, v2.2d\n"
       ".inst 0x4e91a648  // smmla v8.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x20]\n"
+      "ldr q17, [x9, #0x20]\n"
       ".inst 0x4e90a64c  // smmla v12.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x30]\n"
+      "ldr q16, [x9, #0x30]\n"
       ".inst 0x4e91a649  // smmla v9.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x40]\n"
+      "ldr q17, [x9, #0x40]\n"
       ".inst 0x4e90a64d  // smmla v13.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x50]\n"
+      "ldr q16, [x9, #0x50]\n"
       ".inst 0x4e91a64a  // smmla v10.4s, v18.16b, v17.16b\n"
-      "ldr q17, [x10, #0x60]\n"
+      "ldr q17, [x9, #0x60]\n"
       ".inst 0x4e90a64e  // smmla v14.4s, v18.16b, v16.16b\n"
-      "ldr q16, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q16, [x9, #0x70]\n"
       ".inst 0x4e91a64b  // smmla v11.4s, v18.16b, v17.16b\n"
       ".inst 0x4e90a64f  // smmla v15.4s, v18.16b, v16.16b\n"
+      "add x9, x9, #0x80\n"
       "45:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -639,18 +639,18 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
+      "add x25, x11, x20\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x14, x14, #0x40\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "mov v15.16b, v17.16b\n"
+      "add v15.4s, v15.4s, v19.4s\n"
+      "add x14, x14, #0x40\n"
       "add v12.4s, v12.4s, v18.4s\n"
-      "add x26, x9, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "add v13.4s, v13.4s, v5.4s\n"
-      "add v8.4s, v8.4s, v19.4s\n"
-      "add v15.4s, v15.4s, v19.4s\n"
       "add v14.4s, v14.4s, v16.4s\n"
+      "add v8.4s, v8.4s, v19.4s\n"
       "add v9.4s, v9.4s, v18.4s\n"
       "add v10.4s, v10.4s, v5.4s\n"
       "add v11.4s, v11.4s, v16.4s\n"
@@ -667,9 +667,9 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "add x13, x13, #0x40\n"
       "b 47f\n"
       "46:"  // Height 2: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -696,11 +696,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v17.4s, v17.4s, #0x1f\n"
       "sshr v16.4s, v16.4s, #0x1f\n"
       "sqadd v15.4s, v15.4s, v19.4s\n"
-      "and v19.16b, v8.16b, v0.16b\n"
       "sqadd v12.4s, v12.4s, v18.4s\n"
-      "and v18.16b, v9.16b, v1.16b\n"
       "sqadd v13.4s, v13.4s, v17.4s\n"
       "sqadd v14.4s, v14.4s, v16.4s\n"
+      "and v19.16b, v8.16b, v0.16b\n"
+      "and v18.16b, v9.16b, v1.16b\n"
       "and v17.16b, v10.16b, v2.16b\n"
       "and v16.16b, v11.16b, v3.16b\n"
       "sshr v19.4s, v19.4s, #0x1f\n"
@@ -712,21 +712,21 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sqadd v10.4s, v10.4s, v17.4s\n"
       "sqadd v11.4s, v11.4s, v16.4s\n"
       "48:"  // Height 2: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v18.4s }, [x20]\n"
       "srshl v15.4s, v15.4s, v0.4s\n"
       "srshl v12.4s, v12.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v18.4s }, [x21]\n"
-      "ld1r { v17.4s }, [x20]\n"
       "srshl v13.4s, v13.4s, v2.4s\n"
       "srshl v14.4s, v14.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v17.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v16.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "add v15.4s, v15.4s, v18.4s\n"
       "add v12.4s, v12.4s, v18.4s\n"
       "add v13.4s, v13.4s, v18.4s\n"
@@ -758,81 +758,81 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v15.16b, v15.16b, v17.16b\n"
       "uzp1 v8.16b, v8.16b, v16.16b\n"
       "bge 57f\n"
-      "tbz x11, #3, 52f\n"
-      "str d15, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "tbz x11, #2, 50f\n"
-      "st1 { v15.s }[2], [x9], #0x4\n"
-      "st1 { v8.s }[2], [x26], #0x4\n"
-      "tbz x11, #1, 49f\n"
-      "st1 { v15.h }[6], [x9], #0x2\n"
-      "st1 { v8.h }[6], [x26], #0x2\n"
-      "tbz x11, #0, 56f\n"
-      "st1 { v15.b }[14], [x9]\n"
-      "st1 { v8.b }[14], [x26]\n"
+      "tbz x10, #3, 52f\n"
+      "str d15, [x11], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "tbz x10, #2, 50f\n"
+      "st1 { v15.s }[2], [x11], #0x4\n"
+      "st1 { v8.s }[2], [x25], #0x4\n"
+      "tbz x10, #1, 49f\n"
+      "st1 { v15.h }[6], [x11], #0x2\n"
+      "st1 { v8.h }[6], [x25], #0x2\n"
+      "tbz x10, #0, 56f\n"
+      "st1 { v15.b }[14], [x11]\n"
+      "st1 { v8.b }[14], [x25]\n"
       "b 56f\n"
       "49:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 56f\n"
-      "st1 { v15.b }[12], [x9]\n"
-      "st1 { v8.b }[12], [x26]\n"
+      "tbz x10, #0, 56f\n"
+      "st1 { v15.b }[12], [x11]\n"
+      "st1 { v8.b }[12], [x25]\n"
       "b 56f\n"
       "50:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 51f\n"
-      "st1 { v15.h }[4], [x9], #0x2\n"
-      "st1 { v8.h }[4], [x26], #0x2\n"
-      "tbz x11, #0, 56f\n"
-      "st1 { v15.b }[10], [x9]\n"
-      "st1 { v8.b }[10], [x26]\n"
+      "tbz x10, #1, 51f\n"
+      "st1 { v15.h }[4], [x11], #0x2\n"
+      "st1 { v8.h }[4], [x25], #0x2\n"
+      "tbz x10, #0, 56f\n"
+      "st1 { v15.b }[10], [x11]\n"
+      "st1 { v8.b }[10], [x25]\n"
       "b 56f\n"
       "51:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 56f\n"
-      "st1 { v15.b }[8], [x9]\n"
-      "st1 { v8.b }[8], [x26]\n"
+      "tbz x10, #0, 56f\n"
+      "st1 { v15.b }[8], [x11]\n"
+      "st1 { v8.b }[8], [x25]\n"
       "b 56f\n"
       "52:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 54f\n"
-      "str s15, [x9], #0x4\n"
-      "str s8, [x26], #0x4\n"
-      "tbz x11, #1, 53f\n"
-      "st1 { v15.h }[2], [x9], #0x2\n"
-      "st1 { v8.h }[2], [x26], #0x2\n"
-      "tbz x11, #0, 56f\n"
-      "st1 { v15.b }[6], [x9]\n"
-      "st1 { v8.b }[6], [x26]\n"
+      "tbz x10, #2, 54f\n"
+      "str s15, [x11], #0x4\n"
+      "str s8, [x25], #0x4\n"
+      "tbz x10, #1, 53f\n"
+      "st1 { v15.h }[2], [x11], #0x2\n"
+      "st1 { v8.h }[2], [x25], #0x2\n"
+      "tbz x10, #0, 56f\n"
+      "st1 { v15.b }[6], [x11]\n"
+      "st1 { v8.b }[6], [x25]\n"
       "b 56f\n"
       "53:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 56f\n"
-      "st1 { v15.b }[4], [x9]\n"
-      "st1 { v8.b }[4], [x26]\n"
+      "tbz x10, #0, 56f\n"
+      "st1 { v15.b }[4], [x11]\n"
+      "st1 { v8.b }[4], [x25]\n"
       "b 56f\n"
       "54:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 55f\n"
-      "str h15, [x9], #0x2\n"
-      "str h8, [x26], #0x2\n"
-      "tbz x11, #0, 56f\n"
-      "st1 { v15.b }[2], [x9]\n"
-      "st1 { v8.b }[2], [x26]\n"
+      "tbz x10, #1, 55f\n"
+      "str h15, [x11], #0x2\n"
+      "str h8, [x25], #0x2\n"
+      "tbz x10, #0, 56f\n"
+      "st1 { v15.b }[2], [x11]\n"
+      "st1 { v8.b }[2], [x25]\n"
       "b 56f\n"
       "55:"  // Height 2: Partial direct writeback: partial_1_0
-      "str b15, [x9, #0x0]\n"
-      "str b8, [x26, #0x0]\n"
+      "str b15, [x11, #0x0]\n"
+      "str b8, [x25, #0x0]\n"
       "56:"  // Height 2: Partial direct writeback: Done
       "b 58f\n"
       "57:"  // Height 2: Full writeback
-      "str q15, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q8, [x26, #0x0]\n"
+      "str q15, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q8, [x25, #0x0]\n"
       "58:"  // Height 2: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 31b\n"
       "b 176f\n"
       "59:"  // Height 3
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "60:"  // Height 3: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -854,8 +854,8 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "mov x28, #0x0\n"
       "62:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 63f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -879,130 +879,130 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr q2, [x25, #0x0]\n"
       "cmp x27, #0x20\n"
       "ldr q3, [x24, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x0]\n"
+      "ldr q6, [x9, #0x10]\n"
       "blt 66f\n"
       "65:"  // Height 3: Multiply loop: Main loop head
       "trn1 v27.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "trn1 v26.2d, v3.2d, v24.2d\n"
       ".inst 0x4e87a768  // smmla v8.4s, v27.16b, v7.16b\n"
-      ".inst 0x4e86a76c  // smmla v12.4s, v27.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn2 v3.2d, v3.2d, v24.2d\n"
+      "trn1 v26.2d, v3.2d, v28.2d\n"
       ".inst 0x4e87a750  // smmla v16.4s, v26.16b, v7.16b\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
+      ".inst 0x4e86a76c  // smmla v12.4s, v27.16b, v6.16b\n"
       ".inst 0x4e86a754  // smmla v20.4s, v26.16b, v6.16b\n"
-      "ldr q24, [x10, #0x30]\n"
+      "ldr q24, [x9, #0x30]\n"
       ".inst 0x4e99a769  // smmla v9.4s, v27.16b, v25.16b\n"
+      "trn2 v3.2d, v3.2d, v28.2d\n"
       ".inst 0x4e99a751  // smmla v17.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       ".inst 0x4e98a76d  // smmla v13.4s, v27.16b, v24.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e98a755  // smmla v21.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4e99a76a  // smmla v10.4s, v27.16b, v25.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4e99a752  // smmla v18.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4e98a76e  // smmla v14.4s, v27.16b, v24.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e98a756  // smmla v22.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x70]\n"
+      "ldr q24, [x9, #0x70]\n"
       ".inst 0x4e99a76b  // smmla v11.4s, v27.16b, v25.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e99a753  // smmla v19.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x80]\n"
+      "ldr q25, [x9, #0x80]\n"
       ".inst 0x4e98a76f  // smmla v15.4s, v27.16b, v24.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e98a757  // smmla v23.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x90]\n"
+      "ldr q24, [x9, #0x90]\n"
       "ldr q2, [x25, #0x0]\n"
       ".inst 0x4e99a428  // smmla v8.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a470  // smmla v16.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xa0]\n"
+      "ldr q25, [x9, #0xa0]\n"
       ".inst 0x4e98a42c  // smmla v12.4s, v1.16b, v24.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e98a474  // smmla v20.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xb0]\n"
+      "ldr q24, [x9, #0xb0]\n"
       ".inst 0x4e99a429  // smmla v9.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e99a471  // smmla v17.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xc0]\n"
+      "ldr q25, [x9, #0xc0]\n"
       ".inst 0x4e98a42d  // smmla v13.4s, v1.16b, v24.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e98a475  // smmla v21.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xd0]\n"
+      "ldr q24, [x9, #0xd0]\n"
       ".inst 0x4e99a42a  // smmla v10.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a472  // smmla v18.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xe0]\n"
+      "ldr q25, [x9, #0xe0]\n"
       ".inst 0x4e98a42e  // smmla v14.4s, v1.16b, v24.16b\n"
       ".inst 0x4e98a476  // smmla v22.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q24, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4e99a42b  // smmla v11.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a473  // smmla v19.4s, v3.16b, v25.16b\n"
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       ".inst 0x4e98a42f  // smmla v15.4s, v1.16b, v24.16b\n"
       "ldr q1, [x26, #0x0]\n"
       ".inst 0x4e98a477  // smmla v23.4s, v3.16b, v24.16b\n"
       "ldr q3, [x24, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x10]\n"
       "bge 65b\n"
       "66:"  // Height 3: Multiply loop: Single iteration only
       "trn1 v27.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn1 v26.2d, v3.2d, v24.2d\n"
       ".inst 0x4e87a768  // smmla v8.4s, v27.16b, v7.16b\n"
-      ".inst 0x4e86a76c  // smmla v12.4s, v27.16b, v6.16b\n"
-      "trn2 v3.2d, v3.2d, v24.2d\n"
+      "trn1 v26.2d, v3.2d, v25.2d\n"
       ".inst 0x4e87a750  // smmla v16.4s, v26.16b, v7.16b\n"
-      "ldr q24, [x10, #0x20]\n"
+      "ldr q24, [x9, #0x20]\n"
+      ".inst 0x4e86a76c  // smmla v12.4s, v27.16b, v6.16b\n"
       ".inst 0x4e86a754  // smmla v20.4s, v26.16b, v6.16b\n"
-      "ldr q4, [x10, #0x30]\n"
+      "ldr q0, [x9, #0x30]\n"
       ".inst 0x4e98a769  // smmla v9.4s, v27.16b, v24.16b\n"
+      "trn2 v3.2d, v3.2d, v25.2d\n"
       ".inst 0x4e98a751  // smmla v17.4s, v26.16b, v24.16b\n"
-      "ldr q25, [x10, #0x40]\n"
-      ".inst 0x4e84a76d  // smmla v13.4s, v27.16b, v4.16b\n"
-      ".inst 0x4e84a755  // smmla v21.4s, v26.16b, v4.16b\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q25, [x9, #0x40]\n"
+      ".inst 0x4e80a76d  // smmla v13.4s, v27.16b, v0.16b\n"
+      "add x26, x26, #0x10\n"
+      ".inst 0x4e80a755  // smmla v21.4s, v26.16b, v0.16b\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4e99a76a  // smmla v10.4s, v27.16b, v25.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e99a752  // smmla v18.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4e98a76e  // smmla v14.4s, v27.16b, v24.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e98a756  // smmla v22.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x70]\n"
+      "ldr q24, [x9, #0x70]\n"
       ".inst 0x4e99a76b  // smmla v11.4s, v27.16b, v25.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e99a753  // smmla v19.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x80]\n"
+      "ldr q25, [x9, #0x80]\n"
       ".inst 0x4e98a76f  // smmla v15.4s, v27.16b, v24.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e98a757  // smmla v23.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x90]\n"
+      "ldr q24, [x9, #0x90]\n"
       ".inst 0x4e99a428  // smmla v8.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e99a470  // smmla v16.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xa0]\n"
+      "ldr q25, [x9, #0xa0]\n"
       ".inst 0x4e98a42c  // smmla v12.4s, v1.16b, v24.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e98a474  // smmla v20.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xb0]\n"
+      "ldr q24, [x9, #0xb0]\n"
       ".inst 0x4e99a429  // smmla v9.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a471  // smmla v17.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xc0]\n"
+      "ldr q25, [x9, #0xc0]\n"
       ".inst 0x4e98a42d  // smmla v13.4s, v1.16b, v24.16b\n"
       ".inst 0x4e98a475  // smmla v21.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xd0]\n"
+      "ldr q24, [x9, #0xd0]\n"
       ".inst 0x4e99a42a  // smmla v10.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a472  // smmla v18.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xe0]\n"
+      "ldr q25, [x9, #0xe0]\n"
       ".inst 0x4e98a42e  // smmla v14.4s, v1.16b, v24.16b\n"
       ".inst 0x4e98a476  // smmla v22.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q24, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4e99a42b  // smmla v11.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a473  // smmla v19.4s, v3.16b, v25.16b\n"
       ".inst 0x4e98a42f  // smmla v15.4s, v1.16b, v24.16b\n"
@@ -1012,35 +1012,35 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 69f\n"
       "68:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d29, [x26], #0x8\n"
-      "ldr d27, [x25], #0x8\n"
-      "sub x27, x27, #0x8\n"
-      "ldr d26, [x24], #0x8\n"
-      "ldr q25, [x10, #0x0]\n"
-      "cmp x27, #0x8\n"
-      "ldr q24, [x10, #0x10]\n"
-      "trn1 v27.2d, v29.2d, v27.2d\n"
-      "trn1 v26.2d, v26.2d, v28.2d\n"
+      "ldr d25, [x26], #0x8\n"
+      "ldr d24, [x25], #0x8\n"
+      "trn1 v27.2d, v25.2d, v24.2d\n"
+      "ldr d24, [x24], #0x8\n"
+      "ldr q25, [x9, #0x0]\n"
+      "trn1 v26.2d, v24.2d, v26.2d\n"
       ".inst 0x4e99a768  // smmla v8.4s, v27.16b, v25.16b\n"
+      "ldr q24, [x9, #0x10]\n"
       ".inst 0x4e99a750  // smmla v16.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
       ".inst 0x4e98a76c  // smmla v12.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a754  // smmla v20.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x30]\n"
+      "ldr q24, [x9, #0x30]\n"
       ".inst 0x4e99a769  // smmla v9.4s, v27.16b, v25.16b\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x4e99a751  // smmla v17.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       ".inst 0x4e98a76d  // smmla v13.4s, v27.16b, v24.16b\n"
+      "cmp x27, #0x8\n"
       ".inst 0x4e98a755  // smmla v21.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4e99a76a  // smmla v10.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a752  // smmla v18.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4e98a76e  // smmla v14.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a756  // smmla v22.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q24, [x9, #0x70]\n"
       ".inst 0x4e99a76b  // smmla v11.4s, v27.16b, v25.16b\n"
+      "add x9, x9, #0x80\n"
       ".inst 0x4e99a753  // smmla v19.4s, v26.16b, v25.16b\n"
       ".inst 0x4e98a76f  // smmla v15.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a757  // smmla v23.4s, v26.16b, v24.16b\n"
@@ -1081,29 +1081,29 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr b2, [x25, #0x0]\n"
       "ldr b3, [x24, #0x0]\n"
       "73:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "ldr q25, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
+      "ldr q25, [x9, #0x0]\n"
+      "ldr q28, [x9, #0x10]\n"
       "trn1 v27.2d, v1.2d, v2.2d\n"
       "trn1 v26.2d, v3.2d, v24.2d\n"
       ".inst 0x4e99a768  // smmla v8.4s, v27.16b, v25.16b\n"
-      ".inst 0x4e9ca76c  // smmla v12.4s, v27.16b, v28.16b\n"
       ".inst 0x4e99a750  // smmla v16.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
+      ".inst 0x4e9ca76c  // smmla v12.4s, v27.16b, v28.16b\n"
       ".inst 0x4e9ca754  // smmla v20.4s, v26.16b, v28.16b\n"
-      "ldr q24, [x10, #0x30]\n"
+      "ldr q24, [x9, #0x30]\n"
       ".inst 0x4e99a769  // smmla v9.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a751  // smmla v17.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       ".inst 0x4e98a76d  // smmla v13.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a755  // smmla v21.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4e99a76a  // smmla v10.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a752  // smmla v18.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4e98a76e  // smmla v14.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a756  // smmla v22.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q24, [x9, #0x70]\n"
+      "add x9, x9, #0x80\n"
       ".inst 0x4e99a76b  // smmla v11.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a753  // smmla v19.4s, v26.16b, v25.16b\n"
       ".inst 0x4e98a76f  // smmla v15.4s, v27.16b, v24.16b\n"
@@ -1124,23 +1124,23 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
+      "add x25, x11, x20\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x14, x14, #0x40\n"
+      "add x24, x25, x20\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "uzp1 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v17.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v18.2d, v18.2d, v22.2d\n"
       "uzp1 v19.2d, v19.2d, v23.2d\n"
-      "add x26, x9, x20\n"
+      "add x14, x14, #0x40\n"
       "mov v23.16b, v26.16b\n"
-      "add x25, x26, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "add v23.4s, v23.4s, v28.4s\n"
       "add v12.4s, v12.4s, v27.4s\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "add v13.4s, v13.4s, v25.4s\n"
       "add v14.4s, v14.4s, v24.4s\n"
-      "add v23.4s, v23.4s, v28.4s\n"
       "add v8.4s, v8.4s, v28.4s\n"
       "add v9.4s, v9.4s, v27.4s\n"
       "add v10.4s, v10.4s, v25.4s\n"
@@ -1162,9 +1162,9 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "add x13, x13, #0x40\n"
       "b 76f\n"
       "75:"  // Height 3: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -1195,11 +1195,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v21.4s, v21.4s, #0x1f\n"
       "sshr v20.4s, v20.4s, #0x1f\n"
       "sqadd v23.4s, v23.4s, v24.4s\n"
-      "and v24.16b, v8.16b, v0.16b\n"
       "sqadd v12.4s, v12.4s, v22.4s\n"
-      "and v22.16b, v9.16b, v1.16b\n"
       "sqadd v13.4s, v13.4s, v21.4s\n"
       "sqadd v14.4s, v14.4s, v20.4s\n"
+      "and v24.16b, v8.16b, v0.16b\n"
+      "and v22.16b, v9.16b, v1.16b\n"
       "and v21.16b, v10.16b, v2.16b\n"
       "and v20.16b, v11.16b, v3.16b\n"
       "sshr v24.4s, v24.4s, #0x1f\n"
@@ -1207,11 +1207,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v21.4s, v21.4s, #0x1f\n"
       "sshr v20.4s, v20.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v24.4s\n"
-      "and v24.16b, v16.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v22.4s\n"
-      "and v22.16b, v17.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v21.4s\n"
       "sqadd v11.4s, v11.4s, v20.4s\n"
+      "and v24.16b, v16.16b, v0.16b\n"
+      "and v22.16b, v17.16b, v1.16b\n"
       "and v21.16b, v18.16b, v2.16b\n"
       "and v20.16b, v19.16b, v3.16b\n"
       "sshr v24.4s, v24.4s, #0x1f\n"
@@ -1223,21 +1223,21 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sqadd v18.4s, v18.4s, v21.4s\n"
       "sqadd v19.4s, v19.4s, v20.4s\n"
       "77:"  // Height 3: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v22.4s }, [x20]\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "srshl v12.4s, v12.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v22.4s }, [x21]\n"
-      "ld1r { v21.4s }, [x20]\n"
       "srshl v13.4s, v13.4s, v2.4s\n"
       "srshl v14.4s, v14.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v20.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v1.4s\n"
       "srshl v18.4s, v18.4s, v2.4s\n"
@@ -1288,97 +1288,97 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v8.16b, v8.16b, v20.16b\n"
       "uzp1 v16.16b, v16.16b, v17.16b\n"
       "bge 86f\n"
-      "tbz x11, #3, 81f\n"
-      "str d23, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d16, [x25], #0x8\n"
-      "tbz x11, #2, 79f\n"
-      "st1 { v23.s }[2], [x9], #0x4\n"
-      "st1 { v8.s }[2], [x26], #0x4\n"
-      "st1 { v16.s }[2], [x25], #0x4\n"
-      "tbz x11, #1, 78f\n"
-      "st1 { v23.h }[6], [x9], #0x2\n"
-      "st1 { v8.h }[6], [x26], #0x2\n"
-      "st1 { v16.h }[6], [x25], #0x2\n"
-      "tbz x11, #0, 85f\n"
-      "st1 { v23.b }[14], [x9]\n"
-      "st1 { v8.b }[14], [x26]\n"
-      "st1 { v16.b }[14], [x25]\n"
+      "tbz x10, #3, 81f\n"
+      "str d23, [x11], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d16, [x24], #0x8\n"
+      "tbz x10, #2, 79f\n"
+      "st1 { v23.s }[2], [x11], #0x4\n"
+      "st1 { v8.s }[2], [x25], #0x4\n"
+      "st1 { v16.s }[2], [x24], #0x4\n"
+      "tbz x10, #1, 78f\n"
+      "st1 { v23.h }[6], [x11], #0x2\n"
+      "st1 { v8.h }[6], [x25], #0x2\n"
+      "st1 { v16.h }[6], [x24], #0x2\n"
+      "tbz x10, #0, 85f\n"
+      "st1 { v23.b }[14], [x11]\n"
+      "st1 { v8.b }[14], [x25]\n"
+      "st1 { v16.b }[14], [x24]\n"
       "b 85f\n"
       "78:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 85f\n"
-      "st1 { v23.b }[12], [x9]\n"
-      "st1 { v8.b }[12], [x26]\n"
-      "st1 { v16.b }[12], [x25]\n"
+      "tbz x10, #0, 85f\n"
+      "st1 { v23.b }[12], [x11]\n"
+      "st1 { v8.b }[12], [x25]\n"
+      "st1 { v16.b }[12], [x24]\n"
       "b 85f\n"
       "79:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 80f\n"
-      "st1 { v23.h }[4], [x9], #0x2\n"
-      "st1 { v8.h }[4], [x26], #0x2\n"
-      "st1 { v16.h }[4], [x25], #0x2\n"
-      "tbz x11, #0, 85f\n"
-      "st1 { v23.b }[10], [x9]\n"
-      "st1 { v8.b }[10], [x26]\n"
-      "st1 { v16.b }[10], [x25]\n"
+      "tbz x10, #1, 80f\n"
+      "st1 { v23.h }[4], [x11], #0x2\n"
+      "st1 { v8.h }[4], [x25], #0x2\n"
+      "st1 { v16.h }[4], [x24], #0x2\n"
+      "tbz x10, #0, 85f\n"
+      "st1 { v23.b }[10], [x11]\n"
+      "st1 { v8.b }[10], [x25]\n"
+      "st1 { v16.b }[10], [x24]\n"
       "b 85f\n"
       "80:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 85f\n"
-      "st1 { v23.b }[8], [x9]\n"
-      "st1 { v8.b }[8], [x26]\n"
-      "st1 { v16.b }[8], [x25]\n"
+      "tbz x10, #0, 85f\n"
+      "st1 { v23.b }[8], [x11]\n"
+      "st1 { v8.b }[8], [x25]\n"
+      "st1 { v16.b }[8], [x24]\n"
       "b 85f\n"
       "81:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 83f\n"
-      "str s23, [x9], #0x4\n"
-      "str s8, [x26], #0x4\n"
-      "str s16, [x25], #0x4\n"
-      "tbz x11, #1, 82f\n"
-      "st1 { v23.h }[2], [x9], #0x2\n"
-      "st1 { v8.h }[2], [x26], #0x2\n"
-      "st1 { v16.h }[2], [x25], #0x2\n"
-      "tbz x11, #0, 85f\n"
-      "st1 { v23.b }[6], [x9]\n"
-      "st1 { v8.b }[6], [x26]\n"
-      "st1 { v16.b }[6], [x25]\n"
+      "tbz x10, #2, 83f\n"
+      "str s23, [x11], #0x4\n"
+      "str s8, [x25], #0x4\n"
+      "str s16, [x24], #0x4\n"
+      "tbz x10, #1, 82f\n"
+      "st1 { v23.h }[2], [x11], #0x2\n"
+      "st1 { v8.h }[2], [x25], #0x2\n"
+      "st1 { v16.h }[2], [x24], #0x2\n"
+      "tbz x10, #0, 85f\n"
+      "st1 { v23.b }[6], [x11]\n"
+      "st1 { v8.b }[6], [x25]\n"
+      "st1 { v16.b }[6], [x24]\n"
       "b 85f\n"
       "82:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 85f\n"
-      "st1 { v23.b }[4], [x9]\n"
-      "st1 { v8.b }[4], [x26]\n"
-      "st1 { v16.b }[4], [x25]\n"
+      "tbz x10, #0, 85f\n"
+      "st1 { v23.b }[4], [x11]\n"
+      "st1 { v8.b }[4], [x25]\n"
+      "st1 { v16.b }[4], [x24]\n"
       "b 85f\n"
       "83:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 84f\n"
-      "str h23, [x9], #0x2\n"
-      "str h8, [x26], #0x2\n"
-      "str h16, [x25], #0x2\n"
-      "tbz x11, #0, 85f\n"
-      "st1 { v23.b }[2], [x9]\n"
-      "st1 { v8.b }[2], [x26]\n"
-      "st1 { v16.b }[2], [x25]\n"
+      "tbz x10, #1, 84f\n"
+      "str h23, [x11], #0x2\n"
+      "str h8, [x25], #0x2\n"
+      "str h16, [x24], #0x2\n"
+      "tbz x10, #0, 85f\n"
+      "st1 { v23.b }[2], [x11]\n"
+      "st1 { v8.b }[2], [x25]\n"
+      "st1 { v16.b }[2], [x24]\n"
       "b 85f\n"
       "84:"  // Height 3: Partial direct writeback: partial_1_0
-      "str b23, [x9, #0x0]\n"
-      "str b8, [x26, #0x0]\n"
-      "str b16, [x25, #0x0]\n"
+      "str b23, [x11, #0x0]\n"
+      "str b8, [x25, #0x0]\n"
+      "str b16, [x24, #0x0]\n"
       "85:"  // Height 3: Partial direct writeback: Done
       "b 87f\n"
       "86:"  // Height 3: Full writeback
-      "str q23, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q8, [x26, #0x0]\n"
-      "str q16, [x25, #0x0]\n"
+      "str q23, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q8, [x25, #0x0]\n"
+      "str q16, [x24, #0x0]\n"
       "87:"  // Height 3: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 60b\n"
       "b 176f\n"
       "88:"  // Height 4
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "89:"  // Height 4: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -1400,8 +1400,8 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "mov x28, #0x0\n"
       "91:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 92f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1429,135 +1429,135 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "cmp x27, #0x20\n"
       "ldr q3, [x24, #0x0]\n"
       "ldr q4, [x23, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
+      "ldr q7, [x9, #0x0]\n"
+      "ldr q6, [x9, #0x10]\n"
       "blt 95f\n"
       "94:"  // Height 4: Multiply loop: Main loop head
       "trn1 v27.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e87a768  // smmla v8.4s, v27.16b, v7.16b\n"
       "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
       "trn1 v26.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "ldr q4, [x23, #0x0]\n"
-      ".inst 0x4e87a768  // smmla v8.4s, v27.16b, v7.16b\n"
-      ".inst 0x4e86a76c  // smmla v12.4s, v27.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e87a750  // smmla v16.4s, v26.16b, v7.16b\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
+      ".inst 0x4e86a76c  // smmla v12.4s, v27.16b, v6.16b\n"
       ".inst 0x4e86a754  // smmla v20.4s, v26.16b, v6.16b\n"
-      "ldr q24, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "ldr q24, [x9, #0x30]\n"
       ".inst 0x4e99a769  // smmla v9.4s, v27.16b, v25.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x4e99a751  // smmla v17.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       ".inst 0x4e98a76d  // smmla v13.4s, v27.16b, v24.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4e98a755  // smmla v21.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4e99a76a  // smmla v10.4s, v27.16b, v25.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e99a752  // smmla v18.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4e98a76e  // smmla v14.4s, v27.16b, v24.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e98a756  // smmla v22.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x70]\n"
+      "ldr q24, [x9, #0x70]\n"
       ".inst 0x4e99a76b  // smmla v11.4s, v27.16b, v25.16b\n"
+      "add x23, x23, #0x10\n"
+      "ldr q4, [x23, #0x0]\n"
       ".inst 0x4e99a753  // smmla v19.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x80]\n"
+      "ldr q25, [x9, #0x80]\n"
       ".inst 0x4e98a76f  // smmla v15.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a757  // smmla v23.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x90]\n"
+      "ldr q24, [x9, #0x90]\n"
       "ldr q2, [x25, #0x0]\n"
       ".inst 0x4e99a428  // smmla v8.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a470  // smmla v16.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xa0]\n"
+      "ldr q25, [x9, #0xa0]\n"
       ".inst 0x4e98a42c  // smmla v12.4s, v1.16b, v24.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e98a474  // smmla v20.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xb0]\n"
+      "ldr q24, [x9, #0xb0]\n"
       ".inst 0x4e99a429  // smmla v9.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e99a471  // smmla v17.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xc0]\n"
+      "ldr q25, [x9, #0xc0]\n"
       ".inst 0x4e98a42d  // smmla v13.4s, v1.16b, v24.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e98a475  // smmla v21.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xd0]\n"
+      "ldr q24, [x9, #0xd0]\n"
       ".inst 0x4e99a42a  // smmla v10.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e99a472  // smmla v18.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xe0]\n"
+      "ldr q25, [x9, #0xe0]\n"
       ".inst 0x4e98a42e  // smmla v14.4s, v1.16b, v24.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e98a476  // smmla v22.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q24, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4e99a42b  // smmla v11.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a473  // smmla v19.4s, v3.16b, v25.16b\n"
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       ".inst 0x4e98a42f  // smmla v15.4s, v1.16b, v24.16b\n"
       "ldr q1, [x26, #0x0]\n"
       ".inst 0x4e98a477  // smmla v23.4s, v3.16b, v24.16b\n"
       "ldr q3, [x24, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
+      "ldr q6, [x9, #0x10]\n"
       "bge 94b\n"
       "95:"  // Height 4: Multiply loop: Single iteration only
       "trn1 v27.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e87a768  // smmla v8.4s, v27.16b, v7.16b\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v26.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      ".inst 0x4e87a768  // smmla v8.4s, v27.16b, v7.16b\n"
-      ".inst 0x4e86a76c  // smmla v12.4s, v27.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e87a750  // smmla v16.4s, v26.16b, v7.16b\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
+      ".inst 0x4e86a76c  // smmla v12.4s, v27.16b, v6.16b\n"
       ".inst 0x4e86a754  // smmla v20.4s, v26.16b, v6.16b\n"
-      "ldr q24, [x10, #0x30]\n"
+      "ldr q24, [x9, #0x30]\n"
       ".inst 0x4e99a769  // smmla v9.4s, v27.16b, v25.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x4e99a751  // smmla v17.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       ".inst 0x4e98a76d  // smmla v13.4s, v27.16b, v24.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e98a755  // smmla v21.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4e99a76a  // smmla v10.4s, v27.16b, v25.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e99a752  // smmla v18.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4e98a76e  // smmla v14.4s, v27.16b, v24.16b\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e98a756  // smmla v22.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x70]\n"
+      "ldr q24, [x9, #0x70]\n"
       ".inst 0x4e99a76b  // smmla v11.4s, v27.16b, v25.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e99a753  // smmla v19.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x80]\n"
+      "ldr q25, [x9, #0x80]\n"
       ".inst 0x4e98a76f  // smmla v15.4s, v27.16b, v24.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e98a757  // smmla v23.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x90]\n"
+      "ldr q24, [x9, #0x90]\n"
       ".inst 0x4e99a428  // smmla v8.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e99a470  // smmla v16.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xa0]\n"
+      "ldr q25, [x9, #0xa0]\n"
       ".inst 0x4e98a42c  // smmla v12.4s, v1.16b, v24.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e98a474  // smmla v20.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xb0]\n"
+      "ldr q24, [x9, #0xb0]\n"
       ".inst 0x4e99a429  // smmla v9.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e99a471  // smmla v17.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xc0]\n"
+      "ldr q25, [x9, #0xc0]\n"
       ".inst 0x4e98a42d  // smmla v13.4s, v1.16b, v24.16b\n"
       ".inst 0x4e98a475  // smmla v21.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xd0]\n"
+      "ldr q24, [x9, #0xd0]\n"
       ".inst 0x4e99a42a  // smmla v10.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a472  // smmla v18.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xe0]\n"
+      "ldr q25, [x9, #0xe0]\n"
       ".inst 0x4e98a42e  // smmla v14.4s, v1.16b, v24.16b\n"
       ".inst 0x4e98a476  // smmla v22.4s, v3.16b, v24.16b\n"
-      "ldr q24, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q24, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4e99a42b  // smmla v11.4s, v1.16b, v25.16b\n"
       ".inst 0x4e99a473  // smmla v19.4s, v3.16b, v25.16b\n"
       ".inst 0x4e98a42f  // smmla v15.4s, v1.16b, v24.16b\n"
@@ -1567,35 +1567,35 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 98f\n"
       "97:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d29, [x26], #0x8\n"
-      "ldr d27, [x25], #0x8\n"
+      "ldr d25, [x26], #0x8\n"
+      "ldr d24, [x25], #0x8\n"
+      "trn1 v27.2d, v25.2d, v24.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d28, [x24], #0x8\n"
-      "ldr d26, [x23], #0x8\n"
+      "ldr d25, [x24], #0x8\n"
+      "ldr d24, [x23], #0x8\n"
+      "trn1 v26.2d, v25.2d, v24.2d\n"
       "cmp x27, #0x8\n"
-      "ldr q25, [x10, #0x0]\n"
-      "ldr q24, [x10, #0x10]\n"
-      "trn1 v27.2d, v29.2d, v27.2d\n"
-      "trn1 v26.2d, v28.2d, v26.2d\n"
+      "ldr q25, [x9, #0x0]\n"
+      "ldr q24, [x9, #0x10]\n"
       ".inst 0x4e99a768  // smmla v8.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a750  // smmla v16.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
       ".inst 0x4e98a76c  // smmla v12.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a754  // smmla v20.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x30]\n"
+      "ldr q24, [x9, #0x30]\n"
       ".inst 0x4e99a769  // smmla v9.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a751  // smmla v17.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       ".inst 0x4e98a76d  // smmla v13.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a755  // smmla v21.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4e99a76a  // smmla v10.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a752  // smmla v18.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4e98a76e  // smmla v14.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a756  // smmla v22.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q24, [x9, #0x70]\n"
+      "add x9, x9, #0x80\n"
       ".inst 0x4e99a76b  // smmla v11.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a753  // smmla v19.4s, v26.16b, v25.16b\n"
       ".inst 0x4e98a76f  // smmla v15.4s, v27.16b, v24.16b\n"
@@ -1644,29 +1644,29 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr b3, [x24, #0x0]\n"
       "ldr b4, [x23, #0x0]\n"
       "102:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "ldr q25, [x10, #0x0]\n"
-      "ldr q24, [x10, #0x10]\n"
+      "ldr q25, [x9, #0x0]\n"
+      "ldr q24, [x9, #0x10]\n"
       "trn1 v27.2d, v1.2d, v2.2d\n"
       "trn1 v26.2d, v3.2d, v4.2d\n"
       ".inst 0x4e99a768  // smmla v8.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a750  // smmla v16.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x20]\n"
+      "ldr q25, [x9, #0x20]\n"
       ".inst 0x4e98a76c  // smmla v12.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a754  // smmla v20.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x30]\n"
+      "ldr q24, [x9, #0x30]\n"
       ".inst 0x4e99a769  // smmla v9.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a751  // smmla v17.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x40]\n"
+      "ldr q25, [x9, #0x40]\n"
       ".inst 0x4e98a76d  // smmla v13.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a755  // smmla v21.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x50]\n"
+      "ldr q24, [x9, #0x50]\n"
       ".inst 0x4e99a76a  // smmla v10.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a752  // smmla v18.4s, v26.16b, v25.16b\n"
-      "ldr q25, [x10, #0x60]\n"
+      "ldr q25, [x9, #0x60]\n"
       ".inst 0x4e98a76e  // smmla v14.4s, v27.16b, v24.16b\n"
       ".inst 0x4e98a756  // smmla v22.4s, v26.16b, v24.16b\n"
-      "ldr q24, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q24, [x9, #0x70]\n"
+      "add x9, x9, #0x80\n"
       ".inst 0x4e99a76b  // smmla v11.4s, v27.16b, v25.16b\n"
       ".inst 0x4e99a753  // smmla v19.4s, v26.16b, v25.16b\n"
       ".inst 0x4e98a76f  // smmla v15.4s, v27.16b, v24.16b\n"
@@ -1687,29 +1687,29 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
+      "add x25, x11, x20\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x14, x14, #0x40\n"
+      "add x24, x25, x20\n"
+      "add x23, x24, x20\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "add x26, x9, x20\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x25, x26, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "add x14, x14, #0x40\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "add x24, x25, x20\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "mov v23.16b, v26.16b\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "add v23.4s, v23.4s, v28.4s\n"
       "add v12.4s, v12.4s, v27.4s\n"
       "add v13.4s, v13.4s, v25.4s\n"
       "add v14.4s, v14.4s, v24.4s\n"
-      "add v23.4s, v23.4s, v28.4s\n"
       "add v8.4s, v8.4s, v28.4s\n"
       "add v9.4s, v9.4s, v27.4s\n"
       "add v10.4s, v10.4s, v25.4s\n"
@@ -1735,9 +1735,9 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "add x13, x13, #0x40\n"
       "b 105f\n"
       "104:"  // Height 4: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -1772,11 +1772,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v25.4s, v25.4s, #0x1f\n"
       "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v23.4s, v23.4s, v27.4s\n"
-      "and v27.16b, v8.16b, v0.16b\n"
       "sqadd v12.4s, v12.4s, v26.4s\n"
-      "and v26.16b, v9.16b, v1.16b\n"
       "sqadd v13.4s, v13.4s, v25.4s\n"
       "sqadd v14.4s, v14.4s, v24.4s\n"
+      "and v27.16b, v8.16b, v0.16b\n"
+      "and v26.16b, v9.16b, v1.16b\n"
       "and v25.16b, v10.16b, v2.16b\n"
       "and v24.16b, v11.16b, v3.16b\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
@@ -1784,11 +1784,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v25.4s, v25.4s, #0x1f\n"
       "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v27.4s\n"
-      "and v27.16b, v15.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v26.4s\n"
-      "and v26.16b, v20.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v25.4s\n"
       "sqadd v11.4s, v11.4s, v24.4s\n"
+      "and v27.16b, v15.16b, v0.16b\n"
+      "and v26.16b, v20.16b, v1.16b\n"
       "and v25.16b, v21.16b, v2.16b\n"
       "and v24.16b, v22.16b, v3.16b\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
@@ -1796,11 +1796,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v25.4s, v25.4s, #0x1f\n"
       "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v15.4s, v15.4s, v27.4s\n"
-      "and v27.16b, v16.16b, v0.16b\n"
       "sqadd v20.4s, v20.4s, v26.4s\n"
-      "and v26.16b, v17.16b, v1.16b\n"
       "sqadd v21.4s, v21.4s, v25.4s\n"
       "sqadd v22.4s, v22.4s, v24.4s\n"
+      "and v27.16b, v16.16b, v0.16b\n"
+      "and v26.16b, v17.16b, v1.16b\n"
       "and v25.16b, v18.16b, v2.16b\n"
       "and v24.16b, v19.16b, v3.16b\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
@@ -1812,21 +1812,21 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sqadd v18.4s, v18.4s, v25.4s\n"
       "sqadd v19.4s, v19.4s, v24.4s\n"
       "106:"  // Height 4: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "srshl v12.4s, v12.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x21]\n"
-      "ld1r { v25.4s }, [x20]\n"
       "srshl v13.4s, v13.4s, v2.4s\n"
       "srshl v14.4s, v14.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "srshl v15.4s, v15.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v1.4s\n"
       "srshl v21.4s, v21.4s, v2.4s\n"
@@ -1896,113 +1896,113 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v15.16b, v15.16b, v20.16b\n"
       "uzp1 v16.16b, v16.16b, v17.16b\n"
       "bge 115f\n"
-      "tbz x11, #3, 110f\n"
-      "str d23, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "tbz x11, #2, 108f\n"
-      "st1 { v23.s }[2], [x9], #0x4\n"
-      "st1 { v8.s }[2], [x26], #0x4\n"
-      "st1 { v15.s }[2], [x25], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "tbz x11, #1, 107f\n"
-      "st1 { v23.h }[6], [x9], #0x2\n"
-      "st1 { v8.h }[6], [x26], #0x2\n"
-      "st1 { v15.h }[6], [x25], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "tbz x11, #0, 114f\n"
-      "st1 { v23.b }[14], [x9]\n"
-      "st1 { v8.b }[14], [x26]\n"
-      "st1 { v15.b }[14], [x25]\n"
-      "st1 { v16.b }[14], [x24]\n"
+      "tbz x10, #3, 110f\n"
+      "str d23, [x11], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "tbz x10, #2, 108f\n"
+      "st1 { v23.s }[2], [x11], #0x4\n"
+      "st1 { v8.s }[2], [x25], #0x4\n"
+      "st1 { v15.s }[2], [x24], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
+      "tbz x10, #1, 107f\n"
+      "st1 { v23.h }[6], [x11], #0x2\n"
+      "st1 { v8.h }[6], [x25], #0x2\n"
+      "st1 { v15.h }[6], [x24], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
+      "tbz x10, #0, 114f\n"
+      "st1 { v23.b }[14], [x11]\n"
+      "st1 { v8.b }[14], [x25]\n"
+      "st1 { v15.b }[14], [x24]\n"
+      "st1 { v16.b }[14], [x23]\n"
       "b 114f\n"
       "107:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 114f\n"
-      "st1 { v23.b }[12], [x9]\n"
-      "st1 { v8.b }[12], [x26]\n"
-      "st1 { v15.b }[12], [x25]\n"
-      "st1 { v16.b }[12], [x24]\n"
+      "tbz x10, #0, 114f\n"
+      "st1 { v23.b }[12], [x11]\n"
+      "st1 { v8.b }[12], [x25]\n"
+      "st1 { v15.b }[12], [x24]\n"
+      "st1 { v16.b }[12], [x23]\n"
       "b 114f\n"
       "108:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 109f\n"
-      "st1 { v23.h }[4], [x9], #0x2\n"
-      "st1 { v8.h }[4], [x26], #0x2\n"
-      "st1 { v15.h }[4], [x25], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "tbz x11, #0, 114f\n"
-      "st1 { v23.b }[10], [x9]\n"
-      "st1 { v8.b }[10], [x26]\n"
-      "st1 { v15.b }[10], [x25]\n"
-      "st1 { v16.b }[10], [x24]\n"
+      "tbz x10, #1, 109f\n"
+      "st1 { v23.h }[4], [x11], #0x2\n"
+      "st1 { v8.h }[4], [x25], #0x2\n"
+      "st1 { v15.h }[4], [x24], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
+      "tbz x10, #0, 114f\n"
+      "st1 { v23.b }[10], [x11]\n"
+      "st1 { v8.b }[10], [x25]\n"
+      "st1 { v15.b }[10], [x24]\n"
+      "st1 { v16.b }[10], [x23]\n"
       "b 114f\n"
       "109:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 114f\n"
-      "st1 { v23.b }[8], [x9]\n"
-      "st1 { v8.b }[8], [x26]\n"
-      "st1 { v15.b }[8], [x25]\n"
-      "st1 { v16.b }[8], [x24]\n"
+      "tbz x10, #0, 114f\n"
+      "st1 { v23.b }[8], [x11]\n"
+      "st1 { v8.b }[8], [x25]\n"
+      "st1 { v15.b }[8], [x24]\n"
+      "st1 { v16.b }[8], [x23]\n"
       "b 114f\n"
       "110:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 112f\n"
-      "str s23, [x9], #0x4\n"
-      "str s8, [x26], #0x4\n"
-      "str s15, [x25], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "tbz x11, #1, 111f\n"
-      "st1 { v23.h }[2], [x9], #0x2\n"
-      "st1 { v8.h }[2], [x26], #0x2\n"
-      "st1 { v15.h }[2], [x25], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "tbz x11, #0, 114f\n"
-      "st1 { v23.b }[6], [x9]\n"
-      "st1 { v8.b }[6], [x26]\n"
-      "st1 { v15.b }[6], [x25]\n"
-      "st1 { v16.b }[6], [x24]\n"
+      "tbz x10, #2, 112f\n"
+      "str s23, [x11], #0x4\n"
+      "str s8, [x25], #0x4\n"
+      "str s15, [x24], #0x4\n"
+      "str s16, [x23], #0x4\n"
+      "tbz x10, #1, 111f\n"
+      "st1 { v23.h }[2], [x11], #0x2\n"
+      "st1 { v8.h }[2], [x25], #0x2\n"
+      "st1 { v15.h }[2], [x24], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
+      "tbz x10, #0, 114f\n"
+      "st1 { v23.b }[6], [x11]\n"
+      "st1 { v8.b }[6], [x25]\n"
+      "st1 { v15.b }[6], [x24]\n"
+      "st1 { v16.b }[6], [x23]\n"
       "b 114f\n"
       "111:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 114f\n"
-      "st1 { v23.b }[4], [x9]\n"
-      "st1 { v8.b }[4], [x26]\n"
-      "st1 { v15.b }[4], [x25]\n"
-      "st1 { v16.b }[4], [x24]\n"
+      "tbz x10, #0, 114f\n"
+      "st1 { v23.b }[4], [x11]\n"
+      "st1 { v8.b }[4], [x25]\n"
+      "st1 { v15.b }[4], [x24]\n"
+      "st1 { v16.b }[4], [x23]\n"
       "b 114f\n"
       "112:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 113f\n"
-      "str h23, [x9], #0x2\n"
-      "str h8, [x26], #0x2\n"
-      "str h15, [x25], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "tbz x11, #0, 114f\n"
-      "st1 { v23.b }[2], [x9]\n"
-      "st1 { v8.b }[2], [x26]\n"
-      "st1 { v15.b }[2], [x25]\n"
-      "st1 { v16.b }[2], [x24]\n"
+      "tbz x10, #1, 113f\n"
+      "str h23, [x11], #0x2\n"
+      "str h8, [x25], #0x2\n"
+      "str h15, [x24], #0x2\n"
+      "str h16, [x23], #0x2\n"
+      "tbz x10, #0, 114f\n"
+      "st1 { v23.b }[2], [x11]\n"
+      "st1 { v8.b }[2], [x25]\n"
+      "st1 { v15.b }[2], [x24]\n"
+      "st1 { v16.b }[2], [x23]\n"
       "b 114f\n"
       "113:"  // Height 4: Partial direct writeback: partial_1_0
-      "str b23, [x9, #0x0]\n"
-      "str b8, [x26, #0x0]\n"
-      "str b15, [x25, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
+      "str b23, [x11, #0x0]\n"
+      "str b8, [x25, #0x0]\n"
+      "str b15, [x24, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
       "114:"  // Height 4: Partial direct writeback: Done
       "b 116f\n"
       "115:"  // Height 4: Full writeback
-      "str q23, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q8, [x26, #0x0]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q16, [x24, #0x0]\n"
+      "str q23, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q8, [x25, #0x0]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
       "116:"  // Height 4: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 89b\n"
       "b 176f\n"
       "117:"  // Height 5
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "118:"  // Height 5: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -2032,8 +2032,8 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "mov x28, #0x0\n"
       "120:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 121f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2065,91 +2065,91 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr q3, [x24, #0x0]\n"
       "ldr q4, [x23, #0x0]\n"
       "ldr q5, [x22, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       "blt 124f\n"
       "123:"  // Height 5: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
+      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
+      "sub x27, x27, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
-      "ldr q6, [x10, #0x10]\n"
-      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
-      "add x22, x22, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
+      "ldr q6, [x9, #0x10]\n"
       ".inst 0x4e87a498  // smmla v24.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x20]\n"
+      "ldr q7, [x9, #0x20]\n"
       ".inst 0x4e86a40c  // smmla v12.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a454  // smmla v20.4s, v2.16b, v6.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4e86a49c  // smmla v28.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x30]\n"
+      "ldr q6, [x9, #0x30]\n"
       ".inst 0x4e87a409  // smmla v9.4s, v0.16b, v7.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e87a451  // smmla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a499  // smmla v25.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x40]\n"
+      "ldr q7, [x9, #0x40]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e86a40d  // smmla v13.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a455  // smmla v21.4s, v2.16b, v6.16b\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4e86a49d  // smmla v29.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x50]\n"
+      "ldr q6, [x9, #0x50]\n"
       ".inst 0x4e87a40a  // smmla v10.4s, v0.16b, v7.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e87a452  // smmla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49a  // smmla v26.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x60]\n"
+      "ldr q7, [x9, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e86a40e  // smmla v14.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a456  // smmla v22.4s, v2.16b, v6.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e86a49e  // smmla v30.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x70]\n"
+      "ldr q6, [x9, #0x70]\n"
       ".inst 0x4e87a40b  // smmla v11.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e87a453  // smmla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49b  // smmla v27.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x80]\n"
+      "ldr q7, [x9, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e86a40f  // smmla v15.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a457  // smmla v23.4s, v2.16b, v6.16b\n"
       "ldr q2, [x25, #0x0]\n"
       ".inst 0x4e86a49f  // smmla v31.4s, v4.16b, v6.16b\n"
-      "ldr q0, [x10, #0x90]\n"
+      "ldr q0, [x9, #0x90]\n"
       "ldr q4, [x23, #0x0]\n"
       ".inst 0x4e87a428  // smmla v8.4s, v1.16b, v7.16b\n"
       ".inst 0x4e87a470  // smmla v16.4s, v3.16b, v7.16b\n"
       ".inst 0x4e87a4b8  // smmla v24.4s, v5.16b, v7.16b\n"
-      "ldr q6, [x10, #0xa0]\n"
+      "ldr q6, [x9, #0xa0]\n"
       ".inst 0x4e80a42c  // smmla v12.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a474  // smmla v20.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4bc  // smmla v28.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xb0]\n"
+      "ldr q0, [x9, #0xb0]\n"
       ".inst 0x4e86a429  // smmla v9.4s, v1.16b, v6.16b\n"
       ".inst 0x4e86a471  // smmla v17.4s, v3.16b, v6.16b\n"
       ".inst 0x4e86a4b9  // smmla v25.4s, v5.16b, v6.16b\n"
-      "ldr q6, [x10, #0xc0]\n"
+      "ldr q6, [x9, #0xc0]\n"
       ".inst 0x4e80a42d  // smmla v13.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a475  // smmla v21.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4bd  // smmla v29.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xd0]\n"
+      "ldr q0, [x9, #0xd0]\n"
       ".inst 0x4e86a42a  // smmla v10.4s, v1.16b, v6.16b\n"
       ".inst 0x4e86a472  // smmla v18.4s, v3.16b, v6.16b\n"
       ".inst 0x4e86a4ba  // smmla v26.4s, v5.16b, v6.16b\n"
-      "ldr q6, [x10, #0xe0]\n"
+      "ldr q6, [x9, #0xe0]\n"
       ".inst 0x4e80a42e  // smmla v14.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a476  // smmla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4be  // smmla v30.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q0, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4e86a42b  // smmla v11.4s, v1.16b, v6.16b\n"
       ".inst 0x4e86a473  // smmla v19.4s, v3.16b, v6.16b\n"
       ".inst 0x4e86a4bb  // smmla v27.4s, v5.16b, v6.16b\n"
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       ".inst 0x4e80a42f  // smmla v15.4s, v1.16b, v0.16b\n"
       "ldr q1, [x26, #0x0]\n"
       ".inst 0x4e80a477  // smmla v23.4s, v3.16b, v0.16b\n"
@@ -2160,79 +2160,79 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "124:"  // Height 5: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
+      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
-      "ldr q6, [x10, #0x10]\n"
-      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
+      "ldr q6, [x9, #0x10]\n"
       ".inst 0x4e87a498  // smmla v24.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x20]\n"
+      "ldr q7, [x9, #0x20]\n"
       ".inst 0x4e86a40c  // smmla v12.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a454  // smmla v20.4s, v2.16b, v6.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e86a49c  // smmla v28.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x30]\n"
+      "ldr q6, [x9, #0x30]\n"
       ".inst 0x4e87a409  // smmla v9.4s, v0.16b, v7.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e87a451  // smmla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a499  // smmla v25.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x40]\n"
+      "ldr q7, [x9, #0x40]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e86a40d  // smmla v13.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a455  // smmla v21.4s, v2.16b, v6.16b\n"
+      "add x22, x22, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e86a49d  // smmla v29.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x50]\n"
+      "ldr q6, [x9, #0x50]\n"
       ".inst 0x4e87a40a  // smmla v10.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e87a452  // smmla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49a  // smmla v26.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x60]\n"
+      "ldr q7, [x9, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e86a40e  // smmla v14.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a456  // smmla v22.4s, v2.16b, v6.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e86a49e  // smmla v30.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x70]\n"
+      "ldr q6, [x9, #0x70]\n"
       ".inst 0x4e87a40b  // smmla v11.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e87a453  // smmla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49b  // smmla v27.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x80]\n"
+      "ldr q7, [x9, #0x80]\n"
       ".inst 0x4e86a40f  // smmla v15.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a457  // smmla v23.4s, v2.16b, v6.16b\n"
       ".inst 0x4e86a49f  // smmla v31.4s, v4.16b, v6.16b\n"
-      "ldr q0, [x10, #0x90]\n"
+      "ldr q0, [x9, #0x90]\n"
       ".inst 0x4e87a428  // smmla v8.4s, v1.16b, v7.16b\n"
       ".inst 0x4e87a470  // smmla v16.4s, v3.16b, v7.16b\n"
       ".inst 0x4e87a4b8  // smmla v24.4s, v5.16b, v7.16b\n"
-      "ldr q2, [x10, #0xa0]\n"
+      "ldr q2, [x9, #0xa0]\n"
       ".inst 0x4e80a42c  // smmla v12.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a474  // smmla v20.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4bc  // smmla v28.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xb0]\n"
+      "ldr q0, [x9, #0xb0]\n"
       ".inst 0x4e82a429  // smmla v9.4s, v1.16b, v2.16b\n"
       ".inst 0x4e82a471  // smmla v17.4s, v3.16b, v2.16b\n"
       ".inst 0x4e82a4b9  // smmla v25.4s, v5.16b, v2.16b\n"
-      "ldr q2, [x10, #0xc0]\n"
+      "ldr q2, [x9, #0xc0]\n"
       ".inst 0x4e80a42d  // smmla v13.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a475  // smmla v21.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4bd  // smmla v29.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xd0]\n"
+      "ldr q0, [x9, #0xd0]\n"
       ".inst 0x4e82a42a  // smmla v10.4s, v1.16b, v2.16b\n"
       ".inst 0x4e82a472  // smmla v18.4s, v3.16b, v2.16b\n"
       ".inst 0x4e82a4ba  // smmla v26.4s, v5.16b, v2.16b\n"
-      "ldr q2, [x10, #0xe0]\n"
+      "ldr q2, [x9, #0xe0]\n"
       ".inst 0x4e80a42e  // smmla v14.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a476  // smmla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4be  // smmla v30.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q0, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4e82a42b  // smmla v11.4s, v1.16b, v2.16b\n"
       ".inst 0x4e82a473  // smmla v19.4s, v3.16b, v2.16b\n"
       ".inst 0x4e82a4bb  // smmla v27.4s, v5.16b, v2.16b\n"
@@ -2244,44 +2244,44 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 127f\n"
       "126:"  // Height 5: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d3, [x24], #0x8\n"
-      "ldr d2, [x23], #0x8\n"
-      "cmp x27, #0x8\n"
       "ldr d0, [x22], #0x8\n"
-      "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v3.2d, v2.2d\n"
-      "trn1 v2.2d, v0.2d, v5.2d\n"
-      "ldr q0, [x10, #0x10]\n"
+      "ldr q1, [x9, #0x0]\n"
+      "trn1 v2.2d, v0.2d, v2.2d\n"
       ".inst 0x4e81a488  // smmla v8.4s, v4.16b, v1.16b\n"
+      "ldr q0, [x9, #0x10]\n"
       ".inst 0x4e81a470  // smmla v16.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x20]\n"
+      "ldr q1, [x9, #0x20]\n"
       ".inst 0x4e80a48c  // smmla v12.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a474  // smmla v20.4s, v3.16b, v0.16b\n"
+      "cmp x27, #0x8\n"
       ".inst 0x4e80a45c  // smmla v28.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x30]\n"
+      "ldr q0, [x9, #0x30]\n"
       ".inst 0x4e81a489  // smmla v9.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a471  // smmla v17.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a459  // smmla v25.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x40]\n"
+      "ldr q1, [x9, #0x40]\n"
       ".inst 0x4e80a48d  // smmla v13.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a475  // smmla v21.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45d  // smmla v29.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x50]\n"
+      "ldr q0, [x9, #0x50]\n"
       ".inst 0x4e81a48a  // smmla v10.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a472  // smmla v18.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a45a  // smmla v26.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x60]\n"
+      "ldr q1, [x9, #0x60]\n"
       ".inst 0x4e80a48e  // smmla v14.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a476  // smmla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45e  // smmla v30.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q0, [x9, #0x70]\n"
       ".inst 0x4e81a48b  // smmla v11.4s, v4.16b, v1.16b\n"
+      "add x9, x9, #0x80\n"
       ".inst 0x4e81a473  // smmla v19.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a45b  // smmla v27.4s, v2.16b, v1.16b\n"
       ".inst 0x4e80a48f  // smmla v15.4s, v4.16b, v0.16b\n"
@@ -2338,36 +2338,36 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr b4, [x23, #0x0]\n"
       "ldr b5, [x22, #0x0]\n"
       "131:"  // Height 5: Multiply loop: Ragged operand read: Done
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn1 v3.2d, v3.2d, v4.2d\n"
       "trn1 v2.2d, v5.2d, v0.2d\n"
-      "ldr q0, [x10, #0x10]\n"
+      "ldr q0, [x9, #0x10]\n"
       ".inst 0x4e87a4c8  // smmla v8.4s, v6.16b, v7.16b\n"
       ".inst 0x4e87a470  // smmla v16.4s, v3.16b, v7.16b\n"
       ".inst 0x4e87a458  // smmla v24.4s, v2.16b, v7.16b\n"
-      "ldr q1, [x10, #0x20]\n"
+      "ldr q1, [x9, #0x20]\n"
       ".inst 0x4e80a4cc  // smmla v12.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a474  // smmla v20.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45c  // smmla v28.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x30]\n"
+      "ldr q0, [x9, #0x30]\n"
       ".inst 0x4e81a4c9  // smmla v9.4s, v6.16b, v1.16b\n"
       ".inst 0x4e81a471  // smmla v17.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a459  // smmla v25.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x40]\n"
+      "ldr q1, [x9, #0x40]\n"
       ".inst 0x4e80a4cd  // smmla v13.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a475  // smmla v21.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45d  // smmla v29.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x50]\n"
+      "ldr q0, [x9, #0x50]\n"
       ".inst 0x4e81a4ca  // smmla v10.4s, v6.16b, v1.16b\n"
       ".inst 0x4e81a472  // smmla v18.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a45a  // smmla v26.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x60]\n"
+      "ldr q1, [x9, #0x60]\n"
       ".inst 0x4e80a4ce  // smmla v14.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a476  // smmla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45e  // smmla v30.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q0, [x9, #0x70]\n"
+      "add x9, x9, #0x80\n"
       ".inst 0x4e81a4cb  // smmla v11.4s, v6.16b, v1.16b\n"
       ".inst 0x4e81a473  // smmla v19.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a45b  // smmla v27.4s, v2.16b, v1.16b\n"
@@ -2388,37 +2388,37 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x11, x20\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x14, x14, #0x40\n"
+      "add x24, x25, x20\n"
+      "add x23, x24, x20\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "add x26, x9, x20\n"
+      "add x22, x23, x20\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "add x25, x26, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x25, x20\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "add x23, x24, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "add x14, x14, #0x40\n"
       "uzp1 v24.2d, v24.2d, v28.2d\n"
       "uzp1 v25.2d, v25.2d, v29.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v26.2d, v26.2d, v30.2d\n"
       "uzp1 v27.2d, v27.2d, v31.2d\n"
       "mov v31.16b, v2.16b\n"
+      "add v31.4s, v31.4s, v4.4s\n"
       "add v12.4s, v12.4s, v3.4s\n"
       "add v13.4s, v13.4s, v1.4s\n"
       "add v14.4s, v14.4s, v0.4s\n"
-      "add v31.4s, v31.4s, v4.4s\n"
       "add v8.4s, v8.4s, v4.4s\n"
       "add v9.4s, v9.4s, v3.4s\n"
       "add v10.4s, v10.4s, v1.4s\n"
@@ -2448,9 +2448,9 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "add x13, x13, #0x40\n"
       "b 134f\n"
       "133:"  // Height 5: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -2489,11 +2489,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v23.4s, v23.4s, #0x1f\n"
       "sqadd v31.4s, v31.4s, v30.4s\n"
-      "and v30.16b, v8.16b, v0.16b\n"
       "sqadd v12.4s, v12.4s, v29.4s\n"
-      "and v29.16b, v9.16b, v1.16b\n"
       "sqadd v13.4s, v13.4s, v28.4s\n"
       "sqadd v14.4s, v14.4s, v23.4s\n"
+      "and v30.16b, v8.16b, v0.16b\n"
+      "and v29.16b, v9.16b, v1.16b\n"
       "and v28.16b, v10.16b, v2.16b\n"
       "and v23.16b, v11.16b, v3.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
@@ -2501,11 +2501,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v23.4s, v23.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v30.4s\n"
-      "and v30.16b, v15.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v29.4s\n"
-      "and v29.16b, v20.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v28.4s\n"
       "sqadd v11.4s, v11.4s, v23.4s\n"
+      "and v30.16b, v15.16b, v0.16b\n"
+      "and v29.16b, v20.16b, v1.16b\n"
       "and v28.16b, v21.16b, v2.16b\n"
       "and v23.16b, v22.16b, v3.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
@@ -2513,11 +2513,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v23.4s, v23.4s, #0x1f\n"
       "sqadd v15.4s, v15.4s, v30.4s\n"
-      "and v30.16b, v16.16b, v0.16b\n"
       "sqadd v20.4s, v20.4s, v29.4s\n"
-      "and v29.16b, v17.16b, v1.16b\n"
       "sqadd v21.4s, v21.4s, v28.4s\n"
       "sqadd v22.4s, v22.4s, v23.4s\n"
+      "and v30.16b, v16.16b, v0.16b\n"
+      "and v29.16b, v17.16b, v1.16b\n"
       "and v28.16b, v18.16b, v2.16b\n"
       "and v23.16b, v19.16b, v3.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
@@ -2525,11 +2525,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v23.4s, v23.4s, #0x1f\n"
       "sqadd v16.4s, v16.4s, v30.4s\n"
-      "and v30.16b, v24.16b, v0.16b\n"
       "sqadd v17.4s, v17.4s, v29.4s\n"
-      "and v29.16b, v25.16b, v1.16b\n"
       "sqadd v18.4s, v18.4s, v28.4s\n"
       "sqadd v19.4s, v19.4s, v23.4s\n"
+      "and v30.16b, v24.16b, v0.16b\n"
+      "and v29.16b, v25.16b, v1.16b\n"
       "and v28.16b, v26.16b, v2.16b\n"
       "and v23.16b, v27.16b, v3.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
@@ -2541,21 +2541,21 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sqadd v26.4s, v26.4s, v28.4s\n"
       "sqadd v27.4s, v27.4s, v23.4s\n"
       "135:"  // Height 5: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "srshl v31.4s, v31.4s, v0.4s\n"
       "srshl v12.4s, v12.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v29.4s }, [x21]\n"
-      "ld1r { v28.4s }, [x20]\n"
       "srshl v13.4s, v13.4s, v2.4s\n"
       "srshl v14.4s, v14.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v28.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v23.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "srshl v15.4s, v15.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v1.4s\n"
       "srshl v21.4s, v21.4s, v2.4s\n"
@@ -2644,133 +2644,132 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v16.16b, v16.16b, v18.16b\n"
       "uzp1 v24.16b, v24.16b, v17.16b\n"
       "bge 144f\n"
-      "tbz x11, #3, 139f\n"
-      "str d31, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "tbz x11, #2, 137f\n"
-      "st1 { v31.s }[2], [x9], #0x4\n"
-      "st1 { v8.s }[2], [x26], #0x4\n"
-      "st1 { v15.s }[2], [x25], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "tbz x11, #1, 136f\n"
-      "st1 { v31.h }[6], [x9], #0x2\n"
-      "st1 { v8.h }[6], [x26], #0x2\n"
-      "st1 { v15.h }[6], [x25], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "tbz x11, #0, 143f\n"
-      "st1 { v31.b }[14], [x9]\n"
-      "st1 { v8.b }[14], [x26]\n"
-      "st1 { v15.b }[14], [x25]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "tbz x10, #3, 139f\n"
+      "str d31, [x11], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "tbz x10, #2, 137f\n"
+      "st1 { v31.s }[2], [x11], #0x4\n"
+      "st1 { v8.s }[2], [x25], #0x4\n"
+      "st1 { v15.s }[2], [x24], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "tbz x10, #1, 136f\n"
+      "st1 { v31.h }[6], [x11], #0x2\n"
+      "st1 { v8.h }[6], [x25], #0x2\n"
+      "st1 { v15.h }[6], [x24], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "tbz x10, #0, 143f\n"
+      "st1 { v31.b }[14], [x11]\n"
+      "st1 { v8.b }[14], [x25]\n"
+      "st1 { v15.b }[14], [x24]\n"
+      "st1 { v16.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 143f\n"
       "136:"  // Height 5: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 143f\n"
-      "st1 { v31.b }[12], [x9]\n"
-      "st1 { v8.b }[12], [x26]\n"
-      "st1 { v15.b }[12], [x25]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "tbz x10, #0, 143f\n"
+      "st1 { v31.b }[12], [x11]\n"
+      "st1 { v8.b }[12], [x25]\n"
+      "st1 { v15.b }[12], [x24]\n"
+      "st1 { v16.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 143f\n"
       "137:"  // Height 5: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 138f\n"
-      "st1 { v31.h }[4], [x9], #0x2\n"
-      "st1 { v8.h }[4], [x26], #0x2\n"
-      "st1 { v15.h }[4], [x25], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "tbz x11, #0, 143f\n"
-      "st1 { v31.b }[10], [x9]\n"
-      "st1 { v8.b }[10], [x26]\n"
-      "st1 { v15.b }[10], [x25]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "tbz x10, #1, 138f\n"
+      "st1 { v31.h }[4], [x11], #0x2\n"
+      "st1 { v8.h }[4], [x25], #0x2\n"
+      "st1 { v15.h }[4], [x24], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "tbz x10, #0, 143f\n"
+      "st1 { v31.b }[10], [x11]\n"
+      "st1 { v8.b }[10], [x25]\n"
+      "st1 { v15.b }[10], [x24]\n"
+      "st1 { v16.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 143f\n"
       "138:"  // Height 5: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 143f\n"
-      "st1 { v31.b }[8], [x9]\n"
-      "st1 { v8.b }[8], [x26]\n"
-      "st1 { v15.b }[8], [x25]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "tbz x10, #0, 143f\n"
+      "st1 { v31.b }[8], [x11]\n"
+      "st1 { v8.b }[8], [x25]\n"
+      "st1 { v15.b }[8], [x24]\n"
+      "st1 { v16.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 143f\n"
       "139:"  // Height 5: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 141f\n"
-      "str s31, [x9], #0x4\n"
-      "str s8, [x26], #0x4\n"
-      "str s15, [x25], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "tbz x11, #1, 140f\n"
-      "st1 { v31.h }[2], [x9], #0x2\n"
-      "st1 { v8.h }[2], [x26], #0x2\n"
-      "st1 { v15.h }[2], [x25], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "tbz x11, #0, 143f\n"
-      "st1 { v31.b }[6], [x9]\n"
-      "st1 { v8.b }[6], [x26]\n"
-      "st1 { v15.b }[6], [x25]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "tbz x10, #2, 141f\n"
+      "str s31, [x11], #0x4\n"
+      "str s8, [x25], #0x4\n"
+      "str s15, [x24], #0x4\n"
+      "str s16, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "tbz x10, #1, 140f\n"
+      "st1 { v31.h }[2], [x11], #0x2\n"
+      "st1 { v8.h }[2], [x25], #0x2\n"
+      "st1 { v15.h }[2], [x24], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "tbz x10, #0, 143f\n"
+      "st1 { v31.b }[6], [x11]\n"
+      "st1 { v8.b }[6], [x25]\n"
+      "st1 { v15.b }[6], [x24]\n"
+      "st1 { v16.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 143f\n"
       "140:"  // Height 5: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 143f\n"
-      "st1 { v31.b }[4], [x9]\n"
-      "st1 { v8.b }[4], [x26]\n"
-      "st1 { v15.b }[4], [x25]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "tbz x10, #0, 143f\n"
+      "st1 { v31.b }[4], [x11]\n"
+      "st1 { v8.b }[4], [x25]\n"
+      "st1 { v15.b }[4], [x24]\n"
+      "st1 { v16.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 143f\n"
       "141:"  // Height 5: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 142f\n"
-      "str h31, [x9], #0x2\n"
-      "str h8, [x26], #0x2\n"
-      "str h15, [x25], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "tbz x11, #0, 143f\n"
-      "st1 { v31.b }[2], [x9]\n"
-      "st1 { v8.b }[2], [x26]\n"
-      "st1 { v15.b }[2], [x25]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "tbz x10, #1, 142f\n"
+      "str h31, [x11], #0x2\n"
+      "str h8, [x25], #0x2\n"
+      "str h15, [x24], #0x2\n"
+      "str h16, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "tbz x10, #0, 143f\n"
+      "st1 { v31.b }[2], [x11]\n"
+      "st1 { v8.b }[2], [x25]\n"
+      "st1 { v15.b }[2], [x24]\n"
+      "st1 { v16.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 143f\n"
       "142:"  // Height 5: Partial direct writeback: partial_1_0
-      "str b31, [x9, #0x0]\n"
-      "str b8, [x26, #0x0]\n"
-      "str b15, [x25, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b31, [x11, #0x0]\n"
+      "str b8, [x25, #0x0]\n"
+      "str b15, [x24, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "143:"  // Height 5: Partial direct writeback: Done
       "b 145f\n"
       "144:"  // Height 5: Full writeback
-      "str q31, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q8, [x26, #0x0]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q31, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q8, [x25, #0x0]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "145:"  // Height 5: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 118b\n"
       "b 176f\n"
       "146:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x6\n"
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "147:"  // Height 6: Column loop
       "movi v8.4s, #0x0\n"
       "movi v9.4s, #0x0\n"
@@ -2800,8 +2799,8 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "mov x28, #0x0\n"
       "149:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 150f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2837,93 +2836,93 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr q4, [x23, #0x0]\n"
       "ldr q5, [x22, #0x0]\n"
       "ldr q6, [x21, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       "blt 153f\n"
       "152:"  // Height 6: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
       "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
+      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
-      "ldr q6, [x10, #0x10]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
+      "ldr q6, [x9, #0x10]\n"
       ".inst 0x4e87a498  // smmla v24.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x20]\n"
+      "ldr q7, [x9, #0x20]\n"
       ".inst 0x4e86a40c  // smmla v12.4s, v0.16b, v6.16b\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e86a454  // smmla v20.4s, v2.16b, v6.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e86a49c  // smmla v28.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
+      "ldr q6, [x9, #0x30]\n"
       ".inst 0x4e87a409  // smmla v9.4s, v0.16b, v7.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e87a451  // smmla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a499  // smmla v25.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x40]\n"
+      "ldr q7, [x9, #0x40]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e86a40d  // smmla v13.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a455  // smmla v21.4s, v2.16b, v6.16b\n"
+      "add x22, x22, #0x10\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x4e86a49d  // smmla v29.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x50]\n"
+      "ldr q6, [x9, #0x50]\n"
       ".inst 0x4e87a40a  // smmla v10.4s, v0.16b, v7.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e87a452  // smmla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49a  // smmla v26.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x60]\n"
+      "ldr q7, [x9, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e86a40e  // smmla v14.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a456  // smmla v22.4s, v2.16b, v6.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e86a49e  // smmla v30.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x70]\n"
+      "ldr q6, [x9, #0x70]\n"
       ".inst 0x4e87a40b  // smmla v11.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e87a453  // smmla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49b  // smmla v27.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x80]\n"
+      "ldr q7, [x9, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e86a40f  // smmla v15.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a457  // smmla v23.4s, v2.16b, v6.16b\n"
       "ldr q2, [x25, #0x0]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4e86a49f  // smmla v31.4s, v4.16b, v6.16b\n"
-      "ldr q0, [x10, #0x90]\n"
+      "ldr q0, [x9, #0x90]\n"
       "ldr q4, [x23, #0x0]\n"
       ".inst 0x4e87a428  // smmla v8.4s, v1.16b, v7.16b\n"
       ".inst 0x4e87a470  // smmla v16.4s, v3.16b, v7.16b\n"
       ".inst 0x4e87a4b8  // smmla v24.4s, v5.16b, v7.16b\n"
-      "ldr q6, [x10, #0xa0]\n"
+      "ldr q6, [x9, #0xa0]\n"
       ".inst 0x4e80a42c  // smmla v12.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a474  // smmla v20.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4bc  // smmla v28.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xb0]\n"
+      "ldr q0, [x9, #0xb0]\n"
       ".inst 0x4e86a429  // smmla v9.4s, v1.16b, v6.16b\n"
       ".inst 0x4e86a471  // smmla v17.4s, v3.16b, v6.16b\n"
       ".inst 0x4e86a4b9  // smmla v25.4s, v5.16b, v6.16b\n"
-      "ldr q6, [x10, #0xc0]\n"
+      "ldr q6, [x9, #0xc0]\n"
       ".inst 0x4e80a42d  // smmla v13.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a475  // smmla v21.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4bd  // smmla v29.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xd0]\n"
+      "ldr q0, [x9, #0xd0]\n"
       ".inst 0x4e86a42a  // smmla v10.4s, v1.16b, v6.16b\n"
       ".inst 0x4e86a472  // smmla v18.4s, v3.16b, v6.16b\n"
       ".inst 0x4e86a4ba  // smmla v26.4s, v5.16b, v6.16b\n"
-      "ldr q6, [x10, #0xe0]\n"
+      "ldr q6, [x9, #0xe0]\n"
       ".inst 0x4e80a42e  // smmla v14.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a476  // smmla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4be  // smmla v30.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q0, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4e86a42b  // smmla v11.4s, v1.16b, v6.16b\n"
       ".inst 0x4e86a473  // smmla v19.4s, v3.16b, v6.16b\n"
       ".inst 0x4e86a4bb  // smmla v27.4s, v5.16b, v6.16b\n"
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       ".inst 0x4e80a42f  // smmla v15.4s, v1.16b, v0.16b\n"
       "ldr q1, [x26, #0x0]\n"
       ".inst 0x4e80a477  // smmla v23.4s, v3.16b, v0.16b\n"
@@ -2935,81 +2934,81 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "153:"  // Height 6: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
+      "add x25, x25, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
-      "ldr q6, [x10, #0x10]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "ldr q6, [x9, #0x10]\n"
       ".inst 0x4e87a498  // smmla v24.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x20]\n"
+      "ldr q7, [x9, #0x20]\n"
       ".inst 0x4e86a40c  // smmla v12.4s, v0.16b, v6.16b\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e86a454  // smmla v20.4s, v2.16b, v6.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e86a49c  // smmla v28.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
+      "ldr q6, [x9, #0x30]\n"
       ".inst 0x4e87a409  // smmla v9.4s, v0.16b, v7.16b\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e87a451  // smmla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a499  // smmla v25.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x40]\n"
+      "ldr q7, [x9, #0x40]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4e86a40d  // smmla v13.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a455  // smmla v21.4s, v2.16b, v6.16b\n"
+      "add x21, x21, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e86a49d  // smmla v29.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x50]\n"
+      "ldr q6, [x9, #0x50]\n"
       ".inst 0x4e87a40a  // smmla v10.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e87a452  // smmla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49a  // smmla v26.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x60]\n"
+      "ldr q7, [x9, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e86a40e  // smmla v14.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a456  // smmla v22.4s, v2.16b, v6.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e86a49e  // smmla v30.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x70]\n"
+      "ldr q6, [x9, #0x70]\n"
       ".inst 0x4e87a40b  // smmla v11.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e87a453  // smmla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49b  // smmla v27.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x80]\n"
+      "ldr q7, [x9, #0x80]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4e86a40f  // smmla v15.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a457  // smmla v23.4s, v2.16b, v6.16b\n"
       ".inst 0x4e86a49f  // smmla v31.4s, v4.16b, v6.16b\n"
-      "ldr q0, [x10, #0x90]\n"
+      "ldr q0, [x9, #0x90]\n"
       ".inst 0x4e87a428  // smmla v8.4s, v1.16b, v7.16b\n"
       ".inst 0x4e87a470  // smmla v16.4s, v3.16b, v7.16b\n"
       ".inst 0x4e87a4b8  // smmla v24.4s, v5.16b, v7.16b\n"
-      "ldr q2, [x10, #0xa0]\n"
+      "ldr q2, [x9, #0xa0]\n"
       ".inst 0x4e80a42c  // smmla v12.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a474  // smmla v20.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4bc  // smmla v28.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xb0]\n"
+      "ldr q0, [x9, #0xb0]\n"
       ".inst 0x4e82a429  // smmla v9.4s, v1.16b, v2.16b\n"
       ".inst 0x4e82a471  // smmla v17.4s, v3.16b, v2.16b\n"
       ".inst 0x4e82a4b9  // smmla v25.4s, v5.16b, v2.16b\n"
-      "ldr q2, [x10, #0xc0]\n"
+      "ldr q2, [x9, #0xc0]\n"
       ".inst 0x4e80a42d  // smmla v13.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a475  // smmla v21.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4bd  // smmla v29.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xd0]\n"
+      "ldr q0, [x9, #0xd0]\n"
       ".inst 0x4e82a42a  // smmla v10.4s, v1.16b, v2.16b\n"
       ".inst 0x4e82a472  // smmla v18.4s, v3.16b, v2.16b\n"
       ".inst 0x4e82a4ba  // smmla v26.4s, v5.16b, v2.16b\n"
-      "ldr q2, [x10, #0xe0]\n"
+      "ldr q2, [x9, #0xe0]\n"
       ".inst 0x4e80a42e  // smmla v14.4s, v1.16b, v0.16b\n"
       ".inst 0x4e80a476  // smmla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a4be  // smmla v30.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "ldr q0, [x9, #0xf0]\n"
+      "add x9, x9, #0x100\n"
       ".inst 0x4e82a42b  // smmla v11.4s, v1.16b, v2.16b\n"
       ".inst 0x4e82a473  // smmla v19.4s, v3.16b, v2.16b\n"
       ".inst 0x4e82a4bb  // smmla v27.4s, v5.16b, v2.16b\n"
@@ -3021,44 +3020,44 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 156f\n"
       "155:"  // Height 6: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d5, [x24], #0x8\n"
-      "ldr d3, [x23], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "cmp x27, #0x8\n"
-      "ldr d2, [x22], #0x8\n"
+      "ldr d1, [x22], #0x8\n"
       "ldr d0, [x21], #0x8\n"
-      "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v5.2d, v3.2d\n"
-      "trn1 v2.2d, v2.2d, v0.2d\n"
-      "ldr q0, [x10, #0x10]\n"
+      "trn1 v2.2d, v1.2d, v0.2d\n"
+      "ldr q1, [x9, #0x0]\n"
+      "ldr q0, [x9, #0x10]\n"
       ".inst 0x4e81a488  // smmla v8.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a470  // smmla v16.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x20]\n"
+      "ldr q1, [x9, #0x20]\n"
       ".inst 0x4e80a48c  // smmla v12.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a474  // smmla v20.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45c  // smmla v28.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x30]\n"
+      "ldr q0, [x9, #0x30]\n"
       ".inst 0x4e81a489  // smmla v9.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a471  // smmla v17.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a459  // smmla v25.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x40]\n"
+      "ldr q1, [x9, #0x40]\n"
       ".inst 0x4e80a48d  // smmla v13.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a475  // smmla v21.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45d  // smmla v29.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x50]\n"
+      "ldr q0, [x9, #0x50]\n"
       ".inst 0x4e81a48a  // smmla v10.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a472  // smmla v18.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a45a  // smmla v26.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x60]\n"
+      "ldr q1, [x9, #0x60]\n"
       ".inst 0x4e80a48e  // smmla v14.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a476  // smmla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45e  // smmla v30.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q0, [x9, #0x70]\n"
+      "add x9, x9, #0x80\n"
       ".inst 0x4e81a48b  // smmla v11.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a473  // smmla v19.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a45b  // smmla v27.4s, v2.16b, v1.16b\n"
@@ -3123,37 +3122,37 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "ldr b5, [x22, #0x0]\n"
       "ldr b6, [x21, #0x0]\n"
       "160:"  // Height 6: Multiply loop: Ragged operand read: Done
-      "ldr q7, [x10, #0x0]\n"
+      "ldr q7, [x9, #0x0]\n"
       "trn1 v2.2d, v1.2d, v2.2d\n"
       "trn1 v4.2d, v3.2d, v4.2d\n"
-      "trn1 v3.2d, v5.2d, v6.2d\n"
-      "ldr q0, [x10, #0x10]\n"
       ".inst 0x4e87a448  // smmla v8.4s, v2.16b, v7.16b\n"
+      "trn1 v3.2d, v5.2d, v6.2d\n"
+      "ldr q0, [x9, #0x10]\n"
       ".inst 0x4e87a490  // smmla v16.4s, v4.16b, v7.16b\n"
       ".inst 0x4e87a478  // smmla v24.4s, v3.16b, v7.16b\n"
-      "ldr q1, [x10, #0x20]\n"
+      "ldr q1, [x9, #0x20]\n"
       ".inst 0x4e80a44c  // smmla v12.4s, v2.16b, v0.16b\n"
       ".inst 0x4e80a494  // smmla v20.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a47c  // smmla v28.4s, v3.16b, v0.16b\n"
-      "ldr q0, [x10, #0x30]\n"
+      "ldr q0, [x9, #0x30]\n"
       ".inst 0x4e81a449  // smmla v9.4s, v2.16b, v1.16b\n"
       ".inst 0x4e81a491  // smmla v17.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a479  // smmla v25.4s, v3.16b, v1.16b\n"
-      "ldr q1, [x10, #0x40]\n"
+      "ldr q1, [x9, #0x40]\n"
       ".inst 0x4e80a44d  // smmla v13.4s, v2.16b, v0.16b\n"
       ".inst 0x4e80a495  // smmla v21.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a47d  // smmla v29.4s, v3.16b, v0.16b\n"
-      "ldr q0, [x10, #0x50]\n"
+      "ldr q0, [x9, #0x50]\n"
       ".inst 0x4e81a44a  // smmla v10.4s, v2.16b, v1.16b\n"
       ".inst 0x4e81a492  // smmla v18.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a47a  // smmla v26.4s, v3.16b, v1.16b\n"
-      "ldr q1, [x10, #0x60]\n"
+      "ldr q1, [x9, #0x60]\n"
       ".inst 0x4e80a44e  // smmla v14.4s, v2.16b, v0.16b\n"
       ".inst 0x4e80a496  // smmla v22.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a47e  // smmla v30.4s, v3.16b, v0.16b\n"
-      "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "ldr q0, [x9, #0x70]\n"
       ".inst 0x4e81a44b  // smmla v11.4s, v2.16b, v1.16b\n"
+      "add x9, x9, #0x80\n"
       ".inst 0x4e81a493  // smmla v19.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a47b  // smmla v27.4s, v3.16b, v1.16b\n"
       ".inst 0x4e80a44f  // smmla v15.4s, v2.16b, v0.16b\n"
@@ -3173,32 +3172,32 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x11, x20\n"
+      "add x24, x25, x20\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "add x23, x24, x20\n"
+      "add x22, x23, x20\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "add x14, x14, #0x40\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "add x21, x22, x20\n"
+      "prfm pstl1keep, [x11, #0x0]\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x25, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "add x24, x25, x20\n"
-      "prfm pstl1keep, [x26, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x23, x24, x20\n"
-      "prfm pstl1keep, [x25, #0x0]\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
+      "add x14, x14, #0x40\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "add x22, x23, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v23.2d, v24.2d, v28.2d\n"
       "uzp2 v24.2d, v24.2d, v28.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v28.2d, v25.2d, v29.2d\n"
       "uzp2 v25.2d, v25.2d, v29.2d\n"
       "uzp1 v29.2d, v26.2d, v30.2d\n"
@@ -3206,10 +3205,10 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v30.2d, v27.2d, v31.2d\n"
       "uzp2 v27.2d, v27.2d, v31.2d\n"
       "mov v31.16b, v2.16b\n"
+      "add v31.4s, v31.4s, v4.4s\n"
       "add v12.4s, v12.4s, v3.4s\n"
       "add v13.4s, v13.4s, v1.4s\n"
       "add v14.4s, v14.4s, v0.4s\n"
-      "add v31.4s, v31.4s, v4.4s\n"
       "add v8.4s, v8.4s, v4.4s\n"
       "add v9.4s, v9.4s, v3.4s\n"
       "add v10.4s, v10.4s, v1.4s\n"
@@ -3243,9 +3242,9 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "add x13, x13, #0x40\n"
       "b 163f\n"
       "162:"  // Height 6: per layer parameters
-      "add x21, %x[qp], %[per_layer_right_shift]\n"
+      "add x20, %x[qp], %[per_layer_right_shift]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "ld1r { v0.4s }, [x21]\n"
       "ld1r { v4.4s }, [x20]\n"
       "mov v1.16b, v0.16b\n"
       "mov v5.16b, v4.16b\n"
@@ -3288,11 +3287,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v31.4s, v31.4s, v7.4s\n"
-      "and v7.16b, v8.16b, v0.16b\n"
       "sqadd v12.4s, v12.4s, v6.4s\n"
-      "and v6.16b, v9.16b, v1.16b\n"
       "sqadd v13.4s, v13.4s, v5.4s\n"
       "sqadd v14.4s, v14.4s, v4.4s\n"
+      "and v7.16b, v8.16b, v0.16b\n"
+      "and v6.16b, v9.16b, v1.16b\n"
       "and v5.16b, v10.16b, v2.16b\n"
       "and v4.16b, v11.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3300,11 +3299,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v8.4s, v8.4s, v7.4s\n"
-      "and v7.16b, v15.16b, v0.16b\n"
       "sqadd v9.4s, v9.4s, v6.4s\n"
-      "and v6.16b, v20.16b, v1.16b\n"
       "sqadd v10.4s, v10.4s, v5.4s\n"
       "sqadd v11.4s, v11.4s, v4.4s\n"
+      "and v7.16b, v15.16b, v0.16b\n"
+      "and v6.16b, v20.16b, v1.16b\n"
       "and v5.16b, v21.16b, v2.16b\n"
       "and v4.16b, v22.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3312,11 +3311,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v15.4s, v15.4s, v7.4s\n"
-      "and v7.16b, v16.16b, v0.16b\n"
       "sqadd v20.4s, v20.4s, v6.4s\n"
-      "and v6.16b, v17.16b, v1.16b\n"
       "sqadd v21.4s, v21.4s, v5.4s\n"
       "sqadd v22.4s, v22.4s, v4.4s\n"
+      "and v7.16b, v16.16b, v0.16b\n"
+      "and v6.16b, v17.16b, v1.16b\n"
       "and v5.16b, v18.16b, v2.16b\n"
       "and v4.16b, v19.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3324,11 +3323,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v16.4s, v16.4s, v7.4s\n"
-      "and v7.16b, v23.16b, v0.16b\n"
       "sqadd v17.4s, v17.4s, v6.4s\n"
-      "and v6.16b, v28.16b, v1.16b\n"
       "sqadd v18.4s, v18.4s, v5.4s\n"
       "sqadd v19.4s, v19.4s, v4.4s\n"
+      "and v7.16b, v23.16b, v0.16b\n"
+      "and v6.16b, v28.16b, v1.16b\n"
       "and v5.16b, v29.16b, v2.16b\n"
       "and v4.16b, v30.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3336,11 +3335,11 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sqadd v23.4s, v23.4s, v7.4s\n"
-      "and v7.16b, v24.16b, v0.16b\n"
       "sqadd v28.4s, v28.4s, v6.4s\n"
-      "and v6.16b, v25.16b, v1.16b\n"
       "sqadd v29.4s, v29.4s, v5.4s\n"
       "sqadd v30.4s, v30.4s, v4.4s\n"
+      "and v7.16b, v24.16b, v0.16b\n"
+      "and v6.16b, v25.16b, v1.16b\n"
       "and v5.16b, v26.16b, v2.16b\n"
       "and v4.16b, v27.16b, v3.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
@@ -3352,21 +3351,21 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "sqadd v26.4s, v26.4s, v5.4s\n"
       "sqadd v27.4s, v27.4s, v4.4s\n"
       "164:"  // Height 6: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v6.4s }, [x20]\n"
       "srshl v31.4s, v31.4s, v0.4s\n"
       "srshl v12.4s, v12.4s, v1.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v6.4s }, [x21]\n"
-      "ld1r { v5.4s }, [x20]\n"
       "srshl v13.4s, v13.4s, v2.4s\n"
       "srshl v14.4s, v14.4s, v3.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v5.4s }, [x20]\n"
       "srshl v8.4s, v8.4s, v0.4s\n"
       "srshl v9.4s, v9.4s, v1.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x11, #0x10\n"
       "ld1r { v4.4s }, [x20]\n"
       "srshl v10.4s, v10.4s, v2.4s\n"
       "srshl v11.4s, v11.4s, v3.4s\n"
+      "cmp x10, #0x10\n"
       "srshl v15.4s, v15.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v1.4s\n"
       "srshl v21.4s, v21.4s, v2.4s\n"
@@ -3474,136 +3473,136 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "uzp1 v23.16b, v23.16b, v18.16b\n"
       "uzp1 v24.16b, v24.16b, v17.16b\n"
       "bge 173f\n"
-      "tbz x11, #3, 168f\n"
-      "str d31, [x9], #0x8\n"
-      "str d8, [x26], #0x8\n"
-      "str d15, [x25], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
-      "tbz x11, #2, 166f\n"
-      "st1 { v31.s }[2], [x9], #0x4\n"
-      "st1 { v8.s }[2], [x26], #0x4\n"
-      "st1 { v15.s }[2], [x25], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "st1 { v23.s }[2], [x23], #0x4\n"
-      "st1 { v24.s }[2], [x22], #0x4\n"
-      "tbz x11, #1, 165f\n"
-      "st1 { v31.h }[6], [x9], #0x2\n"
-      "st1 { v8.h }[6], [x26], #0x2\n"
-      "st1 { v15.h }[6], [x25], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "st1 { v23.h }[6], [x23], #0x2\n"
-      "st1 { v24.h }[6], [x22], #0x2\n"
-      "tbz x11, #0, 172f\n"
-      "st1 { v31.b }[14], [x9]\n"
-      "st1 { v8.b }[14], [x26]\n"
-      "st1 { v15.b }[14], [x25]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "st1 { v23.b }[14], [x23]\n"
-      "st1 { v24.b }[14], [x22]\n"
+      "tbz x10, #3, 168f\n"
+      "str d31, [x11], #0x8\n"
+      "str d8, [x25], #0x8\n"
+      "str d15, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
+      "str d24, [x21], #0x8\n"
+      "tbz x10, #2, 166f\n"
+      "st1 { v31.s }[2], [x11], #0x4\n"
+      "st1 { v8.s }[2], [x25], #0x4\n"
+      "st1 { v15.s }[2], [x24], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
+      "st1 { v23.s }[2], [x22], #0x4\n"
+      "st1 { v24.s }[2], [x21], #0x4\n"
+      "tbz x10, #1, 165f\n"
+      "st1 { v31.h }[6], [x11], #0x2\n"
+      "st1 { v8.h }[6], [x25], #0x2\n"
+      "st1 { v15.h }[6], [x24], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
+      "st1 { v23.h }[6], [x22], #0x2\n"
+      "st1 { v24.h }[6], [x21], #0x2\n"
+      "tbz x10, #0, 172f\n"
+      "st1 { v31.b }[14], [x11]\n"
+      "st1 { v8.b }[14], [x25]\n"
+      "st1 { v15.b }[14], [x24]\n"
+      "st1 { v16.b }[14], [x23]\n"
+      "st1 { v23.b }[14], [x22]\n"
+      "st1 { v24.b }[14], [x21]\n"
       "b 172f\n"
       "165:"  // Height 6: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 172f\n"
-      "st1 { v31.b }[12], [x9]\n"
-      "st1 { v8.b }[12], [x26]\n"
-      "st1 { v15.b }[12], [x25]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "st1 { v23.b }[12], [x23]\n"
-      "st1 { v24.b }[12], [x22]\n"
+      "tbz x10, #0, 172f\n"
+      "st1 { v31.b }[12], [x11]\n"
+      "st1 { v8.b }[12], [x25]\n"
+      "st1 { v15.b }[12], [x24]\n"
+      "st1 { v16.b }[12], [x23]\n"
+      "st1 { v23.b }[12], [x22]\n"
+      "st1 { v24.b }[12], [x21]\n"
       "b 172f\n"
       "166:"  // Height 6: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 167f\n"
-      "st1 { v31.h }[4], [x9], #0x2\n"
-      "st1 { v8.h }[4], [x26], #0x2\n"
-      "st1 { v15.h }[4], [x25], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "st1 { v23.h }[4], [x23], #0x2\n"
-      "st1 { v24.h }[4], [x22], #0x2\n"
-      "tbz x11, #0, 172f\n"
-      "st1 { v31.b }[10], [x9]\n"
-      "st1 { v8.b }[10], [x26]\n"
-      "st1 { v15.b }[10], [x25]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "st1 { v23.b }[10], [x23]\n"
-      "st1 { v24.b }[10], [x22]\n"
+      "tbz x10, #1, 167f\n"
+      "st1 { v31.h }[4], [x11], #0x2\n"
+      "st1 { v8.h }[4], [x25], #0x2\n"
+      "st1 { v15.h }[4], [x24], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
+      "st1 { v23.h }[4], [x22], #0x2\n"
+      "st1 { v24.h }[4], [x21], #0x2\n"
+      "tbz x10, #0, 172f\n"
+      "st1 { v31.b }[10], [x11]\n"
+      "st1 { v8.b }[10], [x25]\n"
+      "st1 { v15.b }[10], [x24]\n"
+      "st1 { v16.b }[10], [x23]\n"
+      "st1 { v23.b }[10], [x22]\n"
+      "st1 { v24.b }[10], [x21]\n"
       "b 172f\n"
       "167:"  // Height 6: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 172f\n"
-      "st1 { v31.b }[8], [x9]\n"
-      "st1 { v8.b }[8], [x26]\n"
-      "st1 { v15.b }[8], [x25]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "st1 { v23.b }[8], [x23]\n"
-      "st1 { v24.b }[8], [x22]\n"
+      "tbz x10, #0, 172f\n"
+      "st1 { v31.b }[8], [x11]\n"
+      "st1 { v8.b }[8], [x25]\n"
+      "st1 { v15.b }[8], [x24]\n"
+      "st1 { v16.b }[8], [x23]\n"
+      "st1 { v23.b }[8], [x22]\n"
+      "st1 { v24.b }[8], [x21]\n"
       "b 172f\n"
       "168:"  // Height 6: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 170f\n"
-      "str s31, [x9], #0x4\n"
-      "str s8, [x26], #0x4\n"
-      "str s15, [x25], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "str s23, [x23], #0x4\n"
-      "str s24, [x22], #0x4\n"
-      "tbz x11, #1, 169f\n"
-      "st1 { v31.h }[2], [x9], #0x2\n"
-      "st1 { v8.h }[2], [x26], #0x2\n"
-      "st1 { v15.h }[2], [x25], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "st1 { v23.h }[2], [x23], #0x2\n"
-      "st1 { v24.h }[2], [x22], #0x2\n"
-      "tbz x11, #0, 172f\n"
-      "st1 { v31.b }[6], [x9]\n"
-      "st1 { v8.b }[6], [x26]\n"
-      "st1 { v15.b }[6], [x25]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "st1 { v23.b }[6], [x23]\n"
-      "st1 { v24.b }[6], [x22]\n"
+      "tbz x10, #2, 170f\n"
+      "str s31, [x11], #0x4\n"
+      "str s8, [x25], #0x4\n"
+      "str s15, [x24], #0x4\n"
+      "str s16, [x23], #0x4\n"
+      "str s23, [x22], #0x4\n"
+      "str s24, [x21], #0x4\n"
+      "tbz x10, #1, 169f\n"
+      "st1 { v31.h }[2], [x11], #0x2\n"
+      "st1 { v8.h }[2], [x25], #0x2\n"
+      "st1 { v15.h }[2], [x24], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
+      "st1 { v23.h }[2], [x22], #0x2\n"
+      "st1 { v24.h }[2], [x21], #0x2\n"
+      "tbz x10, #0, 172f\n"
+      "st1 { v31.b }[6], [x11]\n"
+      "st1 { v8.b }[6], [x25]\n"
+      "st1 { v15.b }[6], [x24]\n"
+      "st1 { v16.b }[6], [x23]\n"
+      "st1 { v23.b }[6], [x22]\n"
+      "st1 { v24.b }[6], [x21]\n"
       "b 172f\n"
       "169:"  // Height 6: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 172f\n"
-      "st1 { v31.b }[4], [x9]\n"
-      "st1 { v8.b }[4], [x26]\n"
-      "st1 { v15.b }[4], [x25]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "st1 { v23.b }[4], [x23]\n"
-      "st1 { v24.b }[4], [x22]\n"
+      "tbz x10, #0, 172f\n"
+      "st1 { v31.b }[4], [x11]\n"
+      "st1 { v8.b }[4], [x25]\n"
+      "st1 { v15.b }[4], [x24]\n"
+      "st1 { v16.b }[4], [x23]\n"
+      "st1 { v23.b }[4], [x22]\n"
+      "st1 { v24.b }[4], [x21]\n"
       "b 172f\n"
       "170:"  // Height 6: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 171f\n"
-      "str h31, [x9], #0x2\n"
-      "str h8, [x26], #0x2\n"
-      "str h15, [x25], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "str h23, [x23], #0x2\n"
-      "str h24, [x22], #0x2\n"
-      "tbz x11, #0, 172f\n"
-      "st1 { v31.b }[2], [x9]\n"
-      "st1 { v8.b }[2], [x26]\n"
-      "st1 { v15.b }[2], [x25]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "st1 { v23.b }[2], [x23]\n"
-      "st1 { v24.b }[2], [x22]\n"
+      "tbz x10, #1, 171f\n"
+      "str h31, [x11], #0x2\n"
+      "str h8, [x25], #0x2\n"
+      "str h15, [x24], #0x2\n"
+      "str h16, [x23], #0x2\n"
+      "str h23, [x22], #0x2\n"
+      "str h24, [x21], #0x2\n"
+      "tbz x10, #0, 172f\n"
+      "st1 { v31.b }[2], [x11]\n"
+      "st1 { v8.b }[2], [x25]\n"
+      "st1 { v15.b }[2], [x24]\n"
+      "st1 { v16.b }[2], [x23]\n"
+      "st1 { v23.b }[2], [x22]\n"
+      "st1 { v24.b }[2], [x21]\n"
       "b 172f\n"
       "171:"  // Height 6: Partial direct writeback: partial_1_0
-      "str b31, [x9, #0x0]\n"
-      "str b8, [x26, #0x0]\n"
-      "str b15, [x25, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "str b23, [x23, #0x0]\n"
-      "str b24, [x22, #0x0]\n"
+      "str b31, [x11, #0x0]\n"
+      "str b8, [x25, #0x0]\n"
+      "str b15, [x24, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
+      "str b23, [x22, #0x0]\n"
+      "str b24, [x21, #0x0]\n"
       "172:"  // Height 6: Partial direct writeback: Done
       "b 174f\n"
       "173:"  // Height 6: Full writeback
-      "str q31, [x9, #0x0]\n"
-      "add x9, x9, #0x10\n"
-      "str q8, [x26, #0x0]\n"
-      "str q15, [x25, #0x0]\n"
-      "str q16, [x24, #0x0]\n"
-      "str q23, [x23, #0x0]\n"
-      "str q24, [x22, #0x0]\n"
+      "str q31, [x11, #0x0]\n"
+      "add x11, x11, #0x10\n"
+      "str q8, [x25, #0x0]\n"
+      "str q15, [x24, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q23, [x22, #0x0]\n"
+      "str q24, [x21, #0x0]\n"
       "174:"  // Height 6: Writeback done
-      "subs x11, x11, #0x10\n"
+      "subs x10, x10, #0x10\n"
       "bgt 147b\n"
       "subs %x[M], %x[M], #0x6\n"
       "beq 176f\n"
@@ -3617,8 +3616,8 @@ void a64_hybrid_s8qs_mmla_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "176:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16.hpp
index 47b6861f5b..e13f2fb5eb 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp
index 55629a38d0..f5545b4357 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp
@@ -44,18 +44,18 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -87,72 +87,72 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "bgt 69f\n"
       "beq 35f\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "tbz %x[flags], #0, 12f\n"
       "cmp x8, #0x10\n"
       "bge 11f\n"
       "tbz x8, #3, 6f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "tbz x8, #2, 4f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "tbz x8, #1, 3f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "tbz x8, #0, 10f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "b 10f\n"
       "3:"  // Height 1: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 10f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "b 10f\n"
       "4:"  // Height 1: Partial accumulate: partial_2_8
       "tbz x8, #1, 5f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "tbz x8, #0, 10f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "b 10f\n"
       "5:"  // Height 1: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 10f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "b 10f\n"
       "6:"  // Height 1: Partial accumulate: partial_4_0
       "tbz x8, #2, 8f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "tbz x8, #1, 7f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "tbz x8, #0, 10f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "b 10f\n"
       "7:"  // Height 1: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 10f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "b 10f\n"
       "8:"  // Height 1: Partial accumulate: partial_2_0
       "tbz x8, #1, 9f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "tbz x8, #0, 10f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "b 10f\n"
       "9:"  // Height 1: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "10:"  // Height 1: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 13f\n"
       "11:"  // Height 1: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "b 13f\n"
       "12:"  // Height 1: no accumulate
       "movi v8.4s, #0x0\n"
@@ -163,8 +163,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "14:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 15f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -180,118 +180,118 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "blt 19f\n"
       "ldr q0, [x13, #0x0]\n"
       "cmp x14, #0x20\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 18f\n"
       "17:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr d17, [x17, #0x20]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr d17, [x16, #0x20]\n"
+      "ldr x20, [x16, #0x28]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr d16, [x17, #0x30]\n"
-      "add x13, x13, #0x10\n"
-      "ldr x20, [x17, #0x38]\n"
-      "sub x14, x14, #0x10\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x48]\n"
-      "ldr x22, [x13, #0x8]\n"
-      "cmp x14, #0x20\n"
+      "ldr d16, [x16, #0x30]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0x38]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr d17, [x17, #0x40]\n"
+      "ldr d17, [x16, #0x40]\n"
+      "ldr x20, [x16, #0x48]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr d16, [x17, #0x50]\n"
-      "ldr x20, [x17, #0x58]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x68]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr d16, [x16, #0x50]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0x58]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr d17, [x17, #0x60]\n"
+      "ldr d17, [x16, #0x60]\n"
+      "ldr x20, [x16, #0x68]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr d16, [x17, #0x70]\n"
-      "ldr x20, [x17, #0x78]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x88]\n"
+      "ldr d16, [x16, #0x70]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0x78]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr d17, [x17, #0x80]\n"
+      "ldr d17, [x16, #0x80]\n"
+      "ldr x20, [x16, #0x88]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr d16, [x17, #0x90]\n"
-      "ldr x20, [x17, #0x98]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xa8]\n"
+      "ldr d16, [x16, #0x90]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0x98]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr d17, [x17, #0xa0]\n"
+      "ldr d17, [x16, #0xa0]\n"
+      "ldr x20, [x16, #0xa8]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr d16, [x17, #0xb0]\n"
-      "ldr x20, [x17, #0xb8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xc8]\n"
+      "ldr d16, [x16, #0xb0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0xb8]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr d17, [x17, #0xc0]\n"
+      "ldr d17, [x16, #0xc0]\n"
+      "ldr x20, [x16, #0xc8]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr d16, [x17, #0xd0]\n"
-      "ldr x20, [x17, #0xd8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xe8]\n"
+      "ldr d16, [x16, #0xd0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0xd8]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr d17, [x17, #0xe0]\n"
+      "ldr d17, [x16, #0xe0]\n"
+      "ldr x20, [x16, #0xe8]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr d16, [x17, #0xf0]\n"
-      "ldr x20, [x17, #0xf8]\n"
-      "add x17, x17, #0x100\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x8]\n"
+      "ldr d16, [x16, #0xf0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0xf8]\n"
       "mov v16.d[1], x20\n"
+      "add x13, x13, #0x10\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
+      "ldr x20, [x16, #0x8]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
-      "ldr x20, [x17, #0x18]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x22\n"
+      "sub x14, x14, #0x10\n"
+      "ldr d7, [x16, #0x10]\n"
+      "cmp x14, #0x20\n"
+      "ldr x21, [x13, #0x8]\n"
+      "mov v6.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
+      "mov v0.d[1], x21\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "bge 17b\n"
       "18:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q17, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
-      "add x13, x13, #0x10\n"
-      "sub x14, x14, #0x10\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x17, #0x40]\n"
+      "ldr q17, [x16, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x17, #0x50]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr q16, [x16, #0x50]\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x17, #0x60]\n"
+      "ldr q17, [x16, #0x60]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x17, #0x70]\n"
+      "ldr q16, [x16, #0x70]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x17, #0x80]\n"
+      "ldr q17, [x16, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x17, #0x90]\n"
+      "ldr q16, [x16, #0x90]\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x17, #0xa0]\n"
+      "ldr q17, [x16, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x17, #0xb0]\n"
+      "ldr q16, [x16, #0xb0]\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x17, #0xc0]\n"
+      "ldr q17, [x16, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x17, #0xd0]\n"
+      "ldr q16, [x16, #0xd0]\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr q17, [x17, #0xe0]\n"
+      "ldr q17, [x16, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr q16, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "ldr q16, [x16, #0xf0]\n"
+      "add x13, x13, #0x10\n"
+      "sub x14, x14, #0x10\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
+      "add x16, x16, #0x100\n"
       "19:"  // Height 1: Multiply loop: Main loop skip
       "cbz x14, 24f\n"
       "cmp x14, #0x4\n"
@@ -299,16 +299,16 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "20:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x13], #0x4\n"
       "sub x14, x14, #0x4\n"
-      "ldr q17, [x17, #0x0]\n"
-      "cmp x14, #0x4\n"
-      "ldr q16, [x17, #0x10]\n"
-      ".inst 0x4f92e228  // sdot v8.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q16, [x16, #0x0]\n"
+      ".inst 0x4f92e208  // sdot v8.4s, v16.16b, v18.4b[0]\n"
+      "ldr q16, [x16, #0x10]\n"
       ".inst 0x4f92e209  // sdot v9.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
-      "add x17, x17, #0x40\n"
+      "ldr q17, [x16, #0x20]\n"
+      "cmp x14, #0x4\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x4f92e22a  // sdot v10.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x4f92e20b  // sdot v11.4s, v16.16b, v18.4b[0]\n"
+      "add x16, x16, #0x40\n"
       "bge 20b\n"
       "21:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x14, 24f\n"
@@ -320,165 +320,165 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "22:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr b0, [x13, #0x0]\n"
       "23:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x17, #0x0]\n"
-      "ldr q16, [x17, #0x10]\n"
-      ".inst 0x4f80e228  // sdot v8.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q16, [x16, #0x0]\n"
+      ".inst 0x4f80e208  // sdot v8.4s, v16.16b, v0.4b[0]\n"
+      "ldr q16, [x16, #0x10]\n"
       ".inst 0x4f80e209  // sdot v9.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
-      "add x17, x17, #0x40\n"
-      ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
+      "ldr q16, [x16, #0x20]\n"
+      ".inst 0x4f80e20a  // sdot v10.4s, v16.16b, v0.4b[0]\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
+      "add x16, x16, #0x40\n"
       "24:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x15, x15, #0x1\n"
       "cmp x15, x20\n"
       "bne 14b\n"
       "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
       "bge 33f\n"
       "tbz x8, #3, 28f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "tbz x8, #2, 26f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "tbz x8, #1, 25f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "tbz x8, #0, 32f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "b 32f\n"
       "25:"  // Height 1: Partial direct writeback: partial_1_12
       "tbz x8, #0, 32f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "b 32f\n"
       "26:"  // Height 1: Partial direct writeback: partial_2_8
       "tbz x8, #1, 27f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "tbz x8, #0, 32f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "b 32f\n"
       "27:"  // Height 1: Partial direct writeback: partial_1_8
       "tbz x8, #0, 32f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "b 32f\n"
       "28:"  // Height 1: Partial direct writeback: partial_4_0
       "tbz x8, #2, 30f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "tbz x8, #1, 29f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "tbz x8, #0, 32f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "b 32f\n"
       "29:"  // Height 1: Partial direct writeback: partial_1_4
       "tbz x8, #0, 32f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "b 32f\n"
       "30:"  // Height 1: Partial direct writeback: partial_2_0
       "tbz x8, #1, 31f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "tbz x8, #0, 32f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "b 32f\n"
       "31:"  // Height 1: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "32:"  // Height 1: Partial direct writeback: Done
       "b 34f\n"
       "33:"  // Height 1: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "34:"  // Height 1: Writeback done
       "subs x8, x8, #0x10\n"
       "bgt 2b\n"
       "b 206f\n"
       "35:"  // Height 2
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "36:"  // Height 2: Column loop
       "tbz %x[flags], #0, 46f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "bge 45f\n"
       "tbz x8, #3, 40f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "tbz x8, #2, 38f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "tbz x8, #1, 37f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "tbz x8, #0, 44f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "b 44f\n"
       "37:"  // Height 2: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 44f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "b 44f\n"
       "38:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x8, #1, 39f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "tbz x8, #0, 44f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "b 44f\n"
       "39:"  // Height 2: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 44f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "b 44f\n"
       "40:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x8, #2, 42f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "tbz x8, #1, 41f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "tbz x8, #0, 44f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "b 44f\n"
       "41:"  // Height 2: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 44f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "b 44f\n"
       "42:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x8, #1, 43f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "tbz x8, #0, 44f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "b 44f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "44:"  // Height 2: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 47f\n"
       "45:"  // Height 2: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -497,8 +497,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "48:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 49f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -518,154 +518,154 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr q0, [x13, #0x0]\n"
       "cmp x14, #0x20\n"
       "ldr q1, [x12, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 52f\n"
       "51:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x20, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr d17, [x17, #0x20]\n"
+      "ldr d17, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr d16, [x17, #0x30]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x48]\n"
-      "add x13, x13, #0x10\n"
-      "add x12, x12, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x16, #0x30]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr d17, [x17, #0x40]\n"
+      "ldr d17, [x16, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x58]\n"
+      "ldr x20, [x16, #0x48]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr d16, [x17, #0x50]\n"
+      "ldr d16, [x16, #0x50]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x68]\n"
-      "ldr x23, [x13, #0x8]\n"
-      "sub x14, x14, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x16, #0x58]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x4fa1e22c  // sdot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr d17, [x17, #0x60]\n"
+      "ldr d17, [x16, #0x60]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr x21, [x17, #0x78]\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x4fa1e20d  // sdot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr d16, [x17, #0x70]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x88]\n"
-      "ldr x22, [x12, #0x8]\n"
-      "cmp x14, #0x20\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x16, #0x70]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4fa1e22e  // sdot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr d17, [x17, #0x80]\n"
+      "ldr d17, [x16, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr x21, [x17, #0x98]\n"
+      "ldr x20, [x16, #0x88]\n"
       ".inst 0x4fa1e20f  // sdot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr d16, [x17, #0x90]\n"
+      "ldr d16, [x16, #0x90]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xa8]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x16, #0x98]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x4f81ea2c  // sdot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr d17, [x17, #0xa0]\n"
+      "ldr d17, [x16, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr x21, [x17, #0xb8]\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x4f81ea0d  // sdot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr d16, [x17, #0xb0]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xc8]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x16, #0xb0]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4f81ea2e  // sdot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr d17, [x17, #0xc0]\n"
+      "ldr d17, [x16, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr x21, [x17, #0xd8]\n"
+      "ldr x20, [x16, #0xc8]\n"
       ".inst 0x4f81ea0f  // sdot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr d16, [x17, #0xd0]\n"
+      "ldr d16, [x16, #0xd0]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xe8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x16, #0xd8]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x4fa1ea2c  // sdot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr d17, [x17, #0xe0]\n"
+      "ldr d17, [x16, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr x21, [x17, #0xf8]\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x4fa1ea0d  // sdot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr d16, [x17, #0xf0]\n"
-      "mov v17.d[1], x20\n"
-      "add x17, x17, #0x100\n"
-      "ldr x20, [x17, #0x8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x16, #0xf0]\n"
+      "mov v17.d[1], x21\n"
+      "add x13, x13, #0x10\n"
+      "mov v16.d[1], x20\n"
+      "add x12, x12, #0x10\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea2e  // sdot v14.4s, v17.16b, v1.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
+      "ldr x21, [x16, #0x8]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x4fa1ea0f  // sdot v15.4s, v16.16b, v1.4b[3]\n"
       "ldr d1, [x12, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x20\n"
-      "ldr x20, [x17, #0x18]\n"
-      "mov v0.d[1], x23\n"
-      "mov v1.d[1], x22\n"
+      "sub x14, x14, #0x10\n"
+      "ldr d7, [x16, #0x10]\n"
+      "cmp x14, #0x20\n"
+      "ldr x20, [x13, #0x8]\n"
+      "mov v6.d[1], x21\n"
+      "ldr x21, [x12, #0x8]\n"
+      "mov v0.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
+      "mov v1.d[1], x21\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "bge 51b\n"
       "52:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
       "add x13, x13, #0x10\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q17, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       "sub x14, x14, #0x10\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x17, #0x40]\n"
+      "ldr q17, [x16, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x17, #0x50]\n"
+      "ldr q16, [x16, #0x50]\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4fa1e22c  // sdot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x17, #0x60]\n"
+      "ldr q17, [x16, #0x60]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20d  // sdot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x17, #0x70]\n"
+      "ldr q16, [x16, #0x70]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22e  // sdot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x17, #0x80]\n"
+      "ldr q17, [x16, #0x80]\n"
       ".inst 0x4fa0e20b  // sdot v11.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20f  // sdot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x17, #0x90]\n"
+      "ldr q16, [x16, #0x90]\n"
       ".inst 0x4f80ea28  // sdot v8.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x4f81ea2c  // sdot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x17, #0xa0]\n"
+      "ldr q17, [x16, #0xa0]\n"
       ".inst 0x4f80ea09  // sdot v9.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x4f81ea0d  // sdot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x17, #0xb0]\n"
+      "ldr q16, [x16, #0xb0]\n"
       ".inst 0x4f80ea2a  // sdot v10.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x4f81ea2e  // sdot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x17, #0xc0]\n"
+      "ldr q17, [x16, #0xc0]\n"
       ".inst 0x4f80ea0b  // sdot v11.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x4f81ea0f  // sdot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x17, #0xd0]\n"
+      "ldr q16, [x16, #0xd0]\n"
       ".inst 0x4fa0ea28  // sdot v8.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea2c  // sdot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr q17, [x17, #0xe0]\n"
+      "ldr q17, [x16, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea0d  // sdot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr q16, [x17, #0xf0]\n"
+      "ldr q16, [x16, #0xf0]\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa1ea2e  // sdot v14.4s, v17.16b, v1.4b[3]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea0f  // sdot v15.4s, v16.16b, v1.4b[3]\n"
@@ -678,16 +678,16 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "sub x14, x14, #0x4\n"
       "ldr s18, [x12], #0x4\n"
       "cmp x14, #0x4\n"
-      "ldr q17, [x17, #0x0]\n"
-      "ldr q16, [x17, #0x10]\n"
+      "ldr q17, [x16, #0x0]\n"
       ".inst 0x4f93e228  // sdot v8.4s, v17.16b, v19.4b[0]\n"
+      "ldr q16, [x16, #0x10]\n"
       ".inst 0x4f92e22c  // sdot v12.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q17, [x16, #0x20]\n"
       ".inst 0x4f93e209  // sdot v9.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x4f92e20d  // sdot v13.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x4f93e22a  // sdot v10.4s, v17.16b, v19.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f92e22e  // sdot v14.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x4f93e20b  // sdot v11.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x4f92e20f  // sdot v15.4s, v16.16b, v18.4b[0]\n"
@@ -705,16 +705,16 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr b0, [x13, #0x0]\n"
       "ldr b1, [x12, #0x0]\n"
       "57:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x17, #0x0]\n"
-      "ldr q16, [x17, #0x10]\n"
+      "ldr q17, [x16, #0x0]\n"
       ".inst 0x4f80e228  // sdot v8.4s, v17.16b, v0.4b[0]\n"
+      "ldr q16, [x16, #0x10]\n"
       ".inst 0x4f81e22c  // sdot v12.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q17, [x16, #0x20]\n"
       ".inst 0x4f80e209  // sdot v9.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20d  // sdot v13.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
@@ -724,79 +724,79 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 48b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x17, x20, LSL #2\n"
       "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "bge 67f\n"
       "tbz x8, #3, 62f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "tbz x8, #2, 60f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "tbz x8, #1, 59f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "tbz x8, #0, 66f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "b 66f\n"
       "59:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x8, #0, 66f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "b 66f\n"
       "60:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x8, #1, 61f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "tbz x8, #0, 66f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "b 66f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x8, #0, 66f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "b 66f\n"
       "62:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x8, #2, 64f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "tbz x8, #1, 63f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "tbz x8, #0, 66f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "b 66f\n"
       "63:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x8, #0, 66f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "b 66f\n"
       "64:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x8, #1, 65f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "tbz x8, #0, 66f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "b 66f\n"
       "65:"  // Height 2: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "66:"  // Height 2: Partial direct writeback: Done
       "b 68f\n"
       "67:"  // Height 2: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -807,107 +807,107 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 206f\n"
       "69:"  // Height 3
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "70:"  // Height 3: Column loop
       "tbz %x[flags], #0, 80f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x17, x20, LSL #2\n"
       "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "bge 79f\n"
       "tbz x8, #3, 74f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "ld1 { v17.4s }, [x23], #0x10\n"
       "tbz x8, #2, 72f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "ld1 { v18.4s }, [x23], #0x10\n"
       "tbz x8, #1, 71f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "ldr d19, [x23], #0x8\n"
       "tbz x8, #0, 78f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "ld1 { v19.s }[2], [x23]\n"
       "b 78f\n"
       "71:"  // Height 3: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 78f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "ldr s19, [x23, #0x0]\n"
       "b 78f\n"
       "72:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x8, #1, 73f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "ldr d18, [x23], #0x8\n"
       "tbz x8, #0, 78f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "ld1 { v18.s }[2], [x23]\n"
       "b 78f\n"
       "73:"  // Height 3: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 78f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "ldr s18, [x23, #0x0]\n"
       "b 78f\n"
       "74:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x8, #2, 76f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "tbz x8, #1, 75f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "ldr d17, [x23], #0x8\n"
       "tbz x8, #0, 78f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "ld1 { v17.s }[2], [x23]\n"
       "b 78f\n"
       "75:"  // Height 3: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 78f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "ldr s17, [x23, #0x0]\n"
       "b 78f\n"
       "76:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x8, #1, 77f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "ldr d16, [x23], #0x8\n"
       "tbz x8, #0, 78f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "ld1 { v16.s }[2], [x23]\n"
       "b 78f\n"
       "77:"  // Height 3: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "ldr s16, [x23, #0x0]\n"
       "78:"  // Height 3: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 81f\n"
       "79:"  // Height 3: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -934,8 +934,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "82:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 83f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -959,123 +959,123 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "cmp x14, #0x20\n"
       "ldr q1, [x12, #0x0]\n"
       "ldr q2, [x11, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 86f\n"
       "85:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr d21, [x17, #0x20]\n"
+      "ldr d21, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x13, x13, #0x10\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x12, x12, #0x10\n"
+      "ldr x21, [x16, #0x48]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr d20, [x17, #0x30]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x48]\n"
-      "add x11, x11, #0x10\n"
-      "ldr x24, [x13, #0x8]\n"
+      "ldr d20, [x16, #0x30]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x20, [x16, #0x58]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr d21, [x17, #0x40]\n"
+      "ldr d21, [x16, #0x40]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
-      "ldr x23, [x12, #0x8]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4f81e28f  // sdot v15.4s, v20.16b, v1.4b[0]\n"
-      "ldr x22, [x11, #0x8]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x4f82e293  // sdot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr d20, [x17, #0x50]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x68]\n"
-      "sub x14, x14, #0x10\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr d20, [x16, #0x50]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4fa0e2a8  // sdot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ac  // sdot v12.4s, v21.16b, v1.4b[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x4fa2e2b0  // sdot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr d21, [x17, #0x60]\n"
+      "ldr d21, [x16, #0x60]\n"
       ".inst 0x4fa0e289  // sdot v9.4s, v20.16b, v0.4b[1]\n"
-      "cmp x14, #0x20\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4fa1e28d  // sdot v13.4s, v20.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "ldr x21, [x16, #0x88]\n"
       ".inst 0x4fa2e291  // sdot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr d20, [x17, #0x70]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x88]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr d20, [x16, #0x70]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4fa0e2aa  // sdot v10.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ae  // sdot v14.4s, v21.16b, v1.4b[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "ldr x20, [x16, #0x98]\n"
       ".inst 0x4fa2e2b2  // sdot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr d21, [x17, #0x80]\n"
+      "ldr d21, [x16, #0x80]\n"
       ".inst 0x4fa0e28b  // sdot v11.4s, v20.16b, v0.4b[1]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4fa1e28f  // sdot v15.4s, v20.16b, v1.4b[1]\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x4fa2e293  // sdot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr d20, [x17, #0x90]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xa8]\n"
+      "ldr d20, [x16, #0x90]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4f80eaa8  // sdot v8.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaac  // sdot v12.4s, v21.16b, v1.4b[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x4f82eab0  // sdot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr d21, [x17, #0xa0]\n"
+      "ldr d21, [x16, #0xa0]\n"
       ".inst 0x4f80ea89  // sdot v9.4s, v20.16b, v0.4b[2]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4f81ea8d  // sdot v13.4s, v20.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x4f82ea91  // sdot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr d20, [x17, #0xb0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xc8]\n"
+      "ldr d20, [x16, #0xb0]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4f80eaaa  // sdot v10.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaae  // sdot v14.4s, v21.16b, v1.4b[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
+      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x4f82eab2  // sdot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr d21, [x17, #0xc0]\n"
+      "ldr d21, [x16, #0xc0]\n"
       ".inst 0x4f80ea8b  // sdot v11.4s, v20.16b, v0.4b[2]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4f81ea8f  // sdot v15.4s, v20.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x4f82ea93  // sdot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr d20, [x17, #0xd0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xe8]\n"
+      "ldr d20, [x16, #0xd0]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x4fa0eaa8  // sdot v8.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa1eaac  // sdot v12.4s, v21.16b, v1.4b[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x4fa2eab0  // sdot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr d21, [x17, #0xe0]\n"
+      "ldr d21, [x16, #0xe0]\n"
       ".inst 0x4fa0ea89  // sdot v9.4s, v20.16b, v0.4b[3]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x4fa1ea8d  // sdot v13.4s, v20.16b, v1.4b[3]\n"
+      "add x13, x13, #0x10\n"
       ".inst 0x4fa2ea91  // sdot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr d20, [x17, #0xf0]\n"
-      "mov v21.d[1], x21\n"
-      "add x17, x17, #0x100\n"
-      "ldr x21, [x17, #0x8]\n"
+      "ldr d20, [x16, #0xf0]\n"
       "mov v20.d[1], x20\n"
+      "add x12, x12, #0x10\n"
+      "add x11, x11, #0x10\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa0eaaa  // sdot v10.4s, v21.16b, v0.4b[3]\n"
+      "ldr x20, [x16, #0x8]\n"
       ".inst 0x4fa1eaae  // sdot v14.4s, v21.16b, v1.4b[3]\n"
-      "ldr x20, [x17, #0x18]\n"
+      "ldr x23, [x13, #0x8]\n"
       ".inst 0x4fa2eab2  // sdot v18.4s, v21.16b, v2.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
       ".inst 0x4fa0ea8b  // sdot v11.4s, v20.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x4fa1ea8f  // sdot v15.4s, v20.16b, v1.4b[3]\n"
       "ldr d1, [x12, #0x0]\n"
+      "ldr x22, [x12, #0x8]\n"
       ".inst 0x4fa2ea93  // sdot v19.4s, v20.16b, v2.4b[3]\n"
       "ldr d2, [x11, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x24\n"
-      "mov v1.d[1], x23\n"
-      "mov v2.d[1], x22\n"
+      "sub x14, x14, #0x10\n"
+      "ldr d7, [x16, #0x10]\n"
+      "cmp x14, #0x20\n"
+      "ldr x21, [x11, #0x8]\n"
+      "mov v6.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
+      "mov v0.d[1], x23\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
+      "mov v1.d[1], x22\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v2.d[1], x21\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "mov v7.d[1], x20\n"
       "bge 85b\n"
       "86:"  // Height 3: Multiply loop: Single iteration only
@@ -1084,66 +1084,66 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr q21, [x17, #0x20]\n"
+      "ldr q21, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       "sub x14, x14, #0x10\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr q20, [x17, #0x30]\n"
+      "ldr q20, [x16, #0x30]\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x17, #0x40]\n"
+      "ldr q21, [x16, #0x40]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4f81e28f  // sdot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e293  // sdot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x17, #0x50]\n"
+      "ldr q20, [x16, #0x50]\n"
       ".inst 0x4fa0e2a8  // sdot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ac  // sdot v12.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b0  // sdot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x17, #0x60]\n"
+      "ldr q21, [x16, #0x60]\n"
       ".inst 0x4fa0e289  // sdot v9.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x4fa1e28d  // sdot v13.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x4fa2e291  // sdot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x17, #0x70]\n"
+      "ldr q20, [x16, #0x70]\n"
       ".inst 0x4fa0e2aa  // sdot v10.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ae  // sdot v14.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b2  // sdot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x17, #0x80]\n"
+      "ldr q21, [x16, #0x80]\n"
       ".inst 0x4fa0e28b  // sdot v11.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x4fa1e28f  // sdot v15.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x4fa2e293  // sdot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x17, #0x90]\n"
+      "ldr q20, [x16, #0x90]\n"
       ".inst 0x4f80eaa8  // sdot v8.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaac  // sdot v12.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x4f82eab0  // sdot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x17, #0xa0]\n"
+      "ldr q21, [x16, #0xa0]\n"
       ".inst 0x4f80ea89  // sdot v9.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x4f81ea8d  // sdot v13.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x4f82ea91  // sdot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x17, #0xb0]\n"
+      "ldr q20, [x16, #0xb0]\n"
       ".inst 0x4f80eaaa  // sdot v10.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x4f81eaae  // sdot v14.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x4f82eab2  // sdot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x17, #0xc0]\n"
+      "ldr q21, [x16, #0xc0]\n"
       ".inst 0x4f80ea8b  // sdot v11.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x4f81ea8f  // sdot v15.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x4f82ea93  // sdot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x17, #0xd0]\n"
+      "ldr q20, [x16, #0xd0]\n"
       ".inst 0x4fa0eaa8  // sdot v8.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x4fa1eaac  // sdot v12.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x4fa2eab0  // sdot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr q21, [x17, #0xe0]\n"
+      "ldr q21, [x16, #0xe0]\n"
       ".inst 0x4fa0ea89  // sdot v9.4s, v20.16b, v0.4b[3]\n"
       ".inst 0x4fa1ea8d  // sdot v13.4s, v20.16b, v1.4b[3]\n"
       ".inst 0x4fa2ea91  // sdot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr q20, [x17, #0xf0]\n"
+      "ldr q20, [x16, #0xf0]\n"
       ".inst 0x4fa0eaaa  // sdot v10.4s, v21.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa1eaae  // sdot v14.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x4fa2eab2  // sdot v18.4s, v21.16b, v2.4b[3]\n"
       ".inst 0x4fa0ea8b  // sdot v11.4s, v20.16b, v0.4b[3]\n"
@@ -1159,18 +1159,18 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr s23, [x12], #0x4\n"
       "cmp x14, #0x4\n"
       "ldr s22, [x11], #0x4\n"
-      "ldr q21, [x17, #0x0]\n"
-      "ldr q20, [x17, #0x10]\n"
+      "ldr q21, [x16, #0x0]\n"
       ".inst 0x4f98e2a8  // sdot v8.4s, v21.16b, v24.4b[0]\n"
+      "ldr q20, [x16, #0x10]\n"
       ".inst 0x4f97e2ac  // sdot v12.4s, v21.16b, v23.4b[0]\n"
       ".inst 0x4f96e2b0  // sdot v16.4s, v21.16b, v22.4b[0]\n"
-      "ldr q21, [x17, #0x20]\n"
+      "ldr q21, [x16, #0x20]\n"
       ".inst 0x4f98e289  // sdot v9.4s, v20.16b, v24.4b[0]\n"
       ".inst 0x4f97e28d  // sdot v13.4s, v20.16b, v23.4b[0]\n"
       ".inst 0x4f96e291  // sdot v17.4s, v20.16b, v22.4b[0]\n"
-      "ldr q20, [x17, #0x30]\n"
+      "ldr q20, [x16, #0x30]\n"
       ".inst 0x4f98e2aa  // sdot v10.4s, v21.16b, v24.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f97e2ae  // sdot v14.4s, v21.16b, v23.4b[0]\n"
       ".inst 0x4f96e2b2  // sdot v18.4s, v21.16b, v22.4b[0]\n"
       ".inst 0x4f98e28b  // sdot v11.4s, v20.16b, v24.4b[0]\n"
@@ -1193,18 +1193,18 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr b1, [x12, #0x0]\n"
       "ldr b2, [x11, #0x0]\n"
       "91:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "ldr q21, [x17, #0x0]\n"
-      "ldr q20, [x17, #0x10]\n"
+      "ldr q21, [x16, #0x0]\n"
       ".inst 0x4f80e2a8  // sdot v8.4s, v21.16b, v0.4b[0]\n"
+      "ldr q20, [x16, #0x10]\n"
       ".inst 0x4f81e2ac  // sdot v12.4s, v21.16b, v1.4b[0]\n"
       ".inst 0x4f82e2b0  // sdot v16.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x17, #0x20]\n"
+      "ldr q21, [x16, #0x20]\n"
       ".inst 0x4f80e289  // sdot v9.4s, v20.16b, v0.4b[0]\n"
       ".inst 0x4f81e28d  // sdot v13.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e291  // sdot v17.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x17, #0x30]\n"
+      "ldr q20, [x16, #0x30]\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
@@ -1216,97 +1216,97 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 82b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "bge 101f\n"
       "tbz x8, #3, 96f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "st1 { v17.4s }, [x23], #0x10\n"
       "tbz x8, #2, 94f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "st1 { v18.4s }, [x23], #0x10\n"
       "tbz x8, #1, 93f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "str d19, [x23], #0x8\n"
       "tbz x8, #0, 100f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "st1 { v19.s }[2], [x23]\n"
       "b 100f\n"
       "93:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x8, #0, 100f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "str s19, [x23, #0x0]\n"
       "b 100f\n"
       "94:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x8, #1, 95f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "str d18, [x23], #0x8\n"
       "tbz x8, #0, 100f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "st1 { v18.s }[2], [x23]\n"
       "b 100f\n"
       "95:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x8, #0, 100f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "str s18, [x23, #0x0]\n"
       "b 100f\n"
       "96:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x8, #2, 98f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "tbz x8, #1, 97f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "str d17, [x23], #0x8\n"
       "tbz x8, #0, 100f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "st1 { v17.s }[2], [x23]\n"
       "b 100f\n"
       "97:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x8, #0, 100f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "str s17, [x23, #0x0]\n"
       "b 100f\n"
       "98:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x8, #1, 99f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "str d16, [x23], #0x8\n"
       "tbz x8, #0, 100f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "st1 { v16.s }[2], [x23]\n"
       "b 100f\n"
       "99:"  // Height 3: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "str s16, [x23, #0x0]\n"
       "100:"  // Height 3: Partial direct writeback: Done
       "b 102f\n"
       "101:"  // Height 3: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -1321,38 +1321,38 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 206f\n"
       "103:"  // Height 4
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "104:"  // Height 4: Column loop
       "tbz %x[flags], #0, 114f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
       "add x22, x23, x20, LSL #2\n"
       "bge 113f\n"
       "tbz x8, #3, 108f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "ld1 { v17.4s }, [x23], #0x10\n"
       "ld1 { v21.4s }, [x22], #0x10\n"
       "tbz x8, #2, 106f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "ld1 { v18.4s }, [x23], #0x10\n"
       "ld1 { v22.4s }, [x22], #0x10\n"
       "tbz x8, #1, 105f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "ldr d19, [x23], #0x8\n"
       "ldr d23, [x22], #0x8\n"
       "tbz x8, #0, 112f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "ld1 { v19.s }[2], [x23]\n"
       "ld1 { v23.s }[2], [x22]\n"
@@ -1360,20 +1360,20 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "105:"  // Height 4: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 112f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "ldr s19, [x23, #0x0]\n"
       "ldr s23, [x22, #0x0]\n"
       "b 112f\n"
       "106:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x8, #1, 107f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "ldr d18, [x23], #0x8\n"
       "ldr d22, [x22], #0x8\n"
       "tbz x8, #0, 112f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "ld1 { v18.s }[2], [x23]\n"
       "ld1 { v22.s }[2], [x22]\n"
@@ -1381,25 +1381,25 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "107:"  // Height 4: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 112f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "ldr s18, [x23, #0x0]\n"
       "ldr s22, [x22, #0x0]\n"
       "b 112f\n"
       "108:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x8, #2, 110f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "tbz x8, #1, 109f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "ldr d17, [x23], #0x8\n"
       "ldr d21, [x22], #0x8\n"
       "tbz x8, #0, 112f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "ld1 { v17.s }[2], [x23]\n"
       "ld1 { v21.s }[2], [x22]\n"
@@ -1407,38 +1407,38 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "109:"  // Height 4: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 112f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "ldr s17, [x23, #0x0]\n"
       "ldr s21, [x22, #0x0]\n"
       "b 112f\n"
       "110:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x8, #1, 111f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "ldr d16, [x23], #0x8\n"
       "ldr d20, [x22], #0x8\n"
       "tbz x8, #0, 112f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "ld1 { v16.s }[2], [x23]\n"
       "ld1 { v20.s }[2], [x22]\n"
       "b 112f\n"
       "111:"  // Height 4: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "ldr s16, [x23, #0x0]\n"
       "ldr s20, [x22, #0x0]\n"
       "112:"  // Height 4: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 115f\n"
       "113:"  // Height 4: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -1473,8 +1473,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "116:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 117f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1502,129 +1502,130 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr q1, [x12, #0x0]\n"
       "ldr q2, [x11, #0x0]\n"
       "ldr q3, [x10, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 120f\n"
       "119:"  // Height 4: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x20, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x21, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       "add x13, x13, #0x10\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr d25, [x17, #0x20]\n"
+      "ldr d25, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x12, x12, #0x10\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x11, x11, #0x10\n"
+      "ldr x21, [x16, #0x48]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x12, x12, #0x10\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr d24, [x17, #0x30]\n"
+      "ldr d24, [x16, #0x30]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
-      "ldr x20, [x17, #0x48]\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0x58]\n"
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
+      "add x11, x11, #0x10\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr d25, [x17, #0x40]\n"
+      "ldr d25, [x16, #0x40]\n"
       ".inst 0x4f80e30b  // sdot v11.4s, v24.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x58]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4f81e30f  // sdot v15.4s, v24.16b, v1.4b[0]\n"
-      "ldr x25, [x13, #0x8]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x4f82e313  // sdot v19.4s, v24.16b, v2.4b[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x10, x10, #0x10\n"
       ".inst 0x4f83e317  // sdot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr d24, [x17, #0x50]\n"
+      "ldr d24, [x16, #0x50]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4fa0e328  // sdot v8.4s, v25.16b, v0.4b[1]\n"
-      "ldr x20, [x17, #0x68]\n"
       ".inst 0x4fa1e32c  // sdot v12.4s, v25.16b, v1.4b[1]\n"
-      "ldr x24, [x12, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x4fa2e330  // sdot v16.4s, v25.16b, v2.4b[1]\n"
+      "ldr x25, [x13, #0x8]\n"
       ".inst 0x4fa3e334  // sdot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr d25, [x17, #0x60]\n"
+      "ldr d25, [x16, #0x60]\n"
       ".inst 0x4fa0e309  // sdot v9.4s, v24.16b, v0.4b[1]\n"
-      "ldr x21, [x17, #0x78]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4fa1e30d  // sdot v13.4s, v24.16b, v1.4b[1]\n"
-      "ldr x23, [x11, #0x8]\n"
+      "ldr x21, [x16, #0x88]\n"
       ".inst 0x4fa2e311  // sdot v17.4s, v24.16b, v2.4b[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x24, [x12, #0x8]\n"
       ".inst 0x4fa3e315  // sdot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr d24, [x17, #0x70]\n"
+      "ldr d24, [x16, #0x70]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4fa0e32a  // sdot v10.4s, v25.16b, v0.4b[1]\n"
-      "ldr x20, [x17, #0x88]\n"
       ".inst 0x4fa1e32e  // sdot v14.4s, v25.16b, v1.4b[1]\n"
-      "ldr x22, [x10, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0x98]\n"
       ".inst 0x4fa2e332  // sdot v18.4s, v25.16b, v2.4b[1]\n"
+      "ldr x23, [x11, #0x8]\n"
       ".inst 0x4fa3e336  // sdot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr d25, [x17, #0x80]\n"
+      "ldr d25, [x16, #0x80]\n"
       ".inst 0x4fa0e30b  // sdot v11.4s, v24.16b, v0.4b[1]\n"
-      "ldr x21, [x17, #0x98]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4fa1e30f  // sdot v15.4s, v24.16b, v1.4b[1]\n"
-      "sub x14, x14, #0x10\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x4fa2e313  // sdot v19.4s, v24.16b, v2.4b[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x22, [x10, #0x8]\n"
       ".inst 0x4fa3e317  // sdot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr d24, [x17, #0x90]\n"
+      "ldr d24, [x16, #0x90]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4f80eb28  // sdot v8.4s, v25.16b, v0.4b[2]\n"
-      "ldr x20, [x17, #0xa8]\n"
       ".inst 0x4f81eb2c  // sdot v12.4s, v25.16b, v1.4b[2]\n"
-      "cmp x14, #0x20\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x4f82eb30  // sdot v16.4s, v25.16b, v2.4b[2]\n"
+      "sub x14, x14, #0x10\n"
       ".inst 0x4f83eb34  // sdot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr d25, [x17, #0xa0]\n"
+      "ldr d25, [x16, #0xa0]\n"
       ".inst 0x4f80eb09  // sdot v9.4s, v24.16b, v0.4b[2]\n"
-      "ldr x21, [x17, #0xb8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4f81eb0d  // sdot v13.4s, v24.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x4f82eb11  // sdot v17.4s, v24.16b, v2.4b[2]\n"
-      "mov v25.d[1], x20\n"
+      "cmp x14, #0x20\n"
       ".inst 0x4f83eb15  // sdot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr d24, [x17, #0xb0]\n"
+      "ldr d24, [x16, #0xb0]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4f80eb2a  // sdot v10.4s, v25.16b, v0.4b[2]\n"
-      "ldr x20, [x17, #0xc8]\n"
       ".inst 0x4f81eb2e  // sdot v14.4s, v25.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x4f82eb32  // sdot v18.4s, v25.16b, v2.4b[2]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x4f83eb36  // sdot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr d25, [x17, #0xc0]\n"
+      "ldr d25, [x16, #0xc0]\n"
       ".inst 0x4f80eb0b  // sdot v11.4s, v24.16b, v0.4b[2]\n"
-      "ldr x21, [x17, #0xd8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4f81eb0f  // sdot v15.4s, v24.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x4f82eb13  // sdot v19.4s, v24.16b, v2.4b[2]\n"
-      "mov v25.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4f83eb17  // sdot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr d24, [x17, #0xd0]\n"
+      "ldr d24, [x16, #0xd0]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x4fa0eb28  // sdot v8.4s, v25.16b, v0.4b[3]\n"
-      "ldr x20, [x17, #0xe8]\n"
       ".inst 0x4fa1eb2c  // sdot v12.4s, v25.16b, v1.4b[3]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x4fa2eb30  // sdot v16.4s, v25.16b, v2.4b[3]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4fa3eb34  // sdot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr d25, [x17, #0xe0]\n"
+      "ldr d25, [x16, #0xe0]\n"
       ".inst 0x4fa0eb09  // sdot v9.4s, v24.16b, v0.4b[3]\n"
-      "ldr x21, [x17, #0xf8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x4fa1eb0d  // sdot v13.4s, v24.16b, v1.4b[3]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4fa2eb11  // sdot v17.4s, v24.16b, v2.4b[3]\n"
-      "mov v25.d[1], x20\n"
       ".inst 0x4fa3eb15  // sdot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr d24, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "ldr d24, [x16, #0xf0]\n"
+      "mov v24.d[1], x20\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa0eb2a  // sdot v10.4s, v25.16b, v0.4b[3]\n"
+      "ldr x21, [x16, #0x8]\n"
       ".inst 0x4fa1eb2e  // sdot v14.4s, v25.16b, v1.4b[3]\n"
-      "ldr x20, [x17, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0x18]\n"
       ".inst 0x4fa2eb32  // sdot v18.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb36  // sdot v22.4s, v25.16b, v3.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
       ".inst 0x4fa0eb0b  // sdot v11.4s, v24.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x4fa1eb0f  // sdot v15.4s, v24.16b, v1.4b[3]\n"
@@ -1633,9 +1634,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr d2, [x11, #0x0]\n"
       ".inst 0x4fa3eb17  // sdot v23.4s, v24.16b, v3.4b[3]\n"
       "ldr d3, [x10, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x20\n"
-      "ldr x20, [x17, #0x18]\n"
+      "ldr d7, [x16, #0x10]\n"
+      "mov v6.d[1], x21\n"
       "mov v0.d[1], x25\n"
       "mov v1.d[1], x24\n"
       "mov v2.d[1], x23\n"
@@ -1650,7 +1650,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr q25, [x17, #0x20]\n"
+      "ldr q25, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x10, x10, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -1658,7 +1658,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr q24, [x17, #0x30]\n"
+      "ldr q24, [x16, #0x30]\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
@@ -1666,64 +1666,64 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x17, #0x40]\n"
+      "ldr q25, [x16, #0x40]\n"
       ".inst 0x4f80e30b  // sdot v11.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e30f  // sdot v15.4s, v24.16b, v1.4b[0]\n"
       ".inst 0x4f82e313  // sdot v19.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x4f83e317  // sdot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x17, #0x50]\n"
+      "ldr q24, [x16, #0x50]\n"
       ".inst 0x4fa0e328  // sdot v8.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e32c  // sdot v12.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x4fa2e330  // sdot v16.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x4fa3e334  // sdot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x17, #0x60]\n"
+      "ldr q25, [x16, #0x60]\n"
       ".inst 0x4fa0e309  // sdot v9.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x4fa1e30d  // sdot v13.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x4fa2e311  // sdot v17.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x4fa3e315  // sdot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x17, #0x70]\n"
+      "ldr q24, [x16, #0x70]\n"
       ".inst 0x4fa0e32a  // sdot v10.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x4fa1e32e  // sdot v14.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x4fa2e332  // sdot v18.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x4fa3e336  // sdot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x17, #0x80]\n"
+      "ldr q25, [x16, #0x80]\n"
       ".inst 0x4fa0e30b  // sdot v11.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x4fa1e30f  // sdot v15.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x4fa2e313  // sdot v19.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x4fa3e317  // sdot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x17, #0x90]\n"
+      "ldr q24, [x16, #0x90]\n"
       ".inst 0x4f80eb28  // sdot v8.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f81eb2c  // sdot v12.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x4f82eb30  // sdot v16.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x4f83eb34  // sdot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x17, #0xa0]\n"
+      "ldr q25, [x16, #0xa0]\n"
       ".inst 0x4f80eb09  // sdot v9.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb0d  // sdot v13.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x4f82eb11  // sdot v17.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x4f83eb15  // sdot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x17, #0xb0]\n"
+      "ldr q24, [x16, #0xb0]\n"
       ".inst 0x4f80eb2a  // sdot v10.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x4f81eb2e  // sdot v14.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x4f82eb32  // sdot v18.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x4f83eb36  // sdot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x17, #0xc0]\n"
+      "ldr q25, [x16, #0xc0]\n"
       ".inst 0x4f80eb0b  // sdot v11.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x4f81eb0f  // sdot v15.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x4f82eb13  // sdot v19.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x4f83eb17  // sdot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x17, #0xd0]\n"
+      "ldr q24, [x16, #0xd0]\n"
       ".inst 0x4fa0eb28  // sdot v8.4s, v25.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb2c  // sdot v12.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb30  // sdot v16.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb34  // sdot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr q25, [x17, #0xe0]\n"
+      "ldr q25, [x16, #0xe0]\n"
       ".inst 0x4fa0eb09  // sdot v9.4s, v24.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb0d  // sdot v13.4s, v24.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb11  // sdot v17.4s, v24.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb15  // sdot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr q24, [x17, #0xf0]\n"
+      "ldr q24, [x16, #0xf0]\n"
       ".inst 0x4fa0eb2a  // sdot v10.4s, v25.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa1eb2e  // sdot v14.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb32  // sdot v18.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb36  // sdot v22.4s, v25.16b, v3.4b[3]\n"
@@ -1742,20 +1742,20 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "cmp x14, #0x4\n"
       "ldr s27, [x11], #0x4\n"
       "ldr s26, [x10], #0x4\n"
-      "ldr q25, [x17, #0x0]\n"
-      "ldr q24, [x17, #0x10]\n"
+      "ldr q25, [x16, #0x0]\n"
       ".inst 0x4f9de328  // sdot v8.4s, v25.16b, v29.4b[0]\n"
+      "ldr q24, [x16, #0x10]\n"
       ".inst 0x4f9ce32c  // sdot v12.4s, v25.16b, v28.4b[0]\n"
       ".inst 0x4f9be330  // sdot v16.4s, v25.16b, v27.4b[0]\n"
       ".inst 0x4f9ae334  // sdot v20.4s, v25.16b, v26.4b[0]\n"
-      "ldr q25, [x17, #0x20]\n"
+      "ldr q25, [x16, #0x20]\n"
       ".inst 0x4f9de309  // sdot v9.4s, v24.16b, v29.4b[0]\n"
       ".inst 0x4f9ce30d  // sdot v13.4s, v24.16b, v28.4b[0]\n"
       ".inst 0x4f9be311  // sdot v17.4s, v24.16b, v27.4b[0]\n"
       ".inst 0x4f9ae315  // sdot v21.4s, v24.16b, v26.4b[0]\n"
-      "ldr q24, [x17, #0x30]\n"
+      "ldr q24, [x16, #0x30]\n"
       ".inst 0x4f9de32a  // sdot v10.4s, v25.16b, v29.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f9ce32e  // sdot v14.4s, v25.16b, v28.4b[0]\n"
       ".inst 0x4f9be332  // sdot v18.4s, v25.16b, v27.4b[0]\n"
       ".inst 0x4f9ae336  // sdot v22.4s, v25.16b, v26.4b[0]\n"
@@ -1783,20 +1783,20 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr b2, [x11, #0x0]\n"
       "ldr b3, [x10, #0x0]\n"
       "125:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "ldr q25, [x17, #0x0]\n"
-      "ldr q24, [x17, #0x10]\n"
+      "ldr q25, [x16, #0x0]\n"
       ".inst 0x4f80e328  // sdot v8.4s, v25.16b, v0.4b[0]\n"
+      "ldr q24, [x16, #0x10]\n"
       ".inst 0x4f81e32c  // sdot v12.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x4f82e330  // sdot v16.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x4f83e334  // sdot v20.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x17, #0x20]\n"
+      "ldr q25, [x16, #0x20]\n"
       ".inst 0x4f80e309  // sdot v9.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x4f81e30d  // sdot v13.4s, v24.16b, v1.4b[0]\n"
       ".inst 0x4f82e311  // sdot v17.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x4f83e315  // sdot v21.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x17, #0x30]\n"
+      "ldr q24, [x16, #0x30]\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
@@ -1810,18 +1810,18 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 116b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
       "bge 135f\n"
       "tbz x8, #3, 130f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
@@ -1829,96 +1829,96 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "st1 { v20.4s }, [x22], #0x10\n"
       "st1 { v21.4s }, [x22], #0x10\n"
       "tbz x8, #2, 128f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "st1 { v18.4s }, [x23], #0x10\n"
       "st1 { v22.4s }, [x22], #0x10\n"
       "tbz x8, #1, 127f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "str d19, [x23], #0x8\n"
       "str d23, [x22], #0x8\n"
       "tbz x8, #0, 134f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "st1 { v19.s }[2], [x23]\n"
       "st1 { v23.s }[2], [x22]\n"
       "b 134f\n"
       "127:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x8, #0, 134f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "str s19, [x23, #0x0]\n"
       "str s23, [x22, #0x0]\n"
       "b 134f\n"
       "128:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x8, #1, 129f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "str d18, [x23], #0x8\n"
       "str d22, [x22], #0x8\n"
       "tbz x8, #0, 134f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "st1 { v18.s }[2], [x23]\n"
       "st1 { v22.s }[2], [x22]\n"
       "b 134f\n"
       "129:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x8, #0, 134f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "str s18, [x23, #0x0]\n"
       "str s22, [x22, #0x0]\n"
       "b 134f\n"
       "130:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x8, #2, 132f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "st1 { v20.4s }, [x22], #0x10\n"
       "tbz x8, #1, 131f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "str d17, [x23], #0x8\n"
       "str d21, [x22], #0x8\n"
       "tbz x8, #0, 134f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "st1 { v17.s }[2], [x23]\n"
       "st1 { v21.s }[2], [x22]\n"
       "b 134f\n"
       "131:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x8, #0, 134f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "str s17, [x23, #0x0]\n"
       "str s21, [x22, #0x0]\n"
       "b 134f\n"
       "132:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x8, #1, 133f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "str d16, [x23], #0x8\n"
       "str d20, [x22], #0x8\n"
       "tbz x8, #0, 134f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "st1 { v16.s }[2], [x23]\n"
       "st1 { v20.s }[2], [x22]\n"
       "b 134f\n"
       "133:"  // Height 4: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "str s16, [x23, #0x0]\n"
       "str s20, [x22, #0x0]\n"
       "134:"  // Height 4: Partial direct writeback: Done
       "b 136f\n"
       "135:"  // Height 4: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -1937,43 +1937,43 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 206f\n"
       "137:"  // Height 5
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "138:"  // Height 5: Column loop
       "tbz %x[flags], #0, 148f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
       "add x21, x22, x20, LSL #2\n"
       "bge 147f\n"
       "tbz x8, #3, 142f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v24.4s }, [x21], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "ld1 { v17.4s }, [x23], #0x10\n"
       "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v25.4s }, [x21], #0x10\n"
       "tbz x8, #2, 140f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "ld1 { v18.4s }, [x23], #0x10\n"
       "ld1 { v22.4s }, [x22], #0x10\n"
       "ld1 { v26.4s }, [x21], #0x10\n"
       "tbz x8, #1, 139f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "ldr d19, [x23], #0x8\n"
       "ldr d23, [x22], #0x8\n"
       "ldr d27, [x21], #0x8\n"
       "tbz x8, #0, 146f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "ld1 { v19.s }[2], [x23]\n"
       "ld1 { v23.s }[2], [x22]\n"
@@ -1982,7 +1982,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "139:"  // Height 5: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 146f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "ldr s19, [x23, #0x0]\n"
       "ldr s23, [x22, #0x0]\n"
@@ -1990,14 +1990,14 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 146f\n"
       "140:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x8, #1, 141f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "ldr d18, [x23], #0x8\n"
       "ldr d22, [x22], #0x8\n"
       "ldr d26, [x21], #0x8\n"
       "tbz x8, #0, 146f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "ld1 { v18.s }[2], [x23]\n"
       "ld1 { v22.s }[2], [x22]\n"
@@ -2006,7 +2006,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "141:"  // Height 5: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 146f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "ldr s18, [x23, #0x0]\n"
       "ldr s22, [x22, #0x0]\n"
@@ -2014,20 +2014,20 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 146f\n"
       "142:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x8, #2, 144f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v24.4s }, [x21], #0x10\n"
       "tbz x8, #1, 143f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "ldr d17, [x23], #0x8\n"
       "ldr d21, [x22], #0x8\n"
       "ldr d25, [x21], #0x8\n"
       "tbz x8, #0, 146f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "ld1 { v17.s }[2], [x23]\n"
       "ld1 { v21.s }[2], [x22]\n"
@@ -2036,7 +2036,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "143:"  // Height 5: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 146f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "ldr s17, [x23, #0x0]\n"
       "ldr s21, [x22, #0x0]\n"
@@ -2044,34 +2044,34 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 146f\n"
       "144:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x8, #1, 145f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "ldr d16, [x23], #0x8\n"
       "ldr d20, [x22], #0x8\n"
       "ldr d24, [x21], #0x8\n"
       "tbz x8, #0, 146f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "ld1 { v16.s }[2], [x23]\n"
       "ld1 { v20.s }[2], [x22]\n"
       "ld1 { v24.s }[2], [x21]\n"
       "b 146f\n"
       "145:"  // Height 5: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "ldr s16, [x23, #0x0]\n"
       "ldr s20, [x22, #0x0]\n"
       "ldr s24, [x21, #0x0]\n"
       "146:"  // Height 5: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 149f\n"
       "147:"  // Height 5: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -2114,8 +2114,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "150:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 151f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2147,148 +2147,148 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr q2, [x11, #0x0]\n"
       "ldr q3, [x10, #0x0]\n"
       "ldr q4, [x9, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 154f\n"
       "153:"  // Height 5: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       "add x13, x13, #0x10\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr d29, [x17, #0x20]\n"
+      "ldr d29, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x11, x11, #0x10\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "ldr x21, [x16, #0x48]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v29.d[1], x21\n"
+      "add x11, x11, #0x10\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x17, #0x48]\n"
+      "add x10, x10, #0x10\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr d28, [x17, #0x30]\n"
+      "ldr d28, [x16, #0x30]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
-      "add x9, x9, #0x10\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
-      "ldr x26, [x13, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0x58]\n"
       ".inst 0x4f82e3b2  // sdot v18.4s, v29.16b, v2.4b[0]\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x4f83e3b6  // sdot v22.4s, v29.16b, v3.4b[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x26, [x13, #0x8]\n"
       ".inst 0x4f84e3ba  // sdot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr d29, [x17, #0x40]\n"
+      "ldr d29, [x16, #0x40]\n"
       ".inst 0x4f80e38b  // sdot v11.4s, v28.16b, v0.4b[0]\n"
-      "ldr x25, [x12, #0x8]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81e38f  // sdot v15.4s, v28.16b, v1.4b[0]\n"
-      "ldr x24, [x11, #0x8]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x4f82e393  // sdot v19.4s, v28.16b, v2.4b[0]\n"
-      "mov v29.d[1], x21\n"
+      "ldr x25, [x12, #0x8]\n"
       ".inst 0x4f83e397  // sdot v23.4s, v28.16b, v3.4b[0]\n"
-      "ldr x21, [x17, #0x68]\n"
+      "ldr x24, [x11, #0x8]\n"
       ".inst 0x4f84e39b  // sdot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr d28, [x17, #0x50]\n"
+      "ldr d28, [x16, #0x50]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4fa0e3a8  // sdot v8.4s, v29.16b, v0.4b[1]\n"
-      "ldr x23, [x10, #0x8]\n"
       ".inst 0x4fa1e3ac  // sdot v12.4s, v29.16b, v1.4b[1]\n"
-      "ldr x22, [x9, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x4fa2e3b0  // sdot v16.4s, v29.16b, v2.4b[1]\n"
+      "ldr x23, [x10, #0x8]\n"
       ".inst 0x4fa3e3b4  // sdot v20.4s, v29.16b, v3.4b[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "ldr x22, [x9, #0x8]\n"
       ".inst 0x4fa4e3b8  // sdot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr d29, [x17, #0x60]\n"
+      "ldr d29, [x16, #0x60]\n"
       ".inst 0x4fa0e389  // sdot v9.4s, v28.16b, v0.4b[1]\n"
-      "sub x14, x14, #0x10\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4fa1e38d  // sdot v13.4s, v28.16b, v1.4b[1]\n"
-      "cmp x14, #0x20\n"
+      "ldr x21, [x16, #0x88]\n"
       ".inst 0x4fa2e391  // sdot v17.4s, v28.16b, v2.4b[1]\n"
-      "mov v29.d[1], x21\n"
+      "sub x14, x14, #0x10\n"
       ".inst 0x4fa3e395  // sdot v21.4s, v28.16b, v3.4b[1]\n"
-      "ldr x21, [x17, #0x88]\n"
+      "cmp x14, #0x20\n"
       ".inst 0x4fa4e399  // sdot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr d28, [x17, #0x70]\n"
+      "ldr d28, [x16, #0x70]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4fa0e3aa  // sdot v10.4s, v29.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x4fa1e3ae  // sdot v14.4s, v29.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0x98]\n"
       ".inst 0x4fa2e3b2  // sdot v18.4s, v29.16b, v2.4b[1]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x4fa3e3b6  // sdot v22.4s, v29.16b, v3.4b[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4fa4e3ba  // sdot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr d29, [x17, #0x80]\n"
+      "ldr d29, [x16, #0x80]\n"
       ".inst 0x4fa0e38b  // sdot v11.4s, v28.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4fa1e38f  // sdot v15.4s, v28.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x4fa2e393  // sdot v19.4s, v28.16b, v2.4b[1]\n"
-      "mov v29.d[1], x21\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4fa3e397  // sdot v23.4s, v28.16b, v3.4b[1]\n"
-      "ldr x21, [x17, #0xa8]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4fa4e39b  // sdot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr d28, [x17, #0x90]\n"
+      "ldr d28, [x16, #0x90]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4f80eba8  // sdot v8.4s, v29.16b, v0.4b[2]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4f81ebac  // sdot v12.4s, v29.16b, v1.4b[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x4f82ebb0  // sdot v16.4s, v29.16b, v2.4b[2]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4f83ebb4  // sdot v20.4s, v29.16b, v3.4b[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
       ".inst 0x4f84ebb8  // sdot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr d29, [x17, #0xa0]\n"
+      "ldr d29, [x16, #0xa0]\n"
       ".inst 0x4f80eb89  // sdot v9.4s, v28.16b, v0.4b[2]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81eb8d  // sdot v13.4s, v28.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x4f82eb91  // sdot v17.4s, v28.16b, v2.4b[2]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x4f83eb95  // sdot v21.4s, v28.16b, v3.4b[2]\n"
-      "ldr x21, [x17, #0xc8]\n"
       ".inst 0x4f84eb99  // sdot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr d28, [x17, #0xb0]\n"
+      "ldr d28, [x16, #0xb0]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4f80ebaa  // sdot v10.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebae  // sdot v14.4s, v29.16b, v1.4b[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x4f82ebb2  // sdot v18.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb6  // sdot v22.4s, v29.16b, v3.4b[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
       ".inst 0x4f84ebba  // sdot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr d29, [x17, #0xc0]\n"
+      "ldr d29, [x16, #0xc0]\n"
       ".inst 0x4f80eb8b  // sdot v11.4s, v28.16b, v0.4b[2]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4f81eb8f  // sdot v15.4s, v28.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x4f82eb93  // sdot v19.4s, v28.16b, v2.4b[2]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x4f83eb97  // sdot v23.4s, v28.16b, v3.4b[2]\n"
-      "ldr x21, [x17, #0xe8]\n"
       ".inst 0x4f84eb9b  // sdot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr d28, [x17, #0xd0]\n"
+      "ldr d28, [x16, #0xd0]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x4fa0eba8  // sdot v8.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x4fa1ebac  // sdot v12.4s, v29.16b, v1.4b[3]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x4fa2ebb0  // sdot v16.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb4  // sdot v20.4s, v29.16b, v3.4b[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
       ".inst 0x4fa4ebb8  // sdot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr d29, [x17, #0xe0]\n"
+      "ldr d29, [x16, #0xe0]\n"
       ".inst 0x4fa0eb89  // sdot v9.4s, v28.16b, v0.4b[3]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x4fa1eb8d  // sdot v13.4s, v28.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb91  // sdot v17.4s, v28.16b, v2.4b[3]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x4fa3eb95  // sdot v21.4s, v28.16b, v3.4b[3]\n"
       ".inst 0x4fa4eb99  // sdot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr d28, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "ldr d28, [x16, #0xf0]\n"
+      "mov v28.d[1], x20\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa0ebaa  // sdot v10.4s, v29.16b, v0.4b[3]\n"
+      "ldr x21, [x16, #0x8]\n"
       ".inst 0x4fa1ebae  // sdot v14.4s, v29.16b, v1.4b[3]\n"
-      "ldr x21, [x17, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
       ".inst 0x4fa2ebb2  // sdot v18.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb6  // sdot v22.4s, v29.16b, v3.4b[3]\n"
-      "ldr x20, [x17, #0x18]\n"
       ".inst 0x4fa4ebba  // sdot v26.4s, v29.16b, v4.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
       ".inst 0x4fa0eb8b  // sdot v11.4s, v28.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x4fa1eb8f  // sdot v15.4s, v28.16b, v1.4b[3]\n"
@@ -2299,7 +2299,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr d3, [x10, #0x0]\n"
       ".inst 0x4fa4eb9b  // sdot v27.4s, v28.16b, v4.4b[3]\n"
       "ldr d4, [x9, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
+      "ldr d7, [x16, #0x10]\n"
       "mov v6.d[1], x21\n"
       "mov v0.d[1], x26\n"
       "mov v1.d[1], x25\n"
@@ -2318,7 +2318,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
       "add x10, x10, #0x10\n"
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr q29, [x17, #0x20]\n"
+      "ldr q29, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2328,7 +2328,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr q28, [x17, #0x30]\n"
+      "ldr q28, [x16, #0x30]\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
@@ -2337,75 +2337,75 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4f83e3b6  // sdot v22.4s, v29.16b, v3.4b[0]\n"
       ".inst 0x4f84e3ba  // sdot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x17, #0x40]\n"
+      "ldr q29, [x16, #0x40]\n"
       ".inst 0x4f80e38b  // sdot v11.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f81e38f  // sdot v15.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f82e393  // sdot v19.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f83e397  // sdot v23.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x4f84e39b  // sdot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x17, #0x50]\n"
+      "ldr q28, [x16, #0x50]\n"
       ".inst 0x4fa0e3a8  // sdot v8.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3ac  // sdot v12.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3b0  // sdot v16.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x4fa3e3b4  // sdot v20.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x4fa4e3b8  // sdot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x17, #0x60]\n"
+      "ldr q29, [x16, #0x60]\n"
       ".inst 0x4fa0e389  // sdot v9.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x4fa1e38d  // sdot v13.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x4fa2e391  // sdot v17.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x4fa3e395  // sdot v21.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x4fa4e399  // sdot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x17, #0x70]\n"
+      "ldr q28, [x16, #0x70]\n"
       ".inst 0x4fa0e3aa  // sdot v10.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x4fa1e3ae  // sdot v14.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x4fa2e3b2  // sdot v18.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x4fa3e3b6  // sdot v22.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x4fa4e3ba  // sdot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x17, #0x80]\n"
+      "ldr q29, [x16, #0x80]\n"
       ".inst 0x4fa0e38b  // sdot v11.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x4fa1e38f  // sdot v15.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x4fa2e393  // sdot v19.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x4fa3e397  // sdot v23.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x4fa4e39b  // sdot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x17, #0x90]\n"
+      "ldr q28, [x16, #0x90]\n"
       ".inst 0x4f80eba8  // sdot v8.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebac  // sdot v12.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x4f82ebb0  // sdot v16.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb4  // sdot v20.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x4f84ebb8  // sdot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x17, #0xa0]\n"
+      "ldr q29, [x16, #0xa0]\n"
       ".inst 0x4f80eb89  // sdot v9.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb8d  // sdot v13.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb91  // sdot v17.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x4f83eb95  // sdot v21.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x4f84eb99  // sdot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x17, #0xb0]\n"
+      "ldr q28, [x16, #0xb0]\n"
       ".inst 0x4f80ebaa  // sdot v10.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x4f81ebae  // sdot v14.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x4f82ebb2  // sdot v18.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x4f83ebb6  // sdot v22.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x4f84ebba  // sdot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x17, #0xc0]\n"
+      "ldr q29, [x16, #0xc0]\n"
       ".inst 0x4f80eb8b  // sdot v11.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x4f81eb8f  // sdot v15.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x4f82eb93  // sdot v19.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x4f83eb97  // sdot v23.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x4f84eb9b  // sdot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x17, #0xd0]\n"
+      "ldr q28, [x16, #0xd0]\n"
       ".inst 0x4fa0eba8  // sdot v8.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x4fa1ebac  // sdot v12.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x4fa2ebb0  // sdot v16.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb4  // sdot v20.4s, v29.16b, v3.4b[3]\n"
       ".inst 0x4fa4ebb8  // sdot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr q29, [x17, #0xe0]\n"
+      "ldr q29, [x16, #0xe0]\n"
       ".inst 0x4fa0eb89  // sdot v9.4s, v28.16b, v0.4b[3]\n"
       ".inst 0x4fa1eb8d  // sdot v13.4s, v28.16b, v1.4b[3]\n"
       ".inst 0x4fa2eb91  // sdot v17.4s, v28.16b, v2.4b[3]\n"
       ".inst 0x4fa3eb95  // sdot v21.4s, v28.16b, v3.4b[3]\n"
       ".inst 0x4fa4eb99  // sdot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr q28, [x17, #0xf0]\n"
+      "ldr q28, [x16, #0xf0]\n"
       ".inst 0x4fa0ebaa  // sdot v10.4s, v29.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa1ebae  // sdot v14.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x4fa2ebb2  // sdot v18.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x4fa3ebb6  // sdot v22.4s, v29.16b, v3.4b[3]\n"
@@ -2427,22 +2427,22 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr s0, [x11], #0x4\n"
       "ldr s31, [x10], #0x4\n"
       "ldr s30, [x9], #0x4\n"
-      "ldr q29, [x17, #0x0]\n"
-      "ldr q28, [x17, #0x10]\n"
+      "ldr q29, [x16, #0x0]\n"
       ".inst 0x4f82e3a8  // sdot v8.4s, v29.16b, v2.4b[0]\n"
+      "ldr q28, [x16, #0x10]\n"
       ".inst 0x4f81e3ac  // sdot v12.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f80e3b0  // sdot v16.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f9fe3b4  // sdot v20.4s, v29.16b, v31.4b[0]\n"
       ".inst 0x4f9ee3b8  // sdot v24.4s, v29.16b, v30.4b[0]\n"
-      "ldr q29, [x17, #0x20]\n"
+      "ldr q29, [x16, #0x20]\n"
       ".inst 0x4f82e389  // sdot v9.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f81e38d  // sdot v13.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f80e391  // sdot v17.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f9fe395  // sdot v21.4s, v28.16b, v31.4b[0]\n"
       ".inst 0x4f9ee399  // sdot v25.4s, v28.16b, v30.4b[0]\n"
-      "ldr q28, [x17, #0x30]\n"
+      "ldr q28, [x16, #0x30]\n"
       ".inst 0x4f82e3aa  // sdot v10.4s, v29.16b, v2.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f80e3b2  // sdot v18.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f9fe3b6  // sdot v22.4s, v29.16b, v31.4b[0]\n"
@@ -2475,22 +2475,22 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr b3, [x10, #0x0]\n"
       "ldr b4, [x9, #0x0]\n"
       "159:"  // Height 5: Multiply loop: Ragged operand read: Done
-      "ldr q29, [x17, #0x0]\n"
-      "ldr q28, [x17, #0x10]\n"
+      "ldr q29, [x16, #0x0]\n"
       ".inst 0x4f80e3a8  // sdot v8.4s, v29.16b, v0.4b[0]\n"
+      "ldr q28, [x16, #0x10]\n"
       ".inst 0x4f81e3ac  // sdot v12.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f82e3b0  // sdot v16.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x4f83e3b4  // sdot v20.4s, v29.16b, v3.4b[0]\n"
       ".inst 0x4f84e3b8  // sdot v24.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x17, #0x20]\n"
+      "ldr q29, [x16, #0x20]\n"
       ".inst 0x4f80e389  // sdot v9.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x4f81e38d  // sdot v13.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x4f82e391  // sdot v17.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x4f83e395  // sdot v21.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x4f84e399  // sdot v25.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x17, #0x30]\n"
+      "ldr q28, [x16, #0x30]\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x4f82e3b2  // sdot v18.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x4f83e3b6  // sdot v22.4s, v29.16b, v3.4b[0]\n"
@@ -2506,20 +2506,20 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 150b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "add x21, x22, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
       "prfm pstl1keep, [x21, #0x0]\n"
       "bge 169f\n"
       "tbz x8, #3, 164f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
@@ -2529,19 +2529,19 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "st1 { v24.4s }, [x21], #0x10\n"
       "st1 { v25.4s }, [x21], #0x10\n"
       "tbz x8, #2, 162f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "st1 { v18.4s }, [x23], #0x10\n"
       "st1 { v22.4s }, [x22], #0x10\n"
       "st1 { v26.4s }, [x21], #0x10\n"
       "tbz x8, #1, 161f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "str d19, [x23], #0x8\n"
       "str d23, [x22], #0x8\n"
       "str d27, [x21], #0x8\n"
       "tbz x8, #0, 168f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "st1 { v19.s }[2], [x23]\n"
       "st1 { v23.s }[2], [x22]\n"
@@ -2549,7 +2549,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 168f\n"
       "161:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x8, #0, 168f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "str s19, [x23, #0x0]\n"
       "str s23, [x22, #0x0]\n"
@@ -2557,13 +2557,13 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 168f\n"
       "162:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x8, #1, 163f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "str d18, [x23], #0x8\n"
       "str d22, [x22], #0x8\n"
       "str d26, [x21], #0x8\n"
       "tbz x8, #0, 168f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "st1 { v18.s }[2], [x23]\n"
       "st1 { v22.s }[2], [x22]\n"
@@ -2571,7 +2571,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 168f\n"
       "163:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x8, #0, 168f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "str s18, [x23, #0x0]\n"
       "str s22, [x22, #0x0]\n"
@@ -2579,19 +2579,19 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 168f\n"
       "164:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x8, #2, 166f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "st1 { v20.4s }, [x22], #0x10\n"
       "st1 { v24.4s }, [x21], #0x10\n"
       "tbz x8, #1, 165f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "str d17, [x23], #0x8\n"
       "str d21, [x22], #0x8\n"
       "str d25, [x21], #0x8\n"
       "tbz x8, #0, 168f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "st1 { v17.s }[2], [x23]\n"
       "st1 { v21.s }[2], [x22]\n"
@@ -2599,7 +2599,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 168f\n"
       "165:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x8, #0, 168f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "str s17, [x23, #0x0]\n"
       "str s21, [x22, #0x0]\n"
@@ -2607,20 +2607,20 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 168f\n"
       "166:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x8, #1, 167f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "str d16, [x23], #0x8\n"
       "str d20, [x22], #0x8\n"
       "str d24, [x21], #0x8\n"
       "tbz x8, #0, 168f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "st1 { v16.s }[2], [x23]\n"
       "st1 { v20.s }[2], [x22]\n"
       "st1 { v24.s }[2], [x21]\n"
       "b 168f\n"
       "167:"  // Height 5: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "str s16, [x23, #0x0]\n"
       "str s20, [x22, #0x0]\n"
@@ -2628,11 +2628,11 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "168:"  // Height 5: Partial direct writeback: Done
       "b 170f\n"
       "169:"  // Height 5: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -2656,43 +2656,42 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "171:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "mov x20, #0x18\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x16\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "172:"  // Height 6: Column loop
       "tbz %x[flags], #0, 182f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
       "add x20, x21, x20, LSL #2\n"
       "bge 181f\n"
       "tbz x8, #3, 176f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v24.4s }, [x21], #0x10\n"
       "ld1 { v28.4s }, [x20], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "ld1 { v17.4s }, [x23], #0x10\n"
       "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v25.4s }, [x21], #0x10\n"
       "ld1 { v29.4s }, [x20], #0x10\n"
       "tbz x8, #2, 174f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "ld1 { v18.4s }, [x23], #0x10\n"
       "ld1 { v22.4s }, [x22], #0x10\n"
       "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v30.4s }, [x20], #0x10\n"
       "tbz x8, #1, 173f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "ldr d19, [x23], #0x8\n"
@@ -2700,7 +2699,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr d27, [x21], #0x8\n"
       "ldr d31, [x20], #0x8\n"
       "tbz x8, #0, 180f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "ld1 { v19.s }[2], [x23]\n"
       "ld1 { v23.s }[2], [x22]\n"
@@ -2710,7 +2709,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "173:"  // Height 6: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 180f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "ldr s19, [x23, #0x0]\n"
       "ldr s23, [x22, #0x0]\n"
@@ -2719,7 +2718,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 180f\n"
       "174:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x8, #1, 175f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "ldr d18, [x23], #0x8\n"
@@ -2727,7 +2726,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr d26, [x21], #0x8\n"
       "ldr d30, [x20], #0x8\n"
       "tbz x8, #0, 180f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "ld1 { v18.s }[2], [x23]\n"
       "ld1 { v22.s }[2], [x22]\n"
@@ -2737,7 +2736,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "175:"  // Height 6: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 180f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "ldr s18, [x23, #0x0]\n"
       "ldr s22, [x22, #0x0]\n"
@@ -2746,14 +2745,14 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 180f\n"
       "176:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x8, #2, 178f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v24.4s }, [x21], #0x10\n"
       "ld1 { v28.4s }, [x20], #0x10\n"
       "tbz x8, #1, 177f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "ldr d17, [x23], #0x8\n"
@@ -2761,7 +2760,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr d25, [x21], #0x8\n"
       "ldr d29, [x20], #0x8\n"
       "tbz x8, #0, 180f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "ld1 { v17.s }[2], [x23]\n"
       "ld1 { v21.s }[2], [x22]\n"
@@ -2771,7 +2770,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "177:"  // Height 6: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 180f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "ldr s17, [x23, #0x0]\n"
       "ldr s21, [x22, #0x0]\n"
@@ -2780,7 +2779,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 180f\n"
       "178:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x8, #1, 179f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "ldr d16, [x23], #0x8\n"
@@ -2788,7 +2787,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr d24, [x21], #0x8\n"
       "ldr d28, [x20], #0x8\n"
       "tbz x8, #0, 180f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "ld1 { v16.s }[2], [x23]\n"
       "ld1 { v20.s }[2], [x22]\n"
@@ -2796,7 +2795,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ld1 { v28.s }[2], [x20]\n"
       "b 180f\n"
       "179:"  // Height 6: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "ldr s16, [x23, #0x0]\n"
@@ -2804,13 +2803,13 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr s24, [x21, #0x0]\n"
       "ldr s28, [x20, #0x0]\n"
       "180:"  // Height 6: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 183f\n"
       "181:"  // Height 6: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -2861,8 +2860,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "184:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 185f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2898,14 +2897,14 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr q3, [x10, #0x0]\n"
       "ldr q4, [x9, #0x0]\n"
       "ldr q5, [x28, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 188f\n"
       "187:"  // Height 6: Multiply loop: Main loop head
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x4f81e0cc  // sdot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x4f82e0d0  // sdot v16.4s, v6.16b, v2.4b[0]\n"
       "add x13, x13, #0x10\n"
       ".inst 0x4f83e0d4  // sdot v20.4s, v6.16b, v3.4b[0]\n"
@@ -2913,151 +2912,151 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x4f85e0dc  // sdot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr d6, [x17, #0x20]\n"
+      "ldr d6, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x9, x9, #0x10\n"
+      "ldr x21, [x16, #0x48]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v6.d[1], x21\n"
+      "add x10, x10, #0x10\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x17, #0x48]\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
       "add x28, x28, #0x10\n"
       ".inst 0x4f85e0fd  // sdot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr d7, [x17, #0x30]\n"
+      "ldr d7, [x16, #0x30]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4f80e0ca  // sdot v10.4s, v6.16b, v0.4b[0]\n"
-      "ldr x27, [x13, #0x8]\n"
       ".inst 0x4f81e0ce  // sdot v14.4s, v6.16b, v1.4b[0]\n"
-      "ldr x26, [x12, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0x58]\n"
       ".inst 0x4f82e0d2  // sdot v18.4s, v6.16b, v2.4b[0]\n"
+      "ldr x27, [x13, #0x8]\n"
       ".inst 0x4f83e0d6  // sdot v22.4s, v6.16b, v3.4b[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x26, [x12, #0x8]\n"
       ".inst 0x4f84e0da  // sdot v26.4s, v6.16b, v4.4b[0]\n"
       "ldr x25, [x11, #0x8]\n"
       ".inst 0x4f85e0de  // sdot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr d6, [x17, #0x40]\n"
+      "ldr d6, [x16, #0x40]\n"
       ".inst 0x4f80e0eb  // sdot v11.4s, v7.16b, v0.4b[0]\n"
-      "ldr x24, [x10, #0x8]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4f81e0ef  // sdot v15.4s, v7.16b, v1.4b[0]\n"
-      "ldr x23, [x9, #0x8]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x4f82e0f3  // sdot v19.4s, v7.16b, v2.4b[0]\n"
-      "mov v6.d[1], x21\n"
+      "ldr x24, [x10, #0x8]\n"
       ".inst 0x4f83e0f7  // sdot v23.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x17, #0x68]\n"
+      "ldr x23, [x9, #0x8]\n"
       ".inst 0x4f84e0fb  // sdot v27.4s, v7.16b, v4.4b[0]\n"
       "ldr x22, [x28, #0x8]\n"
       ".inst 0x4f85e0ff  // sdot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr d7, [x17, #0x50]\n"
+      "ldr d7, [x16, #0x50]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4fa0e0c8  // sdot v8.4s, v6.16b, v0.4b[1]\n"
-      "sub x14, x14, #0x10\n"
       ".inst 0x4fa1e0cc  // sdot v12.4s, v6.16b, v1.4b[1]\n"
-      "cmp x14, #0x20\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x4fa2e0d0  // sdot v16.4s, v6.16b, v2.4b[1]\n"
+      "sub x14, x14, #0x10\n"
       ".inst 0x4fa3e0d4  // sdot v20.4s, v6.16b, v3.4b[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "cmp x14, #0x20\n"
       ".inst 0x4fa4e0d8  // sdot v24.4s, v6.16b, v4.4b[1]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x4fa5e0dc  // sdot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr d6, [x17, #0x60]\n"
+      "ldr d6, [x16, #0x60]\n"
       ".inst 0x4fa0e0e9  // sdot v9.4s, v7.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4fa1e0ed  // sdot v13.4s, v7.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x16, #0x88]\n"
       ".inst 0x4fa2e0f1  // sdot v17.4s, v7.16b, v2.4b[1]\n"
-      "mov v6.d[1], x21\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x4fa3e0f5  // sdot v21.4s, v7.16b, v3.4b[1]\n"
-      "ldr x21, [x17, #0x88]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4fa4e0f9  // sdot v25.4s, v7.16b, v4.4b[1]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4fa5e0fd  // sdot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr d7, [x17, #0x70]\n"
+      "ldr d7, [x16, #0x70]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4fa0e0ca  // sdot v10.4s, v6.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4fa1e0ce  // sdot v14.4s, v6.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0x98]\n"
       ".inst 0x4fa2e0d2  // sdot v18.4s, v6.16b, v2.4b[1]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x4fa3e0d6  // sdot v22.4s, v6.16b, v3.4b[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       ".inst 0x4fa4e0da  // sdot v26.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0de  // sdot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr d6, [x17, #0x80]\n"
+      "ldr d6, [x16, #0x80]\n"
       ".inst 0x4fa0e0eb  // sdot v11.4s, v7.16b, v0.4b[1]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4fa1e0ef  // sdot v15.4s, v7.16b, v1.4b[1]\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x4fa2e0f3  // sdot v19.4s, v7.16b, v2.4b[1]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x4fa3e0f7  // sdot v23.4s, v7.16b, v3.4b[1]\n"
-      "ldr x21, [x17, #0xa8]\n"
       ".inst 0x4fa4e0fb  // sdot v27.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0ff  // sdot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr d7, [x17, #0x90]\n"
+      "ldr d7, [x16, #0x90]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4f80e8c8  // sdot v8.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8cc  // sdot v12.4s, v6.16b, v1.4b[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x4f82e8d0  // sdot v16.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d4  // sdot v20.4s, v6.16b, v3.4b[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
       ".inst 0x4f84e8d8  // sdot v24.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8dc  // sdot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr d6, [x17, #0xa0]\n"
+      "ldr d6, [x16, #0xa0]\n"
       ".inst 0x4f80e8e9  // sdot v9.4s, v7.16b, v0.4b[2]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4f81e8ed  // sdot v13.4s, v7.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x4f82e8f1  // sdot v17.4s, v7.16b, v2.4b[2]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x4f83e8f5  // sdot v21.4s, v7.16b, v3.4b[2]\n"
-      "ldr x21, [x17, #0xc8]\n"
       ".inst 0x4f84e8f9  // sdot v25.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8fd  // sdot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr d7, [x17, #0xb0]\n"
+      "ldr d7, [x16, #0xb0]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4f80e8ca  // sdot v10.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ce  // sdot v14.4s, v6.16b, v1.4b[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x4f82e8d2  // sdot v18.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d6  // sdot v22.4s, v6.16b, v3.4b[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
       ".inst 0x4f84e8da  // sdot v26.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8de  // sdot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr d6, [x17, #0xc0]\n"
+      "ldr d6, [x16, #0xc0]\n"
       ".inst 0x4f80e8eb  // sdot v11.4s, v7.16b, v0.4b[2]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4f81e8ef  // sdot v15.4s, v7.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x4f82e8f3  // sdot v19.4s, v7.16b, v2.4b[2]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x4f83e8f7  // sdot v23.4s, v7.16b, v3.4b[2]\n"
-      "ldr x21, [x17, #0xe8]\n"
       ".inst 0x4f84e8fb  // sdot v27.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8ff  // sdot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr d7, [x17, #0xd0]\n"
+      "ldr d7, [x16, #0xd0]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x4fa0e8c8  // sdot v8.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8cc  // sdot v12.4s, v6.16b, v1.4b[3]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x4fa2e8d0  // sdot v16.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d4  // sdot v20.4s, v6.16b, v3.4b[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
       ".inst 0x4fa4e8d8  // sdot v24.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8dc  // sdot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr d6, [x17, #0xe0]\n"
+      "ldr d6, [x16, #0xe0]\n"
       ".inst 0x4fa0e8e9  // sdot v9.4s, v7.16b, v0.4b[3]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x4fa1e8ed  // sdot v13.4s, v7.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8f1  // sdot v17.4s, v7.16b, v2.4b[3]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x4fa3e8f5  // sdot v21.4s, v7.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8f9  // sdot v25.4s, v7.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8fd  // sdot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr d7, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "ldr d7, [x16, #0xf0]\n"
+      "mov v7.d[1], x20\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa0e8ca  // sdot v10.4s, v6.16b, v0.4b[3]\n"
+      "ldr x21, [x16, #0x8]\n"
       ".inst 0x4fa1e8ce  // sdot v14.4s, v6.16b, v1.4b[3]\n"
-      "ldr x21, [x17, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
       ".inst 0x4fa2e8d2  // sdot v18.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d6  // sdot v22.4s, v6.16b, v3.4b[3]\n"
-      "ldr x20, [x17, #0x18]\n"
       ".inst 0x4fa4e8da  // sdot v26.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8de  // sdot v30.4s, v6.16b, v5.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
       ".inst 0x4fa0e8eb  // sdot v11.4s, v7.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x4fa1e8ef  // sdot v15.4s, v7.16b, v1.4b[3]\n"
@@ -3070,7 +3069,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr d4, [x9, #0x0]\n"
       ".inst 0x4fa5e8ff  // sdot v31.4s, v7.16b, v5.4b[3]\n"
       "ldr d5, [x28, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
+      "ldr d7, [x16, #0x10]\n"
       "mov v6.d[1], x21\n"
       "mov v0.d[1], x27\n"
       "mov v1.d[1], x26\n"
@@ -3092,7 +3091,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f84e0d8  // sdot v24.4s, v6.16b, v4.4b[0]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x4f85e0dc  // sdot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x17, #0x20]\n"
+      "ldr q6, [x16, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "add x28, x28, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -3104,7 +3103,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x4f85e0fd  // sdot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x17, #0x30]\n"
+      "ldr q7, [x16, #0x30]\n"
       ".inst 0x4f80e0ca  // sdot v10.4s, v6.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x4f81e0ce  // sdot v14.4s, v6.16b, v1.4b[0]\n"
@@ -3114,86 +3113,86 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       ".inst 0x4f83e0d6  // sdot v22.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x4f84e0da  // sdot v26.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0de  // sdot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x17, #0x40]\n"
+      "ldr q6, [x16, #0x40]\n"
       ".inst 0x4f80e0eb  // sdot v11.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ef  // sdot v15.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f3  // sdot v19.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f7  // sdot v23.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0fb  // sdot v27.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0ff  // sdot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x17, #0x50]\n"
+      "ldr q7, [x16, #0x50]\n"
       ".inst 0x4fa0e0c8  // sdot v8.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0cc  // sdot v12.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0d0  // sdot v16.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d4  // sdot v20.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0d8  // sdot v24.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0dc  // sdot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x17, #0x60]\n"
+      "ldr q6, [x16, #0x60]\n"
       ".inst 0x4fa0e0e9  // sdot v9.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ed  // sdot v13.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0f1  // sdot v17.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0f5  // sdot v21.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0f9  // sdot v25.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0fd  // sdot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x17, #0x70]\n"
+      "ldr q7, [x16, #0x70]\n"
       ".inst 0x4fa0e0ca  // sdot v10.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ce  // sdot v14.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0d2  // sdot v18.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0d6  // sdot v22.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0da  // sdot v26.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0de  // sdot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x17, #0x80]\n"
+      "ldr q6, [x16, #0x80]\n"
       ".inst 0x4fa0e0eb  // sdot v11.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x4fa1e0ef  // sdot v15.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x4fa2e0f3  // sdot v19.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x4fa3e0f7  // sdot v23.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x4fa4e0fb  // sdot v27.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x4fa5e0ff  // sdot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x17, #0x90]\n"
+      "ldr q7, [x16, #0x90]\n"
       ".inst 0x4f80e8c8  // sdot v8.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8cc  // sdot v12.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d0  // sdot v16.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d4  // sdot v20.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f84e8d8  // sdot v24.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8dc  // sdot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x17, #0xa0]\n"
+      "ldr q6, [x16, #0xa0]\n"
       ".inst 0x4f80e8e9  // sdot v9.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ed  // sdot v13.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x4f82e8f1  // sdot v17.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x4f83e8f5  // sdot v21.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x4f84e8f9  // sdot v25.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8fd  // sdot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x17, #0xb0]\n"
+      "ldr q7, [x16, #0xb0]\n"
       ".inst 0x4f80e8ca  // sdot v10.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ce  // sdot v14.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x4f82e8d2  // sdot v18.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x4f83e8d6  // sdot v22.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x4f84e8da  // sdot v26.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x4f85e8de  // sdot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x17, #0xc0]\n"
+      "ldr q6, [x16, #0xc0]\n"
       ".inst 0x4f80e8eb  // sdot v11.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x4f81e8ef  // sdot v15.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x4f82e8f3  // sdot v19.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x4f83e8f7  // sdot v23.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x4f84e8fb  // sdot v27.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x4f85e8ff  // sdot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x17, #0xd0]\n"
+      "ldr q7, [x16, #0xd0]\n"
       ".inst 0x4fa0e8c8  // sdot v8.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8cc  // sdot v12.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8d0  // sdot v16.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d4  // sdot v20.4s, v6.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8d8  // sdot v24.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8dc  // sdot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x17, #0xe0]\n"
+      "ldr q6, [x16, #0xe0]\n"
       ".inst 0x4fa0e8e9  // sdot v9.4s, v7.16b, v0.4b[3]\n"
       ".inst 0x4fa1e8ed  // sdot v13.4s, v7.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8f1  // sdot v17.4s, v7.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8f5  // sdot v21.4s, v7.16b, v3.4b[3]\n"
       ".inst 0x4fa4e8f9  // sdot v25.4s, v7.16b, v4.4b[3]\n"
       ".inst 0x4fa5e8fd  // sdot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr q7, [x17, #0xf0]\n"
+      "ldr q7, [x16, #0xf0]\n"
       ".inst 0x4fa0e8ca  // sdot v10.4s, v6.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x4fa1e8ce  // sdot v14.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x4fa2e8d2  // sdot v18.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x4fa3e8d6  // sdot v22.4s, v6.16b, v3.4b[3]\n"
@@ -3218,24 +3217,24 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr s4, [x10], #0x4\n"
       "ldr s3, [x9], #0x4\n"
       "ldr s2, [x28], #0x4\n"
-      "ldr q1, [x17, #0x0]\n"
-      "ldr q0, [x17, #0x10]\n"
+      "ldr q1, [x16, #0x0]\n"
       ".inst 0x4f87e028  // sdot v8.4s, v1.16b, v7.4b[0]\n"
+      "ldr q0, [x16, #0x10]\n"
       ".inst 0x4f86e02c  // sdot v12.4s, v1.16b, v6.4b[0]\n"
       ".inst 0x4f85e030  // sdot v16.4s, v1.16b, v5.4b[0]\n"
       ".inst 0x4f84e034  // sdot v20.4s, v1.16b, v4.4b[0]\n"
       ".inst 0x4f83e038  // sdot v24.4s, v1.16b, v3.4b[0]\n"
       ".inst 0x4f82e03c  // sdot v28.4s, v1.16b, v2.4b[0]\n"
-      "ldr q1, [x17, #0x20]\n"
+      "ldr q1, [x16, #0x20]\n"
       ".inst 0x4f87e009  // sdot v9.4s, v0.16b, v7.4b[0]\n"
       ".inst 0x4f86e00d  // sdot v13.4s, v0.16b, v6.4b[0]\n"
       ".inst 0x4f85e011  // sdot v17.4s, v0.16b, v5.4b[0]\n"
       ".inst 0x4f84e015  // sdot v21.4s, v0.16b, v4.4b[0]\n"
       ".inst 0x4f83e019  // sdot v25.4s, v0.16b, v3.4b[0]\n"
       ".inst 0x4f82e01d  // sdot v29.4s, v0.16b, v2.4b[0]\n"
-      "ldr q0, [x17, #0x30]\n"
+      "ldr q0, [x16, #0x30]\n"
       ".inst 0x4f87e02a  // sdot v10.4s, v1.16b, v7.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f86e02e  // sdot v14.4s, v1.16b, v6.4b[0]\n"
       ".inst 0x4f85e032  // sdot v18.4s, v1.16b, v5.4b[0]\n"
       ".inst 0x4f84e036  // sdot v22.4s, v1.16b, v4.4b[0]\n"
@@ -3273,24 +3272,24 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "ldr b4, [x9, #0x0]\n"
       "ldr b5, [x28, #0x0]\n"
       "193:"  // Height 6: Multiply loop: Ragged operand read: Done
-      "ldr q7, [x17, #0x0]\n"
-      "ldr q6, [x17, #0x10]\n"
+      "ldr q7, [x16, #0x0]\n"
       ".inst 0x4f80e0e8  // sdot v8.4s, v7.16b, v0.4b[0]\n"
+      "ldr q6, [x16, #0x10]\n"
       ".inst 0x4f81e0ec  // sdot v12.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f0  // sdot v16.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f4  // sdot v20.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0f8  // sdot v24.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0fc  // sdot v28.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x17, #0x20]\n"
+      "ldr q7, [x16, #0x20]\n"
       ".inst 0x4f80e0c9  // sdot v9.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0cd  // sdot v13.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x4f82e0d1  // sdot v17.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0d5  // sdot v21.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x4f84e0d9  // sdot v25.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x4f85e0dd  // sdot v29.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x17, #0x30]\n"
+      "ldr q6, [x16, #0x30]\n"
       ".inst 0x4f80e0ea  // sdot v10.4s, v7.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x4f81e0ee  // sdot v14.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f2  // sdot v18.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f6  // sdot v22.4s, v7.16b, v3.4b[0]\n"
@@ -3308,22 +3307,22 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 184b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "add x21, x22, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "add x20, x21, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "prfm pstl1keep, [x21, #0x0]\n"
       "prfm pstl1keep, [x20, #0x0]\n"
       "bge 203f\n"
       "tbz x8, #3, 198f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
@@ -3335,21 +3334,21 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "st1 { v28.4s }, [x20], #0x10\n"
       "st1 { v29.4s }, [x20], #0x10\n"
       "tbz x8, #2, 196f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "st1 { v18.4s }, [x23], #0x10\n"
       "st1 { v22.4s }, [x22], #0x10\n"
       "st1 { v26.4s }, [x21], #0x10\n"
       "st1 { v30.4s }, [x20], #0x10\n"
       "tbz x8, #1, 195f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "str d19, [x23], #0x8\n"
       "str d23, [x22], #0x8\n"
       "str d27, [x21], #0x8\n"
       "str d31, [x20], #0x8\n"
       "tbz x8, #0, 202f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "st1 { v19.s }[2], [x23]\n"
       "st1 { v23.s }[2], [x22]\n"
@@ -3358,7 +3357,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 202f\n"
       "195:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x8, #0, 202f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "str s19, [x23, #0x0]\n"
       "str s23, [x22, #0x0]\n"
@@ -3367,14 +3366,14 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 202f\n"
       "196:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x8, #1, 197f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "str d18, [x23], #0x8\n"
       "str d22, [x22], #0x8\n"
       "str d26, [x21], #0x8\n"
       "str d30, [x20], #0x8\n"
       "tbz x8, #0, 202f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "st1 { v18.s }[2], [x23]\n"
       "st1 { v22.s }[2], [x22]\n"
@@ -3383,7 +3382,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 202f\n"
       "197:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x8, #0, 202f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "str s18, [x23, #0x0]\n"
       "str s22, [x22, #0x0]\n"
@@ -3392,21 +3391,21 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 202f\n"
       "198:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x8, #2, 200f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "st1 { v20.4s }, [x22], #0x10\n"
       "st1 { v24.4s }, [x21], #0x10\n"
       "st1 { v28.4s }, [x20], #0x10\n"
       "tbz x8, #1, 199f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "str d17, [x23], #0x8\n"
       "str d21, [x22], #0x8\n"
       "str d25, [x21], #0x8\n"
       "str d29, [x20], #0x8\n"
       "tbz x8, #0, 202f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "st1 { v17.s }[2], [x23]\n"
       "st1 { v21.s }[2], [x22]\n"
@@ -3415,7 +3414,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 202f\n"
       "199:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x8, #0, 202f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "str s17, [x23, #0x0]\n"
       "str s21, [x22, #0x0]\n"
@@ -3424,14 +3423,14 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "b 202f\n"
       "200:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x8, #1, 201f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "str d16, [x23], #0x8\n"
       "str d20, [x22], #0x8\n"
       "str d24, [x21], #0x8\n"
       "str d28, [x20], #0x8\n"
       "tbz x8, #0, 202f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "st1 { v16.s }[2], [x23]\n"
       "st1 { v20.s }[2], [x22]\n"
@@ -3439,7 +3438,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "st1 { v28.s }[2], [x20]\n"
       "b 202f\n"
       "201:"  // Height 6: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "str s16, [x23, #0x0]\n"
       "str s20, [x22, #0x0]\n"
@@ -3448,11 +3447,11 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "202:"  // Height 6: Partial direct writeback: Done
       "b 204f\n"
       "203:"  // Height 6: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -3488,8 +3487,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "206:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp
index 0950d7d950..ba2f77f541 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp
@@ -44,18 +44,18 @@ void a64_hybrid_s8s32_dot_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -88,7 +88,7 @@ void a64_hybrid_s8s32_dot_6x16 (
       "beq 35f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "tbz %x[flags], #0, 12f\n"
       "cmp x11, #0x10\n"
@@ -163,8 +163,8 @@ void a64_hybrid_s8s32_dot_6x16 (
       "mov x28, #0x0\n"
       "14:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 15f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -188,10 +188,6 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       "ldr q17, [x10, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
@@ -216,21 +212,22 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr q17, [x10, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
       "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr q0, [x26, #0x0]\n"
+      "cmp x27, #0x20\n"
+      "add x10, x10, #0x100\n"
+      "ldr q6, [x10, #0x0]\n"
       "ldr q7, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 17b\n"
       "18:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x4f80e0c8  // sdot v8.4s, v6.16b, v0.4b[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       "ldr q17, [x10, #0x40]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
@@ -255,26 +252,29 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr q17, [x10, #0xe0]\n"
       ".inst 0x4fa0ea09  // sdot v9.4s, v16.16b, v0.4b[3]\n"
       "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4fa0ea2a  // sdot v10.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x4fa0ea0b  // sdot v11.4s, v16.16b, v0.4b[3]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "19:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 24f\n"
       "cmp x27, #0x4\n"
       "blt 21f\n"
       "20:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x26], #0x4\n"
-      "ldr q17, [x10, #0x0]\n"
+      "ldr q16, [x10, #0x0]\n"
+      ".inst 0x4f92e208  // sdot v8.4s, v16.16b, v18.4b[0]\n"
       "sub x27, x27, #0x4\n"
       "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
-      ".inst 0x4f92e228  // sdot v8.4s, v17.16b, v18.4b[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f92e209  // sdot v9.4s, v16.16b, v18.4b[0]\n"
+      "cmp x27, #0x4\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x4f92e22a  // sdot v10.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x4f92e20b  // sdot v11.4s, v16.16b, v18.4b[0]\n"
+      "add x10, x10, #0x40\n"
       "bge 20b\n"
       "21:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 24f\n"
@@ -289,12 +289,12 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr q17, [x10, #0x0]\n"
       "ldr q16, [x10, #0x10]\n"
       ".inst 0x4f80e228  // sdot v8.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f80e209  // sdot v9.4s, v16.16b, v0.4b[0]\n"
+      "ldr q17, [x10, #0x20]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
+      "add x10, x10, #0x40\n"
       "24:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -362,7 +362,7 @@ void a64_hybrid_s8s32_dot_6x16 (
       "35:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "36:"  // Height 2: Column loop
       "tbz %x[flags], #0, 46f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
@@ -463,8 +463,8 @@ void a64_hybrid_s8s32_dot_6x16 (
       "mov x28, #0x0\n"
       "48:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 49f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -496,22 +496,22 @@ void a64_hybrid_s8s32_dot_6x16 (
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
       "ldr q17, [x10, #0x40]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22c  // sdot v12.4s, v17.16b, v1.4b[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20d  // sdot v13.4s, v16.16b, v1.4b[1]\n"
       "ldr q16, [x10, #0x70]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4fa0e22a  // sdot v10.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22e  // sdot v14.4s, v17.16b, v1.4b[1]\n"
       "ldr q17, [x10, #0x80]\n"
@@ -555,18 +555,18 @@ void a64_hybrid_s8s32_dot_6x16 (
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
       "ldr q17, [x10, #0x40]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4fa0e228  // sdot v8.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x4fa1e22c  // sdot v12.4s, v17.16b, v1.4b[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4fa0e209  // sdot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x4fa1e20d  // sdot v13.4s, v16.16b, v1.4b[1]\n"
       "ldr q16, [x10, #0x70]\n"
@@ -607,18 +607,18 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr s19, [x26], #0x4\n"
       "ldr s18, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr q17, [x10, #0x0]\n"
       "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
       ".inst 0x4f93e228  // sdot v8.4s, v17.16b, v19.4b[0]\n"
       ".inst 0x4f92e22c  // sdot v12.4s, v17.16b, v18.4b[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x4f93e209  // sdot v9.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x4f92e20d  // sdot v13.4s, v16.16b, v18.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x4f93e22a  // sdot v10.4s, v17.16b, v19.4b[0]\n"
       ".inst 0x4f92e22e  // sdot v14.4s, v17.16b, v18.4b[0]\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x4f93e20b  // sdot v11.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x4f92e20f  // sdot v15.4s, v16.16b, v18.4b[0]\n"
       "bge 54b\n"
@@ -643,9 +643,9 @@ void a64_hybrid_s8s32_dot_6x16 (
       ".inst 0x4f80e209  // sdot v9.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20d  // sdot v13.4s, v16.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x4f80e22a  // sdot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x4f81e22e  // sdot v14.4s, v17.16b, v1.4b[0]\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x4f80e20b  // sdot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x4f81e20f  // sdot v15.4s, v16.16b, v1.4b[0]\n"
       "58:"  // Height 2: Multiply loop: No odd multiplies
@@ -654,9 +654,9 @@ void a64_hybrid_s8s32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 48b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
       "cmp x11, #0x10\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x24, x9, x20, LSL #2\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "bge 67f\n"
       "tbz x11, #3, 62f\n"
@@ -738,12 +738,12 @@ void a64_hybrid_s8s32_dot_6x16 (
       "69:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "70:"  // Height 3: Column loop
       "tbz %x[flags], #0, 80f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x23, x24, x20, LSL #2\n"
       "bge 79f\n"
       "tbz x11, #3, 74f\n"
@@ -864,8 +864,8 @@ void a64_hybrid_s8s32_dot_6x16 (
       "mov x28, #0x0\n"
       "82:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 83f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -905,18 +905,18 @@ void a64_hybrid_s8s32_dot_6x16 (
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "add x24, x24, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
+      "cmp x27, #0x20\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f81e28f  // sdot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e293  // sdot v19.4s, v20.16b, v2.4b[0]\n"
       "ldr q20, [x10, #0x50]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4fa0e2a8  // sdot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x4fa1e2ac  // sdot v12.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x4fa2e2b0  // sdot v16.4s, v21.16b, v2.4b[1]\n"
@@ -983,14 +983,14 @@ void a64_hybrid_s8s32_dot_6x16 (
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e2aa  // sdot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x4f81e2ae  // sdot v14.4s, v21.16b, v1.4b[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f82e2b2  // sdot v18.4s, v21.16b, v2.4b[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f80e28b  // sdot v11.4s, v20.16b, v0.4b[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f81e28f  // sdot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x4f82e293  // sdot v19.4s, v20.16b, v2.4b[0]\n"
       "ldr q20, [x10, #0x50]\n"
@@ -1049,12 +1049,12 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr s24, [x26], #0x4\n"
       "ldr s23, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s22, [x24], #0x4\n"
       "ldr q21, [x10, #0x0]\n"
-      "cmp x27, #0x4\n"
-      "ldr q20, [x10, #0x10]\n"
       ".inst 0x4f98e2a8  // sdot v8.4s, v21.16b, v24.4b[0]\n"
       ".inst 0x4f97e2ac  // sdot v12.4s, v21.16b, v23.4b[0]\n"
+      "ldr q20, [x10, #0x10]\n"
       ".inst 0x4f96e2b0  // sdot v16.4s, v21.16b, v22.4b[0]\n"
       "ldr q21, [x10, #0x20]\n"
       ".inst 0x4f98e289  // sdot v9.4s, v20.16b, v24.4b[0]\n"
@@ -1108,11 +1108,11 @@ void a64_hybrid_s8s32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 82b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "cmp x11, #0x10\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "bge 101f\n"
       "tbz x11, #3, 96f\n"
@@ -1214,13 +1214,13 @@ void a64_hybrid_s8s32_dot_6x16 (
       "103:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "104:"  // Height 4: Column loop
       "tbz %x[flags], #0, 114f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x22, x23, x20, LSL #2\n"
       "bge 113f\n"
       "tbz x11, #3, 108f\n"
@@ -1365,8 +1365,8 @@ void a64_hybrid_s8s32_dot_6x16 (
       "mov x28, #0x0\n"
       "116:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 117f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1506,14 +1506,14 @@ void a64_hybrid_s8s32_dot_6x16 (
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
       "ldr q24, [x10, #0x30]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4f80e32a  // sdot v10.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x4f81e32e  // sdot v14.4s, v25.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f82e332  // sdot v18.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x4f83e336  // sdot v22.4s, v25.16b, v3.4b[0]\n"
@@ -1591,9 +1591,9 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr s29, [x26], #0x4\n"
       "ldr s28, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s27, [x24], #0x4\n"
       "ldr s26, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
       "ldr q25, [x10, #0x0]\n"
       "ldr q24, [x10, #0x10]\n"
       ".inst 0x4f9de328  // sdot v8.4s, v25.16b, v29.4b[0]\n"
@@ -1662,13 +1662,13 @@ void a64_hybrid_s8s32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 116b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
       "bge 135f\n"
       "tbz x11, #3, 130f\n"
@@ -1790,14 +1790,14 @@ void a64_hybrid_s8s32_dot_6x16 (
       "137:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "138:"  // Height 5: Column loop
       "tbz %x[flags], #0, 148f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x21, x22, x20, LSL #2\n"
       "bge 147f\n"
       "tbz x11, #3, 142f\n"
@@ -1966,8 +1966,8 @@ void a64_hybrid_s8s32_dot_6x16 (
       "mov x28, #0x0\n"
       "150:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 151f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2133,12 +2133,12 @@ void a64_hybrid_s8s32_dot_6x16 (
       "add x22, x22, #0x10\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
+      "sub x27, x27, #0x10\n"
       "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
       "ldr q28, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4f80e3aa  // sdot v10.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f81e3ae  // sdot v14.4s, v29.16b, v1.4b[0]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
@@ -2233,14 +2233,14 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr s2, [x26], #0x4\n"
       "ldr s1, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s0, [x24], #0x4\n"
       "ldr s31, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
       "ldr s30, [x22], #0x4\n"
       "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
       ".inst 0x4f82e3a8  // sdot v8.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x4f81e3ac  // sdot v12.4s, v29.16b, v1.4b[0]\n"
+      "ldr q28, [x10, #0x10]\n"
       ".inst 0x4f80e3b0  // sdot v16.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x4f9fe3b4  // sdot v20.4s, v29.16b, v31.4b[0]\n"
       ".inst 0x4f9ee3b8  // sdot v24.4s, v29.16b, v30.4b[0]\n"
@@ -2316,15 +2316,15 @@ void a64_hybrid_s8s32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 150b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "add x21, x22, x20, LSL #2\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "cmp x11, #0x10\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "prfm pstl1keep, [x21, #0x0]\n"
       "bge 169f\n"
       "tbz x11, #3, 164f\n"
@@ -2465,20 +2465,19 @@ void a64_hybrid_s8s32_dot_6x16 (
       "b 206f\n"
       "171:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "172:"  // Height 6: Column loop
       "tbz %x[flags], #0, 182f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x20, x21, x20, LSL #2\n"
       "bge 181f\n"
       "tbz x11, #3, 176f\n"
@@ -2671,8 +2670,8 @@ void a64_hybrid_s8s32_dot_6x16 (
       "mov x28, #0x0\n"
       "184:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 185f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2862,18 +2861,18 @@ void a64_hybrid_s8s32_dot_6x16 (
       ".inst 0x4f80e0e9  // sdot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ed  // sdot v13.4s, v7.16b, v1.4b[0]\n"
       "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4f82e0f1  // sdot v17.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x4f83e0f5  // sdot v21.4s, v7.16b, v3.4b[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f84e0f9  // sdot v25.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x4f85e0fd  // sdot v29.4s, v7.16b, v5.4b[0]\n"
       "ldr q7, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4f80e0ca  // sdot v10.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x4f81e0ce  // sdot v14.4s, v6.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4f82e0d2  // sdot v18.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x4f83e0d6  // sdot v22.4s, v6.16b, v3.4b[0]\n"
@@ -2979,9 +2978,9 @@ void a64_hybrid_s8s32_dot_6x16 (
       "ldr s7, [x26], #0x4\n"
       "ldr s6, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s5, [x24], #0x4\n"
       "ldr s4, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
       "ldr s3, [x22], #0x4\n"
       "ldr s2, [x21], #0x4\n"
       "ldr q1, [x10, #0x0]\n"
@@ -3074,16 +3073,16 @@ void a64_hybrid_s8s32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 184b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "add x21, x22, x20, LSL #2\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "add x20, x21, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "prfm pstl1keep, [x21, #0x0]\n"
       "prfm pstl1keep, [x20, #0x0]\n"
       "bge 203f\n"
@@ -3254,8 +3253,8 @@ void a64_hybrid_s8s32_dot_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "206:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16.hpp
index c5170553d8..0f7481f0da 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16.hpp
@@ -70,7 +70,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 8> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp
index c6e982e20d..98b4d9b997 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp
@@ -44,18 +44,18 @@ void a64_hybrid_s8s32_mmla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -88,7 +88,7 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "beq 38f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "tbz %x[flags], #0, 13f\n"
       "cmp x11, #0x10\n"
@@ -176,8 +176,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "mov x28, #0x0\n"
       "15:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 16f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -197,12 +197,7 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q6, [x10, #0x10]\n"
       "blt 19f\n"
       "18:"  // Height 1: Multiply loop: Main loop head
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "trn1 v19.2d, v1.2d, v20.2d\n"
-      "trn2 v1.2d, v1.2d, v20.2d\n"
       ".inst 0x4e87a668  // smmla v8.4s, v19.16b, v7.16b\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x4e86a66c  // smmla v12.4s, v19.16b, v6.16b\n"
@@ -215,6 +210,7 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x4e91a66e  // smmla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v20.2d\n"
       ".inst 0x4e92a66b  // smmla v11.4s, v19.16b, v18.16b\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x4e91a66f  // smmla v15.4s, v19.16b, v17.16b\n"
@@ -231,38 +227,39 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x4e91a42e  // smmla v14.4s, v1.16b, v17.16b\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e92a42b  // smmla v11.4s, v1.16b, v18.16b\n"
-      "ldr q7, [x10, #0x0]\n"
       ".inst 0x4e91a42f  // smmla v15.4s, v1.16b, v17.16b\n"
       "ldr q1, [x26, #0x0]\n"
+      "add x10, x10, #0x100\n"
+      "ldr q7, [x10, #0x0]\n"
       "ldr q6, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 18b\n"
       "19:"  // Height 1: Multiply loop: Single iteration only
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "trn1 v19.2d, v1.2d, v17.2d\n"
-      "trn2 v1.2d, v1.2d, v17.2d\n"
-      ".inst 0x4e87a668  // smmla v8.4s, v19.16b, v7.16b\n"
+      "trn1 v20.2d, v1.2d, v21.2d\n"
+      ".inst 0x4e87a688  // smmla v8.4s, v20.16b, v7.16b\n"
       "ldr q18, [x10, #0x20]\n"
-      ".inst 0x4e86a66c  // smmla v12.4s, v19.16b, v6.16b\n"
+      ".inst 0x4e86a68c  // smmla v12.4s, v20.16b, v6.16b\n"
       "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e92a669  // smmla v9.4s, v19.16b, v18.16b\n"
+      ".inst 0x4e92a689  // smmla v9.4s, v20.16b, v18.16b\n"
       "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e91a66d  // smmla v13.4s, v19.16b, v17.16b\n"
+      ".inst 0x4e91a68d  // smmla v13.4s, v20.16b, v17.16b\n"
       "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92a66a  // smmla v10.4s, v19.16b, v18.16b\n"
-      "ldr q20, [x10, #0x60]\n"
-      ".inst 0x4e91a66e  // smmla v14.4s, v19.16b, v17.16b\n"
+      ".inst 0x4e92a68a  // smmla v10.4s, v20.16b, v18.16b\n"
+      "ldr q19, [x10, #0x60]\n"
+      ".inst 0x4e91a68e  // smmla v14.4s, v20.16b, v17.16b\n"
       "ldr q18, [x10, #0x70]\n"
-      ".inst 0x4e94a66b  // smmla v11.4s, v19.16b, v20.16b\n"
+      "trn2 v1.2d, v1.2d, v21.2d\n"
+      ".inst 0x4e93a68b  // smmla v11.4s, v20.16b, v19.16b\n"
       "ldr q17, [x10, #0x80]\n"
-      ".inst 0x4e92a66f  // smmla v15.4s, v19.16b, v18.16b\n"
-      "ldr q20, [x10, #0x90]\n"
+      ".inst 0x4e92a68f  // smmla v15.4s, v20.16b, v18.16b\n"
+      "ldr q19, [x10, #0x90]\n"
       ".inst 0x4e91a428  // smmla v8.4s, v1.16b, v17.16b\n"
       "ldr q18, [x10, #0xa0]\n"
-      ".inst 0x4e94a42c  // smmla v12.4s, v1.16b, v20.16b\n"
+      ".inst 0x4e93a42c  // smmla v12.4s, v1.16b, v19.16b\n"
       "ldr q17, [x10, #0xb0]\n"
       ".inst 0x4e92a429  // smmla v9.4s, v1.16b, v18.16b\n"
       "ldr q18, [x10, #0xc0]\n"
@@ -272,21 +269,22 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x4e91a42e  // smmla v14.4s, v1.16b, v17.16b\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e92a42b  // smmla v11.4s, v1.16b, v18.16b\n"
       ".inst 0x4e91a42f  // smmla v15.4s, v1.16b, v17.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "20:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 27f\n"
       "cmp x27, #0x8\n"
       "blt 22f\n"
       "21:"  // Height 1: Multiply loop: Odd block loop
       "ldr d19, [x26], #0x8\n"
-      "ldr q20, [x10, #0x0]\n"
-      "sub x27, x27, #0x8\n"
+      "ldr q18, [x10, #0x0]\n"
+      "trn1 v19.2d, v19.2d, v17.2d\n"
       "ldr q17, [x10, #0x10]\n"
-      "cmp x27, #0x8\n"
-      "trn1 v19.2d, v19.2d, v18.2d\n"
-      ".inst 0x4e94a668  // smmla v8.4s, v19.16b, v20.16b\n"
+      ".inst 0x4e92a668  // smmla v8.4s, v19.16b, v18.16b\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x4e91a66c  // smmla v12.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x30]\n"
@@ -298,9 +296,11 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x4e91a66e  // smmla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
       ".inst 0x4e92a66b  // smmla v11.4s, v19.16b, v18.16b\n"
       ".inst 0x4e91a66f  // smmla v15.4s, v19.16b, v17.16b\n"
+      "add x10, x10, #0x80\n"
       "bge 21b\n"
       "22:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 27f\n"
@@ -324,24 +324,24 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "25:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr b1, [x26, #0x0]\n"
       "26:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q24, [x10, #0x0]\n"
-      "ldr q20, [x10, #0x10]\n"
+      "ldr q23, [x10, #0x0]\n"
+      "ldr q18, [x10, #0x10]\n"
       "trn1 v19.2d, v1.2d, v17.2d\n"
-      ".inst 0x4e98a668  // smmla v8.4s, v19.16b, v24.16b\n"
+      ".inst 0x4e97a668  // smmla v8.4s, v19.16b, v23.16b\n"
       "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4e94a66c  // smmla v12.4s, v19.16b, v20.16b\n"
-      "ldr q0, [x10, #0x30]\n"
+      ".inst 0x4e92a66c  // smmla v12.4s, v19.16b, v18.16b\n"
+      "ldr q31, [x10, #0x30]\n"
       ".inst 0x4e91a669  // smmla v9.4s, v19.16b, v17.16b\n"
       "ldr q20, [x10, #0x40]\n"
-      ".inst 0x4e80a66d  // smmla v13.4s, v19.16b, v0.16b\n"
+      ".inst 0x4e9fa66d  // smmla v13.4s, v19.16b, v31.16b\n"
       "ldr q17, [x10, #0x50]\n"
       ".inst 0x4e94a66a  // smmla v10.4s, v19.16b, v20.16b\n"
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x4e91a66e  // smmla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x4e92a66b  // smmla v11.4s, v19.16b, v18.16b\n"
       ".inst 0x4e91a66f  // smmla v15.4s, v19.16b, v17.16b\n"
+      "add x10, x10, #0x80\n"
       "27:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -413,7 +413,7 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "38:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "39:"  // Height 2: Column loop
       "tbz %x[flags], #0, 50f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
@@ -523,8 +523,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "mov x28, #0x0\n"
       "52:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 53f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -549,14 +549,6 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "blt 56f\n"
       "55:"  // Height 2: Multiply loop: Main loop head
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e87a668  // smmla v8.4s, v19.16b, v7.16b\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x4e86a66c  // smmla v12.4s, v19.16b, v6.16b\n"
@@ -569,6 +561,7 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x4e91a66e  // smmla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x4e92a66b  // smmla v11.4s, v19.16b, v18.16b\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x4e91a66f  // smmla v15.4s, v19.16b, v17.16b\n"
@@ -585,21 +578,22 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x4e91a42e  // smmla v14.4s, v1.16b, v17.16b\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "ldr q2, [x25, #0x0]\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e92a42b  // smmla v11.4s, v1.16b, v18.16b\n"
+      "add x10, x10, #0x100\n"
       "ldr q7, [x10, #0x0]\n"
       ".inst 0x4e91a42f  // smmla v15.4s, v1.16b, v17.16b\n"
       "ldr q1, [x26, #0x0]\n"
       "ldr q6, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "bge 55b\n"
       "56:"  // Height 2: Multiply loop: Single iteration only
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e87a668  // smmla v8.4s, v19.16b, v7.16b\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x4e86a66c  // smmla v12.4s, v19.16b, v6.16b\n"
@@ -612,6 +606,7 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x4e91a66e  // smmla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x4e92a66b  // smmla v11.4s, v19.16b, v18.16b\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x4e91a66f  // smmla v15.4s, v19.16b, v17.16b\n"
@@ -628,36 +623,41 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x4e91a42e  // smmla v14.4s, v1.16b, v17.16b\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e92a42b  // smmla v11.4s, v1.16b, v18.16b\n"
       ".inst 0x4e91a42f  // smmla v15.4s, v1.16b, v17.16b\n"
+      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "57:"  // Height 2: Multiply loop: Main loop skip
       "cbz x27, 64f\n"
       "cmp x27, #0x8\n"
       "blt 59f\n"
       "58:"  // Height 2: Multiply loop: Odd block loop
-      "ldr d20, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d18, [x26], #0x8\n"
+      "ldr d17, [x25], #0x8\n"
+      "trn1 v19.2d, v18.2d, v17.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr q18, [x10, #0x0]\n"
-      "ldr q17, [x10, #0x10]\n"
-      "cmp x27, #0x8\n"
-      "trn1 v22.2d, v20.2d, v19.2d\n"
-      ".inst 0x4e92a6c8  // smmla v8.4s, v22.16b, v18.16b\n"
-      "ldr q2, [x10, #0x20]\n"
-      ".inst 0x4e91a6cc  // smmla v12.4s, v22.16b, v17.16b\n"
+      "ldr q17, [x10, #0x0]\n"
+      "ldr q22, [x10, #0x10]\n"
+      ".inst 0x4e91a668  // smmla v8.4s, v19.16b, v17.16b\n"
+      ".inst 0x4e96a66c  // smmla v12.4s, v19.16b, v22.16b\n"
+      "ldr q1, [x10, #0x20]\n"
       "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e82a6c9  // smmla v9.4s, v22.16b, v2.16b\n"
+      ".inst 0x4e81a669  // smmla v9.4s, v19.16b, v1.16b\n"
+      ".inst 0x4e91a66d  // smmla v13.4s, v19.16b, v17.16b\n"
       "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e91a6cd  // smmla v13.4s, v22.16b, v17.16b\n"
       "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92a6ca  // smmla v10.4s, v22.16b, v18.16b\n"
+      ".inst 0x4e92a66a  // smmla v10.4s, v19.16b, v18.16b\n"
+      ".inst 0x4e91a66e  // smmla v14.4s, v19.16b, v17.16b\n"
       "ldr q18, [x10, #0x60]\n"
-      ".inst 0x4e91a6ce  // smmla v14.4s, v22.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x4e92a66b  // smmla v11.4s, v19.16b, v18.16b\n"
+      ".inst 0x4e91a66f  // smmla v15.4s, v19.16b, v17.16b\n"
       "add x10, x10, #0x80\n"
-      ".inst 0x4e92a6cb  // smmla v11.4s, v22.16b, v18.16b\n"
-      ".inst 0x4e91a6cf  // smmla v15.4s, v22.16b, v17.16b\n"
       "bge 58b\n"
       "59:"  // Height 2: Multiply loop: Skip odd blocks
       "cbz x27, 64f\n"
@@ -703,27 +703,27 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x4e91a66e  // smmla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x4e92a66b  // smmla v11.4s, v19.16b, v18.16b\n"
       ".inst 0x4e91a66f  // smmla v15.4s, v19.16b, v17.16b\n"
+      "add x10, x10, #0x80\n"
       "64:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 52b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
       "cmp x11, #0x10\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
       "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "bge 73f\n"
       "tbz x11, #3, 68f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
@@ -804,12 +804,12 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "75:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "76:"  // Height 3: Column loop
       "tbz %x[flags], #0, 87f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x23, x24, x20, LSL #2\n"
       "bge 85f\n"
       "tbz x11, #3, 80f\n"
@@ -951,8 +951,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "mov x28, #0x0\n"
       "89:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 90f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -982,38 +982,35 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "92:"  // Height 3: Multiply loop: Main loop head
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x4e87a788  // smmla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x4e86a78c  // smmla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
+      "trn1 v27.2d, v3.2d, v29.2d\n"
       ".inst 0x4e87a770  // smmla v16.4s, v27.16b, v7.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x4e86a78c  // smmla v12.4s, v28.16b, v6.16b\n"
       ".inst 0x4e86a774  // smmla v20.4s, v27.16b, v6.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x4e9aa789  // smmla v9.4s, v28.16b, v26.16b\n"
+      "trn2 v3.2d, v3.2d, v29.2d\n"
       ".inst 0x4e9aa771  // smmla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x4e99a78d  // smmla v13.4s, v28.16b, v25.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e99a775  // smmla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x4e9aa78a  // smmla v10.4s, v28.16b, v26.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4e9aa772  // smmla v18.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x4e99a78e  // smmla v14.4s, v28.16b, v25.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e99a776  // smmla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x4e9aa78b  // smmla v11.4s, v28.16b, v26.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e9aa773  // smmla v19.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x4e99a78f  // smmla v15.4s, v28.16b, v25.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e99a777  // smmla v23.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x90]\n"
       "ldr q2, [x25, #0x0]\n"
@@ -1021,12 +1018,15 @@ void a64_hybrid_s8s32_mmla_6x16 (
       ".inst 0x4e9aa470  // smmla v16.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x4e99a42c  // smmla v12.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e99a474  // smmla v20.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x4e9aa429  // smmla v9.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e9aa471  // smmla v17.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x4e99a42d  // smmla v13.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e99a475  // smmla v21.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xd0]\n"
       ".inst 0x4e9aa42a  // smmla v10.4s, v1.16b, v26.16b\n"
@@ -1048,43 +1048,43 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "93:"  // Height 3: Multiply loop: Single iteration only
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x4e87a788  // smmla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x4e86a78c  // smmla v12.4s, v28.16b, v6.16b\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
+      "trn1 v27.2d, v3.2d, v29.2d\n"
       ".inst 0x4e87a770  // smmla v16.4s, v27.16b, v7.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x4e86a78c  // smmla v12.4s, v28.16b, v6.16b\n"
       ".inst 0x4e86a774  // smmla v20.4s, v27.16b, v6.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x4e9aa789  // smmla v9.4s, v28.16b, v26.16b\n"
+      "trn2 v3.2d, v3.2d, v29.2d\n"
       ".inst 0x4e9aa771  // smmla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x4e99a78d  // smmla v13.4s, v28.16b, v25.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4e99a775  // smmla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x4e9aa78a  // smmla v10.4s, v28.16b, v26.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e9aa772  // smmla v18.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x4e99a78e  // smmla v14.4s, v28.16b, v25.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e99a776  // smmla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x4e9aa78b  // smmla v11.4s, v28.16b, v26.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e9aa773  // smmla v19.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x4e99a78f  // smmla v15.4s, v28.16b, v25.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e99a777  // smmla v23.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x90]\n"
       ".inst 0x4e9aa428  // smmla v8.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e9aa470  // smmla v16.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x4e99a42c  // smmla v12.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e99a474  // smmla v20.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x4e9aa429  // smmla v9.4s, v1.16b, v26.16b\n"
@@ -1109,25 +1109,25 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 96f\n"
       "95:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "sub x27, x27, #0x8\n"
-      "ldr d27, [x24], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "trn1 v28.2d, v26.2d, v25.2d\n"
+      "ldr d25, [x24], #0x8\n"
       "ldr q26, [x10, #0x0]\n"
-      "cmp x27, #0x8\n"
-      "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v27.2d, v29.2d\n"
+      "trn1 v27.2d, v25.2d, v27.2d\n"
       ".inst 0x4e9aa788  // smmla v8.4s, v28.16b, v26.16b\n"
+      "ldr q25, [x10, #0x10]\n"
       ".inst 0x4e9aa770  // smmla v16.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x4e99a78c  // smmla v12.4s, v28.16b, v25.16b\n"
       ".inst 0x4e99a774  // smmla v20.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x4e9aa789  // smmla v9.4s, v28.16b, v26.16b\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x4e9aa771  // smmla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x4e99a78d  // smmla v13.4s, v28.16b, v25.16b\n"
+      "cmp x27, #0x8\n"
       ".inst 0x4e99a775  // smmla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x4e9aa78a  // smmla v10.4s, v28.16b, v26.16b\n"
@@ -1136,8 +1136,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       ".inst 0x4e99a78e  // smmla v14.4s, v28.16b, v25.16b\n"
       ".inst 0x4e99a776  // smmla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x4e9aa78b  // smmla v11.4s, v28.16b, v26.16b\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x4e9aa773  // smmla v19.4s, v27.16b, v26.16b\n"
       ".inst 0x4e99a78f  // smmla v15.4s, v28.16b, v25.16b\n"
       ".inst 0x4e99a777  // smmla v23.4s, v27.16b, v25.16b\n"
@@ -1183,9 +1183,9 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x4e9aa788  // smmla v8.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9da78c  // smmla v12.4s, v28.16b, v29.16b\n"
       ".inst 0x4e9aa770  // smmla v16.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x4e9da78c  // smmla v12.4s, v28.16b, v29.16b\n"
       ".inst 0x4e9da774  // smmla v20.4s, v27.16b, v29.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x4e9aa789  // smmla v9.4s, v28.16b, v26.16b\n"
@@ -1211,20 +1211,20 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 89b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "cmp x11, #0x10\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
       "prfm pstl1keep, [x9, #0x0]\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v18.2d, v18.2d, v22.2d\n"
@@ -1329,13 +1329,13 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "112:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "113:"  // Height 4: Column loop
       "tbz %x[flags], #0, 124f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x22, x23, x20, LSL #2\n"
       "bge 122f\n"
       "tbz x11, #3, 117f\n"
@@ -1497,8 +1497,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "mov x28, #0x0\n"
       "126:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 127f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1532,38 +1532,33 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "129:"  // Height 4: Multiply loop: Main loop head
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e87a788  // smmla v8.4s, v28.16b, v7.16b\n"
       "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "ldr q4, [x23, #0x0]\n"
-      ".inst 0x4e87a788  // smmla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x4e86a78c  // smmla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e87a770  // smmla v16.4s, v27.16b, v7.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x4e86a78c  // smmla v12.4s, v28.16b, v6.16b\n"
       ".inst 0x4e86a774  // smmla v20.4s, v27.16b, v6.16b\n"
       "ldr q25, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e9aa789  // smmla v9.4s, v28.16b, v26.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x4e9aa771  // smmla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x4e99a78d  // smmla v13.4s, v28.16b, v25.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4e99a775  // smmla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x4e9aa78a  // smmla v10.4s, v28.16b, v26.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e9aa772  // smmla v18.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x4e99a78e  // smmla v14.4s, v28.16b, v25.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e99a776  // smmla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x4e9aa78b  // smmla v11.4s, v28.16b, v26.16b\n"
+      "add x23, x23, #0x10\n"
+      "ldr q4, [x23, #0x0]\n"
       ".inst 0x4e9aa773  // smmla v19.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x4e99a78f  // smmla v15.4s, v28.16b, v25.16b\n"
@@ -1574,18 +1569,23 @@ void a64_hybrid_s8s32_mmla_6x16 (
       ".inst 0x4e9aa470  // smmla v16.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x4e99a42c  // smmla v12.4s, v1.16b, v25.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e99a474  // smmla v20.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x4e9aa429  // smmla v9.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e9aa471  // smmla v17.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x4e99a42d  // smmla v13.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e99a475  // smmla v21.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xd0]\n"
       ".inst 0x4e9aa42a  // smmla v10.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e9aa472  // smmla v18.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xe0]\n"
       ".inst 0x4e99a42e  // smmla v14.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e99a476  // smmla v22.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xf0]\n"
       "add x10, x10, #0x100\n"
@@ -1601,48 +1601,48 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "130:"  // Height 4: Multiply loop: Single iteration only
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e87a788  // smmla v8.4s, v28.16b, v7.16b\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      ".inst 0x4e87a788  // smmla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x4e86a78c  // smmla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e87a770  // smmla v16.4s, v27.16b, v7.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x4e86a78c  // smmla v12.4s, v28.16b, v6.16b\n"
       ".inst 0x4e86a774  // smmla v20.4s, v27.16b, v6.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x4e9aa789  // smmla v9.4s, v28.16b, v26.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x4e9aa771  // smmla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x4e99a78d  // smmla v13.4s, v28.16b, v25.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e99a775  // smmla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x4e9aa78a  // smmla v10.4s, v28.16b, v26.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e9aa772  // smmla v18.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x4e99a78e  // smmla v14.4s, v28.16b, v25.16b\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e99a776  // smmla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x4e9aa78b  // smmla v11.4s, v28.16b, v26.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e9aa773  // smmla v19.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x4e99a78f  // smmla v15.4s, v28.16b, v25.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e99a777  // smmla v23.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x90]\n"
       ".inst 0x4e9aa428  // smmla v8.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e9aa470  // smmla v16.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x4e99a42c  // smmla v12.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e99a474  // smmla v20.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x4e9aa429  // smmla v9.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e9aa471  // smmla v17.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x4e99a42d  // smmla v13.4s, v1.16b, v25.16b\n"
@@ -1664,16 +1664,16 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 133f\n"
       "132:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "trn1 v28.2d, v26.2d, v25.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d26, [x24], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "trn1 v27.2d, v26.2d, v25.2d\n"
       "cmp x27, #0x8\n"
       "ldr q26, [x10, #0x0]\n"
       "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v29.2d, v27.2d\n"
       ".inst 0x4e9aa788  // smmla v8.4s, v28.16b, v26.16b\n"
       ".inst 0x4e9aa770  // smmla v16.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x20]\n"
@@ -1774,24 +1774,24 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 126b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
@@ -1918,14 +1918,14 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "149:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "150:"  // Height 5: Column loop
       "tbz %x[flags], #0, 161f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x21, x22, x20, LSL #2\n"
       "bge 159f\n"
       "tbz x11, #3, 154f\n"
@@ -2123,8 +2123,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "mov x28, #0x0\n"
       "163:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 164f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2161,51 +2161,51 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "166:"  // Height 5: Multiply loop: Main loop head
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
+      ".inst 0x4e87a4c8  // smmla v8.4s, v6.16b, v7.16b\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
+      "sub x27, x27, #0x10\n"
       "trn1 v4.2d, v5.2d, v0.2d\n"
       "trn2 v5.2d, v5.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
-      ".inst 0x4e87a4c8  // smmla v8.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
-      "add x22, x22, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e87a498  // smmla v24.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x4e80a4cc  // smmla v12.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a454  // smmla v20.4s, v2.16b, v0.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4e80a49c  // smmla v28.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x4e87a4c9  // smmla v9.4s, v6.16b, v7.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e87a451  // smmla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a499  // smmla v25.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e80a4cd  // smmla v13.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a455  // smmla v21.4s, v2.16b, v0.16b\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4e80a49d  // smmla v29.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x50]\n"
       ".inst 0x4e87a4ca  // smmla v10.4s, v6.16b, v7.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e87a452  // smmla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49a  // smmla v26.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e80a4ce  // smmla v14.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a456  // smmla v22.4s, v2.16b, v0.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e80a49e  // smmla v30.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x70]\n"
       ".inst 0x4e87a4cb  // smmla v11.4s, v6.16b, v7.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e87a453  // smmla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49b  // smmla v27.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e80a4cf  // smmla v15.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a457  // smmla v23.4s, v2.16b, v0.16b\n"
       "ldr q2, [x25, #0x0]\n"
@@ -2251,47 +2251,47 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "167:"  // Height 5: Multiply loop: Single iteration only
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
+      ".inst 0x4e87a4c8  // smmla v8.4s, v6.16b, v7.16b\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v0.2d\n"
       "trn2 v5.2d, v5.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
-      ".inst 0x4e87a4c8  // smmla v8.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e87a498  // smmla v24.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x4e80a4cc  // smmla v12.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a454  // smmla v20.4s, v2.16b, v0.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e80a49c  // smmla v28.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x4e87a4c9  // smmla v9.4s, v6.16b, v7.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e87a451  // smmla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a499  // smmla v25.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e80a4cd  // smmla v13.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a455  // smmla v21.4s, v2.16b, v0.16b\n"
+      "add x22, x22, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e80a49d  // smmla v29.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x50]\n"
       ".inst 0x4e87a4ca  // smmla v10.4s, v6.16b, v7.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e87a452  // smmla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49a  // smmla v26.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e80a4ce  // smmla v14.4s, v6.16b, v0.16b\n"
       ".inst 0x4e80a456  // smmla v22.4s, v2.16b, v0.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e80a49e  // smmla v30.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x70]\n"
       ".inst 0x4e87a4cb  // smmla v11.4s, v6.16b, v7.16b\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e87a453  // smmla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49b  // smmla v27.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x80]\n"
@@ -2335,24 +2335,24 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 170f\n"
       "169:"  // Height 5: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d3, [x24], #0x8\n"
-      "ldr d2, [x23], #0x8\n"
-      "cmp x27, #0x8\n"
       "ldr d0, [x22], #0x8\n"
       "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v3.2d, v2.2d\n"
-      "trn1 v2.2d, v0.2d, v5.2d\n"
-      "ldr q0, [x10, #0x10]\n"
+      "trn1 v2.2d, v0.2d, v2.2d\n"
       ".inst 0x4e81a488  // smmla v8.4s, v4.16b, v1.16b\n"
+      "ldr q0, [x10, #0x10]\n"
       ".inst 0x4e81a470  // smmla v16.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a458  // smmla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x10, #0x20]\n"
       ".inst 0x4e80a48c  // smmla v12.4s, v4.16b, v0.16b\n"
       ".inst 0x4e80a474  // smmla v20.4s, v3.16b, v0.16b\n"
+      "cmp x27, #0x8\n"
       ".inst 0x4e80a45c  // smmla v28.4s, v2.16b, v0.16b\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x4e81a489  // smmla v9.4s, v4.16b, v1.16b\n"
@@ -2371,8 +2371,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       ".inst 0x4e80a476  // smmla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45e  // smmla v30.4s, v2.16b, v0.16b\n"
       "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x4e86a48b  // smmla v11.4s, v4.16b, v6.16b\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x4e86a473  // smmla v19.4s, v3.16b, v6.16b\n"
       ".inst 0x4e86a45b  // smmla v27.4s, v2.16b, v6.16b\n"
       ".inst 0x4e80a48f  // smmla v15.4s, v4.16b, v0.16b\n"
@@ -2471,28 +2471,28 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 163b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "cmp x11, #0x10\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
       "prfm pstl1keep, [x24, #0x0]\n"
-      "add x21, x22, x20, LSL #2\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
@@ -2640,20 +2640,19 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "b 224f\n"
       "186:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "187:"  // Height 6: Column loop
       "tbz %x[flags], #0, 198f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x20, x21, x20, LSL #2\n"
       "bge 196f\n"
       "tbz x11, #3, 191f\n"
@@ -2871,8 +2870,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "mov x28, #0x0\n"
       "200:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 201f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2913,56 +2912,56 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "203:"  // Height 6: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
       "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
+      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
       "ldr q6, [x10, #0x10]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e87a498  // smmla v24.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x4e86a40c  // smmla v12.4s, v0.16b, v6.16b\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e86a454  // smmla v20.4s, v2.16b, v6.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x4e86a49c  // smmla v28.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4e87a409  // smmla v9.4s, v0.16b, v7.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e87a451  // smmla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a499  // smmla v25.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e86a40d  // smmla v13.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a455  // smmla v21.4s, v2.16b, v6.16b\n"
+      "add x22, x22, #0x10\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x4e86a49d  // smmla v29.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x50]\n"
       ".inst 0x4e87a40a  // smmla v10.4s, v0.16b, v7.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x4e87a452  // smmla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49a  // smmla v26.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e86a40e  // smmla v14.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a456  // smmla v22.4s, v2.16b, v6.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x4e86a49e  // smmla v30.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x70]\n"
       ".inst 0x4e87a40b  // smmla v11.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e87a453  // smmla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49b  // smmla v27.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e86a40f  // smmla v15.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a457  // smmla v23.4s, v2.16b, v6.16b\n"
       "ldr q2, [x25, #0x0]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4e86a49f  // smmla v31.4s, v4.16b, v6.16b\n"
       "ldr q0, [x10, #0x90]\n"
       "ldr q4, [x23, #0x0]\n"
@@ -3006,52 +3005,52 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "204:"  // Height 6: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
+      "add x25, x25, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
       "ldr q6, [x10, #0x10]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4e87a450  // smmla v16.4s, v2.16b, v7.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e87a498  // smmla v24.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x4e86a40c  // smmla v12.4s, v0.16b, v6.16b\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e86a454  // smmla v20.4s, v2.16b, v6.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4e86a49c  // smmla v28.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4e87a409  // smmla v9.4s, v0.16b, v7.16b\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x4e87a451  // smmla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a499  // smmla v25.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4e86a40d  // smmla v13.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a455  // smmla v21.4s, v2.16b, v6.16b\n"
+      "add x21, x21, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x4e86a49d  // smmla v29.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x50]\n"
       ".inst 0x4e87a40a  // smmla v10.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x4e87a452  // smmla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49a  // smmla v26.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x4e86a40e  // smmla v14.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a456  // smmla v22.4s, v2.16b, v6.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x4e86a49e  // smmla v30.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x70]\n"
       ".inst 0x4e87a40b  // smmla v11.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x4e87a453  // smmla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x4e87a49b  // smmla v27.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x4e86a40f  // smmla v15.4s, v0.16b, v6.16b\n"
       ".inst 0x4e86a457  // smmla v23.4s, v2.16b, v6.16b\n"
       ".inst 0x4e86a49f  // smmla v31.4s, v4.16b, v6.16b\n"
@@ -3092,18 +3091,18 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 207f\n"
       "206:"  // Height 6: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d5, [x24], #0x8\n"
-      "ldr d3, [x23], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "cmp x27, #0x8\n"
-      "ldr d2, [x22], #0x8\n"
+      "ldr d1, [x22], #0x8\n"
       "ldr d0, [x21], #0x8\n"
+      "trn1 v2.2d, v1.2d, v0.2d\n"
       "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v5.2d, v3.2d\n"
-      "trn1 v2.2d, v2.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
       ".inst 0x4e81a488  // smmla v8.4s, v4.16b, v1.16b\n"
       ".inst 0x4e81a470  // smmla v16.4s, v3.16b, v1.16b\n"
@@ -3197,9 +3196,9 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "ldr q0, [x10, #0x0]\n"
       "trn1 v7.2d, v1.2d, v2.2d\n"
       "trn1 v3.2d, v3.2d, v4.2d\n"
+      ".inst 0x4e80a4e8  // smmla v8.4s, v7.16b, v0.16b\n"
       "trn1 v2.2d, v5.2d, v6.2d\n"
       "ldr q1, [x10, #0x10]\n"
-      ".inst 0x4e80a4e8  // smmla v8.4s, v7.16b, v0.16b\n"
       ".inst 0x4e80a470  // smmla v16.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a458  // smmla v24.4s, v2.16b, v0.16b\n"
       "ldr q0, [x10, #0x20]\n"
@@ -3223,8 +3222,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       ".inst 0x4e81a476  // smmla v22.4s, v3.16b, v1.16b\n"
       ".inst 0x4e81a45e  // smmla v30.4s, v2.16b, v1.16b\n"
       "ldr q6, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x4e80a4eb  // smmla v11.4s, v7.16b, v0.16b\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x4e80a473  // smmla v19.4s, v3.16b, v0.16b\n"
       ".inst 0x4e80a45b  // smmla v27.4s, v2.16b, v0.16b\n"
       ".inst 0x4e86a4ef  // smmla v15.4s, v7.16b, v6.16b\n"
@@ -3236,32 +3235,32 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 200b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "add x20, x21, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
-      "add x21, x22, x20, LSL #2\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
-      "uzp2 v16.2d, v16.2d, v20.2d\n"
       "prfm pstl1keep, [x23, #0x0]\n"
-      "add x20, x21, x20, LSL #2\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
+      "uzp2 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
+      "prfm pstl1keep, [x20, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "prfm pstl1keep, [x21, #0x0]\n"
-      "prfm pstl1keep, [x20, #0x0]\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
       "uzp1 v23.2d, v24.2d, v28.2d\n"
@@ -3440,8 +3439,8 @@ void a64_hybrid_s8s32_mmla_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "224:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16.hpp
index b81e2e8593..926408855d 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16.hpp
@@ -71,7 +71,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 4, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp
index 90b196735a..b2cec742c4 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp
@@ -45,18 +45,18 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -84,133 +84,133 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "cmp %x[M], #0x2\n"
       "bgt 61f\n"
       "beq 31f\n"
-      "mov x16, %x[col_bias]\n"
+      "mov x15, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v15.16b, #0x1\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x13, %x[output_ptr]\n"
+      "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
       "movi v18.4s, #0x0\n"
       "movi v19.4s, #0x0\n"
       "3:"  // Height 1: setup done
-      "mov x12, #0x0\n"
+      "mov x11, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w10, [x20, x11, LSL #0x2]\n"
       "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w11, [x20, x12, LSL #0x2]\n"
       "tbz %x[flags], #3, 5f\n"
-      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
-      "ldr x10, [x20, #0x0]\n"
-      "cbnz x12, 6f\n"
+      "ldr x9, [x20, #0x0]\n"
+      "cbnz x11, 6f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x10, x10, x20\n"
+      "add x9, x9, x20\n"
       "b 6f\n"
       "5:"  // Height 1: setup direct input
-      "mov x10, %x[input_ptr]\n"
+      "mov x9, %x[input_ptr]\n"
       "6:"  // Height 1: input setup done
-      "cmp x11, #0x10\n"
+      "cmp x10, #0x10\n"
       "blt 11f\n"
-      "ldr q0, [x10, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "blt 9f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x6f80e090  // udot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr d21, [x14, #0x70]\n"
-      "ldr x20, [x14, #0x78]\n"
+      "ldr d21, [x12, #0x70]\n"
+      "ldr x20, [x12, #0x78]\n"
       ".inst 0x6f80e0b1  // udot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr d20, [x14, #0x80]\n"
+      "ldr d20, [x12, #0x80]\n"
       ".inst 0x6f80e0d2  // udot v18.4s, v6.16b, v0.4b[0]\n"
-      "ldr d26, [x14, #0x90]\n"
+      "ldr d26, [x12, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr d25, [x14, #0xa0]\n"
+      "ldr d25, [x12, #0xa0]\n"
       "mov v21.d[1], x20\n"
-      "ldr x20, [x14, #0x88]\n"
+      "ldr x20, [x12, #0x88]\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
-      "ldr d24, [x14, #0xb0]\n"
+      "ldr d24, [x12, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
-      "ldr d23, [x14, #0xc0]\n"
+      "ldr d23, [x12, #0xc0]\n"
       ".inst 0x6fa0e152  // udot v18.4s, v10.16b, v0.4b[1]\n"
-      "ldr d22, [x14, #0xd0]\n"
+      "ldr d22, [x12, #0xd0]\n"
       ".inst 0x6fa0e2b3  // udot v19.4s, v21.16b, v0.4b[1]\n"
-      "ldr d21, [x14, #0xe0]\n"
+      "ldr d21, [x12, #0xe0]\n"
       "mov v20.d[1], x20\n"
-      "ldr x22, [x14, #0x98]\n"
-      "add x10, x10, #0x10\n"
-      "ldr x21, [x14, #0xa8]\n"
+      "ldr x20, [x12, #0x98]\n"
+      "mov v26.d[1], x20\n"
+      "ldr x20, [x12, #0xa8]\n"
+      "mov v25.d[1], x20\n"
+      "ldr x20, [x12, #0xb8]\n"
+      "mov v24.d[1], x20\n"
+      "ldr x23, [x12, #0xc8]\n"
       ".inst 0x6f80ea90  // udot v16.4s, v20.16b, v0.4b[2]\n"
-      "ldr d20, [x14, #0xf0]\n"
-      "ldr x20, [x14, #0xb8]\n"
-      "mov v26.d[1], x22\n"
-      "mov v25.d[1], x21\n"
-      "ldr x23, [x14, #0xc8]\n"
-      "ldr x22, [x14, #0xd8]\n"
+      "ldr d20, [x12, #0xf0]\n"
       ".inst 0x6f80eb51  // udot v17.4s, v26.16b, v0.4b[2]\n"
-      "mov v24.d[1], x20\n"
-      "ldr x21, [x14, #0xe8]\n"
-      "ldr x20, [x14, #0xf8]\n"
+      "ldr x22, [x12, #0xd8]\n"
       ".inst 0x6f80eb32  // udot v18.4s, v25.16b, v0.4b[2]\n"
+      "ldr x21, [x12, #0xe8]\n"
       ".inst 0x6f80eb13  // udot v19.4s, v24.16b, v0.4b[2]\n"
+      "ldr x20, [x12, #0xf8]\n"
       "mov v23.d[1], x23\n"
       "mov v22.d[1], x22\n"
-      "add x14, x14, #0x100\n"
+      "add x9, x9, #0x10\n"
       "mov v21.d[1], x21\n"
-      ".inst 0x6fa0eaf0  // udot v16.4s, v23.16b, v0.4b[3]\n"
+      "add x12, x12, #0x100\n"
       "mov v20.d[1], x20\n"
+      ".inst 0x6fa0eaf0  // udot v16.4s, v23.16b, v0.4b[3]\n"
       ".inst 0x6fa0ead1  // udot v17.4s, v22.16b, v0.4b[3]\n"
       ".inst 0x6fa0eab2  // udot v18.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x6fa0ea93  // udot v19.4s, v20.16b, v0.4b[3]\n"
       "tbnz %x[flags], #31, 8f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "8:"  // Height 1: Multiply loop: unique 1: skip row sum
-      "ldr q0, [x10, #0x0]\n"
-      "sub x11, x11, #0x10\n"
-      "ldr q4, [x14, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "sub x10, x10, #0x10\n"
+      "ldr q4, [x12, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "bge 7b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x6f80e090  // udot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr q21, [x14, #0x70]\n"
+      "ldr q21, [x12, #0x70]\n"
       ".inst 0x6f80e0b1  // udot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr q20, [x14, #0x80]\n"
+      "ldr q20, [x12, #0x80]\n"
       ".inst 0x6f80e0d2  // udot v18.4s, v6.16b, v0.4b[0]\n"
-      "ldr q26, [x14, #0x90]\n"
+      "ldr q26, [x12, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr q25, [x14, #0xa0]\n"
+      "ldr q25, [x12, #0xa0]\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
-      "ldr q24, [x14, #0xb0]\n"
+      "ldr q24, [x12, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
-      "ldr q23, [x14, #0xc0]\n"
+      "ldr q23, [x12, #0xc0]\n"
       ".inst 0x6fa0e152  // udot v18.4s, v10.16b, v0.4b[1]\n"
-      "ldr q22, [x14, #0xd0]\n"
+      "ldr q22, [x12, #0xd0]\n"
       ".inst 0x6fa0e2b3  // udot v19.4s, v21.16b, v0.4b[1]\n"
-      "ldr q21, [x14, #0xe0]\n"
+      "ldr q21, [x12, #0xe0]\n"
       ".inst 0x6f80ea90  // udot v16.4s, v20.16b, v0.4b[2]\n"
-      "ldr q20, [x14, #0xf0]\n"
+      "ldr q20, [x12, #0xf0]\n"
       ".inst 0x6f80eb51  // udot v17.4s, v26.16b, v0.4b[2]\n"
-      "sub x11, x11, #0x10\n"
+      "sub x10, x10, #0x10\n"
       ".inst 0x6f80eb32  // udot v18.4s, v25.16b, v0.4b[2]\n"
-      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x6f80eb13  // udot v19.4s, v24.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x6fa0eaf0  // udot v16.4s, v23.16b, v0.4b[3]\n"
       ".inst 0x6fa0ead1  // udot v17.4s, v22.16b, v0.4b[3]\n"
       ".inst 0x6fa0eab2  // udot v18.4s, v21.16b, v0.4b[3]\n"
@@ -218,54 +218,54 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "tbnz %x[flags], #31, 10f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "10:"  // Height 1: Multiply loop: unique 2: skip row sum
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       "11:"  // Height 1: Multiply loop: Main loop skip
-      "cbz x11, 18f\n"
-      "cmp x11, #0x4\n"
+      "cbz x10, 18f\n"
+      "cmp x10, #0x4\n"
       "blt 14f\n"
       "12:"  // Height 1: Multiply loop: Odd block loop
-      "ldr s0, [x10], #0x4\n"
+      "ldr s0, [x9], #0x4\n"
       "tbnz %x[flags], #31, 13f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "13:"  // Height 1: Multiply loop: unique 3: skip row sum
-      "ldr q23, [x14, #0x0]\n"
-      "sub x11, x11, #0x4\n"
-      "ldr q22, [x14, #0x10]\n"
-      "cmp x11, #0x4\n"
-      "ldr q21, [x14, #0x20]\n"
-      "ldr q20, [x14, #0x30]\n"
-      ".inst 0x6f80e2f0  // udot v16.4s, v23.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x0]\n"
+      "sub x10, x10, #0x4\n"
+      "ldr q22, [x12, #0x10]\n"
+      "cmp x10, #0x4\n"
+      "ldr q21, [x12, #0x20]\n"
+      ".inst 0x6f80e290  // udot v16.4s, v20.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x30]\n"
       ".inst 0x6f80e2d1  // udot v17.4s, v22.16b, v0.4b[0]\n"
-      "add x14, x14, #0x40\n"
       ".inst 0x6f80e2b2  // udot v18.4s, v21.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6f80e293  // udot v19.4s, v20.16b, v0.4b[0]\n"
       "bge 12b\n"
       "14:"  // Height 1: Multiply loop: Skip odd blocks
-      "cbz x11, 18f\n"
-      "tbz x11, #1, 15f\n"
-      "ldr h0, [x10], #0x2\n"
-      "tbz x11, #0, 16f\n"
-      "ld1 { v0.b }[2], [x10]\n"
+      "cbz x10, 18f\n"
+      "tbz x10, #1, 15f\n"
+      "ldr h0, [x9], #0x2\n"
+      "tbz x10, #0, 16f\n"
+      "ld1 { v0.b }[2], [x9]\n"
       "b 16f\n"
       "15:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x10, #0x0]\n"
+      "ldr b0, [x9, #0x0]\n"
       "16:"  // Height 1: Multiply loop: Ragged operand read: Done
       "tbnz %x[flags], #31, 17f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "17:"  // Height 1: Multiply loop: unique 4: skip row sum
-      "ldr q23, [x14, #0x0]\n"
-      "ldr q22, [x14, #0x10]\n"
-      "ldr q21, [x14, #0x20]\n"
-      "ldr q20, [x14, #0x30]\n"
-      ".inst 0x6f80e2f0  // udot v16.4s, v23.16b, v0.4b[0]\n"
-      ".inst 0x6f80e2d1  // udot v17.4s, v22.16b, v0.4b[0]\n"
-      "add x14, x14, #0x40\n"
-      ".inst 0x6f80e2b2  // udot v18.4s, v21.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x0]\n"
+      ".inst 0x6f80e290  // udot v16.4s, v20.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x10]\n"
+      ".inst 0x6f80e291  // udot v17.4s, v20.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x20]\n"
+      ".inst 0x6f80e292  // udot v18.4s, v20.16b, v0.4b[0]\n"
+      "ldr q20, [x12, #0x30]\n"
       ".inst 0x6f80e293  // udot v19.4s, v20.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       "18:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x12, x12, #0x1\n"
-      "cmp x12, x20\n"
+      "add x11, x11, #0x1\n"
+      "cmp x11, x20\n"
       "bne 4b\n"
       "prfm pstl1keep, [x13, #0x0]\n"
       "tbnz %x[flags], #31, 19f\n"
@@ -276,28 +276,28 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "addp v11.4s, v11.4s, v11.4s\n"
       "mul v11.4s, v11.4s, v20.4s\n"
       "19:"  // Height 1: skip row sum fixup
-      "ldr q24, [x16, #0x0]\n"
+      "ldr q23, [x15, #0x0]\n"
       "add v16.4s, v16.4s, v11.4s\n"
-      "ldr q23, [x16, #0x10]\n"
+      "ldr q22, [x15, #0x10]\n"
       "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q22, [x16, #0x20]\n"
+      "ldr q21, [x15, #0x20]\n"
       "add v18.4s, v18.4s, v11.4s\n"
-      "ldr q21, [x16, #0x30]\n"
+      "ldr q20, [x15, #0x30]\n"
       "add v19.4s, v19.4s, v11.4s\n"
+      "add v16.4s, v16.4s, v23.4s\n"
+      "add v17.4s, v17.4s, v22.4s\n"
+      "add v18.4s, v18.4s, v21.4s\n"
+      "add v19.4s, v19.4s, v20.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v20.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v24.4s\n"
-      "add v17.4s, v17.4s, v23.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v22.4s\n"
-      "add v19.4s, v19.4s, v21.4s\n"
-      "add x16, x16, #0x40\n"
       "sqrdmulh v16.4s, v16.4s, v20.4s\n"
       "sqrdmulh v17.4s, v17.4s, v20.4s\n"
       "sqrdmulh v18.4s, v18.4s, v20.4s\n"
       "sqrdmulh v19.4s, v19.4s, v20.4s\n"
+      "add x15, x15, #0x40\n"
       "tbz %x[flags], #5, 20f\n"
       "and v23.16b, v16.16b, v0.16b\n"
       "and v22.16b, v17.16b, v0.16b\n"
@@ -317,67 +317,67 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v22.4s }, [x20]\n"
+      "ld1r { v20.4s }, [x20]\n"
+      "add v16.4s, v16.4s, v20.4s\n"
+      "add v17.4s, v17.4s, v20.4s\n"
+      "add v18.4s, v18.4s, v20.4s\n"
+      "add v19.4s, v19.4s, v20.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v20.4s }, [x20]\n"
+      "smin v16.4s, v16.4s, v20.4s\n"
+      "smin v17.4s, v17.4s, v20.4s\n"
+      "smin v18.4s, v18.4s, v20.4s\n"
+      "smin v19.4s, v19.4s, v20.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v21.4s }, [x21]\n"
-      "cmp x15, #0x10\n"
       "ld1r { v20.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v22.4s\n"
-      "add v17.4s, v17.4s, v22.4s\n"
-      "add v18.4s, v18.4s, v22.4s\n"
-      "add v19.4s, v19.4s, v22.4s\n"
-      "smin v16.4s, v16.4s, v21.4s\n"
-      "smin v17.4s, v17.4s, v21.4s\n"
-      "smin v18.4s, v18.4s, v21.4s\n"
-      "smin v19.4s, v19.4s, v21.4s\n"
       "smax v16.4s, v16.4s, v20.4s\n"
       "smax v17.4s, v17.4s, v20.4s\n"
       "smax v18.4s, v18.4s, v20.4s\n"
       "smax v19.4s, v19.4s, v20.4s\n"
       "uzp1 v16.8h, v16.8h, v17.8h\n"
       "uzp1 v17.8h, v18.8h, v19.8h\n"
+      "cmp x14, #0x10\n"
       "uzp1 v16.16b, v16.16b, v17.16b\n"
       "bge 29f\n"
-      "tbz x15, #3, 24f\n"
+      "tbz x14, #3, 24f\n"
       "str d16, [x13], #0x8\n"
-      "tbz x15, #2, 22f\n"
+      "tbz x14, #2, 22f\n"
       "st1 { v16.s }[2], [x13], #0x4\n"
-      "tbz x15, #1, 21f\n"
+      "tbz x14, #1, 21f\n"
       "st1 { v16.h }[6], [x13], #0x2\n"
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[14], [x13]\n"
       "b 28f\n"
       "21:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[12], [x13]\n"
       "b 28f\n"
       "22:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x15, #1, 23f\n"
+      "tbz x14, #1, 23f\n"
       "st1 { v16.h }[4], [x13], #0x2\n"
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[10], [x13]\n"
       "b 28f\n"
       "23:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[8], [x13]\n"
       "b 28f\n"
       "24:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x15, #2, 26f\n"
+      "tbz x14, #2, 26f\n"
       "str s16, [x13], #0x4\n"
-      "tbz x15, #1, 25f\n"
+      "tbz x14, #1, 25f\n"
       "st1 { v16.h }[2], [x13], #0x2\n"
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[6], [x13]\n"
       "b 28f\n"
       "25:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[4], [x13]\n"
       "b 28f\n"
       "26:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x15, #1, 27f\n"
+      "tbz x14, #1, 27f\n"
       "str h16, [x13], #0x2\n"
-      "tbz x15, #0, 28f\n"
+      "tbz x14, #0, 28f\n"
       "st1 { v16.b }[2], [x13]\n"
       "b 28f\n"
       "27:"  // Height 1: Partial direct writeback: partial_1_0
@@ -388,18 +388,18 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "str q16, [x13, #0x0]\n"
       "add x13, x13, #0x10\n"
       "30:"  // Height 1: Writeback done
-      "subs x15, x15, #0x10\n"
+      "subs x14, x14, #0x10\n"
       "bgt 2b\n"
       "b 122f\n"
       "31:"  // Height 2
-      "mov x16, %x[col_bias]\n"
+      "mov x15, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "movi v15.16b, #0x1\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "32:"  // Height 2: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -410,80 +410,80 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "movi v22.4s, #0x0\n"
       "movi v23.4s, #0x0\n"
       "33:"  // Height 2: setup done
-      "mov x12, #0x0\n"
+      "mov x11, #0x0\n"
       "34:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w10, [x20, x11, LSL #0x2]\n"
       "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w11, [x20, x12, LSL #0x2]\n"
       "tbz %x[flags], #3, 35f\n"
-      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
-      "ldr x10, [x20, #0x0]\n"
-      "ldr x9, [x20, #0x8]\n"
-      "cbnz x12, 36f\n"
+      "ldr x9, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "cbnz x11, 36f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x10, x10, x20\n"
       "add x9, x9, x20\n"
+      "add x28, x28, x20\n"
       "b 36f\n"
       "35:"  // Height 2: setup direct input
-      "mov x10, %x[input_ptr]\n"
-      "add x9, x10, x21\n"
+      "mov x9, %x[input_ptr]\n"
+      "add x28, x9, x21\n"
       "36:"  // Height 2: input setup done
-      "cmp x11, #0x10\n"
+      "cmp x10, #0x10\n"
       "blt 41f\n"
-      "ldr q0, [x10, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q1, [x9, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "blt 39f\n"
       "37:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x6f80e090  // udot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr x20, [x14, #0x78]\n"
+      "ldr x20, [x12, #0x78]\n"
       ".inst 0x6f81e094  // udot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr d25, [x14, #0x70]\n"
+      "ldr d25, [x12, #0x70]\n"
       ".inst 0x6f80e0b1  // udot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr x23, [x14, #0x88]\n"
+      "mov v25.d[1], x20\n"
       ".inst 0x6f81e0b5  // udot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr d24, [x14, #0x80]\n"
+      "ldr d24, [x12, #0x80]\n"
       ".inst 0x6f80e0d2  // udot v18.4s, v6.16b, v0.4b[0]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x23, [x12, #0x88]\n"
       ".inst 0x6f81e0d6  // udot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr d30, [x14, #0x90]\n"
+      "ldr d30, [x12, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr x22, [x14, #0x98]\n"
+      "ldr x22, [x12, #0x98]\n"
       ".inst 0x6f81e0f7  // udot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr d29, [x14, #0xa0]\n"
-      "ldr x21, [x14, #0xa8]\n"
+      "ldr d29, [x12, #0xa0]\n"
+      "ldr x21, [x12, #0xa8]\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
       ".inst 0x6fa1e114  // udot v20.4s, v8.16b, v1.4b[1]\n"
-      "ldr d28, [x14, #0xb0]\n"
-      "ldr x20, [x14, #0xb8]\n"
+      "ldr d28, [x12, #0xb0]\n"
+      "ldr x20, [x12, #0xb8]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
       ".inst 0x6fa1e135  // udot v21.4s, v9.16b, v1.4b[1]\n"
-      "ldr d27, [x14, #0xc0]\n"
+      "ldr d27, [x12, #0xc0]\n"
       ".inst 0x6fa0e152  // udot v18.4s, v10.16b, v0.4b[1]\n"
       "mov v24.d[1], x23\n"
       ".inst 0x6fa1e156  // udot v22.4s, v10.16b, v1.4b[1]\n"
-      "ldr d26, [x14, #0xd0]\n"
+      "ldr d26, [x12, #0xd0]\n"
       ".inst 0x6fa0e333  // udot v19.4s, v25.16b, v0.4b[1]\n"
       "mov v30.d[1], x22\n"
       ".inst 0x6fa1e337  // udot v23.4s, v25.16b, v1.4b[1]\n"
-      "ldr d25, [x14, #0xe0]\n"
+      "ldr d25, [x12, #0xe0]\n"
       "mov v29.d[1], x21\n"
-      "ldr x23, [x14, #0xc8]\n"
+      "ldr x23, [x12, #0xc8]\n"
       "mov v28.d[1], x20\n"
-      "ldr x22, [x14, #0xd8]\n"
-      "ldr x21, [x14, #0xe8]\n"
+      "ldr x22, [x12, #0xd8]\n"
+      "ldr x21, [x12, #0xe8]\n"
       ".inst 0x6f80eb10  // udot v16.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x6f81eb14  // udot v20.4s, v24.16b, v1.4b[2]\n"
-      "ldr d24, [x14, #0xf0]\n"
-      "ldr x20, [x14, #0xf8]\n"
+      "ldr d24, [x12, #0xf0]\n"
+      "ldr x20, [x12, #0xf8]\n"
       ".inst 0x6f80ebd1  // udot v17.4s, v30.16b, v0.4b[2]\n"
       ".inst 0x6f81ebd5  // udot v21.4s, v30.16b, v1.4b[2]\n"
       "mov v27.d[1], x23\n"
@@ -494,9 +494,9 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6f80eb93  // udot v19.4s, v28.16b, v0.4b[2]\n"
       "mov v24.d[1], x20\n"
       ".inst 0x6f81eb97  // udot v23.4s, v28.16b, v1.4b[2]\n"
-      "add x10, x10, #0x10\n"
       "add x9, x9, #0x10\n"
-      "add x14, x14, #0x100\n"
+      "add x28, x28, #0x10\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x6fa0eb70  // udot v16.4s, v27.16b, v0.4b[3]\n"
       ".inst 0x6fa1eb74  // udot v20.4s, v27.16b, v1.4b[3]\n"
       ".inst 0x6fa0eb51  // udot v17.4s, v26.16b, v0.4b[3]\n"
@@ -509,53 +509,53 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       "38:"  // Height 2: Multiply loop: unique 5: skip row sum
-      "ldr q0, [x10, #0x0]\n"
-      "sub x11, x11, #0x10\n"
-      "ldr q1, [x9, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "sub x10, x10, #0x10\n"
+      "ldr q1, [x28, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "bge 37b\n"
       "39:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x6f80e090  // udot v16.4s, v4.16b, v0.4b[0]\n"
-      "sub x11, x11, #0x10\n"
+      "sub x10, x10, #0x10\n"
       ".inst 0x6f81e094  // udot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr q25, [x14, #0x70]\n"
+      "ldr q25, [x12, #0x70]\n"
       ".inst 0x6f80e0b1  // udot v17.4s, v5.16b, v0.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x6f81e0b5  // udot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr q24, [x14, #0x80]\n"
+      "ldr q24, [x12, #0x80]\n"
       ".inst 0x6f80e0d2  // udot v18.4s, v6.16b, v0.4b[0]\n"
-      "add x9, x9, #0x10\n"
+      "add x28, x28, #0x10\n"
       ".inst 0x6f81e0d6  // udot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr q30, [x14, #0x90]\n"
+      "ldr q30, [x12, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x6f81e0f7  // udot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr q29, [x14, #0xa0]\n"
+      "ldr q29, [x12, #0xa0]\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
       ".inst 0x6fa1e114  // udot v20.4s, v8.16b, v1.4b[1]\n"
-      "ldr q28, [x14, #0xb0]\n"
+      "ldr q28, [x12, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
       ".inst 0x6fa1e135  // udot v21.4s, v9.16b, v1.4b[1]\n"
-      "ldr q27, [x14, #0xc0]\n"
+      "ldr q27, [x12, #0xc0]\n"
       ".inst 0x6fa0e152  // udot v18.4s, v10.16b, v0.4b[1]\n"
       ".inst 0x6fa1e156  // udot v22.4s, v10.16b, v1.4b[1]\n"
-      "ldr q26, [x14, #0xd0]\n"
+      "ldr q26, [x12, #0xd0]\n"
       ".inst 0x6fa0e333  // udot v19.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x6fa1e337  // udot v23.4s, v25.16b, v1.4b[1]\n"
-      "ldr q25, [x14, #0xe0]\n"
+      "ldr q25, [x12, #0xe0]\n"
       ".inst 0x6f80eb10  // udot v16.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x6f81eb14  // udot v20.4s, v24.16b, v1.4b[2]\n"
-      "ldr q24, [x14, #0xf0]\n"
+      "ldr q24, [x12, #0xf0]\n"
       ".inst 0x6f80ebd1  // udot v17.4s, v30.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x6f81ebd5  // udot v21.4s, v30.16b, v1.4b[2]\n"
       ".inst 0x6f80ebb2  // udot v18.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x6f81ebb6  // udot v22.4s, v29.16b, v1.4b[2]\n"
@@ -573,29 +573,29 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       "40:"  // Height 2: Multiply loop: unique 6: skip row sum
-      "prfm pldl1keep, [x10, #0x80]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       "41:"  // Height 2: Multiply loop: Main loop skip
-      "cbz x11, 48f\n"
-      "cmp x11, #0x4\n"
+      "cbz x10, 48f\n"
+      "cmp x10, #0x4\n"
       "blt 44f\n"
       "42:"  // Height 2: Multiply loop: Odd block loop
-      "ldr s0, [x10], #0x4\n"
-      "ldr s1, [x9], #0x4\n"
+      "ldr s0, [x9], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
       "tbnz %x[flags], #31, 43f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       "43:"  // Height 2: Multiply loop: unique 7: skip row sum
-      "ldr q27, [x14, #0x0]\n"
-      "sub x11, x11, #0x4\n"
-      "ldr q26, [x14, #0x10]\n"
-      "cmp x11, #0x4\n"
-      "ldr q25, [x14, #0x20]\n"
-      "ldr q24, [x14, #0x30]\n"
+      "ldr q27, [x12, #0x0]\n"
+      "sub x10, x10, #0x4\n"
+      "ldr q26, [x12, #0x10]\n"
+      "cmp x10, #0x4\n"
+      "ldr q25, [x12, #0x20]\n"
       ".inst 0x6f80e370  // udot v16.4s, v27.16b, v0.4b[0]\n"
+      "ldr q24, [x12, #0x30]\n"
       ".inst 0x6f81e374  // udot v20.4s, v27.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
       ".inst 0x6f80e351  // udot v17.4s, v26.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6f81e355  // udot v21.4s, v26.16b, v1.4b[0]\n"
       ".inst 0x6f80e332  // udot v18.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x6f81e336  // udot v22.4s, v25.16b, v1.4b[0]\n"
@@ -603,44 +603,44 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6f81e317  // udot v23.4s, v24.16b, v1.4b[0]\n"
       "bge 42b\n"
       "44:"  // Height 2: Multiply loop: Skip odd blocks
-      "cbz x11, 48f\n"
-      "tbz x11, #1, 45f\n"
-      "ldr h0, [x10], #0x2\n"
-      "ldr h1, [x9], #0x2\n"
-      "tbz x11, #0, 46f\n"
-      "ld1 { v0.b }[2], [x10]\n"
-      "ld1 { v1.b }[2], [x9]\n"
+      "cbz x10, 48f\n"
+      "tbz x10, #1, 45f\n"
+      "ldr h0, [x9], #0x2\n"
+      "ldr h1, [x28], #0x2\n"
+      "tbz x10, #0, 46f\n"
+      "ld1 { v0.b }[2], [x9]\n"
+      "ld1 { v1.b }[2], [x28]\n"
       "b 46f\n"
       "45:"  // Height 2: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x10, #0x0]\n"
-      "ldr b1, [x9, #0x0]\n"
+      "ldr b0, [x9, #0x0]\n"
+      "ldr b1, [x28, #0x0]\n"
       "46:"  // Height 2: Multiply loop: Ragged operand read: Done
       "tbnz %x[flags], #31, 47f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       "47:"  // Height 2: Multiply loop: unique 8: skip row sum
-      "ldr q27, [x14, #0x0]\n"
-      "ldr q26, [x14, #0x10]\n"
-      "ldr q25, [x14, #0x20]\n"
-      "ldr q24, [x14, #0x30]\n"
-      ".inst 0x6f80e370  // udot v16.4s, v27.16b, v0.4b[0]\n"
-      ".inst 0x6f81e374  // udot v20.4s, v27.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
+      "ldr q24, [x12, #0x0]\n"
+      ".inst 0x6f80e310  // udot v16.4s, v24.16b, v0.4b[0]\n"
+      "ldr q26, [x12, #0x10]\n"
+      ".inst 0x6f81e314  // udot v20.4s, v24.16b, v1.4b[0]\n"
+      "ldr q25, [x12, #0x20]\n"
       ".inst 0x6f80e351  // udot v17.4s, v26.16b, v0.4b[0]\n"
+      "ldr q24, [x12, #0x30]\n"
       ".inst 0x6f81e355  // udot v21.4s, v26.16b, v1.4b[0]\n"
       ".inst 0x6f80e332  // udot v18.4s, v25.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6f81e336  // udot v22.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x6f80e313  // udot v19.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x6f81e317  // udot v23.4s, v24.16b, v1.4b[0]\n"
       "48:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x12, x12, #0x1\n"
-      "cmp x12, x20\n"
+      "add x11, x11, #0x1\n"
+      "cmp x11, x20\n"
       "bne 34b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20\n"
       "prfm pstl1keep, [x13, #0x0]\n"
-      "add x24, x13, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbnz %x[flags], #31, 49f\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
@@ -652,28 +652,28 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "mul v11.4s, v11.4s, v24.4s\n"
       "mul v12.4s, v12.4s, v24.4s\n"
       "49:"  // Height 2: skip row sum fixup
-      "ldr q28, [x16, #0x0]\n"
+      "ldr q27, [x15, #0x0]\n"
       "add v16.4s, v16.4s, v11.4s\n"
-      "ldr q27, [x16, #0x10]\n"
+      "ldr q26, [x15, #0x10]\n"
       "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q26, [x16, #0x20]\n"
+      "ldr q25, [x15, #0x20]\n"
       "add v18.4s, v18.4s, v11.4s\n"
-      "ldr q25, [x16, #0x30]\n"
+      "ldr q24, [x15, #0x30]\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
       "add v22.4s, v22.4s, v12.4s\n"
       "add v23.4s, v23.4s, v12.4s\n"
-      "add v16.4s, v16.4s, v28.4s\n"
-      "add v17.4s, v17.4s, v27.4s\n"
-      "add v18.4s, v18.4s, v26.4s\n"
+      "add v16.4s, v16.4s, v27.4s\n"
+      "add v17.4s, v17.4s, v26.4s\n"
+      "add v18.4s, v18.4s, v25.4s\n"
+      "add v19.4s, v19.4s, v24.4s\n"
+      "add v20.4s, v20.4s, v27.4s\n"
+      "add v21.4s, v21.4s, v26.4s\n"
+      "add v22.4s, v22.4s, v25.4s\n"
+      "add v23.4s, v23.4s, v24.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
       "ld1r { v24.4s }, [x20]\n"
-      "add v19.4s, v19.4s, v25.4s\n"
-      "add v20.4s, v20.4s, v28.4s\n"
-      "add v21.4s, v21.4s, v27.4s\n"
-      "add v22.4s, v22.4s, v26.4s\n"
-      "add v23.4s, v23.4s, v25.4s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "ld1r { v0.4s }, [x20]\n"
@@ -685,31 +685,31 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "sqrdmulh v21.4s, v21.4s, v24.4s\n"
       "sqrdmulh v22.4s, v22.4s, v24.4s\n"
       "sqrdmulh v23.4s, v23.4s, v24.4s\n"
-      "add x16, x16, #0x40\n"
+      "add x15, x15, #0x40\n"
       "tbz %x[flags], #5, 50f\n"
       "and v24.16b, v16.16b, v0.16b\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
+      "sqadd v16.4s, v16.4s, v24.4s\n"
       "and v30.16b, v17.16b, v0.16b\n"
       "and v29.16b, v18.16b, v0.16b\n"
       "and v28.16b, v19.16b, v0.16b\n"
       "and v27.16b, v20.16b, v0.16b\n"
       "and v26.16b, v21.16b, v0.16b\n"
       "and v25.16b, v22.16b, v0.16b\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
+      "and v24.16b, v23.16b, v0.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
       "sshr v26.4s, v26.4s, #0x1f\n"
       "sshr v25.4s, v25.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v24.4s\n"
-      "and v24.16b, v23.16b, v0.16b\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v17.4s, v17.4s, v30.4s\n"
       "sqadd v18.4s, v18.4s, v29.4s\n"
       "sqadd v19.4s, v19.4s, v28.4s\n"
       "sqadd v20.4s, v20.4s, v27.4s\n"
       "sqadd v21.4s, v21.4s, v26.4s\n"
       "sqadd v22.4s, v22.4s, v25.4s\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v23.4s, v23.4s, v24.4s\n"
       "50:"  // Height 2: no shift correction
       "srshl v16.4s, v16.4s, v0.4s\n"
@@ -721,28 +721,27 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "srshl v22.4s, v22.4s, v0.4s\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x20]\n"
+      "ld1r { v24.4s }, [x20]\n"
+      "add v16.4s, v16.4s, v24.4s\n"
+      "add v17.4s, v17.4s, v24.4s\n"
+      "add v18.4s, v18.4s, v24.4s\n"
+      "add v19.4s, v19.4s, v24.4s\n"
+      "add v20.4s, v20.4s, v24.4s\n"
+      "add v21.4s, v21.4s, v24.4s\n"
+      "add v22.4s, v22.4s, v24.4s\n"
+      "add v23.4s, v23.4s, v24.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v24.4s }, [x20]\n"
+      "smin v16.4s, v16.4s, v24.4s\n"
+      "smin v17.4s, v17.4s, v24.4s\n"
+      "smin v18.4s, v18.4s, v24.4s\n"
+      "smin v19.4s, v19.4s, v24.4s\n"
+      "smin v20.4s, v20.4s, v24.4s\n"
+      "smin v21.4s, v21.4s, v24.4s\n"
+      "smin v22.4s, v22.4s, v24.4s\n"
+      "smin v23.4s, v23.4s, v24.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v25.4s }, [x21]\n"
-      "cmp x15, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v26.4s\n"
-      "add v17.4s, v17.4s, v26.4s\n"
-      "add v18.4s, v18.4s, v26.4s\n"
-      "add v19.4s, v19.4s, v26.4s\n"
-      "add v20.4s, v20.4s, v26.4s\n"
-      "add v21.4s, v21.4s, v26.4s\n"
-      "add v22.4s, v22.4s, v26.4s\n"
-      "add v23.4s, v23.4s, v26.4s\n"
-      "smin v16.4s, v16.4s, v25.4s\n"
-      "smin v17.4s, v17.4s, v25.4s\n"
-      "smin v18.4s, v18.4s, v25.4s\n"
-      "smin v19.4s, v19.4s, v25.4s\n"
-      "smin v20.4s, v20.4s, v25.4s\n"
-      "smin v21.4s, v21.4s, v25.4s\n"
-      "smin v22.4s, v22.4s, v25.4s\n"
-      "smin v23.4s, v23.4s, v25.4s\n"
       "smax v16.4s, v16.4s, v24.4s\n"
       "smax v17.4s, v17.4s, v24.4s\n"
       "smax v18.4s, v18.4s, v24.4s\n"
@@ -755,87 +754,88 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "uzp1 v18.8h, v18.8h, v19.8h\n"
       "uzp1 v20.8h, v20.8h, v21.8h\n"
       "uzp1 v17.8h, v22.8h, v23.8h\n"
+      "cmp x14, #0x10\n"
       "uzp1 v16.16b, v16.16b, v18.16b\n"
       "uzp1 v20.16b, v20.16b, v17.16b\n"
       "bge 59f\n"
-      "tbz x15, #3, 54f\n"
+      "tbz x14, #3, 54f\n"
       "str d16, [x13], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "tbz x15, #2, 52f\n"
+      "str d20, [x23], #0x8\n"
+      "tbz x14, #2, 52f\n"
       "st1 { v16.s }[2], [x13], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "tbz x15, #1, 51f\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "tbz x14, #1, 51f\n"
       "st1 { v16.h }[6], [x13], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "tbz x15, #0, 58f\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[14], [x13]\n"
-      "st1 { v20.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
       "b 58f\n"
       "51:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x15, #0, 58f\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[12], [x13]\n"
-      "st1 { v20.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
       "b 58f\n"
       "52:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x15, #1, 53f\n"
+      "tbz x14, #1, 53f\n"
       "st1 { v16.h }[4], [x13], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "tbz x15, #0, 58f\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[10], [x13]\n"
-      "st1 { v20.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
       "b 58f\n"
       "53:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x15, #0, 58f\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[8], [x13]\n"
-      "st1 { v20.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
       "b 58f\n"
       "54:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x15, #2, 56f\n"
+      "tbz x14, #2, 56f\n"
       "str s16, [x13], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "tbz x15, #1, 55f\n"
+      "str s20, [x23], #0x4\n"
+      "tbz x14, #1, 55f\n"
       "st1 { v16.h }[2], [x13], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "tbz x15, #0, 58f\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[6], [x13]\n"
-      "st1 { v20.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
       "b 58f\n"
       "55:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x15, #0, 58f\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[4], [x13]\n"
-      "st1 { v20.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
       "b 58f\n"
       "56:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x15, #1, 57f\n"
+      "tbz x14, #1, 57f\n"
       "str h16, [x13], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "tbz x15, #0, 58f\n"
+      "str h20, [x23], #0x2\n"
+      "tbz x14, #0, 58f\n"
       "st1 { v16.b }[2], [x13]\n"
-      "st1 { v20.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
       "b 58f\n"
       "57:"  // Height 2: Partial direct writeback: partial_1_0
       "str b16, [x13, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
       "58:"  // Height 2: Partial direct writeback: Done
       "b 60f\n"
       "59:"  // Height 2: Full writeback
       "str q16, [x13, #0x0]\n"
       "add x13, x13, #0x10\n"
-      "str q20, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
       "60:"  // Height 2: Writeback done
-      "subs x15, x15, #0x10\n"
+      "subs x14, x14, #0x10\n"
       "bgt 32b\n"
       "b 122f\n"
       "61:"  // Height 3
-      "mov x16, %x[col_bias]\n"
+      "mov x15, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "movi v13.4s, #0x0\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "movi v15.16b, #0x1\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "62:"  // Height 3: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -850,105 +850,105 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "movi v26.4s, #0x0\n"
       "movi v27.4s, #0x0\n"
       "63:"  // Height 3: setup done
-      "mov x12, #0x0\n"
+      "mov x11, #0x0\n"
       "64:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w10, [x20, x11, LSL #0x2]\n"
       "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w11, [x20, x12, LSL #0x2]\n"
       "tbz %x[flags], #3, 65f\n"
-      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
-      "ldr x10, [x20, #0x0]\n"
-      "ldr x9, [x20, #0x8]\n"
-      "ldr x28, [x20, #0x10]\n"
-      "cbnz x12, 66f\n"
+      "ldr x9, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "ldr x27, [x20, #0x10]\n"
+      "cbnz x11, 66f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x10, x10, x20\n"
       "add x9, x9, x20\n"
       "add x28, x28, x20\n"
+      "add x27, x27, x20\n"
       "b 66f\n"
       "65:"  // Height 3: setup direct input
-      "mov x10, %x[input_ptr]\n"
-      "add x9, x10, x21\n"
+      "mov x9, %x[input_ptr]\n"
       "add x28, x9, x21\n"
+      "add x27, x28, x21\n"
       "66:"  // Height 3: input setup done
-      "cmp x11, #0x10\n"
+      "cmp x10, #0x10\n"
       "blt 71f\n"
-      "ldr q0, [x10, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q1, [x9, #0x0]\n"
-      "ldr q2, [x28, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x27, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "blt 69f\n"
       "67:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x6f80e090  // udot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr x20, [x14, #0x78]\n"
+      "ldr x20, [x12, #0x78]\n"
       ".inst 0x6f81e094  // udot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr x23, [x14, #0x88]\n"
+      "ldr x23, [x12, #0x88]\n"
       ".inst 0x6f82e098  // udot v24.4s, v4.16b, v2.4b[0]\n"
-      "ldr d29, [x14, #0x70]\n"
+      "ldr d29, [x12, #0x70]\n"
       ".inst 0x6f80e0b1  // udot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr x22, [x14, #0x98]\n"
+      "mov v29.d[1], x20\n"
       ".inst 0x6f81e0b5  // udot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr x21, [x14, #0xa8]\n"
+      "ldr x22, [x12, #0x98]\n"
       ".inst 0x6f82e0b9  // udot v25.4s, v5.16b, v2.4b[0]\n"
-      "ldr d28, [x14, #0x80]\n"
+      "ldr d28, [x12, #0x80]\n"
       ".inst 0x6f80e0d2  // udot v18.4s, v6.16b, v0.4b[0]\n"
-      "mov v29.d[1], x20\n"
+      "ldr x21, [x12, #0xa8]\n"
       ".inst 0x6f81e0d6  // udot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x14, #0xb8]\n"
+      "ldr x20, [x12, #0xb8]\n"
       ".inst 0x6f82e0da  // udot v26.4s, v6.16b, v2.4b[0]\n"
-      "ldr d5, [x14, #0x90]\n"
+      "ldr d5, [x12, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
       "mov v28.d[1], x23\n"
       ".inst 0x6f81e0f7  // udot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr x23, [x14, #0xc8]\n"
+      "mov v5.d[1], x22\n"
       ".inst 0x6f82e0fb  // udot v27.4s, v7.16b, v2.4b[0]\n"
-      "ldr d4, [x14, #0xa0]\n"
+      "ldr d4, [x12, #0xa0]\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
-      "mov v5.d[1], x22\n"
+      "mov v4.d[1], x21\n"
       ".inst 0x6fa1e114  // udot v20.4s, v8.16b, v1.4b[1]\n"
-      "ldr x22, [x14, #0xd8]\n"
+      "ldr x23, [x12, #0xc8]\n"
       ".inst 0x6fa2e118  // udot v24.4s, v8.16b, v2.4b[1]\n"
-      "ldr d3, [x14, #0xb0]\n"
+      "ldr d3, [x12, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
-      "mov v4.d[1], x21\n"
+      "mov v3.d[1], x20\n"
       ".inst 0x6fa1e135  // udot v21.4s, v9.16b, v1.4b[1]\n"
-      "ldr x21, [x14, #0xe8]\n"
+      "ldr x22, [x12, #0xd8]\n"
       ".inst 0x6fa2e139  // udot v25.4s, v9.16b, v2.4b[1]\n"
-      "ldr d31, [x14, #0xc0]\n"
+      "ldr d31, [x12, #0xc0]\n"
       ".inst 0x6fa0e152  // udot v18.4s, v10.16b, v0.4b[1]\n"
-      "mov v3.d[1], x20\n"
+      "ldr x21, [x12, #0xe8]\n"
       ".inst 0x6fa1e156  // udot v22.4s, v10.16b, v1.4b[1]\n"
-      "ldr x20, [x14, #0xf8]\n"
+      "ldr x20, [x12, #0xf8]\n"
       ".inst 0x6fa2e15a  // udot v26.4s, v10.16b, v2.4b[1]\n"
-      "ldr d30, [x14, #0xd0]\n"
+      "ldr d30, [x12, #0xd0]\n"
       ".inst 0x6fa0e3b3  // udot v19.4s, v29.16b, v0.4b[1]\n"
       "mov v31.d[1], x23\n"
       ".inst 0x6fa1e3b7  // udot v23.4s, v29.16b, v1.4b[1]\n"
-      "add x10, x10, #0x10\n"
+      "mov v30.d[1], x22\n"
       ".inst 0x6fa2e3bb  // udot v27.4s, v29.16b, v2.4b[1]\n"
-      "ldr d29, [x14, #0xe0]\n"
+      "ldr d29, [x12, #0xe0]\n"
       ".inst 0x6f80eb90  // udot v16.4s, v28.16b, v0.4b[2]\n"
-      "mov v30.d[1], x22\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x6f81eb94  // udot v20.4s, v28.16b, v1.4b[2]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x6f82eb98  // udot v24.4s, v28.16b, v2.4b[2]\n"
-      "ldr d28, [x14, #0xf0]\n"
+      "ldr d28, [x12, #0xf0]\n"
       ".inst 0x6f80e8b1  // udot v17.4s, v5.16b, v0.4b[2]\n"
-      "mov v29.d[1], x21\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x6f81e8b5  // udot v21.4s, v5.16b, v1.4b[2]\n"
       "add x28, x28, #0x10\n"
       ".inst 0x6f82e8b9  // udot v25.4s, v5.16b, v2.4b[2]\n"
-      "mov v28.d[1], x20\n"
+      "add x27, x27, #0x10\n"
       ".inst 0x6f80e892  // udot v18.4s, v4.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x6f81e896  // udot v22.4s, v4.16b, v1.4b[2]\n"
       ".inst 0x6f82e89a  // udot v26.4s, v4.16b, v2.4b[2]\n"
       ".inst 0x6f80e873  // udot v19.4s, v3.16b, v0.4b[2]\n"
@@ -971,65 +971,65 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       "68:"  // Height 3: Multiply loop: unique 9: skip row sum
-      "ldr q0, [x10, #0x0]\n"
-      "sub x11, x11, #0x10\n"
-      "ldr q1, [x9, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q2, [x28, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "sub x10, x10, #0x10\n"
+      "ldr q1, [x28, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q2, [x27, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "bge 67b\n"
       "69:"  // Height 3: Multiply loop: Single iteration only
       ".inst 0x6f80e090  // udot v16.4s, v4.16b, v0.4b[0]\n"
-      "sub x11, x11, #0x10\n"
+      "sub x10, x10, #0x10\n"
       ".inst 0x6f81e094  // udot v20.4s, v4.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x6f82e098  // udot v24.4s, v4.16b, v2.4b[0]\n"
-      "ldr q29, [x14, #0x70]\n"
+      "ldr q29, [x12, #0x70]\n"
       ".inst 0x6f80e0b1  // udot v17.4s, v5.16b, v0.4b[0]\n"
-      "add x9, x9, #0x10\n"
-      ".inst 0x6f81e0b5  // udot v21.4s, v5.16b, v1.4b[0]\n"
       "add x28, x28, #0x10\n"
+      ".inst 0x6f81e0b5  // udot v21.4s, v5.16b, v1.4b[0]\n"
+      "add x27, x27, #0x10\n"
       ".inst 0x6f82e0b9  // udot v25.4s, v5.16b, v2.4b[0]\n"
-      "ldr q28, [x14, #0x80]\n"
+      "ldr q28, [x12, #0x80]\n"
       ".inst 0x6f80e0d2  // udot v18.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x6f81e0d6  // udot v22.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x6f82e0da  // udot v26.4s, v6.16b, v2.4b[0]\n"
-      "ldr q5, [x14, #0x90]\n"
+      "ldr q5, [x12, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x6f81e0f7  // udot v23.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x6f82e0fb  // udot v27.4s, v7.16b, v2.4b[0]\n"
-      "ldr q4, [x14, #0xa0]\n"
+      "ldr q4, [x12, #0xa0]\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
       ".inst 0x6fa1e114  // udot v20.4s, v8.16b, v1.4b[1]\n"
       ".inst 0x6fa2e118  // udot v24.4s, v8.16b, v2.4b[1]\n"
-      "ldr q3, [x14, #0xb0]\n"
+      "ldr q3, [x12, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
       ".inst 0x6fa1e135  // udot v21.4s, v9.16b, v1.4b[1]\n"
       ".inst 0x6fa2e139  // udot v25.4s, v9.16b, v2.4b[1]\n"
-      "ldr q31, [x14, #0xc0]\n"
+      "ldr q31, [x12, #0xc0]\n"
       ".inst 0x6fa0e152  // udot v18.4s, v10.16b, v0.4b[1]\n"
       ".inst 0x6fa1e156  // udot v22.4s, v10.16b, v1.4b[1]\n"
       ".inst 0x6fa2e15a  // udot v26.4s, v10.16b, v2.4b[1]\n"
-      "ldr q30, [x14, #0xd0]\n"
+      "ldr q30, [x12, #0xd0]\n"
       ".inst 0x6fa0e3b3  // udot v19.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x6fa1e3b7  // udot v23.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x6fa2e3bb  // udot v27.4s, v29.16b, v2.4b[1]\n"
-      "ldr q29, [x14, #0xe0]\n"
+      "ldr q29, [x12, #0xe0]\n"
       ".inst 0x6f80eb90  // udot v16.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x6f81eb94  // udot v20.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x6f82eb98  // udot v24.4s, v28.16b, v2.4b[2]\n"
-      "ldr q28, [x14, #0xf0]\n"
+      "ldr q28, [x12, #0xf0]\n"
       ".inst 0x6f80e8b1  // udot v17.4s, v5.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x6f81e8b5  // udot v21.4s, v5.16b, v1.4b[2]\n"
       ".inst 0x6f82e8b9  // udot v25.4s, v5.16b, v2.4b[2]\n"
       ".inst 0x6f80e892  // udot v18.4s, v4.16b, v0.4b[2]\n"
@@ -1055,32 +1055,32 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       "70:"  // Height 3: Multiply loop: unique 10: skip row sum
-      "prfm pldl1keep, [x10, #0x80]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
+      "prfm pldl1keep, [x27, #0x80]\n"
       "71:"  // Height 3: Multiply loop: Main loop skip
-      "cbz x11, 78f\n"
-      "cmp x11, #0x4\n"
+      "cbz x10, 78f\n"
+      "cmp x10, #0x4\n"
       "blt 74f\n"
       "72:"  // Height 3: Multiply loop: Odd block loop
-      "ldr s0, [x10], #0x4\n"
-      "ldr s1, [x9], #0x4\n"
-      "ldr s2, [x28], #0x4\n"
+      "ldr s0, [x9], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr s2, [x27], #0x4\n"
       "tbnz %x[flags], #31, 73f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       "73:"  // Height 3: Multiply loop: unique 11: skip row sum
-      "ldr q31, [x14, #0x0]\n"
-      "sub x11, x11, #0x4\n"
-      "ldr q30, [x14, #0x10]\n"
-      "cmp x11, #0x4\n"
-      "ldr q29, [x14, #0x20]\n"
-      "ldr q28, [x14, #0x30]\n"
+      "ldr q31, [x12, #0x0]\n"
+      "sub x10, x10, #0x4\n"
+      "ldr q30, [x12, #0x10]\n"
+      "cmp x10, #0x4\n"
+      "ldr q29, [x12, #0x20]\n"
       ".inst 0x6f80e3f0  // udot v16.4s, v31.16b, v0.4b[0]\n"
+      "ldr q28, [x12, #0x30]\n"
       ".inst 0x6f81e3f4  // udot v20.4s, v31.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
       ".inst 0x6f82e3f8  // udot v24.4s, v31.16b, v2.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6f80e3d1  // udot v17.4s, v30.16b, v0.4b[0]\n"
       ".inst 0x6f81e3d5  // udot v21.4s, v30.16b, v1.4b[0]\n"
       ".inst 0x6f82e3d9  // udot v25.4s, v30.16b, v2.4b[0]\n"
@@ -1092,36 +1092,36 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6f82e39b  // udot v27.4s, v28.16b, v2.4b[0]\n"
       "bge 72b\n"
       "74:"  // Height 3: Multiply loop: Skip odd blocks
-      "cbz x11, 78f\n"
-      "tbz x11, #1, 75f\n"
-      "ldr h0, [x10], #0x2\n"
-      "ldr h1, [x9], #0x2\n"
-      "ldr h2, [x28], #0x2\n"
-      "tbz x11, #0, 76f\n"
-      "ld1 { v0.b }[2], [x10]\n"
-      "ld1 { v1.b }[2], [x9]\n"
-      "ld1 { v2.b }[2], [x28]\n"
+      "cbz x10, 78f\n"
+      "tbz x10, #1, 75f\n"
+      "ldr h0, [x9], #0x2\n"
+      "ldr h1, [x28], #0x2\n"
+      "ldr h2, [x27], #0x2\n"
+      "tbz x10, #0, 76f\n"
+      "ld1 { v0.b }[2], [x9]\n"
+      "ld1 { v1.b }[2], [x28]\n"
+      "ld1 { v2.b }[2], [x27]\n"
       "b 76f\n"
       "75:"  // Height 3: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x10, #0x0]\n"
-      "ldr b1, [x9, #0x0]\n"
-      "ldr b2, [x28, #0x0]\n"
+      "ldr b0, [x9, #0x0]\n"
+      "ldr b1, [x28, #0x0]\n"
+      "ldr b2, [x27, #0x0]\n"
       "76:"  // Height 3: Multiply loop: Ragged operand read: Done
       "tbnz %x[flags], #31, 77f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       "77:"  // Height 3: Multiply loop: unique 12: skip row sum
-      "ldr q31, [x14, #0x0]\n"
-      "ldr q30, [x14, #0x10]\n"
-      "ldr q29, [x14, #0x20]\n"
-      "ldr q28, [x14, #0x30]\n"
-      ".inst 0x6f80e3f0  // udot v16.4s, v31.16b, v0.4b[0]\n"
-      ".inst 0x6f81e3f4  // udot v20.4s, v31.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
-      ".inst 0x6f82e3f8  // udot v24.4s, v31.16b, v2.4b[0]\n"
+      "ldr q28, [x12, #0x0]\n"
+      ".inst 0x6f80e390  // udot v16.4s, v28.16b, v0.4b[0]\n"
+      "ldr q30, [x12, #0x10]\n"
+      ".inst 0x6f81e394  // udot v20.4s, v28.16b, v1.4b[0]\n"
+      "ldr q29, [x12, #0x20]\n"
+      ".inst 0x6f82e398  // udot v24.4s, v28.16b, v2.4b[0]\n"
+      "ldr q28, [x12, #0x30]\n"
       ".inst 0x6f80e3d1  // udot v17.4s, v30.16b, v0.4b[0]\n"
       ".inst 0x6f81e3d5  // udot v21.4s, v30.16b, v1.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6f82e3d9  // udot v25.4s, v30.16b, v2.4b[0]\n"
       ".inst 0x6f80e3b2  // udot v18.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x6f81e3b6  // udot v22.4s, v29.16b, v1.4b[0]\n"
@@ -1131,15 +1131,15 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6f82e39b  // udot v27.4s, v28.16b, v2.4b[0]\n"
       "78:"  // Height 3: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x12, x12, #0x1\n"
-      "cmp x12, x20\n"
+      "add x11, x11, #0x1\n"
+      "cmp x11, x20\n"
       "bne 64b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20\n"
+      "add x22, x23, x20\n"
       "prfm pstl1keep, [x13, #0x0]\n"
-      "add x24, x13, x20\n"
-      "add x23, x24, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbnz %x[flags], #31, 79f\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
@@ -1154,13 +1154,13 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "mul v12.4s, v12.4s, v28.4s\n"
       "mul v13.4s, v13.4s, v28.4s\n"
       "79:"  // Height 3: skip row sum fixup
-      "ldr q31, [x16, #0x0]\n"
+      "ldr q31, [x15, #0x0]\n"
       "add v16.4s, v16.4s, v11.4s\n"
-      "ldr q30, [x16, #0x10]\n"
+      "ldr q30, [x15, #0x10]\n"
       "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q29, [x16, #0x20]\n"
+      "ldr q29, [x15, #0x20]\n"
       "add v18.4s, v18.4s, v11.4s\n"
-      "ldr q28, [x16, #0x30]\n"
+      "ldr q28, [x15, #0x30]\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
@@ -1183,11 +1183,10 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "add v26.4s, v26.4s, v29.4s\n"
       "add v27.4s, v27.4s, v28.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v28.4s }, [x20]\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add x16, x16, #0x40\n"
       "sqrdmulh v16.4s, v16.4s, v28.4s\n"
       "sqrdmulh v17.4s, v17.4s, v28.4s\n"
       "sqrdmulh v18.4s, v18.4s, v28.4s\n"
@@ -1200,38 +1199,39 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "sqrdmulh v25.4s, v25.4s, v28.4s\n"
       "sqrdmulh v26.4s, v26.4s, v28.4s\n"
       "sqrdmulh v27.4s, v27.4s, v28.4s\n"
+      "add x15, x15, #0x40\n"
       "tbz %x[flags], #5, 80f\n"
       "and v1.16b, v16.16b, v0.16b\n"
       "and v31.16b, v17.16b, v0.16b\n"
       "and v30.16b, v18.16b, v0.16b\n"
       "and v29.16b, v19.16b, v0.16b\n"
       "and v28.16b, v20.16b, v0.16b\n"
-      "and v3.16b, v21.16b, v0.16b\n"
-      "and v2.16b, v22.16b, v0.16b\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
-      "sshr v3.4s, v3.4s, #0x1f\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
       "sqadd v16.4s, v16.4s, v1.4s\n"
       "sqadd v17.4s, v17.4s, v31.4s\n"
       "sqadd v18.4s, v18.4s, v30.4s\n"
       "sqadd v19.4s, v19.4s, v29.4s\n"
       "sqadd v20.4s, v20.4s, v28.4s\n"
+      "and v3.16b, v21.16b, v0.16b\n"
+      "and v2.16b, v22.16b, v0.16b\n"
       "and v1.16b, v23.16b, v0.16b\n"
       "and v31.16b, v24.16b, v0.16b\n"
       "and v30.16b, v25.16b, v0.16b\n"
       "and v29.16b, v26.16b, v0.16b\n"
       "and v28.16b, v27.16b, v0.16b\n"
-      "sqadd v21.4s, v21.4s, v3.4s\n"
-      "sqadd v22.4s, v22.4s, v2.4s\n"
+      "sshr v3.4s, v3.4s, #0x1f\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
+      "sqadd v21.4s, v21.4s, v3.4s\n"
+      "sqadd v22.4s, v22.4s, v2.4s\n"
       "sqadd v23.4s, v23.4s, v1.4s\n"
       "sqadd v24.4s, v24.4s, v31.4s\n"
       "sqadd v25.4s, v25.4s, v30.4s\n"
@@ -1251,36 +1251,35 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "srshl v26.4s, v26.4s, v0.4s\n"
       "srshl v27.4s, v27.4s, v0.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v30.4s }, [x20]\n"
+      "ld1r { v28.4s }, [x20]\n"
+      "add v16.4s, v16.4s, v28.4s\n"
+      "add v17.4s, v17.4s, v28.4s\n"
+      "add v18.4s, v18.4s, v28.4s\n"
+      "add v19.4s, v19.4s, v28.4s\n"
+      "add v20.4s, v20.4s, v28.4s\n"
+      "add v21.4s, v21.4s, v28.4s\n"
+      "add v22.4s, v22.4s, v28.4s\n"
+      "add v23.4s, v23.4s, v28.4s\n"
+      "add v24.4s, v24.4s, v28.4s\n"
+      "add v25.4s, v25.4s, v28.4s\n"
+      "add v26.4s, v26.4s, v28.4s\n"
+      "add v27.4s, v27.4s, v28.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v28.4s }, [x20]\n"
+      "smin v16.4s, v16.4s, v28.4s\n"
+      "smin v17.4s, v17.4s, v28.4s\n"
+      "smin v18.4s, v18.4s, v28.4s\n"
+      "smin v19.4s, v19.4s, v28.4s\n"
+      "smin v20.4s, v20.4s, v28.4s\n"
+      "smin v21.4s, v21.4s, v28.4s\n"
+      "smin v22.4s, v22.4s, v28.4s\n"
+      "smin v23.4s, v23.4s, v28.4s\n"
+      "smin v24.4s, v24.4s, v28.4s\n"
+      "smin v25.4s, v25.4s, v28.4s\n"
+      "smin v26.4s, v26.4s, v28.4s\n"
+      "smin v27.4s, v27.4s, v28.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v29.4s }, [x21]\n"
-      "cmp x15, #0x10\n"
       "ld1r { v28.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v30.4s\n"
-      "add v17.4s, v17.4s, v30.4s\n"
-      "add v18.4s, v18.4s, v30.4s\n"
-      "add v19.4s, v19.4s, v30.4s\n"
-      "add v20.4s, v20.4s, v30.4s\n"
-      "add v21.4s, v21.4s, v30.4s\n"
-      "add v22.4s, v22.4s, v30.4s\n"
-      "add v23.4s, v23.4s, v30.4s\n"
-      "add v24.4s, v24.4s, v30.4s\n"
-      "add v25.4s, v25.4s, v30.4s\n"
-      "add v26.4s, v26.4s, v30.4s\n"
-      "add v27.4s, v27.4s, v30.4s\n"
-      "smin v16.4s, v16.4s, v29.4s\n"
-      "smin v17.4s, v17.4s, v29.4s\n"
-      "smin v18.4s, v18.4s, v29.4s\n"
-      "smin v19.4s, v19.4s, v29.4s\n"
-      "smin v20.4s, v20.4s, v29.4s\n"
-      "smin v21.4s, v21.4s, v29.4s\n"
-      "smin v22.4s, v22.4s, v29.4s\n"
-      "smin v23.4s, v23.4s, v29.4s\n"
-      "smin v24.4s, v24.4s, v29.4s\n"
-      "smin v25.4s, v25.4s, v29.4s\n"
-      "smin v26.4s, v26.4s, v29.4s\n"
-      "smin v27.4s, v27.4s, v29.4s\n"
       "smax v16.4s, v16.4s, v28.4s\n"
       "smax v17.4s, v17.4s, v28.4s\n"
       "smax v18.4s, v18.4s, v28.4s\n"
@@ -1299,109 +1298,109 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "uzp1 v18.8h, v22.8h, v23.8h\n"
       "uzp1 v24.8h, v24.8h, v25.8h\n"
       "uzp1 v17.8h, v26.8h, v27.8h\n"
+      "cmp x14, #0x10\n"
       "uzp1 v16.16b, v16.16b, v19.16b\n"
       "uzp1 v20.16b, v20.16b, v18.16b\n"
       "uzp1 v24.16b, v24.16b, v17.16b\n"
       "bge 89f\n"
-      "tbz x15, #3, 84f\n"
+      "tbz x14, #3, 84f\n"
       "str d16, [x13], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "tbz x15, #2, 82f\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "tbz x14, #2, 82f\n"
       "st1 { v16.s }[2], [x13], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "tbz x15, #1, 81f\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "tbz x14, #1, 81f\n"
       "st1 { v16.h }[6], [x13], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "tbz x15, #0, 88f\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[14], [x13]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 88f\n"
       "81:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x15, #0, 88f\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[12], [x13]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 88f\n"
       "82:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x15, #1, 83f\n"
+      "tbz x14, #1, 83f\n"
       "st1 { v16.h }[4], [x13], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "tbz x15, #0, 88f\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[10], [x13]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 88f\n"
       "83:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x15, #0, 88f\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[8], [x13]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 88f\n"
       "84:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x15, #2, 86f\n"
+      "tbz x14, #2, 86f\n"
       "str s16, [x13], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "tbz x15, #1, 85f\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "tbz x14, #1, 85f\n"
       "st1 { v16.h }[2], [x13], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "tbz x15, #0, 88f\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[6], [x13]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 88f\n"
       "85:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x15, #0, 88f\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[4], [x13]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 88f\n"
       "86:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x15, #1, 87f\n"
+      "tbz x14, #1, 87f\n"
       "str h16, [x13], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "tbz x15, #0, 88f\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "tbz x14, #0, 88f\n"
       "st1 { v16.b }[2], [x13]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 88f\n"
       "87:"  // Height 3: Partial direct writeback: partial_1_0
       "str b16, [x13, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "88:"  // Height 3: Partial direct writeback: Done
       "b 90f\n"
       "89:"  // Height 3: Full writeback
       "str q16, [x13, #0x0]\n"
       "add x13, x13, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "90:"  // Height 3: Writeback done
-      "subs x15, x15, #0x10\n"
+      "subs x14, x14, #0x10\n"
       "bgt 62b\n"
       "b 122f\n"
       "91:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "mov x20, #0x4\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x16, %x[col_bias]\n"
+      "mov x15, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
       "movi v12.4s, #0x0\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_N]]\n"
-      "madd x20, x21, x20, x13\n"
+      "bic %x[flags], %x[flags], #0x80000000\n"
       "movi v13.4s, #0x0\n"
+      "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "movi v14.4s, #0x0\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "movi v15.16b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "92:"  // Height 4: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -1420,118 +1419,118 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "movi v30.4s, #0x0\n"
       "movi v31.4s, #0x0\n"
       "93:"  // Height 4: setup done
-      "mov x12, #0x0\n"
+      "mov x11, #0x0\n"
       "94:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+      "ldr w10, [x20, x11, LSL #0x2]\n"
       "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w11, [x20, x12, LSL #0x2]\n"
       "tbz %x[flags], #3, 95f\n"
-      "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n"
+      "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
-      "ldr x10, [x20, #0x0]\n"
-      "ldr x9, [x20, #0x8]\n"
-      "ldr x28, [x20, #0x10]\n"
-      "ldr x27, [x20, #0x18]\n"
-      "cbnz x12, 96f\n"
+      "ldr x9, [x20, #0x0]\n"
+      "ldr x28, [x20, #0x8]\n"
+      "ldr x27, [x20, #0x10]\n"
+      "ldr x26, [x20, #0x18]\n"
+      "cbnz x11, 96f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x10, x10, x20\n"
       "add x9, x9, x20\n"
       "add x28, x28, x20\n"
       "add x27, x27, x20\n"
+      "add x26, x26, x20\n"
       "b 96f\n"
       "95:"  // Height 4: setup direct input
-      "mov x10, %x[input_ptr]\n"
-      "add x9, x10, x21\n"
+      "mov x9, %x[input_ptr]\n"
       "add x28, x9, x21\n"
       "add x27, x28, x21\n"
+      "add x26, x27, x21\n"
       "96:"  // Height 4: input setup done
-      "cmp x11, #0x10\n"
+      "cmp x10, #0x10\n"
       "blt 101f\n"
-      "ldr q0, [x10, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q1, [x9, #0x0]\n"
-      "ldr q2, [x28, #0x0]\n"
-      "ldr q3, [x27, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q1, [x28, #0x0]\n"
+      "ldr q2, [x27, #0x0]\n"
+      "ldr q3, [x26, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "blt 99f\n"
       "97:"  // Height 4: Multiply loop: Main loop head
       ".inst 0x6f80e090  // udot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr x21, [x14, #0x78]\n"
+      "ldr x22, [x12, #0x78]\n"
       ".inst 0x6f81e094  // udot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr x20, [x14, #0x88]\n"
+      "ldr x21, [x12, #0x88]\n"
       ".inst 0x6f82e098  // udot v24.4s, v4.16b, v2.4b[0]\n"
-      "ldr x26, [x14, #0x98]\n"
+      "ldr x20, [x12, #0x98]\n"
       ".inst 0x6f83e09c  // udot v28.4s, v4.16b, v3.4b[0]\n"
-      "ldr d4, [x14, #0x70]\n"
+      "ldr d4, [x12, #0x70]\n"
       ".inst 0x6f80e0b1  // udot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr x25, [x14, #0xa8]\n"
+      "mov v4.d[1], x22\n"
       ".inst 0x6f81e0b5  // udot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr x24, [x14, #0xb8]\n"
+      "ldr x25, [x12, #0xa8]\n"
       ".inst 0x6f82e0b9  // udot v25.4s, v5.16b, v2.4b[0]\n"
-      "mov v4.d[1], x21\n"
+      "ldr x24, [x12, #0xb8]\n"
       ".inst 0x6f83e0bd  // udot v29.4s, v5.16b, v3.4b[0]\n"
-      "ldr d5, [x14, #0x80]\n"
+      "ldr d5, [x12, #0x80]\n"
       ".inst 0x6f80e0d2  // udot v18.4s, v6.16b, v0.4b[0]\n"
-      "ldr x23, [x14, #0xc8]\n"
+      "mov v5.d[1], x21\n"
       ".inst 0x6f81e0d6  // udot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr x22, [x14, #0xd8]\n"
+      "ldr x23, [x12, #0xc8]\n"
       ".inst 0x6f82e0da  // udot v26.4s, v6.16b, v2.4b[0]\n"
-      "mov v5.d[1], x20\n"
+      "ldr x22, [x12, #0xd8]\n"
       ".inst 0x6f83e0de  // udot v30.4s, v6.16b, v3.4b[0]\n"
-      "ldr d6, [x14, #0x90]\n"
+      "ldr d6, [x12, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr x21, [x14, #0xe8]\n"
+      "mov v6.d[1], x20\n"
       ".inst 0x6f81e0f7  // udot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr x20, [x14, #0xf8]\n"
+      "ldr x21, [x12, #0xe8]\n"
       ".inst 0x6f82e0fb  // udot v27.4s, v7.16b, v2.4b[0]\n"
-      "mov v6.d[1], x26\n"
+      "ldr x20, [x12, #0xf8]\n"
       ".inst 0x6f83e0ff  // udot v31.4s, v7.16b, v3.4b[0]\n"
-      "ldr d7, [x14, #0xa0]\n"
+      "ldr d7, [x12, #0xa0]\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
-      "add x10, x10, #0x10\n"
+      "mov v7.d[1], x25\n"
       ".inst 0x6fa1e114  // udot v20.4s, v8.16b, v1.4b[1]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x6fa2e118  // udot v24.4s, v8.16b, v2.4b[1]\n"
-      "mov v7.d[1], x25\n"
+      "add x28, x28, #0x10\n"
       ".inst 0x6fa3e11c  // udot v28.4s, v8.16b, v3.4b[1]\n"
-      "ldr d8, [x14, #0xb0]\n"
+      "ldr d8, [x12, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
-      "add x28, x28, #0x10\n"
+      "mov v8.d[1], x24\n"
       ".inst 0x6fa1e135  // udot v21.4s, v9.16b, v1.4b[1]\n"
       "add x27, x27, #0x10\n"
       ".inst 0x6fa2e139  // udot v25.4s, v9.16b, v2.4b[1]\n"
-      "mov v8.d[1], x24\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6fa3e13d  // udot v29.4s, v9.16b, v3.4b[1]\n"
-      "ldr d9, [x14, #0xc0]\n"
+      "ldr d9, [x12, #0xc0]\n"
       ".inst 0x6fa0e152  // udot v18.4s, v10.16b, v0.4b[1]\n"
+      "mov v9.d[1], x23\n"
       ".inst 0x6fa1e156  // udot v22.4s, v10.16b, v1.4b[1]\n"
       ".inst 0x6fa2e15a  // udot v26.4s, v10.16b, v2.4b[1]\n"
-      "mov v9.d[1], x23\n"
       ".inst 0x6fa3e15e  // udot v30.4s, v10.16b, v3.4b[1]\n"
-      "ldr d10, [x14, #0xd0]\n"
+      "ldr d10, [x12, #0xd0]\n"
       ".inst 0x6fa0e093  // udot v19.4s, v4.16b, v0.4b[1]\n"
+      "mov v10.d[1], x22\n"
       ".inst 0x6fa1e097  // udot v23.4s, v4.16b, v1.4b[1]\n"
       ".inst 0x6fa2e09b  // udot v27.4s, v4.16b, v2.4b[1]\n"
-      "mov v10.d[1], x22\n"
       ".inst 0x6fa3e09f  // udot v31.4s, v4.16b, v3.4b[1]\n"
-      "ldr d4, [x14, #0xe0]\n"
+      "ldr d4, [x12, #0xe0]\n"
       ".inst 0x6f80e8b0  // udot v16.4s, v5.16b, v0.4b[2]\n"
+      "mov v4.d[1], x21\n"
       ".inst 0x6f81e8b4  // udot v20.4s, v5.16b, v1.4b[2]\n"
       ".inst 0x6f82e8b8  // udot v24.4s, v5.16b, v2.4b[2]\n"
-      "mov v4.d[1], x21\n"
       ".inst 0x6f83e8bc  // udot v28.4s, v5.16b, v3.4b[2]\n"
-      "ldr d5, [x14, #0xf0]\n"
+      "ldr d5, [x12, #0xf0]\n"
       ".inst 0x6f80e8d1  // udot v17.4s, v6.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "mov v5.d[1], x20\n"
       ".inst 0x6f81e8d5  // udot v21.4s, v6.16b, v1.4b[2]\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x6f82e8d9  // udot v25.4s, v6.16b, v2.4b[2]\n"
-      "mov v5.d[1], x20\n"
       ".inst 0x6f83e8dd  // udot v29.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x6f80e8f2  // udot v18.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x6f81e8f6  // udot v22.4s, v7.16b, v1.4b[2]\n"
@@ -1563,77 +1562,77 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       ".inst 0x6e8f946e  // udot v14.4s, v3.16b, v15.16b\n"
       "98:"  // Height 4: Multiply loop: unique 13: skip row sum
-      "ldr q0, [x10, #0x0]\n"
-      "sub x11, x11, #0x10\n"
-      "ldr q1, [x9, #0x0]\n"
-      "cmp x11, #0x20\n"
-      "ldr q2, [x28, #0x0]\n"
-      "ldr q3, [x27, #0x0]\n"
-      "ldr q4, [x14, #0x0]\n"
-      "ldr q5, [x14, #0x10]\n"
-      "ldr q6, [x14, #0x20]\n"
-      "ldr q7, [x14, #0x30]\n"
-      "ldr q8, [x14, #0x40]\n"
-      "ldr q9, [x14, #0x50]\n"
-      "ldr q10, [x14, #0x60]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr q0, [x9, #0x0]\n"
+      "sub x10, x10, #0x10\n"
+      "ldr q1, [x28, #0x0]\n"
+      "cmp x10, #0x20\n"
+      "ldr q2, [x27, #0x0]\n"
+      "ldr q3, [x26, #0x0]\n"
+      "ldr q4, [x12, #0x0]\n"
+      "ldr q5, [x12, #0x10]\n"
+      "ldr q6, [x12, #0x20]\n"
+      "ldr q7, [x12, #0x30]\n"
+      "ldr q8, [x12, #0x40]\n"
+      "ldr q9, [x12, #0x50]\n"
+      "ldr q10, [x12, #0x60]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
       "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 97b\n"
       "99:"  // Height 4: Multiply loop: Single iteration only
       ".inst 0x6f80e090  // udot v16.4s, v4.16b, v0.4b[0]\n"
-      "sub x11, x11, #0x10\n"
+      "sub x10, x10, #0x10\n"
       ".inst 0x6f81e094  // udot v20.4s, v4.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
-      ".inst 0x6f82e098  // udot v24.4s, v4.16b, v2.4b[0]\n"
       "add x9, x9, #0x10\n"
+      ".inst 0x6f82e098  // udot v24.4s, v4.16b, v2.4b[0]\n"
+      "add x28, x28, #0x10\n"
       ".inst 0x6f83e09c  // udot v28.4s, v4.16b, v3.4b[0]\n"
-      "ldr q4, [x14, #0x70]\n"
+      "ldr q4, [x12, #0x70]\n"
       ".inst 0x6f80e0b1  // udot v17.4s, v5.16b, v0.4b[0]\n"
-      "add x28, x28, #0x10\n"
-      ".inst 0x6f81e0b5  // udot v21.4s, v5.16b, v1.4b[0]\n"
       "add x27, x27, #0x10\n"
+      ".inst 0x6f81e0b5  // udot v21.4s, v5.16b, v1.4b[0]\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6f82e0b9  // udot v25.4s, v5.16b, v2.4b[0]\n"
       ".inst 0x6f83e0bd  // udot v29.4s, v5.16b, v3.4b[0]\n"
-      "ldr q5, [x14, #0x80]\n"
+      "ldr q5, [x12, #0x80]\n"
       ".inst 0x6f80e0d2  // udot v18.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x6f81e0d6  // udot v22.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x6f82e0da  // udot v26.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x6f83e0de  // udot v30.4s, v6.16b, v3.4b[0]\n"
-      "ldr q6, [x14, #0x90]\n"
+      "ldr q6, [x12, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x6f81e0f7  // udot v23.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x6f82e0fb  // udot v27.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x6f83e0ff  // udot v31.4s, v7.16b, v3.4b[0]\n"
-      "ldr q7, [x14, #0xa0]\n"
+      "ldr q7, [x12, #0xa0]\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
       ".inst 0x6fa1e114  // udot v20.4s, v8.16b, v1.4b[1]\n"
       ".inst 0x6fa2e118  // udot v24.4s, v8.16b, v2.4b[1]\n"
       ".inst 0x6fa3e11c  // udot v28.4s, v8.16b, v3.4b[1]\n"
-      "ldr q8, [x14, #0xb0]\n"
+      "ldr q8, [x12, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
       ".inst 0x6fa1e135  // udot v21.4s, v9.16b, v1.4b[1]\n"
       ".inst 0x6fa2e139  // udot v25.4s, v9.16b, v2.4b[1]\n"
       ".inst 0x6fa3e13d  // udot v29.4s, v9.16b, v3.4b[1]\n"
-      "ldr q9, [x14, #0xc0]\n"
+      "ldr q9, [x12, #0xc0]\n"
       ".inst 0x6fa0e152  // udot v18.4s, v10.16b, v0.4b[1]\n"
       ".inst 0x6fa1e156  // udot v22.4s, v10.16b, v1.4b[1]\n"
       ".inst 0x6fa2e15a  // udot v26.4s, v10.16b, v2.4b[1]\n"
       ".inst 0x6fa3e15e  // udot v30.4s, v10.16b, v3.4b[1]\n"
-      "ldr q10, [x14, #0xd0]\n"
+      "ldr q10, [x12, #0xd0]\n"
       ".inst 0x6fa0e093  // udot v19.4s, v4.16b, v0.4b[1]\n"
       ".inst 0x6fa1e097  // udot v23.4s, v4.16b, v1.4b[1]\n"
       ".inst 0x6fa2e09b  // udot v27.4s, v4.16b, v2.4b[1]\n"
       ".inst 0x6fa3e09f  // udot v31.4s, v4.16b, v3.4b[1]\n"
-      "ldr q4, [x14, #0xe0]\n"
+      "ldr q4, [x12, #0xe0]\n"
       ".inst 0x6f80e8b0  // udot v16.4s, v5.16b, v0.4b[2]\n"
       ".inst 0x6f81e8b4  // udot v20.4s, v5.16b, v1.4b[2]\n"
       ".inst 0x6f82e8b8  // udot v24.4s, v5.16b, v2.4b[2]\n"
       ".inst 0x6f83e8bc  // udot v28.4s, v5.16b, v3.4b[2]\n"
-      "ldr q5, [x14, #0xf0]\n"
+      "ldr q5, [x12, #0xf0]\n"
       ".inst 0x6f80e8d1  // udot v17.4s, v6.16b, v0.4b[2]\n"
-      "add x14, x14, #0x100\n"
+      "add x12, x12, #0x100\n"
       ".inst 0x6f81e8d5  // udot v21.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x6f82e8d9  // udot v25.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x6f83e8dd  // udot v29.4s, v6.16b, v3.4b[2]\n"
@@ -1667,35 +1666,35 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       ".inst 0x6e8f946e  // udot v14.4s, v3.16b, v15.16b\n"
       "100:"  // Height 4: Multiply loop: unique 14: skip row sum
-      "prfm pldl1keep, [x10, #0x80]\n"
       "prfm pldl1keep, [x9, #0x80]\n"
       "prfm pldl1keep, [x28, #0x80]\n"
       "prfm pldl1keep, [x27, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "101:"  // Height 4: Multiply loop: Main loop skip
-      "cbz x11, 108f\n"
-      "cmp x11, #0x4\n"
+      "cbz x10, 108f\n"
+      "cmp x10, #0x4\n"
       "blt 104f\n"
       "102:"  // Height 4: Multiply loop: Odd block loop
-      "ldr s0, [x10], #0x4\n"
-      "ldr s1, [x9], #0x4\n"
-      "ldr s2, [x28], #0x4\n"
-      "ldr s3, [x27], #0x4\n"
+      "ldr s0, [x9], #0x4\n"
+      "ldr s1, [x28], #0x4\n"
+      "ldr s2, [x27], #0x4\n"
+      "ldr s3, [x26], #0x4\n"
       "tbnz %x[flags], #31, 103f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       ".inst 0x6e8f946e  // udot v14.4s, v3.16b, v15.16b\n"
       "103:"  // Height 4: Multiply loop: unique 15: skip row sum
-      "ldr q7, [x14, #0x0]\n"
-      "sub x11, x11, #0x4\n"
-      "ldr q6, [x14, #0x10]\n"
-      "cmp x11, #0x4\n"
-      "ldr q5, [x14, #0x20]\n"
-      "ldr q4, [x14, #0x30]\n"
+      "ldr q7, [x12, #0x0]\n"
+      "sub x10, x10, #0x4\n"
+      "ldr q6, [x12, #0x10]\n"
+      "cmp x10, #0x4\n"
+      "ldr q5, [x12, #0x20]\n"
       ".inst 0x6f80e0f0  // udot v16.4s, v7.16b, v0.4b[0]\n"
+      "ldr q4, [x12, #0x30]\n"
       ".inst 0x6f81e0f4  // udot v20.4s, v7.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
       ".inst 0x6f82e0f8  // udot v24.4s, v7.16b, v2.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6f83e0fc  // udot v28.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x6f80e0d1  // udot v17.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x6f81e0d5  // udot v21.4s, v6.16b, v1.4b[0]\n"
@@ -1711,23 +1710,23 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6f83e09f  // udot v31.4s, v4.16b, v3.4b[0]\n"
       "bge 102b\n"
       "104:"  // Height 4: Multiply loop: Skip odd blocks
-      "cbz x11, 108f\n"
-      "tbz x11, #1, 105f\n"
-      "ldr h0, [x10], #0x2\n"
-      "ldr h1, [x9], #0x2\n"
-      "ldr h2, [x28], #0x2\n"
-      "ldr h3, [x27], #0x2\n"
-      "tbz x11, #0, 106f\n"
-      "ld1 { v0.b }[2], [x10]\n"
-      "ld1 { v1.b }[2], [x9]\n"
-      "ld1 { v2.b }[2], [x28]\n"
-      "ld1 { v3.b }[2], [x27]\n"
+      "cbz x10, 108f\n"
+      "tbz x10, #1, 105f\n"
+      "ldr h0, [x9], #0x2\n"
+      "ldr h1, [x28], #0x2\n"
+      "ldr h2, [x27], #0x2\n"
+      "ldr h3, [x26], #0x2\n"
+      "tbz x10, #0, 106f\n"
+      "ld1 { v0.b }[2], [x9]\n"
+      "ld1 { v1.b }[2], [x28]\n"
+      "ld1 { v2.b }[2], [x27]\n"
+      "ld1 { v3.b }[2], [x26]\n"
       "b 106f\n"
       "105:"  // Height 4: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x10, #0x0]\n"
-      "ldr b1, [x9, #0x0]\n"
-      "ldr b2, [x28, #0x0]\n"
-      "ldr b3, [x27, #0x0]\n"
+      "ldr b0, [x9, #0x0]\n"
+      "ldr b1, [x28, #0x0]\n"
+      "ldr b2, [x27, #0x0]\n"
+      "ldr b3, [x26, #0x0]\n"
       "106:"  // Height 4: Multiply loop: Ragged operand read: Done
       "tbnz %x[flags], #31, 107f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
@@ -1735,16 +1734,16 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       ".inst 0x6e8f946e  // udot v14.4s, v3.16b, v15.16b\n"
       "107:"  // Height 4: Multiply loop: unique 16: skip row sum
-      "ldr q7, [x14, #0x0]\n"
-      "ldr q6, [x14, #0x10]\n"
-      "ldr q5, [x14, #0x20]\n"
-      "ldr q4, [x14, #0x30]\n"
+      "ldr q7, [x12, #0x0]\n"
       ".inst 0x6f80e0f0  // udot v16.4s, v7.16b, v0.4b[0]\n"
+      "ldr q6, [x12, #0x10]\n"
       ".inst 0x6f81e0f4  // udot v20.4s, v7.16b, v1.4b[0]\n"
-      "add x14, x14, #0x40\n"
+      "ldr q5, [x12, #0x20]\n"
       ".inst 0x6f82e0f8  // udot v24.4s, v7.16b, v2.4b[0]\n"
+      "ldr q4, [x12, #0x30]\n"
       ".inst 0x6f83e0fc  // udot v28.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x6f80e0d1  // udot v17.4s, v6.16b, v0.4b[0]\n"
+      "add x12, x12, #0x40\n"
       ".inst 0x6f81e0d5  // udot v21.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x6f82e0d9  // udot v25.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x6f83e0dd  // udot v29.4s, v6.16b, v3.4b[0]\n"
@@ -1758,17 +1757,17 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       ".inst 0x6f83e09f  // udot v31.4s, v4.16b, v3.4b[0]\n"
       "108:"  // Height 4: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x12, x12, #0x1\n"
-      "cmp x12, x20\n"
+      "add x11, x11, #0x1\n"
+      "cmp x11, x20\n"
       "bne 94b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x13, #0x0]\n"
-      "add x24, x13, x20\n"
-      "add x23, x24, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "add x23, x13, x20\n"
       "add x22, x23, x20\n"
+      "add x21, x22, x20\n"
+      "prfm pstl1keep, [x13, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbnz %x[flags], #31, 109f\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
@@ -1776,9 +1775,9 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "addp v14.4s, v14.4s, v14.4s\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1r { v0.4s }, [x20]\n"
+      "neg v0.4s, v0.4s\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
-      "neg v0.4s, v0.4s\n"
       "addp v13.4s, v13.4s, v13.4s\n"
       "addp v14.4s, v14.4s, v14.4s\n"
       "mul v11.4s, v11.4s, v0.4s\n"
@@ -1786,13 +1785,13 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "mul v13.4s, v13.4s, v0.4s\n"
       "mul v14.4s, v14.4s, v0.4s\n"
       "109:"  // Height 4: skip row sum fixup
-      "ldr q3, [x16, #0x0]\n"
+      "ldr q3, [x15, #0x0]\n"
       "add v16.4s, v16.4s, v11.4s\n"
-      "ldr q2, [x16, #0x10]\n"
+      "ldr q2, [x15, #0x10]\n"
       "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q1, [x16, #0x20]\n"
+      "ldr q1, [x15, #0x20]\n"
       "add v18.4s, v18.4s, v11.4s\n"
-      "ldr q0, [x16, #0x30]\n"
+      "ldr q0, [x15, #0x30]\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
@@ -1823,11 +1822,10 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "add v30.4s, v30.4s, v1.4s\n"
       "add v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v1.4s }, [x20]\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add x16, x16, #0x40\n"
       "sqrdmulh v16.4s, v16.4s, v1.4s\n"
       "sqrdmulh v17.4s, v17.4s, v1.4s\n"
       "sqrdmulh v18.4s, v18.4s, v1.4s\n"
@@ -1844,51 +1842,52 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "sqrdmulh v29.4s, v29.4s, v1.4s\n"
       "sqrdmulh v30.4s, v30.4s, v1.4s\n"
       "sqrdmulh v31.4s, v31.4s, v1.4s\n"
+      "add x15, x15, #0x40\n"
       "tbz %x[flags], #5, 110f\n"
       "and v2.16b, v16.16b, v0.16b\n"
       "and v1.16b, v17.16b, v0.16b\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v16.4s, v16.4s, v2.4s\n"
+      "sqadd v17.4s, v17.4s, v1.4s\n"
       "and v7.16b, v18.16b, v0.16b\n"
       "and v6.16b, v19.16b, v0.16b\n"
       "and v5.16b, v20.16b, v0.16b\n"
       "and v4.16b, v21.16b, v0.16b\n"
       "and v3.16b, v22.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
+      "and v2.16b, v23.16b, v0.16b\n"
+      "and v1.16b, v24.16b, v0.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
       "sshr v6.4s, v6.4s, #0x1f\n"
       "sshr v5.4s, v5.4s, #0x1f\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sshr v3.4s, v3.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v2.4s\n"
-      "sqadd v17.4s, v17.4s, v1.4s\n"
-      "and v2.16b, v23.16b, v0.16b\n"
-      "and v1.16b, v24.16b, v0.16b\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
       "sqadd v18.4s, v18.4s, v7.4s\n"
       "sqadd v19.4s, v19.4s, v6.4s\n"
       "sqadd v20.4s, v20.4s, v5.4s\n"
       "sqadd v21.4s, v21.4s, v4.4s\n"
       "sqadd v22.4s, v22.4s, v3.4s\n"
+      "sqadd v23.4s, v23.4s, v2.4s\n"
+      "sqadd v24.4s, v24.4s, v1.4s\n"
       "and v7.16b, v25.16b, v0.16b\n"
       "and v6.16b, v26.16b, v0.16b\n"
       "and v5.16b, v27.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
       "and v4.16b, v28.16b, v0.16b\n"
       "and v3.16b, v29.16b, v0.16b\n"
+      "and v2.16b, v30.16b, v0.16b\n"
+      "and v1.16b, v31.16b, v0.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
       "sshr v6.4s, v6.4s, #0x1f\n"
       "sshr v5.4s, v5.4s, #0x1f\n"
-      "sqadd v23.4s, v23.4s, v2.4s\n"
-      "sqadd v24.4s, v24.4s, v1.4s\n"
-      "and v2.16b, v30.16b, v0.16b\n"
-      "and v1.16b, v31.16b, v0.16b\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sshr v3.4s, v3.4s, #0x1f\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
       "sqadd v25.4s, v25.4s, v7.4s\n"
       "sqadd v26.4s, v26.4s, v6.4s\n"
       "sqadd v27.4s, v27.4s, v5.4s\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
       "sqadd v28.4s, v28.4s, v4.4s\n"
       "sqadd v29.4s, v29.4s, v3.4s\n"
       "sqadd v30.4s, v30.4s, v2.4s\n"
@@ -1911,44 +1910,43 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "srshl v30.4s, v30.4s, v0.4s\n"
       "srshl v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[maxval]\n"
-      "ld1r { v2.4s }, [x20]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "add v16.4s, v16.4s, v0.4s\n"
+      "add v17.4s, v17.4s, v0.4s\n"
+      "add v18.4s, v18.4s, v0.4s\n"
+      "add v19.4s, v19.4s, v0.4s\n"
+      "add v20.4s, v20.4s, v0.4s\n"
+      "add v21.4s, v21.4s, v0.4s\n"
+      "add v22.4s, v22.4s, v0.4s\n"
+      "add v23.4s, v23.4s, v0.4s\n"
+      "add v24.4s, v24.4s, v0.4s\n"
+      "add v25.4s, v25.4s, v0.4s\n"
+      "add v26.4s, v26.4s, v0.4s\n"
+      "add v27.4s, v27.4s, v0.4s\n"
+      "add v28.4s, v28.4s, v0.4s\n"
+      "add v29.4s, v29.4s, v0.4s\n"
+      "add v30.4s, v30.4s, v0.4s\n"
+      "add v31.4s, v31.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v0.4s }, [x20]\n"
+      "smin v16.4s, v16.4s, v0.4s\n"
+      "smin v17.4s, v17.4s, v0.4s\n"
+      "smin v18.4s, v18.4s, v0.4s\n"
+      "smin v19.4s, v19.4s, v0.4s\n"
+      "smin v20.4s, v20.4s, v0.4s\n"
+      "smin v21.4s, v21.4s, v0.4s\n"
+      "smin v22.4s, v22.4s, v0.4s\n"
+      "smin v23.4s, v23.4s, v0.4s\n"
+      "smin v24.4s, v24.4s, v0.4s\n"
+      "smin v25.4s, v25.4s, v0.4s\n"
+      "smin v26.4s, v26.4s, v0.4s\n"
+      "smin v27.4s, v27.4s, v0.4s\n"
+      "smin v28.4s, v28.4s, v0.4s\n"
+      "smin v29.4s, v29.4s, v0.4s\n"
+      "smin v30.4s, v30.4s, v0.4s\n"
+      "smin v31.4s, v31.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "ld1r { v1.4s }, [x21]\n"
-      "cmp x15, #0x10\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v2.4s\n"
-      "add v17.4s, v17.4s, v2.4s\n"
-      "add v18.4s, v18.4s, v2.4s\n"
-      "add v19.4s, v19.4s, v2.4s\n"
-      "add v20.4s, v20.4s, v2.4s\n"
-      "add v21.4s, v21.4s, v2.4s\n"
-      "add v22.4s, v22.4s, v2.4s\n"
-      "add v23.4s, v23.4s, v2.4s\n"
-      "add v24.4s, v24.4s, v2.4s\n"
-      "add v25.4s, v25.4s, v2.4s\n"
-      "add v26.4s, v26.4s, v2.4s\n"
-      "add v27.4s, v27.4s, v2.4s\n"
-      "add v28.4s, v28.4s, v2.4s\n"
-      "add v29.4s, v29.4s, v2.4s\n"
-      "add v30.4s, v30.4s, v2.4s\n"
-      "add v31.4s, v31.4s, v2.4s\n"
-      "smin v16.4s, v16.4s, v1.4s\n"
-      "smin v17.4s, v17.4s, v1.4s\n"
-      "smin v18.4s, v18.4s, v1.4s\n"
-      "smin v19.4s, v19.4s, v1.4s\n"
-      "smin v20.4s, v20.4s, v1.4s\n"
-      "smin v21.4s, v21.4s, v1.4s\n"
-      "smin v22.4s, v22.4s, v1.4s\n"
-      "smin v23.4s, v23.4s, v1.4s\n"
-      "smin v24.4s, v24.4s, v1.4s\n"
-      "smin v25.4s, v25.4s, v1.4s\n"
-      "smin v26.4s, v26.4s, v1.4s\n"
-      "smin v27.4s, v27.4s, v1.4s\n"
-      "smin v28.4s, v28.4s, v1.4s\n"
-      "smin v29.4s, v29.4s, v1.4s\n"
-      "smin v30.4s, v30.4s, v1.4s\n"
-      "smin v31.4s, v31.4s, v1.4s\n"
       "smax v16.4s, v16.4s, v0.4s\n"
       "smax v17.4s, v17.4s, v0.4s\n"
       "smax v18.4s, v18.4s, v0.4s\n"
@@ -1973,109 +1971,110 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "uzp1 v18.8h, v26.8h, v27.8h\n"
       "uzp1 v28.8h, v28.8h, v29.8h\n"
       "uzp1 v17.8h, v30.8h, v31.8h\n"
+      "cmp x14, #0x10\n"
       "uzp1 v16.16b, v16.16b, v0.16b\n"
       "uzp1 v20.16b, v20.16b, v19.16b\n"
       "uzp1 v24.16b, v24.16b, v18.16b\n"
       "uzp1 v28.16b, v28.16b, v17.16b\n"
       "bge 119f\n"
-      "tbz x15, #3, 114f\n"
+      "tbz x14, #3, 114f\n"
       "str d16, [x13], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
-      "tbz x15, #2, 112f\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
+      "tbz x14, #2, 112f\n"
       "st1 { v16.s }[2], [x13], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
-      "tbz x15, #1, 111f\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
+      "tbz x14, #1, 111f\n"
       "st1 { v16.h }[6], [x13], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "st1 { v28.h }[6], [x22], #0x2\n"
-      "tbz x15, #0, 118f\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "st1 { v28.h }[6], [x21], #0x2\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[14], [x13]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "st1 { v28.b }[14], [x22]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
+      "st1 { v28.b }[14], [x21]\n"
       "b 118f\n"
       "111:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x15, #0, 118f\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[12], [x13]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "st1 { v28.b }[12], [x22]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
+      "st1 { v28.b }[12], [x21]\n"
       "b 118f\n"
       "112:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x15, #1, 113f\n"
+      "tbz x14, #1, 113f\n"
       "st1 { v16.h }[4], [x13], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "st1 { v28.h }[4], [x22], #0x2\n"
-      "tbz x15, #0, 118f\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "st1 { v28.h }[4], [x21], #0x2\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[10], [x13]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "st1 { v28.b }[10], [x22]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
+      "st1 { v28.b }[10], [x21]\n"
       "b 118f\n"
       "113:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x15, #0, 118f\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[8], [x13]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "st1 { v28.b }[8], [x22]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
+      "st1 { v28.b }[8], [x21]\n"
       "b 118f\n"
       "114:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x15, #2, 116f\n"
+      "tbz x14, #2, 116f\n"
       "str s16, [x13], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
-      "tbz x15, #1, 115f\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
+      "tbz x14, #1, 115f\n"
       "st1 { v16.h }[2], [x13], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "st1 { v28.h }[2], [x22], #0x2\n"
-      "tbz x15, #0, 118f\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "st1 { v28.h }[2], [x21], #0x2\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[6], [x13]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "st1 { v28.b }[6], [x22]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
+      "st1 { v28.b }[6], [x21]\n"
       "b 118f\n"
       "115:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x15, #0, 118f\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[4], [x13]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "st1 { v28.b }[4], [x22]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
+      "st1 { v28.b }[4], [x21]\n"
       "b 118f\n"
       "116:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x15, #1, 117f\n"
+      "tbz x14, #1, 117f\n"
       "str h16, [x13], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "str h28, [x22], #0x2\n"
-      "tbz x15, #0, 118f\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "str h28, [x21], #0x2\n"
+      "tbz x14, #0, 118f\n"
       "st1 { v16.b }[2], [x13]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "st1 { v28.b }[2], [x22]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
+      "st1 { v28.b }[2], [x21]\n"
       "b 118f\n"
       "117:"  // Height 4: Partial direct writeback: partial_1_0
       "str b16, [x13, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "str b28, [x22, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
+      "str b28, [x21, #0x0]\n"
       "118:"  // Height 4: Partial direct writeback: Done
       "b 120f\n"
       "119:"  // Height 4: Full writeback
       "str q16, [x13, #0x0]\n"
       "add x13, x13, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q28, [x22, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q28, [x21, #0x0]\n"
       "120:"  // Height 4: Writeback done
-      "subs x15, x15, #0x10\n"
+      "subs x14, x14, #0x10\n"
       "bgt 92b\n"
       "subs %x[M], %x[M], #0x4\n"
       "beq 122f\n"
@@ -2089,9 +2088,9 @@ void a64_hybrid_u8qa_dot_4x16_a55 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "122:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp
index 23315f3c0c..f07902a559 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp
@@ -45,18 +45,18 @@ void a64_hybrid_u8qa_dot_4x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -90,7 +90,7 @@ void a64_hybrid_u8qa_dot_4x16 (
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -100,8 +100,8 @@ void a64_hybrid_u8qa_dot_4x16 (
       "mov x26, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -134,7 +134,6 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q26, [x28, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
       "ldr q25, [x28, #0xa0]\n"
-      "add x24, x24, #0x10\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
       "ldr q24, [x28, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
@@ -145,10 +144,11 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q21, [x28, #0xe0]\n"
       ".inst 0x6f80ea90  // udot v16.4s, v20.16b, v0.4b[2]\n"
       "ldr q20, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
       ".inst 0x6f80eb51  // udot v17.4s, v26.16b, v0.4b[2]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6f80eb32  // udot v18.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x6f80eb13  // udot v19.4s, v24.16b, v0.4b[2]\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x6fa0eaf0  // udot v16.4s, v23.16b, v0.4b[3]\n"
       ".inst 0x6fa0ead1  // udot v17.4s, v22.16b, v0.4b[3]\n"
       ".inst 0x6fa0eab2  // udot v18.4s, v21.16b, v0.4b[3]\n"
@@ -159,9 +159,9 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q0, [x24, #0x0]\n"
       "ldr q4, [x28, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q5, [x28, #0x10]\n"
       "ldr q6, [x28, #0x20]\n"
-      "cmp x25, #0x20\n"
       "ldr q7, [x28, #0x30]\n"
       "ldr q8, [x28, #0x40]\n"
       "ldr q9, [x28, #0x50]\n"
@@ -177,8 +177,6 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q26, [x28, #0x90]\n"
       ".inst 0x6f80e0f3  // udot v19.4s, v7.16b, v0.4b[0]\n"
       "ldr q25, [x28, #0xa0]\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
       ".inst 0x6fa0e110  // udot v16.4s, v8.16b, v0.4b[1]\n"
       "ldr q24, [x28, #0xb0]\n"
       ".inst 0x6fa0e131  // udot v17.4s, v9.16b, v0.4b[1]\n"
@@ -189,10 +187,12 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q21, [x28, #0xe0]\n"
       ".inst 0x6f80ea90  // udot v16.4s, v20.16b, v0.4b[2]\n"
       "ldr q20, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
       ".inst 0x6f80eb51  // udot v17.4s, v26.16b, v0.4b[2]\n"
+      "sub x25, x25, #0x10\n"
       ".inst 0x6f80eb32  // udot v18.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x6f80eb13  // udot v19.4s, v24.16b, v0.4b[2]\n"
+      "add x24, x24, #0x10\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x6fa0eaf0  // udot v16.4s, v23.16b, v0.4b[3]\n"
       ".inst 0x6fa0ead1  // udot v17.4s, v22.16b, v0.4b[3]\n"
       ".inst 0x6fa0eab2  // udot v18.4s, v21.16b, v0.4b[3]\n"
@@ -213,14 +213,14 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q23, [x28, #0x0]\n"
       "ldr q22, [x28, #0x10]\n"
       "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ldr q21, [x28, #0x20]\n"
       "ldr q20, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x6f80e2f0  // udot v16.4s, v23.16b, v0.4b[0]\n"
       ".inst 0x6f80e2d1  // udot v17.4s, v22.16b, v0.4b[0]\n"
       ".inst 0x6f80e2b2  // udot v18.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x6f80e293  // udot v19.4s, v20.16b, v0.4b[0]\n"
+      "add x28, x28, #0x40\n"
       "bge 12b\n"
       "14:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x25, 18f\n"
@@ -235,15 +235,15 @@ void a64_hybrid_u8qa_dot_4x16 (
       "tbnz %x[flags], #31, 17f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "17:"  // Height 1: Multiply loop: unique 4: skip row sum
-      "ldr q23, [x28, #0x0]\n"
-      "ldr q22, [x28, #0x10]\n"
+      "ldr q21, [x28, #0x0]\n"
+      "ldr q20, [x28, #0x10]\n"
+      ".inst 0x6f80e2b0  // udot v16.4s, v21.16b, v0.4b[0]\n"
+      ".inst 0x6f80e291  // udot v17.4s, v20.16b, v0.4b[0]\n"
       "ldr q21, [x28, #0x20]\n"
       "ldr q20, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x6f80e2f0  // udot v16.4s, v23.16b, v0.4b[0]\n"
-      ".inst 0x6f80e2d1  // udot v17.4s, v22.16b, v0.4b[0]\n"
       ".inst 0x6f80e2b2  // udot v18.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x6f80e293  // udot v19.4s, v20.16b, v0.4b[0]\n"
+      "add x28, x28, #0x40\n"
       "18:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -252,8 +252,8 @@ void a64_hybrid_u8qa_dot_4x16 (
       "prfm pstl1keep, [x27, #0x0]\n"
       "tbnz %x[flags], #31, 19f\n"
       "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
       "ld1r { v20.4s }, [x20]\n"
+      "addp v11.4s, v11.4s, v11.4s\n"
       "neg v20.4s, v20.4s\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "mul v11.4s, v11.4s, v20.4s\n"
@@ -267,16 +267,16 @@ void a64_hybrid_u8qa_dot_4x16 (
       "add v18.4s, v18.4s, v11.4s\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v20.4s }, [x20]\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add v16.4s, v16.4s, v24.4s\n"
       "add v17.4s, v17.4s, v23.4s\n"
+      "add v18.4s, v18.4s, v22.4s\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add x10, x10, #0x40\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v22.4s\n"
       "add v19.4s, v19.4s, v21.4s\n"
       "sqrdmulh v16.4s, v16.4s, v20.4s\n"
+      "add x10, x10, #0x40\n"
       "sqrdmulh v17.4s, v17.4s, v20.4s\n"
       "sqrdmulh v18.4s, v18.4s, v20.4s\n"
       "sqrdmulh v19.4s, v19.4s, v20.4s\n"
@@ -294,21 +294,21 @@ void a64_hybrid_u8qa_dot_4x16 (
       "sqadd v18.4s, v18.4s, v21.4s\n"
       "sqadd v19.4s, v19.4s, v20.4s\n"
       "20:"  // Height 1: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v22.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v22.4s }, [x21]\n"
-      "ld1r { v21.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v20.4s }, [x20]\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "add v16.4s, v16.4s, v22.4s\n"
       "add v17.4s, v17.4s, v22.4s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1r { v20.4s }, [x20]\n"
       "add v18.4s, v18.4s, v22.4s\n"
       "add v19.4s, v19.4s, v22.4s\n"
+      "cmp x9, #0x10\n"
       "smin v16.4s, v16.4s, v21.4s\n"
       "smin v17.4s, v17.4s, v21.4s\n"
       "smin v18.4s, v18.4s, v21.4s\n"
@@ -381,7 +381,7 @@ void a64_hybrid_u8qa_dot_4x16 (
       "movi v15.16b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "32:"  // Height 2: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -395,8 +395,8 @@ void a64_hybrid_u8qa_dot_4x16 (
       "mov x26, #0x0\n"
       "34:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 35f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -476,9 +476,9 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q0, [x24, #0x0]\n"
       "ldr q1, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q4, [x28, #0x0]\n"
       "ldr q5, [x28, #0x10]\n"
-      "cmp x25, #0x20\n"
       "ldr q6, [x28, #0x20]\n"
       "ldr q7, [x28, #0x30]\n"
       "ldr q8, [x28, #0x40]\n"
@@ -553,14 +553,14 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q27, [x28, #0x0]\n"
       "ldr q26, [x28, #0x10]\n"
       "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ldr q25, [x28, #0x20]\n"
       "ldr q24, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x6f80e370  // udot v16.4s, v27.16b, v0.4b[0]\n"
       ".inst 0x6f81e374  // udot v20.4s, v27.16b, v1.4b[0]\n"
       ".inst 0x6f80e351  // udot v17.4s, v26.16b, v0.4b[0]\n"
       ".inst 0x6f81e355  // udot v21.4s, v26.16b, v1.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x6f80e332  // udot v18.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x6f81e336  // udot v22.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x6f80e313  // udot v19.4s, v24.16b, v0.4b[0]\n"
@@ -583,17 +583,17 @@ void a64_hybrid_u8qa_dot_4x16 (
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
       "47:"  // Height 2: Multiply loop: unique 8: skip row sum
-      "ldr q27, [x28, #0x0]\n"
+      "ldr q24, [x28, #0x0]\n"
       "ldr q26, [x28, #0x10]\n"
+      ".inst 0x6f80e310  // udot v16.4s, v24.16b, v0.4b[0]\n"
+      ".inst 0x6f81e314  // udot v20.4s, v24.16b, v1.4b[0]\n"
       "ldr q25, [x28, #0x20]\n"
       "ldr q24, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x6f80e370  // udot v16.4s, v27.16b, v0.4b[0]\n"
-      ".inst 0x6f81e374  // udot v20.4s, v27.16b, v1.4b[0]\n"
       ".inst 0x6f80e351  // udot v17.4s, v26.16b, v0.4b[0]\n"
       ".inst 0x6f81e355  // udot v21.4s, v26.16b, v1.4b[0]\n"
       ".inst 0x6f80e332  // udot v18.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x6f81e336  // udot v22.4s, v25.16b, v1.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x6f80e313  // udot v19.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x6f81e317  // udot v23.4s, v24.16b, v1.4b[0]\n"
       "48:"  // Height 2: Multiply loop: No odd multiplies
@@ -602,14 +602,14 @@ void a64_hybrid_u8qa_dot_4x16 (
       "cmp x26, x20\n"
       "bne 34b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "tbnz %x[flags], #31, 49f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v24.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v24.4s }, [x20]\n"
       "neg v24.4s, v24.4s\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
@@ -627,10 +627,10 @@ void a64_hybrid_u8qa_dot_4x16 (
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v24.4s }, [x20]\n"
       "add v22.4s, v22.4s, v12.4s\n"
       "add v23.4s, v23.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v16.4s, v16.4s, v28.4s\n"
       "add v17.4s, v17.4s, v27.4s\n"
@@ -652,45 +652,45 @@ void a64_hybrid_u8qa_dot_4x16 (
       "sqrdmulh v23.4s, v23.4s, v24.4s\n"
       "tbz %x[flags], #5, 50f\n"
       "and v24.16b, v16.16b, v0.16b\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
+      "sqadd v16.4s, v16.4s, v24.4s\n"
       "and v30.16b, v17.16b, v0.16b\n"
       "and v29.16b, v18.16b, v0.16b\n"
       "and v28.16b, v19.16b, v0.16b\n"
       "and v27.16b, v20.16b, v0.16b\n"
       "and v26.16b, v21.16b, v0.16b\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "and v25.16b, v22.16b, v0.16b\n"
+      "and v24.16b, v23.16b, v0.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v24.4s\n"
-      "and v24.16b, v23.16b, v0.16b\n"
       "sshr v26.4s, v26.4s, #0x1f\n"
       "sshr v25.4s, v25.4s, #0x1f\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v17.4s, v17.4s, v30.4s\n"
       "sqadd v18.4s, v18.4s, v29.4s\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v19.4s, v19.4s, v28.4s\n"
       "sqadd v20.4s, v20.4s, v27.4s\n"
       "sqadd v21.4s, v21.4s, v26.4s\n"
       "sqadd v22.4s, v22.4s, v25.4s\n"
       "sqadd v23.4s, v23.4s, v24.4s\n"
       "50:"  // Height 2: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x21]\n"
-      "ld1r { v25.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "add v16.4s, v16.4s, v26.4s\n"
       "add v17.4s, v17.4s, v26.4s\n"
       "add v18.4s, v18.4s, v26.4s\n"
@@ -724,68 +724,68 @@ void a64_hybrid_u8qa_dot_4x16 (
       "bge 59f\n"
       "tbz x9, #3, 54f\n"
       "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
+      "str d20, [x23], #0x8\n"
       "tbz x9, #2, 52f\n"
       "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
       "tbz x9, #1, 51f\n"
       "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
+      "st1 { v20.b }[14], [x23]\n"
       "b 58f\n"
       "51:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
+      "st1 { v20.b }[12], [x23]\n"
       "b 58f\n"
       "52:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x9, #1, 53f\n"
       "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
+      "st1 { v20.b }[10], [x23]\n"
       "b 58f\n"
       "53:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
+      "st1 { v20.b }[8], [x23]\n"
       "b 58f\n"
       "54:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x9, #2, 56f\n"
       "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
+      "str s20, [x23], #0x4\n"
       "tbz x9, #1, 55f\n"
       "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
+      "st1 { v20.b }[6], [x23]\n"
       "b 58f\n"
       "55:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
+      "st1 { v20.b }[4], [x23]\n"
       "b 58f\n"
       "56:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x9, #1, 57f\n"
       "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
+      "str h20, [x23], #0x2\n"
       "tbz x9, #0, 58f\n"
       "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
+      "st1 { v20.b }[2], [x23]\n"
       "b 58f\n"
       "57:"  // Height 2: Partial direct writeback: partial_1_0
       "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
       "58:"  // Height 2: Partial direct writeback: Done
       "b 60f\n"
       "59:"  // Height 2: Full writeback
       "str q16, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
       "60:"  // Height 2: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 32b\n"
@@ -799,7 +799,7 @@ void a64_hybrid_u8qa_dot_4x16 (
       "movi v15.16b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "62:"  // Height 3: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -817,8 +817,8 @@ void a64_hybrid_u8qa_dot_4x16 (
       "mov x26, #0x0\n"
       "64:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 65f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -920,9 +920,9 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q0, [x24, #0x0]\n"
       "ldr q1, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q2, [x22, #0x0]\n"
       "ldr q4, [x28, #0x0]\n"
-      "cmp x25, #0x20\n"
       "ldr q5, [x28, #0x10]\n"
       "ldr q6, [x28, #0x20]\n"
       "ldr q7, [x28, #0x30]\n"
@@ -1020,14 +1020,14 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q31, [x28, #0x0]\n"
       "ldr q30, [x28, #0x10]\n"
       "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ldr q29, [x28, #0x20]\n"
       "ldr q28, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x6f80e3f0  // udot v16.4s, v31.16b, v0.4b[0]\n"
       ".inst 0x6f81e3f4  // udot v20.4s, v31.16b, v1.4b[0]\n"
       ".inst 0x6f82e3f8  // udot v24.4s, v31.16b, v2.4b[0]\n"
       ".inst 0x6f80e3d1  // udot v17.4s, v30.16b, v0.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x6f81e3d5  // udot v21.4s, v30.16b, v1.4b[0]\n"
       ".inst 0x6f82e3d9  // udot v25.4s, v30.16b, v2.4b[0]\n"
       ".inst 0x6f80e3b2  // udot v18.4s, v29.16b, v0.4b[0]\n"
@@ -1060,15 +1060,15 @@ void a64_hybrid_u8qa_dot_4x16 (
       "77:"  // Height 3: Multiply loop: unique 12: skip row sum
       "ldr q31, [x28, #0x0]\n"
       "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x6f80e3f0  // udot v16.4s, v31.16b, v0.4b[0]\n"
       ".inst 0x6f81e3f4  // udot v20.4s, v31.16b, v1.4b[0]\n"
+      "ldr q29, [x28, #0x20]\n"
+      "ldr q28, [x28, #0x30]\n"
       ".inst 0x6f82e3f8  // udot v24.4s, v31.16b, v2.4b[0]\n"
       ".inst 0x6f80e3d1  // udot v17.4s, v30.16b, v0.4b[0]\n"
       ".inst 0x6f81e3d5  // udot v21.4s, v30.16b, v1.4b[0]\n"
       ".inst 0x6f82e3d9  // udot v25.4s, v30.16b, v2.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x6f80e3b2  // udot v18.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x6f81e3b6  // udot v22.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x6f82e3ba  // udot v26.4s, v29.16b, v2.4b[0]\n"
@@ -1081,16 +1081,16 @@ void a64_hybrid_u8qa_dot_4x16 (
       "cmp x26, x20\n"
       "bne 64b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20\n"
       "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "tbnz %x[flags], #31, 79f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v28.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v28.4s }, [x20]\n"
       "addp v13.4s, v13.4s, v13.4s\n"
       "neg v28.4s, v28.4s\n"
       "addp v11.4s, v11.4s, v11.4s\n"
@@ -1111,10 +1111,10 @@ void a64_hybrid_u8qa_dot_4x16 (
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v28.4s }, [x20]\n"
       "add v22.4s, v22.4s, v12.4s\n"
       "add v23.4s, v23.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v24.4s, v24.4s, v13.4s\n"
       "add v25.4s, v25.4s, v13.4s\n"
@@ -1152,18 +1152,18 @@ void a64_hybrid_u8qa_dot_4x16 (
       "and v30.16b, v18.16b, v0.16b\n"
       "and v29.16b, v19.16b, v0.16b\n"
       "and v28.16b, v20.16b, v0.16b\n"
-      "and v3.16b, v21.16b, v0.16b\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
       "sshr v31.4s, v31.4s, #0x1f\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
-      "and v2.16b, v22.16b, v0.16b\n"
       "sqadd v16.4s, v16.4s, v1.4s\n"
       "sqadd v17.4s, v17.4s, v31.4s\n"
       "sqadd v18.4s, v18.4s, v30.4s\n"
       "sqadd v19.4s, v19.4s, v29.4s\n"
       "sqadd v20.4s, v20.4s, v28.4s\n"
+      "and v3.16b, v21.16b, v0.16b\n"
+      "and v2.16b, v22.16b, v0.16b\n"
       "and v1.16b, v23.16b, v0.16b\n"
       "and v31.16b, v24.16b, v0.16b\n"
       "and v30.16b, v25.16b, v0.16b\n"
@@ -1184,21 +1184,21 @@ void a64_hybrid_u8qa_dot_4x16 (
       "sqadd v26.4s, v26.4s, v29.4s\n"
       "sqadd v27.4s, v27.4s, v28.4s\n"
       "80:"  // Height 3: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v30.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v30.4s }, [x21]\n"
-      "ld1r { v29.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v28.4s }, [x20]\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "srshl v24.4s, v24.4s, v0.4s\n"
       "srshl v25.4s, v25.4s, v0.4s\n"
       "srshl v26.4s, v26.4s, v0.4s\n"
@@ -1251,103 +1251,102 @@ void a64_hybrid_u8qa_dot_4x16 (
       "bge 89f\n"
       "tbz x9, #3, 84f\n"
       "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x9, #2, 82f\n"
       "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
       "tbz x9, #1, 81f\n"
       "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 88f\n"
       "81:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 88f\n"
       "82:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x9, #1, 83f\n"
       "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 88f\n"
       "83:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 88f\n"
       "84:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x9, #2, 86f\n"
       "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
       "tbz x9, #1, 85f\n"
       "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 88f\n"
       "85:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 88f\n"
       "86:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x9, #1, 87f\n"
       "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
       "tbz x9, #0, 88f\n"
       "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 88f\n"
       "87:"  // Height 3: Partial direct writeback: partial_1_0
       "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "88:"  // Height 3: Partial direct writeback: Done
       "b 90f\n"
       "89:"  // Height 3: Full writeback
       "str q16, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "90:"  // Height 3: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 62b\n"
       "b 122f\n"
       "91:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x4\n"
       "mov x10, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
+      "movi v13.4s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "movi v13.4s, #0x0\n"
       "movi v14.4s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
       "movi v15.16b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "92:"  // Height 4: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -1369,8 +1368,8 @@ void a64_hybrid_u8qa_dot_4x16 (
       "mov x26, #0x0\n"
       "94:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 95f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1494,9 +1493,9 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q0, [x24, #0x0]\n"
       "ldr q1, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q2, [x22, #0x0]\n"
       "ldr q3, [x21, #0x0]\n"
-      "cmp x25, #0x20\n"
       "ldr q4, [x28, #0x0]\n"
       "ldr q5, [x28, #0x10]\n"
       "ldr q6, [x28, #0x20]\n"
@@ -1617,14 +1616,14 @@ void a64_hybrid_u8qa_dot_4x16 (
       "ldr q7, [x28, #0x0]\n"
       "ldr q6, [x28, #0x10]\n"
       "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ldr q5, [x28, #0x20]\n"
       "ldr q4, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x6f80e0f0  // udot v16.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x6f81e0f4  // udot v20.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x6f82e0f8  // udot v24.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x6f83e0fc  // udot v28.4s, v7.16b, v3.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x6f80e0d1  // udot v17.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x6f81e0d5  // udot v21.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x6f82e0d9  // udot v25.4s, v6.16b, v2.4b[0]\n"
@@ -1665,15 +1664,15 @@ void a64_hybrid_u8qa_dot_4x16 (
       "107:"  // Height 4: Multiply loop: unique 16: skip row sum
       "ldr q7, [x28, #0x0]\n"
       "ldr q6, [x28, #0x10]\n"
-      "ldr q5, [x28, #0x20]\n"
-      "ldr q4, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
       ".inst 0x6f80e0f0  // udot v16.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x6f81e0f4  // udot v20.4s, v7.16b, v1.4b[0]\n"
+      "ldr q5, [x28, #0x20]\n"
+      "ldr q4, [x28, #0x30]\n"
       ".inst 0x6f82e0f8  // udot v24.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x6f83e0fc  // udot v28.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x6f80e0d1  // udot v17.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x6f81e0d5  // udot v21.4s, v6.16b, v1.4b[0]\n"
+      "add x28, x28, #0x40\n"
       ".inst 0x6f82e0d9  // udot v25.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x6f83e0dd  // udot v29.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x6f80e0b2  // udot v18.4s, v5.16b, v0.4b[0]\n"
@@ -1690,18 +1689,18 @@ void a64_hybrid_u8qa_dot_4x16 (
       "cmp x26, x20\n"
       "bne 94b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20\n"
+      "add x21, x22, x20\n"
       "prfm pstl1keep, [x23, #0x0]\n"
-      "add x22, x23, x20\n"
       "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "tbnz %x[flags], #31, 109f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
       "addp v13.4s, v13.4s, v13.4s\n"
       "addp v14.4s, v14.4s, v14.4s\n"
       "neg v0.4s, v0.4s\n"
@@ -1725,10 +1724,10 @@ void a64_hybrid_u8qa_dot_4x16 (
       "add v20.4s, v20.4s, v12.4s\n"
       "add v21.4s, v21.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v1.4s }, [x20]\n"
       "add v22.4s, v22.4s, v12.4s\n"
       "add v23.4s, v23.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v24.4s, v24.4s, v13.4s\n"
       "add v25.4s, v25.4s, v13.4s\n"
@@ -1775,32 +1774,32 @@ void a64_hybrid_u8qa_dot_4x16 (
       "tbz %x[flags], #5, 110f\n"
       "and v2.16b, v16.16b, v0.16b\n"
       "and v1.16b, v17.16b, v0.16b\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v16.4s, v16.4s, v2.4s\n"
+      "sqadd v17.4s, v17.4s, v1.4s\n"
       "and v7.16b, v18.16b, v0.16b\n"
       "and v6.16b, v19.16b, v0.16b\n"
       "and v5.16b, v20.16b, v0.16b\n"
       "and v4.16b, v21.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
       "and v3.16b, v22.16b, v0.16b\n"
+      "and v2.16b, v23.16b, v0.16b\n"
+      "and v1.16b, v24.16b, v0.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
       "sshr v6.4s, v6.4s, #0x1f\n"
       "sshr v5.4s, v5.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v2.4s\n"
-      "sqadd v17.4s, v17.4s, v1.4s\n"
-      "and v2.16b, v23.16b, v0.16b\n"
-      "and v1.16b, v24.16b, v0.16b\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sshr v3.4s, v3.4s, #0x1f\n"
-      "sqadd v18.4s, v18.4s, v7.4s\n"
-      "sqadd v19.4s, v19.4s, v6.4s\n"
       "sshr v2.4s, v2.4s, #0x1f\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v18.4s, v18.4s, v7.4s\n"
+      "sqadd v19.4s, v19.4s, v6.4s\n"
       "sqadd v20.4s, v20.4s, v5.4s\n"
       "sqadd v21.4s, v21.4s, v4.4s\n"
       "sqadd v22.4s, v22.4s, v3.4s\n"
-      "and v7.16b, v25.16b, v0.16b\n"
       "sqadd v23.4s, v23.4s, v2.4s\n"
       "sqadd v24.4s, v24.4s, v1.4s\n"
+      "and v7.16b, v25.16b, v0.16b\n"
       "and v6.16b, v26.16b, v0.16b\n"
       "and v5.16b, v27.16b, v0.16b\n"
       "and v4.16b, v28.16b, v0.16b\n"
@@ -1822,21 +1821,21 @@ void a64_hybrid_u8qa_dot_4x16 (
       "sqadd v30.4s, v30.4s, v2.4s\n"
       "sqadd v31.4s, v31.4s, v1.4s\n"
       "110:"  // Height 4: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v3.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v3.4s }, [x21]\n"
-      "ld1r { v2.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v2.4s }, [x20]\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v1.4s }, [x20]\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "srshl v24.4s, v24.4s, v0.4s\n"
       "srshl v25.4s, v25.4s, v0.4s\n"
       "srshl v26.4s, v26.4s, v0.4s\n"
@@ -1908,100 +1907,100 @@ void a64_hybrid_u8qa_dot_4x16 (
       "bge 119f\n"
       "tbz x9, #3, 114f\n"
       "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
+      "str d20, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
+      "str d28, [x21], #0x8\n"
       "tbz x9, #2, 112f\n"
       "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
+      "st1 { v20.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v28.s }[2], [x21], #0x4\n"
       "tbz x9, #1, 111f\n"
       "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "st1 { v28.h }[6], [x22], #0x2\n"
+      "st1 { v20.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
+      "st1 { v28.h }[6], [x21], #0x2\n"
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "st1 { v28.b }[14], [x22]\n"
+      "st1 { v20.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
+      "st1 { v28.b }[14], [x21]\n"
       "b 118f\n"
       "111:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "st1 { v28.b }[12], [x22]\n"
+      "st1 { v20.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
+      "st1 { v28.b }[12], [x21]\n"
       "b 118f\n"
       "112:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x9, #1, 113f\n"
       "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "st1 { v28.h }[4], [x22], #0x2\n"
+      "st1 { v20.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
+      "st1 { v28.h }[4], [x21], #0x2\n"
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "st1 { v28.b }[10], [x22]\n"
+      "st1 { v20.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
+      "st1 { v28.b }[10], [x21]\n"
       "b 118f\n"
       "113:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "st1 { v28.b }[8], [x22]\n"
+      "st1 { v20.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
+      "st1 { v28.b }[8], [x21]\n"
       "b 118f\n"
       "114:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x9, #2, 116f\n"
       "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
+      "str s20, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
+      "str s28, [x21], #0x4\n"
       "tbz x9, #1, 115f\n"
       "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "st1 { v28.h }[2], [x22], #0x2\n"
+      "st1 { v20.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
+      "st1 { v28.h }[2], [x21], #0x2\n"
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "st1 { v28.b }[6], [x22]\n"
+      "st1 { v20.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
+      "st1 { v28.b }[6], [x21]\n"
       "b 118f\n"
       "115:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "st1 { v28.b }[4], [x22]\n"
+      "st1 { v20.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
+      "st1 { v28.b }[4], [x21]\n"
       "b 118f\n"
       "116:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x9, #1, 117f\n"
       "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "str h28, [x22], #0x2\n"
+      "str h20, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
+      "str h28, [x21], #0x2\n"
       "tbz x9, #0, 118f\n"
       "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "st1 { v28.b }[2], [x22]\n"
+      "st1 { v20.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
+      "st1 { v28.b }[2], [x21]\n"
       "b 118f\n"
       "117:"  // Height 4: Partial direct writeback: partial_1_0
       "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "str b28, [x22, #0x0]\n"
+      "str b20, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
+      "str b28, [x21, #0x0]\n"
       "118:"  // Height 4: Partial direct writeback: Done
       "b 120f\n"
       "119:"  // Height 4: Full writeback
       "str q16, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q28, [x22, #0x0]\n"
+      "str q20, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
+      "str q28, [x21, #0x0]\n"
       "120:"  // Height 4: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 92b\n"
@@ -2017,8 +2016,8 @@ void a64_hybrid_u8qa_dot_4x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "122:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16.hpp
index 84f6ed0553..f8c7f0a549 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 16, 8> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 4, 16, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp
index 2d3af7f9c3..9103ef59ce 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp
@@ -45,18 +45,18 @@ void a64_hybrid_u8qa_mmla_4x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -90,7 +90,7 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -104,8 +104,8 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "mov x26, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -130,7 +130,6 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "ldr q4, [x28, #0x60]\n"
       "blt 9f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
-      "add x24, x24, #0x10\n"
       "trn1 v0.2d, v1.2d, v27.2d\n"
       ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
       "ldr q25, [x28, #0x70]\n"
@@ -152,8 +151,9 @@ void a64_hybrid_u8qa_mmla_4x16 (
       ".inst 0x6e98a430  // ummla v16.4s, v1.16b, v24.16b\n"
       "ldr q24, [x28, #0xf0]\n"
       ".inst 0x6e9ea434  // ummla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e9da431  // ummla v17.4s, v1.16b, v29.16b\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x6e9ca435  // ummla v21.4s, v1.16b, v28.16b\n"
       ".inst 0x6e9ba432  // ummla v18.4s, v1.16b, v27.16b\n"
       ".inst 0x6e9aa436  // ummla v22.4s, v1.16b, v26.16b\n"
@@ -166,9 +166,9 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "ldr q1, [x24, #0x0]\n"
       "ldr q5, [x28, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q6, [x28, #0x10]\n"
       "ldr q7, [x28, #0x20]\n"
-      "cmp x25, #0x20\n"
       "ldr q8, [x28, #0x30]\n"
       "ldr q9, [x28, #0x40]\n"
       "ldr q10, [x28, #0x50]\n"
@@ -176,12 +176,10 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "prfm pldl1keep, [x24, #0x80]\n"
       "bge 7b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
       "trn1 v0.2d, v1.2d, v24.2d\n"
-      "trn2 v1.2d, v1.2d, v24.2d\n"
       ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
       "ldr q25, [x28, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v24.2d\n"
       ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
       "ldr q24, [x28, #0x80]\n"
       ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
@@ -198,9 +196,11 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "ldr q25, [x28, #0xe0]\n"
       ".inst 0x6e98a430  // ummla v16.4s, v1.16b, v24.16b\n"
       "ldr q24, [x28, #0xf0]\n"
+      "sub x25, x25, #0x10\n"
       ".inst 0x6e9ea434  // ummla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
       ".inst 0x6e9da431  // ummla v17.4s, v1.16b, v29.16b\n"
+      "add x24, x24, #0x10\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x6e9ca435  // ummla v21.4s, v1.16b, v28.16b\n"
       ".inst 0x6e9ba432  // ummla v18.4s, v1.16b, v27.16b\n"
       ".inst 0x6e9aa436  // ummla v22.4s, v1.16b, v26.16b\n"
@@ -222,24 +222,24 @@ void a64_hybrid_u8qa_mmla_4x16 (
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "13:"  // Height 1: Multiply loop: unique 3: skip row sum
       "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
+      "ldr q26, [x28, #0x10]\n"
+      ".inst 0x6e98a410  // ummla v16.4s, v0.16b, v24.16b\n"
       "sub x25, x25, #0x8\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
+      "ldr q25, [x28, #0x20]\n"
+      "ldr q24, [x28, #0x30]\n"
       "cmp x25, #0x8\n"
+      ".inst 0x6e9aa414  // ummla v20.4s, v0.16b, v26.16b\n"
       "ldr q27, [x28, #0x40]\n"
       "ldr q26, [x28, #0x50]\n"
+      ".inst 0x6e99a411  // ummla v17.4s, v0.16b, v25.16b\n"
+      ".inst 0x6e98a415  // ummla v21.4s, v0.16b, v24.16b\n"
       "ldr q25, [x28, #0x60]\n"
-      ".inst 0x6e98a410  // ummla v16.4s, v0.16b, v24.16b\n"
       "ldr q24, [x28, #0x70]\n"
-      ".inst 0x6e9ea414  // ummla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x6e9da411  // ummla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x6e9ca415  // ummla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
       ".inst 0x6e9ba412  // ummla v18.4s, v0.16b, v27.16b\n"
       ".inst 0x6e9aa416  // ummla v22.4s, v0.16b, v26.16b\n"
       ".inst 0x6e99a413  // ummla v19.4s, v0.16b, v25.16b\n"
       ".inst 0x6e98a417  // ummla v23.4s, v0.16b, v24.16b\n"
+      "add x28, x28, #0x80\n"
       "bge 12b\n"
       "14:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x25, 20f\n"
@@ -267,23 +267,23 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "tbnz %x[flags], #31, 19f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "19:"  // Height 1: Multiply loop: unique 4: skip row sum
-      "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "ldr q27, [x28, #0x40]\n"
-      "ldr q26, [x28, #0x50]\n"
+      "ldr q25, [x28, #0x0]\n"
+      "ldr q24, [x28, #0x10]\n"
+      ".inst 0x6e99a410  // ummla v16.4s, v0.16b, v25.16b\n"
+      ".inst 0x6e98a414  // ummla v20.4s, v0.16b, v24.16b\n"
+      "ldr q25, [x28, #0x20]\n"
+      "ldr q24, [x28, #0x30]\n"
+      ".inst 0x6e99a411  // ummla v17.4s, v0.16b, v25.16b\n"
+      ".inst 0x6e98a415  // ummla v21.4s, v0.16b, v24.16b\n"
+      "ldr q25, [x28, #0x40]\n"
+      "ldr q24, [x28, #0x50]\n"
+      ".inst 0x6e99a412  // ummla v18.4s, v0.16b, v25.16b\n"
+      ".inst 0x6e98a416  // ummla v22.4s, v0.16b, v24.16b\n"
       "ldr q25, [x28, #0x60]\n"
-      ".inst 0x6e98a410  // ummla v16.4s, v0.16b, v24.16b\n"
       "ldr q24, [x28, #0x70]\n"
-      ".inst 0x6e9ea414  // ummla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x6e9da411  // ummla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x6e9ca415  // ummla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x6e9ba412  // ummla v18.4s, v0.16b, v27.16b\n"
-      ".inst 0x6e9aa416  // ummla v22.4s, v0.16b, v26.16b\n"
       ".inst 0x6e99a413  // ummla v19.4s, v0.16b, v25.16b\n"
       ".inst 0x6e98a417  // ummla v23.4s, v0.16b, v24.16b\n"
+      "add x28, x28, #0x80\n"
       "20:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -297,8 +297,8 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "mov v23.16b, v16.16b\n"
       "tbnz %x[flags], #31, 21f\n"
       "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
       "ld1r { v16.4s }, [x20]\n"
+      "addp v11.4s, v11.4s, v11.4s\n"
       "neg v16.4s, v16.4s\n"
       "dup v11.4s, v11.s[0]\n"
       "mul v11.4s, v11.4s, v16.4s\n"
@@ -312,16 +312,16 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "add v18.4s, v18.4s, v11.4s\n"
       "add v19.4s, v19.4s, v11.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v16.4s }, [x20]\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add v23.4s, v23.4s, v24.4s\n"
       "add v17.4s, v17.4s, v22.4s\n"
+      "add v18.4s, v18.4s, v21.4s\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add x10, x10, #0x40\n"
       "ld1r { v0.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v21.4s\n"
       "add v19.4s, v19.4s, v20.4s\n"
       "sqrdmulh v23.4s, v23.4s, v16.4s\n"
+      "add x10, x10, #0x40\n"
       "sqrdmulh v17.4s, v17.4s, v16.4s\n"
       "sqrdmulh v18.4s, v18.4s, v16.4s\n"
       "sqrdmulh v19.4s, v19.4s, v16.4s\n"
@@ -339,21 +339,21 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "sqadd v18.4s, v18.4s, v20.4s\n"
       "sqadd v19.4s, v19.4s, v16.4s\n"
       "22:"  // Height 1: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v21.4s }, [x20]\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v21.4s }, [x21]\n"
-      "ld1r { v20.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v16.4s }, [x20]\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v20.4s }, [x20]\n"
       "add v23.4s, v23.4s, v21.4s\n"
       "add v17.4s, v17.4s, v21.4s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1r { v16.4s }, [x20]\n"
       "add v18.4s, v18.4s, v21.4s\n"
       "add v19.4s, v19.4s, v21.4s\n"
+      "cmp x9, #0x10\n"
       "smin v23.4s, v23.4s, v20.4s\n"
       "smin v17.4s, v17.4s, v20.4s\n"
       "smin v18.4s, v18.4s, v20.4s\n"
@@ -426,7 +426,7 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "movi v15.16b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "34:"  // Height 2: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -440,8 +440,8 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "mov x26, #0x0\n"
       "36:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 37f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -471,13 +471,11 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "blt 41f\n"
       "39:"  // Height 2: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
       "ldr q25, [x28, #0x70]\n"
       ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
       "ldr q24, [x28, #0x80]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
       "ldr q30, [x28, #0x90]\n"
       ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
@@ -493,9 +491,11 @@ void a64_hybrid_u8qa_mmla_4x16 (
       ".inst 0x6e98a430  // ummla v16.4s, v1.16b, v24.16b\n"
       "ldr q24, [x28, #0xf0]\n"
       ".inst 0x6e9ea434  // ummla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e9da431  // ummla v17.4s, v1.16b, v29.16b\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e9ca435  // ummla v21.4s, v1.16b, v28.16b\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x6e9ba432  // ummla v18.4s, v1.16b, v27.16b\n"
       ".inst 0x6e9aa436  // ummla v22.4s, v1.16b, v26.16b\n"
       ".inst 0x6e99a433  // ummla v19.4s, v1.16b, v25.16b\n"
@@ -507,9 +507,9 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "ldr q1, [x24, #0x0]\n"
       "ldr q2, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q5, [x28, #0x0]\n"
       "ldr q6, [x28, #0x10]\n"
-      "cmp x25, #0x20\n"
       "ldr q7, [x28, #0x20]\n"
       "ldr q8, [x28, #0x30]\n"
       "ldr q9, [x28, #0x40]\n"
@@ -520,14 +520,11 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "bge 39b\n"
       "41:"  // Height 2: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
       "ldr q25, [x28, #0x70]\n"
       ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
       "ldr q24, [x28, #0x80]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
       "ldr q30, [x28, #0x90]\n"
       ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
@@ -542,11 +539,14 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "ldr q25, [x28, #0xe0]\n"
       ".inst 0x6e98a430  // ummla v16.4s, v1.16b, v24.16b\n"
       "ldr q24, [x28, #0xf0]\n"
+      "sub x25, x25, #0x10\n"
       ".inst 0x6e9ea434  // ummla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
       ".inst 0x6e9da431  // ummla v17.4s, v1.16b, v29.16b\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e9ca435  // ummla v21.4s, v1.16b, v28.16b\n"
       ".inst 0x6e9ba432  // ummla v18.4s, v1.16b, v27.16b\n"
+      "add x28, x28, #0x100\n"
       ".inst 0x6e9aa436  // ummla v22.4s, v1.16b, v26.16b\n"
       ".inst 0x6e99a433  // ummla v19.4s, v1.16b, v25.16b\n"
       ".inst 0x6e98a437  // ummla v23.4s, v1.16b, v24.16b\n"
@@ -568,24 +568,24 @@ void a64_hybrid_u8qa_mmla_4x16 (
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "45:"  // Height 2: Multiply loop: unique 7: skip row sum
       "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
+      "ldr q26, [x28, #0x10]\n"
+      ".inst 0x6e98a410  // ummla v16.4s, v0.16b, v24.16b\n"
       "sub x25, x25, #0x8\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
+      "ldr q25, [x28, #0x20]\n"
+      "ldr q24, [x28, #0x30]\n"
       "cmp x25, #0x8\n"
+      ".inst 0x6e9aa414  // ummla v20.4s, v0.16b, v26.16b\n"
       "ldr q27, [x28, #0x40]\n"
       "ldr q26, [x28, #0x50]\n"
+      ".inst 0x6e99a411  // ummla v17.4s, v0.16b, v25.16b\n"
+      ".inst 0x6e98a415  // ummla v21.4s, v0.16b, v24.16b\n"
       "ldr q25, [x28, #0x60]\n"
-      ".inst 0x6e98a410  // ummla v16.4s, v0.16b, v24.16b\n"
       "ldr q24, [x28, #0x70]\n"
-      ".inst 0x6e9ea414  // ummla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x6e9da411  // ummla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x6e9ca415  // ummla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
       ".inst 0x6e9ba412  // ummla v18.4s, v0.16b, v27.16b\n"
       ".inst 0x6e9aa416  // ummla v22.4s, v0.16b, v26.16b\n"
       ".inst 0x6e99a413  // ummla v19.4s, v0.16b, v25.16b\n"
       ".inst 0x6e98a417  // ummla v23.4s, v0.16b, v24.16b\n"
+      "add x28, x28, #0x80\n"
       "bge 44b\n"
       "46:"  // Height 2: Multiply loop: Skip odd blocks
       "cbz x25, 52f\n"
@@ -620,23 +620,23 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "tbnz %x[flags], #31, 51f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       "51:"  // Height 2: Multiply loop: unique 8: skip row sum
-      "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "ldr q27, [x28, #0x40]\n"
-      "ldr q26, [x28, #0x50]\n"
+      "ldr q25, [x28, #0x0]\n"
+      "ldr q24, [x28, #0x10]\n"
+      ".inst 0x6e99a410  // ummla v16.4s, v0.16b, v25.16b\n"
+      ".inst 0x6e98a414  // ummla v20.4s, v0.16b, v24.16b\n"
+      "ldr q25, [x28, #0x20]\n"
+      "ldr q24, [x28, #0x30]\n"
+      ".inst 0x6e99a411  // ummla v17.4s, v0.16b, v25.16b\n"
+      ".inst 0x6e98a415  // ummla v21.4s, v0.16b, v24.16b\n"
+      "ldr q25, [x28, #0x40]\n"
+      "ldr q24, [x28, #0x50]\n"
+      ".inst 0x6e99a412  // ummla v18.4s, v0.16b, v25.16b\n"
+      ".inst 0x6e98a416  // ummla v22.4s, v0.16b, v24.16b\n"
       "ldr q25, [x28, #0x60]\n"
-      ".inst 0x6e98a410  // ummla v16.4s, v0.16b, v24.16b\n"
       "ldr q24, [x28, #0x70]\n"
-      ".inst 0x6e9ea414  // ummla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x6e9da411  // ummla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x6e9ca415  // ummla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x6e9ba412  // ummla v18.4s, v0.16b, v27.16b\n"
-      ".inst 0x6e9aa416  // ummla v22.4s, v0.16b, v26.16b\n"
       ".inst 0x6e99a413  // ummla v19.4s, v0.16b, v25.16b\n"
       ".inst 0x6e98a417  // ummla v23.4s, v0.16b, v24.16b\n"
+      "add x28, x28, #0x80\n"
       "52:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -644,21 +644,21 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "bne 36b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v24.2d, v16.2d, v20.2d\n"
+      "add x23, x27, x20\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "mov v23.16b, v24.16b\n"
       "tbnz %x[flags], #31, 53f\n"
       "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
       "ld1r { v24.4s }, [x20]\n"
+      "addp v11.4s, v11.4s, v11.4s\n"
       "neg v24.4s, v24.4s\n"
       "dup v12.4s, v11.s[3]\n"
       "dup v11.4s, v11.s[0]\n"
@@ -676,10 +676,10 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "add v16.4s, v16.4s, v12.4s\n"
       "add v17.4s, v17.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v24.4s }, [x20]\n"
       "add v18.4s, v18.4s, v12.4s\n"
       "add v19.4s, v19.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v23.4s, v23.4s, v28.4s\n"
       "add v20.4s, v20.4s, v27.4s\n"
@@ -701,45 +701,45 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "sqrdmulh v19.4s, v19.4s, v24.4s\n"
       "tbz %x[flags], #5, 54f\n"
       "and v24.16b, v23.16b, v0.16b\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
+      "sqadd v23.4s, v23.4s, v24.4s\n"
       "and v30.16b, v20.16b, v0.16b\n"
       "and v29.16b, v21.16b, v0.16b\n"
       "and v28.16b, v22.16b, v0.16b\n"
       "and v27.16b, v16.16b, v0.16b\n"
       "and v26.16b, v17.16b, v0.16b\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "and v25.16b, v18.16b, v0.16b\n"
+      "and v24.16b, v19.16b, v0.16b\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v27.4s, v27.4s, #0x1f\n"
-      "sqadd v23.4s, v23.4s, v24.4s\n"
-      "and v24.16b, v19.16b, v0.16b\n"
       "sshr v26.4s, v26.4s, #0x1f\n"
       "sshr v25.4s, v25.4s, #0x1f\n"
+      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v20.4s, v20.4s, v30.4s\n"
       "sqadd v21.4s, v21.4s, v29.4s\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
       "sqadd v22.4s, v22.4s, v28.4s\n"
       "sqadd v16.4s, v16.4s, v27.4s\n"
       "sqadd v17.4s, v17.4s, v26.4s\n"
       "sqadd v18.4s, v18.4s, v25.4s\n"
       "sqadd v19.4s, v19.4s, v24.4s\n"
       "54:"  // Height 2: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v26.4s }, [x20]\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x21]\n"
-      "ld1r { v25.4s }, [x20]\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v25.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v24.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "add v23.4s, v23.4s, v26.4s\n"
       "add v20.4s, v20.4s, v26.4s\n"
       "add v21.4s, v21.4s, v26.4s\n"
@@ -773,68 +773,68 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "bge 63f\n"
       "tbz x9, #3, 58f\n"
       "str d23, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
+      "str d16, [x23], #0x8\n"
       "tbz x9, #2, 56f\n"
       "st1 { v23.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
       "tbz x9, #1, 55f\n"
       "st1 { v23.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
+      "st1 { v16.b }[14], [x23]\n"
       "b 62f\n"
       "55:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
+      "st1 { v16.b }[12], [x23]\n"
       "b 62f\n"
       "56:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x9, #1, 57f\n"
       "st1 { v23.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
+      "st1 { v16.b }[10], [x23]\n"
       "b 62f\n"
       "57:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
+      "st1 { v16.b }[8], [x23]\n"
       "b 62f\n"
       "58:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x9, #2, 60f\n"
       "str s23, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
+      "str s16, [x23], #0x4\n"
       "tbz x9, #1, 59f\n"
       "st1 { v23.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
+      "st1 { v16.b }[6], [x23]\n"
       "b 62f\n"
       "59:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
+      "st1 { v16.b }[4], [x23]\n"
       "b 62f\n"
       "60:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x9, #1, 61f\n"
       "str h23, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
+      "str h16, [x23], #0x2\n"
       "tbz x9, #0, 62f\n"
       "st1 { v23.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
+      "st1 { v16.b }[2], [x23]\n"
       "b 62f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_0
       "str b23, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
       "62:"  // Height 2: Partial direct writeback: Done
       "b 64f\n"
       "63:"  // Height 2: Full writeback
       "str q23, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
       "64:"  // Height 2: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 34b\n"
@@ -848,7 +848,7 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "movi v15.16b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "66:"  // Height 3: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -870,8 +870,8 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "mov x26, #0x0\n"
       "68:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 69f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -905,35 +905,35 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "71:"  // Height 3: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "ldr q14, [x28, #0x60]\n"
       ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
+      "trn1 v2.2d, v3.2d, v4.2d\n"
       ".inst 0x6e85a458  // ummla v24.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0x70]\n"
+      "ldr q14, [x28, #0x70]\n"
+      ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
+      "ldr q5, [x28, #0x60]\n"
       ".inst 0x6e86a45c  // ummla v28.4s, v2.16b, v6.16b\n"
       "ldr q4, [x28, #0x80]\n"
-      ".inst 0x6e89a412  // ummla v18.4s, v0.16b, v9.16b\n"
+      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x6e87a459  // ummla v25.4s, v2.16b, v7.16b\n"
       "ldr q7, [x28, #0x90]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x6e88a45d  // ummla v29.4s, v2.16b, v8.16b\n"
       "ldr q8, [x28, #0xa0]\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6e89a412  // ummla v18.4s, v0.16b, v9.16b\n"
       ".inst 0x6e89a45a  // ummla v26.4s, v2.16b, v9.16b\n"
       "ldr q9, [x28, #0xb0]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e8aa416  // ummla v22.4s, v0.16b, v10.16b\n"
       ".inst 0x6e8aa45e  // ummla v30.4s, v2.16b, v10.16b\n"
       "ldr q10, [x28, #0xc0]\n"
-      ".inst 0x6e8ea413  // ummla v19.4s, v0.16b, v14.16b\n"
-      ".inst 0x6e8ea45b  // ummla v27.4s, v2.16b, v14.16b\n"
+      ".inst 0x6e85a413  // ummla v19.4s, v0.16b, v5.16b\n"
+      ".inst 0x6e85a45b  // ummla v27.4s, v2.16b, v5.16b\n"
       "ldr q6, [x28, #0xd0]\n"
-      ".inst 0x6e85a417  // ummla v23.4s, v0.16b, v5.16b\n"
-      ".inst 0x6e85a45f  // ummla v31.4s, v2.16b, v5.16b\n"
+      ".inst 0x6e8ea417  // ummla v23.4s, v0.16b, v14.16b\n"
+      ".inst 0x6e8ea45f  // ummla v31.4s, v2.16b, v14.16b\n"
       "ldr q5, [x28, #0xe0]\n"
       ".inst 0x6e84a430  // ummla v16.4s, v1.16b, v4.16b\n"
       ".inst 0x6e84a478  // ummla v24.4s, v3.16b, v4.16b\n"
@@ -962,9 +962,9 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "ldr q1, [x24, #0x0]\n"
       "ldr q2, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q3, [x22, #0x0]\n"
       "ldr q5, [x28, #0x0]\n"
-      "cmp x25, #0x20\n"
       "ldr q6, [x28, #0x10]\n"
       "ldr q7, [x28, #0x20]\n"
       "ldr q8, [x28, #0x30]\n"
@@ -977,36 +977,36 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "73:"  // Height 3: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "ldr q14, [x28, #0x60]\n"
       ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
+      "trn1 v2.2d, v3.2d, v4.2d\n"
       ".inst 0x6e85a458  // ummla v24.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0x70]\n"
+      "ldr q14, [x28, #0x70]\n"
+      ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
+      "ldr q5, [x28, #0x60]\n"
       ".inst 0x6e86a45c  // ummla v28.4s, v2.16b, v6.16b\n"
       "ldr q4, [x28, #0x80]\n"
-      ".inst 0x6e89a412  // ummla v18.4s, v0.16b, v9.16b\n"
+      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x6e87a459  // ummla v25.4s, v2.16b, v7.16b\n"
       "ldr q7, [x28, #0x90]\n"
+      "sub x25, x25, #0x10\n"
+      ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x6e88a45d  // ummla v29.4s, v2.16b, v8.16b\n"
       "ldr q8, [x28, #0xa0]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6e89a412  // ummla v18.4s, v0.16b, v9.16b\n"
       ".inst 0x6e89a45a  // ummla v26.4s, v2.16b, v9.16b\n"
       "ldr q9, [x28, #0xb0]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e8aa416  // ummla v22.4s, v0.16b, v10.16b\n"
       ".inst 0x6e8aa45e  // ummla v30.4s, v2.16b, v10.16b\n"
       "ldr q10, [x28, #0xc0]\n"
-      ".inst 0x6e8ea413  // ummla v19.4s, v0.16b, v14.16b\n"
-      ".inst 0x6e8ea45b  // ummla v27.4s, v2.16b, v14.16b\n"
+      "add x22, x22, #0x10\n"
+      ".inst 0x6e85a413  // ummla v19.4s, v0.16b, v5.16b\n"
+      ".inst 0x6e85a45b  // ummla v27.4s, v2.16b, v5.16b\n"
       "ldr q6, [x28, #0xd0]\n"
-      ".inst 0x6e85a417  // ummla v23.4s, v0.16b, v5.16b\n"
-      ".inst 0x6e85a45f  // ummla v31.4s, v2.16b, v5.16b\n"
+      ".inst 0x6e8ea417  // ummla v23.4s, v0.16b, v14.16b\n"
+      ".inst 0x6e8ea45f  // ummla v31.4s, v2.16b, v14.16b\n"
       "ldr q5, [x28, #0xe0]\n"
       ".inst 0x6e84a430  // ummla v16.4s, v1.16b, v4.16b\n"
       ".inst 0x6e84a478  // ummla v24.4s, v3.16b, v4.16b\n"
@@ -1040,34 +1040,34 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "cmp x25, #0x8\n"
       "blt 78f\n"
       "76:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d3, [x24], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
       "ldr d0, [x23], #0x8\n"
+      "trn1 v0.2d, v1.2d, v0.2d\n"
       "ldr d1, [x22], #0x8\n"
-      "trn1 v0.2d, v3.2d, v0.2d\n"
       "trn1 v2.2d, v1.2d, v2.2d\n"
       "tbnz %x[flags], #31, 77f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       "77:"  // Height 3: Multiply loop: unique 11: skip row sum
-      "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "sub x25, x25, #0x8\n"
+      "ldr q3, [x28, #0x0]\n"
+      "ldr q1, [x28, #0x10]\n"
+      ".inst 0x6e83a410  // ummla v16.4s, v0.16b, v3.16b\n"
+      ".inst 0x6e83a458  // ummla v24.4s, v2.16b, v3.16b\n"
       "ldr q7, [x28, #0x20]\n"
       "ldr q6, [x28, #0x30]\n"
+      "sub x25, x25, #0x8\n"
       "cmp x25, #0x8\n"
       "ldr q5, [x28, #0x40]\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x6e81a414  // ummla v20.4s, v0.16b, v1.16b\n"
+      ".inst 0x6e81a45c  // ummla v28.4s, v2.16b, v1.16b\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x6e81a410  // ummla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x6e81a458  // ummla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x28, #0x70]\n"
-      ".inst 0x6e88a414  // ummla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x6e88a45c  // ummla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
       ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x6e87a459  // ummla v25.4s, v2.16b, v7.16b\n"
       ".inst 0x6e86a415  // ummla v21.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a45d  // ummla v29.4s, v2.16b, v6.16b\n"
+      "add x28, x28, #0x80\n"
       ".inst 0x6e85a412  // ummla v18.4s, v0.16b, v5.16b\n"
       ".inst 0x6e85a45a  // ummla v26.4s, v2.16b, v5.16b\n"
       ".inst 0x6e84a416  // ummla v22.4s, v0.16b, v4.16b\n"
@@ -1120,24 +1120,24 @@ void a64_hybrid_u8qa_mmla_4x16 (
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       "83:"  // Height 3: Multiply loop: unique 12: skip row sum
       "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
+      "ldr q3, [x28, #0x10]\n"
+      ".inst 0x6e81a410  // ummla v16.4s, v0.16b, v1.16b\n"
+      ".inst 0x6e81a458  // ummla v24.4s, v2.16b, v1.16b\n"
+      "ldr q1, [x28, #0x20]\n"
       "ldr q6, [x28, #0x30]\n"
+      ".inst 0x6e83a414  // ummla v20.4s, v0.16b, v3.16b\n"
+      ".inst 0x6e83a45c  // ummla v28.4s, v2.16b, v3.16b\n"
       "ldr q5, [x28, #0x40]\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x6e81a411  // ummla v17.4s, v0.16b, v1.16b\n"
+      ".inst 0x6e81a459  // ummla v25.4s, v2.16b, v1.16b\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x6e81a410  // ummla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x6e81a458  // ummla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x28, #0x70]\n"
-      ".inst 0x6e88a414  // ummla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x6e88a45c  // ummla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x6e87a459  // ummla v25.4s, v2.16b, v7.16b\n"
       ".inst 0x6e86a415  // ummla v21.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a45d  // ummla v29.4s, v2.16b, v6.16b\n"
       ".inst 0x6e85a412  // ummla v18.4s, v0.16b, v5.16b\n"
       ".inst 0x6e85a45a  // ummla v26.4s, v2.16b, v5.16b\n"
+      "add x28, x28, #0x80\n"
       ".inst 0x6e84a416  // ummla v22.4s, v0.16b, v4.16b\n"
       ".inst 0x6e84a45e  // ummla v30.4s, v2.16b, v4.16b\n"
       ".inst 0x6e83a413  // ummla v19.4s, v0.16b, v3.16b\n"
@@ -1151,18 +1151,18 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "bne 68b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v0.2d, v16.2d, v20.2d\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v24.2d, v24.2d, v28.2d\n"
       "uzp1 v25.2d, v25.2d, v29.2d\n"
       "uzp1 v26.2d, v26.2d, v30.2d\n"
@@ -1170,9 +1170,9 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "mov v31.16b, v0.16b\n"
       "tbnz %x[flags], #31, 85f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v23.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v13.4s, v13.4s, v13.4s\n"
-      "ld1r { v23.4s }, [x20]\n"
       "neg v23.4s, v23.4s\n"
       "dup v12.4s, v11.s[3]\n"
       "dup v11.4s, v11.s[0]\n"
@@ -1192,10 +1192,10 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "add v16.4s, v16.4s, v12.4s\n"
       "add v17.4s, v17.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v23.4s }, [x20]\n"
       "add v18.4s, v18.4s, v12.4s\n"
       "add v19.4s, v19.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v24.4s, v24.4s, v13.4s\n"
       "add v25.4s, v25.4s, v13.4s\n"
@@ -1233,18 +1233,18 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "and v29.16b, v21.16b, v0.16b\n"
       "and v28.16b, v22.16b, v0.16b\n"
       "and v23.16b, v16.16b, v0.16b\n"
-      "and v3.16b, v17.16b, v0.16b\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
       "sshr v30.4s, v30.4s, #0x1f\n"
       "sshr v29.4s, v29.4s, #0x1f\n"
       "sshr v28.4s, v28.4s, #0x1f\n"
       "sshr v23.4s, v23.4s, #0x1f\n"
-      "and v2.16b, v18.16b, v0.16b\n"
       "sqadd v31.4s, v31.4s, v1.4s\n"
       "sqadd v20.4s, v20.4s, v30.4s\n"
       "sqadd v21.4s, v21.4s, v29.4s\n"
       "sqadd v22.4s, v22.4s, v28.4s\n"
       "sqadd v16.4s, v16.4s, v23.4s\n"
+      "and v3.16b, v17.16b, v0.16b\n"
+      "and v2.16b, v18.16b, v0.16b\n"
       "and v1.16b, v19.16b, v0.16b\n"
       "and v30.16b, v24.16b, v0.16b\n"
       "and v29.16b, v25.16b, v0.16b\n"
@@ -1265,21 +1265,21 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "sqadd v26.4s, v26.4s, v28.4s\n"
       "sqadd v27.4s, v27.4s, v23.4s\n"
       "86:"  // Height 3: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v29.4s }, [x20]\n"
       "srshl v31.4s, v31.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v29.4s }, [x21]\n"
-      "ld1r { v28.4s }, [x20]\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v28.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v23.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "srshl v24.4s, v24.4s, v0.4s\n"
       "srshl v25.4s, v25.4s, v0.4s\n"
       "srshl v26.4s, v26.4s, v0.4s\n"
@@ -1332,103 +1332,102 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "bge 95f\n"
       "tbz x9, #3, 90f\n"
       "str d31, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d24, [x22], #0x8\n"
       "tbz x9, #2, 88f\n"
       "st1 { v31.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
+      "st1 { v24.s }[2], [x22], #0x4\n"
       "tbz x9, #1, 87f\n"
       "st1 { v31.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
+      "st1 { v24.h }[6], [x22], #0x2\n"
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
+      "st1 { v16.b }[14], [x23]\n"
+      "st1 { v24.b }[14], [x22]\n"
       "b 94f\n"
       "87:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
+      "st1 { v16.b }[12], [x23]\n"
+      "st1 { v24.b }[12], [x22]\n"
       "b 94f\n"
       "88:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x9, #1, 89f\n"
       "st1 { v31.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
+      "st1 { v24.h }[4], [x22], #0x2\n"
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
+      "st1 { v16.b }[10], [x23]\n"
+      "st1 { v24.b }[10], [x22]\n"
       "b 94f\n"
       "89:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
+      "st1 { v16.b }[8], [x23]\n"
+      "st1 { v24.b }[8], [x22]\n"
       "b 94f\n"
       "90:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x9, #2, 92f\n"
       "str s31, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
+      "str s16, [x23], #0x4\n"
+      "str s24, [x22], #0x4\n"
       "tbz x9, #1, 91f\n"
       "st1 { v31.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
+      "st1 { v24.h }[2], [x22], #0x2\n"
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
+      "st1 { v16.b }[6], [x23]\n"
+      "st1 { v24.b }[6], [x22]\n"
       "b 94f\n"
       "91:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
+      "st1 { v16.b }[4], [x23]\n"
+      "st1 { v24.b }[4], [x22]\n"
       "b 94f\n"
       "92:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x9, #1, 93f\n"
       "str h31, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
+      "str h16, [x23], #0x2\n"
+      "str h24, [x22], #0x2\n"
       "tbz x9, #0, 94f\n"
       "st1 { v31.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
+      "st1 { v16.b }[2], [x23]\n"
+      "st1 { v24.b }[2], [x22]\n"
       "b 94f\n"
       "93:"  // Height 3: Partial direct writeback: partial_1_0
       "str b31, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
+      "str b24, [x22, #0x0]\n"
       "94:"  // Height 3: Partial direct writeback: Done
       "b 96f\n"
       "95:"  // Height 3: Full writeback
       "str q31, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q24, [x22, #0x0]\n"
       "96:"  // Height 3: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 66b\n"
       "b 130f\n"
       "97:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x4\n"
       "mov x10, %x[col_bias]\n"
       "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
+      "movi v13.4s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "movi v13.4s, #0x0\n"
       "movi v14.4s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
       "movi v15.16b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "98:"  // Height 4: Column loop
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -1450,8 +1449,8 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "mov x26, #0x0\n"
       "100:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 101f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1489,28 +1488,28 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "103:"  // Height 4: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
       "ldr q4, [x28, #0x60]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x6e85a458  // ummla v24.4s, v2.16b, v5.16b\n"
       "ldr q5, [x28, #0x70]\n"
+      ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a45c  // ummla v28.4s, v2.16b, v6.16b\n"
       "ldr q6, [x28, #0x80]\n"
+      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x6e87a459  // ummla v25.4s, v2.16b, v7.16b\n"
       "ldr q7, [x28, #0x90]\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x6e88a45d  // ummla v29.4s, v2.16b, v8.16b\n"
       "ldr q8, [x28, #0xa0]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e89a412  // ummla v18.4s, v0.16b, v9.16b\n"
       ".inst 0x6e89a45a  // ummla v26.4s, v2.16b, v9.16b\n"
       "ldr q9, [x28, #0xb0]\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x6e8aa416  // ummla v22.4s, v0.16b, v10.16b\n"
       ".inst 0x6e8aa45e  // ummla v30.4s, v2.16b, v10.16b\n"
       "ldr q10, [x28, #0xc0]\n"
@@ -1547,9 +1546,9 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "ldr q1, [x24, #0x0]\n"
       "ldr q2, [x23, #0x0]\n"
       "sub x25, x25, #0x10\n"
+      "cmp x25, #0x20\n"
       "ldr q3, [x22, #0x0]\n"
       "ldr q4, [x21, #0x0]\n"
-      "cmp x25, #0x20\n"
       "ldr q5, [x28, #0x0]\n"
       "ldr q6, [x28, #0x10]\n"
       "ldr q7, [x28, #0x20]\n"
@@ -1564,32 +1563,32 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "105:"  // Height 4: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
       "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
       "ldr q4, [x28, #0x60]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x6e85a410  // ummla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x6e85a458  // ummla v24.4s, v2.16b, v5.16b\n"
       "ldr q5, [x28, #0x70]\n"
+      ".inst 0x6e86a414  // ummla v20.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a45c  // ummla v28.4s, v2.16b, v6.16b\n"
       "ldr q6, [x28, #0x80]\n"
+      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x6e87a459  // ummla v25.4s, v2.16b, v7.16b\n"
       "ldr q7, [x28, #0x90]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6e88a415  // ummla v21.4s, v0.16b, v8.16b\n"
       ".inst 0x6e88a45d  // ummla v29.4s, v2.16b, v8.16b\n"
       "ldr q8, [x28, #0xa0]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e89a412  // ummla v18.4s, v0.16b, v9.16b\n"
       ".inst 0x6e89a45a  // ummla v26.4s, v2.16b, v9.16b\n"
       "ldr q9, [x28, #0xb0]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e8aa416  // ummla v22.4s, v0.16b, v10.16b\n"
       ".inst 0x6e8aa45e  // ummla v30.4s, v2.16b, v10.16b\n"
       "ldr q10, [x28, #0xc0]\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x6e84a413  // ummla v19.4s, v0.16b, v4.16b\n"
       ".inst 0x6e84a45b  // ummla v27.4s, v2.16b, v4.16b\n"
       "ldr q4, [x28, #0xd0]\n"
@@ -1629,35 +1628,35 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "cmp x25, #0x8\n"
       "blt 110f\n"
       "108:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d3, [x24], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
       "ldr d0, [x23], #0x8\n"
+      "trn1 v0.2d, v1.2d, v0.2d\n"
       "ldr d2, [x22], #0x8\n"
       "ldr d1, [x21], #0x8\n"
-      "trn1 v0.2d, v3.2d, v0.2d\n"
       "trn1 v2.2d, v2.2d, v1.2d\n"
       "tbnz %x[flags], #31, 109f\n"
       ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       "109:"  // Height 4: Multiply loop: unique 15: skip row sum
-      "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "sub x25, x25, #0x8\n"
+      "ldr q3, [x28, #0x0]\n"
+      "ldr q1, [x28, #0x10]\n"
+      ".inst 0x6e83a410  // ummla v16.4s, v0.16b, v3.16b\n"
+      ".inst 0x6e83a458  // ummla v24.4s, v2.16b, v3.16b\n"
       "ldr q7, [x28, #0x20]\n"
       "ldr q6, [x28, #0x30]\n"
+      "sub x25, x25, #0x8\n"
       "cmp x25, #0x8\n"
       "ldr q5, [x28, #0x40]\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x6e81a414  // ummla v20.4s, v0.16b, v1.16b\n"
+      ".inst 0x6e81a45c  // ummla v28.4s, v2.16b, v1.16b\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x6e81a410  // ummla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x6e81a458  // ummla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x28, #0x70]\n"
-      ".inst 0x6e88a414  // ummla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x6e88a45c  // ummla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
       ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
       ".inst 0x6e87a459  // ummla v25.4s, v2.16b, v7.16b\n"
       ".inst 0x6e86a415  // ummla v21.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a45d  // ummla v29.4s, v2.16b, v6.16b\n"
+      "add x28, x28, #0x80\n"
       ".inst 0x6e85a412  // ummla v18.4s, v0.16b, v5.16b\n"
       ".inst 0x6e85a45a  // ummla v26.4s, v2.16b, v5.16b\n"
       ".inst 0x6e84a416  // ummla v22.4s, v0.16b, v4.16b\n"
@@ -1717,24 +1716,24 @@ void a64_hybrid_u8qa_mmla_4x16 (
       ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
       "115:"  // Height 4: Multiply loop: unique 16: skip row sum
       "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
+      "ldr q3, [x28, #0x10]\n"
+      ".inst 0x6e81a410  // ummla v16.4s, v0.16b, v1.16b\n"
+      ".inst 0x6e81a458  // ummla v24.4s, v2.16b, v1.16b\n"
+      "ldr q1, [x28, #0x20]\n"
       "ldr q6, [x28, #0x30]\n"
+      ".inst 0x6e83a414  // ummla v20.4s, v0.16b, v3.16b\n"
+      ".inst 0x6e83a45c  // ummla v28.4s, v2.16b, v3.16b\n"
       "ldr q5, [x28, #0x40]\n"
       "ldr q4, [x28, #0x50]\n"
+      ".inst 0x6e81a411  // ummla v17.4s, v0.16b, v1.16b\n"
+      ".inst 0x6e81a459  // ummla v25.4s, v2.16b, v1.16b\n"
       "ldr q3, [x28, #0x60]\n"
-      ".inst 0x6e81a410  // ummla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x6e81a458  // ummla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x28, #0x70]\n"
-      ".inst 0x6e88a414  // ummla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x6e88a45c  // ummla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x6e87a411  // ummla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x6e87a459  // ummla v25.4s, v2.16b, v7.16b\n"
       ".inst 0x6e86a415  // ummla v21.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a45d  // ummla v29.4s, v2.16b, v6.16b\n"
       ".inst 0x6e85a412  // ummla v18.4s, v0.16b, v5.16b\n"
       ".inst 0x6e85a45a  // ummla v26.4s, v2.16b, v5.16b\n"
+      "add x28, x28, #0x80\n"
       ".inst 0x6e84a416  // ummla v22.4s, v0.16b, v4.16b\n"
       ".inst 0x6e84a45e  // ummla v30.4s, v2.16b, v4.16b\n"
       ".inst 0x6e83a413  // ummla v19.4s, v0.16b, v3.16b\n"
@@ -1748,22 +1747,22 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "bne 100b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 v0.2d, v16.2d, v20.2d\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
+      "add x21, x22, x20\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x27, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
-      "add x22, x23, x20\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v23.2d, v24.2d, v28.2d\n"
       "uzp2 v24.2d, v24.2d, v28.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v28.2d, v25.2d, v29.2d\n"
       "uzp2 v25.2d, v25.2d, v29.2d\n"
       "uzp1 v29.2d, v26.2d, v30.2d\n"
@@ -1773,9 +1772,9 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "mov v31.16b, v0.16b\n"
       "tbnz %x[flags], #31, 117f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1r { v0.4s }, [x20]\n"
       "addp v11.4s, v11.4s, v11.4s\n"
       "addp v13.4s, v13.4s, v13.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
       "neg v0.4s, v0.4s\n"
       "dup v12.4s, v11.s[3]\n"
       "dup v11.4s, v11.s[0]\n"
@@ -1783,8 +1782,8 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "dup v13.4s, v13.s[0]\n"
       "mul v11.4s, v11.4s, v0.4s\n"
       "mul v12.4s, v12.4s, v0.4s\n"
-      "mul v14.4s, v14.4s, v0.4s\n"
       "mul v13.4s, v13.4s, v0.4s\n"
+      "mul v14.4s, v14.4s, v0.4s\n"
       "117:"  // Height 4: skip row sum fixup
       "ldr q0, [x10, #0x0]\n"
       "ldr q4, [x10, #0x10]\n"
@@ -1797,10 +1796,10 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "add v16.4s, v16.4s, v12.4s\n"
       "add v17.4s, v17.4s, v12.4s\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
       "ld1r { v1.4s }, [x20]\n"
       "add v18.4s, v18.4s, v12.4s\n"
       "add v19.4s, v19.4s, v12.4s\n"
+      "orr %x[flags], %x[flags], #0x80000000\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
       "add v23.4s, v23.4s, v13.4s\n"
       "add v28.4s, v28.4s, v13.4s\n"
@@ -1847,32 +1846,32 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "tbz %x[flags], #5, 118f\n"
       "and v2.16b, v31.16b, v0.16b\n"
       "and v1.16b, v20.16b, v0.16b\n"
+      "sshr v2.4s, v2.4s, #0x1f\n"
+      "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v31.4s, v31.4s, v2.4s\n"
+      "sqadd v20.4s, v20.4s, v1.4s\n"
       "and v7.16b, v21.16b, v0.16b\n"
       "and v6.16b, v22.16b, v0.16b\n"
       "and v5.16b, v16.16b, v0.16b\n"
       "and v4.16b, v17.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
       "and v3.16b, v18.16b, v0.16b\n"
+      "and v2.16b, v19.16b, v0.16b\n"
+      "and v1.16b, v23.16b, v0.16b\n"
       "sshr v7.4s, v7.4s, #0x1f\n"
       "sshr v6.4s, v6.4s, #0x1f\n"
       "sshr v5.4s, v5.4s, #0x1f\n"
-      "sqadd v31.4s, v31.4s, v2.4s\n"
-      "sqadd v20.4s, v20.4s, v1.4s\n"
-      "and v2.16b, v19.16b, v0.16b\n"
-      "and v1.16b, v23.16b, v0.16b\n"
       "sshr v4.4s, v4.4s, #0x1f\n"
       "sshr v3.4s, v3.4s, #0x1f\n"
-      "sqadd v21.4s, v21.4s, v7.4s\n"
-      "sqadd v22.4s, v22.4s, v6.4s\n"
       "sshr v2.4s, v2.4s, #0x1f\n"
       "sshr v1.4s, v1.4s, #0x1f\n"
+      "sqadd v21.4s, v21.4s, v7.4s\n"
+      "sqadd v22.4s, v22.4s, v6.4s\n"
       "sqadd v16.4s, v16.4s, v5.4s\n"
       "sqadd v17.4s, v17.4s, v4.4s\n"
       "sqadd v18.4s, v18.4s, v3.4s\n"
-      "and v7.16b, v28.16b, v0.16b\n"
       "sqadd v19.4s, v19.4s, v2.4s\n"
       "sqadd v23.4s, v23.4s, v1.4s\n"
+      "and v7.16b, v28.16b, v0.16b\n"
       "and v6.16b, v29.16b, v0.16b\n"
       "and v5.16b, v30.16b, v0.16b\n"
       "and v4.16b, v24.16b, v0.16b\n"
@@ -1894,21 +1893,21 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "sqadd v26.4s, v26.4s, v2.4s\n"
       "sqadd v27.4s, v27.4s, v1.4s\n"
       "118:"  // Height 4: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "ld1r { v3.4s }, [x20]\n"
       "srshl v31.4s, v31.4s, v0.4s\n"
       "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v3.4s }, [x21]\n"
-      "ld1r { v2.4s }, [x20]\n"
       "srshl v21.4s, v21.4s, v0.4s\n"
       "srshl v22.4s, v22.4s, v0.4s\n"
+      "add x20, %x[qp], %[maxval]\n"
+      "ld1r { v2.4s }, [x20]\n"
       "srshl v16.4s, v16.4s, v0.4s\n"
       "srshl v17.4s, v17.4s, v0.4s\n"
       "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
       "ld1r { v1.4s }, [x20]\n"
       "srshl v18.4s, v18.4s, v0.4s\n"
       "srshl v19.4s, v19.4s, v0.4s\n"
+      "cmp x9, #0x10\n"
       "srshl v23.4s, v23.4s, v0.4s\n"
       "srshl v28.4s, v28.4s, v0.4s\n"
       "srshl v29.4s, v29.4s, v0.4s\n"
@@ -1980,100 +1979,100 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "bge 127f\n"
       "tbz x9, #3, 122f\n"
       "str d31, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
+      "str d16, [x23], #0x8\n"
+      "str d23, [x22], #0x8\n"
+      "str d24, [x21], #0x8\n"
       "tbz x9, #2, 120f\n"
       "st1 { v31.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "st1 { v23.s }[2], [x23], #0x4\n"
-      "st1 { v24.s }[2], [x22], #0x4\n"
+      "st1 { v16.s }[2], [x23], #0x4\n"
+      "st1 { v23.s }[2], [x22], #0x4\n"
+      "st1 { v24.s }[2], [x21], #0x4\n"
       "tbz x9, #1, 119f\n"
       "st1 { v31.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "st1 { v23.h }[6], [x23], #0x2\n"
-      "st1 { v24.h }[6], [x22], #0x2\n"
+      "st1 { v16.h }[6], [x23], #0x2\n"
+      "st1 { v23.h }[6], [x22], #0x2\n"
+      "st1 { v24.h }[6], [x21], #0x2\n"
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "st1 { v23.b }[14], [x23]\n"
-      "st1 { v24.b }[14], [x22]\n"
+      "st1 { v16.b }[14], [x23]\n"
+      "st1 { v23.b }[14], [x22]\n"
+      "st1 { v24.b }[14], [x21]\n"
       "b 126f\n"
       "119:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "st1 { v23.b }[12], [x23]\n"
-      "st1 { v24.b }[12], [x22]\n"
+      "st1 { v16.b }[12], [x23]\n"
+      "st1 { v23.b }[12], [x22]\n"
+      "st1 { v24.b }[12], [x21]\n"
       "b 126f\n"
       "120:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x9, #1, 121f\n"
       "st1 { v31.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "st1 { v23.h }[4], [x23], #0x2\n"
-      "st1 { v24.h }[4], [x22], #0x2\n"
+      "st1 { v16.h }[4], [x23], #0x2\n"
+      "st1 { v23.h }[4], [x22], #0x2\n"
+      "st1 { v24.h }[4], [x21], #0x2\n"
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "st1 { v23.b }[10], [x23]\n"
-      "st1 { v24.b }[10], [x22]\n"
+      "st1 { v16.b }[10], [x23]\n"
+      "st1 { v23.b }[10], [x22]\n"
+      "st1 { v24.b }[10], [x21]\n"
       "b 126f\n"
       "121:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "st1 { v23.b }[8], [x23]\n"
-      "st1 { v24.b }[8], [x22]\n"
+      "st1 { v16.b }[8], [x23]\n"
+      "st1 { v23.b }[8], [x22]\n"
+      "st1 { v24.b }[8], [x21]\n"
       "b 126f\n"
       "122:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x9, #2, 124f\n"
       "str s31, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "str s23, [x23], #0x4\n"
-      "str s24, [x22], #0x4\n"
+      "str s16, [x23], #0x4\n"
+      "str s23, [x22], #0x4\n"
+      "str s24, [x21], #0x4\n"
       "tbz x9, #1, 123f\n"
       "st1 { v31.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "st1 { v23.h }[2], [x23], #0x2\n"
-      "st1 { v24.h }[2], [x22], #0x2\n"
+      "st1 { v16.h }[2], [x23], #0x2\n"
+      "st1 { v23.h }[2], [x22], #0x2\n"
+      "st1 { v24.h }[2], [x21], #0x2\n"
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "st1 { v23.b }[6], [x23]\n"
-      "st1 { v24.b }[6], [x22]\n"
+      "st1 { v16.b }[6], [x23]\n"
+      "st1 { v23.b }[6], [x22]\n"
+      "st1 { v24.b }[6], [x21]\n"
       "b 126f\n"
       "123:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "st1 { v23.b }[4], [x23]\n"
-      "st1 { v24.b }[4], [x22]\n"
+      "st1 { v16.b }[4], [x23]\n"
+      "st1 { v23.b }[4], [x22]\n"
+      "st1 { v24.b }[4], [x21]\n"
       "b 126f\n"
       "124:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x9, #1, 125f\n"
       "str h31, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "str h23, [x23], #0x2\n"
-      "str h24, [x22], #0x2\n"
+      "str h16, [x23], #0x2\n"
+      "str h23, [x22], #0x2\n"
+      "str h24, [x21], #0x2\n"
       "tbz x9, #0, 126f\n"
       "st1 { v31.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "st1 { v23.b }[2], [x23]\n"
-      "st1 { v24.b }[2], [x22]\n"
+      "st1 { v16.b }[2], [x23]\n"
+      "st1 { v23.b }[2], [x22]\n"
+      "st1 { v24.b }[2], [x21]\n"
       "b 126f\n"
       "125:"  // Height 4: Partial direct writeback: partial_1_0
       "str b31, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "str b23, [x23, #0x0]\n"
-      "str b24, [x22, #0x0]\n"
+      "str b16, [x23, #0x0]\n"
+      "str b23, [x22, #0x0]\n"
+      "str b24, [x21, #0x0]\n"
       "126:"  // Height 4: Partial direct writeback: Done
       "b 128f\n"
       "127:"  // Height 4: Full writeback
       "str q31, [x27, #0x0]\n"
       "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
-      "str q23, [x23, #0x0]\n"
-      "str q24, [x22, #0x0]\n"
+      "str q16, [x23, #0x0]\n"
+      "str q23, [x22, #0x0]\n"
+      "str q24, [x21, #0x0]\n"
       "128:"  // Height 4: Writeback done
       "subs x9, x9, #0x10\n"
       "bgt 98b\n"
@@ -2089,8 +2088,8 @@ void a64_hybrid_u8qa_mmla_4x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "130:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_dot_4x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_dot_4x16.hpp
deleted file mode 100644
index be1947effc..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_dot_4x16.hpp
+++ /dev/null
@@ -1,103 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef __aarch64__
-
-#include "../std_transforms_fixed.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    unsigned int, const unsigned int *, \
-    IndirectInputArg<uint8_t>, \
-    size_t, size_t, \
-    const int8_t *, \
-    IndirectOutputArg<uint8_t>, \
-    const Requantize32 *, const int32_t *, unsigned int
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void a64_hybrid_u8s8qa_dot_4x16( ARGLIST );
-
-class cls_a64_hybrid_u8s8qa_dot_4x16
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef uint8_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 4;
-    }
-
-    static unsigned int out_width()
-    {
-        return 16;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 4;
-    }
-
-    static constexpr bool supports_accumulate()
-    {
-        return false;
-    }
-
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 16, 4> transforms = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                case CPUModel::A55r1:
-                    return { 7.5301 };
-                default:
-                    return { 27.5482 };
-                case CPUModel::A510:
-                    return { 14.81 };
-                case CPUModel::V1:
-                    return { 44.54 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=a64_hybrid_u8s8qa_dot_4x16;
-    cls_a64_hybrid_u8s8qa_dot_4x16(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_dot_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_dot_4x16/generic.cpp
deleted file mode 100644
index e5ca848fb9..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_dot_4x16/generic.cpp
+++ /dev/null
@@ -1,2027 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef __aarch64__
-
-#include "arm_gemm.hpp"
-#include "../../utils.hpp"
-
-#include <cassert>
-#include <cstdint>
-
-namespace arm_gemm {
-
-void a64_hybrid_u8s8qa_dot_4x16 (
-    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg,
-    size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<uint8_t> output_arg,
-    const Requantize32 *qp, const int32_t *col_bias, unsigned int
-)
-{
-    struct KernelArgs {
-        unsigned int num_strings = {};
-        const unsigned int *string_lengths = {};
-        size_t N = {};
-        const int8_t *B_ptr = {};
-        size_t output_offset = {};
-        size_t input_initial_col = {};
-        size_t input_offset = {};
-        void *output_ptr = {};
-    } ka;
-
-    unsigned long flags=0;
-    void *input_ptr;
-
-    if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
-        ka.output_offset=output_arg.indirect.offset;
-        flags |= 0x4;
-    } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
-        ka.output_offset=output_arg.direct.stride;
-    }
-
-    if (A_arg.is_indirect) {
-        input_ptr=(void *)(A_arg.indirect.ptr);
-        ka.input_offset=A_arg.indirect.start_row;
-        ka.input_initial_col=A_arg.indirect.start_col;
-        flags |= 0x8;
-    } else {
-        assert(num_strings==1);
-        input_ptr=(void *)(A_arg.direct.base);
-        ka.input_offset=A_arg.direct.stride;
-    }
-    ka.num_strings = num_strings;
-    ka.string_lengths = string_lengths;
-    ka.N = N;
-    ka.B_ptr = B_ptr;
-    if (qp->c_offset > qp->minval) {
-        flags |= 0x20;
-    }
-    __asm__ __volatile__(
-      "1:"  // Row loop
-      "cmp %x[M], #0x4\n"
-      "bge 91f\n"
-      "cmp %x[M], #0x2\n"
-      "bgt 61f\n"
-      "beq 31f\n"
-      "mov x10, %x[col_bias]\n"
-      "movi v11.4s, #0x0\n"
-      "movi v15.16b, #0x1\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "2:"  // Height 1: Column loop
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "3:"  // Height 1: setup done
-      "mov x26, #0x0\n"
-      "4:"  // Height 1: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 5f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "cbnz x26, 6f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "b 6f\n"
-      "5:"  // Height 1: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "6:"  // Height 1: input setup done
-      "cmp x25, #0x10\n"
-      "blt 11f\n"
-      "ldr q0, [x24, #0x0]\n"
-      "ldr q4, [x28, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q5, [x28, #0x10]\n"
-      "ldr q6, [x28, #0x20]\n"
-      "ldr q7, [x28, #0x30]\n"
-      "ldr q8, [x28, #0x40]\n"
-      "ldr q9, [x28, #0x50]\n"
-      "ldr q10, [x28, #0x60]\n"
-      "blt 9f\n"
-      "7:"  // Height 1: Multiply loop: Main loop head
-      ".inst 0x4f00f090  // sudot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr q21, [x28, #0x70]\n"
-      ".inst 0x4f00f0b1  // sudot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr q20, [x28, #0x80]\n"
-      ".inst 0x4f00f0d2  // sudot v18.4s, v6.16b, v0.4b[0]\n"
-      "ldr q26, [x28, #0x90]\n"
-      ".inst 0x4f00f0f3  // sudot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr q25, [x28, #0xa0]\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f20f110  // sudot v16.4s, v8.16b, v0.4b[1]\n"
-      "ldr q24, [x28, #0xb0]\n"
-      ".inst 0x4f20f131  // sudot v17.4s, v9.16b, v0.4b[1]\n"
-      "ldr q23, [x28, #0xc0]\n"
-      ".inst 0x4f20f152  // sudot v18.4s, v10.16b, v0.4b[1]\n"
-      "ldr q22, [x28, #0xd0]\n"
-      ".inst 0x4f20f2b3  // sudot v19.4s, v21.16b, v0.4b[1]\n"
-      "ldr q21, [x28, #0xe0]\n"
-      ".inst 0x4f00fa90  // sudot v16.4s, v20.16b, v0.4b[2]\n"
-      "ldr q20, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4f00fb51  // sudot v17.4s, v26.16b, v0.4b[2]\n"
-      ".inst 0x4f00fb32  // sudot v18.4s, v25.16b, v0.4b[2]\n"
-      ".inst 0x4f00fb13  // sudot v19.4s, v24.16b, v0.4b[2]\n"
-      ".inst 0x4f20faf0  // sudot v16.4s, v23.16b, v0.4b[3]\n"
-      ".inst 0x4f20fad1  // sudot v17.4s, v22.16b, v0.4b[3]\n"
-      ".inst 0x4f20fab2  // sudot v18.4s, v21.16b, v0.4b[3]\n"
-      ".inst 0x4f20fa93  // sudot v19.4s, v20.16b, v0.4b[3]\n"
-      "tbnz %x[flags], #31, 8f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      "8:"  // Height 1: Multiply loop: unique 1: skip row sum
-      "ldr q0, [x24, #0x0]\n"
-      "ldr q4, [x28, #0x0]\n"
-      "sub x25, x25, #0x10\n"
-      "ldr q5, [x28, #0x10]\n"
-      "ldr q6, [x28, #0x20]\n"
-      "cmp x25, #0x20\n"
-      "ldr q7, [x28, #0x30]\n"
-      "ldr q8, [x28, #0x40]\n"
-      "ldr q9, [x28, #0x50]\n"
-      "ldr q10, [x28, #0x60]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "bge 7b\n"
-      "9:"  // Height 1: Multiply loop: Single iteration only
-      ".inst 0x4f00f090  // sudot v16.4s, v4.16b, v0.4b[0]\n"
-      "ldr q21, [x28, #0x70]\n"
-      ".inst 0x4f00f0b1  // sudot v17.4s, v5.16b, v0.4b[0]\n"
-      "ldr q20, [x28, #0x80]\n"
-      ".inst 0x4f00f0d2  // sudot v18.4s, v6.16b, v0.4b[0]\n"
-      "ldr q26, [x28, #0x90]\n"
-      ".inst 0x4f00f0f3  // sudot v19.4s, v7.16b, v0.4b[0]\n"
-      "ldr q25, [x28, #0xa0]\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f20f110  // sudot v16.4s, v8.16b, v0.4b[1]\n"
-      "ldr q24, [x28, #0xb0]\n"
-      ".inst 0x4f20f131  // sudot v17.4s, v9.16b, v0.4b[1]\n"
-      "ldr q23, [x28, #0xc0]\n"
-      ".inst 0x4f20f152  // sudot v18.4s, v10.16b, v0.4b[1]\n"
-      "ldr q22, [x28, #0xd0]\n"
-      ".inst 0x4f20f2b3  // sudot v19.4s, v21.16b, v0.4b[1]\n"
-      "ldr q21, [x28, #0xe0]\n"
-      ".inst 0x4f00fa90  // sudot v16.4s, v20.16b, v0.4b[2]\n"
-      "ldr q20, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4f00fb51  // sudot v17.4s, v26.16b, v0.4b[2]\n"
-      ".inst 0x4f00fb32  // sudot v18.4s, v25.16b, v0.4b[2]\n"
-      ".inst 0x4f00fb13  // sudot v19.4s, v24.16b, v0.4b[2]\n"
-      ".inst 0x4f20faf0  // sudot v16.4s, v23.16b, v0.4b[3]\n"
-      ".inst 0x4f20fad1  // sudot v17.4s, v22.16b, v0.4b[3]\n"
-      ".inst 0x4f20fab2  // sudot v18.4s, v21.16b, v0.4b[3]\n"
-      ".inst 0x4f20fa93  // sudot v19.4s, v20.16b, v0.4b[3]\n"
-      "tbnz %x[flags], #31, 10f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      "10:"  // Height 1: Multiply loop: unique 2: skip row sum
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "11:"  // Height 1: Multiply loop: Main loop skip
-      "cbz x25, 18f\n"
-      "cmp x25, #0x4\n"
-      "blt 14f\n"
-      "12:"  // Height 1: Multiply loop: Odd block loop
-      "ldr s0, [x24], #0x4\n"
-      "tbnz %x[flags], #31, 13f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      "13:"  // Height 1: Multiply loop: unique 3: skip row sum
-      "ldr q23, [x28, #0x0]\n"
-      "ldr q22, [x28, #0x10]\n"
-      "sub x25, x25, #0x4\n"
-      "ldr q21, [x28, #0x20]\n"
-      "ldr q20, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f00f2f0  // sudot v16.4s, v23.16b, v0.4b[0]\n"
-      ".inst 0x4f00f2d1  // sudot v17.4s, v22.16b, v0.4b[0]\n"
-      ".inst 0x4f00f2b2  // sudot v18.4s, v21.16b, v0.4b[0]\n"
-      ".inst 0x4f00f293  // sudot v19.4s, v20.16b, v0.4b[0]\n"
-      "bge 12b\n"
-      "14:"  // Height 1: Multiply loop: Skip odd blocks
-      "cbz x25, 18f\n"
-      "tbz x25, #1, 15f\n"
-      "ldr h0, [x24], #0x2\n"
-      "tbz x25, #0, 16f\n"
-      "ld1 { v0.b }[2], [x24]\n"
-      "b 16f\n"
-      "15:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x24, #0x0]\n"
-      "16:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "tbnz %x[flags], #31, 17f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      "17:"  // Height 1: Multiply loop: unique 4: skip row sum
-      "ldr q23, [x28, #0x0]\n"
-      "ldr q22, [x28, #0x10]\n"
-      "ldr q21, [x28, #0x20]\n"
-      "ldr q20, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f00f2f0  // sudot v16.4s, v23.16b, v0.4b[0]\n"
-      ".inst 0x4f00f2d1  // sudot v17.4s, v22.16b, v0.4b[0]\n"
-      ".inst 0x4f00f2b2  // sudot v18.4s, v21.16b, v0.4b[0]\n"
-      ".inst 0x4f00f293  // sudot v19.4s, v20.16b, v0.4b[0]\n"
-      "18:"  // Height 1: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 4b\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "tbnz %x[flags], #31, 19f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "ld1r { v20.4s }, [x20]\n"
-      "neg v20.4s, v20.4s\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "mul v11.4s, v11.4s, v20.4s\n"
-      "19:"  // Height 1: skip row sum fixup
-      "ldr q24, [x10, #0x0]\n"
-      "ldr q23, [x10, #0x10]\n"
-      "add v16.4s, v16.4s, v11.4s\n"
-      "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q22, [x10, #0x20]\n"
-      "ldr q21, [x10, #0x30]\n"
-      "add v18.4s, v18.4s, v11.4s\n"
-      "add v19.4s, v19.4s, v11.4s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "ld1r { v20.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v24.4s\n"
-      "add v17.4s, v17.4s, v23.4s\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add x10, x10, #0x40\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v22.4s\n"
-      "add v19.4s, v19.4s, v21.4s\n"
-      "sqrdmulh v16.4s, v16.4s, v20.4s\n"
-      "sqrdmulh v17.4s, v17.4s, v20.4s\n"
-      "sqrdmulh v18.4s, v18.4s, v20.4s\n"
-      "sqrdmulh v19.4s, v19.4s, v20.4s\n"
-      "tbz %x[flags], #5, 20f\n"
-      "and v23.16b, v16.16b, v0.16b\n"
-      "and v22.16b, v17.16b, v0.16b\n"
-      "and v21.16b, v18.16b, v0.16b\n"
-      "and v20.16b, v19.16b, v0.16b\n"
-      "sshr v23.4s, v23.4s, #0x1f\n"
-      "sshr v22.4s, v22.4s, #0x1f\n"
-      "sshr v21.4s, v21.4s, #0x1f\n"
-      "sshr v20.4s, v20.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v23.4s\n"
-      "sqadd v17.4s, v17.4s, v22.4s\n"
-      "sqadd v18.4s, v18.4s, v21.4s\n"
-      "sqadd v19.4s, v19.4s, v20.4s\n"
-      "20:"  // Height 1: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
-      "srshl v16.4s, v16.4s, v0.4s\n"
-      "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v22.4s }, [x21]\n"
-      "ld1r { v21.4s }, [x20]\n"
-      "srshl v18.4s, v18.4s, v0.4s\n"
-      "srshl v19.4s, v19.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v20.4s }, [x20]\n"
-      "add v16.4s, v16.4s, v22.4s\n"
-      "add v17.4s, v17.4s, v22.4s\n"
-      "add v18.4s, v18.4s, v22.4s\n"
-      "add v19.4s, v19.4s, v22.4s\n"
-      "smin v16.4s, v16.4s, v21.4s\n"
-      "smin v17.4s, v17.4s, v21.4s\n"
-      "smin v18.4s, v18.4s, v21.4s\n"
-      "smin v19.4s, v19.4s, v21.4s\n"
-      "smax v16.4s, v16.4s, v20.4s\n"
-      "smax v17.4s, v17.4s, v20.4s\n"
-      "smax v18.4s, v18.4s, v20.4s\n"
-      "smax v19.4s, v19.4s, v20.4s\n"
-      "uzp1 v16.8h, v16.8h, v17.8h\n"
-      "uzp1 v17.8h, v18.8h, v19.8h\n"
-      "uzp1 v16.16b, v16.16b, v17.16b\n"
-      "bge 29f\n"
-      "tbz x9, #3, 24f\n"
-      "str d16, [x27], #0x8\n"
-      "tbz x9, #2, 22f\n"
-      "st1 { v16.s }[2], [x27], #0x4\n"
-      "tbz x9, #1, 21f\n"
-      "st1 { v16.h }[6], [x27], #0x2\n"
-      "tbz x9, #0, 28f\n"
-      "st1 { v16.b }[14], [x27]\n"
-      "b 28f\n"
-      "21:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x9, #0, 28f\n"
-      "st1 { v16.b }[12], [x27]\n"
-      "b 28f\n"
-      "22:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x9, #1, 23f\n"
-      "st1 { v16.h }[4], [x27], #0x2\n"
-      "tbz x9, #0, 28f\n"
-      "st1 { v16.b }[10], [x27]\n"
-      "b 28f\n"
-      "23:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x9, #0, 28f\n"
-      "st1 { v16.b }[8], [x27]\n"
-      "b 28f\n"
-      "24:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x9, #2, 26f\n"
-      "str s16, [x27], #0x4\n"
-      "tbz x9, #1, 25f\n"
-      "st1 { v16.h }[2], [x27], #0x2\n"
-      "tbz x9, #0, 28f\n"
-      "st1 { v16.b }[6], [x27]\n"
-      "b 28f\n"
-      "25:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x9, #0, 28f\n"
-      "st1 { v16.b }[4], [x27]\n"
-      "b 28f\n"
-      "26:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x9, #1, 27f\n"
-      "str h16, [x27], #0x2\n"
-      "tbz x9, #0, 28f\n"
-      "st1 { v16.b }[2], [x27]\n"
-      "b 28f\n"
-      "27:"  // Height 1: Partial direct writeback: partial_1_0
-      "str b16, [x27, #0x0]\n"
-      "28:"  // Height 1: Partial direct writeback: Done
-      "b 30f\n"
-      "29:"  // Height 1: Full writeback
-      "str q16, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "30:"  // Height 1: Writeback done
-      "subs x9, x9, #0x10\n"
-      "bgt 2b\n"
-      "b 122f\n"
-      "31:"  // Height 2
-      "mov x10, %x[col_bias]\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "movi v15.16b, #0x1\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "32:"  // Height 2: Column loop
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "33:"  // Height 2: setup done
-      "mov x26, #0x0\n"
-      "34:"  // Height 2: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 35f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "cbnz x26, 36f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "b 36f\n"
-      "35:"  // Height 2: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "36:"  // Height 2: input setup done
-      "cmp x25, #0x10\n"
-      "blt 41f\n"
-      "ldr q0, [x24, #0x0]\n"
-      "ldr q1, [x23, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q4, [x28, #0x0]\n"
-      "ldr q5, [x28, #0x10]\n"
-      "ldr q6, [x28, #0x20]\n"
-      "ldr q7, [x28, #0x30]\n"
-      "ldr q8, [x28, #0x40]\n"
-      "ldr q9, [x28, #0x50]\n"
-      "ldr q10, [x28, #0x60]\n"
-      "blt 39f\n"
-      "37:"  // Height 2: Multiply loop: Main loop head
-      ".inst 0x4f00f090  // sudot v16.4s, v4.16b, v0.4b[0]\n"
-      ".inst 0x4f01f094  // sudot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr q25, [x28, #0x70]\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f00f0b1  // sudot v17.4s, v5.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0b5  // sudot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr q24, [x28, #0x80]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f00f0d2  // sudot v18.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0d6  // sudot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr q30, [x28, #0x90]\n"
-      ".inst 0x4f00f0f3  // sudot v19.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0f7  // sudot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr q29, [x28, #0xa0]\n"
-      ".inst 0x4f20f110  // sudot v16.4s, v8.16b, v0.4b[1]\n"
-      ".inst 0x4f21f114  // sudot v20.4s, v8.16b, v1.4b[1]\n"
-      "ldr q28, [x28, #0xb0]\n"
-      ".inst 0x4f20f131  // sudot v17.4s, v9.16b, v0.4b[1]\n"
-      ".inst 0x4f21f135  // sudot v21.4s, v9.16b, v1.4b[1]\n"
-      "ldr q27, [x28, #0xc0]\n"
-      ".inst 0x4f20f152  // sudot v18.4s, v10.16b, v0.4b[1]\n"
-      ".inst 0x4f21f156  // sudot v22.4s, v10.16b, v1.4b[1]\n"
-      "ldr q26, [x28, #0xd0]\n"
-      ".inst 0x4f20f333  // sudot v19.4s, v25.16b, v0.4b[1]\n"
-      ".inst 0x4f21f337  // sudot v23.4s, v25.16b, v1.4b[1]\n"
-      "ldr q25, [x28, #0xe0]\n"
-      ".inst 0x4f00fb10  // sudot v16.4s, v24.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb14  // sudot v20.4s, v24.16b, v1.4b[2]\n"
-      "ldr q24, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4f00fbd1  // sudot v17.4s, v30.16b, v0.4b[2]\n"
-      ".inst 0x4f01fbd5  // sudot v21.4s, v30.16b, v1.4b[2]\n"
-      ".inst 0x4f00fbb2  // sudot v18.4s, v29.16b, v0.4b[2]\n"
-      ".inst 0x4f01fbb6  // sudot v22.4s, v29.16b, v1.4b[2]\n"
-      ".inst 0x4f00fb93  // sudot v19.4s, v28.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb97  // sudot v23.4s, v28.16b, v1.4b[2]\n"
-      ".inst 0x4f20fb70  // sudot v16.4s, v27.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb74  // sudot v20.4s, v27.16b, v1.4b[3]\n"
-      ".inst 0x4f20fb51  // sudot v17.4s, v26.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb55  // sudot v21.4s, v26.16b, v1.4b[3]\n"
-      ".inst 0x4f20fb32  // sudot v18.4s, v25.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb36  // sudot v22.4s, v25.16b, v1.4b[3]\n"
-      ".inst 0x4f20fb13  // sudot v19.4s, v24.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb17  // sudot v23.4s, v24.16b, v1.4b[3]\n"
-      "tbnz %x[flags], #31, 38f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      "38:"  // Height 2: Multiply loop: unique 5: skip row sum
-      "ldr q0, [x24, #0x0]\n"
-      "ldr q1, [x23, #0x0]\n"
-      "sub x25, x25, #0x10\n"
-      "ldr q4, [x28, #0x0]\n"
-      "ldr q5, [x28, #0x10]\n"
-      "cmp x25, #0x20\n"
-      "ldr q6, [x28, #0x20]\n"
-      "ldr q7, [x28, #0x30]\n"
-      "ldr q8, [x28, #0x40]\n"
-      "ldr q9, [x28, #0x50]\n"
-      "ldr q10, [x28, #0x60]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "bge 37b\n"
-      "39:"  // Height 2: Multiply loop: Single iteration only
-      ".inst 0x4f00f090  // sudot v16.4s, v4.16b, v0.4b[0]\n"
-      ".inst 0x4f01f094  // sudot v20.4s, v4.16b, v1.4b[0]\n"
-      "ldr q25, [x28, #0x70]\n"
-      "sub x25, x25, #0x10\n"
-      ".inst 0x4f00f0b1  // sudot v17.4s, v5.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0b5  // sudot v21.4s, v5.16b, v1.4b[0]\n"
-      "ldr q24, [x28, #0x80]\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f00f0d2  // sudot v18.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0d6  // sudot v22.4s, v6.16b, v1.4b[0]\n"
-      "ldr q30, [x28, #0x90]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f00f0f3  // sudot v19.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0f7  // sudot v23.4s, v7.16b, v1.4b[0]\n"
-      "ldr q29, [x28, #0xa0]\n"
-      ".inst 0x4f20f110  // sudot v16.4s, v8.16b, v0.4b[1]\n"
-      ".inst 0x4f21f114  // sudot v20.4s, v8.16b, v1.4b[1]\n"
-      "ldr q28, [x28, #0xb0]\n"
-      ".inst 0x4f20f131  // sudot v17.4s, v9.16b, v0.4b[1]\n"
-      ".inst 0x4f21f135  // sudot v21.4s, v9.16b, v1.4b[1]\n"
-      "ldr q27, [x28, #0xc0]\n"
-      ".inst 0x4f20f152  // sudot v18.4s, v10.16b, v0.4b[1]\n"
-      ".inst 0x4f21f156  // sudot v22.4s, v10.16b, v1.4b[1]\n"
-      "ldr q26, [x28, #0xd0]\n"
-      ".inst 0x4f20f333  // sudot v19.4s, v25.16b, v0.4b[1]\n"
-      ".inst 0x4f21f337  // sudot v23.4s, v25.16b, v1.4b[1]\n"
-      "ldr q25, [x28, #0xe0]\n"
-      ".inst 0x4f00fb10  // sudot v16.4s, v24.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb14  // sudot v20.4s, v24.16b, v1.4b[2]\n"
-      "ldr q24, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4f00fbd1  // sudot v17.4s, v30.16b, v0.4b[2]\n"
-      ".inst 0x4f01fbd5  // sudot v21.4s, v30.16b, v1.4b[2]\n"
-      ".inst 0x4f00fbb2  // sudot v18.4s, v29.16b, v0.4b[2]\n"
-      ".inst 0x4f01fbb6  // sudot v22.4s, v29.16b, v1.4b[2]\n"
-      ".inst 0x4f00fb93  // sudot v19.4s, v28.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb97  // sudot v23.4s, v28.16b, v1.4b[2]\n"
-      ".inst 0x4f20fb70  // sudot v16.4s, v27.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb74  // sudot v20.4s, v27.16b, v1.4b[3]\n"
-      ".inst 0x4f20fb51  // sudot v17.4s, v26.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb55  // sudot v21.4s, v26.16b, v1.4b[3]\n"
-      ".inst 0x4f20fb32  // sudot v18.4s, v25.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb36  // sudot v22.4s, v25.16b, v1.4b[3]\n"
-      ".inst 0x4f20fb13  // sudot v19.4s, v24.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb17  // sudot v23.4s, v24.16b, v1.4b[3]\n"
-      "tbnz %x[flags], #31, 40f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      "40:"  // Height 2: Multiply loop: unique 6: skip row sum
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "41:"  // Height 2: Multiply loop: Main loop skip
-      "cbz x25, 48f\n"
-      "cmp x25, #0x4\n"
-      "blt 44f\n"
-      "42:"  // Height 2: Multiply loop: Odd block loop
-      "ldr s0, [x24], #0x4\n"
-      "ldr s1, [x23], #0x4\n"
-      "tbnz %x[flags], #31, 43f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      "43:"  // Height 2: Multiply loop: unique 7: skip row sum
-      "ldr q27, [x28, #0x0]\n"
-      "ldr q26, [x28, #0x10]\n"
-      "sub x25, x25, #0x4\n"
-      "ldr q25, [x28, #0x20]\n"
-      "ldr q24, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f00f370  // sudot v16.4s, v27.16b, v0.4b[0]\n"
-      ".inst 0x4f01f374  // sudot v20.4s, v27.16b, v1.4b[0]\n"
-      ".inst 0x4f00f351  // sudot v17.4s, v26.16b, v0.4b[0]\n"
-      ".inst 0x4f01f355  // sudot v21.4s, v26.16b, v1.4b[0]\n"
-      ".inst 0x4f00f332  // sudot v18.4s, v25.16b, v0.4b[0]\n"
-      ".inst 0x4f01f336  // sudot v22.4s, v25.16b, v1.4b[0]\n"
-      ".inst 0x4f00f313  // sudot v19.4s, v24.16b, v0.4b[0]\n"
-      ".inst 0x4f01f317  // sudot v23.4s, v24.16b, v1.4b[0]\n"
-      "bge 42b\n"
-      "44:"  // Height 2: Multiply loop: Skip odd blocks
-      "cbz x25, 48f\n"
-      "tbz x25, #1, 45f\n"
-      "ldr h0, [x24], #0x2\n"
-      "ldr h1, [x23], #0x2\n"
-      "tbz x25, #0, 46f\n"
-      "ld1 { v0.b }[2], [x24]\n"
-      "ld1 { v1.b }[2], [x23]\n"
-      "b 46f\n"
-      "45:"  // Height 2: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x24, #0x0]\n"
-      "ldr b1, [x23, #0x0]\n"
-      "46:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "tbnz %x[flags], #31, 47f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      "47:"  // Height 2: Multiply loop: unique 8: skip row sum
-      "ldr q27, [x28, #0x0]\n"
-      "ldr q26, [x28, #0x10]\n"
-      "ldr q25, [x28, #0x20]\n"
-      "ldr q24, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f00f370  // sudot v16.4s, v27.16b, v0.4b[0]\n"
-      ".inst 0x4f01f374  // sudot v20.4s, v27.16b, v1.4b[0]\n"
-      ".inst 0x4f00f351  // sudot v17.4s, v26.16b, v0.4b[0]\n"
-      ".inst 0x4f01f355  // sudot v21.4s, v26.16b, v1.4b[0]\n"
-      ".inst 0x4f00f332  // sudot v18.4s, v25.16b, v0.4b[0]\n"
-      ".inst 0x4f01f336  // sudot v22.4s, v25.16b, v1.4b[0]\n"
-      ".inst 0x4f00f313  // sudot v19.4s, v24.16b, v0.4b[0]\n"
-      ".inst 0x4f01f317  // sudot v23.4s, v24.16b, v1.4b[0]\n"
-      "48:"  // Height 2: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 34b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "tbnz %x[flags], #31, 49f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v24.4s }, [x20]\n"
-      "neg v24.4s, v24.4s\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "addp v12.4s, v12.4s, v12.4s\n"
-      "mul v11.4s, v11.4s, v24.4s\n"
-      "mul v12.4s, v12.4s, v24.4s\n"
-      "49:"  // Height 2: skip row sum fixup
-      "ldr q28, [x10, #0x0]\n"
-      "ldr q27, [x10, #0x10]\n"
-      "add v16.4s, v16.4s, v11.4s\n"
-      "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q26, [x10, #0x20]\n"
-      "ldr q25, [x10, #0x30]\n"
-      "add v18.4s, v18.4s, v11.4s\n"
-      "add v19.4s, v19.4s, v11.4s\n"
-      "add v20.4s, v20.4s, v12.4s\n"
-      "add v21.4s, v21.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "ld1r { v24.4s }, [x20]\n"
-      "add v22.4s, v22.4s, v12.4s\n"
-      "add v23.4s, v23.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add v16.4s, v16.4s, v28.4s\n"
-      "add v17.4s, v17.4s, v27.4s\n"
-      "add x10, x10, #0x40\n"
-      "add v18.4s, v18.4s, v26.4s\n"
-      "add v19.4s, v19.4s, v25.4s\n"
-      "add v20.4s, v20.4s, v28.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "add v21.4s, v21.4s, v27.4s\n"
-      "add v22.4s, v22.4s, v26.4s\n"
-      "add v23.4s, v23.4s, v25.4s\n"
-      "sqrdmulh v16.4s, v16.4s, v24.4s\n"
-      "sqrdmulh v17.4s, v17.4s, v24.4s\n"
-      "sqrdmulh v18.4s, v18.4s, v24.4s\n"
-      "sqrdmulh v19.4s, v19.4s, v24.4s\n"
-      "sqrdmulh v20.4s, v20.4s, v24.4s\n"
-      "sqrdmulh v21.4s, v21.4s, v24.4s\n"
-      "sqrdmulh v22.4s, v22.4s, v24.4s\n"
-      "sqrdmulh v23.4s, v23.4s, v24.4s\n"
-      "tbz %x[flags], #5, 50f\n"
-      "and v24.16b, v16.16b, v0.16b\n"
-      "and v30.16b, v17.16b, v0.16b\n"
-      "and v29.16b, v18.16b, v0.16b\n"
-      "and v28.16b, v19.16b, v0.16b\n"
-      "and v27.16b, v20.16b, v0.16b\n"
-      "and v26.16b, v21.16b, v0.16b\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
-      "and v25.16b, v22.16b, v0.16b\n"
-      "sshr v30.4s, v30.4s, #0x1f\n"
-      "sshr v29.4s, v29.4s, #0x1f\n"
-      "sshr v28.4s, v28.4s, #0x1f\n"
-      "sshr v27.4s, v27.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v24.4s\n"
-      "and v24.16b, v23.16b, v0.16b\n"
-      "sshr v26.4s, v26.4s, #0x1f\n"
-      "sshr v25.4s, v25.4s, #0x1f\n"
-      "sqadd v17.4s, v17.4s, v30.4s\n"
-      "sqadd v18.4s, v18.4s, v29.4s\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
-      "sqadd v19.4s, v19.4s, v28.4s\n"
-      "sqadd v20.4s, v20.4s, v27.4s\n"
-      "sqadd v21.4s, v21.4s, v26.4s\n"
-      "sqadd v22.4s, v22.4s, v25.4s\n"
-      "sqadd v23.4s, v23.4s, v24.4s\n"
-      "50:"  // Height 2: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
-      "srshl v16.4s, v16.4s, v0.4s\n"
-      "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x21]\n"
-      "ld1r { v25.4s }, [x20]\n"
-      "srshl v18.4s, v18.4s, v0.4s\n"
-      "srshl v19.4s, v19.4s, v0.4s\n"
-      "srshl v20.4s, v20.4s, v0.4s\n"
-      "srshl v21.4s, v21.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v24.4s }, [x20]\n"
-      "srshl v22.4s, v22.4s, v0.4s\n"
-      "srshl v23.4s, v23.4s, v0.4s\n"
-      "add v16.4s, v16.4s, v26.4s\n"
-      "add v17.4s, v17.4s, v26.4s\n"
-      "add v18.4s, v18.4s, v26.4s\n"
-      "add v19.4s, v19.4s, v26.4s\n"
-      "add v20.4s, v20.4s, v26.4s\n"
-      "add v21.4s, v21.4s, v26.4s\n"
-      "add v22.4s, v22.4s, v26.4s\n"
-      "add v23.4s, v23.4s, v26.4s\n"
-      "smin v16.4s, v16.4s, v25.4s\n"
-      "smin v17.4s, v17.4s, v25.4s\n"
-      "smin v18.4s, v18.4s, v25.4s\n"
-      "smin v19.4s, v19.4s, v25.4s\n"
-      "smin v20.4s, v20.4s, v25.4s\n"
-      "smin v21.4s, v21.4s, v25.4s\n"
-      "smin v22.4s, v22.4s, v25.4s\n"
-      "smin v23.4s, v23.4s, v25.4s\n"
-      "smax v16.4s, v16.4s, v24.4s\n"
-      "smax v17.4s, v17.4s, v24.4s\n"
-      "smax v18.4s, v18.4s, v24.4s\n"
-      "smax v19.4s, v19.4s, v24.4s\n"
-      "smax v20.4s, v20.4s, v24.4s\n"
-      "smax v21.4s, v21.4s, v24.4s\n"
-      "smax v22.4s, v22.4s, v24.4s\n"
-      "smax v23.4s, v23.4s, v24.4s\n"
-      "uzp1 v16.8h, v16.8h, v17.8h\n"
-      "uzp1 v18.8h, v18.8h, v19.8h\n"
-      "uzp1 v20.8h, v20.8h, v21.8h\n"
-      "uzp1 v17.8h, v22.8h, v23.8h\n"
-      "uzp1 v16.16b, v16.16b, v18.16b\n"
-      "uzp1 v20.16b, v20.16b, v17.16b\n"
-      "bge 59f\n"
-      "tbz x9, #3, 54f\n"
-      "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "tbz x9, #2, 52f\n"
-      "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "tbz x9, #1, 51f\n"
-      "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "tbz x9, #0, 58f\n"
-      "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "b 58f\n"
-      "51:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x9, #0, 58f\n"
-      "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "b 58f\n"
-      "52:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x9, #1, 53f\n"
-      "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "tbz x9, #0, 58f\n"
-      "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "b 58f\n"
-      "53:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x9, #0, 58f\n"
-      "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "b 58f\n"
-      "54:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x9, #2, 56f\n"
-      "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "tbz x9, #1, 55f\n"
-      "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "tbz x9, #0, 58f\n"
-      "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "b 58f\n"
-      "55:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x9, #0, 58f\n"
-      "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "b 58f\n"
-      "56:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x9, #1, 57f\n"
-      "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "tbz x9, #0, 58f\n"
-      "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "b 58f\n"
-      "57:"  // Height 2: Partial direct writeback: partial_1_0
-      "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "58:"  // Height 2: Partial direct writeback: Done
-      "b 60f\n"
-      "59:"  // Height 2: Full writeback
-      "str q16, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "60:"  // Height 2: Writeback done
-      "subs x9, x9, #0x10\n"
-      "bgt 32b\n"
-      "b 122f\n"
-      "61:"  // Height 3
-      "mov x10, %x[col_bias]\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "movi v13.4s, #0x0\n"
-      "movi v15.16b, #0x1\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "62:"  // Height 3: Column loop
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "63:"  // Height 3: setup done
-      "mov x26, #0x0\n"
-      "64:"  // Height 3: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 65f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "ldr x22, [x20, #0x10]\n"
-      "cbnz x26, 66f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "b 66f\n"
-      "65:"  // Height 3: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "66:"  // Height 3: input setup done
-      "cmp x25, #0x10\n"
-      "blt 71f\n"
-      "ldr q0, [x24, #0x0]\n"
-      "ldr q1, [x23, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q2, [x22, #0x0]\n"
-      "ldr q4, [x28, #0x0]\n"
-      "ldr q5, [x28, #0x10]\n"
-      "ldr q6, [x28, #0x20]\n"
-      "ldr q7, [x28, #0x30]\n"
-      "ldr q8, [x28, #0x40]\n"
-      "ldr q9, [x28, #0x50]\n"
-      "ldr q10, [x28, #0x60]\n"
-      "blt 69f\n"
-      "67:"  // Height 3: Multiply loop: Main loop head
-      ".inst 0x4f00f090  // sudot v16.4s, v4.16b, v0.4b[0]\n"
-      ".inst 0x4f01f094  // sudot v20.4s, v4.16b, v1.4b[0]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f02f098  // sudot v24.4s, v4.16b, v2.4b[0]\n"
-      "ldr q29, [x28, #0x70]\n"
-      ".inst 0x4f00f0b1  // sudot v17.4s, v5.16b, v0.4b[0]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x4f01f0b5  // sudot v21.4s, v5.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0b9  // sudot v25.4s, v5.16b, v2.4b[0]\n"
-      "ldr q28, [x28, #0x80]\n"
-      ".inst 0x4f00f0d2  // sudot v18.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0d6  // sudot v22.4s, v6.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0da  // sudot v26.4s, v6.16b, v2.4b[0]\n"
-      "ldr q5, [x28, #0x90]\n"
-      ".inst 0x4f00f0f3  // sudot v19.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0f7  // sudot v23.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0fb  // sudot v27.4s, v7.16b, v2.4b[0]\n"
-      "ldr q4, [x28, #0xa0]\n"
-      ".inst 0x4f20f110  // sudot v16.4s, v8.16b, v0.4b[1]\n"
-      ".inst 0x4f21f114  // sudot v20.4s, v8.16b, v1.4b[1]\n"
-      ".inst 0x4f22f118  // sudot v24.4s, v8.16b, v2.4b[1]\n"
-      "ldr q3, [x28, #0xb0]\n"
-      ".inst 0x4f20f131  // sudot v17.4s, v9.16b, v0.4b[1]\n"
-      ".inst 0x4f21f135  // sudot v21.4s, v9.16b, v1.4b[1]\n"
-      ".inst 0x4f22f139  // sudot v25.4s, v9.16b, v2.4b[1]\n"
-      "ldr q31, [x28, #0xc0]\n"
-      ".inst 0x4f20f152  // sudot v18.4s, v10.16b, v0.4b[1]\n"
-      ".inst 0x4f21f156  // sudot v22.4s, v10.16b, v1.4b[1]\n"
-      ".inst 0x4f22f15a  // sudot v26.4s, v10.16b, v2.4b[1]\n"
-      "ldr q30, [x28, #0xd0]\n"
-      ".inst 0x4f20f3b3  // sudot v19.4s, v29.16b, v0.4b[1]\n"
-      ".inst 0x4f21f3b7  // sudot v23.4s, v29.16b, v1.4b[1]\n"
-      ".inst 0x4f22f3bb  // sudot v27.4s, v29.16b, v2.4b[1]\n"
-      "ldr q29, [x28, #0xe0]\n"
-      ".inst 0x4f00fb90  // sudot v16.4s, v28.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb94  // sudot v20.4s, v28.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb98  // sudot v24.4s, v28.16b, v2.4b[2]\n"
-      "ldr q28, [x28, #0xf0]\n"
-      ".inst 0x4f00f8b1  // sudot v17.4s, v5.16b, v0.4b[2]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4f01f8b5  // sudot v21.4s, v5.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8b9  // sudot v25.4s, v5.16b, v2.4b[2]\n"
-      ".inst 0x4f00f892  // sudot v18.4s, v4.16b, v0.4b[2]\n"
-      ".inst 0x4f01f896  // sudot v22.4s, v4.16b, v1.4b[2]\n"
-      ".inst 0x4f02f89a  // sudot v26.4s, v4.16b, v2.4b[2]\n"
-      ".inst 0x4f00f873  // sudot v19.4s, v3.16b, v0.4b[2]\n"
-      ".inst 0x4f01f877  // sudot v23.4s, v3.16b, v1.4b[2]\n"
-      ".inst 0x4f02f87b  // sudot v27.4s, v3.16b, v2.4b[2]\n"
-      ".inst 0x4f20fbf0  // sudot v16.4s, v31.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbf4  // sudot v20.4s, v31.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbf8  // sudot v24.4s, v31.16b, v2.4b[3]\n"
-      ".inst 0x4f20fbd1  // sudot v17.4s, v30.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbd5  // sudot v21.4s, v30.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbd9  // sudot v25.4s, v30.16b, v2.4b[3]\n"
-      ".inst 0x4f20fbb2  // sudot v18.4s, v29.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbb6  // sudot v22.4s, v29.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbba  // sudot v26.4s, v29.16b, v2.4b[3]\n"
-      ".inst 0x4f20fb93  // sudot v19.4s, v28.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb97  // sudot v23.4s, v28.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb9b  // sudot v27.4s, v28.16b, v2.4b[3]\n"
-      "tbnz %x[flags], #31, 68f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      "68:"  // Height 3: Multiply loop: unique 9: skip row sum
-      "ldr q0, [x24, #0x0]\n"
-      "ldr q1, [x23, #0x0]\n"
-      "sub x25, x25, #0x10\n"
-      "ldr q2, [x22, #0x0]\n"
-      "ldr q4, [x28, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q5, [x28, #0x10]\n"
-      "ldr q6, [x28, #0x20]\n"
-      "ldr q7, [x28, #0x30]\n"
-      "ldr q8, [x28, #0x40]\n"
-      "ldr q9, [x28, #0x50]\n"
-      "ldr q10, [x28, #0x60]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "bge 67b\n"
-      "69:"  // Height 3: Multiply loop: Single iteration only
-      ".inst 0x4f00f090  // sudot v16.4s, v4.16b, v0.4b[0]\n"
-      ".inst 0x4f01f094  // sudot v20.4s, v4.16b, v1.4b[0]\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f02f098  // sudot v24.4s, v4.16b, v2.4b[0]\n"
-      "ldr q29, [x28, #0x70]\n"
-      ".inst 0x4f00f0b1  // sudot v17.4s, v5.16b, v0.4b[0]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f01f0b5  // sudot v21.4s, v5.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0b9  // sudot v25.4s, v5.16b, v2.4b[0]\n"
-      "ldr q28, [x28, #0x80]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x4f00f0d2  // sudot v18.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0d6  // sudot v22.4s, v6.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0da  // sudot v26.4s, v6.16b, v2.4b[0]\n"
-      "ldr q5, [x28, #0x90]\n"
-      ".inst 0x4f00f0f3  // sudot v19.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0f7  // sudot v23.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0fb  // sudot v27.4s, v7.16b, v2.4b[0]\n"
-      "ldr q4, [x28, #0xa0]\n"
-      ".inst 0x4f20f110  // sudot v16.4s, v8.16b, v0.4b[1]\n"
-      ".inst 0x4f21f114  // sudot v20.4s, v8.16b, v1.4b[1]\n"
-      ".inst 0x4f22f118  // sudot v24.4s, v8.16b, v2.4b[1]\n"
-      "ldr q3, [x28, #0xb0]\n"
-      ".inst 0x4f20f131  // sudot v17.4s, v9.16b, v0.4b[1]\n"
-      ".inst 0x4f21f135  // sudot v21.4s, v9.16b, v1.4b[1]\n"
-      ".inst 0x4f22f139  // sudot v25.4s, v9.16b, v2.4b[1]\n"
-      "ldr q31, [x28, #0xc0]\n"
-      ".inst 0x4f20f152  // sudot v18.4s, v10.16b, v0.4b[1]\n"
-      ".inst 0x4f21f156  // sudot v22.4s, v10.16b, v1.4b[1]\n"
-      ".inst 0x4f22f15a  // sudot v26.4s, v10.16b, v2.4b[1]\n"
-      "ldr q30, [x28, #0xd0]\n"
-      ".inst 0x4f20f3b3  // sudot v19.4s, v29.16b, v0.4b[1]\n"
-      ".inst 0x4f21f3b7  // sudot v23.4s, v29.16b, v1.4b[1]\n"
-      ".inst 0x4f22f3bb  // sudot v27.4s, v29.16b, v2.4b[1]\n"
-      "ldr q29, [x28, #0xe0]\n"
-      ".inst 0x4f00fb90  // sudot v16.4s, v28.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb94  // sudot v20.4s, v28.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb98  // sudot v24.4s, v28.16b, v2.4b[2]\n"
-      "ldr q28, [x28, #0xf0]\n"
-      ".inst 0x4f00f8b1  // sudot v17.4s, v5.16b, v0.4b[2]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4f01f8b5  // sudot v21.4s, v5.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8b9  // sudot v25.4s, v5.16b, v2.4b[2]\n"
-      ".inst 0x4f00f892  // sudot v18.4s, v4.16b, v0.4b[2]\n"
-      ".inst 0x4f01f896  // sudot v22.4s, v4.16b, v1.4b[2]\n"
-      ".inst 0x4f02f89a  // sudot v26.4s, v4.16b, v2.4b[2]\n"
-      ".inst 0x4f00f873  // sudot v19.4s, v3.16b, v0.4b[2]\n"
-      ".inst 0x4f01f877  // sudot v23.4s, v3.16b, v1.4b[2]\n"
-      ".inst 0x4f02f87b  // sudot v27.4s, v3.16b, v2.4b[2]\n"
-      ".inst 0x4f20fbf0  // sudot v16.4s, v31.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbf4  // sudot v20.4s, v31.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbf8  // sudot v24.4s, v31.16b, v2.4b[3]\n"
-      ".inst 0x4f20fbd1  // sudot v17.4s, v30.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbd5  // sudot v21.4s, v30.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbd9  // sudot v25.4s, v30.16b, v2.4b[3]\n"
-      ".inst 0x4f20fbb2  // sudot v18.4s, v29.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbb6  // sudot v22.4s, v29.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbba  // sudot v26.4s, v29.16b, v2.4b[3]\n"
-      ".inst 0x4f20fb93  // sudot v19.4s, v28.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb97  // sudot v23.4s, v28.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb9b  // sudot v27.4s, v28.16b, v2.4b[3]\n"
-      "tbnz %x[flags], #31, 70f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      "70:"  // Height 3: Multiply loop: unique 10: skip row sum
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "71:"  // Height 3: Multiply loop: Main loop skip
-      "cbz x25, 78f\n"
-      "cmp x25, #0x4\n"
-      "blt 74f\n"
-      "72:"  // Height 3: Multiply loop: Odd block loop
-      "ldr s0, [x24], #0x4\n"
-      "ldr s1, [x23], #0x4\n"
-      "ldr s2, [x22], #0x4\n"
-      "tbnz %x[flags], #31, 73f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      "73:"  // Height 3: Multiply loop: unique 11: skip row sum
-      "ldr q31, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "sub x25, x25, #0x4\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f00f3f0  // sudot v16.4s, v31.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3f4  // sudot v20.4s, v31.16b, v1.4b[0]\n"
-      ".inst 0x4f02f3f8  // sudot v24.4s, v31.16b, v2.4b[0]\n"
-      ".inst 0x4f00f3d1  // sudot v17.4s, v30.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3d5  // sudot v21.4s, v30.16b, v1.4b[0]\n"
-      ".inst 0x4f02f3d9  // sudot v25.4s, v30.16b, v2.4b[0]\n"
-      ".inst 0x4f00f3b2  // sudot v18.4s, v29.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3b6  // sudot v22.4s, v29.16b, v1.4b[0]\n"
-      ".inst 0x4f02f3ba  // sudot v26.4s, v29.16b, v2.4b[0]\n"
-      ".inst 0x4f00f393  // sudot v19.4s, v28.16b, v0.4b[0]\n"
-      ".inst 0x4f01f397  // sudot v23.4s, v28.16b, v1.4b[0]\n"
-      ".inst 0x4f02f39b  // sudot v27.4s, v28.16b, v2.4b[0]\n"
-      "bge 72b\n"
-      "74:"  // Height 3: Multiply loop: Skip odd blocks
-      "cbz x25, 78f\n"
-      "tbz x25, #1, 75f\n"
-      "ldr h0, [x24], #0x2\n"
-      "ldr h1, [x23], #0x2\n"
-      "ldr h2, [x22], #0x2\n"
-      "tbz x25, #0, 76f\n"
-      "ld1 { v0.b }[2], [x24]\n"
-      "ld1 { v1.b }[2], [x23]\n"
-      "ld1 { v2.b }[2], [x22]\n"
-      "b 76f\n"
-      "75:"  // Height 3: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x24, #0x0]\n"
-      "ldr b1, [x23, #0x0]\n"
-      "ldr b2, [x22, #0x0]\n"
-      "76:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "tbnz %x[flags], #31, 77f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      "77:"  // Height 3: Multiply loop: unique 12: skip row sum
-      "ldr q31, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f00f3f0  // sudot v16.4s, v31.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3f4  // sudot v20.4s, v31.16b, v1.4b[0]\n"
-      ".inst 0x4f02f3f8  // sudot v24.4s, v31.16b, v2.4b[0]\n"
-      ".inst 0x4f00f3d1  // sudot v17.4s, v30.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3d5  // sudot v21.4s, v30.16b, v1.4b[0]\n"
-      ".inst 0x4f02f3d9  // sudot v25.4s, v30.16b, v2.4b[0]\n"
-      ".inst 0x4f00f3b2  // sudot v18.4s, v29.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3b6  // sudot v22.4s, v29.16b, v1.4b[0]\n"
-      ".inst 0x4f02f3ba  // sudot v26.4s, v29.16b, v2.4b[0]\n"
-      ".inst 0x4f00f393  // sudot v19.4s, v28.16b, v0.4b[0]\n"
-      ".inst 0x4f01f397  // sudot v23.4s, v28.16b, v1.4b[0]\n"
-      ".inst 0x4f02f39b  // sudot v27.4s, v28.16b, v2.4b[0]\n"
-      "78:"  // Height 3: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 64b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "tbnz %x[flags], #31, 79f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v28.4s }, [x20]\n"
-      "addp v13.4s, v13.4s, v13.4s\n"
-      "neg v28.4s, v28.4s\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "addp v12.4s, v12.4s, v12.4s\n"
-      "addp v13.4s, v13.4s, v13.4s\n"
-      "mul v11.4s, v11.4s, v28.4s\n"
-      "mul v12.4s, v12.4s, v28.4s\n"
-      "mul v13.4s, v13.4s, v28.4s\n"
-      "79:"  // Height 3: skip row sum fixup
-      "ldr q0, [x10, #0x0]\n"
-      "ldr q31, [x10, #0x10]\n"
-      "add v16.4s, v16.4s, v11.4s\n"
-      "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q30, [x10, #0x20]\n"
-      "ldr q29, [x10, #0x30]\n"
-      "add v18.4s, v18.4s, v11.4s\n"
-      "add v19.4s, v19.4s, v11.4s\n"
-      "add v20.4s, v20.4s, v12.4s\n"
-      "add v21.4s, v21.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "ld1r { v28.4s }, [x20]\n"
-      "add v22.4s, v22.4s, v12.4s\n"
-      "add v23.4s, v23.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add v24.4s, v24.4s, v13.4s\n"
-      "add v25.4s, v25.4s, v13.4s\n"
-      "add x10, x10, #0x40\n"
-      "add v26.4s, v26.4s, v13.4s\n"
-      "add v27.4s, v27.4s, v13.4s\n"
-      "add v16.4s, v16.4s, v0.4s\n"
-      "add v17.4s, v17.4s, v31.4s\n"
-      "add v18.4s, v18.4s, v30.4s\n"
-      "add v19.4s, v19.4s, v29.4s\n"
-      "add v20.4s, v20.4s, v0.4s\n"
-      "add v21.4s, v21.4s, v31.4s\n"
-      "add v22.4s, v22.4s, v30.4s\n"
-      "add v23.4s, v23.4s, v29.4s\n"
-      "add v24.4s, v24.4s, v0.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "add v25.4s, v25.4s, v31.4s\n"
-      "add v26.4s, v26.4s, v30.4s\n"
-      "add v27.4s, v27.4s, v29.4s\n"
-      "sqrdmulh v16.4s, v16.4s, v28.4s\n"
-      "sqrdmulh v17.4s, v17.4s, v28.4s\n"
-      "sqrdmulh v18.4s, v18.4s, v28.4s\n"
-      "sqrdmulh v19.4s, v19.4s, v28.4s\n"
-      "sqrdmulh v20.4s, v20.4s, v28.4s\n"
-      "sqrdmulh v21.4s, v21.4s, v28.4s\n"
-      "sqrdmulh v22.4s, v22.4s, v28.4s\n"
-      "sqrdmulh v23.4s, v23.4s, v28.4s\n"
-      "sqrdmulh v24.4s, v24.4s, v28.4s\n"
-      "sqrdmulh v25.4s, v25.4s, v28.4s\n"
-      "sqrdmulh v26.4s, v26.4s, v28.4s\n"
-      "sqrdmulh v27.4s, v27.4s, v28.4s\n"
-      "tbz %x[flags], #5, 80f\n"
-      "and v1.16b, v16.16b, v0.16b\n"
-      "and v31.16b, v17.16b, v0.16b\n"
-      "and v30.16b, v18.16b, v0.16b\n"
-      "and v29.16b, v19.16b, v0.16b\n"
-      "and v28.16b, v20.16b, v0.16b\n"
-      "and v3.16b, v21.16b, v0.16b\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "sshr v31.4s, v31.4s, #0x1f\n"
-      "sshr v30.4s, v30.4s, #0x1f\n"
-      "sshr v29.4s, v29.4s, #0x1f\n"
-      "sshr v28.4s, v28.4s, #0x1f\n"
-      "and v2.16b, v22.16b, v0.16b\n"
-      "sqadd v16.4s, v16.4s, v1.4s\n"
-      "sqadd v17.4s, v17.4s, v31.4s\n"
-      "sqadd v18.4s, v18.4s, v30.4s\n"
-      "sqadd v19.4s, v19.4s, v29.4s\n"
-      "sqadd v20.4s, v20.4s, v28.4s\n"
-      "and v1.16b, v23.16b, v0.16b\n"
-      "and v31.16b, v24.16b, v0.16b\n"
-      "and v30.16b, v25.16b, v0.16b\n"
-      "and v29.16b, v26.16b, v0.16b\n"
-      "and v28.16b, v27.16b, v0.16b\n"
-      "sshr v3.4s, v3.4s, #0x1f\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "sshr v31.4s, v31.4s, #0x1f\n"
-      "sshr v30.4s, v30.4s, #0x1f\n"
-      "sshr v29.4s, v29.4s, #0x1f\n"
-      "sshr v28.4s, v28.4s, #0x1f\n"
-      "sqadd v21.4s, v21.4s, v3.4s\n"
-      "sqadd v22.4s, v22.4s, v2.4s\n"
-      "sqadd v23.4s, v23.4s, v1.4s\n"
-      "sqadd v24.4s, v24.4s, v31.4s\n"
-      "sqadd v25.4s, v25.4s, v30.4s\n"
-      "sqadd v26.4s, v26.4s, v29.4s\n"
-      "sqadd v27.4s, v27.4s, v28.4s\n"
-      "80:"  // Height 3: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
-      "srshl v16.4s, v16.4s, v0.4s\n"
-      "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v30.4s }, [x21]\n"
-      "ld1r { v29.4s }, [x20]\n"
-      "srshl v18.4s, v18.4s, v0.4s\n"
-      "srshl v19.4s, v19.4s, v0.4s\n"
-      "srshl v20.4s, v20.4s, v0.4s\n"
-      "srshl v21.4s, v21.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v28.4s }, [x20]\n"
-      "srshl v22.4s, v22.4s, v0.4s\n"
-      "srshl v23.4s, v23.4s, v0.4s\n"
-      "srshl v24.4s, v24.4s, v0.4s\n"
-      "srshl v25.4s, v25.4s, v0.4s\n"
-      "srshl v26.4s, v26.4s, v0.4s\n"
-      "srshl v27.4s, v27.4s, v0.4s\n"
-      "add v16.4s, v16.4s, v30.4s\n"
-      "add v17.4s, v17.4s, v30.4s\n"
-      "add v18.4s, v18.4s, v30.4s\n"
-      "add v19.4s, v19.4s, v30.4s\n"
-      "add v20.4s, v20.4s, v30.4s\n"
-      "add v21.4s, v21.4s, v30.4s\n"
-      "add v22.4s, v22.4s, v30.4s\n"
-      "add v23.4s, v23.4s, v30.4s\n"
-      "add v24.4s, v24.4s, v30.4s\n"
-      "add v25.4s, v25.4s, v30.4s\n"
-      "add v26.4s, v26.4s, v30.4s\n"
-      "add v27.4s, v27.4s, v30.4s\n"
-      "smin v16.4s, v16.4s, v29.4s\n"
-      "smin v17.4s, v17.4s, v29.4s\n"
-      "smin v18.4s, v18.4s, v29.4s\n"
-      "smin v19.4s, v19.4s, v29.4s\n"
-      "smin v20.4s, v20.4s, v29.4s\n"
-      "smin v21.4s, v21.4s, v29.4s\n"
-      "smin v22.4s, v22.4s, v29.4s\n"
-      "smin v23.4s, v23.4s, v29.4s\n"
-      "smin v24.4s, v24.4s, v29.4s\n"
-      "smin v25.4s, v25.4s, v29.4s\n"
-      "smin v26.4s, v26.4s, v29.4s\n"
-      "smin v27.4s, v27.4s, v29.4s\n"
-      "smax v16.4s, v16.4s, v28.4s\n"
-      "smax v17.4s, v17.4s, v28.4s\n"
-      "smax v18.4s, v18.4s, v28.4s\n"
-      "smax v19.4s, v19.4s, v28.4s\n"
-      "smax v20.4s, v20.4s, v28.4s\n"
-      "smax v21.4s, v21.4s, v28.4s\n"
-      "smax v22.4s, v22.4s, v28.4s\n"
-      "smax v23.4s, v23.4s, v28.4s\n"
-      "smax v24.4s, v24.4s, v28.4s\n"
-      "smax v25.4s, v25.4s, v28.4s\n"
-      "smax v26.4s, v26.4s, v28.4s\n"
-      "smax v27.4s, v27.4s, v28.4s\n"
-      "uzp1 v16.8h, v16.8h, v17.8h\n"
-      "uzp1 v19.8h, v18.8h, v19.8h\n"
-      "uzp1 v20.8h, v20.8h, v21.8h\n"
-      "uzp1 v18.8h, v22.8h, v23.8h\n"
-      "uzp1 v24.8h, v24.8h, v25.8h\n"
-      "uzp1 v17.8h, v26.8h, v27.8h\n"
-      "uzp1 v16.16b, v16.16b, v19.16b\n"
-      "uzp1 v20.16b, v20.16b, v18.16b\n"
-      "uzp1 v24.16b, v24.16b, v17.16b\n"
-      "bge 89f\n"
-      "tbz x9, #3, 84f\n"
-      "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "tbz x9, #2, 82f\n"
-      "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "tbz x9, #1, 81f\n"
-      "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "tbz x9, #0, 88f\n"
-      "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "b 88f\n"
-      "81:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x9, #0, 88f\n"
-      "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "b 88f\n"
-      "82:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x9, #1, 83f\n"
-      "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "tbz x9, #0, 88f\n"
-      "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "b 88f\n"
-      "83:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x9, #0, 88f\n"
-      "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "b 88f\n"
-      "84:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x9, #2, 86f\n"
-      "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "tbz x9, #1, 85f\n"
-      "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "tbz x9, #0, 88f\n"
-      "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "b 88f\n"
-      "85:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x9, #0, 88f\n"
-      "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "b 88f\n"
-      "86:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x9, #1, 87f\n"
-      "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "tbz x9, #0, 88f\n"
-      "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "b 88f\n"
-      "87:"  // Height 3: Partial direct writeback: partial_1_0
-      "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "88:"  // Height 3: Partial direct writeback: Done
-      "b 90f\n"
-      "89:"  // Height 3: Full writeback
-      "str q16, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "90:"  // Height 3: Writeback done
-      "subs x9, x9, #0x10\n"
-      "bgt 62b\n"
-      "b 122f\n"
-      "91:"  // Height 4
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x20, #0x4\n"
-      "mov x10, %x[col_bias]\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
-      "movi v15.16b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "92:"  // Height 4: Column loop
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "movi v28.4s, #0x0\n"
-      "movi v29.4s, #0x0\n"
-      "movi v30.4s, #0x0\n"
-      "movi v31.4s, #0x0\n"
-      "93:"  // Height 4: setup done
-      "mov x26, #0x0\n"
-      "94:"  // Height 4: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 95f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "ldr x22, [x20, #0x10]\n"
-      "ldr x21, [x20, #0x18]\n"
-      "cbnz x26, 96f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "add x21, x21, x20\n"
-      "b 96f\n"
-      "95:"  // Height 4: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "add x21, x22, x21\n"
-      "96:"  // Height 4: input setup done
-      "cmp x25, #0x10\n"
-      "blt 101f\n"
-      "ldr q0, [x24, #0x0]\n"
-      "ldr q1, [x23, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q2, [x22, #0x0]\n"
-      "ldr q3, [x21, #0x0]\n"
-      "ldr q4, [x28, #0x0]\n"
-      "ldr q5, [x28, #0x10]\n"
-      "ldr q6, [x28, #0x20]\n"
-      "ldr q7, [x28, #0x30]\n"
-      "ldr q8, [x28, #0x40]\n"
-      "ldr q9, [x28, #0x50]\n"
-      "ldr q10, [x28, #0x60]\n"
-      "blt 99f\n"
-      "97:"  // Height 4: Multiply loop: Main loop head
-      ".inst 0x4f00f090  // sudot v16.4s, v4.16b, v0.4b[0]\n"
-      ".inst 0x4f01f094  // sudot v20.4s, v4.16b, v1.4b[0]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f02f098  // sudot v24.4s, v4.16b, v2.4b[0]\n"
-      ".inst 0x4f03f09c  // sudot v28.4s, v4.16b, v3.4b[0]\n"
-      "ldr q4, [x28, #0x70]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x4f00f0b1  // sudot v17.4s, v5.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0b5  // sudot v21.4s, v5.16b, v1.4b[0]\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x4f02f0b9  // sudot v25.4s, v5.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0bd  // sudot v29.4s, v5.16b, v3.4b[0]\n"
-      "ldr q5, [x28, #0x80]\n"
-      ".inst 0x4f00f0d2  // sudot v18.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0d6  // sudot v22.4s, v6.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0da  // sudot v26.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0de  // sudot v30.4s, v6.16b, v3.4b[0]\n"
-      "ldr q6, [x28, #0x90]\n"
-      ".inst 0x4f00f0f3  // sudot v19.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0f7  // sudot v23.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0fb  // sudot v27.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0ff  // sudot v31.4s, v7.16b, v3.4b[0]\n"
-      "ldr q7, [x28, #0xa0]\n"
-      ".inst 0x4f20f110  // sudot v16.4s, v8.16b, v0.4b[1]\n"
-      ".inst 0x4f21f114  // sudot v20.4s, v8.16b, v1.4b[1]\n"
-      ".inst 0x4f22f118  // sudot v24.4s, v8.16b, v2.4b[1]\n"
-      ".inst 0x4f23f11c  // sudot v28.4s, v8.16b, v3.4b[1]\n"
-      "ldr q8, [x28, #0xb0]\n"
-      ".inst 0x4f20f131  // sudot v17.4s, v9.16b, v0.4b[1]\n"
-      ".inst 0x4f21f135  // sudot v21.4s, v9.16b, v1.4b[1]\n"
-      ".inst 0x4f22f139  // sudot v25.4s, v9.16b, v2.4b[1]\n"
-      ".inst 0x4f23f13d  // sudot v29.4s, v9.16b, v3.4b[1]\n"
-      "ldr q9, [x28, #0xc0]\n"
-      ".inst 0x4f20f152  // sudot v18.4s, v10.16b, v0.4b[1]\n"
-      ".inst 0x4f21f156  // sudot v22.4s, v10.16b, v1.4b[1]\n"
-      ".inst 0x4f22f15a  // sudot v26.4s, v10.16b, v2.4b[1]\n"
-      ".inst 0x4f23f15e  // sudot v30.4s, v10.16b, v3.4b[1]\n"
-      "ldr q10, [x28, #0xd0]\n"
-      ".inst 0x4f20f093  // sudot v19.4s, v4.16b, v0.4b[1]\n"
-      ".inst 0x4f21f097  // sudot v23.4s, v4.16b, v1.4b[1]\n"
-      ".inst 0x4f22f09b  // sudot v27.4s, v4.16b, v2.4b[1]\n"
-      ".inst 0x4f23f09f  // sudot v31.4s, v4.16b, v3.4b[1]\n"
-      "ldr q4, [x28, #0xe0]\n"
-      ".inst 0x4f00f8b0  // sudot v16.4s, v5.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8b4  // sudot v20.4s, v5.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8b8  // sudot v24.4s, v5.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8bc  // sudot v28.4s, v5.16b, v3.4b[2]\n"
-      "ldr q5, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4f00f8d1  // sudot v17.4s, v6.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8d5  // sudot v21.4s, v6.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8d9  // sudot v25.4s, v6.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8dd  // sudot v29.4s, v6.16b, v3.4b[2]\n"
-      ".inst 0x4f00f8f2  // sudot v18.4s, v7.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8f6  // sudot v22.4s, v7.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8fa  // sudot v26.4s, v7.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8fe  // sudot v30.4s, v7.16b, v3.4b[2]\n"
-      ".inst 0x4f00f913  // sudot v19.4s, v8.16b, v0.4b[2]\n"
-      ".inst 0x4f01f917  // sudot v23.4s, v8.16b, v1.4b[2]\n"
-      ".inst 0x4f02f91b  // sudot v27.4s, v8.16b, v2.4b[2]\n"
-      ".inst 0x4f03f91f  // sudot v31.4s, v8.16b, v3.4b[2]\n"
-      ".inst 0x4f20f930  // sudot v16.4s, v9.16b, v0.4b[3]\n"
-      ".inst 0x4f21f934  // sudot v20.4s, v9.16b, v1.4b[3]\n"
-      ".inst 0x4f22f938  // sudot v24.4s, v9.16b, v2.4b[3]\n"
-      ".inst 0x4f23f93c  // sudot v28.4s, v9.16b, v3.4b[3]\n"
-      ".inst 0x4f20f951  // sudot v17.4s, v10.16b, v0.4b[3]\n"
-      ".inst 0x4f21f955  // sudot v21.4s, v10.16b, v1.4b[3]\n"
-      ".inst 0x4f22f959  // sudot v25.4s, v10.16b, v2.4b[3]\n"
-      ".inst 0x4f23f95d  // sudot v29.4s, v10.16b, v3.4b[3]\n"
-      ".inst 0x4f20f892  // sudot v18.4s, v4.16b, v0.4b[3]\n"
-      ".inst 0x4f21f896  // sudot v22.4s, v4.16b, v1.4b[3]\n"
-      ".inst 0x4f22f89a  // sudot v26.4s, v4.16b, v2.4b[3]\n"
-      ".inst 0x4f23f89e  // sudot v30.4s, v4.16b, v3.4b[3]\n"
-      ".inst 0x4f20f8b3  // sudot v19.4s, v5.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8b7  // sudot v23.4s, v5.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8bb  // sudot v27.4s, v5.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8bf  // sudot v31.4s, v5.16b, v3.4b[3]\n"
-      "tbnz %x[flags], #31, 98f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      ".inst 0x6e8f946e  // udot v14.4s, v3.16b, v15.16b\n"
-      "98:"  // Height 4: Multiply loop: unique 13: skip row sum
-      "ldr q0, [x24, #0x0]\n"
-      "ldr q1, [x23, #0x0]\n"
-      "sub x25, x25, #0x10\n"
-      "ldr q2, [x22, #0x0]\n"
-      "ldr q3, [x21, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q4, [x28, #0x0]\n"
-      "ldr q5, [x28, #0x10]\n"
-      "ldr q6, [x28, #0x20]\n"
-      "ldr q7, [x28, #0x30]\n"
-      "ldr q8, [x28, #0x40]\n"
-      "ldr q9, [x28, #0x50]\n"
-      "ldr q10, [x28, #0x60]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      "bge 97b\n"
-      "99:"  // Height 4: Multiply loop: Single iteration only
-      ".inst 0x4f00f090  // sudot v16.4s, v4.16b, v0.4b[0]\n"
-      ".inst 0x4f01f094  // sudot v20.4s, v4.16b, v1.4b[0]\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f02f098  // sudot v24.4s, v4.16b, v2.4b[0]\n"
-      ".inst 0x4f03f09c  // sudot v28.4s, v4.16b, v3.4b[0]\n"
-      "ldr q4, [x28, #0x70]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f00f0b1  // sudot v17.4s, v5.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0b5  // sudot v21.4s, v5.16b, v1.4b[0]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x4f02f0b9  // sudot v25.4s, v5.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0bd  // sudot v29.4s, v5.16b, v3.4b[0]\n"
-      "ldr q5, [x28, #0x80]\n"
-      ".inst 0x4f00f0d2  // sudot v18.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0d6  // sudot v22.4s, v6.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0da  // sudot v26.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0de  // sudot v30.4s, v6.16b, v3.4b[0]\n"
-      "ldr q6, [x28, #0x90]\n"
-      ".inst 0x4f00f0f3  // sudot v19.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0f7  // sudot v23.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0fb  // sudot v27.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0ff  // sudot v31.4s, v7.16b, v3.4b[0]\n"
-      "ldr q7, [x28, #0xa0]\n"
-      ".inst 0x4f20f110  // sudot v16.4s, v8.16b, v0.4b[1]\n"
-      ".inst 0x4f21f114  // sudot v20.4s, v8.16b, v1.4b[1]\n"
-      ".inst 0x4f22f118  // sudot v24.4s, v8.16b, v2.4b[1]\n"
-      ".inst 0x4f23f11c  // sudot v28.4s, v8.16b, v3.4b[1]\n"
-      "ldr q8, [x28, #0xb0]\n"
-      ".inst 0x4f20f131  // sudot v17.4s, v9.16b, v0.4b[1]\n"
-      ".inst 0x4f21f135  // sudot v21.4s, v9.16b, v1.4b[1]\n"
-      ".inst 0x4f22f139  // sudot v25.4s, v9.16b, v2.4b[1]\n"
-      ".inst 0x4f23f13d  // sudot v29.4s, v9.16b, v3.4b[1]\n"
-      "ldr q9, [x28, #0xc0]\n"
-      ".inst 0x4f20f152  // sudot v18.4s, v10.16b, v0.4b[1]\n"
-      ".inst 0x4f21f156  // sudot v22.4s, v10.16b, v1.4b[1]\n"
-      ".inst 0x4f22f15a  // sudot v26.4s, v10.16b, v2.4b[1]\n"
-      ".inst 0x4f23f15e  // sudot v30.4s, v10.16b, v3.4b[1]\n"
-      "ldr q10, [x28, #0xd0]\n"
-      ".inst 0x4f20f093  // sudot v19.4s, v4.16b, v0.4b[1]\n"
-      ".inst 0x4f21f097  // sudot v23.4s, v4.16b, v1.4b[1]\n"
-      ".inst 0x4f22f09b  // sudot v27.4s, v4.16b, v2.4b[1]\n"
-      ".inst 0x4f23f09f  // sudot v31.4s, v4.16b, v3.4b[1]\n"
-      "ldr q4, [x28, #0xe0]\n"
-      ".inst 0x4f00f8b0  // sudot v16.4s, v5.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8b4  // sudot v20.4s, v5.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8b8  // sudot v24.4s, v5.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8bc  // sudot v28.4s, v5.16b, v3.4b[2]\n"
-      "ldr q5, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4f00f8d1  // sudot v17.4s, v6.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8d5  // sudot v21.4s, v6.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8d9  // sudot v25.4s, v6.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8dd  // sudot v29.4s, v6.16b, v3.4b[2]\n"
-      ".inst 0x4f00f8f2  // sudot v18.4s, v7.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8f6  // sudot v22.4s, v7.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8fa  // sudot v26.4s, v7.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8fe  // sudot v30.4s, v7.16b, v3.4b[2]\n"
-      ".inst 0x4f00f913  // sudot v19.4s, v8.16b, v0.4b[2]\n"
-      ".inst 0x4f01f917  // sudot v23.4s, v8.16b, v1.4b[2]\n"
-      ".inst 0x4f02f91b  // sudot v27.4s, v8.16b, v2.4b[2]\n"
-      ".inst 0x4f03f91f  // sudot v31.4s, v8.16b, v3.4b[2]\n"
-      ".inst 0x4f20f930  // sudot v16.4s, v9.16b, v0.4b[3]\n"
-      ".inst 0x4f21f934  // sudot v20.4s, v9.16b, v1.4b[3]\n"
-      ".inst 0x4f22f938  // sudot v24.4s, v9.16b, v2.4b[3]\n"
-      ".inst 0x4f23f93c  // sudot v28.4s, v9.16b, v3.4b[3]\n"
-      ".inst 0x4f20f951  // sudot v17.4s, v10.16b, v0.4b[3]\n"
-      ".inst 0x4f21f955  // sudot v21.4s, v10.16b, v1.4b[3]\n"
-      ".inst 0x4f22f959  // sudot v25.4s, v10.16b, v2.4b[3]\n"
-      ".inst 0x4f23f95d  // sudot v29.4s, v10.16b, v3.4b[3]\n"
-      ".inst 0x4f20f892  // sudot v18.4s, v4.16b, v0.4b[3]\n"
-      ".inst 0x4f21f896  // sudot v22.4s, v4.16b, v1.4b[3]\n"
-      ".inst 0x4f22f89a  // sudot v26.4s, v4.16b, v2.4b[3]\n"
-      ".inst 0x4f23f89e  // sudot v30.4s, v4.16b, v3.4b[3]\n"
-      ".inst 0x4f20f8b3  // sudot v19.4s, v5.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8b7  // sudot v23.4s, v5.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8bb  // sudot v27.4s, v5.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8bf  // sudot v31.4s, v5.16b, v3.4b[3]\n"
-      "tbnz %x[flags], #31, 100f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      ".inst 0x6e8f946e  // udot v14.4s, v3.16b, v15.16b\n"
-      "100:"  // Height 4: Multiply loop: unique 14: skip row sum
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      "101:"  // Height 4: Multiply loop: Main loop skip
-      "cbz x25, 108f\n"
-      "cmp x25, #0x4\n"
-      "blt 104f\n"
-      "102:"  // Height 4: Multiply loop: Odd block loop
-      "ldr s0, [x24], #0x4\n"
-      "ldr s1, [x23], #0x4\n"
-      "ldr s2, [x22], #0x4\n"
-      "ldr s3, [x21], #0x4\n"
-      "tbnz %x[flags], #31, 103f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      ".inst 0x6e8f946e  // udot v14.4s, v3.16b, v15.16b\n"
-      "103:"  // Height 4: Multiply loop: unique 15: skip row sum
-      "ldr q7, [x28, #0x0]\n"
-      "ldr q6, [x28, #0x10]\n"
-      "sub x25, x25, #0x4\n"
-      "ldr q5, [x28, #0x20]\n"
-      "ldr q4, [x28, #0x30]\n"
-      "cmp x25, #0x4\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f00f0f0  // sudot v16.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0f4  // sudot v20.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f8  // sudot v24.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0fc  // sudot v28.4s, v7.16b, v3.4b[0]\n"
-      ".inst 0x4f00f0d1  // sudot v17.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0d5  // sudot v21.4s, v6.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0d9  // sudot v25.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0dd  // sudot v29.4s, v6.16b, v3.4b[0]\n"
-      ".inst 0x4f00f0b2  // sudot v18.4s, v5.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0b6  // sudot v22.4s, v5.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0ba  // sudot v26.4s, v5.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0be  // sudot v30.4s, v5.16b, v3.4b[0]\n"
-      ".inst 0x4f00f093  // sudot v19.4s, v4.16b, v0.4b[0]\n"
-      ".inst 0x4f01f097  // sudot v23.4s, v4.16b, v1.4b[0]\n"
-      ".inst 0x4f02f09b  // sudot v27.4s, v4.16b, v2.4b[0]\n"
-      ".inst 0x4f03f09f  // sudot v31.4s, v4.16b, v3.4b[0]\n"
-      "bge 102b\n"
-      "104:"  // Height 4: Multiply loop: Skip odd blocks
-      "cbz x25, 108f\n"
-      "tbz x25, #1, 105f\n"
-      "ldr h0, [x24], #0x2\n"
-      "ldr h1, [x23], #0x2\n"
-      "ldr h2, [x22], #0x2\n"
-      "ldr h3, [x21], #0x2\n"
-      "tbz x25, #0, 106f\n"
-      "ld1 { v0.b }[2], [x24]\n"
-      "ld1 { v1.b }[2], [x23]\n"
-      "ld1 { v2.b }[2], [x22]\n"
-      "ld1 { v3.b }[2], [x21]\n"
-      "b 106f\n"
-      "105:"  // Height 4: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x24, #0x0]\n"
-      "ldr b1, [x23, #0x0]\n"
-      "ldr b2, [x22, #0x0]\n"
-      "ldr b3, [x21, #0x0]\n"
-      "106:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "tbnz %x[flags], #31, 107f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942c  // udot v12.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      ".inst 0x6e8f946e  // udot v14.4s, v3.16b, v15.16b\n"
-      "107:"  // Height 4: Multiply loop: unique 16: skip row sum
-      "ldr q7, [x28, #0x0]\n"
-      "ldr q6, [x28, #0x10]\n"
-      "ldr q5, [x28, #0x20]\n"
-      "ldr q4, [x28, #0x30]\n"
-      "add x28, x28, #0x40\n"
-      ".inst 0x4f00f0f0  // sudot v16.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0f4  // sudot v20.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f8  // sudot v24.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0fc  // sudot v28.4s, v7.16b, v3.4b[0]\n"
-      ".inst 0x4f00f0d1  // sudot v17.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0d5  // sudot v21.4s, v6.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0d9  // sudot v25.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0dd  // sudot v29.4s, v6.16b, v3.4b[0]\n"
-      ".inst 0x4f00f0b2  // sudot v18.4s, v5.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0b6  // sudot v22.4s, v5.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0ba  // sudot v26.4s, v5.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0be  // sudot v30.4s, v5.16b, v3.4b[0]\n"
-      ".inst 0x4f00f093  // sudot v19.4s, v4.16b, v0.4b[0]\n"
-      ".inst 0x4f01f097  // sudot v23.4s, v4.16b, v1.4b[0]\n"
-      ".inst 0x4f02f09b  // sudot v27.4s, v4.16b, v2.4b[0]\n"
-      ".inst 0x4f03f09f  // sudot v31.4s, v4.16b, v3.4b[0]\n"
-      "108:"  // Height 4: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 94b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "add x24, x27, x20\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "add x22, x23, x20\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "tbnz %x[flags], #31, 109f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "addp v12.4s, v12.4s, v12.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "addp v13.4s, v13.4s, v13.4s\n"
-      "addp v14.4s, v14.4s, v14.4s\n"
-      "neg v0.4s, v0.4s\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "addp v12.4s, v12.4s, v12.4s\n"
-      "addp v13.4s, v13.4s, v13.4s\n"
-      "addp v14.4s, v14.4s, v14.4s\n"
-      "mul v11.4s, v11.4s, v0.4s\n"
-      "mul v12.4s, v12.4s, v0.4s\n"
-      "mul v13.4s, v13.4s, v0.4s\n"
-      "mul v14.4s, v14.4s, v0.4s\n"
-      "109:"  // Height 4: skip row sum fixup
-      "ldr q0, [x10, #0x0]\n"
-      "ldr q4, [x10, #0x10]\n"
-      "add v16.4s, v16.4s, v11.4s\n"
-      "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q3, [x10, #0x20]\n"
-      "ldr q2, [x10, #0x30]\n"
-      "add v18.4s, v18.4s, v11.4s\n"
-      "add v19.4s, v19.4s, v11.4s\n"
-      "add v20.4s, v20.4s, v12.4s\n"
-      "add v21.4s, v21.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "ld1r { v1.4s }, [x20]\n"
-      "add v22.4s, v22.4s, v12.4s\n"
-      "add v23.4s, v23.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add v24.4s, v24.4s, v13.4s\n"
-      "add v25.4s, v25.4s, v13.4s\n"
-      "add x10, x10, #0x40\n"
-      "add v26.4s, v26.4s, v13.4s\n"
-      "add v27.4s, v27.4s, v13.4s\n"
-      "add v28.4s, v28.4s, v14.4s\n"
-      "add v29.4s, v29.4s, v14.4s\n"
-      "add v30.4s, v30.4s, v14.4s\n"
-      "add v31.4s, v31.4s, v14.4s\n"
-      "add v16.4s, v16.4s, v0.4s\n"
-      "add v17.4s, v17.4s, v4.4s\n"
-      "add v18.4s, v18.4s, v3.4s\n"
-      "add v19.4s, v19.4s, v2.4s\n"
-      "add v20.4s, v20.4s, v0.4s\n"
-      "add v21.4s, v21.4s, v4.4s\n"
-      "add v22.4s, v22.4s, v3.4s\n"
-      "add v23.4s, v23.4s, v2.4s\n"
-      "add v24.4s, v24.4s, v0.4s\n"
-      "add v25.4s, v25.4s, v4.4s\n"
-      "add v26.4s, v26.4s, v3.4s\n"
-      "add v27.4s, v27.4s, v2.4s\n"
-      "add v28.4s, v28.4s, v0.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "add v29.4s, v29.4s, v4.4s\n"
-      "add v30.4s, v30.4s, v3.4s\n"
-      "add v31.4s, v31.4s, v2.4s\n"
-      "sqrdmulh v16.4s, v16.4s, v1.4s\n"
-      "sqrdmulh v17.4s, v17.4s, v1.4s\n"
-      "sqrdmulh v18.4s, v18.4s, v1.4s\n"
-      "sqrdmulh v19.4s, v19.4s, v1.4s\n"
-      "sqrdmulh v20.4s, v20.4s, v1.4s\n"
-      "sqrdmulh v21.4s, v21.4s, v1.4s\n"
-      "sqrdmulh v22.4s, v22.4s, v1.4s\n"
-      "sqrdmulh v23.4s, v23.4s, v1.4s\n"
-      "sqrdmulh v24.4s, v24.4s, v1.4s\n"
-      "sqrdmulh v25.4s, v25.4s, v1.4s\n"
-      "sqrdmulh v26.4s, v26.4s, v1.4s\n"
-      "sqrdmulh v27.4s, v27.4s, v1.4s\n"
-      "sqrdmulh v28.4s, v28.4s, v1.4s\n"
-      "sqrdmulh v29.4s, v29.4s, v1.4s\n"
-      "sqrdmulh v30.4s, v30.4s, v1.4s\n"
-      "sqrdmulh v31.4s, v31.4s, v1.4s\n"
-      "tbz %x[flags], #5, 110f\n"
-      "and v2.16b, v16.16b, v0.16b\n"
-      "and v1.16b, v17.16b, v0.16b\n"
-      "and v7.16b, v18.16b, v0.16b\n"
-      "and v6.16b, v19.16b, v0.16b\n"
-      "and v5.16b, v20.16b, v0.16b\n"
-      "and v4.16b, v21.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "and v3.16b, v22.16b, v0.16b\n"
-      "sshr v7.4s, v7.4s, #0x1f\n"
-      "sshr v6.4s, v6.4s, #0x1f\n"
-      "sshr v5.4s, v5.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v2.4s\n"
-      "sqadd v17.4s, v17.4s, v1.4s\n"
-      "and v2.16b, v23.16b, v0.16b\n"
-      "and v1.16b, v24.16b, v0.16b\n"
-      "sshr v4.4s, v4.4s, #0x1f\n"
-      "sshr v3.4s, v3.4s, #0x1f\n"
-      "sqadd v18.4s, v18.4s, v7.4s\n"
-      "sqadd v19.4s, v19.4s, v6.4s\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "sqadd v20.4s, v20.4s, v5.4s\n"
-      "sqadd v21.4s, v21.4s, v4.4s\n"
-      "sqadd v22.4s, v22.4s, v3.4s\n"
-      "and v7.16b, v25.16b, v0.16b\n"
-      "sqadd v23.4s, v23.4s, v2.4s\n"
-      "sqadd v24.4s, v24.4s, v1.4s\n"
-      "and v6.16b, v26.16b, v0.16b\n"
-      "and v5.16b, v27.16b, v0.16b\n"
-      "and v4.16b, v28.16b, v0.16b\n"
-      "and v3.16b, v29.16b, v0.16b\n"
-      "and v2.16b, v30.16b, v0.16b\n"
-      "and v1.16b, v31.16b, v0.16b\n"
-      "sshr v7.4s, v7.4s, #0x1f\n"
-      "sshr v6.4s, v6.4s, #0x1f\n"
-      "sshr v5.4s, v5.4s, #0x1f\n"
-      "sshr v4.4s, v4.4s, #0x1f\n"
-      "sshr v3.4s, v3.4s, #0x1f\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "sqadd v25.4s, v25.4s, v7.4s\n"
-      "sqadd v26.4s, v26.4s, v6.4s\n"
-      "sqadd v27.4s, v27.4s, v5.4s\n"
-      "sqadd v28.4s, v28.4s, v4.4s\n"
-      "sqadd v29.4s, v29.4s, v3.4s\n"
-      "sqadd v30.4s, v30.4s, v2.4s\n"
-      "sqadd v31.4s, v31.4s, v1.4s\n"
-      "110:"  // Height 4: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
-      "srshl v16.4s, v16.4s, v0.4s\n"
-      "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v3.4s }, [x21]\n"
-      "ld1r { v2.4s }, [x20]\n"
-      "srshl v18.4s, v18.4s, v0.4s\n"
-      "srshl v19.4s, v19.4s, v0.4s\n"
-      "srshl v20.4s, v20.4s, v0.4s\n"
-      "srshl v21.4s, v21.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v1.4s }, [x20]\n"
-      "srshl v22.4s, v22.4s, v0.4s\n"
-      "srshl v23.4s, v23.4s, v0.4s\n"
-      "srshl v24.4s, v24.4s, v0.4s\n"
-      "srshl v25.4s, v25.4s, v0.4s\n"
-      "srshl v26.4s, v26.4s, v0.4s\n"
-      "srshl v27.4s, v27.4s, v0.4s\n"
-      "srshl v28.4s, v28.4s, v0.4s\n"
-      "srshl v29.4s, v29.4s, v0.4s\n"
-      "srshl v30.4s, v30.4s, v0.4s\n"
-      "srshl v31.4s, v31.4s, v0.4s\n"
-      "add v16.4s, v16.4s, v3.4s\n"
-      "add v17.4s, v17.4s, v3.4s\n"
-      "add v18.4s, v18.4s, v3.4s\n"
-      "add v19.4s, v19.4s, v3.4s\n"
-      "add v20.4s, v20.4s, v3.4s\n"
-      "add v21.4s, v21.4s, v3.4s\n"
-      "add v22.4s, v22.4s, v3.4s\n"
-      "add v23.4s, v23.4s, v3.4s\n"
-      "add v24.4s, v24.4s, v3.4s\n"
-      "add v25.4s, v25.4s, v3.4s\n"
-      "add v26.4s, v26.4s, v3.4s\n"
-      "add v27.4s, v27.4s, v3.4s\n"
-      "add v28.4s, v28.4s, v3.4s\n"
-      "add v29.4s, v29.4s, v3.4s\n"
-      "add v30.4s, v30.4s, v3.4s\n"
-      "add v31.4s, v31.4s, v3.4s\n"
-      "smin v16.4s, v16.4s, v2.4s\n"
-      "smin v17.4s, v17.4s, v2.4s\n"
-      "smin v18.4s, v18.4s, v2.4s\n"
-      "smin v19.4s, v19.4s, v2.4s\n"
-      "smin v20.4s, v20.4s, v2.4s\n"
-      "smin v21.4s, v21.4s, v2.4s\n"
-      "smin v22.4s, v22.4s, v2.4s\n"
-      "smin v23.4s, v23.4s, v2.4s\n"
-      "smin v24.4s, v24.4s, v2.4s\n"
-      "smin v25.4s, v25.4s, v2.4s\n"
-      "smin v26.4s, v26.4s, v2.4s\n"
-      "smin v27.4s, v27.4s, v2.4s\n"
-      "smin v28.4s, v28.4s, v2.4s\n"
-      "smin v29.4s, v29.4s, v2.4s\n"
-      "smin v30.4s, v30.4s, v2.4s\n"
-      "smin v31.4s, v31.4s, v2.4s\n"
-      "smax v16.4s, v16.4s, v1.4s\n"
-      "smax v17.4s, v17.4s, v1.4s\n"
-      "smax v18.4s, v18.4s, v1.4s\n"
-      "smax v19.4s, v19.4s, v1.4s\n"
-      "smax v20.4s, v20.4s, v1.4s\n"
-      "smax v21.4s, v21.4s, v1.4s\n"
-      "smax v22.4s, v22.4s, v1.4s\n"
-      "smax v23.4s, v23.4s, v1.4s\n"
-      "smax v24.4s, v24.4s, v1.4s\n"
-      "smax v25.4s, v25.4s, v1.4s\n"
-      "smax v26.4s, v26.4s, v1.4s\n"
-      "smax v27.4s, v27.4s, v1.4s\n"
-      "smax v28.4s, v28.4s, v1.4s\n"
-      "smax v29.4s, v29.4s, v1.4s\n"
-      "smax v30.4s, v30.4s, v1.4s\n"
-      "smax v31.4s, v31.4s, v1.4s\n"
-      "uzp1 v16.8h, v16.8h, v17.8h\n"
-      "uzp1 v0.8h, v18.8h, v19.8h\n"
-      "uzp1 v20.8h, v20.8h, v21.8h\n"
-      "uzp1 v19.8h, v22.8h, v23.8h\n"
-      "uzp1 v24.8h, v24.8h, v25.8h\n"
-      "uzp1 v18.8h, v26.8h, v27.8h\n"
-      "uzp1 v28.8h, v28.8h, v29.8h\n"
-      "uzp1 v17.8h, v30.8h, v31.8h\n"
-      "uzp1 v16.16b, v16.16b, v0.16b\n"
-      "uzp1 v20.16b, v20.16b, v19.16b\n"
-      "uzp1 v24.16b, v24.16b, v18.16b\n"
-      "uzp1 v28.16b, v28.16b, v17.16b\n"
-      "bge 119f\n"
-      "tbz x9, #3, 114f\n"
-      "str d16, [x27], #0x8\n"
-      "str d20, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "str d28, [x22], #0x8\n"
-      "tbz x9, #2, 112f\n"
-      "st1 { v16.s }[2], [x27], #0x4\n"
-      "st1 { v20.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "st1 { v28.s }[2], [x22], #0x4\n"
-      "tbz x9, #1, 111f\n"
-      "st1 { v16.h }[6], [x27], #0x2\n"
-      "st1 { v20.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "st1 { v28.h }[6], [x22], #0x2\n"
-      "tbz x9, #0, 118f\n"
-      "st1 { v16.b }[14], [x27]\n"
-      "st1 { v20.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "st1 { v28.b }[14], [x22]\n"
-      "b 118f\n"
-      "111:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x9, #0, 118f\n"
-      "st1 { v16.b }[12], [x27]\n"
-      "st1 { v20.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "st1 { v28.b }[12], [x22]\n"
-      "b 118f\n"
-      "112:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x9, #1, 113f\n"
-      "st1 { v16.h }[4], [x27], #0x2\n"
-      "st1 { v20.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "st1 { v28.h }[4], [x22], #0x2\n"
-      "tbz x9, #0, 118f\n"
-      "st1 { v16.b }[10], [x27]\n"
-      "st1 { v20.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "st1 { v28.b }[10], [x22]\n"
-      "b 118f\n"
-      "113:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x9, #0, 118f\n"
-      "st1 { v16.b }[8], [x27]\n"
-      "st1 { v20.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "st1 { v28.b }[8], [x22]\n"
-      "b 118f\n"
-      "114:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x9, #2, 116f\n"
-      "str s16, [x27], #0x4\n"
-      "str s20, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "str s28, [x22], #0x4\n"
-      "tbz x9, #1, 115f\n"
-      "st1 { v16.h }[2], [x27], #0x2\n"
-      "st1 { v20.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "st1 { v28.h }[2], [x22], #0x2\n"
-      "tbz x9, #0, 118f\n"
-      "st1 { v16.b }[6], [x27]\n"
-      "st1 { v20.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "st1 { v28.b }[6], [x22]\n"
-      "b 118f\n"
-      "115:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x9, #0, 118f\n"
-      "st1 { v16.b }[4], [x27]\n"
-      "st1 { v20.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "st1 { v28.b }[4], [x22]\n"
-      "b 118f\n"
-      "116:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x9, #1, 117f\n"
-      "str h16, [x27], #0x2\n"
-      "str h20, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "str h28, [x22], #0x2\n"
-      "tbz x9, #0, 118f\n"
-      "st1 { v16.b }[2], [x27]\n"
-      "st1 { v20.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "st1 { v28.b }[2], [x22]\n"
-      "b 118f\n"
-      "117:"  // Height 4: Partial direct writeback: partial_1_0
-      "str b16, [x27, #0x0]\n"
-      "str b20, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "str b28, [x22, #0x0]\n"
-      "118:"  // Height 4: Partial direct writeback: Done
-      "b 120f\n"
-      "119:"  // Height 4: Full writeback
-      "str q16, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "str q20, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "str q28, [x22, #0x0]\n"
-      "120:"  // Height 4: Writeback done
-      "subs x9, x9, #0x10\n"
-      "bgt 92b\n"
-      "subs %x[M], %x[M], #0x4\n"
-      "beq 122f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "tbz %x[flags], #3, 121f\n"
-      "add x21, x21, #0x4\n"
-      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "b 1b\n"
-      "121:"  // Update direct input
-      "mov x20, #0x4\n"
-      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
-      "b 1b\n"
-      "122:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
-    );
-}
-
-} // namespace arm_gemm
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_mmla_4x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_mmla_4x16.hpp
deleted file mode 100644
index ee1297fc7b..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_mmla_4x16.hpp
+++ /dev/null
@@ -1,101 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef __aarch64__
-
-#include "../std_transforms_fixed.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    unsigned int, const unsigned int *, \
-    IndirectInputArg<uint8_t>, \
-    size_t, size_t, \
-    const int8_t *, \
-    IndirectOutputArg<uint8_t>, \
-    const Requantize32 *, const int32_t *, unsigned int
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void a64_hybrid_u8s8qa_mmla_4x16( ARGLIST );
-
-class cls_a64_hybrid_u8s8qa_mmla_4x16
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef uint8_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 4;
-    }
-
-    static unsigned int out_width()
-    {
-        return 16;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 8;
-    }
-
-    static constexpr bool supports_accumulate()
-    {
-        return false;
-    }
-
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 4, 16, 8> transforms = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 47.74 };
-                case CPUModel::A510:
-                    return { 27.99 };
-                case CPUModel::V1:
-                    return { 62.26 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=a64_hybrid_u8s8qa_mmla_4x16;
-    cls_a64_hybrid_u8s8qa_mmla_4x16(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_mmla_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_mmla_4x16/generic.cpp
deleted file mode 100644
index 00b9db05c0..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8qa_mmla_4x16/generic.cpp
+++ /dev/null
@@ -1,2099 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef __aarch64__
-
-#include "arm_gemm.hpp"
-#include "../../utils.hpp"
-
-#include <cassert>
-#include <cstdint>
-
-namespace arm_gemm {
-
-void a64_hybrid_u8s8qa_mmla_4x16 (
-    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg,
-    size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<uint8_t> output_arg,
-    const Requantize32 *qp, const int32_t *col_bias, unsigned int
-)
-{
-    struct KernelArgs {
-        unsigned int num_strings = {};
-        const unsigned int *string_lengths = {};
-        size_t N = {};
-        const int8_t *B_ptr = {};
-        size_t output_offset = {};
-        size_t input_initial_col = {};
-        size_t input_offset = {};
-        void *output_ptr = {};
-    } ka;
-
-    unsigned long flags=0;
-    void *input_ptr;
-
-    if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
-        ka.output_offset=output_arg.indirect.offset;
-        flags |= 0x4;
-    } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
-        ka.output_offset=output_arg.direct.stride;
-    }
-
-    if (A_arg.is_indirect) {
-        input_ptr=(void *)(A_arg.indirect.ptr);
-        ka.input_offset=A_arg.indirect.start_row;
-        ka.input_initial_col=A_arg.indirect.start_col;
-        flags |= 0x8;
-    } else {
-        assert(num_strings==1);
-        input_ptr=(void *)(A_arg.direct.base);
-        ka.input_offset=A_arg.direct.stride;
-    }
-    ka.num_strings = num_strings;
-    ka.string_lengths = string_lengths;
-    ka.N = N;
-    ka.B_ptr = B_ptr;
-    if (qp->c_offset > qp->minval) {
-        flags |= 0x20;
-    }
-    __asm__ __volatile__(
-      "1:"  // Row loop
-      "cmp %x[M], #0x4\n"
-      "bge 97f\n"
-      "cmp %x[M], #0x2\n"
-      "bgt 65f\n"
-      "beq 33f\n"
-      "mov x10, %x[col_bias]\n"
-      "movi v11.4s, #0x0\n"
-      "movi v15.16b, #0x1\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "2:"  // Height 1: Column loop
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "3:"  // Height 1: setup done
-      "mov x26, #0x0\n"
-      "4:"  // Height 1: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 5f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "cbnz x26, 6f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "b 6f\n"
-      "5:"  // Height 1: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "6:"  // Height 1: input setup done
-      "cmp x25, #0x10\n"
-      "blt 11f\n"
-      "ldr q1, [x24, #0x0]\n"
-      "ldr q5, [x28, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q6, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q8, [x28, #0x30]\n"
-      "ldr q9, [x28, #0x40]\n"
-      "ldr q10, [x28, #0x50]\n"
-      "ldr q4, [x28, #0x60]\n"
-      "blt 9f\n"
-      "7:"  // Height 1: Multiply loop: Main loop head
-      "add x24, x24, #0x10\n"
-      "trn1 v0.2d, v1.2d, v27.2d\n"
-      ".inst 0x4e85ac10  // usmmla v16.4s, v0.16b, v5.16b\n"
-      "ldr q25, [x28, #0x70]\n"
-      "trn2 v1.2d, v1.2d, v27.2d\n"
-      ".inst 0x4e86ac14  // usmmla v20.4s, v0.16b, v6.16b\n"
-      "ldr q24, [x28, #0x80]\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      "ldr q30, [x28, #0x90]\n"
-      ".inst 0x4e88ac15  // usmmla v21.4s, v0.16b, v8.16b\n"
-      "ldr q29, [x28, #0xa0]\n"
-      ".inst 0x4e89ac12  // usmmla v18.4s, v0.16b, v9.16b\n"
-      "ldr q28, [x28, #0xb0]\n"
-      ".inst 0x4e8aac16  // usmmla v22.4s, v0.16b, v10.16b\n"
-      "ldr q27, [x28, #0xc0]\n"
-      ".inst 0x4e84ac13  // usmmla v19.4s, v0.16b, v4.16b\n"
-      "ldr q26, [x28, #0xd0]\n"
-      ".inst 0x4e99ac17  // usmmla v23.4s, v0.16b, v25.16b\n"
-      "ldr q25, [x28, #0xe0]\n"
-      ".inst 0x4e98ac30  // usmmla v16.4s, v1.16b, v24.16b\n"
-      "ldr q24, [x28, #0xf0]\n"
-      ".inst 0x4e9eac34  // usmmla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4e9dac31  // usmmla v17.4s, v1.16b, v29.16b\n"
-      ".inst 0x4e9cac35  // usmmla v21.4s, v1.16b, v28.16b\n"
-      ".inst 0x4e9bac32  // usmmla v18.4s, v1.16b, v27.16b\n"
-      ".inst 0x4e9aac36  // usmmla v22.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e99ac33  // usmmla v19.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e98ac37  // usmmla v23.4s, v1.16b, v24.16b\n"
-      "tbnz %x[flags], #31, 8f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942b  // udot v11.4s, v1.16b, v15.16b\n"
-      "8:"  // Height 1: Multiply loop: unique 1: skip row sum
-      "ldr q1, [x24, #0x0]\n"
-      "ldr q5, [x28, #0x0]\n"
-      "sub x25, x25, #0x10\n"
-      "ldr q6, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "cmp x25, #0x20\n"
-      "ldr q8, [x28, #0x30]\n"
-      "ldr q9, [x28, #0x40]\n"
-      "ldr q10, [x28, #0x50]\n"
-      "ldr q4, [x28, #0x60]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "bge 7b\n"
-      "9:"  // Height 1: Multiply loop: Single iteration only
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "trn1 v0.2d, v1.2d, v24.2d\n"
-      "trn2 v1.2d, v1.2d, v24.2d\n"
-      ".inst 0x4e85ac10  // usmmla v16.4s, v0.16b, v5.16b\n"
-      "ldr q25, [x28, #0x70]\n"
-      ".inst 0x4e86ac14  // usmmla v20.4s, v0.16b, v6.16b\n"
-      "ldr q24, [x28, #0x80]\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      "ldr q30, [x28, #0x90]\n"
-      ".inst 0x4e88ac15  // usmmla v21.4s, v0.16b, v8.16b\n"
-      "ldr q29, [x28, #0xa0]\n"
-      ".inst 0x4e89ac12  // usmmla v18.4s, v0.16b, v9.16b\n"
-      "ldr q28, [x28, #0xb0]\n"
-      ".inst 0x4e8aac16  // usmmla v22.4s, v0.16b, v10.16b\n"
-      "ldr q27, [x28, #0xc0]\n"
-      ".inst 0x4e84ac13  // usmmla v19.4s, v0.16b, v4.16b\n"
-      "ldr q26, [x28, #0xd0]\n"
-      ".inst 0x4e99ac17  // usmmla v23.4s, v0.16b, v25.16b\n"
-      "ldr q25, [x28, #0xe0]\n"
-      ".inst 0x4e98ac30  // usmmla v16.4s, v1.16b, v24.16b\n"
-      "ldr q24, [x28, #0xf0]\n"
-      ".inst 0x4e9eac34  // usmmla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4e9dac31  // usmmla v17.4s, v1.16b, v29.16b\n"
-      ".inst 0x4e9cac35  // usmmla v21.4s, v1.16b, v28.16b\n"
-      ".inst 0x4e9bac32  // usmmla v18.4s, v1.16b, v27.16b\n"
-      ".inst 0x4e9aac36  // usmmla v22.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e99ac33  // usmmla v19.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e98ac37  // usmmla v23.4s, v1.16b, v24.16b\n"
-      "tbnz %x[flags], #31, 10f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942b  // udot v11.4s, v1.16b, v15.16b\n"
-      "10:"  // Height 1: Multiply loop: unique 2: skip row sum
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "11:"  // Height 1: Multiply loop: Main loop skip
-      "cbz x25, 20f\n"
-      "cmp x25, #0x8\n"
-      "blt 14f\n"
-      "12:"  // Height 1: Multiply loop: Odd block loop
-      "ldr d25, [x24], #0x8\n"
-      "trn1 v0.2d, v25.2d, v24.2d\n"
-      "tbnz %x[flags], #31, 13f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      "13:"  // Height 1: Multiply loop: unique 3: skip row sum
-      "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "sub x25, x25, #0x8\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "cmp x25, #0x8\n"
-      "ldr q27, [x28, #0x40]\n"
-      "ldr q26, [x28, #0x50]\n"
-      "ldr q25, [x28, #0x60]\n"
-      ".inst 0x4e98ac10  // usmmla v16.4s, v0.16b, v24.16b\n"
-      "ldr q24, [x28, #0x70]\n"
-      ".inst 0x4e9eac14  // usmmla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x4e9dac11  // usmmla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x4e9cac15  // usmmla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e9bac12  // usmmla v18.4s, v0.16b, v27.16b\n"
-      ".inst 0x4e9aac16  // usmmla v22.4s, v0.16b, v26.16b\n"
-      ".inst 0x4e99ac13  // usmmla v19.4s, v0.16b, v25.16b\n"
-      ".inst 0x4e98ac17  // usmmla v23.4s, v0.16b, v24.16b\n"
-      "bge 12b\n"
-      "14:"  // Height 1: Multiply loop: Skip odd blocks
-      "cbz x25, 20f\n"
-      "tbz x25, #2, 16f\n"
-      "ldr s1, [x24], #0x4\n"
-      "tbz x25, #1, 15f\n"
-      "ld1 { v1.h }[2], [x24], #0x2\n"
-      "tbz x25, #0, 18f\n"
-      "ld1 { v1.b }[6], [x24]\n"
-      "b 18f\n"
-      "15:"  // Height 1: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x25, #0, 18f\n"
-      "ld1 { v1.b }[4], [x24]\n"
-      "b 18f\n"
-      "16:"  // Height 1: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x25, #1, 17f\n"
-      "ldr h1, [x24], #0x2\n"
-      "tbz x25, #0, 18f\n"
-      "ld1 { v1.b }[2], [x24]\n"
-      "b 18f\n"
-      "17:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x24, #0x0]\n"
-      "18:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "trn1 v0.2d, v1.2d, v24.2d\n"
-      "tbnz %x[flags], #31, 19f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      "19:"  // Height 1: Multiply loop: unique 4: skip row sum
-      "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "ldr q27, [x28, #0x40]\n"
-      "ldr q26, [x28, #0x50]\n"
-      "ldr q25, [x28, #0x60]\n"
-      ".inst 0x4e98ac10  // usmmla v16.4s, v0.16b, v24.16b\n"
-      "ldr q24, [x28, #0x70]\n"
-      ".inst 0x4e9eac14  // usmmla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x4e9dac11  // usmmla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x4e9cac15  // usmmla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e9bac12  // usmmla v18.4s, v0.16b, v27.16b\n"
-      ".inst 0x4e9aac16  // usmmla v22.4s, v0.16b, v26.16b\n"
-      ".inst 0x4e99ac13  // usmmla v19.4s, v0.16b, v25.16b\n"
-      ".inst 0x4e98ac17  // usmmla v23.4s, v0.16b, v24.16b\n"
-      "20:"  // Height 1: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 4b\n"
-      "uzp1 v16.2d, v16.2d, v20.2d\n"
-      "uzp1 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "uzp1 v18.2d, v18.2d, v22.2d\n"
-      "uzp1 v19.2d, v19.2d, v23.2d\n"
-      "mov v23.16b, v16.16b\n"
-      "tbnz %x[flags], #31, 21f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "neg v16.4s, v16.4s\n"
-      "dup v11.4s, v11.s[0]\n"
-      "mul v11.4s, v11.4s, v16.4s\n"
-      "21:"  // Height 1: skip row sum fixup
-      "ldr q24, [x10, #0x0]\n"
-      "ldr q22, [x10, #0x10]\n"
-      "add v23.4s, v23.4s, v11.4s\n"
-      "add v17.4s, v17.4s, v11.4s\n"
-      "ldr q21, [x10, #0x20]\n"
-      "ldr q20, [x10, #0x30]\n"
-      "add v18.4s, v18.4s, v11.4s\n"
-      "add v19.4s, v19.4s, v11.4s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "add v23.4s, v23.4s, v24.4s\n"
-      "add v17.4s, v17.4s, v22.4s\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add x10, x10, #0x40\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v21.4s\n"
-      "add v19.4s, v19.4s, v20.4s\n"
-      "sqrdmulh v23.4s, v23.4s, v16.4s\n"
-      "sqrdmulh v17.4s, v17.4s, v16.4s\n"
-      "sqrdmulh v18.4s, v18.4s, v16.4s\n"
-      "sqrdmulh v19.4s, v19.4s, v16.4s\n"
-      "tbz %x[flags], #5, 22f\n"
-      "and v22.16b, v23.16b, v0.16b\n"
-      "and v21.16b, v17.16b, v0.16b\n"
-      "and v20.16b, v18.16b, v0.16b\n"
-      "and v16.16b, v19.16b, v0.16b\n"
-      "sshr v22.4s, v22.4s, #0x1f\n"
-      "sshr v21.4s, v21.4s, #0x1f\n"
-      "sshr v20.4s, v20.4s, #0x1f\n"
-      "sshr v16.4s, v16.4s, #0x1f\n"
-      "sqadd v23.4s, v23.4s, v22.4s\n"
-      "sqadd v17.4s, v17.4s, v21.4s\n"
-      "sqadd v18.4s, v18.4s, v20.4s\n"
-      "sqadd v19.4s, v19.4s, v16.4s\n"
-      "22:"  // Height 1: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
-      "srshl v23.4s, v23.4s, v0.4s\n"
-      "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v21.4s }, [x21]\n"
-      "ld1r { v20.4s }, [x20]\n"
-      "srshl v18.4s, v18.4s, v0.4s\n"
-      "srshl v19.4s, v19.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v16.4s }, [x20]\n"
-      "add v23.4s, v23.4s, v21.4s\n"
-      "add v17.4s, v17.4s, v21.4s\n"
-      "add v18.4s, v18.4s, v21.4s\n"
-      "add v19.4s, v19.4s, v21.4s\n"
-      "smin v23.4s, v23.4s, v20.4s\n"
-      "smin v17.4s, v17.4s, v20.4s\n"
-      "smin v18.4s, v18.4s, v20.4s\n"
-      "smin v19.4s, v19.4s, v20.4s\n"
-      "smax v23.4s, v23.4s, v16.4s\n"
-      "smax v17.4s, v17.4s, v16.4s\n"
-      "smax v18.4s, v18.4s, v16.4s\n"
-      "smax v19.4s, v19.4s, v16.4s\n"
-      "uzp1 v23.8h, v23.8h, v17.8h\n"
-      "uzp1 v16.8h, v18.8h, v19.8h\n"
-      "uzp1 v23.16b, v23.16b, v16.16b\n"
-      "bge 31f\n"
-      "tbz x9, #3, 26f\n"
-      "str d23, [x27], #0x8\n"
-      "tbz x9, #2, 24f\n"
-      "st1 { v23.s }[2], [x27], #0x4\n"
-      "tbz x9, #1, 23f\n"
-      "st1 { v23.h }[6], [x27], #0x2\n"
-      "tbz x9, #0, 30f\n"
-      "st1 { v23.b }[14], [x27]\n"
-      "b 30f\n"
-      "23:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x9, #0, 30f\n"
-      "st1 { v23.b }[12], [x27]\n"
-      "b 30f\n"
-      "24:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x9, #1, 25f\n"
-      "st1 { v23.h }[4], [x27], #0x2\n"
-      "tbz x9, #0, 30f\n"
-      "st1 { v23.b }[10], [x27]\n"
-      "b 30f\n"
-      "25:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x9, #0, 30f\n"
-      "st1 { v23.b }[8], [x27]\n"
-      "b 30f\n"
-      "26:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x9, #2, 28f\n"
-      "str s23, [x27], #0x4\n"
-      "tbz x9, #1, 27f\n"
-      "st1 { v23.h }[2], [x27], #0x2\n"
-      "tbz x9, #0, 30f\n"
-      "st1 { v23.b }[6], [x27]\n"
-      "b 30f\n"
-      "27:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x9, #0, 30f\n"
-      "st1 { v23.b }[4], [x27]\n"
-      "b 30f\n"
-      "28:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x9, #1, 29f\n"
-      "str h23, [x27], #0x2\n"
-      "tbz x9, #0, 30f\n"
-      "st1 { v23.b }[2], [x27]\n"
-      "b 30f\n"
-      "29:"  // Height 1: Partial direct writeback: partial_1_0
-      "str b23, [x27, #0x0]\n"
-      "30:"  // Height 1: Partial direct writeback: Done
-      "b 32f\n"
-      "31:"  // Height 1: Full writeback
-      "str q23, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "32:"  // Height 1: Writeback done
-      "subs x9, x9, #0x10\n"
-      "bgt 2b\n"
-      "b 130f\n"
-      "33:"  // Height 2
-      "mov x10, %x[col_bias]\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "movi v15.16b, #0x1\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "34:"  // Height 2: Column loop
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "35:"  // Height 2: setup done
-      "mov x26, #0x0\n"
-      "36:"  // Height 2: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 37f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "cbnz x26, 38f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "b 38f\n"
-      "37:"  // Height 2: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "38:"  // Height 2: input setup done
-      "cmp x25, #0x10\n"
-      "blt 43f\n"
-      "ldr q1, [x24, #0x0]\n"
-      "ldr q2, [x23, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q5, [x28, #0x0]\n"
-      "ldr q6, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q8, [x28, #0x30]\n"
-      "ldr q9, [x28, #0x40]\n"
-      "ldr q10, [x28, #0x50]\n"
-      "ldr q4, [x28, #0x60]\n"
-      "blt 41f\n"
-      "39:"  // Height 2: Multiply loop: Main loop head
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4e85ac10  // usmmla v16.4s, v0.16b, v5.16b\n"
-      "ldr q25, [x28, #0x70]\n"
-      ".inst 0x4e86ac14  // usmmla v20.4s, v0.16b, v6.16b\n"
-      "ldr q24, [x28, #0x80]\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      "ldr q30, [x28, #0x90]\n"
-      ".inst 0x4e88ac15  // usmmla v21.4s, v0.16b, v8.16b\n"
-      "ldr q29, [x28, #0xa0]\n"
-      ".inst 0x4e89ac12  // usmmla v18.4s, v0.16b, v9.16b\n"
-      "ldr q28, [x28, #0xb0]\n"
-      ".inst 0x4e8aac16  // usmmla v22.4s, v0.16b, v10.16b\n"
-      "ldr q27, [x28, #0xc0]\n"
-      ".inst 0x4e84ac13  // usmmla v19.4s, v0.16b, v4.16b\n"
-      "ldr q26, [x28, #0xd0]\n"
-      ".inst 0x4e99ac17  // usmmla v23.4s, v0.16b, v25.16b\n"
-      "ldr q25, [x28, #0xe0]\n"
-      ".inst 0x4e98ac30  // usmmla v16.4s, v1.16b, v24.16b\n"
-      "ldr q24, [x28, #0xf0]\n"
-      ".inst 0x4e9eac34  // usmmla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4e9dac31  // usmmla v17.4s, v1.16b, v29.16b\n"
-      ".inst 0x4e9cac35  // usmmla v21.4s, v1.16b, v28.16b\n"
-      ".inst 0x4e9bac32  // usmmla v18.4s, v1.16b, v27.16b\n"
-      ".inst 0x4e9aac36  // usmmla v22.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e99ac33  // usmmla v19.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e98ac37  // usmmla v23.4s, v1.16b, v24.16b\n"
-      "tbnz %x[flags], #31, 40f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942b  // udot v11.4s, v1.16b, v15.16b\n"
-      "40:"  // Height 2: Multiply loop: unique 5: skip row sum
-      "ldr q1, [x24, #0x0]\n"
-      "ldr q2, [x23, #0x0]\n"
-      "sub x25, x25, #0x10\n"
-      "ldr q5, [x28, #0x0]\n"
-      "ldr q6, [x28, #0x10]\n"
-      "cmp x25, #0x20\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q8, [x28, #0x30]\n"
-      "ldr q9, [x28, #0x40]\n"
-      "ldr q10, [x28, #0x50]\n"
-      "ldr q4, [x28, #0x60]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "bge 39b\n"
-      "41:"  // Height 2: Multiply loop: Single iteration only
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4e85ac10  // usmmla v16.4s, v0.16b, v5.16b\n"
-      "ldr q25, [x28, #0x70]\n"
-      ".inst 0x4e86ac14  // usmmla v20.4s, v0.16b, v6.16b\n"
-      "ldr q24, [x28, #0x80]\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      "ldr q30, [x28, #0x90]\n"
-      ".inst 0x4e88ac15  // usmmla v21.4s, v0.16b, v8.16b\n"
-      "ldr q29, [x28, #0xa0]\n"
-      ".inst 0x4e89ac12  // usmmla v18.4s, v0.16b, v9.16b\n"
-      "ldr q28, [x28, #0xb0]\n"
-      ".inst 0x4e8aac16  // usmmla v22.4s, v0.16b, v10.16b\n"
-      "ldr q27, [x28, #0xc0]\n"
-      ".inst 0x4e84ac13  // usmmla v19.4s, v0.16b, v4.16b\n"
-      "ldr q26, [x28, #0xd0]\n"
-      ".inst 0x4e99ac17  // usmmla v23.4s, v0.16b, v25.16b\n"
-      "ldr q25, [x28, #0xe0]\n"
-      ".inst 0x4e98ac30  // usmmla v16.4s, v1.16b, v24.16b\n"
-      "ldr q24, [x28, #0xf0]\n"
-      ".inst 0x4e9eac34  // usmmla v20.4s, v1.16b, v30.16b\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4e9dac31  // usmmla v17.4s, v1.16b, v29.16b\n"
-      ".inst 0x4e9cac35  // usmmla v21.4s, v1.16b, v28.16b\n"
-      ".inst 0x4e9bac32  // usmmla v18.4s, v1.16b, v27.16b\n"
-      ".inst 0x4e9aac36  // usmmla v22.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e99ac33  // usmmla v19.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e98ac37  // usmmla v23.4s, v1.16b, v24.16b\n"
-      "tbnz %x[flags], #31, 42f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f942b  // udot v11.4s, v1.16b, v15.16b\n"
-      "42:"  // Height 2: Multiply loop: unique 6: skip row sum
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "43:"  // Height 2: Multiply loop: Main loop skip
-      "cbz x25, 52f\n"
-      "cmp x25, #0x8\n"
-      "blt 46f\n"
-      "44:"  // Height 2: Multiply loop: Odd block loop
-      "ldr d25, [x24], #0x8\n"
-      "ldr d24, [x23], #0x8\n"
-      "trn1 v0.2d, v25.2d, v24.2d\n"
-      "tbnz %x[flags], #31, 45f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      "45:"  // Height 2: Multiply loop: unique 7: skip row sum
-      "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "sub x25, x25, #0x8\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "cmp x25, #0x8\n"
-      "ldr q27, [x28, #0x40]\n"
-      "ldr q26, [x28, #0x50]\n"
-      "ldr q25, [x28, #0x60]\n"
-      ".inst 0x4e98ac10  // usmmla v16.4s, v0.16b, v24.16b\n"
-      "ldr q24, [x28, #0x70]\n"
-      ".inst 0x4e9eac14  // usmmla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x4e9dac11  // usmmla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x4e9cac15  // usmmla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e9bac12  // usmmla v18.4s, v0.16b, v27.16b\n"
-      ".inst 0x4e9aac16  // usmmla v22.4s, v0.16b, v26.16b\n"
-      ".inst 0x4e99ac13  // usmmla v19.4s, v0.16b, v25.16b\n"
-      ".inst 0x4e98ac17  // usmmla v23.4s, v0.16b, v24.16b\n"
-      "bge 44b\n"
-      "46:"  // Height 2: Multiply loop: Skip odd blocks
-      "cbz x25, 52f\n"
-      "tbz x25, #2, 48f\n"
-      "ldr s1, [x24], #0x4\n"
-      "ldr s2, [x23], #0x4\n"
-      "tbz x25, #1, 47f\n"
-      "ld1 { v1.h }[2], [x24], #0x2\n"
-      "ld1 { v2.h }[2], [x23], #0x2\n"
-      "tbz x25, #0, 50f\n"
-      "ld1 { v1.b }[6], [x24]\n"
-      "ld1 { v2.b }[6], [x23]\n"
-      "b 50f\n"
-      "47:"  // Height 2: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x25, #0, 50f\n"
-      "ld1 { v1.b }[4], [x24]\n"
-      "ld1 { v2.b }[4], [x23]\n"
-      "b 50f\n"
-      "48:"  // Height 2: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x25, #1, 49f\n"
-      "ldr h1, [x24], #0x2\n"
-      "ldr h2, [x23], #0x2\n"
-      "tbz x25, #0, 50f\n"
-      "ld1 { v1.b }[2], [x24]\n"
-      "ld1 { v2.b }[2], [x23]\n"
-      "b 50f\n"
-      "49:"  // Height 2: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x24, #0x0]\n"
-      "ldr b2, [x23, #0x0]\n"
-      "50:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "tbnz %x[flags], #31, 51f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      "51:"  // Height 2: Multiply loop: unique 8: skip row sum
-      "ldr q24, [x28, #0x0]\n"
-      "ldr q30, [x28, #0x10]\n"
-      "ldr q29, [x28, #0x20]\n"
-      "ldr q28, [x28, #0x30]\n"
-      "ldr q27, [x28, #0x40]\n"
-      "ldr q26, [x28, #0x50]\n"
-      "ldr q25, [x28, #0x60]\n"
-      ".inst 0x4e98ac10  // usmmla v16.4s, v0.16b, v24.16b\n"
-      "ldr q24, [x28, #0x70]\n"
-      ".inst 0x4e9eac14  // usmmla v20.4s, v0.16b, v30.16b\n"
-      ".inst 0x4e9dac11  // usmmla v17.4s, v0.16b, v29.16b\n"
-      ".inst 0x4e9cac15  // usmmla v21.4s, v0.16b, v28.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e9bac12  // usmmla v18.4s, v0.16b, v27.16b\n"
-      ".inst 0x4e9aac16  // usmmla v22.4s, v0.16b, v26.16b\n"
-      ".inst 0x4e99ac13  // usmmla v19.4s, v0.16b, v25.16b\n"
-      ".inst 0x4e98ac17  // usmmla v23.4s, v0.16b, v24.16b\n"
-      "52:"  // Height 2: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 36b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 v24.2d, v16.2d, v20.2d\n"
-      "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "uzp1 v20.2d, v17.2d, v21.2d\n"
-      "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "uzp1 v21.2d, v18.2d, v22.2d\n"
-      "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
-      "uzp1 v22.2d, v19.2d, v23.2d\n"
-      "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "mov v23.16b, v24.16b\n"
-      "tbnz %x[flags], #31, 53f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "ld1r { v24.4s }, [x20]\n"
-      "neg v24.4s, v24.4s\n"
-      "dup v12.4s, v11.s[3]\n"
-      "dup v11.4s, v11.s[0]\n"
-      "mul v11.4s, v11.4s, v24.4s\n"
-      "mul v12.4s, v12.4s, v24.4s\n"
-      "53:"  // Height 2: skip row sum fixup
-      "ldr q28, [x10, #0x0]\n"
-      "ldr q27, [x10, #0x10]\n"
-      "add v23.4s, v23.4s, v11.4s\n"
-      "add v20.4s, v20.4s, v11.4s\n"
-      "ldr q26, [x10, #0x20]\n"
-      "ldr q25, [x10, #0x30]\n"
-      "add v21.4s, v21.4s, v11.4s\n"
-      "add v22.4s, v22.4s, v11.4s\n"
-      "add v16.4s, v16.4s, v12.4s\n"
-      "add v17.4s, v17.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "ld1r { v24.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v12.4s\n"
-      "add v19.4s, v19.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add v23.4s, v23.4s, v28.4s\n"
-      "add v20.4s, v20.4s, v27.4s\n"
-      "add x10, x10, #0x40\n"
-      "add v21.4s, v21.4s, v26.4s\n"
-      "add v22.4s, v22.4s, v25.4s\n"
-      "add v16.4s, v16.4s, v28.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "add v17.4s, v17.4s, v27.4s\n"
-      "add v18.4s, v18.4s, v26.4s\n"
-      "add v19.4s, v19.4s, v25.4s\n"
-      "sqrdmulh v23.4s, v23.4s, v24.4s\n"
-      "sqrdmulh v20.4s, v20.4s, v24.4s\n"
-      "sqrdmulh v21.4s, v21.4s, v24.4s\n"
-      "sqrdmulh v22.4s, v22.4s, v24.4s\n"
-      "sqrdmulh v16.4s, v16.4s, v24.4s\n"
-      "sqrdmulh v17.4s, v17.4s, v24.4s\n"
-      "sqrdmulh v18.4s, v18.4s, v24.4s\n"
-      "sqrdmulh v19.4s, v19.4s, v24.4s\n"
-      "tbz %x[flags], #5, 54f\n"
-      "and v24.16b, v23.16b, v0.16b\n"
-      "and v30.16b, v20.16b, v0.16b\n"
-      "and v29.16b, v21.16b, v0.16b\n"
-      "and v28.16b, v22.16b, v0.16b\n"
-      "and v27.16b, v16.16b, v0.16b\n"
-      "and v26.16b, v17.16b, v0.16b\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
-      "and v25.16b, v18.16b, v0.16b\n"
-      "sshr v30.4s, v30.4s, #0x1f\n"
-      "sshr v29.4s, v29.4s, #0x1f\n"
-      "sshr v28.4s, v28.4s, #0x1f\n"
-      "sshr v27.4s, v27.4s, #0x1f\n"
-      "sqadd v23.4s, v23.4s, v24.4s\n"
-      "and v24.16b, v19.16b, v0.16b\n"
-      "sshr v26.4s, v26.4s, #0x1f\n"
-      "sshr v25.4s, v25.4s, #0x1f\n"
-      "sqadd v20.4s, v20.4s, v30.4s\n"
-      "sqadd v21.4s, v21.4s, v29.4s\n"
-      "sshr v24.4s, v24.4s, #0x1f\n"
-      "sqadd v22.4s, v22.4s, v28.4s\n"
-      "sqadd v16.4s, v16.4s, v27.4s\n"
-      "sqadd v17.4s, v17.4s, v26.4s\n"
-      "sqadd v18.4s, v18.4s, v25.4s\n"
-      "sqadd v19.4s, v19.4s, v24.4s\n"
-      "54:"  // Height 2: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
-      "srshl v23.4s, v23.4s, v0.4s\n"
-      "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v26.4s }, [x21]\n"
-      "ld1r { v25.4s }, [x20]\n"
-      "srshl v21.4s, v21.4s, v0.4s\n"
-      "srshl v22.4s, v22.4s, v0.4s\n"
-      "srshl v16.4s, v16.4s, v0.4s\n"
-      "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v24.4s }, [x20]\n"
-      "srshl v18.4s, v18.4s, v0.4s\n"
-      "srshl v19.4s, v19.4s, v0.4s\n"
-      "add v23.4s, v23.4s, v26.4s\n"
-      "add v20.4s, v20.4s, v26.4s\n"
-      "add v21.4s, v21.4s, v26.4s\n"
-      "add v22.4s, v22.4s, v26.4s\n"
-      "add v16.4s, v16.4s, v26.4s\n"
-      "add v17.4s, v17.4s, v26.4s\n"
-      "add v18.4s, v18.4s, v26.4s\n"
-      "add v19.4s, v19.4s, v26.4s\n"
-      "smin v23.4s, v23.4s, v25.4s\n"
-      "smin v20.4s, v20.4s, v25.4s\n"
-      "smin v21.4s, v21.4s, v25.4s\n"
-      "smin v22.4s, v22.4s, v25.4s\n"
-      "smin v16.4s, v16.4s, v25.4s\n"
-      "smin v17.4s, v17.4s, v25.4s\n"
-      "smin v18.4s, v18.4s, v25.4s\n"
-      "smin v19.4s, v19.4s, v25.4s\n"
-      "smax v23.4s, v23.4s, v24.4s\n"
-      "smax v20.4s, v20.4s, v24.4s\n"
-      "smax v21.4s, v21.4s, v24.4s\n"
-      "smax v22.4s, v22.4s, v24.4s\n"
-      "smax v16.4s, v16.4s, v24.4s\n"
-      "smax v17.4s, v17.4s, v24.4s\n"
-      "smax v18.4s, v18.4s, v24.4s\n"
-      "smax v19.4s, v19.4s, v24.4s\n"
-      "uzp1 v23.8h, v23.8h, v20.8h\n"
-      "uzp1 v20.8h, v21.8h, v22.8h\n"
-      "uzp1 v16.8h, v16.8h, v17.8h\n"
-      "uzp1 v17.8h, v18.8h, v19.8h\n"
-      "uzp1 v23.16b, v23.16b, v20.16b\n"
-      "uzp1 v16.16b, v16.16b, v17.16b\n"
-      "bge 63f\n"
-      "tbz x9, #3, 58f\n"
-      "str d23, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "tbz x9, #2, 56f\n"
-      "st1 { v23.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "tbz x9, #1, 55f\n"
-      "st1 { v23.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "tbz x9, #0, 62f\n"
-      "st1 { v23.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "b 62f\n"
-      "55:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x9, #0, 62f\n"
-      "st1 { v23.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "b 62f\n"
-      "56:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x9, #1, 57f\n"
-      "st1 { v23.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "tbz x9, #0, 62f\n"
-      "st1 { v23.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "b 62f\n"
-      "57:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x9, #0, 62f\n"
-      "st1 { v23.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "b 62f\n"
-      "58:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x9, #2, 60f\n"
-      "str s23, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "tbz x9, #1, 59f\n"
-      "st1 { v23.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "tbz x9, #0, 62f\n"
-      "st1 { v23.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "b 62f\n"
-      "59:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x9, #0, 62f\n"
-      "st1 { v23.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "b 62f\n"
-      "60:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x9, #1, 61f\n"
-      "str h23, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "tbz x9, #0, 62f\n"
-      "st1 { v23.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "b 62f\n"
-      "61:"  // Height 2: Partial direct writeback: partial_1_0
-      "str b23, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "62:"  // Height 2: Partial direct writeback: Done
-      "b 64f\n"
-      "63:"  // Height 2: Full writeback
-      "str q23, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
-      "64:"  // Height 2: Writeback done
-      "subs x9, x9, #0x10\n"
-      "bgt 34b\n"
-      "b 130f\n"
-      "65:"  // Height 3
-      "mov x10, %x[col_bias]\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "movi v13.4s, #0x0\n"
-      "movi v15.16b, #0x1\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "66:"  // Height 3: Column loop
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "movi v28.4s, #0x0\n"
-      "movi v29.4s, #0x0\n"
-      "movi v30.4s, #0x0\n"
-      "movi v31.4s, #0x0\n"
-      "67:"  // Height 3: setup done
-      "mov x26, #0x0\n"
-      "68:"  // Height 3: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 69f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "ldr x22, [x20, #0x10]\n"
-      "cbnz x26, 70f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "b 70f\n"
-      "69:"  // Height 3: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "70:"  // Height 3: input setup done
-      "cmp x25, #0x10\n"
-      "blt 75f\n"
-      "ldr q1, [x24, #0x0]\n"
-      "ldr q2, [x23, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q3, [x22, #0x0]\n"
-      "ldr q5, [x28, #0x0]\n"
-      "ldr q6, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q8, [x28, #0x30]\n"
-      "ldr q9, [x28, #0x40]\n"
-      "ldr q10, [x28, #0x50]\n"
-      "blt 73f\n"
-      "71:"  // Height 3: Multiply loop: Main loop head
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "ldr q14, [x28, #0x60]\n"
-      ".inst 0x4e85ac10  // usmmla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e86ac14  // usmmla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e88ac15  // usmmla v21.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e85ac58  // usmmla v24.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0x70]\n"
-      ".inst 0x4e86ac5c  // usmmla v28.4s, v2.16b, v6.16b\n"
-      "ldr q4, [x28, #0x80]\n"
-      ".inst 0x4e89ac12  // usmmla v18.4s, v0.16b, v9.16b\n"
-      ".inst 0x4e87ac59  // usmmla v25.4s, v2.16b, v7.16b\n"
-      "ldr q7, [x28, #0x90]\n"
-      ".inst 0x4e88ac5d  // usmmla v29.4s, v2.16b, v8.16b\n"
-      "ldr q8, [x28, #0xa0]\n"
-      ".inst 0x4e89ac5a  // usmmla v26.4s, v2.16b, v9.16b\n"
-      "ldr q9, [x28, #0xb0]\n"
-      ".inst 0x4e8aac16  // usmmla v22.4s, v0.16b, v10.16b\n"
-      ".inst 0x4e8aac5e  // usmmla v30.4s, v2.16b, v10.16b\n"
-      "ldr q10, [x28, #0xc0]\n"
-      ".inst 0x4e8eac13  // usmmla v19.4s, v0.16b, v14.16b\n"
-      ".inst 0x4e8eac5b  // usmmla v27.4s, v2.16b, v14.16b\n"
-      "ldr q6, [x28, #0xd0]\n"
-      ".inst 0x4e85ac17  // usmmla v23.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85ac5f  // usmmla v31.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0xe0]\n"
-      ".inst 0x4e84ac30  // usmmla v16.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e84ac78  // usmmla v24.4s, v3.16b, v4.16b\n"
-      "ldr q4, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4e87ac34  // usmmla v20.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e87ac7c  // usmmla v28.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e88ac31  // usmmla v17.4s, v1.16b, v8.16b\n"
-      ".inst 0x4e88ac79  // usmmla v25.4s, v3.16b, v8.16b\n"
-      ".inst 0x4e89ac35  // usmmla v21.4s, v1.16b, v9.16b\n"
-      ".inst 0x4e89ac7d  // usmmla v29.4s, v3.16b, v9.16b\n"
-      ".inst 0x4e8aac32  // usmmla v18.4s, v1.16b, v10.16b\n"
-      ".inst 0x4e8aac7a  // usmmla v26.4s, v3.16b, v10.16b\n"
-      ".inst 0x4e86ac36  // usmmla v22.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac7e  // usmmla v30.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e85ac33  // usmmla v19.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e85ac7b  // usmmla v27.4s, v3.16b, v5.16b\n"
-      ".inst 0x4e84ac37  // usmmla v23.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e84ac7f  // usmmla v31.4s, v3.16b, v4.16b\n"
-      "tbnz %x[flags], #31, 72f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      ".inst 0x6e8f942b  // udot v11.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f946d  // udot v13.4s, v3.16b, v15.16b\n"
-      "72:"  // Height 3: Multiply loop: unique 9: skip row sum
-      "ldr q1, [x24, #0x0]\n"
-      "ldr q2, [x23, #0x0]\n"
-      "sub x25, x25, #0x10\n"
-      "ldr q3, [x22, #0x0]\n"
-      "ldr q5, [x28, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q6, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q8, [x28, #0x30]\n"
-      "ldr q9, [x28, #0x40]\n"
-      "ldr q10, [x28, #0x50]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "bge 71b\n"
-      "73:"  // Height 3: Multiply loop: Single iteration only
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "ldr q14, [x28, #0x60]\n"
-      ".inst 0x4e85ac10  // usmmla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e86ac14  // usmmla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e88ac15  // usmmla v21.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e85ac58  // usmmla v24.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0x70]\n"
-      ".inst 0x4e86ac5c  // usmmla v28.4s, v2.16b, v6.16b\n"
-      "ldr q4, [x28, #0x80]\n"
-      ".inst 0x4e89ac12  // usmmla v18.4s, v0.16b, v9.16b\n"
-      ".inst 0x4e87ac59  // usmmla v25.4s, v2.16b, v7.16b\n"
-      "ldr q7, [x28, #0x90]\n"
-      ".inst 0x4e88ac5d  // usmmla v29.4s, v2.16b, v8.16b\n"
-      "ldr q8, [x28, #0xa0]\n"
-      ".inst 0x4e89ac5a  // usmmla v26.4s, v2.16b, v9.16b\n"
-      "ldr q9, [x28, #0xb0]\n"
-      ".inst 0x4e8aac16  // usmmla v22.4s, v0.16b, v10.16b\n"
-      ".inst 0x4e8aac5e  // usmmla v30.4s, v2.16b, v10.16b\n"
-      "ldr q10, [x28, #0xc0]\n"
-      ".inst 0x4e8eac13  // usmmla v19.4s, v0.16b, v14.16b\n"
-      ".inst 0x4e8eac5b  // usmmla v27.4s, v2.16b, v14.16b\n"
-      "ldr q6, [x28, #0xd0]\n"
-      ".inst 0x4e85ac17  // usmmla v23.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85ac5f  // usmmla v31.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0xe0]\n"
-      ".inst 0x4e84ac30  // usmmla v16.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e84ac78  // usmmla v24.4s, v3.16b, v4.16b\n"
-      "ldr q4, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4e87ac34  // usmmla v20.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e87ac7c  // usmmla v28.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e88ac31  // usmmla v17.4s, v1.16b, v8.16b\n"
-      ".inst 0x4e88ac79  // usmmla v25.4s, v3.16b, v8.16b\n"
-      ".inst 0x4e89ac35  // usmmla v21.4s, v1.16b, v9.16b\n"
-      ".inst 0x4e89ac7d  // usmmla v29.4s, v3.16b, v9.16b\n"
-      ".inst 0x4e8aac32  // usmmla v18.4s, v1.16b, v10.16b\n"
-      ".inst 0x4e8aac7a  // usmmla v26.4s, v3.16b, v10.16b\n"
-      ".inst 0x4e86ac36  // usmmla v22.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac7e  // usmmla v30.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e85ac33  // usmmla v19.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e85ac7b  // usmmla v27.4s, v3.16b, v5.16b\n"
-      ".inst 0x4e84ac37  // usmmla v23.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e84ac7f  // usmmla v31.4s, v3.16b, v4.16b\n"
-      "tbnz %x[flags], #31, 74f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      ".inst 0x6e8f942b  // udot v11.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f946d  // udot v13.4s, v3.16b, v15.16b\n"
-      "74:"  // Height 3: Multiply loop: unique 10: skip row sum
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "75:"  // Height 3: Multiply loop: Main loop skip
-      "cbz x25, 84f\n"
-      "cmp x25, #0x8\n"
-      "blt 78f\n"
-      "76:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d3, [x24], #0x8\n"
-      "ldr d0, [x23], #0x8\n"
-      "ldr d1, [x22], #0x8\n"
-      "trn1 v0.2d, v3.2d, v0.2d\n"
-      "trn1 v2.2d, v1.2d, v2.2d\n"
-      "tbnz %x[flags], #31, 77f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      "77:"  // Height 3: Multiply loop: unique 11: skip row sum
-      "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "sub x25, x25, #0x8\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q6, [x28, #0x30]\n"
-      "cmp x25, #0x8\n"
-      "ldr q5, [x28, #0x40]\n"
-      "ldr q4, [x28, #0x50]\n"
-      "ldr q3, [x28, #0x60]\n"
-      ".inst 0x4e81ac10  // usmmla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81ac58  // usmmla v24.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x28, #0x70]\n"
-      ".inst 0x4e88ac14  // usmmla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e88ac5c  // usmmla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac59  // usmmla v25.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e86ac15  // usmmla v21.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac5d  // usmmla v29.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e85ac12  // usmmla v18.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85ac5a  // usmmla v26.4s, v2.16b, v5.16b\n"
-      ".inst 0x4e84ac16  // usmmla v22.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e84ac5e  // usmmla v30.4s, v2.16b, v4.16b\n"
-      ".inst 0x4e83ac13  // usmmla v19.4s, v0.16b, v3.16b\n"
-      ".inst 0x4e83ac5b  // usmmla v27.4s, v2.16b, v3.16b\n"
-      ".inst 0x4e81ac17  // usmmla v23.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81ac5f  // usmmla v31.4s, v2.16b, v1.16b\n"
-      "bge 76b\n"
-      "78:"  // Height 3: Multiply loop: Skip odd blocks
-      "cbz x25, 84f\n"
-      "tbz x25, #2, 80f\n"
-      "ldr s1, [x24], #0x4\n"
-      "ldr s2, [x23], #0x4\n"
-      "ldr s3, [x22], #0x4\n"
-      "tbz x25, #1, 79f\n"
-      "ld1 { v1.h }[2], [x24], #0x2\n"
-      "ld1 { v2.h }[2], [x23], #0x2\n"
-      "ld1 { v3.h }[2], [x22], #0x2\n"
-      "tbz x25, #0, 82f\n"
-      "ld1 { v1.b }[6], [x24]\n"
-      "ld1 { v2.b }[6], [x23]\n"
-      "ld1 { v3.b }[6], [x22]\n"
-      "b 82f\n"
-      "79:"  // Height 3: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x25, #0, 82f\n"
-      "ld1 { v1.b }[4], [x24]\n"
-      "ld1 { v2.b }[4], [x23]\n"
-      "ld1 { v3.b }[4], [x22]\n"
-      "b 82f\n"
-      "80:"  // Height 3: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x25, #1, 81f\n"
-      "ldr h1, [x24], #0x2\n"
-      "ldr h2, [x23], #0x2\n"
-      "ldr h3, [x22], #0x2\n"
-      "tbz x25, #0, 82f\n"
-      "ld1 { v1.b }[2], [x24]\n"
-      "ld1 { v2.b }[2], [x23]\n"
-      "ld1 { v3.b }[2], [x22]\n"
-      "b 82f\n"
-      "81:"  // Height 3: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x24, #0x0]\n"
-      "ldr b2, [x23, #0x0]\n"
-      "ldr b3, [x22, #0x0]\n"
-      "82:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "tbnz %x[flags], #31, 83f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      "83:"  // Height 3: Multiply loop: unique 12: skip row sum
-      "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q6, [x28, #0x30]\n"
-      "ldr q5, [x28, #0x40]\n"
-      "ldr q4, [x28, #0x50]\n"
-      "ldr q3, [x28, #0x60]\n"
-      ".inst 0x4e81ac10  // usmmla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81ac58  // usmmla v24.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x28, #0x70]\n"
-      ".inst 0x4e88ac14  // usmmla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e88ac5c  // usmmla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac59  // usmmla v25.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e86ac15  // usmmla v21.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac5d  // usmmla v29.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e85ac12  // usmmla v18.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85ac5a  // usmmla v26.4s, v2.16b, v5.16b\n"
-      ".inst 0x4e84ac16  // usmmla v22.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e84ac5e  // usmmla v30.4s, v2.16b, v4.16b\n"
-      ".inst 0x4e83ac13  // usmmla v19.4s, v0.16b, v3.16b\n"
-      ".inst 0x4e83ac5b  // usmmla v27.4s, v2.16b, v3.16b\n"
-      ".inst 0x4e81ac17  // usmmla v23.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81ac5f  // usmmla v31.4s, v2.16b, v1.16b\n"
-      "84:"  // Height 3: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 68b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 v0.2d, v16.2d, v20.2d\n"
-      "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "uzp1 v20.2d, v17.2d, v21.2d\n"
-      "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "uzp1 v21.2d, v18.2d, v22.2d\n"
-      "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
-      "uzp1 v22.2d, v19.2d, v23.2d\n"
-      "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "uzp1 v24.2d, v24.2d, v28.2d\n"
-      "uzp1 v25.2d, v25.2d, v29.2d\n"
-      "uzp1 v26.2d, v26.2d, v30.2d\n"
-      "uzp1 v27.2d, v27.2d, v31.2d\n"
-      "mov v31.16b, v0.16b\n"
-      "tbnz %x[flags], #31, 85f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "addp v13.4s, v13.4s, v13.4s\n"
-      "ld1r { v23.4s }, [x20]\n"
-      "neg v23.4s, v23.4s\n"
-      "dup v12.4s, v11.s[3]\n"
-      "dup v11.4s, v11.s[0]\n"
-      "dup v13.4s, v13.s[0]\n"
-      "mul v11.4s, v11.4s, v23.4s\n"
-      "mul v12.4s, v12.4s, v23.4s\n"
-      "mul v13.4s, v13.4s, v23.4s\n"
-      "85:"  // Height 3: skip row sum fixup
-      "ldr q0, [x10, #0x0]\n"
-      "ldr q30, [x10, #0x10]\n"
-      "add v31.4s, v31.4s, v11.4s\n"
-      "add v20.4s, v20.4s, v11.4s\n"
-      "ldr q29, [x10, #0x20]\n"
-      "ldr q28, [x10, #0x30]\n"
-      "add v21.4s, v21.4s, v11.4s\n"
-      "add v22.4s, v22.4s, v11.4s\n"
-      "add v16.4s, v16.4s, v12.4s\n"
-      "add v17.4s, v17.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "ld1r { v23.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v12.4s\n"
-      "add v19.4s, v19.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add v24.4s, v24.4s, v13.4s\n"
-      "add v25.4s, v25.4s, v13.4s\n"
-      "add x10, x10, #0x40\n"
-      "add v26.4s, v26.4s, v13.4s\n"
-      "add v27.4s, v27.4s, v13.4s\n"
-      "add v31.4s, v31.4s, v0.4s\n"
-      "add v20.4s, v20.4s, v30.4s\n"
-      "add v21.4s, v21.4s, v29.4s\n"
-      "add v22.4s, v22.4s, v28.4s\n"
-      "add v16.4s, v16.4s, v0.4s\n"
-      "add v17.4s, v17.4s, v30.4s\n"
-      "add v18.4s, v18.4s, v29.4s\n"
-      "add v19.4s, v19.4s, v28.4s\n"
-      "add v24.4s, v24.4s, v0.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "add v25.4s, v25.4s, v30.4s\n"
-      "add v26.4s, v26.4s, v29.4s\n"
-      "add v27.4s, v27.4s, v28.4s\n"
-      "sqrdmulh v31.4s, v31.4s, v23.4s\n"
-      "sqrdmulh v20.4s, v20.4s, v23.4s\n"
-      "sqrdmulh v21.4s, v21.4s, v23.4s\n"
-      "sqrdmulh v22.4s, v22.4s, v23.4s\n"
-      "sqrdmulh v16.4s, v16.4s, v23.4s\n"
-      "sqrdmulh v17.4s, v17.4s, v23.4s\n"
-      "sqrdmulh v18.4s, v18.4s, v23.4s\n"
-      "sqrdmulh v19.4s, v19.4s, v23.4s\n"
-      "sqrdmulh v24.4s, v24.4s, v23.4s\n"
-      "sqrdmulh v25.4s, v25.4s, v23.4s\n"
-      "sqrdmulh v26.4s, v26.4s, v23.4s\n"
-      "sqrdmulh v27.4s, v27.4s, v23.4s\n"
-      "tbz %x[flags], #5, 86f\n"
-      "and v1.16b, v31.16b, v0.16b\n"
-      "and v30.16b, v20.16b, v0.16b\n"
-      "and v29.16b, v21.16b, v0.16b\n"
-      "and v28.16b, v22.16b, v0.16b\n"
-      "and v23.16b, v16.16b, v0.16b\n"
-      "and v3.16b, v17.16b, v0.16b\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "sshr v30.4s, v30.4s, #0x1f\n"
-      "sshr v29.4s, v29.4s, #0x1f\n"
-      "sshr v28.4s, v28.4s, #0x1f\n"
-      "sshr v23.4s, v23.4s, #0x1f\n"
-      "and v2.16b, v18.16b, v0.16b\n"
-      "sqadd v31.4s, v31.4s, v1.4s\n"
-      "sqadd v20.4s, v20.4s, v30.4s\n"
-      "sqadd v21.4s, v21.4s, v29.4s\n"
-      "sqadd v22.4s, v22.4s, v28.4s\n"
-      "sqadd v16.4s, v16.4s, v23.4s\n"
-      "and v1.16b, v19.16b, v0.16b\n"
-      "and v30.16b, v24.16b, v0.16b\n"
-      "and v29.16b, v25.16b, v0.16b\n"
-      "and v28.16b, v26.16b, v0.16b\n"
-      "and v23.16b, v27.16b, v0.16b\n"
-      "sshr v3.4s, v3.4s, #0x1f\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "sshr v30.4s, v30.4s, #0x1f\n"
-      "sshr v29.4s, v29.4s, #0x1f\n"
-      "sshr v28.4s, v28.4s, #0x1f\n"
-      "sshr v23.4s, v23.4s, #0x1f\n"
-      "sqadd v17.4s, v17.4s, v3.4s\n"
-      "sqadd v18.4s, v18.4s, v2.4s\n"
-      "sqadd v19.4s, v19.4s, v1.4s\n"
-      "sqadd v24.4s, v24.4s, v30.4s\n"
-      "sqadd v25.4s, v25.4s, v29.4s\n"
-      "sqadd v26.4s, v26.4s, v28.4s\n"
-      "sqadd v27.4s, v27.4s, v23.4s\n"
-      "86:"  // Height 3: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
-      "srshl v31.4s, v31.4s, v0.4s\n"
-      "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v29.4s }, [x21]\n"
-      "ld1r { v28.4s }, [x20]\n"
-      "srshl v21.4s, v21.4s, v0.4s\n"
-      "srshl v22.4s, v22.4s, v0.4s\n"
-      "srshl v16.4s, v16.4s, v0.4s\n"
-      "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v23.4s }, [x20]\n"
-      "srshl v18.4s, v18.4s, v0.4s\n"
-      "srshl v19.4s, v19.4s, v0.4s\n"
-      "srshl v24.4s, v24.4s, v0.4s\n"
-      "srshl v25.4s, v25.4s, v0.4s\n"
-      "srshl v26.4s, v26.4s, v0.4s\n"
-      "srshl v27.4s, v27.4s, v0.4s\n"
-      "add v31.4s, v31.4s, v29.4s\n"
-      "add v20.4s, v20.4s, v29.4s\n"
-      "add v21.4s, v21.4s, v29.4s\n"
-      "add v22.4s, v22.4s, v29.4s\n"
-      "add v16.4s, v16.4s, v29.4s\n"
-      "add v17.4s, v17.4s, v29.4s\n"
-      "add v18.4s, v18.4s, v29.4s\n"
-      "add v19.4s, v19.4s, v29.4s\n"
-      "add v24.4s, v24.4s, v29.4s\n"
-      "add v25.4s, v25.4s, v29.4s\n"
-      "add v26.4s, v26.4s, v29.4s\n"
-      "add v27.4s, v27.4s, v29.4s\n"
-      "smin v31.4s, v31.4s, v28.4s\n"
-      "smin v20.4s, v20.4s, v28.4s\n"
-      "smin v21.4s, v21.4s, v28.4s\n"
-      "smin v22.4s, v22.4s, v28.4s\n"
-      "smin v16.4s, v16.4s, v28.4s\n"
-      "smin v17.4s, v17.4s, v28.4s\n"
-      "smin v18.4s, v18.4s, v28.4s\n"
-      "smin v19.4s, v19.4s, v28.4s\n"
-      "smin v24.4s, v24.4s, v28.4s\n"
-      "smin v25.4s, v25.4s, v28.4s\n"
-      "smin v26.4s, v26.4s, v28.4s\n"
-      "smin v27.4s, v27.4s, v28.4s\n"
-      "smax v31.4s, v31.4s, v23.4s\n"
-      "smax v20.4s, v20.4s, v23.4s\n"
-      "smax v21.4s, v21.4s, v23.4s\n"
-      "smax v22.4s, v22.4s, v23.4s\n"
-      "smax v16.4s, v16.4s, v23.4s\n"
-      "smax v17.4s, v17.4s, v23.4s\n"
-      "smax v18.4s, v18.4s, v23.4s\n"
-      "smax v19.4s, v19.4s, v23.4s\n"
-      "smax v24.4s, v24.4s, v23.4s\n"
-      "smax v25.4s, v25.4s, v23.4s\n"
-      "smax v26.4s, v26.4s, v23.4s\n"
-      "smax v27.4s, v27.4s, v23.4s\n"
-      "uzp1 v31.8h, v31.8h, v20.8h\n"
-      "uzp1 v20.8h, v21.8h, v22.8h\n"
-      "uzp1 v16.8h, v16.8h, v17.8h\n"
-      "uzp1 v18.8h, v18.8h, v19.8h\n"
-      "uzp1 v24.8h, v24.8h, v25.8h\n"
-      "uzp1 v17.8h, v26.8h, v27.8h\n"
-      "uzp1 v31.16b, v31.16b, v20.16b\n"
-      "uzp1 v16.16b, v16.16b, v18.16b\n"
-      "uzp1 v24.16b, v24.16b, v17.16b\n"
-      "bge 95f\n"
-      "tbz x9, #3, 90f\n"
-      "str d31, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d24, [x23], #0x8\n"
-      "tbz x9, #2, 88f\n"
-      "st1 { v31.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "st1 { v24.s }[2], [x23], #0x4\n"
-      "tbz x9, #1, 87f\n"
-      "st1 { v31.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "st1 { v24.h }[6], [x23], #0x2\n"
-      "tbz x9, #0, 94f\n"
-      "st1 { v31.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "st1 { v24.b }[14], [x23]\n"
-      "b 94f\n"
-      "87:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x9, #0, 94f\n"
-      "st1 { v31.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "st1 { v24.b }[12], [x23]\n"
-      "b 94f\n"
-      "88:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x9, #1, 89f\n"
-      "st1 { v31.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "st1 { v24.h }[4], [x23], #0x2\n"
-      "tbz x9, #0, 94f\n"
-      "st1 { v31.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "st1 { v24.b }[10], [x23]\n"
-      "b 94f\n"
-      "89:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x9, #0, 94f\n"
-      "st1 { v31.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "st1 { v24.b }[8], [x23]\n"
-      "b 94f\n"
-      "90:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x9, #2, 92f\n"
-      "str s31, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "str s24, [x23], #0x4\n"
-      "tbz x9, #1, 91f\n"
-      "st1 { v31.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "st1 { v24.h }[2], [x23], #0x2\n"
-      "tbz x9, #0, 94f\n"
-      "st1 { v31.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "st1 { v24.b }[6], [x23]\n"
-      "b 94f\n"
-      "91:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x9, #0, 94f\n"
-      "st1 { v31.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "st1 { v24.b }[4], [x23]\n"
-      "b 94f\n"
-      "92:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x9, #1, 93f\n"
-      "str h31, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "str h24, [x23], #0x2\n"
-      "tbz x9, #0, 94f\n"
-      "st1 { v31.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "st1 { v24.b }[2], [x23]\n"
-      "b 94f\n"
-      "93:"  // Height 3: Partial direct writeback: partial_1_0
-      "str b31, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "str b24, [x23, #0x0]\n"
-      "94:"  // Height 3: Partial direct writeback: Done
-      "b 96f\n"
-      "95:"  // Height 3: Full writeback
-      "str q31, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
-      "str q24, [x23, #0x0]\n"
-      "96:"  // Height 3: Writeback done
-      "subs x9, x9, #0x10\n"
-      "bgt 66b\n"
-      "b 130f\n"
-      "97:"  // Height 4
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x20, #0x4\n"
-      "mov x10, %x[col_bias]\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
-      "movi v15.16b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "98:"  // Height 4: Column loop
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "movi v28.4s, #0x0\n"
-      "movi v29.4s, #0x0\n"
-      "movi v30.4s, #0x0\n"
-      "movi v31.4s, #0x0\n"
-      "99:"  // Height 4: setup done
-      "mov x26, #0x0\n"
-      "100:"  // Height 4: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 101f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "ldr x22, [x20, #0x10]\n"
-      "ldr x21, [x20, #0x18]\n"
-      "cbnz x26, 102f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "add x21, x21, x20\n"
-      "b 102f\n"
-      "101:"  // Height 4: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "add x21, x22, x21\n"
-      "102:"  // Height 4: input setup done
-      "cmp x25, #0x10\n"
-      "blt 107f\n"
-      "ldr q1, [x24, #0x0]\n"
-      "ldr q2, [x23, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q3, [x22, #0x0]\n"
-      "ldr q4, [x21, #0x0]\n"
-      "ldr q5, [x28, #0x0]\n"
-      "ldr q6, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q8, [x28, #0x30]\n"
-      "ldr q9, [x28, #0x40]\n"
-      "ldr q10, [x28, #0x50]\n"
-      "blt 105f\n"
-      "103:"  // Height 4: Multiply loop: Main loop head
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "ldr q4, [x28, #0x60]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x4e85ac10  // usmmla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e86ac14  // usmmla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e88ac15  // usmmla v21.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e85ac58  // usmmla v24.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0x70]\n"
-      ".inst 0x4e86ac5c  // usmmla v28.4s, v2.16b, v6.16b\n"
-      "ldr q6, [x28, #0x80]\n"
-      ".inst 0x4e87ac59  // usmmla v25.4s, v2.16b, v7.16b\n"
-      "ldr q7, [x28, #0x90]\n"
-      ".inst 0x4e88ac5d  // usmmla v29.4s, v2.16b, v8.16b\n"
-      "ldr q8, [x28, #0xa0]\n"
-      ".inst 0x4e89ac12  // usmmla v18.4s, v0.16b, v9.16b\n"
-      ".inst 0x4e89ac5a  // usmmla v26.4s, v2.16b, v9.16b\n"
-      "ldr q9, [x28, #0xb0]\n"
-      ".inst 0x4e8aac16  // usmmla v22.4s, v0.16b, v10.16b\n"
-      ".inst 0x4e8aac5e  // usmmla v30.4s, v2.16b, v10.16b\n"
-      "ldr q10, [x28, #0xc0]\n"
-      ".inst 0x4e84ac13  // usmmla v19.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e84ac5b  // usmmla v27.4s, v2.16b, v4.16b\n"
-      "ldr q4, [x28, #0xd0]\n"
-      ".inst 0x4e85ac17  // usmmla v23.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85ac5f  // usmmla v31.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0xe0]\n"
-      ".inst 0x4e86ac30  // usmmla v16.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac78  // usmmla v24.4s, v3.16b, v6.16b\n"
-      "ldr q6, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4e87ac34  // usmmla v20.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e87ac7c  // usmmla v28.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e88ac31  // usmmla v17.4s, v1.16b, v8.16b\n"
-      ".inst 0x4e88ac79  // usmmla v25.4s, v3.16b, v8.16b\n"
-      ".inst 0x4e89ac35  // usmmla v21.4s, v1.16b, v9.16b\n"
-      ".inst 0x4e89ac7d  // usmmla v29.4s, v3.16b, v9.16b\n"
-      ".inst 0x4e8aac32  // usmmla v18.4s, v1.16b, v10.16b\n"
-      ".inst 0x4e8aac7a  // usmmla v26.4s, v3.16b, v10.16b\n"
-      ".inst 0x4e84ac36  // usmmla v22.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e84ac7e  // usmmla v30.4s, v3.16b, v4.16b\n"
-      ".inst 0x4e85ac33  // usmmla v19.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e85ac7b  // usmmla v27.4s, v3.16b, v5.16b\n"
-      ".inst 0x4e86ac37  // usmmla v23.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac7f  // usmmla v31.4s, v3.16b, v6.16b\n"
-      "tbnz %x[flags], #31, 104f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      ".inst 0x6e8f942b  // udot v11.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f946d  // udot v13.4s, v3.16b, v15.16b\n"
-      "104:"  // Height 4: Multiply loop: unique 13: skip row sum
-      "ldr q1, [x24, #0x0]\n"
-      "ldr q2, [x23, #0x0]\n"
-      "sub x25, x25, #0x10\n"
-      "ldr q3, [x22, #0x0]\n"
-      "ldr q4, [x21, #0x0]\n"
-      "cmp x25, #0x20\n"
-      "ldr q5, [x28, #0x0]\n"
-      "ldr q6, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q8, [x28, #0x30]\n"
-      "ldr q9, [x28, #0x40]\n"
-      "ldr q10, [x28, #0x50]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      "bge 103b\n"
-      "105:"  // Height 4: Multiply loop: Single iteration only
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "ldr q4, [x28, #0x60]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x4e85ac10  // usmmla v16.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e86ac14  // usmmla v20.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e88ac15  // usmmla v21.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e85ac58  // usmmla v24.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0x70]\n"
-      ".inst 0x4e86ac5c  // usmmla v28.4s, v2.16b, v6.16b\n"
-      "ldr q6, [x28, #0x80]\n"
-      ".inst 0x4e87ac59  // usmmla v25.4s, v2.16b, v7.16b\n"
-      "ldr q7, [x28, #0x90]\n"
-      ".inst 0x4e88ac5d  // usmmla v29.4s, v2.16b, v8.16b\n"
-      "ldr q8, [x28, #0xa0]\n"
-      ".inst 0x4e89ac12  // usmmla v18.4s, v0.16b, v9.16b\n"
-      ".inst 0x4e89ac5a  // usmmla v26.4s, v2.16b, v9.16b\n"
-      "ldr q9, [x28, #0xb0]\n"
-      ".inst 0x4e8aac16  // usmmla v22.4s, v0.16b, v10.16b\n"
-      ".inst 0x4e8aac5e  // usmmla v30.4s, v2.16b, v10.16b\n"
-      "ldr q10, [x28, #0xc0]\n"
-      ".inst 0x4e84ac13  // usmmla v19.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e84ac5b  // usmmla v27.4s, v2.16b, v4.16b\n"
-      "ldr q4, [x28, #0xd0]\n"
-      ".inst 0x4e85ac17  // usmmla v23.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85ac5f  // usmmla v31.4s, v2.16b, v5.16b\n"
-      "ldr q5, [x28, #0xe0]\n"
-      ".inst 0x4e86ac30  // usmmla v16.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac78  // usmmla v24.4s, v3.16b, v6.16b\n"
-      "ldr q6, [x28, #0xf0]\n"
-      "add x28, x28, #0x100\n"
-      ".inst 0x4e87ac34  // usmmla v20.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e87ac7c  // usmmla v28.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e88ac31  // usmmla v17.4s, v1.16b, v8.16b\n"
-      ".inst 0x4e88ac79  // usmmla v25.4s, v3.16b, v8.16b\n"
-      ".inst 0x4e89ac35  // usmmla v21.4s, v1.16b, v9.16b\n"
-      ".inst 0x4e89ac7d  // usmmla v29.4s, v3.16b, v9.16b\n"
-      ".inst 0x4e8aac32  // usmmla v18.4s, v1.16b, v10.16b\n"
-      ".inst 0x4e8aac7a  // usmmla v26.4s, v3.16b, v10.16b\n"
-      ".inst 0x4e84ac36  // usmmla v22.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e84ac7e  // usmmla v30.4s, v3.16b, v4.16b\n"
-      ".inst 0x4e85ac33  // usmmla v19.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e85ac7b  // usmmla v27.4s, v3.16b, v5.16b\n"
-      ".inst 0x4e86ac37  // usmmla v23.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac7f  // usmmla v31.4s, v3.16b, v6.16b\n"
-      "tbnz %x[flags], #31, 106f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      ".inst 0x6e8f942b  // udot v11.4s, v1.16b, v15.16b\n"
-      ".inst 0x6e8f946d  // udot v13.4s, v3.16b, v15.16b\n"
-      "106:"  // Height 4: Multiply loop: unique 14: skip row sum
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      "107:"  // Height 4: Multiply loop: Main loop skip
-      "cbz x25, 116f\n"
-      "cmp x25, #0x8\n"
-      "blt 110f\n"
-      "108:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d3, [x24], #0x8\n"
-      "ldr d0, [x23], #0x8\n"
-      "ldr d2, [x22], #0x8\n"
-      "ldr d1, [x21], #0x8\n"
-      "trn1 v0.2d, v3.2d, v0.2d\n"
-      "trn1 v2.2d, v2.2d, v1.2d\n"
-      "tbnz %x[flags], #31, 109f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      "109:"  // Height 4: Multiply loop: unique 15: skip row sum
-      "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "sub x25, x25, #0x8\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q6, [x28, #0x30]\n"
-      "cmp x25, #0x8\n"
-      "ldr q5, [x28, #0x40]\n"
-      "ldr q4, [x28, #0x50]\n"
-      "ldr q3, [x28, #0x60]\n"
-      ".inst 0x4e81ac10  // usmmla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81ac58  // usmmla v24.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x28, #0x70]\n"
-      ".inst 0x4e88ac14  // usmmla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e88ac5c  // usmmla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac59  // usmmla v25.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e86ac15  // usmmla v21.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac5d  // usmmla v29.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e85ac12  // usmmla v18.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85ac5a  // usmmla v26.4s, v2.16b, v5.16b\n"
-      ".inst 0x4e84ac16  // usmmla v22.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e84ac5e  // usmmla v30.4s, v2.16b, v4.16b\n"
-      ".inst 0x4e83ac13  // usmmla v19.4s, v0.16b, v3.16b\n"
-      ".inst 0x4e83ac5b  // usmmla v27.4s, v2.16b, v3.16b\n"
-      ".inst 0x4e81ac17  // usmmla v23.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81ac5f  // usmmla v31.4s, v2.16b, v1.16b\n"
-      "bge 108b\n"
-      "110:"  // Height 4: Multiply loop: Skip odd blocks
-      "cbz x25, 116f\n"
-      "tbz x25, #2, 112f\n"
-      "ldr s1, [x24], #0x4\n"
-      "ldr s2, [x23], #0x4\n"
-      "ldr s3, [x22], #0x4\n"
-      "ldr s9, [x21], #0x4\n"
-      "tbz x25, #1, 111f\n"
-      "ld1 { v1.h }[2], [x24], #0x2\n"
-      "ld1 { v2.h }[2], [x23], #0x2\n"
-      "ld1 { v3.h }[2], [x22], #0x2\n"
-      "ld1 { v9.h }[2], [x21], #0x2\n"
-      "tbz x25, #0, 114f\n"
-      "ld1 { v1.b }[6], [x24]\n"
-      "ld1 { v2.b }[6], [x23]\n"
-      "ld1 { v3.b }[6], [x22]\n"
-      "ld1 { v9.b }[6], [x21]\n"
-      "b 114f\n"
-      "111:"  // Height 4: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x25, #0, 114f\n"
-      "ld1 { v1.b }[4], [x24]\n"
-      "ld1 { v2.b }[4], [x23]\n"
-      "ld1 { v3.b }[4], [x22]\n"
-      "ld1 { v9.b }[4], [x21]\n"
-      "b 114f\n"
-      "112:"  // Height 4: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x25, #1, 113f\n"
-      "ldr h1, [x24], #0x2\n"
-      "ldr h2, [x23], #0x2\n"
-      "ldr h3, [x22], #0x2\n"
-      "ldr h9, [x21], #0x2\n"
-      "tbz x25, #0, 114f\n"
-      "ld1 { v1.b }[2], [x24]\n"
-      "ld1 { v2.b }[2], [x23]\n"
-      "ld1 { v3.b }[2], [x22]\n"
-      "ld1 { v9.b }[2], [x21]\n"
-      "b 114f\n"
-      "113:"  // Height 4: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x24, #0x0]\n"
-      "ldr b2, [x23, #0x0]\n"
-      "ldr b3, [x22, #0x0]\n"
-      "ldr b9, [x21, #0x0]\n"
-      "114:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn1 v2.2d, v3.2d, v9.2d\n"
-      "tbnz %x[flags], #31, 115f\n"
-      ".inst 0x6e8f940b  // udot v11.4s, v0.16b, v15.16b\n"
-      ".inst 0x6e8f944d  // udot v13.4s, v2.16b, v15.16b\n"
-      "115:"  // Height 4: Multiply loop: unique 16: skip row sum
-      "ldr q1, [x28, #0x0]\n"
-      "ldr q8, [x28, #0x10]\n"
-      "ldr q7, [x28, #0x20]\n"
-      "ldr q6, [x28, #0x30]\n"
-      "ldr q5, [x28, #0x40]\n"
-      "ldr q4, [x28, #0x50]\n"
-      "ldr q3, [x28, #0x60]\n"
-      ".inst 0x4e81ac10  // usmmla v16.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81ac58  // usmmla v24.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x28, #0x70]\n"
-      ".inst 0x4e88ac14  // usmmla v20.4s, v0.16b, v8.16b\n"
-      ".inst 0x4e88ac5c  // usmmla v28.4s, v2.16b, v8.16b\n"
-      "add x28, x28, #0x80\n"
-      ".inst 0x4e87ac11  // usmmla v17.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac59  // usmmla v25.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e86ac15  // usmmla v21.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac5d  // usmmla v29.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e85ac12  // usmmla v18.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e85ac5a  // usmmla v26.4s, v2.16b, v5.16b\n"
-      ".inst 0x4e84ac16  // usmmla v22.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e84ac5e  // usmmla v30.4s, v2.16b, v4.16b\n"
-      ".inst 0x4e83ac13  // usmmla v19.4s, v0.16b, v3.16b\n"
-      ".inst 0x4e83ac5b  // usmmla v27.4s, v2.16b, v3.16b\n"
-      ".inst 0x4e81ac17  // usmmla v23.4s, v0.16b, v1.16b\n"
-      ".inst 0x4e81ac5f  // usmmla v31.4s, v2.16b, v1.16b\n"
-      "116:"  // Height 4: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 100b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 v0.2d, v16.2d, v20.2d\n"
-      "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x27, #0x0]\n"
-      "uzp1 v20.2d, v17.2d, v21.2d\n"
-      "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "uzp1 v21.2d, v18.2d, v22.2d\n"
-      "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
-      "add x22, x23, x20\n"
-      "uzp1 v22.2d, v19.2d, v23.2d\n"
-      "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "uzp1 v23.2d, v24.2d, v28.2d\n"
-      "uzp2 v24.2d, v24.2d, v28.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "uzp1 v28.2d, v25.2d, v29.2d\n"
-      "uzp2 v25.2d, v25.2d, v29.2d\n"
-      "uzp1 v29.2d, v26.2d, v30.2d\n"
-      "uzp2 v26.2d, v26.2d, v30.2d\n"
-      "uzp1 v30.2d, v27.2d, v31.2d\n"
-      "uzp2 v27.2d, v27.2d, v31.2d\n"
-      "mov v31.16b, v0.16b\n"
-      "tbnz %x[flags], #31, 117f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "addp v11.4s, v11.4s, v11.4s\n"
-      "addp v13.4s, v13.4s, v13.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "neg v0.4s, v0.4s\n"
-      "dup v12.4s, v11.s[3]\n"
-      "dup v11.4s, v11.s[0]\n"
-      "dup v14.4s, v13.s[3]\n"
-      "dup v13.4s, v13.s[0]\n"
-      "mul v11.4s, v11.4s, v0.4s\n"
-      "mul v12.4s, v12.4s, v0.4s\n"
-      "mul v14.4s, v14.4s, v0.4s\n"
-      "mul v13.4s, v13.4s, v0.4s\n"
-      "117:"  // Height 4: skip row sum fixup
-      "ldr q0, [x10, #0x0]\n"
-      "ldr q4, [x10, #0x10]\n"
-      "add v31.4s, v31.4s, v11.4s\n"
-      "add v20.4s, v20.4s, v11.4s\n"
-      "ldr q3, [x10, #0x20]\n"
-      "ldr q2, [x10, #0x30]\n"
-      "add v21.4s, v21.4s, v11.4s\n"
-      "add v22.4s, v22.4s, v11.4s\n"
-      "add v16.4s, v16.4s, v12.4s\n"
-      "add v17.4s, v17.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "ld1r { v1.4s }, [x20]\n"
-      "add v18.4s, v18.4s, v12.4s\n"
-      "add v19.4s, v19.4s, v12.4s\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add v23.4s, v23.4s, v13.4s\n"
-      "add v28.4s, v28.4s, v13.4s\n"
-      "add x10, x10, #0x40\n"
-      "add v29.4s, v29.4s, v13.4s\n"
-      "add v30.4s, v30.4s, v13.4s\n"
-      "add v24.4s, v24.4s, v14.4s\n"
-      "add v25.4s, v25.4s, v14.4s\n"
-      "add v26.4s, v26.4s, v14.4s\n"
-      "add v27.4s, v27.4s, v14.4s\n"
-      "add v31.4s, v31.4s, v0.4s\n"
-      "add v20.4s, v20.4s, v4.4s\n"
-      "add v21.4s, v21.4s, v3.4s\n"
-      "add v22.4s, v22.4s, v2.4s\n"
-      "add v16.4s, v16.4s, v0.4s\n"
-      "add v17.4s, v17.4s, v4.4s\n"
-      "add v18.4s, v18.4s, v3.4s\n"
-      "add v19.4s, v19.4s, v2.4s\n"
-      "add v23.4s, v23.4s, v0.4s\n"
-      "add v28.4s, v28.4s, v4.4s\n"
-      "add v29.4s, v29.4s, v3.4s\n"
-      "add v30.4s, v30.4s, v2.4s\n"
-      "add v24.4s, v24.4s, v0.4s\n"
-      "ld1r { v0.4s }, [x20]\n"
-      "add v25.4s, v25.4s, v4.4s\n"
-      "add v26.4s, v26.4s, v3.4s\n"
-      "add v27.4s, v27.4s, v2.4s\n"
-      "sqrdmulh v31.4s, v31.4s, v1.4s\n"
-      "sqrdmulh v20.4s, v20.4s, v1.4s\n"
-      "sqrdmulh v21.4s, v21.4s, v1.4s\n"
-      "sqrdmulh v22.4s, v22.4s, v1.4s\n"
-      "sqrdmulh v16.4s, v16.4s, v1.4s\n"
-      "sqrdmulh v17.4s, v17.4s, v1.4s\n"
-      "sqrdmulh v18.4s, v18.4s, v1.4s\n"
-      "sqrdmulh v19.4s, v19.4s, v1.4s\n"
-      "sqrdmulh v23.4s, v23.4s, v1.4s\n"
-      "sqrdmulh v28.4s, v28.4s, v1.4s\n"
-      "sqrdmulh v29.4s, v29.4s, v1.4s\n"
-      "sqrdmulh v30.4s, v30.4s, v1.4s\n"
-      "sqrdmulh v24.4s, v24.4s, v1.4s\n"
-      "sqrdmulh v25.4s, v25.4s, v1.4s\n"
-      "sqrdmulh v26.4s, v26.4s, v1.4s\n"
-      "sqrdmulh v27.4s, v27.4s, v1.4s\n"
-      "tbz %x[flags], #5, 118f\n"
-      "and v2.16b, v31.16b, v0.16b\n"
-      "and v1.16b, v20.16b, v0.16b\n"
-      "and v7.16b, v21.16b, v0.16b\n"
-      "and v6.16b, v22.16b, v0.16b\n"
-      "and v5.16b, v16.16b, v0.16b\n"
-      "and v4.16b, v17.16b, v0.16b\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "and v3.16b, v18.16b, v0.16b\n"
-      "sshr v7.4s, v7.4s, #0x1f\n"
-      "sshr v6.4s, v6.4s, #0x1f\n"
-      "sshr v5.4s, v5.4s, #0x1f\n"
-      "sqadd v31.4s, v31.4s, v2.4s\n"
-      "sqadd v20.4s, v20.4s, v1.4s\n"
-      "and v2.16b, v19.16b, v0.16b\n"
-      "and v1.16b, v23.16b, v0.16b\n"
-      "sshr v4.4s, v4.4s, #0x1f\n"
-      "sshr v3.4s, v3.4s, #0x1f\n"
-      "sqadd v21.4s, v21.4s, v7.4s\n"
-      "sqadd v22.4s, v22.4s, v6.4s\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "sqadd v16.4s, v16.4s, v5.4s\n"
-      "sqadd v17.4s, v17.4s, v4.4s\n"
-      "sqadd v18.4s, v18.4s, v3.4s\n"
-      "and v7.16b, v28.16b, v0.16b\n"
-      "sqadd v19.4s, v19.4s, v2.4s\n"
-      "sqadd v23.4s, v23.4s, v1.4s\n"
-      "and v6.16b, v29.16b, v0.16b\n"
-      "and v5.16b, v30.16b, v0.16b\n"
-      "and v4.16b, v24.16b, v0.16b\n"
-      "and v3.16b, v25.16b, v0.16b\n"
-      "and v2.16b, v26.16b, v0.16b\n"
-      "and v1.16b, v27.16b, v0.16b\n"
-      "sshr v7.4s, v7.4s, #0x1f\n"
-      "sshr v6.4s, v6.4s, #0x1f\n"
-      "sshr v5.4s, v5.4s, #0x1f\n"
-      "sshr v4.4s, v4.4s, #0x1f\n"
-      "sshr v3.4s, v3.4s, #0x1f\n"
-      "sshr v2.4s, v2.4s, #0x1f\n"
-      "sshr v1.4s, v1.4s, #0x1f\n"
-      "sqadd v28.4s, v28.4s, v7.4s\n"
-      "sqadd v29.4s, v29.4s, v6.4s\n"
-      "sqadd v30.4s, v30.4s, v5.4s\n"
-      "sqadd v24.4s, v24.4s, v4.4s\n"
-      "sqadd v25.4s, v25.4s, v3.4s\n"
-      "sqadd v26.4s, v26.4s, v2.4s\n"
-      "sqadd v27.4s, v27.4s, v1.4s\n"
-      "118:"  // Height 4: no shift correction
-      "add x21, %x[qp], %[c_offset]\n"
-      "srshl v31.4s, v31.4s, v0.4s\n"
-      "srshl v20.4s, v20.4s, v0.4s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1r { v3.4s }, [x21]\n"
-      "ld1r { v2.4s }, [x20]\n"
-      "srshl v21.4s, v21.4s, v0.4s\n"
-      "srshl v22.4s, v22.4s, v0.4s\n"
-      "srshl v16.4s, v16.4s, v0.4s\n"
-      "srshl v17.4s, v17.4s, v0.4s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "cmp x9, #0x10\n"
-      "ld1r { v1.4s }, [x20]\n"
-      "srshl v18.4s, v18.4s, v0.4s\n"
-      "srshl v19.4s, v19.4s, v0.4s\n"
-      "srshl v23.4s, v23.4s, v0.4s\n"
-      "srshl v28.4s, v28.4s, v0.4s\n"
-      "srshl v29.4s, v29.4s, v0.4s\n"
-      "srshl v30.4s, v30.4s, v0.4s\n"
-      "srshl v24.4s, v24.4s, v0.4s\n"
-      "srshl v25.4s, v25.4s, v0.4s\n"
-      "srshl v26.4s, v26.4s, v0.4s\n"
-      "srshl v27.4s, v27.4s, v0.4s\n"
-      "add v31.4s, v31.4s, v3.4s\n"
-      "add v20.4s, v20.4s, v3.4s\n"
-      "add v21.4s, v21.4s, v3.4s\n"
-      "add v22.4s, v22.4s, v3.4s\n"
-      "add v16.4s, v16.4s, v3.4s\n"
-      "add v17.4s, v17.4s, v3.4s\n"
-      "add v18.4s, v18.4s, v3.4s\n"
-      "add v19.4s, v19.4s, v3.4s\n"
-      "add v23.4s, v23.4s, v3.4s\n"
-      "add v28.4s, v28.4s, v3.4s\n"
-      "add v29.4s, v29.4s, v3.4s\n"
-      "add v30.4s, v30.4s, v3.4s\n"
-      "add v24.4s, v24.4s, v3.4s\n"
-      "add v25.4s, v25.4s, v3.4s\n"
-      "add v26.4s, v26.4s, v3.4s\n"
-      "add v27.4s, v27.4s, v3.4s\n"
-      "smin v31.4s, v31.4s, v2.4s\n"
-      "smin v20.4s, v20.4s, v2.4s\n"
-      "smin v21.4s, v21.4s, v2.4s\n"
-      "smin v22.4s, v22.4s, v2.4s\n"
-      "smin v16.4s, v16.4s, v2.4s\n"
-      "smin v17.4s, v17.4s, v2.4s\n"
-      "smin v18.4s, v18.4s, v2.4s\n"
-      "smin v19.4s, v19.4s, v2.4s\n"
-      "smin v23.4s, v23.4s, v2.4s\n"
-      "smin v28.4s, v28.4s, v2.4s\n"
-      "smin v29.4s, v29.4s, v2.4s\n"
-      "smin v30.4s, v30.4s, v2.4s\n"
-      "smin v24.4s, v24.4s, v2.4s\n"
-      "smin v25.4s, v25.4s, v2.4s\n"
-      "smin v26.4s, v26.4s, v2.4s\n"
-      "smin v27.4s, v27.4s, v2.4s\n"
-      "smax v31.4s, v31.4s, v1.4s\n"
-      "smax v20.4s, v20.4s, v1.4s\n"
-      "smax v21.4s, v21.4s, v1.4s\n"
-      "smax v22.4s, v22.4s, v1.4s\n"
-      "smax v16.4s, v16.4s, v1.4s\n"
-      "smax v17.4s, v17.4s, v1.4s\n"
-      "smax v18.4s, v18.4s, v1.4s\n"
-      "smax v19.4s, v19.4s, v1.4s\n"
-      "smax v23.4s, v23.4s, v1.4s\n"
-      "smax v28.4s, v28.4s, v1.4s\n"
-      "smax v29.4s, v29.4s, v1.4s\n"
-      "smax v30.4s, v30.4s, v1.4s\n"
-      "smax v24.4s, v24.4s, v1.4s\n"
-      "smax v25.4s, v25.4s, v1.4s\n"
-      "smax v26.4s, v26.4s, v1.4s\n"
-      "smax v27.4s, v27.4s, v1.4s\n"
-      "uzp1 v31.8h, v31.8h, v20.8h\n"
-      "uzp1 v20.8h, v21.8h, v22.8h\n"
-      "uzp1 v16.8h, v16.8h, v17.8h\n"
-      "uzp1 v19.8h, v18.8h, v19.8h\n"
-      "uzp1 v23.8h, v23.8h, v28.8h\n"
-      "uzp1 v18.8h, v29.8h, v30.8h\n"
-      "uzp1 v24.8h, v24.8h, v25.8h\n"
-      "uzp1 v17.8h, v26.8h, v27.8h\n"
-      "uzp1 v31.16b, v31.16b, v20.16b\n"
-      "uzp1 v16.16b, v16.16b, v19.16b\n"
-      "uzp1 v23.16b, v23.16b, v18.16b\n"
-      "uzp1 v24.16b, v24.16b, v17.16b\n"
-      "bge 127f\n"
-      "tbz x9, #3, 122f\n"
-      "str d31, [x27], #0x8\n"
-      "str d16, [x24], #0x8\n"
-      "str d23, [x23], #0x8\n"
-      "str d24, [x22], #0x8\n"
-      "tbz x9, #2, 120f\n"
-      "st1 { v31.s }[2], [x27], #0x4\n"
-      "st1 { v16.s }[2], [x24], #0x4\n"
-      "st1 { v23.s }[2], [x23], #0x4\n"
-      "st1 { v24.s }[2], [x22], #0x4\n"
-      "tbz x9, #1, 119f\n"
-      "st1 { v31.h }[6], [x27], #0x2\n"
-      "st1 { v16.h }[6], [x24], #0x2\n"
-      "st1 { v23.h }[6], [x23], #0x2\n"
-      "st1 { v24.h }[6], [x22], #0x2\n"
-      "tbz x9, #0, 126f\n"
-      "st1 { v31.b }[14], [x27]\n"
-      "st1 { v16.b }[14], [x24]\n"
-      "st1 { v23.b }[14], [x23]\n"
-      "st1 { v24.b }[14], [x22]\n"
-      "b 126f\n"
-      "119:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x9, #0, 126f\n"
-      "st1 { v31.b }[12], [x27]\n"
-      "st1 { v16.b }[12], [x24]\n"
-      "st1 { v23.b }[12], [x23]\n"
-      "st1 { v24.b }[12], [x22]\n"
-      "b 126f\n"
-      "120:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x9, #1, 121f\n"
-      "st1 { v31.h }[4], [x27], #0x2\n"
-      "st1 { v16.h }[4], [x24], #0x2\n"
-      "st1 { v23.h }[4], [x23], #0x2\n"
-      "st1 { v24.h }[4], [x22], #0x2\n"
-      "tbz x9, #0, 126f\n"
-      "st1 { v31.b }[10], [x27]\n"
-      "st1 { v16.b }[10], [x24]\n"
-      "st1 { v23.b }[10], [x23]\n"
-      "st1 { v24.b }[10], [x22]\n"
-      "b 126f\n"
-      "121:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x9, #0, 126f\n"
-      "st1 { v31.b }[8], [x27]\n"
-      "st1 { v16.b }[8], [x24]\n"
-      "st1 { v23.b }[8], [x23]\n"
-      "st1 { v24.b }[8], [x22]\n"
-      "b 126f\n"
-      "122:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x9, #2, 124f\n"
-      "str s31, [x27], #0x4\n"
-      "str s16, [x24], #0x4\n"
-      "str s23, [x23], #0x4\n"
-      "str s24, [x22], #0x4\n"
-      "tbz x9, #1, 123f\n"
-      "st1 { v31.h }[2], [x27], #0x2\n"
-      "st1 { v16.h }[2], [x24], #0x2\n"
-      "st1 { v23.h }[2], [x23], #0x2\n"
-      "st1 { v24.h }[2], [x22], #0x2\n"
-      "tbz x9, #0, 126f\n"
-      "st1 { v31.b }[6], [x27]\n"
-      "st1 { v16.b }[6], [x24]\n"
-      "st1 { v23.b }[6], [x23]\n"
-      "st1 { v24.b }[6], [x22]\n"
-      "b 126f\n"
-      "123:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x9, #0, 126f\n"
-      "st1 { v31.b }[4], [x27]\n"
-      "st1 { v16.b }[4], [x24]\n"
-      "st1 { v23.b }[4], [x23]\n"
-      "st1 { v24.b }[4], [x22]\n"
-      "b 126f\n"
-      "124:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x9, #1, 125f\n"
-      "str h31, [x27], #0x2\n"
-      "str h16, [x24], #0x2\n"
-      "str h23, [x23], #0x2\n"
-      "str h24, [x22], #0x2\n"
-      "tbz x9, #0, 126f\n"
-      "st1 { v31.b }[2], [x27]\n"
-      "st1 { v16.b }[2], [x24]\n"
-      "st1 { v23.b }[2], [x23]\n"
-      "st1 { v24.b }[2], [x22]\n"
-      "b 126f\n"
-      "125:"  // Height 4: Partial direct writeback: partial_1_0
-      "str b31, [x27, #0x0]\n"
-      "str b16, [x24, #0x0]\n"
-      "str b23, [x23, #0x0]\n"
-      "str b24, [x22, #0x0]\n"
-      "126:"  // Height 4: Partial direct writeback: Done
-      "b 128f\n"
-      "127:"  // Height 4: Full writeback
-      "str q31, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "str q16, [x24, #0x0]\n"
-      "str q23, [x23, #0x0]\n"
-      "str q24, [x22, #0x0]\n"
-      "128:"  // Height 4: Writeback done
-      "subs x9, x9, #0x10\n"
-      "bgt 98b\n"
-      "subs %x[M], %x[M], #0x4\n"
-      "beq 130f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "tbz %x[flags], #3, 129f\n"
-      "add x21, x21, #0x4\n"
-      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "b 1b\n"
-      "129:"  // Update direct input
-      "mov x20, #0x4\n"
-      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
-      "b 1b\n"
-      "130:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
-    );
-}
-
-} // namespace arm_gemm
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_dot_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_dot_6x16.hpp
deleted file mode 100644
index 4f963124cb..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_dot_6x16.hpp
+++ /dev/null
@@ -1,116 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef __aarch64__
-
-#include "../std_transforms_fixed.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    unsigned int, const unsigned int *, \
-    IndirectInputArg<uint8_t>, \
-    size_t, size_t, \
-    const int8_t *, \
-    IndirectOutputArg<int32_t>, \
-    const int32_t *, Activation, bool
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void a64_hybrid_u8s8s32_dot_6x16( ARGLIST );
-
-class cls_a64_hybrid_u8s8s32_dot_6x16
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef int32_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 6;
-    }
-
-    static unsigned int out_width()
-    {
-        return 16;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 4;
-    }
-
-    static constexpr bool supports_accumulate()
-    {
-        return true;
-    }
-
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 4> transforms = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-        if (std::is_same<T, uint32_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 31.63 };
-                case CPUModel::A510:
-                    return { 15.89 };
-                case CPUModel::V1:
-                    return { 53.87 };
-                case CPUModel::A55r1:
-                    return { 9.217 };
-            }
-        }
-
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                case CPUModel::A55r1:
-                    return { 9.5238, 2.0799, 0.2279 };
-                default:
-                    return { 29.6736, 11.4025, 0.5591 };
-                case CPUModel::A510:
-                    return { 16.65, 3.92, 0.48 };
-                case CPUModel::V1:
-                    return { 42.62, 16.32, 0.83 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=a64_hybrid_u8s8s32_dot_6x16;
-    cls_a64_hybrid_u8s8s32_dot_6x16(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_dot_6x16/generic.cpp
deleted file mode 100644
index 074a9585d8..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_dot_6x16/generic.cpp
+++ /dev/null
@@ -1,3264 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef __aarch64__
-
-#include "arm_gemm.hpp"
-#include "../../utils.hpp"
-
-#include <cassert>
-
-namespace arm_gemm {
-
-void a64_hybrid_u8s8s32_dot_6x16 (
-    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg,
-    size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<int32_t> output_arg,
-    const int32_t *, Activation, bool accumulate
-)
-{
-    struct KernelArgs {
-        unsigned int num_strings = {};
-        const unsigned int *string_lengths = {};
-        size_t N = {};
-        const int8_t *B_ptr = {};
-        size_t output_offset = {};
-        size_t input_initial_col = {};
-        size_t input_offset = {};
-        void *output_ptr = {};
-    } ka;
-
-    unsigned long flags=0;
-    void *input_ptr;
-
-    if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
-        ka.output_offset=output_arg.indirect.offset;
-        flags |= 0x4;
-    } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
-        ka.output_offset=output_arg.direct.stride;
-    }
-
-    if (A_arg.is_indirect) {
-        input_ptr=(void *)(A_arg.indirect.ptr);
-        ka.input_offset=A_arg.indirect.start_row;
-        ka.input_initial_col=A_arg.indirect.start_col;
-        flags |= 0x8;
-    } else {
-        assert(num_strings==1);
-        input_ptr=(void *)(A_arg.direct.base);
-        ka.input_offset=A_arg.direct.stride;
-    }
-    if (accumulate) {
-        flags |= 0x1;
-    }
-    ka.num_strings = num_strings;
-    ka.string_lengths = string_lengths;
-    ka.N = N;
-    ka.B_ptr = B_ptr;
-    __asm__ __volatile__(
-      "1:"  // Row loop
-      "cmp %x[M], #0x6\n"
-      "bge 171f\n"
-      "cmp %x[M], #0x4\n"
-      "bgt 137f\n"
-      "beq 103f\n"
-      "cmp %x[M], #0x2\n"
-      "bgt 69f\n"
-      "beq 35f\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "2:"  // Height 1: Column loop
-      "tbz %x[flags], #0, 12f\n"
-      "cmp x11, #0x10\n"
-      "bge 11f\n"
-      "tbz x11, #3, 6f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "tbz x11, #2, 4f\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "tbz x11, #1, 3f\n"
-      "ldr d11, [x9], #0x8\n"
-      "mov x25, #0x38\n"
-      "tbz x11, #0, 10f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "b 10f\n"
-      "3:"  // Height 1: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 10f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "b 10f\n"
-      "4:"  // Height 1: Partial accumulate: partial_2_8
-      "tbz x11, #1, 5f\n"
-      "ldr d10, [x9], #0x8\n"
-      "mov x25, #0x28\n"
-      "tbz x11, #0, 10f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "b 10f\n"
-      "5:"  // Height 1: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 10f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "b 10f\n"
-      "6:"  // Height 1: Partial accumulate: partial_4_0
-      "tbz x11, #2, 8f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "tbz x11, #1, 7f\n"
-      "ldr d9, [x9], #0x8\n"
-      "mov x25, #0x18\n"
-      "tbz x11, #0, 10f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "b 10f\n"
-      "7:"  // Height 1: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 10f\n"
-      "ldr s9, [x9, #0x0]\n"
-      "b 10f\n"
-      "8:"  // Height 1: Partial accumulate: partial_2_0
-      "tbz x11, #1, 9f\n"
-      "ldr d8, [x9], #0x8\n"
-      "mov x25, #0x8\n"
-      "tbz x11, #0, 10f\n"
-      "ld1 { v8.s }[2], [x9]\n"
-      "b 10f\n"
-      "9:"  // Height 1: Partial accumulate: partial_1_0
-      "ldr s8, [x9, #0x0]\n"
-      "mov x25, #0x0\n"
-      "10:"  // Height 1: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 13f\n"
-      "11:"  // Height 1: full accumulate
-      "ldr q8, [x9, #0x0]\n"
-      "ldr q9, [x9, #0x10]\n"
-      "ldr q10, [x9, #0x20]\n"
-      "ldr q11, [x9, #0x30]\n"
-      "b 13f\n"
-      "12:"  // Height 1: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "13:"  // Height 1: setup done
-      "mov x28, #0x0\n"
-      "14:"  // Height 1: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 15f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "cbnz x28, 16f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "b 16f\n"
-      "15:"  // Height 1: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "16:"  // Height 1: input setup done
-      "cmp x27, #0x10\n"
-      "blt 19f\n"
-      "ldr q0, [x26, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q7, [x10, #0x10]\n"
-      "blt 18f\n"
-      "17:"  // Height 1: Multiply loop: Main loop head
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4f00f22a  // sudot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x40]\n"
-      ".inst 0x4f00f20b  // sudot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x50]\n"
-      ".inst 0x4f20f228  // sudot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x10, #0x60]\n"
-      ".inst 0x4f20f209  // sudot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x10, #0x70]\n"
-      ".inst 0x4f20f22a  // sudot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x10, #0x80]\n"
-      ".inst 0x4f20f20b  // sudot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x10, #0x90]\n"
-      ".inst 0x4f00fa28  // sudot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x10, #0xa0]\n"
-      ".inst 0x4f00fa09  // sudot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x10, #0xb0]\n"
-      ".inst 0x4f00fa2a  // sudot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x10, #0xc0]\n"
-      ".inst 0x4f00fa0b  // sudot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x10, #0xd0]\n"
-      ".inst 0x4f20fa28  // sudot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr q17, [x10, #0xe0]\n"
-      ".inst 0x4f20fa09  // sudot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20fa2a  // sudot v10.4s, v17.16b, v0.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
-      ".inst 0x4f20fa0b  // sudot v11.4s, v16.16b, v0.4b[3]\n"
-      "ldr q0, [x26, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "bge 17b\n"
-      "18:"  // Height 1: Multiply loop: Single iteration only
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4f00f22a  // sudot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x40]\n"
-      ".inst 0x4f00f20b  // sudot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x50]\n"
-      ".inst 0x4f20f228  // sudot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x10, #0x60]\n"
-      ".inst 0x4f20f209  // sudot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x10, #0x70]\n"
-      ".inst 0x4f20f22a  // sudot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x10, #0x80]\n"
-      ".inst 0x4f20f20b  // sudot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x10, #0x90]\n"
-      ".inst 0x4f00fa28  // sudot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x10, #0xa0]\n"
-      ".inst 0x4f00fa09  // sudot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x10, #0xb0]\n"
-      ".inst 0x4f00fa2a  // sudot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x10, #0xc0]\n"
-      ".inst 0x4f00fa0b  // sudot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x10, #0xd0]\n"
-      ".inst 0x4f20fa28  // sudot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr q17, [x10, #0xe0]\n"
-      ".inst 0x4f20fa09  // sudot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20fa2a  // sudot v10.4s, v17.16b, v0.4b[3]\n"
-      ".inst 0x4f20fa0b  // sudot v11.4s, v16.16b, v0.4b[3]\n"
-      "19:"  // Height 1: Multiply loop: Main loop skip
-      "cbz x27, 24f\n"
-      "cmp x27, #0x4\n"
-      "blt 21f\n"
-      "20:"  // Height 1: Multiply loop: Odd block loop
-      "ldr s18, [x26], #0x4\n"
-      "ldr q17, [x10, #0x0]\n"
-      "sub x27, x27, #0x4\n"
-      "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
-      ".inst 0x4f12f228  // sudot v8.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4f12f209  // sudot v9.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f12f22a  // sudot v10.4s, v17.16b, v18.4b[0]\n"
-      ".inst 0x4f12f20b  // sudot v11.4s, v16.16b, v18.4b[0]\n"
-      "bge 20b\n"
-      "21:"  // Height 1: Multiply loop: Skip odd blocks
-      "cbz x27, 24f\n"
-      "tbz x27, #1, 22f\n"
-      "ldr h0, [x26], #0x2\n"
-      "tbz x27, #0, 23f\n"
-      "ld1 { v0.b }[2], [x26]\n"
-      "b 23f\n"
-      "22:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x26, #0x0]\n"
-      "23:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q16, [x10, #0x10]\n"
-      ".inst 0x4f00f228  // sudot v8.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4f00f209  // sudot v9.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f00f22a  // sudot v10.4s, v17.16b, v0.4b[0]\n"
-      ".inst 0x4f00f20b  // sudot v11.4s, v16.16b, v0.4b[0]\n"
-      "24:"  // Height 1: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 14b\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "bge 33f\n"
-      "tbz x11, #3, 28f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v9.4s }, [x9], #0x10\n"
-      "tbz x11, #2, 26f\n"
-      "st1 { v10.4s }, [x9], #0x10\n"
-      "tbz x11, #1, 25f\n"
-      "str d11, [x9], #0x8\n"
-      "tbz x11, #0, 32f\n"
-      "st1 { v11.s }[2], [x9]\n"
-      "b 32f\n"
-      "25:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 32f\n"
-      "str s11, [x9, #0x0]\n"
-      "b 32f\n"
-      "26:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 27f\n"
-      "str d10, [x9], #0x8\n"
-      "tbz x11, #0, 32f\n"
-      "st1 { v10.s }[2], [x9]\n"
-      "b 32f\n"
-      "27:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 32f\n"
-      "str s10, [x9, #0x0]\n"
-      "b 32f\n"
-      "28:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 30f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "tbz x11, #1, 29f\n"
-      "str d9, [x9], #0x8\n"
-      "tbz x11, #0, 32f\n"
-      "st1 { v9.s }[2], [x9]\n"
-      "b 32f\n"
-      "29:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 32f\n"
-      "str s9, [x9, #0x0]\n"
-      "b 32f\n"
-      "30:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 31f\n"
-      "str d8, [x9], #0x8\n"
-      "tbz x11, #0, 32f\n"
-      "st1 { v8.s }[2], [x9]\n"
-      "b 32f\n"
-      "31:"  // Height 1: Partial direct writeback: partial_1_0
-      "str s8, [x9, #0x0]\n"
-      "32:"  // Height 1: Partial direct writeback: Done
-      "b 34f\n"
-      "33:"  // Height 1: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "str q9, [x9, #0x10]\n"
-      "str q10, [x9, #0x20]\n"
-      "str q11, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "34:"  // Height 1: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 2b\n"
-      "b 206f\n"
-      "35:"  // Height 2
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "36:"  // Height 2: Column loop
-      "tbz %x[flags], #0, 46f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "bge 45f\n"
-      "tbz x11, #3, 40f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "tbz x11, #2, 38f\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "tbz x11, #1, 37f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "tbz x11, #0, 44f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "b 44f\n"
-      "37:"  // Height 2: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 44f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "b 44f\n"
-      "38:"  // Height 2: Partial accumulate: partial_2_8
-      "tbz x11, #1, 39f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "tbz x11, #0, 44f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "b 44f\n"
-      "39:"  // Height 2: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 44f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "b 44f\n"
-      "40:"  // Height 2: Partial accumulate: partial_4_0
-      "tbz x11, #2, 42f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "tbz x11, #1, 41f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "tbz x11, #0, 44f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "b 44f\n"
-      "41:"  // Height 2: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 44f\n"
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "b 44f\n"
-      "42:"  // Height 2: Partial accumulate: partial_2_0
-      "tbz x11, #1, 43f\n"
-      "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "tbz x11, #0, 44f\n"
-      "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "b 44f\n"
-      "43:"  // Height 2: Partial accumulate: partial_1_0
-      "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "44:"  // Height 2: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 47f\n"
-      "45:"  // Height 2: full accumulate
-      "ldr q8, [x9, #0x0]\n"
-      "ldr q9, [x9, #0x10]\n"
-      "ldr q10, [x9, #0x20]\n"
-      "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "b 47f\n"
-      "46:"  // Height 2: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "47:"  // Height 2: setup done
-      "mov x28, #0x0\n"
-      "48:"  // Height 2: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 49f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "cbnz x28, 50f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "b 50f\n"
-      "49:"  // Height 2: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "50:"  // Height 2: input setup done
-      "cmp x27, #0x10\n"
-      "blt 53f\n"
-      "ldr q0, [x26, #0x0]\n"
-      "ldr q1, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "blt 52f\n"
-      "51:"  // Height 2: Multiply loop: Main loop head
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      "sub x27, x27, #0x10\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4f00f22a  // sudot v10.4s, v17.16b, v0.4b[0]\n"
-      ".inst 0x4f01f22e  // sudot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x40]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4f00f20b  // sudot v11.4s, v16.16b, v0.4b[0]\n"
-      ".inst 0x4f01f20f  // sudot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x50]\n"
-      ".inst 0x4f20f228  // sudot v8.4s, v17.16b, v0.4b[1]\n"
-      ".inst 0x4f21f22c  // sudot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x10, #0x60]\n"
-      ".inst 0x4f20f209  // sudot v9.4s, v16.16b, v0.4b[1]\n"
-      ".inst 0x4f21f20d  // sudot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x10, #0x70]\n"
-      ".inst 0x4f20f22a  // sudot v10.4s, v17.16b, v0.4b[1]\n"
-      ".inst 0x4f21f22e  // sudot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x10, #0x80]\n"
-      ".inst 0x4f20f20b  // sudot v11.4s, v16.16b, v0.4b[1]\n"
-      ".inst 0x4f21f20f  // sudot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x10, #0x90]\n"
-      ".inst 0x4f00fa28  // sudot v8.4s, v17.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa2c  // sudot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x10, #0xa0]\n"
-      ".inst 0x4f00fa09  // sudot v9.4s, v16.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa0d  // sudot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x10, #0xb0]\n"
-      ".inst 0x4f00fa2a  // sudot v10.4s, v17.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa2e  // sudot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x10, #0xc0]\n"
-      ".inst 0x4f00fa0b  // sudot v11.4s, v16.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa0f  // sudot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x10, #0xd0]\n"
-      ".inst 0x4f20fa28  // sudot v8.4s, v17.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa2c  // sudot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr q17, [x10, #0xe0]\n"
-      ".inst 0x4f20fa09  // sudot v9.4s, v16.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa0d  // sudot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20fa2a  // sudot v10.4s, v17.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa2e  // sudot v14.4s, v17.16b, v1.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
-      ".inst 0x4f20fa0b  // sudot v11.4s, v16.16b, v0.4b[3]\n"
-      "ldr q0, [x26, #0x0]\n"
-      ".inst 0x4f21fa0f  // sudot v15.4s, v16.16b, v1.4b[3]\n"
-      "ldr q1, [x25, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "bge 51b\n"
-      "52:"  // Height 2: Multiply loop: Single iteration only
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4f00f22a  // sudot v10.4s, v17.16b, v0.4b[0]\n"
-      ".inst 0x4f01f22e  // sudot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x40]\n"
-      ".inst 0x4f00f20b  // sudot v11.4s, v16.16b, v0.4b[0]\n"
-      ".inst 0x4f01f20f  // sudot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x50]\n"
-      ".inst 0x4f20f228  // sudot v8.4s, v17.16b, v0.4b[1]\n"
-      ".inst 0x4f21f22c  // sudot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x10, #0x60]\n"
-      ".inst 0x4f20f209  // sudot v9.4s, v16.16b, v0.4b[1]\n"
-      ".inst 0x4f21f20d  // sudot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x10, #0x70]\n"
-      ".inst 0x4f20f22a  // sudot v10.4s, v17.16b, v0.4b[1]\n"
-      ".inst 0x4f21f22e  // sudot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x10, #0x80]\n"
-      ".inst 0x4f20f20b  // sudot v11.4s, v16.16b, v0.4b[1]\n"
-      ".inst 0x4f21f20f  // sudot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x10, #0x90]\n"
-      ".inst 0x4f00fa28  // sudot v8.4s, v17.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa2c  // sudot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x10, #0xa0]\n"
-      ".inst 0x4f00fa09  // sudot v9.4s, v16.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa0d  // sudot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x10, #0xb0]\n"
-      ".inst 0x4f00fa2a  // sudot v10.4s, v17.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa2e  // sudot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x10, #0xc0]\n"
-      ".inst 0x4f00fa0b  // sudot v11.4s, v16.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa0f  // sudot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x10, #0xd0]\n"
-      ".inst 0x4f20fa28  // sudot v8.4s, v17.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa2c  // sudot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr q17, [x10, #0xe0]\n"
-      ".inst 0x4f20fa09  // sudot v9.4s, v16.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa0d  // sudot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20fa2a  // sudot v10.4s, v17.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa2e  // sudot v14.4s, v17.16b, v1.4b[3]\n"
-      ".inst 0x4f20fa0b  // sudot v11.4s, v16.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa0f  // sudot v15.4s, v16.16b, v1.4b[3]\n"
-      "53:"  // Height 2: Multiply loop: Main loop skip
-      "cbz x27, 58f\n"
-      "cmp x27, #0x4\n"
-      "blt 55f\n"
-      "54:"  // Height 2: Multiply loop: Odd block loop
-      "ldr s19, [x26], #0x4\n"
-      "ldr s18, [x25], #0x4\n"
-      "sub x27, x27, #0x4\n"
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
-      ".inst 0x4f13f228  // sudot v8.4s, v17.16b, v19.4b[0]\n"
-      ".inst 0x4f12f22c  // sudot v12.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4f13f209  // sudot v9.4s, v16.16b, v19.4b[0]\n"
-      ".inst 0x4f12f20d  // sudot v13.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f13f22a  // sudot v10.4s, v17.16b, v19.4b[0]\n"
-      ".inst 0x4f12f22e  // sudot v14.4s, v17.16b, v18.4b[0]\n"
-      ".inst 0x4f13f20b  // sudot v11.4s, v16.16b, v19.4b[0]\n"
-      ".inst 0x4f12f20f  // sudot v15.4s, v16.16b, v18.4b[0]\n"
-      "bge 54b\n"
-      "55:"  // Height 2: Multiply loop: Skip odd blocks
-      "cbz x27, 58f\n"
-      "tbz x27, #1, 56f\n"
-      "ldr h0, [x26], #0x2\n"
-      "ldr h1, [x25], #0x2\n"
-      "tbz x27, #0, 57f\n"
-      "ld1 { v0.b }[2], [x26]\n"
-      "ld1 { v1.b }[2], [x25]\n"
-      "b 57f\n"
-      "56:"  // Height 2: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x26, #0x0]\n"
-      "ldr b1, [x25, #0x0]\n"
-      "57:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x10, #0x0]\n"
-      "ldr q16, [x10, #0x10]\n"
-      ".inst 0x4f00f228  // sudot v8.4s, v17.16b, v0.4b[0]\n"
-      ".inst 0x4f01f22c  // sudot v12.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4f00f209  // sudot v9.4s, v16.16b, v0.4b[0]\n"
-      ".inst 0x4f01f20d  // sudot v13.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f00f22a  // sudot v10.4s, v17.16b, v0.4b[0]\n"
-      ".inst 0x4f01f22e  // sudot v14.4s, v17.16b, v1.4b[0]\n"
-      ".inst 0x4f00f20b  // sudot v11.4s, v16.16b, v0.4b[0]\n"
-      ".inst 0x4f01f20f  // sudot v15.4s, v16.16b, v1.4b[0]\n"
-      "58:"  // Height 2: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 48b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "bge 67f\n"
-      "tbz x11, #3, 62f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v13.4s }, [x24], #0x10\n"
-      "tbz x11, #2, 60f\n"
-      "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "tbz x11, #1, 59f\n"
-      "str d11, [x9], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "tbz x11, #0, 66f\n"
-      "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "b 66f\n"
-      "59:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 66f\n"
-      "str s11, [x9, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "b 66f\n"
-      "60:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 61f\n"
-      "str d10, [x9], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "tbz x11, #0, 66f\n"
-      "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "b 66f\n"
-      "61:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 66f\n"
-      "str s10, [x9, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "b 66f\n"
-      "62:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 64f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "tbz x11, #1, 63f\n"
-      "str d9, [x9], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "tbz x11, #0, 66f\n"
-      "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "b 66f\n"
-      "63:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 66f\n"
-      "str s9, [x9, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "b 66f\n"
-      "64:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 65f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "tbz x11, #0, 66f\n"
-      "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "b 66f\n"
-      "65:"  // Height 2: Partial direct writeback: partial_1_0
-      "str s8, [x9, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "66:"  // Height 2: Partial direct writeback: Done
-      "b 68f\n"
-      "67:"  // Height 2: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "str q9, [x9, #0x10]\n"
-      "str q10, [x9, #0x20]\n"
-      "str q11, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q12, [x24, #0x0]\n"
-      "str q13, [x24, #0x10]\n"
-      "str q14, [x24, #0x20]\n"
-      "str q15, [x24, #0x30]\n"
-      "68:"  // Height 2: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 36b\n"
-      "b 206f\n"
-      "69:"  // Height 3
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "70:"  // Height 3: Column loop
-      "tbz %x[flags], #0, 80f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "bge 79f\n"
-      "tbz x11, #3, 74f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "tbz x11, #2, 72f\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v18.4s }, [x23], #0x10\n"
-      "tbz x11, #1, 71f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "ldr d19, [x23], #0x8\n"
-      "tbz x11, #0, 78f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v19.s }[2], [x23]\n"
-      "b 78f\n"
-      "71:"  // Height 3: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 78f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s19, [x23, #0x0]\n"
-      "b 78f\n"
-      "72:"  // Height 3: Partial accumulate: partial_2_8
-      "tbz x11, #1, 73f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "ldr d18, [x23], #0x8\n"
-      "tbz x11, #0, 78f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v18.s }[2], [x23]\n"
-      "b 78f\n"
-      "73:"  // Height 3: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 78f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s18, [x23, #0x0]\n"
-      "b 78f\n"
-      "74:"  // Height 3: Partial accumulate: partial_4_0
-      "tbz x11, #2, 76f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v16.4s }, [x23], #0x10\n"
-      "tbz x11, #1, 75f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "ldr d17, [x23], #0x8\n"
-      "tbz x11, #0, 78f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "ld1 { v17.s }[2], [x23]\n"
-      "b 78f\n"
-      "75:"  // Height 3: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 78f\n"
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "ldr s17, [x23, #0x0]\n"
-      "b 78f\n"
-      "76:"  // Height 3: Partial accumulate: partial_2_0
-      "tbz x11, #1, 77f\n"
-      "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "ldr d16, [x23], #0x8\n"
-      "tbz x11, #0, 78f\n"
-      "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "ld1 { v16.s }[2], [x23]\n"
-      "b 78f\n"
-      "77:"  // Height 3: Partial accumulate: partial_1_0
-      "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "ldr s16, [x23, #0x0]\n"
-      "78:"  // Height 3: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 81f\n"
-      "79:"  // Height 3: full accumulate
-      "ldr q8, [x9, #0x0]\n"
-      "ldr q9, [x9, #0x10]\n"
-      "ldr q10, [x9, #0x20]\n"
-      "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "ldr q16, [x23, #0x0]\n"
-      "ldr q17, [x23, #0x10]\n"
-      "ldr q18, [x23, #0x20]\n"
-      "ldr q19, [x23, #0x30]\n"
-      "b 81f\n"
-      "80:"  // Height 3: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "81:"  // Height 3: setup done
-      "mov x28, #0x0\n"
-      "82:"  // Height 3: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 83f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "cbnz x28, 84f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "b 84f\n"
-      "83:"  // Height 3: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "84:"  // Height 3: input setup done
-      "cmp x27, #0x10\n"
-      "blt 87f\n"
-      "ldr q0, [x26, #0x0]\n"
-      "ldr q1, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q2, [x24, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "blt 86f\n"
-      "85:"  // Height 3: Multiply loop: Main loop head
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0x4f02f0d0  // sudot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x20]\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f1  // sudot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x30]\n"
-      "add x24, x24, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4f00f2aa  // sudot v10.4s, v21.16b, v0.4b[0]\n"
-      ".inst 0x4f01f2ae  // sudot v14.4s, v21.16b, v1.4b[0]\n"
-      ".inst 0x4f02f2b2  // sudot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      ".inst 0x4f00f28b  // sudot v11.4s, v20.16b, v0.4b[0]\n"
-      ".inst 0x4f01f28f  // sudot v15.4s, v20.16b, v1.4b[0]\n"
-      ".inst 0x4f02f293  // sudot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x50]\n"
-      ".inst 0x4f20f2a8  // sudot v8.4s, v21.16b, v0.4b[1]\n"
-      ".inst 0x4f21f2ac  // sudot v12.4s, v21.16b, v1.4b[1]\n"
-      ".inst 0x4f22f2b0  // sudot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x10, #0x60]\n"
-      ".inst 0x4f20f289  // sudot v9.4s, v20.16b, v0.4b[1]\n"
-      ".inst 0x4f21f28d  // sudot v13.4s, v20.16b, v1.4b[1]\n"
-      ".inst 0x4f22f291  // sudot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x10, #0x70]\n"
-      ".inst 0x4f20f2aa  // sudot v10.4s, v21.16b, v0.4b[1]\n"
-      ".inst 0x4f21f2ae  // sudot v14.4s, v21.16b, v1.4b[1]\n"
-      ".inst 0x4f22f2b2  // sudot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x10, #0x80]\n"
-      ".inst 0x4f20f28b  // sudot v11.4s, v20.16b, v0.4b[1]\n"
-      ".inst 0x4f21f28f  // sudot v15.4s, v20.16b, v1.4b[1]\n"
-      ".inst 0x4f22f293  // sudot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x10, #0x90]\n"
-      ".inst 0x4f00faa8  // sudot v8.4s, v21.16b, v0.4b[2]\n"
-      ".inst 0x4f01faac  // sudot v12.4s, v21.16b, v1.4b[2]\n"
-      ".inst 0x4f02fab0  // sudot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x10, #0xa0]\n"
-      ".inst 0x4f00fa89  // sudot v9.4s, v20.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa8d  // sudot v13.4s, v20.16b, v1.4b[2]\n"
-      ".inst 0x4f02fa91  // sudot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x10, #0xb0]\n"
-      ".inst 0x4f00faaa  // sudot v10.4s, v21.16b, v0.4b[2]\n"
-      ".inst 0x4f01faae  // sudot v14.4s, v21.16b, v1.4b[2]\n"
-      ".inst 0x4f02fab2  // sudot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x10, #0xc0]\n"
-      ".inst 0x4f00fa8b  // sudot v11.4s, v20.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa8f  // sudot v15.4s, v20.16b, v1.4b[2]\n"
-      ".inst 0x4f02fa93  // sudot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x10, #0xd0]\n"
-      ".inst 0x4f20faa8  // sudot v8.4s, v21.16b, v0.4b[3]\n"
-      ".inst 0x4f21faac  // sudot v12.4s, v21.16b, v1.4b[3]\n"
-      ".inst 0x4f22fab0  // sudot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr q21, [x10, #0xe0]\n"
-      ".inst 0x4f20fa89  // sudot v9.4s, v20.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa8d  // sudot v13.4s, v20.16b, v1.4b[3]\n"
-      ".inst 0x4f22fa91  // sudot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr q20, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20faaa  // sudot v10.4s, v21.16b, v0.4b[3]\n"
-      ".inst 0x4f21faae  // sudot v14.4s, v21.16b, v1.4b[3]\n"
-      ".inst 0x4f22fab2  // sudot v18.4s, v21.16b, v2.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
-      ".inst 0x4f20fa8b  // sudot v11.4s, v20.16b, v0.4b[3]\n"
-      "ldr q0, [x26, #0x0]\n"
-      ".inst 0x4f21fa8f  // sudot v15.4s, v20.16b, v1.4b[3]\n"
-      "ldr q1, [x25, #0x0]\n"
-      ".inst 0x4f22fa93  // sudot v19.4s, v20.16b, v2.4b[3]\n"
-      "ldr q2, [x24, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "bge 85b\n"
-      "86:"  // Height 3: Multiply loop: Single iteration only
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0x4f02f0d0  // sudot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x20]\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f1  // sudot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4f00f2aa  // sudot v10.4s, v21.16b, v0.4b[0]\n"
-      ".inst 0x4f01f2ae  // sudot v14.4s, v21.16b, v1.4b[0]\n"
-      ".inst 0x4f02f2b2  // sudot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      ".inst 0x4f00f28b  // sudot v11.4s, v20.16b, v0.4b[0]\n"
-      ".inst 0x4f01f28f  // sudot v15.4s, v20.16b, v1.4b[0]\n"
-      ".inst 0x4f02f293  // sudot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x50]\n"
-      ".inst 0x4f20f2a8  // sudot v8.4s, v21.16b, v0.4b[1]\n"
-      ".inst 0x4f21f2ac  // sudot v12.4s, v21.16b, v1.4b[1]\n"
-      ".inst 0x4f22f2b0  // sudot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x10, #0x60]\n"
-      ".inst 0x4f20f289  // sudot v9.4s, v20.16b, v0.4b[1]\n"
-      ".inst 0x4f21f28d  // sudot v13.4s, v20.16b, v1.4b[1]\n"
-      ".inst 0x4f22f291  // sudot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x10, #0x70]\n"
-      ".inst 0x4f20f2aa  // sudot v10.4s, v21.16b, v0.4b[1]\n"
-      ".inst 0x4f21f2ae  // sudot v14.4s, v21.16b, v1.4b[1]\n"
-      ".inst 0x4f22f2b2  // sudot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x10, #0x80]\n"
-      ".inst 0x4f20f28b  // sudot v11.4s, v20.16b, v0.4b[1]\n"
-      ".inst 0x4f21f28f  // sudot v15.4s, v20.16b, v1.4b[1]\n"
-      ".inst 0x4f22f293  // sudot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x10, #0x90]\n"
-      ".inst 0x4f00faa8  // sudot v8.4s, v21.16b, v0.4b[2]\n"
-      ".inst 0x4f01faac  // sudot v12.4s, v21.16b, v1.4b[2]\n"
-      ".inst 0x4f02fab0  // sudot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x10, #0xa0]\n"
-      ".inst 0x4f00fa89  // sudot v9.4s, v20.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa8d  // sudot v13.4s, v20.16b, v1.4b[2]\n"
-      ".inst 0x4f02fa91  // sudot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x10, #0xb0]\n"
-      ".inst 0x4f00faaa  // sudot v10.4s, v21.16b, v0.4b[2]\n"
-      ".inst 0x4f01faae  // sudot v14.4s, v21.16b, v1.4b[2]\n"
-      ".inst 0x4f02fab2  // sudot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x10, #0xc0]\n"
-      ".inst 0x4f00fa8b  // sudot v11.4s, v20.16b, v0.4b[2]\n"
-      ".inst 0x4f01fa8f  // sudot v15.4s, v20.16b, v1.4b[2]\n"
-      ".inst 0x4f02fa93  // sudot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x10, #0xd0]\n"
-      ".inst 0x4f20faa8  // sudot v8.4s, v21.16b, v0.4b[3]\n"
-      ".inst 0x4f21faac  // sudot v12.4s, v21.16b, v1.4b[3]\n"
-      ".inst 0x4f22fab0  // sudot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr q21, [x10, #0xe0]\n"
-      ".inst 0x4f20fa89  // sudot v9.4s, v20.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa8d  // sudot v13.4s, v20.16b, v1.4b[3]\n"
-      ".inst 0x4f22fa91  // sudot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr q20, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20faaa  // sudot v10.4s, v21.16b, v0.4b[3]\n"
-      ".inst 0x4f21faae  // sudot v14.4s, v21.16b, v1.4b[3]\n"
-      ".inst 0x4f22fab2  // sudot v18.4s, v21.16b, v2.4b[3]\n"
-      ".inst 0x4f20fa8b  // sudot v11.4s, v20.16b, v0.4b[3]\n"
-      ".inst 0x4f21fa8f  // sudot v15.4s, v20.16b, v1.4b[3]\n"
-      ".inst 0x4f22fa93  // sudot v19.4s, v20.16b, v2.4b[3]\n"
-      "87:"  // Height 3: Multiply loop: Main loop skip
-      "cbz x27, 92f\n"
-      "cmp x27, #0x4\n"
-      "blt 89f\n"
-      "88:"  // Height 3: Multiply loop: Odd block loop
-      "ldr s24, [x26], #0x4\n"
-      "ldr s23, [x25], #0x4\n"
-      "sub x27, x27, #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr q21, [x10, #0x0]\n"
-      "cmp x27, #0x4\n"
-      "ldr q20, [x10, #0x10]\n"
-      ".inst 0x4f18f2a8  // sudot v8.4s, v21.16b, v24.4b[0]\n"
-      ".inst 0x4f17f2ac  // sudot v12.4s, v21.16b, v23.4b[0]\n"
-      ".inst 0x4f16f2b0  // sudot v16.4s, v21.16b, v22.4b[0]\n"
-      "ldr q21, [x10, #0x20]\n"
-      ".inst 0x4f18f289  // sudot v9.4s, v20.16b, v24.4b[0]\n"
-      ".inst 0x4f17f28d  // sudot v13.4s, v20.16b, v23.4b[0]\n"
-      ".inst 0x4f16f291  // sudot v17.4s, v20.16b, v22.4b[0]\n"
-      "ldr q20, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f18f2aa  // sudot v10.4s, v21.16b, v24.4b[0]\n"
-      ".inst 0x4f17f2ae  // sudot v14.4s, v21.16b, v23.4b[0]\n"
-      ".inst 0x4f16f2b2  // sudot v18.4s, v21.16b, v22.4b[0]\n"
-      ".inst 0x4f18f28b  // sudot v11.4s, v20.16b, v24.4b[0]\n"
-      ".inst 0x4f17f28f  // sudot v15.4s, v20.16b, v23.4b[0]\n"
-      ".inst 0x4f16f293  // sudot v19.4s, v20.16b, v22.4b[0]\n"
-      "bge 88b\n"
-      "89:"  // Height 3: Multiply loop: Skip odd blocks
-      "cbz x27, 92f\n"
-      "tbz x27, #1, 90f\n"
-      "ldr h0, [x26], #0x2\n"
-      "ldr h1, [x25], #0x2\n"
-      "ldr h2, [x24], #0x2\n"
-      "tbz x27, #0, 91f\n"
-      "ld1 { v0.b }[2], [x26]\n"
-      "ld1 { v1.b }[2], [x25]\n"
-      "ld1 { v2.b }[2], [x24]\n"
-      "b 91f\n"
-      "90:"  // Height 3: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x26, #0x0]\n"
-      "ldr b1, [x25, #0x0]\n"
-      "ldr b2, [x24, #0x0]\n"
-      "91:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "ldr q21, [x10, #0x0]\n"
-      "ldr q20, [x10, #0x10]\n"
-      ".inst 0x4f00f2a8  // sudot v8.4s, v21.16b, v0.4b[0]\n"
-      ".inst 0x4f01f2ac  // sudot v12.4s, v21.16b, v1.4b[0]\n"
-      ".inst 0x4f02f2b0  // sudot v16.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x10, #0x20]\n"
-      ".inst 0x4f00f289  // sudot v9.4s, v20.16b, v0.4b[0]\n"
-      ".inst 0x4f01f28d  // sudot v13.4s, v20.16b, v1.4b[0]\n"
-      ".inst 0x4f02f291  // sudot v17.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f00f2aa  // sudot v10.4s, v21.16b, v0.4b[0]\n"
-      ".inst 0x4f01f2ae  // sudot v14.4s, v21.16b, v1.4b[0]\n"
-      ".inst 0x4f02f2b2  // sudot v18.4s, v21.16b, v2.4b[0]\n"
-      ".inst 0x4f00f28b  // sudot v11.4s, v20.16b, v0.4b[0]\n"
-      ".inst 0x4f01f28f  // sudot v15.4s, v20.16b, v1.4b[0]\n"
-      ".inst 0x4f02f293  // sudot v19.4s, v20.16b, v2.4b[0]\n"
-      "92:"  // Height 3: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 82b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "bge 101f\n"
-      "tbz x11, #3, 96f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v13.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "st1 { v17.4s }, [x23], #0x10\n"
-      "tbz x11, #2, 94f\n"
-      "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v18.4s }, [x23], #0x10\n"
-      "tbz x11, #1, 93f\n"
-      "str d11, [x9], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "str d19, [x23], #0x8\n"
-      "tbz x11, #0, 100f\n"
-      "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "st1 { v19.s }[2], [x23]\n"
-      "b 100f\n"
-      "93:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 100f\n"
-      "str s11, [x9, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "str s19, [x23, #0x0]\n"
-      "b 100f\n"
-      "94:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 95f\n"
-      "str d10, [x9], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "str d18, [x23], #0x8\n"
-      "tbz x11, #0, 100f\n"
-      "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "st1 { v18.s }[2], [x23]\n"
-      "b 100f\n"
-      "95:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 100f\n"
-      "str s10, [x9, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "str s18, [x23, #0x0]\n"
-      "b 100f\n"
-      "96:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 98f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "tbz x11, #1, 97f\n"
-      "str d9, [x9], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "str d17, [x23], #0x8\n"
-      "tbz x11, #0, 100f\n"
-      "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "st1 { v17.s }[2], [x23]\n"
-      "b 100f\n"
-      "97:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 100f\n"
-      "str s9, [x9, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "str s17, [x23, #0x0]\n"
-      "b 100f\n"
-      "98:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 99f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "str d16, [x23], #0x8\n"
-      "tbz x11, #0, 100f\n"
-      "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "st1 { v16.s }[2], [x23]\n"
-      "b 100f\n"
-      "99:"  // Height 3: Partial direct writeback: partial_1_0
-      "str s8, [x9, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "str s16, [x23, #0x0]\n"
-      "100:"  // Height 3: Partial direct writeback: Done
-      "b 102f\n"
-      "101:"  // Height 3: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "str q9, [x9, #0x10]\n"
-      "str q10, [x9, #0x20]\n"
-      "str q11, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q12, [x24, #0x0]\n"
-      "str q13, [x24, #0x10]\n"
-      "str q14, [x24, #0x20]\n"
-      "str q15, [x24, #0x30]\n"
-      "str q16, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q18, [x23, #0x20]\n"
-      "str q19, [x23, #0x30]\n"
-      "102:"  // Height 3: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 70b\n"
-      "b 206f\n"
-      "103:"  // Height 4
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "104:"  // Height 4: Column loop
-      "tbz %x[flags], #0, 114f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "bge 113f\n"
-      "tbz x11, #3, 108f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v21.4s }, [x22], #0x10\n"
-      "tbz x11, #2, 106f\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v18.4s }, [x23], #0x10\n"
-      "ld1 { v22.4s }, [x22], #0x10\n"
-      "tbz x11, #1, 105f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "ldr d19, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "tbz x11, #0, 112f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v19.s }[2], [x23]\n"
-      "ld1 { v23.s }[2], [x22]\n"
-      "b 112f\n"
-      "105:"  // Height 4: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 112f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s19, [x23, #0x0]\n"
-      "ldr s23, [x22, #0x0]\n"
-      "b 112f\n"
-      "106:"  // Height 4: Partial accumulate: partial_2_8
-      "tbz x11, #1, 107f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d22, [x22], #0x8\n"
-      "tbz x11, #0, 112f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v18.s }[2], [x23]\n"
-      "ld1 { v22.s }[2], [x22]\n"
-      "b 112f\n"
-      "107:"  // Height 4: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 112f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s18, [x23, #0x0]\n"
-      "ldr s22, [x22, #0x0]\n"
-      "b 112f\n"
-      "108:"  // Height 4: Partial accumulate: partial_4_0
-      "tbz x11, #2, 110f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "tbz x11, #1, 109f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "ldr d17, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "tbz x11, #0, 112f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "ld1 { v17.s }[2], [x23]\n"
-      "ld1 { v21.s }[2], [x22]\n"
-      "b 112f\n"
-      "109:"  // Height 4: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 112f\n"
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "ldr s17, [x23, #0x0]\n"
-      "ldr s21, [x22, #0x0]\n"
-      "b 112f\n"
-      "110:"  // Height 4: Partial accumulate: partial_2_0
-      "tbz x11, #1, 111f\n"
-      "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "ldr d16, [x23], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "tbz x11, #0, 112f\n"
-      "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "ld1 { v16.s }[2], [x23]\n"
-      "ld1 { v20.s }[2], [x22]\n"
-      "b 112f\n"
-      "111:"  // Height 4: Partial accumulate: partial_1_0
-      "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "ldr s16, [x23, #0x0]\n"
-      "ldr s20, [x22, #0x0]\n"
-      "112:"  // Height 4: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 115f\n"
-      "113:"  // Height 4: full accumulate
-      "ldr q8, [x9, #0x0]\n"
-      "ldr q9, [x9, #0x10]\n"
-      "ldr q10, [x9, #0x20]\n"
-      "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "ldr q16, [x23, #0x0]\n"
-      "ldr q17, [x23, #0x10]\n"
-      "ldr q18, [x23, #0x20]\n"
-      "ldr q19, [x23, #0x30]\n"
-      "ldr q20, [x22, #0x0]\n"
-      "ldr q21, [x22, #0x10]\n"
-      "ldr q22, [x22, #0x20]\n"
-      "ldr q23, [x22, #0x30]\n"
-      "b 115f\n"
-      "114:"  // Height 4: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "115:"  // Height 4: setup done
-      "mov x28, #0x0\n"
-      "116:"  // Height 4: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 117f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "cbnz x28, 118f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "b 118f\n"
-      "117:"  // Height 4: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "118:"  // Height 4: input setup done
-      "cmp x27, #0x10\n"
-      "blt 121f\n"
-      "ldr q0, [x26, #0x0]\n"
-      "ldr q1, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q2, [x24, #0x0]\n"
-      "ldr q3, [x23, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "blt 120f\n"
-      "119:"  // Height 4: Multiply loop: Main loop head
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0x4f02f0d0  // sudot v16.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d4  // sudot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x20]\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f02f0f1  // sudot v17.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0f5  // sudot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x30]\n"
-      "cmp x27, #0x20\n"
-      ".inst 0x4f00f32a  // sudot v10.4s, v25.16b, v0.4b[0]\n"
-      ".inst 0x4f01f32e  // sudot v14.4s, v25.16b, v1.4b[0]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4f02f332  // sudot v18.4s, v25.16b, v2.4b[0]\n"
-      ".inst 0x4f03f336  // sudot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      ".inst 0x4f00f30b  // sudot v11.4s, v24.16b, v0.4b[0]\n"
-      ".inst 0x4f01f30f  // sudot v15.4s, v24.16b, v1.4b[0]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4f02f313  // sudot v19.4s, v24.16b, v2.4b[0]\n"
-      ".inst 0x4f03f317  // sudot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x50]\n"
-      ".inst 0x4f20f328  // sudot v8.4s, v25.16b, v0.4b[1]\n"
-      ".inst 0x4f21f32c  // sudot v12.4s, v25.16b, v1.4b[1]\n"
-      ".inst 0x4f22f330  // sudot v16.4s, v25.16b, v2.4b[1]\n"
-      ".inst 0x4f23f334  // sudot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x10, #0x60]\n"
-      ".inst 0x4f20f309  // sudot v9.4s, v24.16b, v0.4b[1]\n"
-      ".inst 0x4f21f30d  // sudot v13.4s, v24.16b, v1.4b[1]\n"
-      ".inst 0x4f22f311  // sudot v17.4s, v24.16b, v2.4b[1]\n"
-      ".inst 0x4f23f315  // sudot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x10, #0x70]\n"
-      ".inst 0x4f20f32a  // sudot v10.4s, v25.16b, v0.4b[1]\n"
-      ".inst 0x4f21f32e  // sudot v14.4s, v25.16b, v1.4b[1]\n"
-      ".inst 0x4f22f332  // sudot v18.4s, v25.16b, v2.4b[1]\n"
-      ".inst 0x4f23f336  // sudot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x10, #0x80]\n"
-      ".inst 0x4f20f30b  // sudot v11.4s, v24.16b, v0.4b[1]\n"
-      ".inst 0x4f21f30f  // sudot v15.4s, v24.16b, v1.4b[1]\n"
-      ".inst 0x4f22f313  // sudot v19.4s, v24.16b, v2.4b[1]\n"
-      ".inst 0x4f23f317  // sudot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x10, #0x90]\n"
-      ".inst 0x4f00fb28  // sudot v8.4s, v25.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb2c  // sudot v12.4s, v25.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb30  // sudot v16.4s, v25.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb34  // sudot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x10, #0xa0]\n"
-      ".inst 0x4f00fb09  // sudot v9.4s, v24.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb0d  // sudot v13.4s, v24.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb11  // sudot v17.4s, v24.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb15  // sudot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x10, #0xb0]\n"
-      ".inst 0x4f00fb2a  // sudot v10.4s, v25.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb2e  // sudot v14.4s, v25.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb32  // sudot v18.4s, v25.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb36  // sudot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x10, #0xc0]\n"
-      ".inst 0x4f00fb0b  // sudot v11.4s, v24.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb0f  // sudot v15.4s, v24.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb13  // sudot v19.4s, v24.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb17  // sudot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x10, #0xd0]\n"
-      ".inst 0x4f20fb28  // sudot v8.4s, v25.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb2c  // sudot v12.4s, v25.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb30  // sudot v16.4s, v25.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb34  // sudot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr q25, [x10, #0xe0]\n"
-      ".inst 0x4f20fb09  // sudot v9.4s, v24.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb0d  // sudot v13.4s, v24.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb11  // sudot v17.4s, v24.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb15  // sudot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr q24, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20fb2a  // sudot v10.4s, v25.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb2e  // sudot v14.4s, v25.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb32  // sudot v18.4s, v25.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb36  // sudot v22.4s, v25.16b, v3.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
-      ".inst 0x4f20fb0b  // sudot v11.4s, v24.16b, v0.4b[3]\n"
-      "ldr q0, [x26, #0x0]\n"
-      ".inst 0x4f21fb0f  // sudot v15.4s, v24.16b, v1.4b[3]\n"
-      "ldr q1, [x25, #0x0]\n"
-      ".inst 0x4f22fb13  // sudot v19.4s, v24.16b, v2.4b[3]\n"
-      "ldr q2, [x24, #0x0]\n"
-      ".inst 0x4f23fb17  // sudot v23.4s, v24.16b, v3.4b[3]\n"
-      "ldr q3, [x23, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "bge 119b\n"
-      "120:"  // Height 4: Multiply loop: Single iteration only
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0x4f02f0d0  // sudot v16.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d4  // sudot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x20]\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4f02f0f1  // sudot v17.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0f5  // sudot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x30]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4f00f32a  // sudot v10.4s, v25.16b, v0.4b[0]\n"
-      ".inst 0x4f01f32e  // sudot v14.4s, v25.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      ".inst 0x4f02f332  // sudot v18.4s, v25.16b, v2.4b[0]\n"
-      ".inst 0x4f03f336  // sudot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x40]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4f00f30b  // sudot v11.4s, v24.16b, v0.4b[0]\n"
-      ".inst 0x4f01f30f  // sudot v15.4s, v24.16b, v1.4b[0]\n"
-      ".inst 0x4f02f313  // sudot v19.4s, v24.16b, v2.4b[0]\n"
-      ".inst 0x4f03f317  // sudot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x50]\n"
-      ".inst 0x4f20f328  // sudot v8.4s, v25.16b, v0.4b[1]\n"
-      ".inst 0x4f21f32c  // sudot v12.4s, v25.16b, v1.4b[1]\n"
-      ".inst 0x4f22f330  // sudot v16.4s, v25.16b, v2.4b[1]\n"
-      ".inst 0x4f23f334  // sudot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x10, #0x60]\n"
-      ".inst 0x4f20f309  // sudot v9.4s, v24.16b, v0.4b[1]\n"
-      ".inst 0x4f21f30d  // sudot v13.4s, v24.16b, v1.4b[1]\n"
-      ".inst 0x4f22f311  // sudot v17.4s, v24.16b, v2.4b[1]\n"
-      ".inst 0x4f23f315  // sudot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x10, #0x70]\n"
-      ".inst 0x4f20f32a  // sudot v10.4s, v25.16b, v0.4b[1]\n"
-      ".inst 0x4f21f32e  // sudot v14.4s, v25.16b, v1.4b[1]\n"
-      ".inst 0x4f22f332  // sudot v18.4s, v25.16b, v2.4b[1]\n"
-      ".inst 0x4f23f336  // sudot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x10, #0x80]\n"
-      ".inst 0x4f20f30b  // sudot v11.4s, v24.16b, v0.4b[1]\n"
-      ".inst 0x4f21f30f  // sudot v15.4s, v24.16b, v1.4b[1]\n"
-      ".inst 0x4f22f313  // sudot v19.4s, v24.16b, v2.4b[1]\n"
-      ".inst 0x4f23f317  // sudot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x10, #0x90]\n"
-      ".inst 0x4f00fb28  // sudot v8.4s, v25.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb2c  // sudot v12.4s, v25.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb30  // sudot v16.4s, v25.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb34  // sudot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x10, #0xa0]\n"
-      ".inst 0x4f00fb09  // sudot v9.4s, v24.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb0d  // sudot v13.4s, v24.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb11  // sudot v17.4s, v24.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb15  // sudot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x10, #0xb0]\n"
-      ".inst 0x4f00fb2a  // sudot v10.4s, v25.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb2e  // sudot v14.4s, v25.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb32  // sudot v18.4s, v25.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb36  // sudot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x10, #0xc0]\n"
-      ".inst 0x4f00fb0b  // sudot v11.4s, v24.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb0f  // sudot v15.4s, v24.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb13  // sudot v19.4s, v24.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb17  // sudot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x10, #0xd0]\n"
-      ".inst 0x4f20fb28  // sudot v8.4s, v25.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb2c  // sudot v12.4s, v25.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb30  // sudot v16.4s, v25.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb34  // sudot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr q25, [x10, #0xe0]\n"
-      ".inst 0x4f20fb09  // sudot v9.4s, v24.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb0d  // sudot v13.4s, v24.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb11  // sudot v17.4s, v24.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb15  // sudot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr q24, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20fb2a  // sudot v10.4s, v25.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb2e  // sudot v14.4s, v25.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb32  // sudot v18.4s, v25.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb36  // sudot v22.4s, v25.16b, v3.4b[3]\n"
-      ".inst 0x4f20fb0b  // sudot v11.4s, v24.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb0f  // sudot v15.4s, v24.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb13  // sudot v19.4s, v24.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb17  // sudot v23.4s, v24.16b, v3.4b[3]\n"
-      "121:"  // Height 4: Multiply loop: Main loop skip
-      "cbz x27, 126f\n"
-      "cmp x27, #0x4\n"
-      "blt 123f\n"
-      "122:"  // Height 4: Multiply loop: Odd block loop
-      "ldr s29, [x26], #0x4\n"
-      "ldr s28, [x25], #0x4\n"
-      "sub x27, x27, #0x4\n"
-      "ldr s27, [x24], #0x4\n"
-      "ldr s26, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
-      "ldr q25, [x10, #0x0]\n"
-      "ldr q24, [x10, #0x10]\n"
-      ".inst 0x4f1df328  // sudot v8.4s, v25.16b, v29.4b[0]\n"
-      ".inst 0x4f1cf32c  // sudot v12.4s, v25.16b, v28.4b[0]\n"
-      ".inst 0x4f1bf330  // sudot v16.4s, v25.16b, v27.4b[0]\n"
-      ".inst 0x4f1af334  // sudot v20.4s, v25.16b, v26.4b[0]\n"
-      "ldr q25, [x10, #0x20]\n"
-      ".inst 0x4f1df309  // sudot v9.4s, v24.16b, v29.4b[0]\n"
-      ".inst 0x4f1cf30d  // sudot v13.4s, v24.16b, v28.4b[0]\n"
-      ".inst 0x4f1bf311  // sudot v17.4s, v24.16b, v27.4b[0]\n"
-      ".inst 0x4f1af315  // sudot v21.4s, v24.16b, v26.4b[0]\n"
-      "ldr q24, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f1df32a  // sudot v10.4s, v25.16b, v29.4b[0]\n"
-      ".inst 0x4f1cf32e  // sudot v14.4s, v25.16b, v28.4b[0]\n"
-      ".inst 0x4f1bf332  // sudot v18.4s, v25.16b, v27.4b[0]\n"
-      ".inst 0x4f1af336  // sudot v22.4s, v25.16b, v26.4b[0]\n"
-      ".inst 0x4f1df30b  // sudot v11.4s, v24.16b, v29.4b[0]\n"
-      ".inst 0x4f1cf30f  // sudot v15.4s, v24.16b, v28.4b[0]\n"
-      ".inst 0x4f1bf313  // sudot v19.4s, v24.16b, v27.4b[0]\n"
-      ".inst 0x4f1af317  // sudot v23.4s, v24.16b, v26.4b[0]\n"
-      "bge 122b\n"
-      "123:"  // Height 4: Multiply loop: Skip odd blocks
-      "cbz x27, 126f\n"
-      "tbz x27, #1, 124f\n"
-      "ldr h0, [x26], #0x2\n"
-      "ldr h1, [x25], #0x2\n"
-      "ldr h2, [x24], #0x2\n"
-      "ldr h3, [x23], #0x2\n"
-      "tbz x27, #0, 125f\n"
-      "ld1 { v0.b }[2], [x26]\n"
-      "ld1 { v1.b }[2], [x25]\n"
-      "ld1 { v2.b }[2], [x24]\n"
-      "ld1 { v3.b }[2], [x23]\n"
-      "b 125f\n"
-      "124:"  // Height 4: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x26, #0x0]\n"
-      "ldr b1, [x25, #0x0]\n"
-      "ldr b2, [x24, #0x0]\n"
-      "ldr b3, [x23, #0x0]\n"
-      "125:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "ldr q25, [x10, #0x0]\n"
-      "ldr q24, [x10, #0x10]\n"
-      ".inst 0x4f00f328  // sudot v8.4s, v25.16b, v0.4b[0]\n"
-      ".inst 0x4f01f32c  // sudot v12.4s, v25.16b, v1.4b[0]\n"
-      ".inst 0x4f02f330  // sudot v16.4s, v25.16b, v2.4b[0]\n"
-      ".inst 0x4f03f334  // sudot v20.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x10, #0x20]\n"
-      ".inst 0x4f00f309  // sudot v9.4s, v24.16b, v0.4b[0]\n"
-      ".inst 0x4f01f30d  // sudot v13.4s, v24.16b, v1.4b[0]\n"
-      ".inst 0x4f02f311  // sudot v17.4s, v24.16b, v2.4b[0]\n"
-      ".inst 0x4f03f315  // sudot v21.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f00f32a  // sudot v10.4s, v25.16b, v0.4b[0]\n"
-      ".inst 0x4f01f32e  // sudot v14.4s, v25.16b, v1.4b[0]\n"
-      ".inst 0x4f02f332  // sudot v18.4s, v25.16b, v2.4b[0]\n"
-      ".inst 0x4f03f336  // sudot v22.4s, v25.16b, v3.4b[0]\n"
-      ".inst 0x4f00f30b  // sudot v11.4s, v24.16b, v0.4b[0]\n"
-      ".inst 0x4f01f30f  // sudot v15.4s, v24.16b, v1.4b[0]\n"
-      ".inst 0x4f02f313  // sudot v19.4s, v24.16b, v2.4b[0]\n"
-      ".inst 0x4f03f317  // sudot v23.4s, v24.16b, v3.4b[0]\n"
-      "126:"  // Height 4: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 116b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "bge 135f\n"
-      "tbz x11, #3, 130f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v13.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "st1 { v17.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v21.4s }, [x22], #0x10\n"
-      "tbz x11, #2, 128f\n"
-      "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v18.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x22], #0x10\n"
-      "tbz x11, #1, 127f\n"
-      "str d11, [x9], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "str d19, [x23], #0x8\n"
-      "str d23, [x22], #0x8\n"
-      "tbz x11, #0, 134f\n"
-      "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "st1 { v19.s }[2], [x23]\n"
-      "st1 { v23.s }[2], [x22]\n"
-      "b 134f\n"
-      "127:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 134f\n"
-      "str s11, [x9, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "str s19, [x23, #0x0]\n"
-      "str s23, [x22, #0x0]\n"
-      "b 134f\n"
-      "128:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 129f\n"
-      "str d10, [x9], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "str d18, [x23], #0x8\n"
-      "str d22, [x22], #0x8\n"
-      "tbz x11, #0, 134f\n"
-      "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "st1 { v18.s }[2], [x23]\n"
-      "st1 { v22.s }[2], [x22]\n"
-      "b 134f\n"
-      "129:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 134f\n"
-      "str s10, [x9, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "str s18, [x23, #0x0]\n"
-      "str s22, [x22, #0x0]\n"
-      "b 134f\n"
-      "130:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 132f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "tbz x11, #1, 131f\n"
-      "str d9, [x9], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "str d17, [x23], #0x8\n"
-      "str d21, [x22], #0x8\n"
-      "tbz x11, #0, 134f\n"
-      "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "st1 { v17.s }[2], [x23]\n"
-      "st1 { v21.s }[2], [x22]\n"
-      "b 134f\n"
-      "131:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 134f\n"
-      "str s9, [x9, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "str s17, [x23, #0x0]\n"
-      "str s21, [x22, #0x0]\n"
-      "b 134f\n"
-      "132:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 133f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "str d16, [x23], #0x8\n"
-      "str d20, [x22], #0x8\n"
-      "tbz x11, #0, 134f\n"
-      "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "st1 { v16.s }[2], [x23]\n"
-      "st1 { v20.s }[2], [x22]\n"
-      "b 134f\n"
-      "133:"  // Height 4: Partial direct writeback: partial_1_0
-      "str s8, [x9, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "str s16, [x23, #0x0]\n"
-      "str s20, [x22, #0x0]\n"
-      "134:"  // Height 4: Partial direct writeback: Done
-      "b 136f\n"
-      "135:"  // Height 4: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "str q9, [x9, #0x10]\n"
-      "str q10, [x9, #0x20]\n"
-      "str q11, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q12, [x24, #0x0]\n"
-      "str q13, [x24, #0x10]\n"
-      "str q14, [x24, #0x20]\n"
-      "str q15, [x24, #0x30]\n"
-      "str q16, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q18, [x23, #0x20]\n"
-      "str q19, [x23, #0x30]\n"
-      "str q20, [x22, #0x0]\n"
-      "str q21, [x22, #0x10]\n"
-      "str q22, [x22, #0x20]\n"
-      "str q23, [x22, #0x30]\n"
-      "136:"  // Height 4: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 104b\n"
-      "b 206f\n"
-      "137:"  // Height 5
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "138:"  // Height 5: Column loop
-      "tbz %x[flags], #0, 148f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "bge 147f\n"
-      "tbz x11, #3, 142f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v24.4s }, [x21], #0x10\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v21.4s }, [x22], #0x10\n"
-      "ld1 { v25.4s }, [x21], #0x10\n"
-      "tbz x11, #2, 140f\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v18.4s }, [x23], #0x10\n"
-      "ld1 { v22.4s }, [x22], #0x10\n"
-      "ld1 { v26.4s }, [x21], #0x10\n"
-      "tbz x11, #1, 139f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "ldr d19, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "ldr d27, [x21], #0x8\n"
-      "tbz x11, #0, 146f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v19.s }[2], [x23]\n"
-      "ld1 { v23.s }[2], [x22]\n"
-      "ld1 { v27.s }[2], [x21]\n"
-      "b 146f\n"
-      "139:"  // Height 5: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 146f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s19, [x23, #0x0]\n"
-      "ldr s23, [x22, #0x0]\n"
-      "ldr s27, [x21, #0x0]\n"
-      "b 146f\n"
-      "140:"  // Height 5: Partial accumulate: partial_2_8
-      "tbz x11, #1, 141f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d22, [x22], #0x8\n"
-      "ldr d26, [x21], #0x8\n"
-      "tbz x11, #0, 146f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v18.s }[2], [x23]\n"
-      "ld1 { v22.s }[2], [x22]\n"
-      "ld1 { v26.s }[2], [x21]\n"
-      "b 146f\n"
-      "141:"  // Height 5: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 146f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s18, [x23, #0x0]\n"
-      "ldr s22, [x22, #0x0]\n"
-      "ldr s26, [x21, #0x0]\n"
-      "b 146f\n"
-      "142:"  // Height 5: Partial accumulate: partial_4_0
-      "tbz x11, #2, 144f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v24.4s }, [x21], #0x10\n"
-      "tbz x11, #1, 143f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "ldr d17, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d25, [x21], #0x8\n"
-      "tbz x11, #0, 146f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "ld1 { v17.s }[2], [x23]\n"
-      "ld1 { v21.s }[2], [x22]\n"
-      "ld1 { v25.s }[2], [x21]\n"
-      "b 146f\n"
-      "143:"  // Height 5: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 146f\n"
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "ldr s17, [x23, #0x0]\n"
-      "ldr s21, [x22, #0x0]\n"
-      "ldr s25, [x21, #0x0]\n"
-      "b 146f\n"
-      "144:"  // Height 5: Partial accumulate: partial_2_0
-      "tbz x11, #1, 145f\n"
-      "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "ldr d16, [x23], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "ldr d24, [x21], #0x8\n"
-      "tbz x11, #0, 146f\n"
-      "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "ld1 { v16.s }[2], [x23]\n"
-      "ld1 { v20.s }[2], [x22]\n"
-      "ld1 { v24.s }[2], [x21]\n"
-      "b 146f\n"
-      "145:"  // Height 5: Partial accumulate: partial_1_0
-      "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "ldr s16, [x23, #0x0]\n"
-      "ldr s20, [x22, #0x0]\n"
-      "ldr s24, [x21, #0x0]\n"
-      "146:"  // Height 5: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 149f\n"
-      "147:"  // Height 5: full accumulate
-      "ldr q8, [x9, #0x0]\n"
-      "ldr q9, [x9, #0x10]\n"
-      "ldr q10, [x9, #0x20]\n"
-      "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "ldr q16, [x23, #0x0]\n"
-      "ldr q17, [x23, #0x10]\n"
-      "ldr q18, [x23, #0x20]\n"
-      "ldr q19, [x23, #0x30]\n"
-      "ldr q20, [x22, #0x0]\n"
-      "ldr q21, [x22, #0x10]\n"
-      "ldr q22, [x22, #0x20]\n"
-      "ldr q23, [x22, #0x30]\n"
-      "ldr q24, [x21, #0x0]\n"
-      "ldr q25, [x21, #0x10]\n"
-      "ldr q26, [x21, #0x20]\n"
-      "ldr q27, [x21, #0x30]\n"
-      "b 149f\n"
-      "148:"  // Height 5: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "149:"  // Height 5: setup done
-      "mov x28, #0x0\n"
-      "150:"  // Height 5: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 151f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "ldr x22, [x20, #0x20]\n"
-      "cbnz x28, 152f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "b 152f\n"
-      "151:"  // Height 5: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "152:"  // Height 5: input setup done
-      "cmp x27, #0x10\n"
-      "blt 155f\n"
-      "ldr q0, [x26, #0x0]\n"
-      "ldr q1, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q2, [x24, #0x0]\n"
-      "ldr q3, [x23, #0x0]\n"
-      "ldr q4, [x22, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "blt 154f\n"
-      "153:"  // Height 5: Multiply loop: Main loop head
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0x4f02f0d0  // sudot v16.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d4  // sudot v20.4s, v6.16b, v3.4b[0]\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f04f0d8  // sudot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x20]\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f1  // sudot v17.4s, v7.16b, v2.4b[0]\n"
-      "add x22, x22, #0x10\n"
-      "cmp x27, #0x20\n"
-      ".inst 0x4f03f0f5  // sudot v21.4s, v7.16b, v3.4b[0]\n"
-      ".inst 0x4f04f0f9  // sudot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x30]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4f00f3aa  // sudot v10.4s, v29.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3ae  // sudot v14.4s, v29.16b, v1.4b[0]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      ".inst 0x4f02f3b2  // sudot v18.4s, v29.16b, v2.4b[0]\n"
-      ".inst 0x4f03f3b6  // sudot v22.4s, v29.16b, v3.4b[0]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      ".inst 0x4f04f3ba  // sudot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x40]\n"
-      ".inst 0x4f00f38b  // sudot v11.4s, v28.16b, v0.4b[0]\n"
-      ".inst 0x4f01f38f  // sudot v15.4s, v28.16b, v1.4b[0]\n"
-      ".inst 0x4f02f393  // sudot v19.4s, v28.16b, v2.4b[0]\n"
-      ".inst 0x4f03f397  // sudot v23.4s, v28.16b, v3.4b[0]\n"
-      ".inst 0x4f04f39b  // sudot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x50]\n"
-      ".inst 0x4f20f3a8  // sudot v8.4s, v29.16b, v0.4b[1]\n"
-      ".inst 0x4f21f3ac  // sudot v12.4s, v29.16b, v1.4b[1]\n"
-      ".inst 0x4f22f3b0  // sudot v16.4s, v29.16b, v2.4b[1]\n"
-      ".inst 0x4f23f3b4  // sudot v20.4s, v29.16b, v3.4b[1]\n"
-      ".inst 0x4f24f3b8  // sudot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x10, #0x60]\n"
-      ".inst 0x4f20f389  // sudot v9.4s, v28.16b, v0.4b[1]\n"
-      ".inst 0x4f21f38d  // sudot v13.4s, v28.16b, v1.4b[1]\n"
-      ".inst 0x4f22f391  // sudot v17.4s, v28.16b, v2.4b[1]\n"
-      ".inst 0x4f23f395  // sudot v21.4s, v28.16b, v3.4b[1]\n"
-      ".inst 0x4f24f399  // sudot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x10, #0x70]\n"
-      ".inst 0x4f20f3aa  // sudot v10.4s, v29.16b, v0.4b[1]\n"
-      ".inst 0x4f21f3ae  // sudot v14.4s, v29.16b, v1.4b[1]\n"
-      ".inst 0x4f22f3b2  // sudot v18.4s, v29.16b, v2.4b[1]\n"
-      ".inst 0x4f23f3b6  // sudot v22.4s, v29.16b, v3.4b[1]\n"
-      ".inst 0x4f24f3ba  // sudot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x10, #0x80]\n"
-      ".inst 0x4f20f38b  // sudot v11.4s, v28.16b, v0.4b[1]\n"
-      ".inst 0x4f21f38f  // sudot v15.4s, v28.16b, v1.4b[1]\n"
-      ".inst 0x4f22f393  // sudot v19.4s, v28.16b, v2.4b[1]\n"
-      ".inst 0x4f23f397  // sudot v23.4s, v28.16b, v3.4b[1]\n"
-      ".inst 0x4f24f39b  // sudot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x10, #0x90]\n"
-      ".inst 0x4f00fba8  // sudot v8.4s, v29.16b, v0.4b[2]\n"
-      ".inst 0x4f01fbac  // sudot v12.4s, v29.16b, v1.4b[2]\n"
-      ".inst 0x4f02fbb0  // sudot v16.4s, v29.16b, v2.4b[2]\n"
-      ".inst 0x4f03fbb4  // sudot v20.4s, v29.16b, v3.4b[2]\n"
-      ".inst 0x4f04fbb8  // sudot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x10, #0xa0]\n"
-      ".inst 0x4f00fb89  // sudot v9.4s, v28.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb8d  // sudot v13.4s, v28.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb91  // sudot v17.4s, v28.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb95  // sudot v21.4s, v28.16b, v3.4b[2]\n"
-      ".inst 0x4f04fb99  // sudot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x10, #0xb0]\n"
-      ".inst 0x4f00fbaa  // sudot v10.4s, v29.16b, v0.4b[2]\n"
-      ".inst 0x4f01fbae  // sudot v14.4s, v29.16b, v1.4b[2]\n"
-      ".inst 0x4f02fbb2  // sudot v18.4s, v29.16b, v2.4b[2]\n"
-      ".inst 0x4f03fbb6  // sudot v22.4s, v29.16b, v3.4b[2]\n"
-      ".inst 0x4f04fbba  // sudot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x10, #0xc0]\n"
-      ".inst 0x4f00fb8b  // sudot v11.4s, v28.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb8f  // sudot v15.4s, v28.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb93  // sudot v19.4s, v28.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb97  // sudot v23.4s, v28.16b, v3.4b[2]\n"
-      ".inst 0x4f04fb9b  // sudot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x10, #0xd0]\n"
-      ".inst 0x4f20fba8  // sudot v8.4s, v29.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbac  // sudot v12.4s, v29.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbb0  // sudot v16.4s, v29.16b, v2.4b[3]\n"
-      ".inst 0x4f23fbb4  // sudot v20.4s, v29.16b, v3.4b[3]\n"
-      ".inst 0x4f24fbb8  // sudot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr q29, [x10, #0xe0]\n"
-      ".inst 0x4f20fb89  // sudot v9.4s, v28.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb8d  // sudot v13.4s, v28.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb91  // sudot v17.4s, v28.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb95  // sudot v21.4s, v28.16b, v3.4b[3]\n"
-      ".inst 0x4f24fb99  // sudot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr q28, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20fbaa  // sudot v10.4s, v29.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbae  // sudot v14.4s, v29.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbb2  // sudot v18.4s, v29.16b, v2.4b[3]\n"
-      ".inst 0x4f23fbb6  // sudot v22.4s, v29.16b, v3.4b[3]\n"
-      ".inst 0x4f24fbba  // sudot v26.4s, v29.16b, v4.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
-      ".inst 0x4f20fb8b  // sudot v11.4s, v28.16b, v0.4b[3]\n"
-      "ldr q0, [x26, #0x0]\n"
-      ".inst 0x4f21fb8f  // sudot v15.4s, v28.16b, v1.4b[3]\n"
-      "ldr q1, [x25, #0x0]\n"
-      ".inst 0x4f22fb93  // sudot v19.4s, v28.16b, v2.4b[3]\n"
-      "ldr q2, [x24, #0x0]\n"
-      ".inst 0x4f23fb97  // sudot v23.4s, v28.16b, v3.4b[3]\n"
-      "ldr q3, [x23, #0x0]\n"
-      ".inst 0x4f24fb9b  // sudot v27.4s, v28.16b, v4.4b[3]\n"
-      "ldr q4, [x22, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "bge 153b\n"
-      "154:"  // Height 5: Multiply loop: Single iteration only
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0x4f02f0d0  // sudot v16.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d4  // sudot v20.4s, v6.16b, v3.4b[0]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f04f0d8  // sudot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x20]\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f1  // sudot v17.4s, v7.16b, v2.4b[0]\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4f03f0f5  // sudot v21.4s, v7.16b, v3.4b[0]\n"
-      ".inst 0x4f04f0f9  // sudot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
-      ".inst 0x4f00f3aa  // sudot v10.4s, v29.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3ae  // sudot v14.4s, v29.16b, v1.4b[0]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4f02f3b2  // sudot v18.4s, v29.16b, v2.4b[0]\n"
-      ".inst 0x4f03f3b6  // sudot v22.4s, v29.16b, v3.4b[0]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      ".inst 0x4f04f3ba  // sudot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x40]\n"
-      ".inst 0x4f00f38b  // sudot v11.4s, v28.16b, v0.4b[0]\n"
-      ".inst 0x4f01f38f  // sudot v15.4s, v28.16b, v1.4b[0]\n"
-      ".inst 0x4f02f393  // sudot v19.4s, v28.16b, v2.4b[0]\n"
-      ".inst 0x4f03f397  // sudot v23.4s, v28.16b, v3.4b[0]\n"
-      ".inst 0x4f04f39b  // sudot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x50]\n"
-      ".inst 0x4f20f3a8  // sudot v8.4s, v29.16b, v0.4b[1]\n"
-      ".inst 0x4f21f3ac  // sudot v12.4s, v29.16b, v1.4b[1]\n"
-      ".inst 0x4f22f3b0  // sudot v16.4s, v29.16b, v2.4b[1]\n"
-      ".inst 0x4f23f3b4  // sudot v20.4s, v29.16b, v3.4b[1]\n"
-      ".inst 0x4f24f3b8  // sudot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x10, #0x60]\n"
-      ".inst 0x4f20f389  // sudot v9.4s, v28.16b, v0.4b[1]\n"
-      ".inst 0x4f21f38d  // sudot v13.4s, v28.16b, v1.4b[1]\n"
-      ".inst 0x4f22f391  // sudot v17.4s, v28.16b, v2.4b[1]\n"
-      ".inst 0x4f23f395  // sudot v21.4s, v28.16b, v3.4b[1]\n"
-      ".inst 0x4f24f399  // sudot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x10, #0x70]\n"
-      ".inst 0x4f20f3aa  // sudot v10.4s, v29.16b, v0.4b[1]\n"
-      ".inst 0x4f21f3ae  // sudot v14.4s, v29.16b, v1.4b[1]\n"
-      ".inst 0x4f22f3b2  // sudot v18.4s, v29.16b, v2.4b[1]\n"
-      ".inst 0x4f23f3b6  // sudot v22.4s, v29.16b, v3.4b[1]\n"
-      ".inst 0x4f24f3ba  // sudot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x10, #0x80]\n"
-      ".inst 0x4f20f38b  // sudot v11.4s, v28.16b, v0.4b[1]\n"
-      ".inst 0x4f21f38f  // sudot v15.4s, v28.16b, v1.4b[1]\n"
-      ".inst 0x4f22f393  // sudot v19.4s, v28.16b, v2.4b[1]\n"
-      ".inst 0x4f23f397  // sudot v23.4s, v28.16b, v3.4b[1]\n"
-      ".inst 0x4f24f39b  // sudot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x10, #0x90]\n"
-      ".inst 0x4f00fba8  // sudot v8.4s, v29.16b, v0.4b[2]\n"
-      ".inst 0x4f01fbac  // sudot v12.4s, v29.16b, v1.4b[2]\n"
-      ".inst 0x4f02fbb0  // sudot v16.4s, v29.16b, v2.4b[2]\n"
-      ".inst 0x4f03fbb4  // sudot v20.4s, v29.16b, v3.4b[2]\n"
-      ".inst 0x4f04fbb8  // sudot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x10, #0xa0]\n"
-      ".inst 0x4f00fb89  // sudot v9.4s, v28.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb8d  // sudot v13.4s, v28.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb91  // sudot v17.4s, v28.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb95  // sudot v21.4s, v28.16b, v3.4b[2]\n"
-      ".inst 0x4f04fb99  // sudot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x10, #0xb0]\n"
-      ".inst 0x4f00fbaa  // sudot v10.4s, v29.16b, v0.4b[2]\n"
-      ".inst 0x4f01fbae  // sudot v14.4s, v29.16b, v1.4b[2]\n"
-      ".inst 0x4f02fbb2  // sudot v18.4s, v29.16b, v2.4b[2]\n"
-      ".inst 0x4f03fbb6  // sudot v22.4s, v29.16b, v3.4b[2]\n"
-      ".inst 0x4f04fbba  // sudot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x10, #0xc0]\n"
-      ".inst 0x4f00fb8b  // sudot v11.4s, v28.16b, v0.4b[2]\n"
-      ".inst 0x4f01fb8f  // sudot v15.4s, v28.16b, v1.4b[2]\n"
-      ".inst 0x4f02fb93  // sudot v19.4s, v28.16b, v2.4b[2]\n"
-      ".inst 0x4f03fb97  // sudot v23.4s, v28.16b, v3.4b[2]\n"
-      ".inst 0x4f04fb9b  // sudot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x10, #0xd0]\n"
-      ".inst 0x4f20fba8  // sudot v8.4s, v29.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbac  // sudot v12.4s, v29.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbb0  // sudot v16.4s, v29.16b, v2.4b[3]\n"
-      ".inst 0x4f23fbb4  // sudot v20.4s, v29.16b, v3.4b[3]\n"
-      ".inst 0x4f24fbb8  // sudot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr q29, [x10, #0xe0]\n"
-      ".inst 0x4f20fb89  // sudot v9.4s, v28.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb8d  // sudot v13.4s, v28.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb91  // sudot v17.4s, v28.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb95  // sudot v21.4s, v28.16b, v3.4b[3]\n"
-      ".inst 0x4f24fb99  // sudot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr q28, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20fbaa  // sudot v10.4s, v29.16b, v0.4b[3]\n"
-      ".inst 0x4f21fbae  // sudot v14.4s, v29.16b, v1.4b[3]\n"
-      ".inst 0x4f22fbb2  // sudot v18.4s, v29.16b, v2.4b[3]\n"
-      ".inst 0x4f23fbb6  // sudot v22.4s, v29.16b, v3.4b[3]\n"
-      ".inst 0x4f24fbba  // sudot v26.4s, v29.16b, v4.4b[3]\n"
-      ".inst 0x4f20fb8b  // sudot v11.4s, v28.16b, v0.4b[3]\n"
-      ".inst 0x4f21fb8f  // sudot v15.4s, v28.16b, v1.4b[3]\n"
-      ".inst 0x4f22fb93  // sudot v19.4s, v28.16b, v2.4b[3]\n"
-      ".inst 0x4f23fb97  // sudot v23.4s, v28.16b, v3.4b[3]\n"
-      ".inst 0x4f24fb9b  // sudot v27.4s, v28.16b, v4.4b[3]\n"
-      "155:"  // Height 5: Multiply loop: Main loop skip
-      "cbz x27, 160f\n"
-      "cmp x27, #0x4\n"
-      "blt 157f\n"
-      "156:"  // Height 5: Multiply loop: Odd block loop
-      "ldr s2, [x26], #0x4\n"
-      "ldr s1, [x25], #0x4\n"
-      "sub x27, x27, #0x4\n"
-      "ldr s0, [x24], #0x4\n"
-      "ldr s31, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
-      "ldr s30, [x22], #0x4\n"
-      "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
-      ".inst 0x4f02f3a8  // sudot v8.4s, v29.16b, v2.4b[0]\n"
-      ".inst 0x4f01f3ac  // sudot v12.4s, v29.16b, v1.4b[0]\n"
-      ".inst 0x4f00f3b0  // sudot v16.4s, v29.16b, v0.4b[0]\n"
-      ".inst 0x4f1ff3b4  // sudot v20.4s, v29.16b, v31.4b[0]\n"
-      ".inst 0x4f1ef3b8  // sudot v24.4s, v29.16b, v30.4b[0]\n"
-      "ldr q29, [x10, #0x20]\n"
-      ".inst 0x4f02f389  // sudot v9.4s, v28.16b, v2.4b[0]\n"
-      ".inst 0x4f01f38d  // sudot v13.4s, v28.16b, v1.4b[0]\n"
-      ".inst 0x4f00f391  // sudot v17.4s, v28.16b, v0.4b[0]\n"
-      ".inst 0x4f1ff395  // sudot v21.4s, v28.16b, v31.4b[0]\n"
-      ".inst 0x4f1ef399  // sudot v25.4s, v28.16b, v30.4b[0]\n"
-      "ldr q28, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f02f3aa  // sudot v10.4s, v29.16b, v2.4b[0]\n"
-      ".inst 0x4f01f3ae  // sudot v14.4s, v29.16b, v1.4b[0]\n"
-      ".inst 0x4f00f3b2  // sudot v18.4s, v29.16b, v0.4b[0]\n"
-      ".inst 0x4f1ff3b6  // sudot v22.4s, v29.16b, v31.4b[0]\n"
-      ".inst 0x4f1ef3ba  // sudot v26.4s, v29.16b, v30.4b[0]\n"
-      ".inst 0x4f02f38b  // sudot v11.4s, v28.16b, v2.4b[0]\n"
-      ".inst 0x4f01f38f  // sudot v15.4s, v28.16b, v1.4b[0]\n"
-      ".inst 0x4f00f393  // sudot v19.4s, v28.16b, v0.4b[0]\n"
-      ".inst 0x4f1ff397  // sudot v23.4s, v28.16b, v31.4b[0]\n"
-      ".inst 0x4f1ef39b  // sudot v27.4s, v28.16b, v30.4b[0]\n"
-      "bge 156b\n"
-      "157:"  // Height 5: Multiply loop: Skip odd blocks
-      "cbz x27, 160f\n"
-      "tbz x27, #1, 158f\n"
-      "ldr h0, [x26], #0x2\n"
-      "ldr h1, [x25], #0x2\n"
-      "ldr h2, [x24], #0x2\n"
-      "ldr h3, [x23], #0x2\n"
-      "ldr h4, [x22], #0x2\n"
-      "tbz x27, #0, 159f\n"
-      "ld1 { v0.b }[2], [x26]\n"
-      "ld1 { v1.b }[2], [x25]\n"
-      "ld1 { v2.b }[2], [x24]\n"
-      "ld1 { v3.b }[2], [x23]\n"
-      "ld1 { v4.b }[2], [x22]\n"
-      "b 159f\n"
-      "158:"  // Height 5: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x26, #0x0]\n"
-      "ldr b1, [x25, #0x0]\n"
-      "ldr b2, [x24, #0x0]\n"
-      "ldr b3, [x23, #0x0]\n"
-      "ldr b4, [x22, #0x0]\n"
-      "159:"  // Height 5: Multiply loop: Ragged operand read: Done
-      "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
-      ".inst 0x4f00f3a8  // sudot v8.4s, v29.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3ac  // sudot v12.4s, v29.16b, v1.4b[0]\n"
-      ".inst 0x4f02f3b0  // sudot v16.4s, v29.16b, v2.4b[0]\n"
-      ".inst 0x4f03f3b4  // sudot v20.4s, v29.16b, v3.4b[0]\n"
-      ".inst 0x4f04f3b8  // sudot v24.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x10, #0x20]\n"
-      ".inst 0x4f00f389  // sudot v9.4s, v28.16b, v0.4b[0]\n"
-      ".inst 0x4f01f38d  // sudot v13.4s, v28.16b, v1.4b[0]\n"
-      ".inst 0x4f02f391  // sudot v17.4s, v28.16b, v2.4b[0]\n"
-      ".inst 0x4f03f395  // sudot v21.4s, v28.16b, v3.4b[0]\n"
-      ".inst 0x4f04f399  // sudot v25.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f00f3aa  // sudot v10.4s, v29.16b, v0.4b[0]\n"
-      ".inst 0x4f01f3ae  // sudot v14.4s, v29.16b, v1.4b[0]\n"
-      ".inst 0x4f02f3b2  // sudot v18.4s, v29.16b, v2.4b[0]\n"
-      ".inst 0x4f03f3b6  // sudot v22.4s, v29.16b, v3.4b[0]\n"
-      ".inst 0x4f04f3ba  // sudot v26.4s, v29.16b, v4.4b[0]\n"
-      ".inst 0x4f00f38b  // sudot v11.4s, v28.16b, v0.4b[0]\n"
-      ".inst 0x4f01f38f  // sudot v15.4s, v28.16b, v1.4b[0]\n"
-      ".inst 0x4f02f393  // sudot v19.4s, v28.16b, v2.4b[0]\n"
-      ".inst 0x4f03f397  // sudot v23.4s, v28.16b, v3.4b[0]\n"
-      ".inst 0x4f04f39b  // sudot v27.4s, v28.16b, v4.4b[0]\n"
-      "160:"  // Height 5: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 150b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "prfm pstl1keep, [x21, #0x0]\n"
-      "bge 169f\n"
-      "tbz x11, #3, 164f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v13.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "st1 { v17.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v21.4s }, [x22], #0x10\n"
-      "st1 { v24.4s }, [x21], #0x10\n"
-      "st1 { v25.4s }, [x21], #0x10\n"
-      "tbz x11, #2, 162f\n"
-      "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v18.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x22], #0x10\n"
-      "st1 { v26.4s }, [x21], #0x10\n"
-      "tbz x11, #1, 161f\n"
-      "str d11, [x9], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "str d19, [x23], #0x8\n"
-      "str d23, [x22], #0x8\n"
-      "str d27, [x21], #0x8\n"
-      "tbz x11, #0, 168f\n"
-      "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "st1 { v19.s }[2], [x23]\n"
-      "st1 { v23.s }[2], [x22]\n"
-      "st1 { v27.s }[2], [x21]\n"
-      "b 168f\n"
-      "161:"  // Height 5: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 168f\n"
-      "str s11, [x9, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "str s19, [x23, #0x0]\n"
-      "str s23, [x22, #0x0]\n"
-      "str s27, [x21, #0x0]\n"
-      "b 168f\n"
-      "162:"  // Height 5: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 163f\n"
-      "str d10, [x9], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "str d18, [x23], #0x8\n"
-      "str d22, [x22], #0x8\n"
-      "str d26, [x21], #0x8\n"
-      "tbz x11, #0, 168f\n"
-      "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "st1 { v18.s }[2], [x23]\n"
-      "st1 { v22.s }[2], [x22]\n"
-      "st1 { v26.s }[2], [x21]\n"
-      "b 168f\n"
-      "163:"  // Height 5: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 168f\n"
-      "str s10, [x9, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "str s18, [x23, #0x0]\n"
-      "str s22, [x22, #0x0]\n"
-      "str s26, [x21, #0x0]\n"
-      "b 168f\n"
-      "164:"  // Height 5: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 166f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v24.4s }, [x21], #0x10\n"
-      "tbz x11, #1, 165f\n"
-      "str d9, [x9], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "str d17, [x23], #0x8\n"
-      "str d21, [x22], #0x8\n"
-      "str d25, [x21], #0x8\n"
-      "tbz x11, #0, 168f\n"
-      "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "st1 { v17.s }[2], [x23]\n"
-      "st1 { v21.s }[2], [x22]\n"
-      "st1 { v25.s }[2], [x21]\n"
-      "b 168f\n"
-      "165:"  // Height 5: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 168f\n"
-      "str s9, [x9, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "str s17, [x23, #0x0]\n"
-      "str s21, [x22, #0x0]\n"
-      "str s25, [x21, #0x0]\n"
-      "b 168f\n"
-      "166:"  // Height 5: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 167f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "str d16, [x23], #0x8\n"
-      "str d20, [x22], #0x8\n"
-      "str d24, [x21], #0x8\n"
-      "tbz x11, #0, 168f\n"
-      "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "st1 { v16.s }[2], [x23]\n"
-      "st1 { v20.s }[2], [x22]\n"
-      "st1 { v24.s }[2], [x21]\n"
-      "b 168f\n"
-      "167:"  // Height 5: Partial direct writeback: partial_1_0
-      "str s8, [x9, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "str s16, [x23, #0x0]\n"
-      "str s20, [x22, #0x0]\n"
-      "str s24, [x21, #0x0]\n"
-      "168:"  // Height 5: Partial direct writeback: Done
-      "b 170f\n"
-      "169:"  // Height 5: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "str q9, [x9, #0x10]\n"
-      "str q10, [x9, #0x20]\n"
-      "str q11, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q12, [x24, #0x0]\n"
-      "str q13, [x24, #0x10]\n"
-      "str q14, [x24, #0x20]\n"
-      "str q15, [x24, #0x30]\n"
-      "str q16, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q18, [x23, #0x20]\n"
-      "str q19, [x23, #0x30]\n"
-      "str q20, [x22, #0x0]\n"
-      "str q21, [x22, #0x10]\n"
-      "str q22, [x22, #0x20]\n"
-      "str q23, [x22, #0x30]\n"
-      "str q24, [x21, #0x0]\n"
-      "str q25, [x21, #0x10]\n"
-      "str q26, [x21, #0x20]\n"
-      "str q27, [x21, #0x30]\n"
-      "170:"  // Height 5: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 138b\n"
-      "b 206f\n"
-      "171:"  // Height 6
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x20, #0x18\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "172:"  // Height 6: Column loop
-      "tbz %x[flags], #0, 182f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "bge 181f\n"
-      "tbz x11, #3, 176f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v24.4s }, [x21], #0x10\n"
-      "ld1 { v28.4s }, [x20], #0x10\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v21.4s }, [x22], #0x10\n"
-      "ld1 { v25.4s }, [x21], #0x10\n"
-      "ld1 { v29.4s }, [x20], #0x10\n"
-      "tbz x11, #2, 174f\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v18.4s }, [x23], #0x10\n"
-      "ld1 { v22.4s }, [x22], #0x10\n"
-      "ld1 { v26.4s }, [x21], #0x10\n"
-      "ld1 { v30.4s }, [x20], #0x10\n"
-      "tbz x11, #1, 173f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "ldr d19, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "ldr d27, [x21], #0x8\n"
-      "ldr d31, [x20], #0x8\n"
-      "tbz x11, #0, 180f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v19.s }[2], [x23]\n"
-      "ld1 { v23.s }[2], [x22]\n"
-      "ld1 { v27.s }[2], [x21]\n"
-      "ld1 { v31.s }[2], [x20]\n"
-      "b 180f\n"
-      "173:"  // Height 6: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 180f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s19, [x23, #0x0]\n"
-      "ldr s23, [x22, #0x0]\n"
-      "ldr s27, [x21, #0x0]\n"
-      "ldr s31, [x20, #0x0]\n"
-      "b 180f\n"
-      "174:"  // Height 6: Partial accumulate: partial_2_8
-      "tbz x11, #1, 175f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d22, [x22], #0x8\n"
-      "ldr d26, [x21], #0x8\n"
-      "ldr d30, [x20], #0x8\n"
-      "tbz x11, #0, 180f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v18.s }[2], [x23]\n"
-      "ld1 { v22.s }[2], [x22]\n"
-      "ld1 { v26.s }[2], [x21]\n"
-      "ld1 { v30.s }[2], [x20]\n"
-      "b 180f\n"
-      "175:"  // Height 6: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 180f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s18, [x23, #0x0]\n"
-      "ldr s22, [x22, #0x0]\n"
-      "ldr s26, [x21, #0x0]\n"
-      "ldr s30, [x20, #0x0]\n"
-      "b 180f\n"
-      "176:"  // Height 6: Partial accumulate: partial_4_0
-      "tbz x11, #2, 178f\n"
-      "ld1 { v8.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v24.4s }, [x21], #0x10\n"
-      "ld1 { v28.4s }, [x20], #0x10\n"
-      "tbz x11, #1, 177f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "ldr d17, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d25, [x21], #0x8\n"
-      "ldr d29, [x20], #0x8\n"
-      "tbz x11, #0, 180f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "ld1 { v17.s }[2], [x23]\n"
-      "ld1 { v21.s }[2], [x22]\n"
-      "ld1 { v25.s }[2], [x21]\n"
-      "ld1 { v29.s }[2], [x20]\n"
-      "b 180f\n"
-      "177:"  // Height 6: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 180f\n"
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "ldr s17, [x23, #0x0]\n"
-      "ldr s21, [x22, #0x0]\n"
-      "ldr s25, [x21, #0x0]\n"
-      "ldr s29, [x20, #0x0]\n"
-      "b 180f\n"
-      "178:"  // Height 6: Partial accumulate: partial_2_0
-      "tbz x11, #1, 179f\n"
-      "ldr d8, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "ldr d16, [x23], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "ldr d24, [x21], #0x8\n"
-      "ldr d28, [x20], #0x8\n"
-      "tbz x11, #0, 180f\n"
-      "ld1 { v8.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "ld1 { v16.s }[2], [x23]\n"
-      "ld1 { v20.s }[2], [x22]\n"
-      "ld1 { v24.s }[2], [x21]\n"
-      "ld1 { v28.s }[2], [x20]\n"
-      "b 180f\n"
-      "179:"  // Height 6: Partial accumulate: partial_1_0
-      "ldr s8, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "ldr s16, [x23, #0x0]\n"
-      "ldr s20, [x22, #0x0]\n"
-      "ldr s24, [x21, #0x0]\n"
-      "ldr s28, [x20, #0x0]\n"
-      "180:"  // Height 6: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 183f\n"
-      "181:"  // Height 6: full accumulate
-      "ldr q8, [x9, #0x0]\n"
-      "ldr q9, [x9, #0x10]\n"
-      "ldr q10, [x9, #0x20]\n"
-      "ldr q11, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "ldr q16, [x23, #0x0]\n"
-      "ldr q17, [x23, #0x10]\n"
-      "ldr q18, [x23, #0x20]\n"
-      "ldr q19, [x23, #0x30]\n"
-      "ldr q20, [x22, #0x0]\n"
-      "ldr q21, [x22, #0x10]\n"
-      "ldr q22, [x22, #0x20]\n"
-      "ldr q23, [x22, #0x30]\n"
-      "ldr q24, [x21, #0x0]\n"
-      "ldr q25, [x21, #0x10]\n"
-      "ldr q26, [x21, #0x20]\n"
-      "ldr q27, [x21, #0x30]\n"
-      "ldr q28, [x20, #0x0]\n"
-      "ldr q29, [x20, #0x10]\n"
-      "ldr q30, [x20, #0x20]\n"
-      "ldr q31, [x20, #0x30]\n"
-      "b 183f\n"
-      "182:"  // Height 6: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "movi v28.4s, #0x0\n"
-      "movi v29.4s, #0x0\n"
-      "movi v30.4s, #0x0\n"
-      "movi v31.4s, #0x0\n"
-      "183:"  // Height 6: setup done
-      "mov x28, #0x0\n"
-      "184:"  // Height 6: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 185f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "ldr x22, [x20, #0x20]\n"
-      "ldr x21, [x20, #0x28]\n"
-      "cbnz x28, 186f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "add x21, x21, x20\n"
-      "b 186f\n"
-      "185:"  // Height 6: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "add x21, x22, x21\n"
-      "186:"  // Height 6: input setup done
-      "cmp x27, #0x10\n"
-      "blt 189f\n"
-      "ldr q0, [x26, #0x0]\n"
-      "ldr q1, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q2, [x24, #0x0]\n"
-      "ldr q3, [x23, #0x0]\n"
-      "ldr q4, [x22, #0x0]\n"
-      "ldr q5, [x21, #0x0]\n"
-      "ldr q6, [x10, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "blt 188f\n"
-      "187:"  // Height 6: Multiply loop: Main loop head
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0x4f02f0d0  // sudot v16.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d4  // sudot v20.4s, v6.16b, v3.4b[0]\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x4f04f0d8  // sudot v24.4s, v6.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0dc  // sudot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x20]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      ".inst 0x4f02f0f1  // sudot v17.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0f5  // sudot v21.4s, v7.16b, v3.4b[0]\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4f04f0f9  // sudot v25.4s, v7.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0fd  // sudot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x30]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4f00f0ca  // sudot v10.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ce  // sudot v14.4s, v6.16b, v1.4b[0]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4f02f0d2  // sudot v18.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d6  // sudot v22.4s, v6.16b, v3.4b[0]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      ".inst 0x4f04f0da  // sudot v26.4s, v6.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0de  // sudot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x40]\n"
-      ".inst 0x4f00f0eb  // sudot v11.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ef  // sudot v15.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f3  // sudot v19.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0f7  // sudot v23.4s, v7.16b, v3.4b[0]\n"
-      ".inst 0x4f04f0fb  // sudot v27.4s, v7.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0ff  // sudot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x50]\n"
-      ".inst 0x4f20f0c8  // sudot v8.4s, v6.16b, v0.4b[1]\n"
-      ".inst 0x4f21f0cc  // sudot v12.4s, v6.16b, v1.4b[1]\n"
-      ".inst 0x4f22f0d0  // sudot v16.4s, v6.16b, v2.4b[1]\n"
-      ".inst 0x4f23f0d4  // sudot v20.4s, v6.16b, v3.4b[1]\n"
-      ".inst 0x4f24f0d8  // sudot v24.4s, v6.16b, v4.4b[1]\n"
-      ".inst 0x4f25f0dc  // sudot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x10, #0x60]\n"
-      ".inst 0x4f20f0e9  // sudot v9.4s, v7.16b, v0.4b[1]\n"
-      ".inst 0x4f21f0ed  // sudot v13.4s, v7.16b, v1.4b[1]\n"
-      ".inst 0x4f22f0f1  // sudot v17.4s, v7.16b, v2.4b[1]\n"
-      ".inst 0x4f23f0f5  // sudot v21.4s, v7.16b, v3.4b[1]\n"
-      ".inst 0x4f24f0f9  // sudot v25.4s, v7.16b, v4.4b[1]\n"
-      ".inst 0x4f25f0fd  // sudot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x10, #0x70]\n"
-      ".inst 0x4f20f0ca  // sudot v10.4s, v6.16b, v0.4b[1]\n"
-      ".inst 0x4f21f0ce  // sudot v14.4s, v6.16b, v1.4b[1]\n"
-      ".inst 0x4f22f0d2  // sudot v18.4s, v6.16b, v2.4b[1]\n"
-      ".inst 0x4f23f0d6  // sudot v22.4s, v6.16b, v3.4b[1]\n"
-      ".inst 0x4f24f0da  // sudot v26.4s, v6.16b, v4.4b[1]\n"
-      ".inst 0x4f25f0de  // sudot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x10, #0x80]\n"
-      ".inst 0x4f20f0eb  // sudot v11.4s, v7.16b, v0.4b[1]\n"
-      ".inst 0x4f21f0ef  // sudot v15.4s, v7.16b, v1.4b[1]\n"
-      ".inst 0x4f22f0f3  // sudot v19.4s, v7.16b, v2.4b[1]\n"
-      ".inst 0x4f23f0f7  // sudot v23.4s, v7.16b, v3.4b[1]\n"
-      ".inst 0x4f24f0fb  // sudot v27.4s, v7.16b, v4.4b[1]\n"
-      ".inst 0x4f25f0ff  // sudot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x10, #0x90]\n"
-      ".inst 0x4f00f8c8  // sudot v8.4s, v6.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8cc  // sudot v12.4s, v6.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8d0  // sudot v16.4s, v6.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8d4  // sudot v20.4s, v6.16b, v3.4b[2]\n"
-      ".inst 0x4f04f8d8  // sudot v24.4s, v6.16b, v4.4b[2]\n"
-      ".inst 0x4f05f8dc  // sudot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x10, #0xa0]\n"
-      ".inst 0x4f00f8e9  // sudot v9.4s, v7.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8ed  // sudot v13.4s, v7.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8f1  // sudot v17.4s, v7.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8f5  // sudot v21.4s, v7.16b, v3.4b[2]\n"
-      ".inst 0x4f04f8f9  // sudot v25.4s, v7.16b, v4.4b[2]\n"
-      ".inst 0x4f05f8fd  // sudot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x10, #0xb0]\n"
-      ".inst 0x4f00f8ca  // sudot v10.4s, v6.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8ce  // sudot v14.4s, v6.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8d2  // sudot v18.4s, v6.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8d6  // sudot v22.4s, v6.16b, v3.4b[2]\n"
-      ".inst 0x4f04f8da  // sudot v26.4s, v6.16b, v4.4b[2]\n"
-      ".inst 0x4f05f8de  // sudot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x10, #0xc0]\n"
-      ".inst 0x4f00f8eb  // sudot v11.4s, v7.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8ef  // sudot v15.4s, v7.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8f3  // sudot v19.4s, v7.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8f7  // sudot v23.4s, v7.16b, v3.4b[2]\n"
-      ".inst 0x4f04f8fb  // sudot v27.4s, v7.16b, v4.4b[2]\n"
-      ".inst 0x4f05f8ff  // sudot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x10, #0xd0]\n"
-      ".inst 0x4f20f8c8  // sudot v8.4s, v6.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8cc  // sudot v12.4s, v6.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8d0  // sudot v16.4s, v6.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8d4  // sudot v20.4s, v6.16b, v3.4b[3]\n"
-      ".inst 0x4f24f8d8  // sudot v24.4s, v6.16b, v4.4b[3]\n"
-      ".inst 0x4f25f8dc  // sudot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x10, #0xe0]\n"
-      ".inst 0x4f20f8e9  // sudot v9.4s, v7.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8ed  // sudot v13.4s, v7.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8f1  // sudot v17.4s, v7.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8f5  // sudot v21.4s, v7.16b, v3.4b[3]\n"
-      ".inst 0x4f24f8f9  // sudot v25.4s, v7.16b, v4.4b[3]\n"
-      ".inst 0x4f25f8fd  // sudot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr q7, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20f8ca  // sudot v10.4s, v6.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8ce  // sudot v14.4s, v6.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8d2  // sudot v18.4s, v6.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8d6  // sudot v22.4s, v6.16b, v3.4b[3]\n"
-      ".inst 0x4f24f8da  // sudot v26.4s, v6.16b, v4.4b[3]\n"
-      ".inst 0x4f25f8de  // sudot v30.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
-      ".inst 0x4f20f8eb  // sudot v11.4s, v7.16b, v0.4b[3]\n"
-      "ldr q0, [x26, #0x0]\n"
-      ".inst 0x4f21f8ef  // sudot v15.4s, v7.16b, v1.4b[3]\n"
-      "ldr q1, [x25, #0x0]\n"
-      ".inst 0x4f22f8f3  // sudot v19.4s, v7.16b, v2.4b[3]\n"
-      "ldr q2, [x24, #0x0]\n"
-      ".inst 0x4f23f8f7  // sudot v23.4s, v7.16b, v3.4b[3]\n"
-      "ldr q3, [x23, #0x0]\n"
-      ".inst 0x4f24f8fb  // sudot v27.4s, v7.16b, v4.4b[3]\n"
-      "ldr q4, [x22, #0x0]\n"
-      ".inst 0x4f25f8ff  // sudot v31.4s, v7.16b, v5.4b[3]\n"
-      "ldr q5, [x21, #0x0]\n"
-      "ldr q7, [x10, #0x10]\n"
-      "bge 187b\n"
-      "188:"  // Height 6: Multiply loop: Single iteration only
-      ".inst 0x4f00f0c8  // sudot v8.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cc  // sudot v12.4s, v6.16b, v1.4b[0]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      ".inst 0x4f02f0d0  // sudot v16.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d4  // sudot v20.4s, v6.16b, v3.4b[0]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4f04f0d8  // sudot v24.4s, v6.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0dc  // sudot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x20]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x4f00f0e9  // sudot v9.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ed  // sudot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4f02f0f1  // sudot v17.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0f5  // sudot v21.4s, v7.16b, v3.4b[0]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      ".inst 0x4f04f0f9  // sudot v25.4s, v7.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0fd  // sudot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4f00f0ca  // sudot v10.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ce  // sudot v14.4s, v6.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      ".inst 0x4f02f0d2  // sudot v18.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d6  // sudot v22.4s, v6.16b, v3.4b[0]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      ".inst 0x4f04f0da  // sudot v26.4s, v6.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0de  // sudot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x40]\n"
-      ".inst 0x4f00f0eb  // sudot v11.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ef  // sudot v15.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f3  // sudot v19.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0f7  // sudot v23.4s, v7.16b, v3.4b[0]\n"
-      ".inst 0x4f04f0fb  // sudot v27.4s, v7.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0ff  // sudot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x50]\n"
-      ".inst 0x4f20f0c8  // sudot v8.4s, v6.16b, v0.4b[1]\n"
-      ".inst 0x4f21f0cc  // sudot v12.4s, v6.16b, v1.4b[1]\n"
-      ".inst 0x4f22f0d0  // sudot v16.4s, v6.16b, v2.4b[1]\n"
-      ".inst 0x4f23f0d4  // sudot v20.4s, v6.16b, v3.4b[1]\n"
-      ".inst 0x4f24f0d8  // sudot v24.4s, v6.16b, v4.4b[1]\n"
-      ".inst 0x4f25f0dc  // sudot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x10, #0x60]\n"
-      ".inst 0x4f20f0e9  // sudot v9.4s, v7.16b, v0.4b[1]\n"
-      ".inst 0x4f21f0ed  // sudot v13.4s, v7.16b, v1.4b[1]\n"
-      ".inst 0x4f22f0f1  // sudot v17.4s, v7.16b, v2.4b[1]\n"
-      ".inst 0x4f23f0f5  // sudot v21.4s, v7.16b, v3.4b[1]\n"
-      ".inst 0x4f24f0f9  // sudot v25.4s, v7.16b, v4.4b[1]\n"
-      ".inst 0x4f25f0fd  // sudot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x10, #0x70]\n"
-      ".inst 0x4f20f0ca  // sudot v10.4s, v6.16b, v0.4b[1]\n"
-      ".inst 0x4f21f0ce  // sudot v14.4s, v6.16b, v1.4b[1]\n"
-      ".inst 0x4f22f0d2  // sudot v18.4s, v6.16b, v2.4b[1]\n"
-      ".inst 0x4f23f0d6  // sudot v22.4s, v6.16b, v3.4b[1]\n"
-      ".inst 0x4f24f0da  // sudot v26.4s, v6.16b, v4.4b[1]\n"
-      ".inst 0x4f25f0de  // sudot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x10, #0x80]\n"
-      ".inst 0x4f20f0eb  // sudot v11.4s, v7.16b, v0.4b[1]\n"
-      ".inst 0x4f21f0ef  // sudot v15.4s, v7.16b, v1.4b[1]\n"
-      ".inst 0x4f22f0f3  // sudot v19.4s, v7.16b, v2.4b[1]\n"
-      ".inst 0x4f23f0f7  // sudot v23.4s, v7.16b, v3.4b[1]\n"
-      ".inst 0x4f24f0fb  // sudot v27.4s, v7.16b, v4.4b[1]\n"
-      ".inst 0x4f25f0ff  // sudot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x10, #0x90]\n"
-      ".inst 0x4f00f8c8  // sudot v8.4s, v6.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8cc  // sudot v12.4s, v6.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8d0  // sudot v16.4s, v6.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8d4  // sudot v20.4s, v6.16b, v3.4b[2]\n"
-      ".inst 0x4f04f8d8  // sudot v24.4s, v6.16b, v4.4b[2]\n"
-      ".inst 0x4f05f8dc  // sudot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x10, #0xa0]\n"
-      ".inst 0x4f00f8e9  // sudot v9.4s, v7.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8ed  // sudot v13.4s, v7.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8f1  // sudot v17.4s, v7.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8f5  // sudot v21.4s, v7.16b, v3.4b[2]\n"
-      ".inst 0x4f04f8f9  // sudot v25.4s, v7.16b, v4.4b[2]\n"
-      ".inst 0x4f05f8fd  // sudot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x10, #0xb0]\n"
-      ".inst 0x4f00f8ca  // sudot v10.4s, v6.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8ce  // sudot v14.4s, v6.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8d2  // sudot v18.4s, v6.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8d6  // sudot v22.4s, v6.16b, v3.4b[2]\n"
-      ".inst 0x4f04f8da  // sudot v26.4s, v6.16b, v4.4b[2]\n"
-      ".inst 0x4f05f8de  // sudot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x10, #0xc0]\n"
-      ".inst 0x4f00f8eb  // sudot v11.4s, v7.16b, v0.4b[2]\n"
-      ".inst 0x4f01f8ef  // sudot v15.4s, v7.16b, v1.4b[2]\n"
-      ".inst 0x4f02f8f3  // sudot v19.4s, v7.16b, v2.4b[2]\n"
-      ".inst 0x4f03f8f7  // sudot v23.4s, v7.16b, v3.4b[2]\n"
-      ".inst 0x4f04f8fb  // sudot v27.4s, v7.16b, v4.4b[2]\n"
-      ".inst 0x4f05f8ff  // sudot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x10, #0xd0]\n"
-      ".inst 0x4f20f8c8  // sudot v8.4s, v6.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8cc  // sudot v12.4s, v6.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8d0  // sudot v16.4s, v6.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8d4  // sudot v20.4s, v6.16b, v3.4b[3]\n"
-      ".inst 0x4f24f8d8  // sudot v24.4s, v6.16b, v4.4b[3]\n"
-      ".inst 0x4f25f8dc  // sudot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x10, #0xe0]\n"
-      ".inst 0x4f20f8e9  // sudot v9.4s, v7.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8ed  // sudot v13.4s, v7.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8f1  // sudot v17.4s, v7.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8f5  // sudot v21.4s, v7.16b, v3.4b[3]\n"
-      ".inst 0x4f24f8f9  // sudot v25.4s, v7.16b, v4.4b[3]\n"
-      ".inst 0x4f25f8fd  // sudot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr q7, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4f20f8ca  // sudot v10.4s, v6.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8ce  // sudot v14.4s, v6.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8d2  // sudot v18.4s, v6.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8d6  // sudot v22.4s, v6.16b, v3.4b[3]\n"
-      ".inst 0x4f24f8da  // sudot v26.4s, v6.16b, v4.4b[3]\n"
-      ".inst 0x4f25f8de  // sudot v30.4s, v6.16b, v5.4b[3]\n"
-      ".inst 0x4f20f8eb  // sudot v11.4s, v7.16b, v0.4b[3]\n"
-      ".inst 0x4f21f8ef  // sudot v15.4s, v7.16b, v1.4b[3]\n"
-      ".inst 0x4f22f8f3  // sudot v19.4s, v7.16b, v2.4b[3]\n"
-      ".inst 0x4f23f8f7  // sudot v23.4s, v7.16b, v3.4b[3]\n"
-      ".inst 0x4f24f8fb  // sudot v27.4s, v7.16b, v4.4b[3]\n"
-      ".inst 0x4f25f8ff  // sudot v31.4s, v7.16b, v5.4b[3]\n"
-      "189:"  // Height 6: Multiply loop: Main loop skip
-      "cbz x27, 194f\n"
-      "cmp x27, #0x4\n"
-      "blt 191f\n"
-      "190:"  // Height 6: Multiply loop: Odd block loop
-      "ldr s7, [x26], #0x4\n"
-      "ldr s6, [x25], #0x4\n"
-      "sub x27, x27, #0x4\n"
-      "ldr s5, [x24], #0x4\n"
-      "ldr s4, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
-      "ldr s3, [x22], #0x4\n"
-      "ldr s2, [x21], #0x4\n"
-      "ldr q1, [x10, #0x0]\n"
-      "ldr q0, [x10, #0x10]\n"
-      ".inst 0x4f07f028  // sudot v8.4s, v1.16b, v7.4b[0]\n"
-      ".inst 0x4f06f02c  // sudot v12.4s, v1.16b, v6.4b[0]\n"
-      ".inst 0x4f05f030  // sudot v16.4s, v1.16b, v5.4b[0]\n"
-      ".inst 0x4f04f034  // sudot v20.4s, v1.16b, v4.4b[0]\n"
-      ".inst 0x4f03f038  // sudot v24.4s, v1.16b, v3.4b[0]\n"
-      ".inst 0x4f02f03c  // sudot v28.4s, v1.16b, v2.4b[0]\n"
-      "ldr q1, [x10, #0x20]\n"
-      ".inst 0x4f07f009  // sudot v9.4s, v0.16b, v7.4b[0]\n"
-      ".inst 0x4f06f00d  // sudot v13.4s, v0.16b, v6.4b[0]\n"
-      ".inst 0x4f05f011  // sudot v17.4s, v0.16b, v5.4b[0]\n"
-      ".inst 0x4f04f015  // sudot v21.4s, v0.16b, v4.4b[0]\n"
-      ".inst 0x4f03f019  // sudot v25.4s, v0.16b, v3.4b[0]\n"
-      ".inst 0x4f02f01d  // sudot v29.4s, v0.16b, v2.4b[0]\n"
-      "ldr q0, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f07f02a  // sudot v10.4s, v1.16b, v7.4b[0]\n"
-      ".inst 0x4f06f02e  // sudot v14.4s, v1.16b, v6.4b[0]\n"
-      ".inst 0x4f05f032  // sudot v18.4s, v1.16b, v5.4b[0]\n"
-      ".inst 0x4f04f036  // sudot v22.4s, v1.16b, v4.4b[0]\n"
-      ".inst 0x4f03f03a  // sudot v26.4s, v1.16b, v3.4b[0]\n"
-      ".inst 0x4f02f03e  // sudot v30.4s, v1.16b, v2.4b[0]\n"
-      ".inst 0x4f07f00b  // sudot v11.4s, v0.16b, v7.4b[0]\n"
-      ".inst 0x4f06f00f  // sudot v15.4s, v0.16b, v6.4b[0]\n"
-      ".inst 0x4f05f013  // sudot v19.4s, v0.16b, v5.4b[0]\n"
-      ".inst 0x4f04f017  // sudot v23.4s, v0.16b, v4.4b[0]\n"
-      ".inst 0x4f03f01b  // sudot v27.4s, v0.16b, v3.4b[0]\n"
-      ".inst 0x4f02f01f  // sudot v31.4s, v0.16b, v2.4b[0]\n"
-      "bge 190b\n"
-      "191:"  // Height 6: Multiply loop: Skip odd blocks
-      "cbz x27, 194f\n"
-      "tbz x27, #1, 192f\n"
-      "ldr h0, [x26], #0x2\n"
-      "ldr h1, [x25], #0x2\n"
-      "ldr h2, [x24], #0x2\n"
-      "ldr h3, [x23], #0x2\n"
-      "ldr h4, [x22], #0x2\n"
-      "ldr h5, [x21], #0x2\n"
-      "tbz x27, #0, 193f\n"
-      "ld1 { v0.b }[2], [x26]\n"
-      "ld1 { v1.b }[2], [x25]\n"
-      "ld1 { v2.b }[2], [x24]\n"
-      "ld1 { v3.b }[2], [x23]\n"
-      "ld1 { v4.b }[2], [x22]\n"
-      "ld1 { v5.b }[2], [x21]\n"
-      "b 193f\n"
-      "192:"  // Height 6: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b0, [x26, #0x0]\n"
-      "ldr b1, [x25, #0x0]\n"
-      "ldr b2, [x24, #0x0]\n"
-      "ldr b3, [x23, #0x0]\n"
-      "ldr b4, [x22, #0x0]\n"
-      "ldr b5, [x21, #0x0]\n"
-      "193:"  // Height 6: Multiply loop: Ragged operand read: Done
-      "ldr q7, [x10, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
-      ".inst 0x4f00f0e8  // sudot v8.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ec  // sudot v12.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f0  // sudot v16.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0f4  // sudot v20.4s, v7.16b, v3.4b[0]\n"
-      ".inst 0x4f04f0f8  // sudot v24.4s, v7.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0fc  // sudot v28.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x10, #0x20]\n"
-      ".inst 0x4f00f0c9  // sudot v9.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cd  // sudot v13.4s, v6.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0d1  // sudot v17.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d5  // sudot v21.4s, v6.16b, v3.4b[0]\n"
-      ".inst 0x4f04f0d9  // sudot v25.4s, v6.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0dd  // sudot v29.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
-      ".inst 0x4f00f0ea  // sudot v10.4s, v7.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0ee  // sudot v14.4s, v7.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0f2  // sudot v18.4s, v7.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0f6  // sudot v22.4s, v7.16b, v3.4b[0]\n"
-      ".inst 0x4f04f0fa  // sudot v26.4s, v7.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0fe  // sudot v30.4s, v7.16b, v5.4b[0]\n"
-      ".inst 0x4f00f0cb  // sudot v11.4s, v6.16b, v0.4b[0]\n"
-      ".inst 0x4f01f0cf  // sudot v15.4s, v6.16b, v1.4b[0]\n"
-      ".inst 0x4f02f0d3  // sudot v19.4s, v6.16b, v2.4b[0]\n"
-      ".inst 0x4f03f0d7  // sudot v23.4s, v6.16b, v3.4b[0]\n"
-      ".inst 0x4f04f0db  // sudot v27.4s, v6.16b, v4.4b[0]\n"
-      ".inst 0x4f05f0df  // sudot v31.4s, v6.16b, v5.4b[0]\n"
-      "194:"  // Height 6: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 184b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "prfm pstl1keep, [x21, #0x0]\n"
-      "prfm pstl1keep, [x20, #0x0]\n"
-      "bge 203f\n"
-      "tbz x11, #3, 198f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v9.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v13.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "st1 { v17.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v21.4s }, [x22], #0x10\n"
-      "st1 { v24.4s }, [x21], #0x10\n"
-      "st1 { v25.4s }, [x21], #0x10\n"
-      "st1 { v28.4s }, [x20], #0x10\n"
-      "st1 { v29.4s }, [x20], #0x10\n"
-      "tbz x11, #2, 196f\n"
-      "st1 { v10.4s }, [x9], #0x10\n"
-      "st1 { v14.4s }, [x24], #0x10\n"
-      "st1 { v18.4s }, [x23], #0x10\n"
-      "st1 { v22.4s }, [x22], #0x10\n"
-      "st1 { v26.4s }, [x21], #0x10\n"
-      "st1 { v30.4s }, [x20], #0x10\n"
-      "tbz x11, #1, 195f\n"
-      "str d11, [x9], #0x8\n"
-      "str d15, [x24], #0x8\n"
-      "str d19, [x23], #0x8\n"
-      "str d23, [x22], #0x8\n"
-      "str d27, [x21], #0x8\n"
-      "str d31, [x20], #0x8\n"
-      "tbz x11, #0, 202f\n"
-      "st1 { v11.s }[2], [x9]\n"
-      "st1 { v15.s }[2], [x24]\n"
-      "st1 { v19.s }[2], [x23]\n"
-      "st1 { v23.s }[2], [x22]\n"
-      "st1 { v27.s }[2], [x21]\n"
-      "st1 { v31.s }[2], [x20]\n"
-      "b 202f\n"
-      "195:"  // Height 6: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 202f\n"
-      "str s11, [x9, #0x0]\n"
-      "str s15, [x24, #0x0]\n"
-      "str s19, [x23, #0x0]\n"
-      "str s23, [x22, #0x0]\n"
-      "str s27, [x21, #0x0]\n"
-      "str s31, [x20, #0x0]\n"
-      "b 202f\n"
-      "196:"  // Height 6: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 197f\n"
-      "str d10, [x9], #0x8\n"
-      "str d14, [x24], #0x8\n"
-      "str d18, [x23], #0x8\n"
-      "str d22, [x22], #0x8\n"
-      "str d26, [x21], #0x8\n"
-      "str d30, [x20], #0x8\n"
-      "tbz x11, #0, 202f\n"
-      "st1 { v10.s }[2], [x9]\n"
-      "st1 { v14.s }[2], [x24]\n"
-      "st1 { v18.s }[2], [x23]\n"
-      "st1 { v22.s }[2], [x22]\n"
-      "st1 { v26.s }[2], [x21]\n"
-      "st1 { v30.s }[2], [x20]\n"
-      "b 202f\n"
-      "197:"  // Height 6: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 202f\n"
-      "str s10, [x9, #0x0]\n"
-      "str s14, [x24, #0x0]\n"
-      "str s18, [x23, #0x0]\n"
-      "str s22, [x22, #0x0]\n"
-      "str s26, [x21, #0x0]\n"
-      "str s30, [x20, #0x0]\n"
-      "b 202f\n"
-      "198:"  // Height 6: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 200f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x22], #0x10\n"
-      "st1 { v24.4s }, [x21], #0x10\n"
-      "st1 { v28.4s }, [x20], #0x10\n"
-      "tbz x11, #1, 199f\n"
-      "str d9, [x9], #0x8\n"
-      "str d13, [x24], #0x8\n"
-      "str d17, [x23], #0x8\n"
-      "str d21, [x22], #0x8\n"
-      "str d25, [x21], #0x8\n"
-      "str d29, [x20], #0x8\n"
-      "tbz x11, #0, 202f\n"
-      "st1 { v9.s }[2], [x9]\n"
-      "st1 { v13.s }[2], [x24]\n"
-      "st1 { v17.s }[2], [x23]\n"
-      "st1 { v21.s }[2], [x22]\n"
-      "st1 { v25.s }[2], [x21]\n"
-      "st1 { v29.s }[2], [x20]\n"
-      "b 202f\n"
-      "199:"  // Height 6: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 202f\n"
-      "str s9, [x9, #0x0]\n"
-      "str s13, [x24, #0x0]\n"
-      "str s17, [x23, #0x0]\n"
-      "str s21, [x22, #0x0]\n"
-      "str s25, [x21, #0x0]\n"
-      "str s29, [x20, #0x0]\n"
-      "b 202f\n"
-      "200:"  // Height 6: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 201f\n"
-      "str d8, [x9], #0x8\n"
-      "str d12, [x24], #0x8\n"
-      "str d16, [x23], #0x8\n"
-      "str d20, [x22], #0x8\n"
-      "str d24, [x21], #0x8\n"
-      "str d28, [x20], #0x8\n"
-      "tbz x11, #0, 202f\n"
-      "st1 { v8.s }[2], [x9]\n"
-      "st1 { v12.s }[2], [x24]\n"
-      "st1 { v16.s }[2], [x23]\n"
-      "st1 { v20.s }[2], [x22]\n"
-      "st1 { v24.s }[2], [x21]\n"
-      "st1 { v28.s }[2], [x20]\n"
-      "b 202f\n"
-      "201:"  // Height 6: Partial direct writeback: partial_1_0
-      "str s8, [x9, #0x0]\n"
-      "str s12, [x24, #0x0]\n"
-      "str s16, [x23, #0x0]\n"
-      "str s20, [x22, #0x0]\n"
-      "str s24, [x21, #0x0]\n"
-      "str s28, [x20, #0x0]\n"
-      "202:"  // Height 6: Partial direct writeback: Done
-      "b 204f\n"
-      "203:"  // Height 6: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "str q9, [x9, #0x10]\n"
-      "str q10, [x9, #0x20]\n"
-      "str q11, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q12, [x24, #0x0]\n"
-      "str q13, [x24, #0x10]\n"
-      "str q14, [x24, #0x20]\n"
-      "str q15, [x24, #0x30]\n"
-      "str q16, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q18, [x23, #0x20]\n"
-      "str q19, [x23, #0x30]\n"
-      "str q20, [x22, #0x0]\n"
-      "str q21, [x22, #0x10]\n"
-      "str q22, [x22, #0x20]\n"
-      "str q23, [x22, #0x30]\n"
-      "str q24, [x21, #0x0]\n"
-      "str q25, [x21, #0x10]\n"
-      "str q26, [x21, #0x20]\n"
-      "str q27, [x21, #0x30]\n"
-      "str q28, [x20, #0x0]\n"
-      "str q29, [x20, #0x10]\n"
-      "str q30, [x20, #0x20]\n"
-      "str q31, [x20, #0x30]\n"
-      "204:"  // Height 6: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 172b\n"
-      "subs %x[M], %x[M], #0x6\n"
-      "beq 206f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "tbz %x[flags], #3, 205f\n"
-      "add x21, x21, #0x6\n"
-      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "b 1b\n"
-      "205:"  // Update direct input
-      "mov x20, #0x6\n"
-      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
-      "b 1b\n"
-      "206:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
-    );
-}
-
-} // namespace arm_gemm
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_mmla_6x16.hpp
deleted file mode 100644
index feda7d707a..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_mmla_6x16.hpp
+++ /dev/null
@@ -1,112 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef __aarch64__
-
-#include "../std_transforms_fixed.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    unsigned int, const unsigned int *, \
-    IndirectInputArg<uint8_t>, \
-    size_t, size_t, \
-    const int8_t *, \
-    IndirectOutputArg<int32_t>, \
-    const int32_t *, Activation, bool
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void a64_hybrid_u8s8s32_mmla_6x16( ARGLIST );
-
-class cls_a64_hybrid_u8s8s32_mmla_6x16
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef int32_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 6;
-    }
-
-    static unsigned int out_width()
-    {
-        return 16;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 8;
-    }
-
-    static constexpr bool supports_accumulate()
-    {
-        return true;
-    }
-
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 8> transforms = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-        if (std::is_same<T, uint32_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 55.05 };
-                case CPUModel::A510:
-                    return { 30.34 };
-                case CPUModel::V1:
-                    return { 83.77 };
-            }
-        }
-
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 55.31, 15.72, 0.62 };
-                case CPUModel::A510:
-                    return { 33.64, 3.92, 0.48 };
-                case CPUModel::V1:
-                    return { 63.94, 16.18, 0.83 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=a64_hybrid_u8s8s32_mmla_6x16;
-    cls_a64_hybrid_u8s8s32_mmla_6x16(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_mmla_6x16/generic.cpp
deleted file mode 100644
index 32fa470d9e..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8s8s32_mmla_6x16/generic.cpp
+++ /dev/null
@@ -1,3450 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef __aarch64__
-
-#include "arm_gemm.hpp"
-#include "../../utils.hpp"
-
-#include <cassert>
-
-namespace arm_gemm {
-
-void a64_hybrid_u8s8s32_mmla_6x16 (
-    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg,
-    size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<int32_t> output_arg,
-    const int32_t *, Activation, bool accumulate
-)
-{
-    struct KernelArgs {
-        unsigned int num_strings = {};
-        const unsigned int *string_lengths = {};
-        size_t N = {};
-        const int8_t *B_ptr = {};
-        size_t output_offset = {};
-        size_t input_initial_col = {};
-        size_t input_offset = {};
-        void *output_ptr = {};
-    } ka;
-
-    unsigned long flags=0;
-    void *input_ptr;
-
-    if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
-        ka.output_offset=output_arg.indirect.offset;
-        flags |= 0x4;
-    } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
-        ka.output_offset=output_arg.direct.stride;
-    }
-
-    if (A_arg.is_indirect) {
-        input_ptr=(void *)(A_arg.indirect.ptr);
-        ka.input_offset=A_arg.indirect.start_row;
-        ka.input_initial_col=A_arg.indirect.start_col;
-        flags |= 0x8;
-    } else {
-        assert(num_strings==1);
-        input_ptr=(void *)(A_arg.direct.base);
-        ka.input_offset=A_arg.direct.stride;
-    }
-    if (accumulate) {
-        flags |= 0x1;
-    }
-    ka.num_strings = num_strings;
-    ka.string_lengths = string_lengths;
-    ka.N = N;
-    ka.B_ptr = B_ptr;
-    __asm__ __volatile__(
-      "1:"  // Row loop
-      "cmp %x[M], #0x6\n"
-      "bge 186f\n"
-      "cmp %x[M], #0x4\n"
-      "bgt 149f\n"
-      "beq 112f\n"
-      "cmp %x[M], #0x2\n"
-      "bgt 75f\n"
-      "beq 38f\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "2:"  // Height 1: Column loop
-      "tbz %x[flags], #0, 13f\n"
-      "cmp x11, #0x10\n"
-      "bge 11f\n"
-      "tbz x11, #3, 6f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "tbz x11, #2, 4f\n"
-      "ld1 { v11.4s }, [x9], #0x10\n"
-      "tbz x11, #1, 3f\n"
-      "ldr d16, [x9], #0x8\n"
-      "mov x25, #0x38\n"
-      "tbz x11, #0, 10f\n"
-      "ld1 { v16.s }[2], [x9]\n"
-      "b 10f\n"
-      "3:"  // Height 1: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 10f\n"
-      "ldr s16, [x9, #0x0]\n"
-      "b 10f\n"
-      "4:"  // Height 1: Partial accumulate: partial_2_8
-      "tbz x11, #1, 5f\n"
-      "ldr d11, [x9], #0x8\n"
-      "mov x25, #0x28\n"
-      "tbz x11, #0, 10f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "b 10f\n"
-      "5:"  // Height 1: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 10f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "b 10f\n"
-      "6:"  // Height 1: Partial accumulate: partial_4_0
-      "tbz x11, #2, 8f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "tbz x11, #1, 7f\n"
-      "ldr d10, [x9], #0x8\n"
-      "mov x25, #0x18\n"
-      "tbz x11, #0, 10f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "b 10f\n"
-      "7:"  // Height 1: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 10f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "b 10f\n"
-      "8:"  // Height 1: Partial accumulate: partial_2_0
-      "tbz x11, #1, 9f\n"
-      "ldr d9, [x9], #0x8\n"
-      "mov x25, #0x8\n"
-      "tbz x11, #0, 10f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "b 10f\n"
-      "9:"  // Height 1: Partial accumulate: partial_1_0
-      "ldr s9, [x9, #0x0]\n"
-      "mov x25, #0x0\n"
-      "10:"  // Height 1: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 12f\n"
-      "11:"  // Height 1: full accumulate
-      "ldr q9, [x9, #0x0]\n"
-      "ldr q10, [x9, #0x10]\n"
-      "ldr q11, [x9, #0x20]\n"
-      "ldr q16, [x9, #0x30]\n"
-      "12:"  // Height 1: MMLA fixup
-      "zip1 v8.2d, v9.2d, v12.2d\n"
-      "zip2 v12.2d, v9.2d, v12.2d\n"
-      "zip1 v9.2d, v10.2d, v13.2d\n"
-      "zip2 v13.2d, v10.2d, v13.2d\n"
-      "zip1 v10.2d, v11.2d, v14.2d\n"
-      "zip2 v14.2d, v11.2d, v14.2d\n"
-      "zip1 v11.2d, v16.2d, v15.2d\n"
-      "zip2 v15.2d, v16.2d, v15.2d\n"
-      "b 14f\n"
-      "13:"  // Height 1: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "14:"  // Height 1: setup done
-      "mov x28, #0x0\n"
-      "15:"  // Height 1: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 16f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "cbnz x28, 17f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "b 17f\n"
-      "16:"  // Height 1: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "17:"  // Height 1: input setup done
-      "cmp x27, #0x10\n"
-      "blt 20f\n"
-      "ldr q1, [x26, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q6, [x10, #0x10]\n"
-      "blt 19f\n"
-      "18:"  // Height 1: Multiply loop: Main loop head
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "trn1 v19.2d, v1.2d, v20.2d\n"
-      "trn2 v1.2d, v1.2d, v20.2d\n"
-      ".inst 0x4e87ae68  // usmmla v8.4s, v19.16b, v7.16b\n"
-      "ldr q18, [x10, #0x20]\n"
-      ".inst 0x4e86ae6c  // usmmla v12.4s, v19.16b, v6.16b\n"
-      "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e92ae69  // usmmla v9.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e91ae6d  // usmmla v13.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92ae6a  // usmmla v10.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x60]\n"
-      ".inst 0x4e91ae6e  // usmmla v14.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x70]\n"
-      ".inst 0x4e92ae6b  // usmmla v11.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x80]\n"
-      ".inst 0x4e91ae6f  // usmmla v15.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x90]\n"
-      ".inst 0x4e92ac28  // usmmla v8.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xa0]\n"
-      ".inst 0x4e91ac2c  // usmmla v12.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xb0]\n"
-      ".inst 0x4e92ac29  // usmmla v9.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xc0]\n"
-      ".inst 0x4e91ac2d  // usmmla v13.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xd0]\n"
-      ".inst 0x4e92ac2a  // usmmla v10.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xe0]\n"
-      ".inst 0x4e91ac2e  // usmmla v14.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e92ac2b  // usmmla v11.4s, v1.16b, v18.16b\n"
-      "ldr q7, [x10, #0x0]\n"
-      ".inst 0x4e91ac2f  // usmmla v15.4s, v1.16b, v17.16b\n"
-      "ldr q1, [x26, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
-      "bge 18b\n"
-      "19:"  // Height 1: Multiply loop: Single iteration only
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "trn1 v19.2d, v1.2d, v17.2d\n"
-      "trn2 v1.2d, v1.2d, v17.2d\n"
-      ".inst 0x4e87ae68  // usmmla v8.4s, v19.16b, v7.16b\n"
-      "ldr q18, [x10, #0x20]\n"
-      ".inst 0x4e86ae6c  // usmmla v12.4s, v19.16b, v6.16b\n"
-      "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e92ae69  // usmmla v9.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e91ae6d  // usmmla v13.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92ae6a  // usmmla v10.4s, v19.16b, v18.16b\n"
-      "ldr q20, [x10, #0x60]\n"
-      ".inst 0x4e91ae6e  // usmmla v14.4s, v19.16b, v17.16b\n"
-      "ldr q18, [x10, #0x70]\n"
-      ".inst 0x4e94ae6b  // usmmla v11.4s, v19.16b, v20.16b\n"
-      "ldr q17, [x10, #0x80]\n"
-      ".inst 0x4e92ae6f  // usmmla v15.4s, v19.16b, v18.16b\n"
-      "ldr q20, [x10, #0x90]\n"
-      ".inst 0x4e91ac28  // usmmla v8.4s, v1.16b, v17.16b\n"
-      "ldr q18, [x10, #0xa0]\n"
-      ".inst 0x4e94ac2c  // usmmla v12.4s, v1.16b, v20.16b\n"
-      "ldr q17, [x10, #0xb0]\n"
-      ".inst 0x4e92ac29  // usmmla v9.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xc0]\n"
-      ".inst 0x4e91ac2d  // usmmla v13.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xd0]\n"
-      ".inst 0x4e92ac2a  // usmmla v10.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xe0]\n"
-      ".inst 0x4e91ac2e  // usmmla v14.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e92ac2b  // usmmla v11.4s, v1.16b, v18.16b\n"
-      ".inst 0x4e91ac2f  // usmmla v15.4s, v1.16b, v17.16b\n"
-      "20:"  // Height 1: Multiply loop: Main loop skip
-      "cbz x27, 27f\n"
-      "cmp x27, #0x8\n"
-      "blt 22f\n"
-      "21:"  // Height 1: Multiply loop: Odd block loop
-      "ldr d19, [x26], #0x8\n"
-      "ldr q20, [x10, #0x0]\n"
-      "sub x27, x27, #0x8\n"
-      "ldr q17, [x10, #0x10]\n"
-      "cmp x27, #0x8\n"
-      "trn1 v19.2d, v19.2d, v18.2d\n"
-      ".inst 0x4e94ae68  // usmmla v8.4s, v19.16b, v20.16b\n"
-      "ldr q18, [x10, #0x20]\n"
-      ".inst 0x4e91ae6c  // usmmla v12.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e92ae69  // usmmla v9.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e91ae6d  // usmmla v13.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92ae6a  // usmmla v10.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x60]\n"
-      ".inst 0x4e91ae6e  // usmmla v14.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e92ae6b  // usmmla v11.4s, v19.16b, v18.16b\n"
-      ".inst 0x4e91ae6f  // usmmla v15.4s, v19.16b, v17.16b\n"
-      "bge 21b\n"
-      "22:"  // Height 1: Multiply loop: Skip odd blocks
-      "cbz x27, 27f\n"
-      "tbz x27, #2, 24f\n"
-      "ldr s1, [x26], #0x4\n"
-      "tbz x27, #1, 23f\n"
-      "ld1 { v1.h }[2], [x26], #0x2\n"
-      "tbz x27, #0, 26f\n"
-      "ld1 { v1.b }[6], [x26]\n"
-      "b 26f\n"
-      "23:"  // Height 1: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x27, #0, 26f\n"
-      "ld1 { v1.b }[4], [x26]\n"
-      "b 26f\n"
-      "24:"  // Height 1: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x27, #1, 25f\n"
-      "ldr h1, [x26], #0x2\n"
-      "tbz x27, #0, 26f\n"
-      "ld1 { v1.b }[2], [x26]\n"
-      "b 26f\n"
-      "25:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x26, #0x0]\n"
-      "26:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q24, [x10, #0x0]\n"
-      "ldr q20, [x10, #0x10]\n"
-      "trn1 v19.2d, v1.2d, v17.2d\n"
-      ".inst 0x4e98ae68  // usmmla v8.4s, v19.16b, v24.16b\n"
-      "ldr q17, [x10, #0x20]\n"
-      ".inst 0x4e94ae6c  // usmmla v12.4s, v19.16b, v20.16b\n"
-      "ldr q0, [x10, #0x30]\n"
-      ".inst 0x4e91ae69  // usmmla v9.4s, v19.16b, v17.16b\n"
-      "ldr q20, [x10, #0x40]\n"
-      ".inst 0x4e80ae6d  // usmmla v13.4s, v19.16b, v0.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e94ae6a  // usmmla v10.4s, v19.16b, v20.16b\n"
-      "ldr q18, [x10, #0x60]\n"
-      ".inst 0x4e91ae6e  // usmmla v14.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e92ae6b  // usmmla v11.4s, v19.16b, v18.16b\n"
-      ".inst 0x4e91ae6f  // usmmla v15.4s, v19.16b, v17.16b\n"
-      "27:"  // Height 1: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 15b\n"
-      "cmp x11, #0x10\n"
-      "uzp1 v8.2d, v8.2d, v12.2d\n"
-      "uzp1 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "uzp1 v10.2d, v10.2d, v14.2d\n"
-      "uzp1 v11.2d, v11.2d, v15.2d\n"
-      "bge 36f\n"
-      "tbz x11, #3, 31f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "st1 { v9.4s }, [x9], #0x10\n"
-      "tbz x11, #2, 29f\n"
-      "st1 { v10.4s }, [x9], #0x10\n"
-      "tbz x11, #1, 28f\n"
-      "str d11, [x9], #0x8\n"
-      "tbz x11, #0, 35f\n"
-      "st1 { v11.s }[2], [x9]\n"
-      "b 35f\n"
-      "28:"  // Height 1: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 35f\n"
-      "str s11, [x9, #0x0]\n"
-      "b 35f\n"
-      "29:"  // Height 1: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 30f\n"
-      "str d10, [x9], #0x8\n"
-      "tbz x11, #0, 35f\n"
-      "st1 { v10.s }[2], [x9]\n"
-      "b 35f\n"
-      "30:"  // Height 1: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 35f\n"
-      "str s10, [x9, #0x0]\n"
-      "b 35f\n"
-      "31:"  // Height 1: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 33f\n"
-      "st1 { v8.4s }, [x9], #0x10\n"
-      "tbz x11, #1, 32f\n"
-      "str d9, [x9], #0x8\n"
-      "tbz x11, #0, 35f\n"
-      "st1 { v9.s }[2], [x9]\n"
-      "b 35f\n"
-      "32:"  // Height 1: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 35f\n"
-      "str s9, [x9, #0x0]\n"
-      "b 35f\n"
-      "33:"  // Height 1: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 34f\n"
-      "str d8, [x9], #0x8\n"
-      "tbz x11, #0, 35f\n"
-      "st1 { v8.s }[2], [x9]\n"
-      "b 35f\n"
-      "34:"  // Height 1: Partial direct writeback: partial_1_0
-      "str s8, [x9, #0x0]\n"
-      "35:"  // Height 1: Partial direct writeback: Done
-      "b 37f\n"
-      "36:"  // Height 1: Full writeback
-      "str q8, [x9, #0x0]\n"
-      "str q9, [x9, #0x10]\n"
-      "str q10, [x9, #0x20]\n"
-      "str q11, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "37:"  // Height 1: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 2b\n"
-      "b 224f\n"
-      "38:"  // Height 2
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "39:"  // Height 2: Column loop
-      "tbz %x[flags], #0, 50f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "bge 48f\n"
-      "tbz x11, #3, 43f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "tbz x11, #2, 41f\n"
-      "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "tbz x11, #1, 40f\n"
-      "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "tbz x11, #0, 47f\n"
-      "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "b 47f\n"
-      "40:"  // Height 2: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 47f\n"
-      "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "b 47f\n"
-      "41:"  // Height 2: Partial accumulate: partial_2_8
-      "tbz x11, #1, 42f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "tbz x11, #0, 47f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "b 47f\n"
-      "42:"  // Height 2: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 47f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "b 47f\n"
-      "43:"  // Height 2: Partial accumulate: partial_4_0
-      "tbz x11, #2, 45f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "tbz x11, #1, 44f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "tbz x11, #0, 47f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "b 47f\n"
-      "44:"  // Height 2: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 47f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "b 47f\n"
-      "45:"  // Height 2: Partial accumulate: partial_2_0
-      "tbz x11, #1, 46f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "tbz x11, #0, 47f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "b 47f\n"
-      "46:"  // Height 2: Partial accumulate: partial_1_0
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "47:"  // Height 2: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 49f\n"
-      "48:"  // Height 2: full accumulate
-      "ldr q9, [x9, #0x0]\n"
-      "ldr q10, [x9, #0x10]\n"
-      "ldr q11, [x9, #0x20]\n"
-      "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "49:"  // Height 2: MMLA fixup
-      "zip1 v8.2d, v9.2d, v12.2d\n"
-      "zip2 v12.2d, v9.2d, v12.2d\n"
-      "zip1 v9.2d, v10.2d, v13.2d\n"
-      "zip2 v13.2d, v10.2d, v13.2d\n"
-      "zip1 v10.2d, v11.2d, v14.2d\n"
-      "zip2 v14.2d, v11.2d, v14.2d\n"
-      "zip1 v11.2d, v16.2d, v15.2d\n"
-      "zip2 v15.2d, v16.2d, v15.2d\n"
-      "b 51f\n"
-      "50:"  // Height 2: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "51:"  // Height 2: setup done
-      "mov x28, #0x0\n"
-      "52:"  // Height 2: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 53f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "cbnz x28, 54f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "b 54f\n"
-      "53:"  // Height 2: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "54:"  // Height 2: input setup done
-      "cmp x27, #0x10\n"
-      "blt 57f\n"
-      "ldr q1, [x26, #0x0]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q7, [x10, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
-      "blt 56f\n"
-      "55:"  // Height 2: Multiply loop: Main loop head
-      "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4e87ae68  // usmmla v8.4s, v19.16b, v7.16b\n"
-      "ldr q18, [x10, #0x20]\n"
-      ".inst 0x4e86ae6c  // usmmla v12.4s, v19.16b, v6.16b\n"
-      "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e92ae69  // usmmla v9.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e91ae6d  // usmmla v13.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92ae6a  // usmmla v10.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x60]\n"
-      ".inst 0x4e91ae6e  // usmmla v14.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x70]\n"
-      ".inst 0x4e92ae6b  // usmmla v11.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x80]\n"
-      ".inst 0x4e91ae6f  // usmmla v15.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x90]\n"
-      ".inst 0x4e92ac28  // usmmla v8.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xa0]\n"
-      ".inst 0x4e91ac2c  // usmmla v12.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xb0]\n"
-      ".inst 0x4e92ac29  // usmmla v9.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xc0]\n"
-      ".inst 0x4e91ac2d  // usmmla v13.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xd0]\n"
-      ".inst 0x4e92ac2a  // usmmla v10.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xe0]\n"
-      ".inst 0x4e91ac2e  // usmmla v14.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e92ac2b  // usmmla v11.4s, v1.16b, v18.16b\n"
-      "ldr q7, [x10, #0x0]\n"
-      ".inst 0x4e91ac2f  // usmmla v15.4s, v1.16b, v17.16b\n"
-      "ldr q1, [x26, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
-      "bge 55b\n"
-      "56:"  // Height 2: Multiply loop: Single iteration only
-      "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4e87ae68  // usmmla v8.4s, v19.16b, v7.16b\n"
-      "ldr q18, [x10, #0x20]\n"
-      ".inst 0x4e86ae6c  // usmmla v12.4s, v19.16b, v6.16b\n"
-      "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e92ae69  // usmmla v9.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e91ae6d  // usmmla v13.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92ae6a  // usmmla v10.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x60]\n"
-      ".inst 0x4e91ae6e  // usmmla v14.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x70]\n"
-      ".inst 0x4e92ae6b  // usmmla v11.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x80]\n"
-      ".inst 0x4e91ae6f  // usmmla v15.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x90]\n"
-      ".inst 0x4e92ac28  // usmmla v8.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xa0]\n"
-      ".inst 0x4e91ac2c  // usmmla v12.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xb0]\n"
-      ".inst 0x4e92ac29  // usmmla v9.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xc0]\n"
-      ".inst 0x4e91ac2d  // usmmla v13.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xd0]\n"
-      ".inst 0x4e92ac2a  // usmmla v10.4s, v1.16b, v18.16b\n"
-      "ldr q18, [x10, #0xe0]\n"
-      ".inst 0x4e91ac2e  // usmmla v14.4s, v1.16b, v17.16b\n"
-      "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e92ac2b  // usmmla v11.4s, v1.16b, v18.16b\n"
-      ".inst 0x4e91ac2f  // usmmla v15.4s, v1.16b, v17.16b\n"
-      "57:"  // Height 2: Multiply loop: Main loop skip
-      "cbz x27, 64f\n"
-      "cmp x27, #0x8\n"
-      "blt 59f\n"
-      "58:"  // Height 2: Multiply loop: Odd block loop
-      "ldr d20, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
-      "sub x27, x27, #0x8\n"
-      "ldr q18, [x10, #0x0]\n"
-      "ldr q17, [x10, #0x10]\n"
-      "cmp x27, #0x8\n"
-      "trn1 v22.2d, v20.2d, v19.2d\n"
-      ".inst 0x4e92aec8  // usmmla v8.4s, v22.16b, v18.16b\n"
-      "ldr q2, [x10, #0x20]\n"
-      ".inst 0x4e91aecc  // usmmla v12.4s, v22.16b, v17.16b\n"
-      "ldr q17, [x10, #0x30]\n"
-      ".inst 0x4e82aec9  // usmmla v9.4s, v22.16b, v2.16b\n"
-      "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e91aecd  // usmmla v13.4s, v22.16b, v17.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92aeca  // usmmla v10.4s, v22.16b, v18.16b\n"
-      "ldr q18, [x10, #0x60]\n"
-      ".inst 0x4e91aece  // usmmla v14.4s, v22.16b, v17.16b\n"
-      "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e92aecb  // usmmla v11.4s, v22.16b, v18.16b\n"
-      ".inst 0x4e91aecf  // usmmla v15.4s, v22.16b, v17.16b\n"
-      "bge 58b\n"
-      "59:"  // Height 2: Multiply loop: Skip odd blocks
-      "cbz x27, 64f\n"
-      "tbz x27, #2, 61f\n"
-      "ldr s1, [x26], #0x4\n"
-      "ldr s2, [x25], #0x4\n"
-      "tbz x27, #1, 60f\n"
-      "ld1 { v1.h }[2], [x26], #0x2\n"
-      "ld1 { v2.h }[2], [x25], #0x2\n"
-      "tbz x27, #0, 63f\n"
-      "ld1 { v1.b }[6], [x26]\n"
-      "ld1 { v2.b }[6], [x25]\n"
-      "b 63f\n"
-      "60:"  // Height 2: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x27, #0, 63f\n"
-      "ld1 { v1.b }[4], [x26]\n"
-      "ld1 { v2.b }[4], [x25]\n"
-      "b 63f\n"
-      "61:"  // Height 2: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x27, #1, 62f\n"
-      "ldr h1, [x26], #0x2\n"
-      "ldr h2, [x25], #0x2\n"
-      "tbz x27, #0, 63f\n"
-      "ld1 { v1.b }[2], [x26]\n"
-      "ld1 { v2.b }[2], [x25]\n"
-      "b 63f\n"
-      "62:"  // Height 2: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x26, #0x0]\n"
-      "ldr b2, [x25, #0x0]\n"
-      "63:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "ldr q18, [x10, #0x0]\n"
-      "ldr q17, [x10, #0x10]\n"
-      "trn1 v19.2d, v1.2d, v2.2d\n"
-      ".inst 0x4e92ae68  // usmmla v8.4s, v19.16b, v18.16b\n"
-      "ldr q5, [x10, #0x20]\n"
-      ".inst 0x4e91ae6c  // usmmla v12.4s, v19.16b, v17.16b\n"
-      "ldr q21, [x10, #0x30]\n"
-      ".inst 0x4e85ae69  // usmmla v9.4s, v19.16b, v5.16b\n"
-      "ldr q18, [x10, #0x40]\n"
-      ".inst 0x4e95ae6d  // usmmla v13.4s, v19.16b, v21.16b\n"
-      "ldr q17, [x10, #0x50]\n"
-      ".inst 0x4e92ae6a  // usmmla v10.4s, v19.16b, v18.16b\n"
-      "ldr q18, [x10, #0x60]\n"
-      ".inst 0x4e91ae6e  // usmmla v14.4s, v19.16b, v17.16b\n"
-      "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e92ae6b  // usmmla v11.4s, v19.16b, v18.16b\n"
-      ".inst 0x4e91ae6f  // usmmla v15.4s, v19.16b, v17.16b\n"
-      "64:"  // Height 2: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 52b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "uzp1 v7.2d, v8.2d, v12.2d\n"
-      "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "uzp1 v13.2d, v10.2d, v14.2d\n"
-      "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "bge 73f\n"
-      "tbz x11, #3, 68f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "tbz x11, #2, 66f\n"
-      "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "tbz x11, #1, 65f\n"
-      "str d14, [x9], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "tbz x11, #0, 72f\n"
-      "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "b 72f\n"
-      "65:"  // Height 2: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 72f\n"
-      "str s14, [x9, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "b 72f\n"
-      "66:"  // Height 2: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 67f\n"
-      "str d13, [x9], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "tbz x11, #0, 72f\n"
-      "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "b 72f\n"
-      "67:"  // Height 2: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 72f\n"
-      "str s13, [x9, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "b 72f\n"
-      "68:"  // Height 2: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 70f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "tbz x11, #1, 69f\n"
-      "str d12, [x9], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "tbz x11, #0, 72f\n"
-      "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "b 72f\n"
-      "69:"  // Height 2: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 72f\n"
-      "str s12, [x9, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "b 72f\n"
-      "70:"  // Height 2: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 71f\n"
-      "str d7, [x9], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "tbz x11, #0, 72f\n"
-      "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "b 72f\n"
-      "71:"  // Height 2: Partial direct writeback: partial_1_0
-      "str s7, [x9, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "72:"  // Height 2: Partial direct writeback: Done
-      "b 74f\n"
-      "73:"  // Height 2: Full writeback
-      "str q7, [x9, #0x0]\n"
-      "str q12, [x9, #0x10]\n"
-      "str q13, [x9, #0x20]\n"
-      "str q14, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "74:"  // Height 2: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 39b\n"
-      "b 224f\n"
-      "75:"  // Height 3
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "76:"  // Height 3: Column loop
-      "tbz %x[flags], #0, 87f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "bge 85f\n"
-      "tbz x11, #3, 80f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "ld1 { v18.4s }, [x23], #0x10\n"
-      "tbz x11, #2, 78f\n"
-      "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v19.4s }, [x23], #0x10\n"
-      "tbz x11, #1, 77f\n"
-      "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "ldr d24, [x23], #0x8\n"
-      "tbz x11, #0, 84f\n"
-      "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "b 84f\n"
-      "77:"  // Height 3: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 84f\n"
-      "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "b 84f\n"
-      "78:"  // Height 3: Partial accumulate: partial_2_8
-      "tbz x11, #1, 79f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "ldr d19, [x23], #0x8\n"
-      "tbz x11, #0, 84f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v19.s }[2], [x23]\n"
-      "b 84f\n"
-      "79:"  // Height 3: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 84f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s19, [x23, #0x0]\n"
-      "b 84f\n"
-      "80:"  // Height 3: Partial accumulate: partial_4_0
-      "tbz x11, #2, 82f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "tbz x11, #1, 81f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "ldr d18, [x23], #0x8\n"
-      "tbz x11, #0, 84f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "ld1 { v18.s }[2], [x23]\n"
-      "b 84f\n"
-      "81:"  // Height 3: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 84f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "ldr s18, [x23, #0x0]\n"
-      "b 84f\n"
-      "82:"  // Height 3: Partial accumulate: partial_2_0
-      "tbz x11, #1, 83f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "ldr d17, [x23], #0x8\n"
-      "tbz x11, #0, 84f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "ld1 { v17.s }[2], [x23]\n"
-      "b 84f\n"
-      "83:"  // Height 3: Partial accumulate: partial_1_0
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "ldr s17, [x23, #0x0]\n"
-      "84:"  // Height 3: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 86f\n"
-      "85:"  // Height 3: full accumulate
-      "ldr q9, [x9, #0x0]\n"
-      "ldr q10, [x9, #0x10]\n"
-      "ldr q11, [x9, #0x20]\n"
-      "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "ldr q17, [x23, #0x0]\n"
-      "ldr q18, [x23, #0x10]\n"
-      "ldr q19, [x23, #0x20]\n"
-      "ldr q24, [x23, #0x30]\n"
-      "86:"  // Height 3: MMLA fixup
-      "zip1 v8.2d, v9.2d, v12.2d\n"
-      "zip2 v12.2d, v9.2d, v12.2d\n"
-      "zip1 v9.2d, v10.2d, v13.2d\n"
-      "zip2 v13.2d, v10.2d, v13.2d\n"
-      "zip1 v10.2d, v11.2d, v14.2d\n"
-      "zip2 v14.2d, v11.2d, v14.2d\n"
-      "zip1 v11.2d, v16.2d, v15.2d\n"
-      "zip2 v15.2d, v16.2d, v15.2d\n"
-      "zip1 v16.2d, v17.2d, v20.2d\n"
-      "zip2 v20.2d, v17.2d, v20.2d\n"
-      "zip1 v17.2d, v18.2d, v21.2d\n"
-      "zip2 v21.2d, v18.2d, v21.2d\n"
-      "zip1 v18.2d, v19.2d, v22.2d\n"
-      "zip2 v22.2d, v19.2d, v22.2d\n"
-      "zip1 v19.2d, v24.2d, v23.2d\n"
-      "zip2 v23.2d, v24.2d, v23.2d\n"
-      "b 88f\n"
-      "87:"  // Height 3: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "88:"  // Height 3: setup done
-      "mov x28, #0x0\n"
-      "89:"  // Height 3: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 90f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "cbnz x28, 91f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "b 91f\n"
-      "90:"  // Height 3: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "91:"  // Height 3: input setup done
-      "cmp x27, #0x10\n"
-      "blt 94f\n"
-      "ldr q1, [x26, #0x0]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q3, [x24, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
-      "blt 93f\n"
-      "92:"  // Height 3: Multiply loop: Main loop head
-      "trn1 v28.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
-      ".inst 0x4e87af88  // usmmla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x4e86af8c  // usmmla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
-      ".inst 0x4e87af70  // usmmla v16.4s, v27.16b, v7.16b\n"
-      "ldr q26, [x10, #0x20]\n"
-      ".inst 0x4e86af74  // usmmla v20.4s, v27.16b, v6.16b\n"
-      "ldr q25, [x10, #0x30]\n"
-      ".inst 0x4e9aaf89  // usmmla v9.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf71  // usmmla v17.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x40]\n"
-      ".inst 0x4e99af8d  // usmmla v13.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af75  // usmmla v21.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x50]\n"
-      ".inst 0x4e9aaf8a  // usmmla v10.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf72  // usmmla v18.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x60]\n"
-      ".inst 0x4e99af8e  // usmmla v14.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af76  // usmmla v22.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x70]\n"
-      ".inst 0x4e9aaf8b  // usmmla v11.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf73  // usmmla v19.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x80]\n"
-      ".inst 0x4e99af8f  // usmmla v15.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af77  // usmmla v23.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x90]\n"
-      "ldr q2, [x25, #0x0]\n"
-      ".inst 0x4e9aac28  // usmmla v8.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac70  // usmmla v16.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xa0]\n"
-      ".inst 0x4e99ac2c  // usmmla v12.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac74  // usmmla v20.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xb0]\n"
-      ".inst 0x4e9aac29  // usmmla v9.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac71  // usmmla v17.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xc0]\n"
-      ".inst 0x4e99ac2d  // usmmla v13.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac75  // usmmla v21.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xd0]\n"
-      ".inst 0x4e9aac2a  // usmmla v10.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac72  // usmmla v18.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xe0]\n"
-      ".inst 0x4e99ac2e  // usmmla v14.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac76  // usmmla v22.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e9aac2b  // usmmla v11.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac73  // usmmla v19.4s, v3.16b, v26.16b\n"
-      "ldr q7, [x10, #0x0]\n"
-      ".inst 0x4e99ac2f  // usmmla v15.4s, v1.16b, v25.16b\n"
-      "ldr q1, [x26, #0x0]\n"
-      ".inst 0x4e99ac77  // usmmla v23.4s, v3.16b, v25.16b\n"
-      "ldr q3, [x24, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
-      "bge 92b\n"
-      "93:"  // Height 3: Multiply loop: Single iteration only
-      "trn1 v28.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
-      ".inst 0x4e87af88  // usmmla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x4e86af8c  // usmmla v12.4s, v28.16b, v6.16b\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
-      ".inst 0x4e87af70  // usmmla v16.4s, v27.16b, v7.16b\n"
-      "ldr q26, [x10, #0x20]\n"
-      ".inst 0x4e86af74  // usmmla v20.4s, v27.16b, v6.16b\n"
-      "ldr q25, [x10, #0x30]\n"
-      ".inst 0x4e9aaf89  // usmmla v9.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf71  // usmmla v17.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x40]\n"
-      ".inst 0x4e99af8d  // usmmla v13.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af75  // usmmla v21.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x50]\n"
-      ".inst 0x4e9aaf8a  // usmmla v10.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf72  // usmmla v18.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x60]\n"
-      ".inst 0x4e99af8e  // usmmla v14.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af76  // usmmla v22.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x70]\n"
-      ".inst 0x4e9aaf8b  // usmmla v11.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf73  // usmmla v19.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x80]\n"
-      ".inst 0x4e99af8f  // usmmla v15.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af77  // usmmla v23.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x90]\n"
-      ".inst 0x4e9aac28  // usmmla v8.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac70  // usmmla v16.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xa0]\n"
-      ".inst 0x4e99ac2c  // usmmla v12.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac74  // usmmla v20.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xb0]\n"
-      ".inst 0x4e9aac29  // usmmla v9.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac71  // usmmla v17.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xc0]\n"
-      ".inst 0x4e99ac2d  // usmmla v13.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac75  // usmmla v21.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xd0]\n"
-      ".inst 0x4e9aac2a  // usmmla v10.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac72  // usmmla v18.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xe0]\n"
-      ".inst 0x4e99ac2e  // usmmla v14.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac76  // usmmla v22.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e9aac2b  // usmmla v11.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac73  // usmmla v19.4s, v3.16b, v26.16b\n"
-      ".inst 0x4e99ac2f  // usmmla v15.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac77  // usmmla v23.4s, v3.16b, v25.16b\n"
-      "94:"  // Height 3: Multiply loop: Main loop skip
-      "cbz x27, 101f\n"
-      "cmp x27, #0x8\n"
-      "blt 96f\n"
-      "95:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "sub x27, x27, #0x8\n"
-      "ldr d27, [x24], #0x8\n"
-      "ldr q26, [x10, #0x0]\n"
-      "cmp x27, #0x8\n"
-      "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v27.2d, v29.2d\n"
-      ".inst 0x4e9aaf88  // usmmla v8.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf70  // usmmla v16.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x20]\n"
-      ".inst 0x4e99af8c  // usmmla v12.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af74  // usmmla v20.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x30]\n"
-      ".inst 0x4e9aaf89  // usmmla v9.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf71  // usmmla v17.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x40]\n"
-      ".inst 0x4e99af8d  // usmmla v13.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af75  // usmmla v21.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x50]\n"
-      ".inst 0x4e9aaf8a  // usmmla v10.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf72  // usmmla v18.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x60]\n"
-      ".inst 0x4e99af8e  // usmmla v14.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af76  // usmmla v22.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e9aaf8b  // usmmla v11.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf73  // usmmla v19.4s, v27.16b, v26.16b\n"
-      ".inst 0x4e99af8f  // usmmla v15.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af77  // usmmla v23.4s, v27.16b, v25.16b\n"
-      "bge 95b\n"
-      "96:"  // Height 3: Multiply loop: Skip odd blocks
-      "cbz x27, 101f\n"
-      "tbz x27, #2, 98f\n"
-      "ldr s1, [x26], #0x4\n"
-      "ldr s2, [x25], #0x4\n"
-      "ldr s3, [x24], #0x4\n"
-      "tbz x27, #1, 97f\n"
-      "ld1 { v1.h }[2], [x26], #0x2\n"
-      "ld1 { v2.h }[2], [x25], #0x2\n"
-      "ld1 { v3.h }[2], [x24], #0x2\n"
-      "tbz x27, #0, 100f\n"
-      "ld1 { v1.b }[6], [x26]\n"
-      "ld1 { v2.b }[6], [x25]\n"
-      "ld1 { v3.b }[6], [x24]\n"
-      "b 100f\n"
-      "97:"  // Height 3: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x27, #0, 100f\n"
-      "ld1 { v1.b }[4], [x26]\n"
-      "ld1 { v2.b }[4], [x25]\n"
-      "ld1 { v3.b }[4], [x24]\n"
-      "b 100f\n"
-      "98:"  // Height 3: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x27, #1, 99f\n"
-      "ldr h1, [x26], #0x2\n"
-      "ldr h2, [x25], #0x2\n"
-      "ldr h3, [x24], #0x2\n"
-      "tbz x27, #0, 100f\n"
-      "ld1 { v1.b }[2], [x26]\n"
-      "ld1 { v2.b }[2], [x25]\n"
-      "ld1 { v3.b }[2], [x24]\n"
-      "b 100f\n"
-      "99:"  // Height 3: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x26, #0x0]\n"
-      "ldr b2, [x25, #0x0]\n"
-      "ldr b3, [x24, #0x0]\n"
-      "100:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "ldr q26, [x10, #0x0]\n"
-      "ldr q29, [x10, #0x10]\n"
-      "trn1 v28.2d, v1.2d, v2.2d\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
-      ".inst 0x4e9aaf88  // usmmla v8.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9daf8c  // usmmla v12.4s, v28.16b, v29.16b\n"
-      ".inst 0x4e9aaf70  // usmmla v16.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x20]\n"
-      ".inst 0x4e9daf74  // usmmla v20.4s, v27.16b, v29.16b\n"
-      "ldr q25, [x10, #0x30]\n"
-      ".inst 0x4e9aaf89  // usmmla v9.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf71  // usmmla v17.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x40]\n"
-      ".inst 0x4e99af8d  // usmmla v13.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af75  // usmmla v21.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x50]\n"
-      ".inst 0x4e9aaf8a  // usmmla v10.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf72  // usmmla v18.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x60]\n"
-      ".inst 0x4e99af8e  // usmmla v14.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af76  // usmmla v22.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e9aaf8b  // usmmla v11.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf73  // usmmla v19.4s, v27.16b, v26.16b\n"
-      ".inst 0x4e99af8f  // usmmla v15.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af77  // usmmla v23.4s, v27.16b, v25.16b\n"
-      "101:"  // Height 3: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 89b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "uzp1 v7.2d, v8.2d, v12.2d\n"
-      "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "uzp1 v13.2d, v10.2d, v14.2d\n"
-      "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "uzp1 v16.2d, v16.2d, v20.2d\n"
-      "uzp1 v17.2d, v17.2d, v21.2d\n"
-      "uzp1 v18.2d, v18.2d, v22.2d\n"
-      "uzp1 v19.2d, v19.2d, v23.2d\n"
-      "bge 110f\n"
-      "tbz x11, #3, 105f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "st1 { v17.4s }, [x23], #0x10\n"
-      "tbz x11, #2, 103f\n"
-      "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v18.4s }, [x23], #0x10\n"
-      "tbz x11, #1, 102f\n"
-      "str d14, [x9], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "str d19, [x23], #0x8\n"
-      "tbz x11, #0, 109f\n"
-      "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "st1 { v19.s }[2], [x23]\n"
-      "b 109f\n"
-      "102:"  // Height 3: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 109f\n"
-      "str s14, [x9, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "str s19, [x23, #0x0]\n"
-      "b 109f\n"
-      "103:"  // Height 3: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 104f\n"
-      "str d13, [x9], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "str d18, [x23], #0x8\n"
-      "tbz x11, #0, 109f\n"
-      "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "st1 { v18.s }[2], [x23]\n"
-      "b 109f\n"
-      "104:"  // Height 3: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 109f\n"
-      "str s13, [x9, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "str s18, [x23, #0x0]\n"
-      "b 109f\n"
-      "105:"  // Height 3: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 107f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v16.4s }, [x23], #0x10\n"
-      "tbz x11, #1, 106f\n"
-      "str d12, [x9], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "str d17, [x23], #0x8\n"
-      "tbz x11, #0, 109f\n"
-      "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "st1 { v17.s }[2], [x23]\n"
-      "b 109f\n"
-      "106:"  // Height 3: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 109f\n"
-      "str s12, [x9, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "str s17, [x23, #0x0]\n"
-      "b 109f\n"
-      "107:"  // Height 3: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 108f\n"
-      "str d7, [x9], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "str d16, [x23], #0x8\n"
-      "tbz x11, #0, 109f\n"
-      "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "st1 { v16.s }[2], [x23]\n"
-      "b 109f\n"
-      "108:"  // Height 3: Partial direct writeback: partial_1_0
-      "str s7, [x9, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "str s16, [x23, #0x0]\n"
-      "109:"  // Height 3: Partial direct writeback: Done
-      "b 111f\n"
-      "110:"  // Height 3: Full writeback
-      "str q7, [x9, #0x0]\n"
-      "str q12, [x9, #0x10]\n"
-      "str q13, [x9, #0x20]\n"
-      "str q14, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q16, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q18, [x23, #0x20]\n"
-      "str q19, [x23, #0x30]\n"
-      "111:"  // Height 3: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 76b\n"
-      "b 224f\n"
-      "112:"  // Height 4
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "113:"  // Height 4: Column loop
-      "tbz %x[flags], #0, 124f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "bge 122f\n"
-      "tbz x11, #3, 117f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "ld1 { v18.4s }, [x23], #0x10\n"
-      "ld1 { v21.4s }, [x22], #0x10\n"
-      "tbz x11, #2, 115f\n"
-      "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v19.4s }, [x23], #0x10\n"
-      "ld1 { v22.4s }, [x22], #0x10\n"
-      "tbz x11, #1, 114f\n"
-      "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "tbz x11, #0, 121f\n"
-      "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v23.s }[2], [x22]\n"
-      "b 121f\n"
-      "114:"  // Height 4: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 121f\n"
-      "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s23, [x22, #0x0]\n"
-      "b 121f\n"
-      "115:"  // Height 4: Partial accumulate: partial_2_8
-      "tbz x11, #1, 116f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "ldr d19, [x23], #0x8\n"
-      "ldr d22, [x22], #0x8\n"
-      "tbz x11, #0, 121f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v19.s }[2], [x23]\n"
-      "ld1 { v22.s }[2], [x22]\n"
-      "b 121f\n"
-      "116:"  // Height 4: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 121f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s19, [x23, #0x0]\n"
-      "ldr s22, [x22, #0x0]\n"
-      "b 121f\n"
-      "117:"  // Height 4: Partial accumulate: partial_4_0
-      "tbz x11, #2, 119f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "tbz x11, #1, 118f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "tbz x11, #0, 121f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "ld1 { v18.s }[2], [x23]\n"
-      "ld1 { v21.s }[2], [x22]\n"
-      "b 121f\n"
-      "118:"  // Height 4: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 121f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "ldr s18, [x23, #0x0]\n"
-      "ldr s21, [x22, #0x0]\n"
-      "b 121f\n"
-      "119:"  // Height 4: Partial accumulate: partial_2_0
-      "tbz x11, #1, 120f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "ldr d17, [x23], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "tbz x11, #0, 121f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "ld1 { v17.s }[2], [x23]\n"
-      "ld1 { v20.s }[2], [x22]\n"
-      "b 121f\n"
-      "120:"  // Height 4: Partial accumulate: partial_1_0
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "ldr s17, [x23, #0x0]\n"
-      "ldr s20, [x22, #0x0]\n"
-      "121:"  // Height 4: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 123f\n"
-      "122:"  // Height 4: full accumulate
-      "ldr q9, [x9, #0x0]\n"
-      "ldr q10, [x9, #0x10]\n"
-      "ldr q11, [x9, #0x20]\n"
-      "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "ldr q17, [x23, #0x0]\n"
-      "ldr q18, [x23, #0x10]\n"
-      "ldr q19, [x23, #0x20]\n"
-      "ldr q24, [x23, #0x30]\n"
-      "ldr q20, [x22, #0x0]\n"
-      "ldr q21, [x22, #0x10]\n"
-      "ldr q22, [x22, #0x20]\n"
-      "ldr q23, [x22, #0x30]\n"
-      "123:"  // Height 4: MMLA fixup
-      "zip1 v8.2d, v9.2d, v12.2d\n"
-      "zip2 v12.2d, v9.2d, v12.2d\n"
-      "zip1 v9.2d, v10.2d, v13.2d\n"
-      "zip2 v13.2d, v10.2d, v13.2d\n"
-      "zip1 v10.2d, v11.2d, v14.2d\n"
-      "zip2 v14.2d, v11.2d, v14.2d\n"
-      "zip1 v11.2d, v16.2d, v15.2d\n"
-      "zip2 v15.2d, v16.2d, v15.2d\n"
-      "zip1 v16.2d, v17.2d, v20.2d\n"
-      "zip2 v20.2d, v17.2d, v20.2d\n"
-      "zip1 v17.2d, v18.2d, v21.2d\n"
-      "zip2 v21.2d, v18.2d, v21.2d\n"
-      "zip1 v18.2d, v19.2d, v22.2d\n"
-      "zip2 v22.2d, v19.2d, v22.2d\n"
-      "zip1 v19.2d, v24.2d, v23.2d\n"
-      "zip2 v23.2d, v24.2d, v23.2d\n"
-      "b 125f\n"
-      "124:"  // Height 4: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "125:"  // Height 4: setup done
-      "mov x28, #0x0\n"
-      "126:"  // Height 4: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 127f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "cbnz x28, 128f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "b 128f\n"
-      "127:"  // Height 4: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "128:"  // Height 4: input setup done
-      "cmp x27, #0x10\n"
-      "blt 131f\n"
-      "ldr q1, [x26, #0x0]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q3, [x24, #0x0]\n"
-      "ldr q4, [x23, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
-      "blt 130f\n"
-      "129:"  // Height 4: Multiply loop: Main loop head
-      "trn1 v28.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "ldr q4, [x23, #0x0]\n"
-      ".inst 0x4e87af88  // usmmla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x4e86af8c  // usmmla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      ".inst 0x4e87af70  // usmmla v16.4s, v27.16b, v7.16b\n"
-      "ldr q26, [x10, #0x20]\n"
-      ".inst 0x4e86af74  // usmmla v20.4s, v27.16b, v6.16b\n"
-      "ldr q25, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4e9aaf89  // usmmla v9.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf71  // usmmla v17.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x40]\n"
-      ".inst 0x4e99af8d  // usmmla v13.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af75  // usmmla v21.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x50]\n"
-      ".inst 0x4e9aaf8a  // usmmla v10.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf72  // usmmla v18.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x60]\n"
-      ".inst 0x4e99af8e  // usmmla v14.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af76  // usmmla v22.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x70]\n"
-      ".inst 0x4e9aaf8b  // usmmla v11.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf73  // usmmla v19.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x80]\n"
-      ".inst 0x4e99af8f  // usmmla v15.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af77  // usmmla v23.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x90]\n"
-      "ldr q2, [x25, #0x0]\n"
-      ".inst 0x4e9aac28  // usmmla v8.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac70  // usmmla v16.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xa0]\n"
-      ".inst 0x4e99ac2c  // usmmla v12.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac74  // usmmla v20.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xb0]\n"
-      ".inst 0x4e9aac29  // usmmla v9.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac71  // usmmla v17.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xc0]\n"
-      ".inst 0x4e99ac2d  // usmmla v13.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac75  // usmmla v21.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xd0]\n"
-      ".inst 0x4e9aac2a  // usmmla v10.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac72  // usmmla v18.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xe0]\n"
-      ".inst 0x4e99ac2e  // usmmla v14.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac76  // usmmla v22.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e9aac2b  // usmmla v11.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac73  // usmmla v19.4s, v3.16b, v26.16b\n"
-      "ldr q7, [x10, #0x0]\n"
-      ".inst 0x4e99ac2f  // usmmla v15.4s, v1.16b, v25.16b\n"
-      "ldr q1, [x26, #0x0]\n"
-      ".inst 0x4e99ac77  // usmmla v23.4s, v3.16b, v25.16b\n"
-      "ldr q3, [x24, #0x0]\n"
-      "ldr q6, [x10, #0x10]\n"
-      "bge 129b\n"
-      "130:"  // Height 4: Multiply loop: Single iteration only
-      "trn1 v28.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      ".inst 0x4e87af88  // usmmla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x4e86af8c  // usmmla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4e87af70  // usmmla v16.4s, v27.16b, v7.16b\n"
-      "ldr q26, [x10, #0x20]\n"
-      ".inst 0x4e86af74  // usmmla v20.4s, v27.16b, v6.16b\n"
-      "ldr q25, [x10, #0x30]\n"
-      ".inst 0x4e9aaf89  // usmmla v9.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf71  // usmmla v17.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x40]\n"
-      ".inst 0x4e99af8d  // usmmla v13.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af75  // usmmla v21.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x50]\n"
-      ".inst 0x4e9aaf8a  // usmmla v10.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf72  // usmmla v18.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x60]\n"
-      ".inst 0x4e99af8e  // usmmla v14.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af76  // usmmla v22.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x70]\n"
-      ".inst 0x4e9aaf8b  // usmmla v11.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf73  // usmmla v19.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x80]\n"
-      ".inst 0x4e99af8f  // usmmla v15.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af77  // usmmla v23.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x90]\n"
-      ".inst 0x4e9aac28  // usmmla v8.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac70  // usmmla v16.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xa0]\n"
-      ".inst 0x4e99ac2c  // usmmla v12.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac74  // usmmla v20.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xb0]\n"
-      ".inst 0x4e9aac29  // usmmla v9.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac71  // usmmla v17.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xc0]\n"
-      ".inst 0x4e99ac2d  // usmmla v13.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac75  // usmmla v21.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xd0]\n"
-      ".inst 0x4e9aac2a  // usmmla v10.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac72  // usmmla v18.4s, v3.16b, v26.16b\n"
-      "ldr q26, [x10, #0xe0]\n"
-      ".inst 0x4e99ac2e  // usmmla v14.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac76  // usmmla v22.4s, v3.16b, v25.16b\n"
-      "ldr q25, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e9aac2b  // usmmla v11.4s, v1.16b, v26.16b\n"
-      ".inst 0x4e9aac73  // usmmla v19.4s, v3.16b, v26.16b\n"
-      ".inst 0x4e99ac2f  // usmmla v15.4s, v1.16b, v25.16b\n"
-      ".inst 0x4e99ac77  // usmmla v23.4s, v3.16b, v25.16b\n"
-      "131:"  // Height 4: Multiply loop: Main loop skip
-      "cbz x27, 138f\n"
-      "cmp x27, #0x8\n"
-      "blt 133f\n"
-      "132:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "sub x27, x27, #0x8\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
-      "cmp x27, #0x8\n"
-      "ldr q26, [x10, #0x0]\n"
-      "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v29.2d, v27.2d\n"
-      ".inst 0x4e9aaf88  // usmmla v8.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf70  // usmmla v16.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x20]\n"
-      ".inst 0x4e99af8c  // usmmla v12.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af74  // usmmla v20.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x30]\n"
-      ".inst 0x4e9aaf89  // usmmla v9.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf71  // usmmla v17.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x40]\n"
-      ".inst 0x4e99af8d  // usmmla v13.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af75  // usmmla v21.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x50]\n"
-      ".inst 0x4e9aaf8a  // usmmla v10.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf72  // usmmla v18.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x60]\n"
-      ".inst 0x4e99af8e  // usmmla v14.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af76  // usmmla v22.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e9aaf8b  // usmmla v11.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf73  // usmmla v19.4s, v27.16b, v26.16b\n"
-      ".inst 0x4e99af8f  // usmmla v15.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af77  // usmmla v23.4s, v27.16b, v25.16b\n"
-      "bge 132b\n"
-      "133:"  // Height 4: Multiply loop: Skip odd blocks
-      "cbz x27, 138f\n"
-      "tbz x27, #2, 135f\n"
-      "ldr s1, [x26], #0x4\n"
-      "ldr s2, [x25], #0x4\n"
-      "ldr s3, [x24], #0x4\n"
-      "ldr s4, [x23], #0x4\n"
-      "tbz x27, #1, 134f\n"
-      "ld1 { v1.h }[2], [x26], #0x2\n"
-      "ld1 { v2.h }[2], [x25], #0x2\n"
-      "ld1 { v3.h }[2], [x24], #0x2\n"
-      "ld1 { v4.h }[2], [x23], #0x2\n"
-      "tbz x27, #0, 137f\n"
-      "ld1 { v1.b }[6], [x26]\n"
-      "ld1 { v2.b }[6], [x25]\n"
-      "ld1 { v3.b }[6], [x24]\n"
-      "ld1 { v4.b }[6], [x23]\n"
-      "b 137f\n"
-      "134:"  // Height 4: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x27, #0, 137f\n"
-      "ld1 { v1.b }[4], [x26]\n"
-      "ld1 { v2.b }[4], [x25]\n"
-      "ld1 { v3.b }[4], [x24]\n"
-      "ld1 { v4.b }[4], [x23]\n"
-      "b 137f\n"
-      "135:"  // Height 4: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x27, #1, 136f\n"
-      "ldr h1, [x26], #0x2\n"
-      "ldr h2, [x25], #0x2\n"
-      "ldr h3, [x24], #0x2\n"
-      "ldr h4, [x23], #0x2\n"
-      "tbz x27, #0, 137f\n"
-      "ld1 { v1.b }[2], [x26]\n"
-      "ld1 { v2.b }[2], [x25]\n"
-      "ld1 { v3.b }[2], [x24]\n"
-      "ld1 { v4.b }[2], [x23]\n"
-      "b 137f\n"
-      "136:"  // Height 4: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x26, #0x0]\n"
-      "ldr b2, [x25, #0x0]\n"
-      "ldr b3, [x24, #0x0]\n"
-      "ldr b4, [x23, #0x0]\n"
-      "137:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "ldr q26, [x10, #0x0]\n"
-      "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v1.2d, v2.2d\n"
-      "trn1 v27.2d, v3.2d, v4.2d\n"
-      ".inst 0x4e9aaf88  // usmmla v8.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf70  // usmmla v16.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x20]\n"
-      ".inst 0x4e99af8c  // usmmla v12.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af74  // usmmla v20.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x30]\n"
-      ".inst 0x4e9aaf89  // usmmla v9.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf71  // usmmla v17.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x40]\n"
-      ".inst 0x4e99af8d  // usmmla v13.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af75  // usmmla v21.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x50]\n"
-      ".inst 0x4e9aaf8a  // usmmla v10.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf72  // usmmla v18.4s, v27.16b, v26.16b\n"
-      "ldr q26, [x10, #0x60]\n"
-      ".inst 0x4e99af8e  // usmmla v14.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af76  // usmmla v22.4s, v27.16b, v25.16b\n"
-      "ldr q25, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e9aaf8b  // usmmla v11.4s, v28.16b, v26.16b\n"
-      ".inst 0x4e9aaf73  // usmmla v19.4s, v27.16b, v26.16b\n"
-      ".inst 0x4e99af8f  // usmmla v15.4s, v28.16b, v25.16b\n"
-      ".inst 0x4e99af77  // usmmla v23.4s, v27.16b, v25.16b\n"
-      "138:"  // Height 4: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 126b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "uzp1 v7.2d, v8.2d, v12.2d\n"
-      "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "uzp1 v13.2d, v10.2d, v14.2d\n"
-      "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "uzp1 v15.2d, v16.2d, v20.2d\n"
-      "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "uzp1 v20.2d, v17.2d, v21.2d\n"
-      "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "uzp1 v21.2d, v18.2d, v22.2d\n"
-      "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "uzp1 v22.2d, v19.2d, v23.2d\n"
-      "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "bge 147f\n"
-      "tbz x11, #3, 142f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v16.4s }, [x22], #0x10\n"
-      "st1 { v17.4s }, [x22], #0x10\n"
-      "tbz x11, #2, 140f\n"
-      "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v18.4s }, [x22], #0x10\n"
-      "tbz x11, #1, 139f\n"
-      "str d14, [x9], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
-      "str d19, [x22], #0x8\n"
-      "tbz x11, #0, 146f\n"
-      "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
-      "st1 { v19.s }[2], [x22]\n"
-      "b 146f\n"
-      "139:"  // Height 4: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 146f\n"
-      "str s14, [x9, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
-      "str s19, [x22, #0x0]\n"
-      "b 146f\n"
-      "140:"  // Height 4: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 141f\n"
-      "str d13, [x9], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
-      "str d18, [x22], #0x8\n"
-      "tbz x11, #0, 146f\n"
-      "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
-      "st1 { v18.s }[2], [x22]\n"
-      "b 146f\n"
-      "141:"  // Height 4: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 146f\n"
-      "str s13, [x9, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
-      "str s18, [x22, #0x0]\n"
-      "b 146f\n"
-      "142:"  // Height 4: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 144f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x23], #0x10\n"
-      "st1 { v16.4s }, [x22], #0x10\n"
-      "tbz x11, #1, 143f\n"
-      "str d12, [x9], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
-      "str d17, [x22], #0x8\n"
-      "tbz x11, #0, 146f\n"
-      "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
-      "st1 { v17.s }[2], [x22]\n"
-      "b 146f\n"
-      "143:"  // Height 4: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 146f\n"
-      "str s12, [x9, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
-      "str s17, [x22, #0x0]\n"
-      "b 146f\n"
-      "144:"  // Height 4: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 145f\n"
-      "str d7, [x9], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "str d15, [x23], #0x8\n"
-      "str d16, [x22], #0x8\n"
-      "tbz x11, #0, 146f\n"
-      "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "st1 { v15.s }[2], [x23]\n"
-      "st1 { v16.s }[2], [x22]\n"
-      "b 146f\n"
-      "145:"  // Height 4: Partial direct writeback: partial_1_0
-      "str s7, [x9, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "str s15, [x23, #0x0]\n"
-      "str s16, [x22, #0x0]\n"
-      "146:"  // Height 4: Partial direct writeback: Done
-      "b 148f\n"
-      "147:"  // Height 4: Full writeback
-      "str q7, [x9, #0x0]\n"
-      "str q12, [x9, #0x10]\n"
-      "str q13, [x9, #0x20]\n"
-      "str q14, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q15, [x23, #0x0]\n"
-      "str q20, [x23, #0x10]\n"
-      "str q21, [x23, #0x20]\n"
-      "str q22, [x23, #0x30]\n"
-      "str q16, [x22, #0x0]\n"
-      "str q17, [x22, #0x10]\n"
-      "str q18, [x22, #0x20]\n"
-      "str q19, [x22, #0x30]\n"
-      "148:"  // Height 4: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 113b\n"
-      "b 224f\n"
-      "149:"  // Height 5
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "150:"  // Height 5: Column loop
-      "tbz %x[flags], #0, 161f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "bge 159f\n"
-      "tbz x11, #3, 154f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v25.4s }, [x21], #0x10\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "ld1 { v18.4s }, [x23], #0x10\n"
-      "ld1 { v21.4s }, [x22], #0x10\n"
-      "ld1 { v26.4s }, [x21], #0x10\n"
-      "tbz x11, #2, 152f\n"
-      "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v19.4s }, [x23], #0x10\n"
-      "ld1 { v22.4s }, [x22], #0x10\n"
-      "ld1 { v27.4s }, [x21], #0x10\n"
-      "tbz x11, #1, 151f\n"
-      "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "ldr d6, [x21], #0x8\n"
-      "tbz x11, #0, 158f\n"
-      "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v23.s }[2], [x22]\n"
-      "ld1 { v6.s }[2], [x21]\n"
-      "b 158f\n"
-      "151:"  // Height 5: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 158f\n"
-      "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s23, [x22, #0x0]\n"
-      "ldr s6, [x21, #0x0]\n"
-      "b 158f\n"
-      "152:"  // Height 5: Partial accumulate: partial_2_8
-      "tbz x11, #1, 153f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "ldr d19, [x23], #0x8\n"
-      "ldr d22, [x22], #0x8\n"
-      "ldr d27, [x21], #0x8\n"
-      "tbz x11, #0, 158f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v19.s }[2], [x23]\n"
-      "ld1 { v22.s }[2], [x22]\n"
-      "ld1 { v27.s }[2], [x21]\n"
-      "b 158f\n"
-      "153:"  // Height 5: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 158f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s19, [x23, #0x0]\n"
-      "ldr s22, [x22, #0x0]\n"
-      "ldr s27, [x21, #0x0]\n"
-      "b 158f\n"
-      "154:"  // Height 5: Partial accumulate: partial_4_0
-      "tbz x11, #2, 156f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v25.4s }, [x21], #0x10\n"
-      "tbz x11, #1, 155f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d26, [x21], #0x8\n"
-      "tbz x11, #0, 158f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "ld1 { v18.s }[2], [x23]\n"
-      "ld1 { v21.s }[2], [x22]\n"
-      "ld1 { v26.s }[2], [x21]\n"
-      "b 158f\n"
-      "155:"  // Height 5: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 158f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "ldr s18, [x23, #0x0]\n"
-      "ldr s21, [x22, #0x0]\n"
-      "ldr s26, [x21, #0x0]\n"
-      "b 158f\n"
-      "156:"  // Height 5: Partial accumulate: partial_2_0
-      "tbz x11, #1, 157f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "ldr d17, [x23], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "ldr d25, [x21], #0x8\n"
-      "tbz x11, #0, 158f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "ld1 { v17.s }[2], [x23]\n"
-      "ld1 { v20.s }[2], [x22]\n"
-      "ld1 { v25.s }[2], [x21]\n"
-      "b 158f\n"
-      "157:"  // Height 5: Partial accumulate: partial_1_0
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "ldr s17, [x23, #0x0]\n"
-      "ldr s20, [x22, #0x0]\n"
-      "ldr s25, [x21, #0x0]\n"
-      "158:"  // Height 5: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 160f\n"
-      "159:"  // Height 5: full accumulate
-      "ldr q9, [x9, #0x0]\n"
-      "ldr q10, [x9, #0x10]\n"
-      "ldr q11, [x9, #0x20]\n"
-      "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "ldr q17, [x23, #0x0]\n"
-      "ldr q18, [x23, #0x10]\n"
-      "ldr q19, [x23, #0x20]\n"
-      "ldr q24, [x23, #0x30]\n"
-      "ldr q20, [x22, #0x0]\n"
-      "ldr q21, [x22, #0x10]\n"
-      "ldr q22, [x22, #0x20]\n"
-      "ldr q23, [x22, #0x30]\n"
-      "ldr q25, [x21, #0x0]\n"
-      "ldr q26, [x21, #0x10]\n"
-      "ldr q27, [x21, #0x20]\n"
-      "ldr q6, [x21, #0x30]\n"
-      "160:"  // Height 5: MMLA fixup
-      "zip1 v8.2d, v9.2d, v12.2d\n"
-      "zip2 v12.2d, v9.2d, v12.2d\n"
-      "zip1 v9.2d, v10.2d, v13.2d\n"
-      "zip2 v13.2d, v10.2d, v13.2d\n"
-      "zip1 v10.2d, v11.2d, v14.2d\n"
-      "zip2 v14.2d, v11.2d, v14.2d\n"
-      "zip1 v11.2d, v16.2d, v15.2d\n"
-      "zip2 v15.2d, v16.2d, v15.2d\n"
-      "zip1 v16.2d, v17.2d, v20.2d\n"
-      "zip2 v20.2d, v17.2d, v20.2d\n"
-      "zip1 v17.2d, v18.2d, v21.2d\n"
-      "zip2 v21.2d, v18.2d, v21.2d\n"
-      "zip1 v18.2d, v19.2d, v22.2d\n"
-      "zip2 v22.2d, v19.2d, v22.2d\n"
-      "zip1 v19.2d, v24.2d, v23.2d\n"
-      "zip2 v23.2d, v24.2d, v23.2d\n"
-      "zip1 v24.2d, v25.2d, v28.2d\n"
-      "zip2 v28.2d, v25.2d, v28.2d\n"
-      "zip1 v25.2d, v26.2d, v29.2d\n"
-      "zip2 v29.2d, v26.2d, v29.2d\n"
-      "zip1 v26.2d, v27.2d, v30.2d\n"
-      "zip2 v30.2d, v27.2d, v30.2d\n"
-      "zip1 v27.2d, v6.2d, v31.2d\n"
-      "zip2 v31.2d, v6.2d, v31.2d\n"
-      "b 162f\n"
-      "161:"  // Height 5: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "movi v28.4s, #0x0\n"
-      "movi v29.4s, #0x0\n"
-      "movi v30.4s, #0x0\n"
-      "movi v31.4s, #0x0\n"
-      "162:"  // Height 5: setup done
-      "mov x28, #0x0\n"
-      "163:"  // Height 5: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 164f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "ldr x22, [x20, #0x20]\n"
-      "cbnz x28, 165f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "b 165f\n"
-      "164:"  // Height 5: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "165:"  // Height 5: input setup done
-      "cmp x27, #0x10\n"
-      "blt 168f\n"
-      "ldr q1, [x26, #0x0]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q3, [x24, #0x0]\n"
-      "ldr q4, [x23, #0x0]\n"
-      "ldr q5, [x22, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
-      "blt 167f\n"
-      "166:"  // Height 5: Multiply loop: Main loop head
-      "trn1 v6.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "trn1 v4.2d, v5.2d, v0.2d\n"
-      "trn2 v5.2d, v5.2d, v0.2d\n"
-      "ldr q0, [x10, #0x10]\n"
-      ".inst 0x4e87acc8  // usmmla v8.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87ac50  // usmmla v16.4s, v2.16b, v7.16b\n"
-      "add x22, x22, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      ".inst 0x4e87ac98  // usmmla v24.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x20]\n"
-      ".inst 0x4e80accc  // usmmla v12.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e80ac54  // usmmla v20.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e80ac9c  // usmmla v28.4s, v4.16b, v0.16b\n"
-      "ldr q0, [x10, #0x30]\n"
-      ".inst 0x4e87acc9  // usmmla v9.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87ac51  // usmmla v17.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac99  // usmmla v25.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x40]\n"
-      ".inst 0x4e80accd  // usmmla v13.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e80ac55  // usmmla v21.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e80ac9d  // usmmla v29.4s, v4.16b, v0.16b\n"
-      "ldr q0, [x10, #0x50]\n"
-      ".inst 0x4e87acca  // usmmla v10.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87ac52  // usmmla v18.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac9a  // usmmla v26.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x60]\n"
-      ".inst 0x4e80acce  // usmmla v14.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e80ac56  // usmmla v22.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e80ac9e  // usmmla v30.4s, v4.16b, v0.16b\n"
-      "ldr q0, [x10, #0x70]\n"
-      ".inst 0x4e87accb  // usmmla v11.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87ac53  // usmmla v19.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac9b  // usmmla v27.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x80]\n"
-      ".inst 0x4e80accf  // usmmla v15.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e80ac57  // usmmla v23.4s, v2.16b, v0.16b\n"
-      "ldr q2, [x25, #0x0]\n"
-      ".inst 0x4e80ac9f  // usmmla v31.4s, v4.16b, v0.16b\n"
-      "ldr q0, [x10, #0x90]\n"
-      "ldr q4, [x23, #0x0]\n"
-      ".inst 0x4e87ac28  // usmmla v8.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e87ac70  // usmmla v16.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e87acb8  // usmmla v24.4s, v5.16b, v7.16b\n"
-      "ldr q6, [x10, #0xa0]\n"
-      ".inst 0x4e80ac2c  // usmmla v12.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac74  // usmmla v20.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acbc  // usmmla v28.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xb0]\n"
-      ".inst 0x4e86ac29  // usmmla v9.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac71  // usmmla v17.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86acb9  // usmmla v25.4s, v5.16b, v6.16b\n"
-      "ldr q6, [x10, #0xc0]\n"
-      ".inst 0x4e80ac2d  // usmmla v13.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac75  // usmmla v21.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acbd  // usmmla v29.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xd0]\n"
-      ".inst 0x4e86ac2a  // usmmla v10.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac72  // usmmla v18.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86acba  // usmmla v26.4s, v5.16b, v6.16b\n"
-      "ldr q6, [x10, #0xe0]\n"
-      ".inst 0x4e80ac2e  // usmmla v14.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac76  // usmmla v22.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acbe  // usmmla v30.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e86ac2b  // usmmla v11.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac73  // usmmla v19.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86acbb  // usmmla v27.4s, v5.16b, v6.16b\n"
-      "ldr q7, [x10, #0x0]\n"
-      ".inst 0x4e80ac2f  // usmmla v15.4s, v1.16b, v0.16b\n"
-      "ldr q1, [x26, #0x0]\n"
-      ".inst 0x4e80ac77  // usmmla v23.4s, v3.16b, v0.16b\n"
-      "ldr q3, [x24, #0x0]\n"
-      ".inst 0x4e80acbf  // usmmla v31.4s, v5.16b, v0.16b\n"
-      "ldr q5, [x22, #0x0]\n"
-      "bge 166b\n"
-      "167:"  // Height 5: Multiply loop: Single iteration only
-      "trn1 v6.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "trn1 v4.2d, v5.2d, v0.2d\n"
-      "trn2 v5.2d, v5.2d, v0.2d\n"
-      "ldr q0, [x10, #0x10]\n"
-      ".inst 0x4e87acc8  // usmmla v8.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87ac50  // usmmla v16.4s, v2.16b, v7.16b\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      ".inst 0x4e87ac98  // usmmla v24.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x20]\n"
-      ".inst 0x4e80accc  // usmmla v12.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e80ac54  // usmmla v20.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e80ac9c  // usmmla v28.4s, v4.16b, v0.16b\n"
-      "ldr q0, [x10, #0x30]\n"
-      ".inst 0x4e87acc9  // usmmla v9.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87ac51  // usmmla v17.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac99  // usmmla v25.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x40]\n"
-      ".inst 0x4e80accd  // usmmla v13.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e80ac55  // usmmla v21.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e80ac9d  // usmmla v29.4s, v4.16b, v0.16b\n"
-      "ldr q0, [x10, #0x50]\n"
-      ".inst 0x4e87acca  // usmmla v10.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87ac52  // usmmla v18.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac9a  // usmmla v26.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x60]\n"
-      ".inst 0x4e80acce  // usmmla v14.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e80ac56  // usmmla v22.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e80ac9e  // usmmla v30.4s, v4.16b, v0.16b\n"
-      "ldr q0, [x10, #0x70]\n"
-      ".inst 0x4e87accb  // usmmla v11.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e87ac53  // usmmla v19.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac9b  // usmmla v27.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x80]\n"
-      ".inst 0x4e80accf  // usmmla v15.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e80ac57  // usmmla v23.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e80ac9f  // usmmla v31.4s, v4.16b, v0.16b\n"
-      "ldr q2, [x10, #0x90]\n"
-      ".inst 0x4e87ac28  // usmmla v8.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e87ac70  // usmmla v16.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e87acb8  // usmmla v24.4s, v5.16b, v7.16b\n"
-      "ldr q0, [x10, #0xa0]\n"
-      ".inst 0x4e82ac2c  // usmmla v12.4s, v1.16b, v2.16b\n"
-      ".inst 0x4e82ac74  // usmmla v20.4s, v3.16b, v2.16b\n"
-      ".inst 0x4e82acbc  // usmmla v28.4s, v5.16b, v2.16b\n"
-      "ldr q2, [x10, #0xb0]\n"
-      ".inst 0x4e80ac29  // usmmla v9.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac71  // usmmla v17.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acb9  // usmmla v25.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xc0]\n"
-      ".inst 0x4e82ac2d  // usmmla v13.4s, v1.16b, v2.16b\n"
-      ".inst 0x4e82ac75  // usmmla v21.4s, v3.16b, v2.16b\n"
-      ".inst 0x4e82acbd  // usmmla v29.4s, v5.16b, v2.16b\n"
-      "ldr q2, [x10, #0xd0]\n"
-      ".inst 0x4e80ac2a  // usmmla v10.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac72  // usmmla v18.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acba  // usmmla v26.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xe0]\n"
-      ".inst 0x4e82ac2e  // usmmla v14.4s, v1.16b, v2.16b\n"
-      ".inst 0x4e82ac76  // usmmla v22.4s, v3.16b, v2.16b\n"
-      ".inst 0x4e82acbe  // usmmla v30.4s, v5.16b, v2.16b\n"
-      "ldr q6, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e80ac2b  // usmmla v11.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac73  // usmmla v19.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acbb  // usmmla v27.4s, v5.16b, v0.16b\n"
-      ".inst 0x4e86ac2f  // usmmla v15.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac77  // usmmla v23.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86acbf  // usmmla v31.4s, v5.16b, v6.16b\n"
-      "168:"  // Height 5: Multiply loop: Main loop skip
-      "cbz x27, 175f\n"
-      "cmp x27, #0x8\n"
-      "blt 170f\n"
-      "169:"  // Height 5: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
-      "sub x27, x27, #0x8\n"
-      "ldr d3, [x24], #0x8\n"
-      "ldr d2, [x23], #0x8\n"
-      "cmp x27, #0x8\n"
-      "ldr d0, [x22], #0x8\n"
-      "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v3.2d, v2.2d\n"
-      "trn1 v2.2d, v0.2d, v5.2d\n"
-      "ldr q0, [x10, #0x10]\n"
-      ".inst 0x4e81ac88  // usmmla v8.4s, v4.16b, v1.16b\n"
-      ".inst 0x4e81ac70  // usmmla v16.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac58  // usmmla v24.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x20]\n"
-      ".inst 0x4e80ac8c  // usmmla v12.4s, v4.16b, v0.16b\n"
-      ".inst 0x4e80ac74  // usmmla v20.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5c  // usmmla v28.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x30]\n"
-      ".inst 0x4e81ac89  // usmmla v9.4s, v4.16b, v1.16b\n"
-      ".inst 0x4e81ac71  // usmmla v17.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac59  // usmmla v25.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x40]\n"
-      ".inst 0x4e80ac8d  // usmmla v13.4s, v4.16b, v0.16b\n"
-      ".inst 0x4e80ac75  // usmmla v21.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5d  // usmmla v29.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x50]\n"
-      ".inst 0x4e81ac8a  // usmmla v10.4s, v4.16b, v1.16b\n"
-      ".inst 0x4e81ac72  // usmmla v18.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac5a  // usmmla v26.4s, v2.16b, v1.16b\n"
-      "ldr q6, [x10, #0x60]\n"
-      ".inst 0x4e80ac8e  // usmmla v14.4s, v4.16b, v0.16b\n"
-      ".inst 0x4e80ac76  // usmmla v22.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5e  // usmmla v30.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e86ac8b  // usmmla v11.4s, v4.16b, v6.16b\n"
-      ".inst 0x4e86ac73  // usmmla v19.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86ac5b  // usmmla v27.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e80ac8f  // usmmla v15.4s, v4.16b, v0.16b\n"
-      ".inst 0x4e80ac77  // usmmla v23.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5f  // usmmla v31.4s, v2.16b, v0.16b\n"
-      "bge 169b\n"
-      "170:"  // Height 5: Multiply loop: Skip odd blocks
-      "cbz x27, 175f\n"
-      "tbz x27, #2, 172f\n"
-      "ldr s1, [x26], #0x4\n"
-      "ldr s2, [x25], #0x4\n"
-      "ldr s3, [x24], #0x4\n"
-      "ldr s4, [x23], #0x4\n"
-      "ldr s5, [x22], #0x4\n"
-      "tbz x27, #1, 171f\n"
-      "ld1 { v1.h }[2], [x26], #0x2\n"
-      "ld1 { v2.h }[2], [x25], #0x2\n"
-      "ld1 { v3.h }[2], [x24], #0x2\n"
-      "ld1 { v4.h }[2], [x23], #0x2\n"
-      "ld1 { v5.h }[2], [x22], #0x2\n"
-      "tbz x27, #0, 174f\n"
-      "ld1 { v1.b }[6], [x26]\n"
-      "ld1 { v2.b }[6], [x25]\n"
-      "ld1 { v3.b }[6], [x24]\n"
-      "ld1 { v4.b }[6], [x23]\n"
-      "ld1 { v5.b }[6], [x22]\n"
-      "b 174f\n"
-      "171:"  // Height 5: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x27, #0, 174f\n"
-      "ld1 { v1.b }[4], [x26]\n"
-      "ld1 { v2.b }[4], [x25]\n"
-      "ld1 { v3.b }[4], [x24]\n"
-      "ld1 { v4.b }[4], [x23]\n"
-      "ld1 { v5.b }[4], [x22]\n"
-      "b 174f\n"
-      "172:"  // Height 5: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x27, #1, 173f\n"
-      "ldr h1, [x26], #0x2\n"
-      "ldr h2, [x25], #0x2\n"
-      "ldr h3, [x24], #0x2\n"
-      "ldr h4, [x23], #0x2\n"
-      "ldr h5, [x22], #0x2\n"
-      "tbz x27, #0, 174f\n"
-      "ld1 { v1.b }[2], [x26]\n"
-      "ld1 { v2.b }[2], [x25]\n"
-      "ld1 { v3.b }[2], [x24]\n"
-      "ld1 { v4.b }[2], [x23]\n"
-      "ld1 { v5.b }[2], [x22]\n"
-      "b 174f\n"
-      "173:"  // Height 5: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x26, #0x0]\n"
-      "ldr b2, [x25, #0x0]\n"
-      "ldr b3, [x24, #0x0]\n"
-      "ldr b4, [x23, #0x0]\n"
-      "ldr b5, [x22, #0x0]\n"
-      "174:"  // Height 5: Multiply loop: Ragged operand read: Done
-      "ldr q6, [x10, #0x0]\n"
-      "trn1 v7.2d, v1.2d, v2.2d\n"
-      "trn1 v3.2d, v3.2d, v4.2d\n"
-      "trn1 v2.2d, v5.2d, v0.2d\n"
-      "ldr q1, [x10, #0x10]\n"
-      ".inst 0x4e86ace8  // usmmla v8.4s, v7.16b, v6.16b\n"
-      ".inst 0x4e86ac70  // usmmla v16.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86ac58  // usmmla v24.4s, v2.16b, v6.16b\n"
-      "ldr q0, [x10, #0x20]\n"
-      ".inst 0x4e81acec  // usmmla v12.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e81ac74  // usmmla v20.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac5c  // usmmla v28.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x30]\n"
-      ".inst 0x4e80ace9  // usmmla v9.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e80ac71  // usmmla v17.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac59  // usmmla v25.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x40]\n"
-      ".inst 0x4e81aced  // usmmla v13.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e81ac75  // usmmla v21.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac5d  // usmmla v29.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x50]\n"
-      ".inst 0x4e80acea  // usmmla v10.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e80ac72  // usmmla v18.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5a  // usmmla v26.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x60]\n"
-      ".inst 0x4e81acee  // usmmla v14.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e81ac76  // usmmla v22.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac5e  // usmmla v30.4s, v2.16b, v1.16b\n"
-      "ldr q6, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e80aceb  // usmmla v11.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e80ac73  // usmmla v19.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5b  // usmmla v27.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e86acef  // usmmla v15.4s, v7.16b, v6.16b\n"
-      ".inst 0x4e86ac77  // usmmla v23.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86ac5f  // usmmla v31.4s, v2.16b, v6.16b\n"
-      "175:"  // Height 5: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 163b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "uzp1 v7.2d, v8.2d, v12.2d\n"
-      "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "uzp1 v13.2d, v10.2d, v14.2d\n"
-      "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "uzp1 v15.2d, v16.2d, v20.2d\n"
-      "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "uzp1 v20.2d, v17.2d, v21.2d\n"
-      "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "prfm pstl1keep, [x21, #0x0]\n"
-      "uzp1 v21.2d, v18.2d, v22.2d\n"
-      "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "uzp1 v22.2d, v19.2d, v23.2d\n"
-      "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "uzp1 v24.2d, v24.2d, v28.2d\n"
-      "uzp1 v25.2d, v25.2d, v29.2d\n"
-      "uzp1 v26.2d, v26.2d, v30.2d\n"
-      "uzp1 v27.2d, v27.2d, v31.2d\n"
-      "bge 184f\n"
-      "tbz x11, #3, 179f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v16.4s }, [x22], #0x10\n"
-      "st1 { v17.4s }, [x22], #0x10\n"
-      "st1 { v24.4s }, [x21], #0x10\n"
-      "st1 { v25.4s }, [x21], #0x10\n"
-      "tbz x11, #2, 177f\n"
-      "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v18.4s }, [x22], #0x10\n"
-      "st1 { v26.4s }, [x21], #0x10\n"
-      "tbz x11, #1, 176f\n"
-      "str d14, [x9], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
-      "str d19, [x22], #0x8\n"
-      "str d27, [x21], #0x8\n"
-      "tbz x11, #0, 183f\n"
-      "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
-      "st1 { v19.s }[2], [x22]\n"
-      "st1 { v27.s }[2], [x21]\n"
-      "b 183f\n"
-      "176:"  // Height 5: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 183f\n"
-      "str s14, [x9, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
-      "str s19, [x22, #0x0]\n"
-      "str s27, [x21, #0x0]\n"
-      "b 183f\n"
-      "177:"  // Height 5: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 178f\n"
-      "str d13, [x9], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
-      "str d18, [x22], #0x8\n"
-      "str d26, [x21], #0x8\n"
-      "tbz x11, #0, 183f\n"
-      "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
-      "st1 { v18.s }[2], [x22]\n"
-      "st1 { v26.s }[2], [x21]\n"
-      "b 183f\n"
-      "178:"  // Height 5: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 183f\n"
-      "str s13, [x9, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
-      "str s18, [x22, #0x0]\n"
-      "str s26, [x21, #0x0]\n"
-      "b 183f\n"
-      "179:"  // Height 5: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 181f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x23], #0x10\n"
-      "st1 { v16.4s }, [x22], #0x10\n"
-      "st1 { v24.4s }, [x21], #0x10\n"
-      "tbz x11, #1, 180f\n"
-      "str d12, [x9], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
-      "str d17, [x22], #0x8\n"
-      "str d25, [x21], #0x8\n"
-      "tbz x11, #0, 183f\n"
-      "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
-      "st1 { v17.s }[2], [x22]\n"
-      "st1 { v25.s }[2], [x21]\n"
-      "b 183f\n"
-      "180:"  // Height 5: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 183f\n"
-      "str s12, [x9, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
-      "str s17, [x22, #0x0]\n"
-      "str s25, [x21, #0x0]\n"
-      "b 183f\n"
-      "181:"  // Height 5: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 182f\n"
-      "str d7, [x9], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "str d15, [x23], #0x8\n"
-      "str d16, [x22], #0x8\n"
-      "str d24, [x21], #0x8\n"
-      "tbz x11, #0, 183f\n"
-      "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "st1 { v15.s }[2], [x23]\n"
-      "st1 { v16.s }[2], [x22]\n"
-      "st1 { v24.s }[2], [x21]\n"
-      "b 183f\n"
-      "182:"  // Height 5: Partial direct writeback: partial_1_0
-      "str s7, [x9, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "str s15, [x23, #0x0]\n"
-      "str s16, [x22, #0x0]\n"
-      "str s24, [x21, #0x0]\n"
-      "183:"  // Height 5: Partial direct writeback: Done
-      "b 185f\n"
-      "184:"  // Height 5: Full writeback
-      "str q7, [x9, #0x0]\n"
-      "str q12, [x9, #0x10]\n"
-      "str q13, [x9, #0x20]\n"
-      "str q14, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q15, [x23, #0x0]\n"
-      "str q20, [x23, #0x10]\n"
-      "str q21, [x23, #0x20]\n"
-      "str q22, [x23, #0x30]\n"
-      "str q16, [x22, #0x0]\n"
-      "str q17, [x22, #0x10]\n"
-      "str q18, [x22, #0x20]\n"
-      "str q19, [x22, #0x30]\n"
-      "str q24, [x21, #0x0]\n"
-      "str q25, [x21, #0x10]\n"
-      "str q26, [x21, #0x20]\n"
-      "str q27, [x21, #0x30]\n"
-      "185:"  // Height 5: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 150b\n"
-      "b 224f\n"
-      "186:"  // Height 6
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x20, #0x18\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "187:"  // Height 6: Column loop
-      "tbz %x[flags], #0, 198f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "bge 196f\n"
-      "tbz x11, #3, 191f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v25.4s }, [x21], #0x10\n"
-      "ld1 { v28.4s }, [x20], #0x10\n"
-      "ld1 { v10.4s }, [x9], #0x10\n"
-      "ld1 { v13.4s }, [x24], #0x10\n"
-      "ld1 { v18.4s }, [x23], #0x10\n"
-      "ld1 { v21.4s }, [x22], #0x10\n"
-      "ld1 { v26.4s }, [x21], #0x10\n"
-      "ld1 { v29.4s }, [x20], #0x10\n"
-      "tbz x11, #2, 189f\n"
-      "ld1 { v11.4s }, [x9], #0x10\n"
-      "ld1 { v14.4s }, [x24], #0x10\n"
-      "ld1 { v19.4s }, [x23], #0x10\n"
-      "ld1 { v22.4s }, [x22], #0x10\n"
-      "ld1 { v27.4s }, [x21], #0x10\n"
-      "ld1 { v30.4s }, [x20], #0x10\n"
-      "tbz x11, #1, 188f\n"
-      "ldr d16, [x9], #0x8\n"
-      "ldr d15, [x24], #0x8\n"
-      "mov x25, #0x38\n"
-      "ldr d24, [x23], #0x8\n"
-      "ldr d23, [x22], #0x8\n"
-      "ldr d6, [x21], #0x8\n"
-      "ldr d31, [x20], #0x8\n"
-      "tbz x11, #0, 195f\n"
-      "ld1 { v16.s }[2], [x9]\n"
-      "ld1 { v15.s }[2], [x24]\n"
-      "ld1 { v24.s }[2], [x23]\n"
-      "ld1 { v23.s }[2], [x22]\n"
-      "ld1 { v6.s }[2], [x21]\n"
-      "ld1 { v31.s }[2], [x20]\n"
-      "b 195f\n"
-      "188:"  // Height 6: Partial accumulate: partial_1_12
-      "mov x25, #0x30\n"
-      "tbz x11, #0, 195f\n"
-      "ldr s16, [x9, #0x0]\n"
-      "ldr s15, [x24, #0x0]\n"
-      "ldr s24, [x23, #0x0]\n"
-      "ldr s23, [x22, #0x0]\n"
-      "ldr s6, [x21, #0x0]\n"
-      "ldr s31, [x20, #0x0]\n"
-      "b 195f\n"
-      "189:"  // Height 6: Partial accumulate: partial_2_8
-      "tbz x11, #1, 190f\n"
-      "ldr d11, [x9], #0x8\n"
-      "ldr d14, [x24], #0x8\n"
-      "mov x25, #0x28\n"
-      "ldr d19, [x23], #0x8\n"
-      "ldr d22, [x22], #0x8\n"
-      "ldr d27, [x21], #0x8\n"
-      "ldr d30, [x20], #0x8\n"
-      "tbz x11, #0, 195f\n"
-      "ld1 { v11.s }[2], [x9]\n"
-      "ld1 { v14.s }[2], [x24]\n"
-      "ld1 { v19.s }[2], [x23]\n"
-      "ld1 { v22.s }[2], [x22]\n"
-      "ld1 { v27.s }[2], [x21]\n"
-      "ld1 { v30.s }[2], [x20]\n"
-      "b 195f\n"
-      "190:"  // Height 6: Partial accumulate: partial_1_8
-      "mov x25, #0x20\n"
-      "tbz x11, #0, 195f\n"
-      "ldr s11, [x9, #0x0]\n"
-      "ldr s14, [x24, #0x0]\n"
-      "ldr s19, [x23, #0x0]\n"
-      "ldr s22, [x22, #0x0]\n"
-      "ldr s27, [x21, #0x0]\n"
-      "ldr s30, [x20, #0x0]\n"
-      "b 195f\n"
-      "191:"  // Height 6: Partial accumulate: partial_4_0
-      "tbz x11, #2, 193f\n"
-      "ld1 { v9.4s }, [x9], #0x10\n"
-      "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v17.4s }, [x23], #0x10\n"
-      "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v25.4s }, [x21], #0x10\n"
-      "ld1 { v28.4s }, [x20], #0x10\n"
-      "tbz x11, #1, 192f\n"
-      "ldr d10, [x9], #0x8\n"
-      "ldr d13, [x24], #0x8\n"
-      "mov x25, #0x18\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d26, [x21], #0x8\n"
-      "ldr d29, [x20], #0x8\n"
-      "tbz x11, #0, 195f\n"
-      "ld1 { v10.s }[2], [x9]\n"
-      "ld1 { v13.s }[2], [x24]\n"
-      "ld1 { v18.s }[2], [x23]\n"
-      "ld1 { v21.s }[2], [x22]\n"
-      "ld1 { v26.s }[2], [x21]\n"
-      "ld1 { v29.s }[2], [x20]\n"
-      "b 195f\n"
-      "192:"  // Height 6: Partial accumulate: partial_1_4
-      "mov x25, #0x10\n"
-      "tbz x11, #0, 195f\n"
-      "ldr s10, [x9, #0x0]\n"
-      "ldr s13, [x24, #0x0]\n"
-      "ldr s18, [x23, #0x0]\n"
-      "ldr s21, [x22, #0x0]\n"
-      "ldr s26, [x21, #0x0]\n"
-      "ldr s29, [x20, #0x0]\n"
-      "b 195f\n"
-      "193:"  // Height 6: Partial accumulate: partial_2_0
-      "tbz x11, #1, 194f\n"
-      "ldr d9, [x9], #0x8\n"
-      "ldr d12, [x24], #0x8\n"
-      "mov x25, #0x8\n"
-      "ldr d17, [x23], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "ldr d25, [x21], #0x8\n"
-      "ldr d28, [x20], #0x8\n"
-      "tbz x11, #0, 195f\n"
-      "ld1 { v9.s }[2], [x9]\n"
-      "ld1 { v12.s }[2], [x24]\n"
-      "ld1 { v17.s }[2], [x23]\n"
-      "ld1 { v20.s }[2], [x22]\n"
-      "ld1 { v25.s }[2], [x21]\n"
-      "ld1 { v28.s }[2], [x20]\n"
-      "b 195f\n"
-      "194:"  // Height 6: Partial accumulate: partial_1_0
-      "ldr s9, [x9, #0x0]\n"
-      "ldr s12, [x24, #0x0]\n"
-      "mov x25, #0x0\n"
-      "ldr s17, [x23, #0x0]\n"
-      "ldr s20, [x22, #0x0]\n"
-      "ldr s25, [x21, #0x0]\n"
-      "ldr s28, [x20, #0x0]\n"
-      "195:"  // Height 6: Partial accumulate: Done
-      "sub x9, x9, x25\n"
-      "b 197f\n"
-      "196:"  // Height 6: full accumulate
-      "ldr q9, [x9, #0x0]\n"
-      "ldr q10, [x9, #0x10]\n"
-      "ldr q11, [x9, #0x20]\n"
-      "ldr q16, [x9, #0x30]\n"
-      "ldr q12, [x24, #0x0]\n"
-      "ldr q13, [x24, #0x10]\n"
-      "ldr q14, [x24, #0x20]\n"
-      "ldr q15, [x24, #0x30]\n"
-      "ldr q17, [x23, #0x0]\n"
-      "ldr q18, [x23, #0x10]\n"
-      "ldr q19, [x23, #0x20]\n"
-      "ldr q24, [x23, #0x30]\n"
-      "ldr q20, [x22, #0x0]\n"
-      "ldr q21, [x22, #0x10]\n"
-      "ldr q22, [x22, #0x20]\n"
-      "ldr q23, [x22, #0x30]\n"
-      "ldr q25, [x21, #0x0]\n"
-      "ldr q26, [x21, #0x10]\n"
-      "ldr q27, [x21, #0x20]\n"
-      "ldr q6, [x21, #0x30]\n"
-      "ldr q28, [x20, #0x0]\n"
-      "ldr q29, [x20, #0x10]\n"
-      "ldr q30, [x20, #0x20]\n"
-      "ldr q31, [x20, #0x30]\n"
-      "197:"  // Height 6: MMLA fixup
-      "zip1 v8.2d, v9.2d, v12.2d\n"
-      "zip2 v12.2d, v9.2d, v12.2d\n"
-      "zip1 v9.2d, v10.2d, v13.2d\n"
-      "zip2 v13.2d, v10.2d, v13.2d\n"
-      "zip1 v10.2d, v11.2d, v14.2d\n"
-      "zip2 v14.2d, v11.2d, v14.2d\n"
-      "zip1 v11.2d, v16.2d, v15.2d\n"
-      "zip2 v15.2d, v16.2d, v15.2d\n"
-      "zip1 v16.2d, v17.2d, v20.2d\n"
-      "zip2 v20.2d, v17.2d, v20.2d\n"
-      "zip1 v17.2d, v18.2d, v21.2d\n"
-      "zip2 v21.2d, v18.2d, v21.2d\n"
-      "zip1 v18.2d, v19.2d, v22.2d\n"
-      "zip2 v22.2d, v19.2d, v22.2d\n"
-      "zip1 v19.2d, v24.2d, v23.2d\n"
-      "zip2 v23.2d, v24.2d, v23.2d\n"
-      "zip1 v24.2d, v25.2d, v28.2d\n"
-      "zip2 v28.2d, v25.2d, v28.2d\n"
-      "zip1 v25.2d, v26.2d, v29.2d\n"
-      "zip2 v29.2d, v26.2d, v29.2d\n"
-      "zip1 v26.2d, v27.2d, v30.2d\n"
-      "zip2 v30.2d, v27.2d, v30.2d\n"
-      "zip1 v27.2d, v6.2d, v31.2d\n"
-      "zip2 v31.2d, v6.2d, v31.2d\n"
-      "b 199f\n"
-      "198:"  // Height 6: no accumulate
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "movi v12.4s, #0x0\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "movi v28.4s, #0x0\n"
-      "movi v29.4s, #0x0\n"
-      "movi v30.4s, #0x0\n"
-      "movi v31.4s, #0x0\n"
-      "199:"  // Height 6: setup done
-      "mov x28, #0x0\n"
-      "200:"  // Height 6: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 201f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "ldr x22, [x20, #0x20]\n"
-      "ldr x21, [x20, #0x28]\n"
-      "cbnz x28, 202f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "add x21, x21, x20\n"
-      "b 202f\n"
-      "201:"  // Height 6: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "add x21, x22, x21\n"
-      "202:"  // Height 6: input setup done
-      "cmp x27, #0x10\n"
-      "blt 205f\n"
-      "ldr q1, [x26, #0x0]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "cmp x27, #0x20\n"
-      "ldr q3, [x24, #0x0]\n"
-      "ldr q4, [x23, #0x0]\n"
-      "ldr q5, [x22, #0x0]\n"
-      "ldr q6, [x21, #0x0]\n"
-      "ldr q7, [x10, #0x0]\n"
-      "blt 204f\n"
-      "203:"  // Height 6: Multiply loop: Main loop head
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "trn1 v4.2d, v5.2d, v6.2d\n"
-      "trn2 v5.2d, v5.2d, v6.2d\n"
-      "ldr q6, [x10, #0x10]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x4e87ac08  // usmmla v8.4s, v0.16b, v7.16b\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x4e87ac50  // usmmla v16.4s, v2.16b, v7.16b\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      ".inst 0x4e87ac98  // usmmla v24.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x20]\n"
-      ".inst 0x4e86ac0c  // usmmla v12.4s, v0.16b, v6.16b\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4e86ac54  // usmmla v20.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e86ac9c  // usmmla v28.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      ".inst 0x4e87ac09  // usmmla v9.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac51  // usmmla v17.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac99  // usmmla v25.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x40]\n"
-      ".inst 0x4e86ac0d  // usmmla v13.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac55  // usmmla v21.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e86ac9d  // usmmla v29.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x50]\n"
-      ".inst 0x4e87ac0a  // usmmla v10.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac52  // usmmla v18.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac9a  // usmmla v26.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x60]\n"
-      ".inst 0x4e86ac0e  // usmmla v14.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac56  // usmmla v22.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e86ac9e  // usmmla v30.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x70]\n"
-      ".inst 0x4e87ac0b  // usmmla v11.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac53  // usmmla v19.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac9b  // usmmla v27.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x80]\n"
-      ".inst 0x4e86ac0f  // usmmla v15.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac57  // usmmla v23.4s, v2.16b, v6.16b\n"
-      "ldr q2, [x25, #0x0]\n"
-      ".inst 0x4e86ac9f  // usmmla v31.4s, v4.16b, v6.16b\n"
-      "ldr q0, [x10, #0x90]\n"
-      "ldr q4, [x23, #0x0]\n"
-      ".inst 0x4e87ac28  // usmmla v8.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e87ac70  // usmmla v16.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e87acb8  // usmmla v24.4s, v5.16b, v7.16b\n"
-      "ldr q6, [x10, #0xa0]\n"
-      ".inst 0x4e80ac2c  // usmmla v12.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac74  // usmmla v20.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acbc  // usmmla v28.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xb0]\n"
-      ".inst 0x4e86ac29  // usmmla v9.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac71  // usmmla v17.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86acb9  // usmmla v25.4s, v5.16b, v6.16b\n"
-      "ldr q6, [x10, #0xc0]\n"
-      ".inst 0x4e80ac2d  // usmmla v13.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac75  // usmmla v21.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acbd  // usmmla v29.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xd0]\n"
-      ".inst 0x4e86ac2a  // usmmla v10.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac72  // usmmla v18.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86acba  // usmmla v26.4s, v5.16b, v6.16b\n"
-      "ldr q6, [x10, #0xe0]\n"
-      ".inst 0x4e80ac2e  // usmmla v14.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac76  // usmmla v22.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acbe  // usmmla v30.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e86ac2b  // usmmla v11.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac73  // usmmla v19.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86acbb  // usmmla v27.4s, v5.16b, v6.16b\n"
-      "ldr q7, [x10, #0x0]\n"
-      ".inst 0x4e80ac2f  // usmmla v15.4s, v1.16b, v0.16b\n"
-      "ldr q1, [x26, #0x0]\n"
-      ".inst 0x4e80ac77  // usmmla v23.4s, v3.16b, v0.16b\n"
-      "ldr q3, [x24, #0x0]\n"
-      ".inst 0x4e80acbf  // usmmla v31.4s, v5.16b, v0.16b\n"
-      "ldr q5, [x22, #0x0]\n"
-      "ldr q6, [x21, #0x0]\n"
-      "bge 203b\n"
-      "204:"  // Height 6: Multiply loop: Single iteration only
-      "trn1 v0.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "trn1 v2.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "trn1 v4.2d, v5.2d, v6.2d\n"
-      "trn2 v5.2d, v5.2d, v6.2d\n"
-      "ldr q6, [x10, #0x10]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x4e87ac08  // usmmla v8.4s, v0.16b, v7.16b\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x4e87ac50  // usmmla v16.4s, v2.16b, v7.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      ".inst 0x4e87ac98  // usmmla v24.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x20]\n"
-      ".inst 0x4e86ac0c  // usmmla v12.4s, v0.16b, v6.16b\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      ".inst 0x4e86ac54  // usmmla v20.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e86ac9c  // usmmla v28.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
-      ".inst 0x4e87ac09  // usmmla v9.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac51  // usmmla v17.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac99  // usmmla v25.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x40]\n"
-      ".inst 0x4e86ac0d  // usmmla v13.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac55  // usmmla v21.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e86ac9d  // usmmla v29.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x50]\n"
-      ".inst 0x4e87ac0a  // usmmla v10.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac52  // usmmla v18.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac9a  // usmmla v26.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x60]\n"
-      ".inst 0x4e86ac0e  // usmmla v14.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac56  // usmmla v22.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e86ac9e  // usmmla v30.4s, v4.16b, v6.16b\n"
-      "ldr q6, [x10, #0x70]\n"
-      ".inst 0x4e87ac0b  // usmmla v11.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e87ac53  // usmmla v19.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e87ac9b  // usmmla v27.4s, v4.16b, v7.16b\n"
-      "ldr q7, [x10, #0x80]\n"
-      ".inst 0x4e86ac0f  // usmmla v15.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e86ac57  // usmmla v23.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e86ac9f  // usmmla v31.4s, v4.16b, v6.16b\n"
-      "ldr q2, [x10, #0x90]\n"
-      ".inst 0x4e87ac28  // usmmla v8.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e87ac70  // usmmla v16.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e87acb8  // usmmla v24.4s, v5.16b, v7.16b\n"
-      "ldr q0, [x10, #0xa0]\n"
-      ".inst 0x4e82ac2c  // usmmla v12.4s, v1.16b, v2.16b\n"
-      ".inst 0x4e82ac74  // usmmla v20.4s, v3.16b, v2.16b\n"
-      ".inst 0x4e82acbc  // usmmla v28.4s, v5.16b, v2.16b\n"
-      "ldr q2, [x10, #0xb0]\n"
-      ".inst 0x4e80ac29  // usmmla v9.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac71  // usmmla v17.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acb9  // usmmla v25.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xc0]\n"
-      ".inst 0x4e82ac2d  // usmmla v13.4s, v1.16b, v2.16b\n"
-      ".inst 0x4e82ac75  // usmmla v21.4s, v3.16b, v2.16b\n"
-      ".inst 0x4e82acbd  // usmmla v29.4s, v5.16b, v2.16b\n"
-      "ldr q2, [x10, #0xd0]\n"
-      ".inst 0x4e80ac2a  // usmmla v10.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac72  // usmmla v18.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acba  // usmmla v26.4s, v5.16b, v0.16b\n"
-      "ldr q0, [x10, #0xe0]\n"
-      ".inst 0x4e82ac2e  // usmmla v14.4s, v1.16b, v2.16b\n"
-      ".inst 0x4e82ac76  // usmmla v22.4s, v3.16b, v2.16b\n"
-      ".inst 0x4e82acbe  // usmmla v30.4s, v5.16b, v2.16b\n"
-      "ldr q6, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
-      ".inst 0x4e80ac2b  // usmmla v11.4s, v1.16b, v0.16b\n"
-      ".inst 0x4e80ac73  // usmmla v19.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80acbb  // usmmla v27.4s, v5.16b, v0.16b\n"
-      ".inst 0x4e86ac2f  // usmmla v15.4s, v1.16b, v6.16b\n"
-      ".inst 0x4e86ac77  // usmmla v23.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86acbf  // usmmla v31.4s, v5.16b, v6.16b\n"
-      "205:"  // Height 6: Multiply loop: Main loop skip
-      "cbz x27, 212f\n"
-      "cmp x27, #0x8\n"
-      "blt 207f\n"
-      "206:"  // Height 6: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
-      "sub x27, x27, #0x8\n"
-      "ldr d5, [x24], #0x8\n"
-      "ldr d3, [x23], #0x8\n"
-      "cmp x27, #0x8\n"
-      "ldr d2, [x22], #0x8\n"
-      "ldr d0, [x21], #0x8\n"
-      "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v5.2d, v3.2d\n"
-      "trn1 v2.2d, v2.2d, v0.2d\n"
-      "ldr q0, [x10, #0x10]\n"
-      ".inst 0x4e81ac88  // usmmla v8.4s, v4.16b, v1.16b\n"
-      ".inst 0x4e81ac70  // usmmla v16.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac58  // usmmla v24.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x20]\n"
-      ".inst 0x4e80ac8c  // usmmla v12.4s, v4.16b, v0.16b\n"
-      ".inst 0x4e80ac74  // usmmla v20.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5c  // usmmla v28.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x30]\n"
-      ".inst 0x4e81ac89  // usmmla v9.4s, v4.16b, v1.16b\n"
-      ".inst 0x4e81ac71  // usmmla v17.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac59  // usmmla v25.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x40]\n"
-      ".inst 0x4e80ac8d  // usmmla v13.4s, v4.16b, v0.16b\n"
-      ".inst 0x4e80ac75  // usmmla v21.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5d  // usmmla v29.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x50]\n"
-      ".inst 0x4e81ac8a  // usmmla v10.4s, v4.16b, v1.16b\n"
-      ".inst 0x4e81ac72  // usmmla v18.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac5a  // usmmla v26.4s, v2.16b, v1.16b\n"
-      "ldr q6, [x10, #0x60]\n"
-      ".inst 0x4e80ac8e  // usmmla v14.4s, v4.16b, v0.16b\n"
-      ".inst 0x4e80ac76  // usmmla v22.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5e  // usmmla v30.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e86ac8b  // usmmla v11.4s, v4.16b, v6.16b\n"
-      ".inst 0x4e86ac73  // usmmla v19.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86ac5b  // usmmla v27.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e80ac8f  // usmmla v15.4s, v4.16b, v0.16b\n"
-      ".inst 0x4e80ac77  // usmmla v23.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5f  // usmmla v31.4s, v2.16b, v0.16b\n"
-      "bge 206b\n"
-      "207:"  // Height 6: Multiply loop: Skip odd blocks
-      "cbz x27, 212f\n"
-      "tbz x27, #2, 209f\n"
-      "ldr s1, [x26], #0x4\n"
-      "ldr s2, [x25], #0x4\n"
-      "ldr s3, [x24], #0x4\n"
-      "ldr s4, [x23], #0x4\n"
-      "ldr s5, [x22], #0x4\n"
-      "ldr s6, [x21], #0x4\n"
-      "tbz x27, #1, 208f\n"
-      "ld1 { v1.h }[2], [x26], #0x2\n"
-      "ld1 { v2.h }[2], [x25], #0x2\n"
-      "ld1 { v3.h }[2], [x24], #0x2\n"
-      "ld1 { v4.h }[2], [x23], #0x2\n"
-      "ld1 { v5.h }[2], [x22], #0x2\n"
-      "ld1 { v6.h }[2], [x21], #0x2\n"
-      "tbz x27, #0, 211f\n"
-      "ld1 { v1.b }[6], [x26]\n"
-      "ld1 { v2.b }[6], [x25]\n"
-      "ld1 { v3.b }[6], [x24]\n"
-      "ld1 { v4.b }[6], [x23]\n"
-      "ld1 { v5.b }[6], [x22]\n"
-      "ld1 { v6.b }[6], [x21]\n"
-      "b 211f\n"
-      "208:"  // Height 6: Multiply loop: Ragged operand read: partial_1_4
-      "tbz x27, #0, 211f\n"
-      "ld1 { v1.b }[4], [x26]\n"
-      "ld1 { v2.b }[4], [x25]\n"
-      "ld1 { v3.b }[4], [x24]\n"
-      "ld1 { v4.b }[4], [x23]\n"
-      "ld1 { v5.b }[4], [x22]\n"
-      "ld1 { v6.b }[4], [x21]\n"
-      "b 211f\n"
-      "209:"  // Height 6: Multiply loop: Ragged operand read: partial_2_0
-      "tbz x27, #1, 210f\n"
-      "ldr h1, [x26], #0x2\n"
-      "ldr h2, [x25], #0x2\n"
-      "ldr h3, [x24], #0x2\n"
-      "ldr h4, [x23], #0x2\n"
-      "ldr h5, [x22], #0x2\n"
-      "ldr h6, [x21], #0x2\n"
-      "tbz x27, #0, 211f\n"
-      "ld1 { v1.b }[2], [x26]\n"
-      "ld1 { v2.b }[2], [x25]\n"
-      "ld1 { v3.b }[2], [x24]\n"
-      "ld1 { v4.b }[2], [x23]\n"
-      "ld1 { v5.b }[2], [x22]\n"
-      "ld1 { v6.b }[2], [x21]\n"
-      "b 211f\n"
-      "210:"  // Height 6: Multiply loop: Ragged operand read: partial_1_0
-      "ldr b1, [x26, #0x0]\n"
-      "ldr b2, [x25, #0x0]\n"
-      "ldr b3, [x24, #0x0]\n"
-      "ldr b4, [x23, #0x0]\n"
-      "ldr b5, [x22, #0x0]\n"
-      "ldr b6, [x21, #0x0]\n"
-      "211:"  // Height 6: Multiply loop: Ragged operand read: Done
-      "ldr q0, [x10, #0x0]\n"
-      "trn1 v7.2d, v1.2d, v2.2d\n"
-      "trn1 v3.2d, v3.2d, v4.2d\n"
-      "trn1 v2.2d, v5.2d, v6.2d\n"
-      "ldr q1, [x10, #0x10]\n"
-      ".inst 0x4e80ace8  // usmmla v8.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e80ac70  // usmmla v16.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac58  // usmmla v24.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x20]\n"
-      ".inst 0x4e81acec  // usmmla v12.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e81ac74  // usmmla v20.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac5c  // usmmla v28.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x30]\n"
-      ".inst 0x4e80ace9  // usmmla v9.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e80ac71  // usmmla v17.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac59  // usmmla v25.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x40]\n"
-      ".inst 0x4e81aced  // usmmla v13.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e81ac75  // usmmla v21.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac5d  // usmmla v29.4s, v2.16b, v1.16b\n"
-      "ldr q1, [x10, #0x50]\n"
-      ".inst 0x4e80acea  // usmmla v10.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e80ac72  // usmmla v18.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5a  // usmmla v26.4s, v2.16b, v0.16b\n"
-      "ldr q0, [x10, #0x60]\n"
-      ".inst 0x4e81acee  // usmmla v14.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e81ac76  // usmmla v22.4s, v3.16b, v1.16b\n"
-      ".inst 0x4e81ac5e  // usmmla v30.4s, v2.16b, v1.16b\n"
-      "ldr q6, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
-      ".inst 0x4e80aceb  // usmmla v11.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e80ac73  // usmmla v19.4s, v3.16b, v0.16b\n"
-      ".inst 0x4e80ac5b  // usmmla v27.4s, v2.16b, v0.16b\n"
-      ".inst 0x4e86acef  // usmmla v15.4s, v7.16b, v6.16b\n"
-      ".inst 0x4e86ac77  // usmmla v23.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e86ac5f  // usmmla v31.4s, v2.16b, v6.16b\n"
-      "212:"  // Height 6: Multiply loop: No odd multiplies
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 200b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "uzp1 v7.2d, v8.2d, v12.2d\n"
-      "uzp2 v8.2d, v8.2d, v12.2d\n"
-      "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
-      "uzp1 v13.2d, v10.2d, v14.2d\n"
-      "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "uzp1 v15.2d, v16.2d, v20.2d\n"
-      "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "add x20, x21, x20, LSL #2\n"
-      "uzp1 v20.2d, v17.2d, v21.2d\n"
-      "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "uzp1 v21.2d, v18.2d, v22.2d\n"
-      "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "prfm pstl1keep, [x21, #0x0]\n"
-      "prfm pstl1keep, [x20, #0x0]\n"
-      "uzp1 v22.2d, v19.2d, v23.2d\n"
-      "uzp2 v19.2d, v19.2d, v23.2d\n"
-      "uzp1 v23.2d, v24.2d, v28.2d\n"
-      "uzp2 v24.2d, v24.2d, v28.2d\n"
-      "uzp1 v28.2d, v25.2d, v29.2d\n"
-      "uzp2 v25.2d, v25.2d, v29.2d\n"
-      "uzp1 v29.2d, v26.2d, v30.2d\n"
-      "uzp2 v26.2d, v26.2d, v30.2d\n"
-      "uzp1 v30.2d, v27.2d, v31.2d\n"
-      "uzp2 v27.2d, v27.2d, v31.2d\n"
-      "bge 221f\n"
-      "tbz x11, #3, 216f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v12.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v9.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x23], #0x10\n"
-      "st1 { v20.4s }, [x23], #0x10\n"
-      "st1 { v16.4s }, [x22], #0x10\n"
-      "st1 { v17.4s }, [x22], #0x10\n"
-      "st1 { v23.4s }, [x21], #0x10\n"
-      "st1 { v28.4s }, [x21], #0x10\n"
-      "st1 { v24.4s }, [x20], #0x10\n"
-      "st1 { v25.4s }, [x20], #0x10\n"
-      "tbz x11, #2, 214f\n"
-      "st1 { v13.4s }, [x9], #0x10\n"
-      "st1 { v10.4s }, [x24], #0x10\n"
-      "st1 { v21.4s }, [x23], #0x10\n"
-      "st1 { v18.4s }, [x22], #0x10\n"
-      "st1 { v29.4s }, [x21], #0x10\n"
-      "st1 { v26.4s }, [x20], #0x10\n"
-      "tbz x11, #1, 213f\n"
-      "str d14, [x9], #0x8\n"
-      "str d11, [x24], #0x8\n"
-      "str d22, [x23], #0x8\n"
-      "str d19, [x22], #0x8\n"
-      "str d30, [x21], #0x8\n"
-      "str d27, [x20], #0x8\n"
-      "tbz x11, #0, 220f\n"
-      "st1 { v14.s }[2], [x9]\n"
-      "st1 { v11.s }[2], [x24]\n"
-      "st1 { v22.s }[2], [x23]\n"
-      "st1 { v19.s }[2], [x22]\n"
-      "st1 { v30.s }[2], [x21]\n"
-      "st1 { v27.s }[2], [x20]\n"
-      "b 220f\n"
-      "213:"  // Height 6: Partial direct writeback: partial_1_12
-      "tbz x11, #0, 220f\n"
-      "str s14, [x9, #0x0]\n"
-      "str s11, [x24, #0x0]\n"
-      "str s22, [x23, #0x0]\n"
-      "str s19, [x22, #0x0]\n"
-      "str s30, [x21, #0x0]\n"
-      "str s27, [x20, #0x0]\n"
-      "b 220f\n"
-      "214:"  // Height 6: Partial direct writeback: partial_2_8
-      "tbz x11, #1, 215f\n"
-      "str d13, [x9], #0x8\n"
-      "str d10, [x24], #0x8\n"
-      "str d21, [x23], #0x8\n"
-      "str d18, [x22], #0x8\n"
-      "str d29, [x21], #0x8\n"
-      "str d26, [x20], #0x8\n"
-      "tbz x11, #0, 220f\n"
-      "st1 { v13.s }[2], [x9]\n"
-      "st1 { v10.s }[2], [x24]\n"
-      "st1 { v21.s }[2], [x23]\n"
-      "st1 { v18.s }[2], [x22]\n"
-      "st1 { v29.s }[2], [x21]\n"
-      "st1 { v26.s }[2], [x20]\n"
-      "b 220f\n"
-      "215:"  // Height 6: Partial direct writeback: partial_1_8
-      "tbz x11, #0, 220f\n"
-      "str s13, [x9, #0x0]\n"
-      "str s10, [x24, #0x0]\n"
-      "str s21, [x23, #0x0]\n"
-      "str s18, [x22, #0x0]\n"
-      "str s29, [x21, #0x0]\n"
-      "str s26, [x20, #0x0]\n"
-      "b 220f\n"
-      "216:"  // Height 6: Partial direct writeback: partial_4_0
-      "tbz x11, #2, 218f\n"
-      "st1 { v7.4s }, [x9], #0x10\n"
-      "st1 { v8.4s }, [x24], #0x10\n"
-      "st1 { v15.4s }, [x23], #0x10\n"
-      "st1 { v16.4s }, [x22], #0x10\n"
-      "st1 { v23.4s }, [x21], #0x10\n"
-      "st1 { v24.4s }, [x20], #0x10\n"
-      "tbz x11, #1, 217f\n"
-      "str d12, [x9], #0x8\n"
-      "str d9, [x24], #0x8\n"
-      "str d20, [x23], #0x8\n"
-      "str d17, [x22], #0x8\n"
-      "str d28, [x21], #0x8\n"
-      "str d25, [x20], #0x8\n"
-      "tbz x11, #0, 220f\n"
-      "st1 { v12.s }[2], [x9]\n"
-      "st1 { v9.s }[2], [x24]\n"
-      "st1 { v20.s }[2], [x23]\n"
-      "st1 { v17.s }[2], [x22]\n"
-      "st1 { v28.s }[2], [x21]\n"
-      "st1 { v25.s }[2], [x20]\n"
-      "b 220f\n"
-      "217:"  // Height 6: Partial direct writeback: partial_1_4
-      "tbz x11, #0, 220f\n"
-      "str s12, [x9, #0x0]\n"
-      "str s9, [x24, #0x0]\n"
-      "str s20, [x23, #0x0]\n"
-      "str s17, [x22, #0x0]\n"
-      "str s28, [x21, #0x0]\n"
-      "str s25, [x20, #0x0]\n"
-      "b 220f\n"
-      "218:"  // Height 6: Partial direct writeback: partial_2_0
-      "tbz x11, #1, 219f\n"
-      "str d7, [x9], #0x8\n"
-      "str d8, [x24], #0x8\n"
-      "str d15, [x23], #0x8\n"
-      "str d16, [x22], #0x8\n"
-      "str d23, [x21], #0x8\n"
-      "str d24, [x20], #0x8\n"
-      "tbz x11, #0, 220f\n"
-      "st1 { v7.s }[2], [x9]\n"
-      "st1 { v8.s }[2], [x24]\n"
-      "st1 { v15.s }[2], [x23]\n"
-      "st1 { v16.s }[2], [x22]\n"
-      "st1 { v23.s }[2], [x21]\n"
-      "st1 { v24.s }[2], [x20]\n"
-      "b 220f\n"
-      "219:"  // Height 6: Partial direct writeback: partial_1_0
-      "str s7, [x9, #0x0]\n"
-      "str s8, [x24, #0x0]\n"
-      "str s15, [x23, #0x0]\n"
-      "str s16, [x22, #0x0]\n"
-      "str s23, [x21, #0x0]\n"
-      "str s24, [x20, #0x0]\n"
-      "220:"  // Height 6: Partial direct writeback: Done
-      "b 222f\n"
-      "221:"  // Height 6: Full writeback
-      "str q7, [x9, #0x0]\n"
-      "str q12, [x9, #0x10]\n"
-      "str q13, [x9, #0x20]\n"
-      "str q14, [x9, #0x30]\n"
-      "add x9, x9, #0x40\n"
-      "str q8, [x24, #0x0]\n"
-      "str q9, [x24, #0x10]\n"
-      "str q10, [x24, #0x20]\n"
-      "str q11, [x24, #0x30]\n"
-      "str q15, [x23, #0x0]\n"
-      "str q20, [x23, #0x10]\n"
-      "str q21, [x23, #0x20]\n"
-      "str q22, [x23, #0x30]\n"
-      "str q16, [x22, #0x0]\n"
-      "str q17, [x22, #0x10]\n"
-      "str q18, [x22, #0x20]\n"
-      "str q19, [x22, #0x30]\n"
-      "str q23, [x21, #0x0]\n"
-      "str q28, [x21, #0x10]\n"
-      "str q29, [x21, #0x20]\n"
-      "str q30, [x21, #0x30]\n"
-      "str q24, [x20, #0x0]\n"
-      "str q25, [x20, #0x10]\n"
-      "str q26, [x20, #0x20]\n"
-      "str q27, [x20, #0x30]\n"
-      "222:"  // Height 6: Writeback done
-      "subs x11, x11, #0x10\n"
-      "bgt 187b\n"
-      "subs %x[M], %x[M], #0x6\n"
-      "beq 224f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "tbz %x[flags], #3, 223f\n"
-      "add x21, x21, #0x6\n"
-      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "b 1b\n"
-      "223:"  // Update direct input
-      "mov x20, #0x6\n"
-      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
-      "b 1b\n"
-      "224:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
-    );
-}
-
-} // namespace arm_gemm
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16.hpp
index 4bb4c31577..16d241ff02 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 4> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp
index 4ec75191b9..f12269be58 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp
@@ -44,18 +44,18 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -87,72 +87,72 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "bgt 69f\n"
       "beq 35f\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "tbz %x[flags], #0, 12f\n"
       "cmp x8, #0x10\n"
       "bge 11f\n"
       "tbz x8, #3, 6f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "tbz x8, #2, 4f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "tbz x8, #1, 3f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "tbz x8, #0, 10f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "b 10f\n"
       "3:"  // Height 1: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 10f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "b 10f\n"
       "4:"  // Height 1: Partial accumulate: partial_2_8
       "tbz x8, #1, 5f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "tbz x8, #0, 10f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "b 10f\n"
       "5:"  // Height 1: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 10f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "b 10f\n"
       "6:"  // Height 1: Partial accumulate: partial_4_0
       "tbz x8, #2, 8f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "tbz x8, #1, 7f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "tbz x8, #0, 10f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "b 10f\n"
       "7:"  // Height 1: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 10f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "b 10f\n"
       "8:"  // Height 1: Partial accumulate: partial_2_0
       "tbz x8, #1, 9f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "tbz x8, #0, 10f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "b 10f\n"
       "9:"  // Height 1: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "10:"  // Height 1: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 13f\n"
       "11:"  // Height 1: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "b 13f\n"
       "12:"  // Height 1: no accumulate
       "movi v8.4s, #0x0\n"
@@ -163,8 +163,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "14:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 15f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -180,118 +180,118 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "blt 19f\n"
       "ldr q0, [x13, #0x0]\n"
       "cmp x14, #0x20\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 18f\n"
       "17:"  // Height 1: Multiply loop: Main loop head
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr d17, [x17, #0x20]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr d17, [x16, #0x20]\n"
+      "ldr x20, [x16, #0x28]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr d16, [x17, #0x30]\n"
-      "add x13, x13, #0x10\n"
-      "ldr x20, [x17, #0x38]\n"
-      "sub x14, x14, #0x10\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x48]\n"
-      "ldr x22, [x13, #0x8]\n"
-      "cmp x14, #0x20\n"
+      "ldr d16, [x16, #0x30]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0x38]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr d17, [x17, #0x40]\n"
+      "ldr d17, [x16, #0x40]\n"
+      "ldr x20, [x16, #0x48]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr d16, [x17, #0x50]\n"
-      "ldr x20, [x17, #0x58]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x68]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr d16, [x16, #0x50]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0x58]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x6fa0e228  // udot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr d17, [x17, #0x60]\n"
+      "ldr d17, [x16, #0x60]\n"
+      "ldr x20, [x16, #0x68]\n"
       ".inst 0x6fa0e209  // udot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr d16, [x17, #0x70]\n"
-      "ldr x20, [x17, #0x78]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x88]\n"
+      "ldr d16, [x16, #0x70]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0x78]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x6fa0e22a  // udot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr d17, [x17, #0x80]\n"
+      "ldr d17, [x16, #0x80]\n"
+      "ldr x20, [x16, #0x88]\n"
       ".inst 0x6fa0e20b  // udot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr d16, [x17, #0x90]\n"
-      "ldr x20, [x17, #0x98]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xa8]\n"
+      "ldr d16, [x16, #0x90]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0x98]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x6f80ea28  // udot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr d17, [x17, #0xa0]\n"
+      "ldr d17, [x16, #0xa0]\n"
+      "ldr x20, [x16, #0xa8]\n"
       ".inst 0x6f80ea09  // udot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr d16, [x17, #0xb0]\n"
-      "ldr x20, [x17, #0xb8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xc8]\n"
+      "ldr d16, [x16, #0xb0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0xb8]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x6f80ea2a  // udot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr d17, [x17, #0xc0]\n"
+      "ldr d17, [x16, #0xc0]\n"
+      "ldr x20, [x16, #0xc8]\n"
       ".inst 0x6f80ea0b  // udot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr d16, [x17, #0xd0]\n"
-      "ldr x20, [x17, #0xd8]\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0xe8]\n"
+      "ldr d16, [x16, #0xd0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0xd8]\n"
       "mov v16.d[1], x20\n"
       ".inst 0x6fa0ea28  // udot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr d17, [x17, #0xe0]\n"
+      "ldr d17, [x16, #0xe0]\n"
+      "ldr x20, [x16, #0xe8]\n"
       ".inst 0x6fa0ea09  // udot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr d16, [x17, #0xf0]\n"
-      "ldr x20, [x17, #0xf8]\n"
-      "add x17, x17, #0x100\n"
-      "mov v17.d[1], x21\n"
-      "ldr x21, [x17, #0x8]\n"
+      "ldr d16, [x16, #0xf0]\n"
+      "mov v17.d[1], x20\n"
+      "ldr x20, [x16, #0xf8]\n"
       "mov v16.d[1], x20\n"
+      "add x13, x13, #0x10\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa0ea2a  // udot v10.4s, v17.16b, v0.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
+      "ldr x20, [x16, #0x8]\n"
       ".inst 0x6fa0ea0b  // udot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
-      "ldr x20, [x17, #0x18]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x22\n"
+      "sub x14, x14, #0x10\n"
+      "ldr d7, [x16, #0x10]\n"
+      "cmp x14, #0x20\n"
+      "ldr x21, [x13, #0x8]\n"
+      "mov v6.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
+      "mov v0.d[1], x21\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "bge 17b\n"
       "18:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q17, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
-      "add x13, x13, #0x10\n"
-      "sub x14, x14, #0x10\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x17, #0x40]\n"
+      "ldr q17, [x16, #0x40]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x17, #0x50]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr q16, [x16, #0x50]\n"
       ".inst 0x6fa0e228  // udot v8.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x17, #0x60]\n"
+      "ldr q17, [x16, #0x60]\n"
       ".inst 0x6fa0e209  // udot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x17, #0x70]\n"
+      "ldr q16, [x16, #0x70]\n"
       ".inst 0x6fa0e22a  // udot v10.4s, v17.16b, v0.4b[1]\n"
-      "ldr q17, [x17, #0x80]\n"
+      "ldr q17, [x16, #0x80]\n"
       ".inst 0x6fa0e20b  // udot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr q16, [x17, #0x90]\n"
+      "ldr q16, [x16, #0x90]\n"
       ".inst 0x6f80ea28  // udot v8.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x17, #0xa0]\n"
+      "ldr q17, [x16, #0xa0]\n"
       ".inst 0x6f80ea09  // udot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x17, #0xb0]\n"
+      "ldr q16, [x16, #0xb0]\n"
       ".inst 0x6f80ea2a  // udot v10.4s, v17.16b, v0.4b[2]\n"
-      "ldr q17, [x17, #0xc0]\n"
+      "ldr q17, [x16, #0xc0]\n"
       ".inst 0x6f80ea0b  // udot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr q16, [x17, #0xd0]\n"
+      "ldr q16, [x16, #0xd0]\n"
       ".inst 0x6fa0ea28  // udot v8.4s, v17.16b, v0.4b[3]\n"
-      "ldr q17, [x17, #0xe0]\n"
+      "ldr q17, [x16, #0xe0]\n"
       ".inst 0x6fa0ea09  // udot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr q16, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "ldr q16, [x16, #0xf0]\n"
+      "add x13, x13, #0x10\n"
+      "sub x14, x14, #0x10\n"
       ".inst 0x6fa0ea2a  // udot v10.4s, v17.16b, v0.4b[3]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x6fa0ea0b  // udot v11.4s, v16.16b, v0.4b[3]\n"
+      "add x16, x16, #0x100\n"
       "19:"  // Height 1: Multiply loop: Main loop skip
       "cbz x14, 24f\n"
       "cmp x14, #0x4\n"
@@ -299,16 +299,16 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "20:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x13], #0x4\n"
       "sub x14, x14, #0x4\n"
-      "ldr q17, [x17, #0x0]\n"
-      "cmp x14, #0x4\n"
-      "ldr q16, [x17, #0x10]\n"
-      ".inst 0x6f92e228  // udot v8.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q16, [x16, #0x0]\n"
+      ".inst 0x6f92e208  // udot v8.4s, v16.16b, v18.4b[0]\n"
+      "ldr q16, [x16, #0x10]\n"
       ".inst 0x6f92e209  // udot v9.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
-      "add x17, x17, #0x40\n"
+      "ldr q17, [x16, #0x20]\n"
+      "cmp x14, #0x4\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x6f92e22a  // udot v10.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x6f92e20b  // udot v11.4s, v16.16b, v18.4b[0]\n"
+      "add x16, x16, #0x40\n"
       "bge 20b\n"
       "21:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x14, 24f\n"
@@ -320,165 +320,165 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "22:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr b0, [x13, #0x0]\n"
       "23:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x17, #0x0]\n"
-      "ldr q16, [x17, #0x10]\n"
-      ".inst 0x6f80e228  // udot v8.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q16, [x16, #0x0]\n"
+      ".inst 0x6f80e208  // udot v8.4s, v16.16b, v0.4b[0]\n"
+      "ldr q16, [x16, #0x10]\n"
       ".inst 0x6f80e209  // udot v9.4s, v16.16b, v0.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
-      "add x17, x17, #0x40\n"
-      ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
+      "ldr q16, [x16, #0x20]\n"
+      ".inst 0x6f80e20a  // udot v10.4s, v16.16b, v0.4b[0]\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
+      "add x16, x16, #0x40\n"
       "24:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x15, x15, #0x1\n"
       "cmp x15, x20\n"
       "bne 14b\n"
       "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
       "bge 33f\n"
       "tbz x8, #3, 28f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "tbz x8, #2, 26f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "tbz x8, #1, 25f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "tbz x8, #0, 32f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "b 32f\n"
       "25:"  // Height 1: Partial direct writeback: partial_1_12
       "tbz x8, #0, 32f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "b 32f\n"
       "26:"  // Height 1: Partial direct writeback: partial_2_8
       "tbz x8, #1, 27f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "tbz x8, #0, 32f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "b 32f\n"
       "27:"  // Height 1: Partial direct writeback: partial_1_8
       "tbz x8, #0, 32f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "b 32f\n"
       "28:"  // Height 1: Partial direct writeback: partial_4_0
       "tbz x8, #2, 30f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "tbz x8, #1, 29f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "tbz x8, #0, 32f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "b 32f\n"
       "29:"  // Height 1: Partial direct writeback: partial_1_4
       "tbz x8, #0, 32f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "b 32f\n"
       "30:"  // Height 1: Partial direct writeback: partial_2_0
       "tbz x8, #1, 31f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "tbz x8, #0, 32f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "b 32f\n"
       "31:"  // Height 1: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "32:"  // Height 1: Partial direct writeback: Done
       "b 34f\n"
       "33:"  // Height 1: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "34:"  // Height 1: Writeback done
       "subs x8, x8, #0x10\n"
       "bgt 2b\n"
       "b 206f\n"
       "35:"  // Height 2
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "36:"  // Height 2: Column loop
       "tbz %x[flags], #0, 46f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "bge 45f\n"
       "tbz x8, #3, 40f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "tbz x8, #2, 38f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "tbz x8, #1, 37f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "tbz x8, #0, 44f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "b 44f\n"
       "37:"  // Height 2: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 44f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "b 44f\n"
       "38:"  // Height 2: Partial accumulate: partial_2_8
       "tbz x8, #1, 39f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "tbz x8, #0, 44f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "b 44f\n"
       "39:"  // Height 2: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 44f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "b 44f\n"
       "40:"  // Height 2: Partial accumulate: partial_4_0
       "tbz x8, #2, 42f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "tbz x8, #1, 41f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "tbz x8, #0, 44f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "b 44f\n"
       "41:"  // Height 2: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 44f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "b 44f\n"
       "42:"  // Height 2: Partial accumulate: partial_2_0
       "tbz x8, #1, 43f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "tbz x8, #0, 44f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "b 44f\n"
       "43:"  // Height 2: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "44:"  // Height 2: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 47f\n"
       "45:"  // Height 2: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -497,8 +497,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "48:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 49f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -518,154 +518,154 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr q0, [x13, #0x0]\n"
       "cmp x14, #0x20\n"
       "ldr q1, [x12, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 52f\n"
       "51:"  // Height 2: Multiply loop: Main loop head
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x20, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x6f81e0cc  // udot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr d17, [x17, #0x20]\n"
+      "ldr d17, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr d16, [x17, #0x30]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x48]\n"
-      "add x13, x13, #0x10\n"
-      "add x12, x12, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x16, #0x30]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x6f81e22e  // udot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr d17, [x17, #0x40]\n"
+      "ldr d17, [x16, #0x40]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x58]\n"
+      "ldr x20, [x16, #0x48]\n"
       ".inst 0x6f81e20f  // udot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr d16, [x17, #0x50]\n"
+      "ldr d16, [x16, #0x50]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x68]\n"
-      "ldr x23, [x13, #0x8]\n"
-      "sub x14, x14, #0x10\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x16, #0x58]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x6fa0e228  // udot v8.4s, v17.16b, v0.4b[1]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x6fa1e22c  // udot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr d17, [x17, #0x60]\n"
+      "ldr d17, [x16, #0x60]\n"
       ".inst 0x6fa0e209  // udot v9.4s, v16.16b, v0.4b[1]\n"
-      "ldr x21, [x17, #0x78]\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x6fa1e20d  // udot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr d16, [x17, #0x70]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0x88]\n"
-      "ldr x22, [x12, #0x8]\n"
-      "cmp x14, #0x20\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x16, #0x70]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x6fa0e22a  // udot v10.4s, v17.16b, v0.4b[1]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x6fa1e22e  // udot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr d17, [x17, #0x80]\n"
+      "ldr d17, [x16, #0x80]\n"
       ".inst 0x6fa0e20b  // udot v11.4s, v16.16b, v0.4b[1]\n"
-      "ldr x21, [x17, #0x98]\n"
+      "ldr x20, [x16, #0x88]\n"
       ".inst 0x6fa1e20f  // udot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr d16, [x17, #0x90]\n"
+      "ldr d16, [x16, #0x90]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xa8]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x16, #0x98]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x6f80ea28  // udot v8.4s, v17.16b, v0.4b[2]\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x6f81ea2c  // udot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr d17, [x17, #0xa0]\n"
+      "ldr d17, [x16, #0xa0]\n"
       ".inst 0x6f80ea09  // udot v9.4s, v16.16b, v0.4b[2]\n"
-      "ldr x21, [x17, #0xb8]\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x6f81ea0d  // udot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr d16, [x17, #0xb0]\n"
-      "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xc8]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x16, #0xb0]\n"
+      "mov v17.d[1], x21\n"
       ".inst 0x6f80ea2a  // udot v10.4s, v17.16b, v0.4b[2]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x6f81ea2e  // udot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr d17, [x17, #0xc0]\n"
+      "ldr d17, [x16, #0xc0]\n"
       ".inst 0x6f80ea0b  // udot v11.4s, v16.16b, v0.4b[2]\n"
-      "ldr x21, [x17, #0xd8]\n"
+      "ldr x20, [x16, #0xc8]\n"
       ".inst 0x6f81ea0f  // udot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr d16, [x17, #0xd0]\n"
+      "ldr d16, [x16, #0xd0]\n"
       "mov v17.d[1], x20\n"
-      "ldr x20, [x17, #0xe8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr x20, [x16, #0xd8]\n"
+      "mov v16.d[1], x20\n"
       ".inst 0x6fa0ea28  // udot v8.4s, v17.16b, v0.4b[3]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x6fa1ea2c  // udot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr d17, [x17, #0xe0]\n"
+      "ldr d17, [x16, #0xe0]\n"
       ".inst 0x6fa0ea09  // udot v9.4s, v16.16b, v0.4b[3]\n"
-      "ldr x21, [x17, #0xf8]\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x6fa1ea0d  // udot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr d16, [x17, #0xf0]\n"
-      "mov v17.d[1], x20\n"
-      "add x17, x17, #0x100\n"
-      "ldr x20, [x17, #0x8]\n"
-      "mov v16.d[1], x21\n"
+      "ldr d16, [x16, #0xf0]\n"
+      "mov v17.d[1], x21\n"
+      "add x13, x13, #0x10\n"
+      "mov v16.d[1], x20\n"
+      "add x12, x12, #0x10\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa0ea2a  // udot v10.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x6fa1ea2e  // udot v14.4s, v17.16b, v1.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
+      "ldr x21, [x16, #0x8]\n"
       ".inst 0x6fa0ea0b  // udot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x6fa1ea0f  // udot v15.4s, v16.16b, v1.4b[3]\n"
       "ldr d1, [x12, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x20\n"
-      "ldr x20, [x17, #0x18]\n"
-      "mov v0.d[1], x23\n"
-      "mov v1.d[1], x22\n"
+      "sub x14, x14, #0x10\n"
+      "ldr d7, [x16, #0x10]\n"
+      "cmp x14, #0x20\n"
+      "ldr x20, [x13, #0x8]\n"
+      "mov v6.d[1], x21\n"
+      "ldr x21, [x12, #0x8]\n"
+      "mov v0.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
+      "mov v1.d[1], x21\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       "mov v7.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       "bge 51b\n"
       "52:"  // Height 2: Multiply loop: Single iteration only
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
       "add x13, x13, #0x10\n"
       ".inst 0x6f81e0cc  // udot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q17, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
       "sub x14, x14, #0x10\n"
       ".inst 0x6f81e22e  // udot v14.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x17, #0x40]\n"
+      "ldr q17, [x16, #0x40]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x6f81e20f  // udot v15.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x17, #0x50]\n"
+      "ldr q16, [x16, #0x50]\n"
       ".inst 0x6fa0e228  // udot v8.4s, v17.16b, v0.4b[1]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x6fa1e22c  // udot v12.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x17, #0x60]\n"
+      "ldr q17, [x16, #0x60]\n"
       ".inst 0x6fa0e209  // udot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x6fa1e20d  // udot v13.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x17, #0x70]\n"
+      "ldr q16, [x16, #0x70]\n"
       ".inst 0x6fa0e22a  // udot v10.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x6fa1e22e  // udot v14.4s, v17.16b, v1.4b[1]\n"
-      "ldr q17, [x17, #0x80]\n"
+      "ldr q17, [x16, #0x80]\n"
       ".inst 0x6fa0e20b  // udot v11.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x6fa1e20f  // udot v15.4s, v16.16b, v1.4b[1]\n"
-      "ldr q16, [x17, #0x90]\n"
+      "ldr q16, [x16, #0x90]\n"
       ".inst 0x6f80ea28  // udot v8.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x6f81ea2c  // udot v12.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x17, #0xa0]\n"
+      "ldr q17, [x16, #0xa0]\n"
       ".inst 0x6f80ea09  // udot v9.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x6f81ea0d  // udot v13.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x17, #0xb0]\n"
+      "ldr q16, [x16, #0xb0]\n"
       ".inst 0x6f80ea2a  // udot v10.4s, v17.16b, v0.4b[2]\n"
       ".inst 0x6f81ea2e  // udot v14.4s, v17.16b, v1.4b[2]\n"
-      "ldr q17, [x17, #0xc0]\n"
+      "ldr q17, [x16, #0xc0]\n"
       ".inst 0x6f80ea0b  // udot v11.4s, v16.16b, v0.4b[2]\n"
       ".inst 0x6f81ea0f  // udot v15.4s, v16.16b, v1.4b[2]\n"
-      "ldr q16, [x17, #0xd0]\n"
+      "ldr q16, [x16, #0xd0]\n"
       ".inst 0x6fa0ea28  // udot v8.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x6fa1ea2c  // udot v12.4s, v17.16b, v1.4b[3]\n"
-      "ldr q17, [x17, #0xe0]\n"
+      "ldr q17, [x16, #0xe0]\n"
       ".inst 0x6fa0ea09  // udot v9.4s, v16.16b, v0.4b[3]\n"
       ".inst 0x6fa1ea0d  // udot v13.4s, v16.16b, v1.4b[3]\n"
-      "ldr q16, [x17, #0xf0]\n"
+      "ldr q16, [x16, #0xf0]\n"
       ".inst 0x6fa0ea2a  // udot v10.4s, v17.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa1ea2e  // udot v14.4s, v17.16b, v1.4b[3]\n"
       ".inst 0x6fa0ea0b  // udot v11.4s, v16.16b, v0.4b[3]\n"
       ".inst 0x6fa1ea0f  // udot v15.4s, v16.16b, v1.4b[3]\n"
@@ -678,16 +678,16 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "sub x14, x14, #0x4\n"
       "ldr s18, [x12], #0x4\n"
       "cmp x14, #0x4\n"
-      "ldr q17, [x17, #0x0]\n"
-      "ldr q16, [x17, #0x10]\n"
+      "ldr q17, [x16, #0x0]\n"
       ".inst 0x6f93e228  // udot v8.4s, v17.16b, v19.4b[0]\n"
+      "ldr q16, [x16, #0x10]\n"
       ".inst 0x6f92e22c  // udot v12.4s, v17.16b, v18.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q17, [x16, #0x20]\n"
       ".inst 0x6f93e209  // udot v9.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x6f92e20d  // udot v13.4s, v16.16b, v18.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x6f93e22a  // udot v10.4s, v17.16b, v19.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f92e22e  // udot v14.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x6f93e20b  // udot v11.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x6f92e20f  // udot v15.4s, v16.16b, v18.4b[0]\n"
@@ -705,16 +705,16 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr b0, [x13, #0x0]\n"
       "ldr b1, [x12, #0x0]\n"
       "57:"  // Height 2: Multiply loop: Ragged operand read: Done
-      "ldr q17, [x17, #0x0]\n"
-      "ldr q16, [x17, #0x10]\n"
+      "ldr q17, [x16, #0x0]\n"
       ".inst 0x6f80e228  // udot v8.4s, v17.16b, v0.4b[0]\n"
+      "ldr q16, [x16, #0x10]\n"
       ".inst 0x6f81e22c  // udot v12.4s, v17.16b, v1.4b[0]\n"
-      "ldr q17, [x17, #0x20]\n"
+      "ldr q17, [x16, #0x20]\n"
       ".inst 0x6f80e209  // udot v9.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x6f81e20d  // udot v13.4s, v16.16b, v1.4b[0]\n"
-      "ldr q16, [x17, #0x30]\n"
+      "ldr q16, [x16, #0x30]\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f81e22e  // udot v14.4s, v17.16b, v1.4b[0]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x6f81e20f  // udot v15.4s, v16.16b, v1.4b[0]\n"
@@ -724,79 +724,79 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 48b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x17, x20, LSL #2\n"
       "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "bge 67f\n"
       "tbz x8, #3, 62f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "tbz x8, #2, 60f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "tbz x8, #1, 59f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "tbz x8, #0, 66f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "b 66f\n"
       "59:"  // Height 2: Partial direct writeback: partial_1_12
       "tbz x8, #0, 66f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "b 66f\n"
       "60:"  // Height 2: Partial direct writeback: partial_2_8
       "tbz x8, #1, 61f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "tbz x8, #0, 66f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "b 66f\n"
       "61:"  // Height 2: Partial direct writeback: partial_1_8
       "tbz x8, #0, 66f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "b 66f\n"
       "62:"  // Height 2: Partial direct writeback: partial_4_0
       "tbz x8, #2, 64f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "tbz x8, #1, 63f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "tbz x8, #0, 66f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "b 66f\n"
       "63:"  // Height 2: Partial direct writeback: partial_1_4
       "tbz x8, #0, 66f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "b 66f\n"
       "64:"  // Height 2: Partial direct writeback: partial_2_0
       "tbz x8, #1, 65f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "tbz x8, #0, 66f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "b 66f\n"
       "65:"  // Height 2: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "66:"  // Height 2: Partial direct writeback: Done
       "b 68f\n"
       "67:"  // Height 2: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -807,107 +807,107 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 206f\n"
       "69:"  // Height 3
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "70:"  // Height 3: Column loop
       "tbz %x[flags], #0, 80f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x17, x20, LSL #2\n"
       "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "bge 79f\n"
       "tbz x8, #3, 74f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "ld1 { v17.4s }, [x23], #0x10\n"
       "tbz x8, #2, 72f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "ld1 { v18.4s }, [x23], #0x10\n"
       "tbz x8, #1, 71f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "ldr d19, [x23], #0x8\n"
       "tbz x8, #0, 78f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "ld1 { v19.s }[2], [x23]\n"
       "b 78f\n"
       "71:"  // Height 3: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 78f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "ldr s19, [x23, #0x0]\n"
       "b 78f\n"
       "72:"  // Height 3: Partial accumulate: partial_2_8
       "tbz x8, #1, 73f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "ldr d18, [x23], #0x8\n"
       "tbz x8, #0, 78f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "ld1 { v18.s }[2], [x23]\n"
       "b 78f\n"
       "73:"  // Height 3: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 78f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "ldr s18, [x23, #0x0]\n"
       "b 78f\n"
       "74:"  // Height 3: Partial accumulate: partial_4_0
       "tbz x8, #2, 76f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "tbz x8, #1, 75f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "ldr d17, [x23], #0x8\n"
       "tbz x8, #0, 78f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "ld1 { v17.s }[2], [x23]\n"
       "b 78f\n"
       "75:"  // Height 3: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 78f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "ldr s17, [x23, #0x0]\n"
       "b 78f\n"
       "76:"  // Height 3: Partial accumulate: partial_2_0
       "tbz x8, #1, 77f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "ldr d16, [x23], #0x8\n"
       "tbz x8, #0, 78f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "ld1 { v16.s }[2], [x23]\n"
       "b 78f\n"
       "77:"  // Height 3: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "ldr s16, [x23, #0x0]\n"
       "78:"  // Height 3: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 81f\n"
       "79:"  // Height 3: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -934,8 +934,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "82:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 83f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -959,123 +959,123 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "cmp x14, #0x20\n"
       "ldr q1, [x12, #0x0]\n"
       "ldr q2, [x11, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 86f\n"
       "85:"  // Height 3: Multiply loop: Main loop head
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x6f81e0cc  // udot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x6f82e0d0  // udot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr d21, [x17, #0x20]\n"
+      "ldr d21, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x13, x13, #0x10\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x12, x12, #0x10\n"
+      "ldr x21, [x16, #0x48]\n"
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr d20, [x17, #0x30]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x48]\n"
-      "add x11, x11, #0x10\n"
-      "ldr x24, [x13, #0x8]\n"
+      "ldr d20, [x16, #0x30]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x6f80e2aa  // udot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x6f81e2ae  // udot v14.4s, v21.16b, v1.4b[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x20, [x16, #0x58]\n"
       ".inst 0x6f82e2b2  // udot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr d21, [x17, #0x40]\n"
+      "ldr d21, [x16, #0x40]\n"
       ".inst 0x6f80e28b  // udot v11.4s, v20.16b, v0.4b[0]\n"
-      "ldr x23, [x12, #0x8]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x6f81e28f  // udot v15.4s, v20.16b, v1.4b[0]\n"
-      "ldr x22, [x11, #0x8]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x6f82e293  // udot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr d20, [x17, #0x50]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x68]\n"
-      "sub x14, x14, #0x10\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr d20, [x16, #0x50]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x6fa0e2a8  // udot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x6fa1e2ac  // udot v12.4s, v21.16b, v1.4b[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x6fa2e2b0  // udot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr d21, [x17, #0x60]\n"
+      "ldr d21, [x16, #0x60]\n"
       ".inst 0x6fa0e289  // udot v9.4s, v20.16b, v0.4b[1]\n"
-      "cmp x14, #0x20\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x6fa1e28d  // udot v13.4s, v20.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "ldr x21, [x16, #0x88]\n"
       ".inst 0x6fa2e291  // udot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr d20, [x17, #0x70]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0x88]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr d20, [x16, #0x70]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x6fa0e2aa  // udot v10.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x6fa1e2ae  // udot v14.4s, v21.16b, v1.4b[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "ldr x20, [x16, #0x98]\n"
       ".inst 0x6fa2e2b2  // udot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr d21, [x17, #0x80]\n"
+      "ldr d21, [x16, #0x80]\n"
       ".inst 0x6fa0e28b  // udot v11.4s, v20.16b, v0.4b[1]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x6fa1e28f  // udot v15.4s, v20.16b, v1.4b[1]\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x6fa2e293  // udot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr d20, [x17, #0x90]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xa8]\n"
+      "ldr d20, [x16, #0x90]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x6f80eaa8  // udot v8.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x6f81eaac  // udot v12.4s, v21.16b, v1.4b[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x6f82eab0  // udot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr d21, [x17, #0xa0]\n"
+      "ldr d21, [x16, #0xa0]\n"
       ".inst 0x6f80ea89  // udot v9.4s, v20.16b, v0.4b[2]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x6f81ea8d  // udot v13.4s, v20.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x6f82ea91  // udot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr d20, [x17, #0xb0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xc8]\n"
+      "ldr d20, [x16, #0xb0]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x6f80eaaa  // udot v10.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x6f81eaae  // udot v14.4s, v21.16b, v1.4b[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
+      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x6f82eab2  // udot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr d21, [x17, #0xc0]\n"
+      "ldr d21, [x16, #0xc0]\n"
       ".inst 0x6f80ea8b  // udot v11.4s, v20.16b, v0.4b[2]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x6f81ea8f  // udot v15.4s, v20.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x6f82ea93  // udot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr d20, [x17, #0xd0]\n"
-      "mov v21.d[1], x21\n"
-      "ldr x21, [x17, #0xe8]\n"
+      "ldr d20, [x16, #0xd0]\n"
       "mov v20.d[1], x20\n"
       ".inst 0x6fa0eaa8  // udot v8.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x6fa1eaac  // udot v12.4s, v21.16b, v1.4b[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x6fa2eab0  // udot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr d21, [x17, #0xe0]\n"
+      "ldr d21, [x16, #0xe0]\n"
       ".inst 0x6fa0ea89  // udot v9.4s, v20.16b, v0.4b[3]\n"
+      "mov v21.d[1], x21\n"
       ".inst 0x6fa1ea8d  // udot v13.4s, v20.16b, v1.4b[3]\n"
+      "add x13, x13, #0x10\n"
       ".inst 0x6fa2ea91  // udot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr d20, [x17, #0xf0]\n"
-      "mov v21.d[1], x21\n"
-      "add x17, x17, #0x100\n"
-      "ldr x21, [x17, #0x8]\n"
+      "ldr d20, [x16, #0xf0]\n"
       "mov v20.d[1], x20\n"
+      "add x12, x12, #0x10\n"
+      "add x11, x11, #0x10\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa0eaaa  // udot v10.4s, v21.16b, v0.4b[3]\n"
+      "ldr x20, [x16, #0x8]\n"
       ".inst 0x6fa1eaae  // udot v14.4s, v21.16b, v1.4b[3]\n"
-      "ldr x20, [x17, #0x18]\n"
+      "ldr x23, [x13, #0x8]\n"
       ".inst 0x6fa2eab2  // udot v18.4s, v21.16b, v2.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
       ".inst 0x6fa0ea8b  // udot v11.4s, v20.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x6fa1ea8f  // udot v15.4s, v20.16b, v1.4b[3]\n"
       "ldr d1, [x12, #0x0]\n"
+      "ldr x22, [x12, #0x8]\n"
       ".inst 0x6fa2ea93  // udot v19.4s, v20.16b, v2.4b[3]\n"
       "ldr d2, [x11, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x21\n"
-      "mov v0.d[1], x24\n"
-      "mov v1.d[1], x23\n"
-      "mov v2.d[1], x22\n"
+      "sub x14, x14, #0x10\n"
+      "ldr d7, [x16, #0x10]\n"
+      "cmp x14, #0x20\n"
+      "ldr x21, [x11, #0x8]\n"
+      "mov v6.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
+      "mov v0.d[1], x23\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
+      "mov v1.d[1], x22\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v2.d[1], x21\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       "mov v7.d[1], x20\n"
       "bge 85b\n"
       "86:"  // Height 3: Multiply loop: Single iteration only
@@ -1084,66 +1084,66 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f81e0cc  // udot v12.4s, v6.16b, v1.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x6f82e0d0  // udot v16.4s, v6.16b, v2.4b[0]\n"
-      "ldr q21, [x17, #0x20]\n"
+      "ldr q21, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
       "sub x14, x14, #0x10\n"
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
-      "ldr q20, [x17, #0x30]\n"
+      "ldr q20, [x16, #0x30]\n"
       ".inst 0x6f80e2aa  // udot v10.4s, v21.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x6f81e2ae  // udot v14.4s, v21.16b, v1.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x6f82e2b2  // udot v18.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x17, #0x40]\n"
+      "ldr q21, [x16, #0x40]\n"
       ".inst 0x6f80e28b  // udot v11.4s, v20.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x6f81e28f  // udot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x6f82e293  // udot v19.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x17, #0x50]\n"
+      "ldr q20, [x16, #0x50]\n"
       ".inst 0x6fa0e2a8  // udot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x6fa1e2ac  // udot v12.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x6fa2e2b0  // udot v16.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x17, #0x60]\n"
+      "ldr q21, [x16, #0x60]\n"
       ".inst 0x6fa0e289  // udot v9.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x6fa1e28d  // udot v13.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x6fa2e291  // udot v17.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x17, #0x70]\n"
+      "ldr q20, [x16, #0x70]\n"
       ".inst 0x6fa0e2aa  // udot v10.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x6fa1e2ae  // udot v14.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x6fa2e2b2  // udot v18.4s, v21.16b, v2.4b[1]\n"
-      "ldr q21, [x17, #0x80]\n"
+      "ldr q21, [x16, #0x80]\n"
       ".inst 0x6fa0e28b  // udot v11.4s, v20.16b, v0.4b[1]\n"
       ".inst 0x6fa1e28f  // udot v15.4s, v20.16b, v1.4b[1]\n"
       ".inst 0x6fa2e293  // udot v19.4s, v20.16b, v2.4b[1]\n"
-      "ldr q20, [x17, #0x90]\n"
+      "ldr q20, [x16, #0x90]\n"
       ".inst 0x6f80eaa8  // udot v8.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x6f81eaac  // udot v12.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x6f82eab0  // udot v16.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x17, #0xa0]\n"
+      "ldr q21, [x16, #0xa0]\n"
       ".inst 0x6f80ea89  // udot v9.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x6f81ea8d  // udot v13.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x6f82ea91  // udot v17.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x17, #0xb0]\n"
+      "ldr q20, [x16, #0xb0]\n"
       ".inst 0x6f80eaaa  // udot v10.4s, v21.16b, v0.4b[2]\n"
       ".inst 0x6f81eaae  // udot v14.4s, v21.16b, v1.4b[2]\n"
       ".inst 0x6f82eab2  // udot v18.4s, v21.16b, v2.4b[2]\n"
-      "ldr q21, [x17, #0xc0]\n"
+      "ldr q21, [x16, #0xc0]\n"
       ".inst 0x6f80ea8b  // udot v11.4s, v20.16b, v0.4b[2]\n"
       ".inst 0x6f81ea8f  // udot v15.4s, v20.16b, v1.4b[2]\n"
       ".inst 0x6f82ea93  // udot v19.4s, v20.16b, v2.4b[2]\n"
-      "ldr q20, [x17, #0xd0]\n"
+      "ldr q20, [x16, #0xd0]\n"
       ".inst 0x6fa0eaa8  // udot v8.4s, v21.16b, v0.4b[3]\n"
       ".inst 0x6fa1eaac  // udot v12.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x6fa2eab0  // udot v16.4s, v21.16b, v2.4b[3]\n"
-      "ldr q21, [x17, #0xe0]\n"
+      "ldr q21, [x16, #0xe0]\n"
       ".inst 0x6fa0ea89  // udot v9.4s, v20.16b, v0.4b[3]\n"
       ".inst 0x6fa1ea8d  // udot v13.4s, v20.16b, v1.4b[3]\n"
       ".inst 0x6fa2ea91  // udot v17.4s, v20.16b, v2.4b[3]\n"
-      "ldr q20, [x17, #0xf0]\n"
+      "ldr q20, [x16, #0xf0]\n"
       ".inst 0x6fa0eaaa  // udot v10.4s, v21.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa1eaae  // udot v14.4s, v21.16b, v1.4b[3]\n"
       ".inst 0x6fa2eab2  // udot v18.4s, v21.16b, v2.4b[3]\n"
       ".inst 0x6fa0ea8b  // udot v11.4s, v20.16b, v0.4b[3]\n"
@@ -1159,18 +1159,18 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr s23, [x12], #0x4\n"
       "cmp x14, #0x4\n"
       "ldr s22, [x11], #0x4\n"
-      "ldr q21, [x17, #0x0]\n"
-      "ldr q20, [x17, #0x10]\n"
+      "ldr q21, [x16, #0x0]\n"
       ".inst 0x6f98e2a8  // udot v8.4s, v21.16b, v24.4b[0]\n"
+      "ldr q20, [x16, #0x10]\n"
       ".inst 0x6f97e2ac  // udot v12.4s, v21.16b, v23.4b[0]\n"
       ".inst 0x6f96e2b0  // udot v16.4s, v21.16b, v22.4b[0]\n"
-      "ldr q21, [x17, #0x20]\n"
+      "ldr q21, [x16, #0x20]\n"
       ".inst 0x6f98e289  // udot v9.4s, v20.16b, v24.4b[0]\n"
       ".inst 0x6f97e28d  // udot v13.4s, v20.16b, v23.4b[0]\n"
       ".inst 0x6f96e291  // udot v17.4s, v20.16b, v22.4b[0]\n"
-      "ldr q20, [x17, #0x30]\n"
+      "ldr q20, [x16, #0x30]\n"
       ".inst 0x6f98e2aa  // udot v10.4s, v21.16b, v24.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f97e2ae  // udot v14.4s, v21.16b, v23.4b[0]\n"
       ".inst 0x6f96e2b2  // udot v18.4s, v21.16b, v22.4b[0]\n"
       ".inst 0x6f98e28b  // udot v11.4s, v20.16b, v24.4b[0]\n"
@@ -1193,18 +1193,18 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr b1, [x12, #0x0]\n"
       "ldr b2, [x11, #0x0]\n"
       "91:"  // Height 3: Multiply loop: Ragged operand read: Done
-      "ldr q21, [x17, #0x0]\n"
-      "ldr q20, [x17, #0x10]\n"
+      "ldr q21, [x16, #0x0]\n"
       ".inst 0x6f80e2a8  // udot v8.4s, v21.16b, v0.4b[0]\n"
+      "ldr q20, [x16, #0x10]\n"
       ".inst 0x6f81e2ac  // udot v12.4s, v21.16b, v1.4b[0]\n"
       ".inst 0x6f82e2b0  // udot v16.4s, v21.16b, v2.4b[0]\n"
-      "ldr q21, [x17, #0x20]\n"
+      "ldr q21, [x16, #0x20]\n"
       ".inst 0x6f80e289  // udot v9.4s, v20.16b, v0.4b[0]\n"
       ".inst 0x6f81e28d  // udot v13.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x6f82e291  // udot v17.4s, v20.16b, v2.4b[0]\n"
-      "ldr q20, [x17, #0x30]\n"
+      "ldr q20, [x16, #0x30]\n"
       ".inst 0x6f80e2aa  // udot v10.4s, v21.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f81e2ae  // udot v14.4s, v21.16b, v1.4b[0]\n"
       ".inst 0x6f82e2b2  // udot v18.4s, v21.16b, v2.4b[0]\n"
       ".inst 0x6f80e28b  // udot v11.4s, v20.16b, v0.4b[0]\n"
@@ -1216,97 +1216,97 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 82b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "bge 101f\n"
       "tbz x8, #3, 96f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "st1 { v17.4s }, [x23], #0x10\n"
       "tbz x8, #2, 94f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "st1 { v18.4s }, [x23], #0x10\n"
       "tbz x8, #1, 93f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "str d19, [x23], #0x8\n"
       "tbz x8, #0, 100f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "st1 { v19.s }[2], [x23]\n"
       "b 100f\n"
       "93:"  // Height 3: Partial direct writeback: partial_1_12
       "tbz x8, #0, 100f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "str s19, [x23, #0x0]\n"
       "b 100f\n"
       "94:"  // Height 3: Partial direct writeback: partial_2_8
       "tbz x8, #1, 95f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "str d18, [x23], #0x8\n"
       "tbz x8, #0, 100f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "st1 { v18.s }[2], [x23]\n"
       "b 100f\n"
       "95:"  // Height 3: Partial direct writeback: partial_1_8
       "tbz x8, #0, 100f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "str s18, [x23, #0x0]\n"
       "b 100f\n"
       "96:"  // Height 3: Partial direct writeback: partial_4_0
       "tbz x8, #2, 98f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "tbz x8, #1, 97f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "str d17, [x23], #0x8\n"
       "tbz x8, #0, 100f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "st1 { v17.s }[2], [x23]\n"
       "b 100f\n"
       "97:"  // Height 3: Partial direct writeback: partial_1_4
       "tbz x8, #0, 100f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "str s17, [x23, #0x0]\n"
       "b 100f\n"
       "98:"  // Height 3: Partial direct writeback: partial_2_0
       "tbz x8, #1, 99f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "str d16, [x23], #0x8\n"
       "tbz x8, #0, 100f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "st1 { v16.s }[2], [x23]\n"
       "b 100f\n"
       "99:"  // Height 3: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "str s16, [x23, #0x0]\n"
       "100:"  // Height 3: Partial direct writeback: Done
       "b 102f\n"
       "101:"  // Height 3: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -1321,38 +1321,38 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 206f\n"
       "103:"  // Height 4
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "104:"  // Height 4: Column loop
       "tbz %x[flags], #0, 114f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
       "add x22, x23, x20, LSL #2\n"
       "bge 113f\n"
       "tbz x8, #3, 108f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "ld1 { v17.4s }, [x23], #0x10\n"
       "ld1 { v21.4s }, [x22], #0x10\n"
       "tbz x8, #2, 106f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "ld1 { v18.4s }, [x23], #0x10\n"
       "ld1 { v22.4s }, [x22], #0x10\n"
       "tbz x8, #1, 105f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "ldr d19, [x23], #0x8\n"
       "ldr d23, [x22], #0x8\n"
       "tbz x8, #0, 112f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "ld1 { v19.s }[2], [x23]\n"
       "ld1 { v23.s }[2], [x22]\n"
@@ -1360,20 +1360,20 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "105:"  // Height 4: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 112f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "ldr s19, [x23, #0x0]\n"
       "ldr s23, [x22, #0x0]\n"
       "b 112f\n"
       "106:"  // Height 4: Partial accumulate: partial_2_8
       "tbz x8, #1, 107f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "ldr d18, [x23], #0x8\n"
       "ldr d22, [x22], #0x8\n"
       "tbz x8, #0, 112f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "ld1 { v18.s }[2], [x23]\n"
       "ld1 { v22.s }[2], [x22]\n"
@@ -1381,25 +1381,25 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "107:"  // Height 4: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 112f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "ldr s18, [x23, #0x0]\n"
       "ldr s22, [x22, #0x0]\n"
       "b 112f\n"
       "108:"  // Height 4: Partial accumulate: partial_4_0
       "tbz x8, #2, 110f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "tbz x8, #1, 109f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "ldr d17, [x23], #0x8\n"
       "ldr d21, [x22], #0x8\n"
       "tbz x8, #0, 112f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "ld1 { v17.s }[2], [x23]\n"
       "ld1 { v21.s }[2], [x22]\n"
@@ -1407,38 +1407,38 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "109:"  // Height 4: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 112f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "ldr s17, [x23, #0x0]\n"
       "ldr s21, [x22, #0x0]\n"
       "b 112f\n"
       "110:"  // Height 4: Partial accumulate: partial_2_0
       "tbz x8, #1, 111f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "ldr d16, [x23], #0x8\n"
       "ldr d20, [x22], #0x8\n"
       "tbz x8, #0, 112f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "ld1 { v16.s }[2], [x23]\n"
       "ld1 { v20.s }[2], [x22]\n"
       "b 112f\n"
       "111:"  // Height 4: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "ldr s16, [x23, #0x0]\n"
       "ldr s20, [x22, #0x0]\n"
       "112:"  // Height 4: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 115f\n"
       "113:"  // Height 4: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -1473,8 +1473,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "116:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 117f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1502,129 +1502,130 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr q1, [x12, #0x0]\n"
       "ldr q2, [x11, #0x0]\n"
       "ldr q3, [x10, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 120f\n"
       "119:"  // Height 4: Multiply loop: Main loop head
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x20, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x6f81e0cc  // udot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x21, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x6f82e0d0  // udot v16.4s, v6.16b, v2.4b[0]\n"
       "add x13, x13, #0x10\n"
       ".inst 0x6f83e0d4  // udot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr d25, [x17, #0x20]\n"
+      "ldr d25, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x12, x12, #0x10\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x11, x11, #0x10\n"
+      "ldr x21, [x16, #0x48]\n"
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x12, x12, #0x10\n"
       ".inst 0x6f83e0f5  // udot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr d24, [x17, #0x30]\n"
+      "ldr d24, [x16, #0x30]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x6f80e32a  // udot v10.4s, v25.16b, v0.4b[0]\n"
-      "ldr x20, [x17, #0x48]\n"
       ".inst 0x6f81e32e  // udot v14.4s, v25.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0x58]\n"
       ".inst 0x6f82e332  // udot v18.4s, v25.16b, v2.4b[0]\n"
+      "add x11, x11, #0x10\n"
       ".inst 0x6f83e336  // udot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr d25, [x17, #0x40]\n"
+      "ldr d25, [x16, #0x40]\n"
       ".inst 0x6f80e30b  // udot v11.4s, v24.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x58]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x6f81e30f  // udot v15.4s, v24.16b, v1.4b[0]\n"
-      "ldr x25, [x13, #0x8]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x6f82e313  // udot v19.4s, v24.16b, v2.4b[0]\n"
-      "mov v25.d[1], x20\n"
+      "add x10, x10, #0x10\n"
       ".inst 0x6f83e317  // udot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr d24, [x17, #0x50]\n"
+      "ldr d24, [x16, #0x50]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x6fa0e328  // udot v8.4s, v25.16b, v0.4b[1]\n"
-      "ldr x20, [x17, #0x68]\n"
       ".inst 0x6fa1e32c  // udot v12.4s, v25.16b, v1.4b[1]\n"
-      "ldr x24, [x12, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x6fa2e330  // udot v16.4s, v25.16b, v2.4b[1]\n"
+      "ldr x25, [x13, #0x8]\n"
       ".inst 0x6fa3e334  // udot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr d25, [x17, #0x60]\n"
+      "ldr d25, [x16, #0x60]\n"
       ".inst 0x6fa0e309  // udot v9.4s, v24.16b, v0.4b[1]\n"
-      "ldr x21, [x17, #0x78]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x6fa1e30d  // udot v13.4s, v24.16b, v1.4b[1]\n"
-      "ldr x23, [x11, #0x8]\n"
+      "ldr x21, [x16, #0x88]\n"
       ".inst 0x6fa2e311  // udot v17.4s, v24.16b, v2.4b[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x24, [x12, #0x8]\n"
       ".inst 0x6fa3e315  // udot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr d24, [x17, #0x70]\n"
+      "ldr d24, [x16, #0x70]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x6fa0e32a  // udot v10.4s, v25.16b, v0.4b[1]\n"
-      "ldr x20, [x17, #0x88]\n"
       ".inst 0x6fa1e32e  // udot v14.4s, v25.16b, v1.4b[1]\n"
-      "ldr x22, [x10, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0x98]\n"
       ".inst 0x6fa2e332  // udot v18.4s, v25.16b, v2.4b[1]\n"
+      "ldr x23, [x11, #0x8]\n"
       ".inst 0x6fa3e336  // udot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr d25, [x17, #0x80]\n"
+      "ldr d25, [x16, #0x80]\n"
       ".inst 0x6fa0e30b  // udot v11.4s, v24.16b, v0.4b[1]\n"
-      "ldr x21, [x17, #0x98]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x6fa1e30f  // udot v15.4s, v24.16b, v1.4b[1]\n"
-      "sub x14, x14, #0x10\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x6fa2e313  // udot v19.4s, v24.16b, v2.4b[1]\n"
-      "mov v25.d[1], x20\n"
+      "ldr x22, [x10, #0x8]\n"
       ".inst 0x6fa3e317  // udot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr d24, [x17, #0x90]\n"
+      "ldr d24, [x16, #0x90]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x6f80eb28  // udot v8.4s, v25.16b, v0.4b[2]\n"
-      "ldr x20, [x17, #0xa8]\n"
       ".inst 0x6f81eb2c  // udot v12.4s, v25.16b, v1.4b[2]\n"
-      "cmp x14, #0x20\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x6f82eb30  // udot v16.4s, v25.16b, v2.4b[2]\n"
+      "sub x14, x14, #0x10\n"
       ".inst 0x6f83eb34  // udot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr d25, [x17, #0xa0]\n"
+      "ldr d25, [x16, #0xa0]\n"
       ".inst 0x6f80eb09  // udot v9.4s, v24.16b, v0.4b[2]\n"
-      "ldr x21, [x17, #0xb8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x6f81eb0d  // udot v13.4s, v24.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
+      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x6f82eb11  // udot v17.4s, v24.16b, v2.4b[2]\n"
-      "mov v25.d[1], x20\n"
+      "cmp x14, #0x20\n"
       ".inst 0x6f83eb15  // udot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr d24, [x17, #0xb0]\n"
+      "ldr d24, [x16, #0xb0]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x6f80eb2a  // udot v10.4s, v25.16b, v0.4b[2]\n"
-      "ldr x20, [x17, #0xc8]\n"
       ".inst 0x6f81eb2e  // udot v14.4s, v25.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x6f82eb32  // udot v18.4s, v25.16b, v2.4b[2]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x6f83eb36  // udot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr d25, [x17, #0xc0]\n"
+      "ldr d25, [x16, #0xc0]\n"
       ".inst 0x6f80eb0b  // udot v11.4s, v24.16b, v0.4b[2]\n"
-      "ldr x21, [x17, #0xd8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x6f81eb0f  // udot v15.4s, v24.16b, v1.4b[2]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x6f82eb13  // udot v19.4s, v24.16b, v2.4b[2]\n"
-      "mov v25.d[1], x20\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x6f83eb17  // udot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr d24, [x17, #0xd0]\n"
+      "ldr d24, [x16, #0xd0]\n"
+      "mov v24.d[1], x20\n"
       ".inst 0x6fa0eb28  // udot v8.4s, v25.16b, v0.4b[3]\n"
-      "ldr x20, [x17, #0xe8]\n"
       ".inst 0x6fa1eb2c  // udot v12.4s, v25.16b, v1.4b[3]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x6fa2eb30  // udot v16.4s, v25.16b, v2.4b[3]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x6fa3eb34  // udot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr d25, [x17, #0xe0]\n"
+      "ldr d25, [x16, #0xe0]\n"
       ".inst 0x6fa0eb09  // udot v9.4s, v24.16b, v0.4b[3]\n"
-      "ldr x21, [x17, #0xf8]\n"
+      "mov v25.d[1], x21\n"
       ".inst 0x6fa1eb0d  // udot v13.4s, v24.16b, v1.4b[3]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x6fa2eb11  // udot v17.4s, v24.16b, v2.4b[3]\n"
-      "mov v25.d[1], x20\n"
       ".inst 0x6fa3eb15  // udot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr d24, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "ldr d24, [x16, #0xf0]\n"
+      "mov v24.d[1], x20\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa0eb2a  // udot v10.4s, v25.16b, v0.4b[3]\n"
+      "ldr x21, [x16, #0x8]\n"
       ".inst 0x6fa1eb2e  // udot v14.4s, v25.16b, v1.4b[3]\n"
-      "ldr x20, [x17, #0x8]\n"
-      "mov v24.d[1], x21\n"
+      "ldr x20, [x16, #0x18]\n"
       ".inst 0x6fa2eb32  // udot v18.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x6fa3eb36  // udot v22.4s, v25.16b, v3.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
       ".inst 0x6fa0eb0b  // udot v11.4s, v24.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x6fa1eb0f  // udot v15.4s, v24.16b, v1.4b[3]\n"
@@ -1633,9 +1634,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr d2, [x11, #0x0]\n"
       ".inst 0x6fa3eb17  // udot v23.4s, v24.16b, v3.4b[3]\n"
       "ldr d3, [x10, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
-      "mov v6.d[1], x20\n"
-      "ldr x20, [x17, #0x18]\n"
+      "ldr d7, [x16, #0x10]\n"
+      "mov v6.d[1], x21\n"
       "mov v0.d[1], x25\n"
       "mov v1.d[1], x24\n"
       "mov v2.d[1], x23\n"
@@ -1650,7 +1650,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f82e0d0  // udot v16.4s, v6.16b, v2.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x6f83e0d4  // udot v20.4s, v6.16b, v3.4b[0]\n"
-      "ldr q25, [x17, #0x20]\n"
+      "ldr q25, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       "add x10, x10, #0x10\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
@@ -1658,7 +1658,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x6f83e0f5  // udot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr q24, [x17, #0x30]\n"
+      "ldr q24, [x16, #0x30]\n"
       ".inst 0x6f80e32a  // udot v10.4s, v25.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x6f81e32e  // udot v14.4s, v25.16b, v1.4b[0]\n"
@@ -1666,64 +1666,64 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f82e332  // udot v18.4s, v25.16b, v2.4b[0]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x6f83e336  // udot v22.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x17, #0x40]\n"
+      "ldr q25, [x16, #0x40]\n"
       ".inst 0x6f80e30b  // udot v11.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x6f81e30f  // udot v15.4s, v24.16b, v1.4b[0]\n"
       ".inst 0x6f82e313  // udot v19.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x6f83e317  // udot v23.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x17, #0x50]\n"
+      "ldr q24, [x16, #0x50]\n"
       ".inst 0x6fa0e328  // udot v8.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x6fa1e32c  // udot v12.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x6fa2e330  // udot v16.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x6fa3e334  // udot v20.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x17, #0x60]\n"
+      "ldr q25, [x16, #0x60]\n"
       ".inst 0x6fa0e309  // udot v9.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x6fa1e30d  // udot v13.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x6fa2e311  // udot v17.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x6fa3e315  // udot v21.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x17, #0x70]\n"
+      "ldr q24, [x16, #0x70]\n"
       ".inst 0x6fa0e32a  // udot v10.4s, v25.16b, v0.4b[1]\n"
       ".inst 0x6fa1e32e  // udot v14.4s, v25.16b, v1.4b[1]\n"
       ".inst 0x6fa2e332  // udot v18.4s, v25.16b, v2.4b[1]\n"
       ".inst 0x6fa3e336  // udot v22.4s, v25.16b, v3.4b[1]\n"
-      "ldr q25, [x17, #0x80]\n"
+      "ldr q25, [x16, #0x80]\n"
       ".inst 0x6fa0e30b  // udot v11.4s, v24.16b, v0.4b[1]\n"
       ".inst 0x6fa1e30f  // udot v15.4s, v24.16b, v1.4b[1]\n"
       ".inst 0x6fa2e313  // udot v19.4s, v24.16b, v2.4b[1]\n"
       ".inst 0x6fa3e317  // udot v23.4s, v24.16b, v3.4b[1]\n"
-      "ldr q24, [x17, #0x90]\n"
+      "ldr q24, [x16, #0x90]\n"
       ".inst 0x6f80eb28  // udot v8.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x6f81eb2c  // udot v12.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x6f82eb30  // udot v16.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x6f83eb34  // udot v20.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x17, #0xa0]\n"
+      "ldr q25, [x16, #0xa0]\n"
       ".inst 0x6f80eb09  // udot v9.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x6f81eb0d  // udot v13.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x6f82eb11  // udot v17.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x6f83eb15  // udot v21.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x17, #0xb0]\n"
+      "ldr q24, [x16, #0xb0]\n"
       ".inst 0x6f80eb2a  // udot v10.4s, v25.16b, v0.4b[2]\n"
       ".inst 0x6f81eb2e  // udot v14.4s, v25.16b, v1.4b[2]\n"
       ".inst 0x6f82eb32  // udot v18.4s, v25.16b, v2.4b[2]\n"
       ".inst 0x6f83eb36  // udot v22.4s, v25.16b, v3.4b[2]\n"
-      "ldr q25, [x17, #0xc0]\n"
+      "ldr q25, [x16, #0xc0]\n"
       ".inst 0x6f80eb0b  // udot v11.4s, v24.16b, v0.4b[2]\n"
       ".inst 0x6f81eb0f  // udot v15.4s, v24.16b, v1.4b[2]\n"
       ".inst 0x6f82eb13  // udot v19.4s, v24.16b, v2.4b[2]\n"
       ".inst 0x6f83eb17  // udot v23.4s, v24.16b, v3.4b[2]\n"
-      "ldr q24, [x17, #0xd0]\n"
+      "ldr q24, [x16, #0xd0]\n"
       ".inst 0x6fa0eb28  // udot v8.4s, v25.16b, v0.4b[3]\n"
       ".inst 0x6fa1eb2c  // udot v12.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x6fa2eb30  // udot v16.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x6fa3eb34  // udot v20.4s, v25.16b, v3.4b[3]\n"
-      "ldr q25, [x17, #0xe0]\n"
+      "ldr q25, [x16, #0xe0]\n"
       ".inst 0x6fa0eb09  // udot v9.4s, v24.16b, v0.4b[3]\n"
       ".inst 0x6fa1eb0d  // udot v13.4s, v24.16b, v1.4b[3]\n"
       ".inst 0x6fa2eb11  // udot v17.4s, v24.16b, v2.4b[3]\n"
       ".inst 0x6fa3eb15  // udot v21.4s, v24.16b, v3.4b[3]\n"
-      "ldr q24, [x17, #0xf0]\n"
+      "ldr q24, [x16, #0xf0]\n"
       ".inst 0x6fa0eb2a  // udot v10.4s, v25.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa1eb2e  // udot v14.4s, v25.16b, v1.4b[3]\n"
       ".inst 0x6fa2eb32  // udot v18.4s, v25.16b, v2.4b[3]\n"
       ".inst 0x6fa3eb36  // udot v22.4s, v25.16b, v3.4b[3]\n"
@@ -1742,20 +1742,20 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "cmp x14, #0x4\n"
       "ldr s27, [x11], #0x4\n"
       "ldr s26, [x10], #0x4\n"
-      "ldr q25, [x17, #0x0]\n"
-      "ldr q24, [x17, #0x10]\n"
+      "ldr q25, [x16, #0x0]\n"
       ".inst 0x6f9de328  // udot v8.4s, v25.16b, v29.4b[0]\n"
+      "ldr q24, [x16, #0x10]\n"
       ".inst 0x6f9ce32c  // udot v12.4s, v25.16b, v28.4b[0]\n"
       ".inst 0x6f9be330  // udot v16.4s, v25.16b, v27.4b[0]\n"
       ".inst 0x6f9ae334  // udot v20.4s, v25.16b, v26.4b[0]\n"
-      "ldr q25, [x17, #0x20]\n"
+      "ldr q25, [x16, #0x20]\n"
       ".inst 0x6f9de309  // udot v9.4s, v24.16b, v29.4b[0]\n"
       ".inst 0x6f9ce30d  // udot v13.4s, v24.16b, v28.4b[0]\n"
       ".inst 0x6f9be311  // udot v17.4s, v24.16b, v27.4b[0]\n"
       ".inst 0x6f9ae315  // udot v21.4s, v24.16b, v26.4b[0]\n"
-      "ldr q24, [x17, #0x30]\n"
+      "ldr q24, [x16, #0x30]\n"
       ".inst 0x6f9de32a  // udot v10.4s, v25.16b, v29.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f9ce32e  // udot v14.4s, v25.16b, v28.4b[0]\n"
       ".inst 0x6f9be332  // udot v18.4s, v25.16b, v27.4b[0]\n"
       ".inst 0x6f9ae336  // udot v22.4s, v25.16b, v26.4b[0]\n"
@@ -1783,20 +1783,20 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr b2, [x11, #0x0]\n"
       "ldr b3, [x10, #0x0]\n"
       "125:"  // Height 4: Multiply loop: Ragged operand read: Done
-      "ldr q25, [x17, #0x0]\n"
-      "ldr q24, [x17, #0x10]\n"
+      "ldr q25, [x16, #0x0]\n"
       ".inst 0x6f80e328  // udot v8.4s, v25.16b, v0.4b[0]\n"
+      "ldr q24, [x16, #0x10]\n"
       ".inst 0x6f81e32c  // udot v12.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x6f82e330  // udot v16.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x6f83e334  // udot v20.4s, v25.16b, v3.4b[0]\n"
-      "ldr q25, [x17, #0x20]\n"
+      "ldr q25, [x16, #0x20]\n"
       ".inst 0x6f80e309  // udot v9.4s, v24.16b, v0.4b[0]\n"
       ".inst 0x6f81e30d  // udot v13.4s, v24.16b, v1.4b[0]\n"
       ".inst 0x6f82e311  // udot v17.4s, v24.16b, v2.4b[0]\n"
       ".inst 0x6f83e315  // udot v21.4s, v24.16b, v3.4b[0]\n"
-      "ldr q24, [x17, #0x30]\n"
+      "ldr q24, [x16, #0x30]\n"
       ".inst 0x6f80e32a  // udot v10.4s, v25.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f81e32e  // udot v14.4s, v25.16b, v1.4b[0]\n"
       ".inst 0x6f82e332  // udot v18.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x6f83e336  // udot v22.4s, v25.16b, v3.4b[0]\n"
@@ -1810,18 +1810,18 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 116b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
       "bge 135f\n"
       "tbz x8, #3, 130f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
@@ -1829,96 +1829,96 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "st1 { v20.4s }, [x22], #0x10\n"
       "st1 { v21.4s }, [x22], #0x10\n"
       "tbz x8, #2, 128f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "st1 { v18.4s }, [x23], #0x10\n"
       "st1 { v22.4s }, [x22], #0x10\n"
       "tbz x8, #1, 127f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "str d19, [x23], #0x8\n"
       "str d23, [x22], #0x8\n"
       "tbz x8, #0, 134f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "st1 { v19.s }[2], [x23]\n"
       "st1 { v23.s }[2], [x22]\n"
       "b 134f\n"
       "127:"  // Height 4: Partial direct writeback: partial_1_12
       "tbz x8, #0, 134f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "str s19, [x23, #0x0]\n"
       "str s23, [x22, #0x0]\n"
       "b 134f\n"
       "128:"  // Height 4: Partial direct writeback: partial_2_8
       "tbz x8, #1, 129f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "str d18, [x23], #0x8\n"
       "str d22, [x22], #0x8\n"
       "tbz x8, #0, 134f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "st1 { v18.s }[2], [x23]\n"
       "st1 { v22.s }[2], [x22]\n"
       "b 134f\n"
       "129:"  // Height 4: Partial direct writeback: partial_1_8
       "tbz x8, #0, 134f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "str s18, [x23, #0x0]\n"
       "str s22, [x22, #0x0]\n"
       "b 134f\n"
       "130:"  // Height 4: Partial direct writeback: partial_4_0
       "tbz x8, #2, 132f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "st1 { v20.4s }, [x22], #0x10\n"
       "tbz x8, #1, 131f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "str d17, [x23], #0x8\n"
       "str d21, [x22], #0x8\n"
       "tbz x8, #0, 134f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "st1 { v17.s }[2], [x23]\n"
       "st1 { v21.s }[2], [x22]\n"
       "b 134f\n"
       "131:"  // Height 4: Partial direct writeback: partial_1_4
       "tbz x8, #0, 134f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "str s17, [x23, #0x0]\n"
       "str s21, [x22, #0x0]\n"
       "b 134f\n"
       "132:"  // Height 4: Partial direct writeback: partial_2_0
       "tbz x8, #1, 133f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "str d16, [x23], #0x8\n"
       "str d20, [x22], #0x8\n"
       "tbz x8, #0, 134f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "st1 { v16.s }[2], [x23]\n"
       "st1 { v20.s }[2], [x22]\n"
       "b 134f\n"
       "133:"  // Height 4: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "str s16, [x23, #0x0]\n"
       "str s20, [x22, #0x0]\n"
       "134:"  // Height 4: Partial direct writeback: Done
       "b 136f\n"
       "135:"  // Height 4: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -1937,43 +1937,43 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 206f\n"
       "137:"  // Height 5
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "138:"  // Height 5: Column loop
       "tbz %x[flags], #0, 148f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
       "add x21, x22, x20, LSL #2\n"
       "bge 147f\n"
       "tbz x8, #3, 142f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v24.4s }, [x21], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "ld1 { v17.4s }, [x23], #0x10\n"
       "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v25.4s }, [x21], #0x10\n"
       "tbz x8, #2, 140f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "ld1 { v18.4s }, [x23], #0x10\n"
       "ld1 { v22.4s }, [x22], #0x10\n"
       "ld1 { v26.4s }, [x21], #0x10\n"
       "tbz x8, #1, 139f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "ldr d19, [x23], #0x8\n"
       "ldr d23, [x22], #0x8\n"
       "ldr d27, [x21], #0x8\n"
       "tbz x8, #0, 146f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "ld1 { v19.s }[2], [x23]\n"
       "ld1 { v23.s }[2], [x22]\n"
@@ -1982,7 +1982,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "139:"  // Height 5: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 146f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "ldr s19, [x23, #0x0]\n"
       "ldr s23, [x22, #0x0]\n"
@@ -1990,14 +1990,14 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 146f\n"
       "140:"  // Height 5: Partial accumulate: partial_2_8
       "tbz x8, #1, 141f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "ldr d18, [x23], #0x8\n"
       "ldr d22, [x22], #0x8\n"
       "ldr d26, [x21], #0x8\n"
       "tbz x8, #0, 146f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "ld1 { v18.s }[2], [x23]\n"
       "ld1 { v22.s }[2], [x22]\n"
@@ -2006,7 +2006,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "141:"  // Height 5: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 146f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "ldr s18, [x23, #0x0]\n"
       "ldr s22, [x22, #0x0]\n"
@@ -2014,20 +2014,20 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 146f\n"
       "142:"  // Height 5: Partial accumulate: partial_4_0
       "tbz x8, #2, 144f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v24.4s }, [x21], #0x10\n"
       "tbz x8, #1, 143f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "ldr d17, [x23], #0x8\n"
       "ldr d21, [x22], #0x8\n"
       "ldr d25, [x21], #0x8\n"
       "tbz x8, #0, 146f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "ld1 { v17.s }[2], [x23]\n"
       "ld1 { v21.s }[2], [x22]\n"
@@ -2036,7 +2036,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "143:"  // Height 5: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 146f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "ldr s17, [x23, #0x0]\n"
       "ldr s21, [x22, #0x0]\n"
@@ -2044,34 +2044,34 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 146f\n"
       "144:"  // Height 5: Partial accumulate: partial_2_0
       "tbz x8, #1, 145f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "ldr d16, [x23], #0x8\n"
       "ldr d20, [x22], #0x8\n"
       "ldr d24, [x21], #0x8\n"
       "tbz x8, #0, 146f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "ld1 { v16.s }[2], [x23]\n"
       "ld1 { v20.s }[2], [x22]\n"
       "ld1 { v24.s }[2], [x21]\n"
       "b 146f\n"
       "145:"  // Height 5: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "ldr s16, [x23, #0x0]\n"
       "ldr s20, [x22, #0x0]\n"
       "ldr s24, [x21, #0x0]\n"
       "146:"  // Height 5: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 149f\n"
       "147:"  // Height 5: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -2114,8 +2114,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "150:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 151f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2147,148 +2147,148 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr q2, [x11, #0x0]\n"
       "ldr q3, [x10, #0x0]\n"
       "ldr q4, [x9, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 154f\n"
       "153:"  // Height 5: Multiply loop: Main loop head
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x6f81e0cc  // udot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x6f82e0d0  // udot v16.4s, v6.16b, v2.4b[0]\n"
       "add x13, x13, #0x10\n"
       ".inst 0x6f83e0d4  // udot v20.4s, v6.16b, v3.4b[0]\n"
       "add x12, x12, #0x10\n"
       ".inst 0x6f84e0d8  // udot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr d29, [x17, #0x20]\n"
+      "ldr d29, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x11, x11, #0x10\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "ldr x21, [x16, #0x48]\n"
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v29.d[1], x21\n"
+      "add x11, x11, #0x10\n"
       ".inst 0x6f83e0f5  // udot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x17, #0x48]\n"
+      "add x10, x10, #0x10\n"
       ".inst 0x6f84e0f9  // udot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr d28, [x17, #0x30]\n"
+      "ldr d28, [x16, #0x30]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x6f80e3aa  // udot v10.4s, v29.16b, v0.4b[0]\n"
-      "add x9, x9, #0x10\n"
       ".inst 0x6f81e3ae  // udot v14.4s, v29.16b, v1.4b[0]\n"
-      "ldr x26, [x13, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0x58]\n"
       ".inst 0x6f82e3b2  // udot v18.4s, v29.16b, v2.4b[0]\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x6f83e3b6  // udot v22.4s, v29.16b, v3.4b[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x26, [x13, #0x8]\n"
       ".inst 0x6f84e3ba  // udot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr d29, [x17, #0x40]\n"
+      "ldr d29, [x16, #0x40]\n"
       ".inst 0x6f80e38b  // udot v11.4s, v28.16b, v0.4b[0]\n"
-      "ldr x25, [x12, #0x8]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x6f81e38f  // udot v15.4s, v28.16b, v1.4b[0]\n"
-      "ldr x24, [x11, #0x8]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x6f82e393  // udot v19.4s, v28.16b, v2.4b[0]\n"
-      "mov v29.d[1], x21\n"
+      "ldr x25, [x12, #0x8]\n"
       ".inst 0x6f83e397  // udot v23.4s, v28.16b, v3.4b[0]\n"
-      "ldr x21, [x17, #0x68]\n"
+      "ldr x24, [x11, #0x8]\n"
       ".inst 0x6f84e39b  // udot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr d28, [x17, #0x50]\n"
+      "ldr d28, [x16, #0x50]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x6fa0e3a8  // udot v8.4s, v29.16b, v0.4b[1]\n"
-      "ldr x23, [x10, #0x8]\n"
       ".inst 0x6fa1e3ac  // udot v12.4s, v29.16b, v1.4b[1]\n"
-      "ldr x22, [x9, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x6fa2e3b0  // udot v16.4s, v29.16b, v2.4b[1]\n"
+      "ldr x23, [x10, #0x8]\n"
       ".inst 0x6fa3e3b4  // udot v20.4s, v29.16b, v3.4b[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "ldr x22, [x9, #0x8]\n"
       ".inst 0x6fa4e3b8  // udot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr d29, [x17, #0x60]\n"
+      "ldr d29, [x16, #0x60]\n"
       ".inst 0x6fa0e389  // udot v9.4s, v28.16b, v0.4b[1]\n"
-      "sub x14, x14, #0x10\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x6fa1e38d  // udot v13.4s, v28.16b, v1.4b[1]\n"
-      "cmp x14, #0x20\n"
+      "ldr x21, [x16, #0x88]\n"
       ".inst 0x6fa2e391  // udot v17.4s, v28.16b, v2.4b[1]\n"
-      "mov v29.d[1], x21\n"
+      "sub x14, x14, #0x10\n"
       ".inst 0x6fa3e395  // udot v21.4s, v28.16b, v3.4b[1]\n"
-      "ldr x21, [x17, #0x88]\n"
+      "cmp x14, #0x20\n"
       ".inst 0x6fa4e399  // udot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr d28, [x17, #0x70]\n"
+      "ldr d28, [x16, #0x70]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x6fa0e3aa  // udot v10.4s, v29.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x6fa1e3ae  // udot v14.4s, v29.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0x98]\n"
       ".inst 0x6fa2e3b2  // udot v18.4s, v29.16b, v2.4b[1]\n"
+      "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x6fa3e3b6  // udot v22.4s, v29.16b, v3.4b[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x6fa4e3ba  // udot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr d29, [x17, #0x80]\n"
+      "ldr d29, [x16, #0x80]\n"
       ".inst 0x6fa0e38b  // udot v11.4s, v28.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x6fa1e38f  // udot v15.4s, v28.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x10, #0x80]\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x6fa2e393  // udot v19.4s, v28.16b, v2.4b[1]\n"
-      "mov v29.d[1], x21\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x6fa3e397  // udot v23.4s, v28.16b, v3.4b[1]\n"
-      "ldr x21, [x17, #0xa8]\n"
+      "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x6fa4e39b  // udot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr d28, [x17, #0x90]\n"
+      "ldr d28, [x16, #0x90]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x6f80eba8  // udot v8.4s, v29.16b, v0.4b[2]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x6f81ebac  // udot v12.4s, v29.16b, v1.4b[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x6f82ebb0  // udot v16.4s, v29.16b, v2.4b[2]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x6f83ebb4  // udot v20.4s, v29.16b, v3.4b[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
       ".inst 0x6f84ebb8  // udot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr d29, [x17, #0xa0]\n"
+      "ldr d29, [x16, #0xa0]\n"
       ".inst 0x6f80eb89  // udot v9.4s, v28.16b, v0.4b[2]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x6f81eb8d  // udot v13.4s, v28.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x6f82eb91  // udot v17.4s, v28.16b, v2.4b[2]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x6f83eb95  // udot v21.4s, v28.16b, v3.4b[2]\n"
-      "ldr x21, [x17, #0xc8]\n"
       ".inst 0x6f84eb99  // udot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr d28, [x17, #0xb0]\n"
+      "ldr d28, [x16, #0xb0]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x6f80ebaa  // udot v10.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x6f81ebae  // udot v14.4s, v29.16b, v1.4b[2]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x6f82ebb2  // udot v18.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x6f83ebb6  // udot v22.4s, v29.16b, v3.4b[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
       ".inst 0x6f84ebba  // udot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr d29, [x17, #0xc0]\n"
+      "ldr d29, [x16, #0xc0]\n"
       ".inst 0x6f80eb8b  // udot v11.4s, v28.16b, v0.4b[2]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x6f81eb8f  // udot v15.4s, v28.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x6f82eb93  // udot v19.4s, v28.16b, v2.4b[2]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x6f83eb97  // udot v23.4s, v28.16b, v3.4b[2]\n"
-      "ldr x21, [x17, #0xe8]\n"
       ".inst 0x6f84eb9b  // udot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr d28, [x17, #0xd0]\n"
+      "ldr d28, [x16, #0xd0]\n"
+      "mov v28.d[1], x20\n"
       ".inst 0x6fa0eba8  // udot v8.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x6fa1ebac  // udot v12.4s, v29.16b, v1.4b[3]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x6fa2ebb0  // udot v16.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x6fa3ebb4  // udot v20.4s, v29.16b, v3.4b[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
       ".inst 0x6fa4ebb8  // udot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr d29, [x17, #0xe0]\n"
+      "ldr d29, [x16, #0xe0]\n"
       ".inst 0x6fa0eb89  // udot v9.4s, v28.16b, v0.4b[3]\n"
+      "mov v29.d[1], x21\n"
       ".inst 0x6fa1eb8d  // udot v13.4s, v28.16b, v1.4b[3]\n"
       ".inst 0x6fa2eb91  // udot v17.4s, v28.16b, v2.4b[3]\n"
-      "mov v29.d[1], x21\n"
       ".inst 0x6fa3eb95  // udot v21.4s, v28.16b, v3.4b[3]\n"
       ".inst 0x6fa4eb99  // udot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr d28, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "ldr d28, [x16, #0xf0]\n"
+      "mov v28.d[1], x20\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa0ebaa  // udot v10.4s, v29.16b, v0.4b[3]\n"
+      "ldr x21, [x16, #0x8]\n"
       ".inst 0x6fa1ebae  // udot v14.4s, v29.16b, v1.4b[3]\n"
-      "ldr x21, [x17, #0x8]\n"
-      "mov v28.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
       ".inst 0x6fa2ebb2  // udot v18.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x6fa3ebb6  // udot v22.4s, v29.16b, v3.4b[3]\n"
-      "ldr x20, [x17, #0x18]\n"
       ".inst 0x6fa4ebba  // udot v26.4s, v29.16b, v4.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
       ".inst 0x6fa0eb8b  // udot v11.4s, v28.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x6fa1eb8f  // udot v15.4s, v28.16b, v1.4b[3]\n"
@@ -2299,7 +2299,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr d3, [x10, #0x0]\n"
       ".inst 0x6fa4eb9b  // udot v27.4s, v28.16b, v4.4b[3]\n"
       "ldr d4, [x9, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
+      "ldr d7, [x16, #0x10]\n"
       "mov v6.d[1], x21\n"
       "mov v0.d[1], x26\n"
       "mov v1.d[1], x25\n"
@@ -2318,7 +2318,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f83e0d4  // udot v20.4s, v6.16b, v3.4b[0]\n"
       "add x10, x10, #0x10\n"
       ".inst 0x6f84e0d8  // udot v24.4s, v6.16b, v4.4b[0]\n"
-      "ldr q29, [x17, #0x20]\n"
+      "ldr q29, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2328,7 +2328,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f83e0f5  // udot v21.4s, v7.16b, v3.4b[0]\n"
       "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x6f84e0f9  // udot v25.4s, v7.16b, v4.4b[0]\n"
-      "ldr q28, [x17, #0x30]\n"
+      "ldr q28, [x16, #0x30]\n"
       ".inst 0x6f80e3aa  // udot v10.4s, v29.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x6f81e3ae  // udot v14.4s, v29.16b, v1.4b[0]\n"
@@ -2337,75 +2337,75 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x6f83e3b6  // udot v22.4s, v29.16b, v3.4b[0]\n"
       ".inst 0x6f84e3ba  // udot v26.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x17, #0x40]\n"
+      "ldr q29, [x16, #0x40]\n"
       ".inst 0x6f80e38b  // udot v11.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x6f81e38f  // udot v15.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x6f82e393  // udot v19.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x6f83e397  // udot v23.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x6f84e39b  // udot v27.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x17, #0x50]\n"
+      "ldr q28, [x16, #0x50]\n"
       ".inst 0x6fa0e3a8  // udot v8.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x6fa1e3ac  // udot v12.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x6fa2e3b0  // udot v16.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x6fa3e3b4  // udot v20.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x6fa4e3b8  // udot v24.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x17, #0x60]\n"
+      "ldr q29, [x16, #0x60]\n"
       ".inst 0x6fa0e389  // udot v9.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x6fa1e38d  // udot v13.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x6fa2e391  // udot v17.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x6fa3e395  // udot v21.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x6fa4e399  // udot v25.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x17, #0x70]\n"
+      "ldr q28, [x16, #0x70]\n"
       ".inst 0x6fa0e3aa  // udot v10.4s, v29.16b, v0.4b[1]\n"
       ".inst 0x6fa1e3ae  // udot v14.4s, v29.16b, v1.4b[1]\n"
       ".inst 0x6fa2e3b2  // udot v18.4s, v29.16b, v2.4b[1]\n"
       ".inst 0x6fa3e3b6  // udot v22.4s, v29.16b, v3.4b[1]\n"
       ".inst 0x6fa4e3ba  // udot v26.4s, v29.16b, v4.4b[1]\n"
-      "ldr q29, [x17, #0x80]\n"
+      "ldr q29, [x16, #0x80]\n"
       ".inst 0x6fa0e38b  // udot v11.4s, v28.16b, v0.4b[1]\n"
       ".inst 0x6fa1e38f  // udot v15.4s, v28.16b, v1.4b[1]\n"
       ".inst 0x6fa2e393  // udot v19.4s, v28.16b, v2.4b[1]\n"
       ".inst 0x6fa3e397  // udot v23.4s, v28.16b, v3.4b[1]\n"
       ".inst 0x6fa4e39b  // udot v27.4s, v28.16b, v4.4b[1]\n"
-      "ldr q28, [x17, #0x90]\n"
+      "ldr q28, [x16, #0x90]\n"
       ".inst 0x6f80eba8  // udot v8.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x6f81ebac  // udot v12.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x6f82ebb0  // udot v16.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x6f83ebb4  // udot v20.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x6f84ebb8  // udot v24.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x17, #0xa0]\n"
+      "ldr q29, [x16, #0xa0]\n"
       ".inst 0x6f80eb89  // udot v9.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x6f81eb8d  // udot v13.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x6f82eb91  // udot v17.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x6f83eb95  // udot v21.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x6f84eb99  // udot v25.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x17, #0xb0]\n"
+      "ldr q28, [x16, #0xb0]\n"
       ".inst 0x6f80ebaa  // udot v10.4s, v29.16b, v0.4b[2]\n"
       ".inst 0x6f81ebae  // udot v14.4s, v29.16b, v1.4b[2]\n"
       ".inst 0x6f82ebb2  // udot v18.4s, v29.16b, v2.4b[2]\n"
       ".inst 0x6f83ebb6  // udot v22.4s, v29.16b, v3.4b[2]\n"
       ".inst 0x6f84ebba  // udot v26.4s, v29.16b, v4.4b[2]\n"
-      "ldr q29, [x17, #0xc0]\n"
+      "ldr q29, [x16, #0xc0]\n"
       ".inst 0x6f80eb8b  // udot v11.4s, v28.16b, v0.4b[2]\n"
       ".inst 0x6f81eb8f  // udot v15.4s, v28.16b, v1.4b[2]\n"
       ".inst 0x6f82eb93  // udot v19.4s, v28.16b, v2.4b[2]\n"
       ".inst 0x6f83eb97  // udot v23.4s, v28.16b, v3.4b[2]\n"
       ".inst 0x6f84eb9b  // udot v27.4s, v28.16b, v4.4b[2]\n"
-      "ldr q28, [x17, #0xd0]\n"
+      "ldr q28, [x16, #0xd0]\n"
       ".inst 0x6fa0eba8  // udot v8.4s, v29.16b, v0.4b[3]\n"
       ".inst 0x6fa1ebac  // udot v12.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x6fa2ebb0  // udot v16.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x6fa3ebb4  // udot v20.4s, v29.16b, v3.4b[3]\n"
       ".inst 0x6fa4ebb8  // udot v24.4s, v29.16b, v4.4b[3]\n"
-      "ldr q29, [x17, #0xe0]\n"
+      "ldr q29, [x16, #0xe0]\n"
       ".inst 0x6fa0eb89  // udot v9.4s, v28.16b, v0.4b[3]\n"
       ".inst 0x6fa1eb8d  // udot v13.4s, v28.16b, v1.4b[3]\n"
       ".inst 0x6fa2eb91  // udot v17.4s, v28.16b, v2.4b[3]\n"
       ".inst 0x6fa3eb95  // udot v21.4s, v28.16b, v3.4b[3]\n"
       ".inst 0x6fa4eb99  // udot v25.4s, v28.16b, v4.4b[3]\n"
-      "ldr q28, [x17, #0xf0]\n"
+      "ldr q28, [x16, #0xf0]\n"
       ".inst 0x6fa0ebaa  // udot v10.4s, v29.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa1ebae  // udot v14.4s, v29.16b, v1.4b[3]\n"
       ".inst 0x6fa2ebb2  // udot v18.4s, v29.16b, v2.4b[3]\n"
       ".inst 0x6fa3ebb6  // udot v22.4s, v29.16b, v3.4b[3]\n"
@@ -2427,22 +2427,22 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr s0, [x11], #0x4\n"
       "ldr s31, [x10], #0x4\n"
       "ldr s30, [x9], #0x4\n"
-      "ldr q29, [x17, #0x0]\n"
-      "ldr q28, [x17, #0x10]\n"
+      "ldr q29, [x16, #0x0]\n"
       ".inst 0x6f82e3a8  // udot v8.4s, v29.16b, v2.4b[0]\n"
+      "ldr q28, [x16, #0x10]\n"
       ".inst 0x6f81e3ac  // udot v12.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x6f80e3b0  // udot v16.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x6f9fe3b4  // udot v20.4s, v29.16b, v31.4b[0]\n"
       ".inst 0x6f9ee3b8  // udot v24.4s, v29.16b, v30.4b[0]\n"
-      "ldr q29, [x17, #0x20]\n"
+      "ldr q29, [x16, #0x20]\n"
       ".inst 0x6f82e389  // udot v9.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x6f81e38d  // udot v13.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x6f80e391  // udot v17.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x6f9fe395  // udot v21.4s, v28.16b, v31.4b[0]\n"
       ".inst 0x6f9ee399  // udot v25.4s, v28.16b, v30.4b[0]\n"
-      "ldr q28, [x17, #0x30]\n"
+      "ldr q28, [x16, #0x30]\n"
       ".inst 0x6f82e3aa  // udot v10.4s, v29.16b, v2.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f81e3ae  // udot v14.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x6f80e3b2  // udot v18.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x6f9fe3b6  // udot v22.4s, v29.16b, v31.4b[0]\n"
@@ -2475,22 +2475,22 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr b3, [x10, #0x0]\n"
       "ldr b4, [x9, #0x0]\n"
       "159:"  // Height 5: Multiply loop: Ragged operand read: Done
-      "ldr q29, [x17, #0x0]\n"
-      "ldr q28, [x17, #0x10]\n"
+      "ldr q29, [x16, #0x0]\n"
       ".inst 0x6f80e3a8  // udot v8.4s, v29.16b, v0.4b[0]\n"
+      "ldr q28, [x16, #0x10]\n"
       ".inst 0x6f81e3ac  // udot v12.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x6f82e3b0  // udot v16.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x6f83e3b4  // udot v20.4s, v29.16b, v3.4b[0]\n"
       ".inst 0x6f84e3b8  // udot v24.4s, v29.16b, v4.4b[0]\n"
-      "ldr q29, [x17, #0x20]\n"
+      "ldr q29, [x16, #0x20]\n"
       ".inst 0x6f80e389  // udot v9.4s, v28.16b, v0.4b[0]\n"
       ".inst 0x6f81e38d  // udot v13.4s, v28.16b, v1.4b[0]\n"
       ".inst 0x6f82e391  // udot v17.4s, v28.16b, v2.4b[0]\n"
       ".inst 0x6f83e395  // udot v21.4s, v28.16b, v3.4b[0]\n"
       ".inst 0x6f84e399  // udot v25.4s, v28.16b, v4.4b[0]\n"
-      "ldr q28, [x17, #0x30]\n"
+      "ldr q28, [x16, #0x30]\n"
       ".inst 0x6f80e3aa  // udot v10.4s, v29.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f81e3ae  // udot v14.4s, v29.16b, v1.4b[0]\n"
       ".inst 0x6f82e3b2  // udot v18.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x6f83e3b6  // udot v22.4s, v29.16b, v3.4b[0]\n"
@@ -2506,20 +2506,20 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 150b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "add x21, x22, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
       "prfm pstl1keep, [x21, #0x0]\n"
       "bge 169f\n"
       "tbz x8, #3, 164f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
@@ -2529,19 +2529,19 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "st1 { v24.4s }, [x21], #0x10\n"
       "st1 { v25.4s }, [x21], #0x10\n"
       "tbz x8, #2, 162f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "st1 { v18.4s }, [x23], #0x10\n"
       "st1 { v22.4s }, [x22], #0x10\n"
       "st1 { v26.4s }, [x21], #0x10\n"
       "tbz x8, #1, 161f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "str d19, [x23], #0x8\n"
       "str d23, [x22], #0x8\n"
       "str d27, [x21], #0x8\n"
       "tbz x8, #0, 168f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "st1 { v19.s }[2], [x23]\n"
       "st1 { v23.s }[2], [x22]\n"
@@ -2549,7 +2549,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 168f\n"
       "161:"  // Height 5: Partial direct writeback: partial_1_12
       "tbz x8, #0, 168f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "str s19, [x23, #0x0]\n"
       "str s23, [x22, #0x0]\n"
@@ -2557,13 +2557,13 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 168f\n"
       "162:"  // Height 5: Partial direct writeback: partial_2_8
       "tbz x8, #1, 163f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "str d18, [x23], #0x8\n"
       "str d22, [x22], #0x8\n"
       "str d26, [x21], #0x8\n"
       "tbz x8, #0, 168f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "st1 { v18.s }[2], [x23]\n"
       "st1 { v22.s }[2], [x22]\n"
@@ -2571,7 +2571,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 168f\n"
       "163:"  // Height 5: Partial direct writeback: partial_1_8
       "tbz x8, #0, 168f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "str s18, [x23, #0x0]\n"
       "str s22, [x22, #0x0]\n"
@@ -2579,19 +2579,19 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 168f\n"
       "164:"  // Height 5: Partial direct writeback: partial_4_0
       "tbz x8, #2, 166f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "st1 { v20.4s }, [x22], #0x10\n"
       "st1 { v24.4s }, [x21], #0x10\n"
       "tbz x8, #1, 165f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "str d17, [x23], #0x8\n"
       "str d21, [x22], #0x8\n"
       "str d25, [x21], #0x8\n"
       "tbz x8, #0, 168f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "st1 { v17.s }[2], [x23]\n"
       "st1 { v21.s }[2], [x22]\n"
@@ -2599,7 +2599,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 168f\n"
       "165:"  // Height 5: Partial direct writeback: partial_1_4
       "tbz x8, #0, 168f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "str s17, [x23, #0x0]\n"
       "str s21, [x22, #0x0]\n"
@@ -2607,20 +2607,20 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 168f\n"
       "166:"  // Height 5: Partial direct writeback: partial_2_0
       "tbz x8, #1, 167f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "str d16, [x23], #0x8\n"
       "str d20, [x22], #0x8\n"
       "str d24, [x21], #0x8\n"
       "tbz x8, #0, 168f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "st1 { v16.s }[2], [x23]\n"
       "st1 { v20.s }[2], [x22]\n"
       "st1 { v24.s }[2], [x21]\n"
       "b 168f\n"
       "167:"  // Height 5: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "str s16, [x23, #0x0]\n"
       "str s20, [x22, #0x0]\n"
@@ -2628,11 +2628,11 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "168:"  // Height 5: Partial direct writeback: Done
       "b 170f\n"
       "169:"  // Height 5: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -2656,43 +2656,42 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "171:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "mov x20, #0x18\n"
-      "ldr x16, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x16\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x17, %x[output_ptr]\n"
+      "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "172:"  // Height 6: Column loop
       "tbz %x[flags], #0, 182f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
       "add x20, x21, x20, LSL #2\n"
       "bge 181f\n"
       "tbz x8, #3, 176f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v24.4s }, [x21], #0x10\n"
       "ld1 { v28.4s }, [x20], #0x10\n"
-      "ld1 { v9.4s }, [x16], #0x10\n"
+      "ld1 { v9.4s }, [x17], #0x10\n"
       "ld1 { v13.4s }, [x24], #0x10\n"
       "ld1 { v17.4s }, [x23], #0x10\n"
       "ld1 { v21.4s }, [x22], #0x10\n"
       "ld1 { v25.4s }, [x21], #0x10\n"
       "ld1 { v29.4s }, [x20], #0x10\n"
       "tbz x8, #2, 174f\n"
-      "ld1 { v10.4s }, [x16], #0x10\n"
+      "ld1 { v10.4s }, [x17], #0x10\n"
       "ld1 { v14.4s }, [x24], #0x10\n"
       "ld1 { v18.4s }, [x23], #0x10\n"
       "ld1 { v22.4s }, [x22], #0x10\n"
       "ld1 { v26.4s }, [x21], #0x10\n"
       "ld1 { v30.4s }, [x20], #0x10\n"
       "tbz x8, #1, 173f\n"
-      "ldr d11, [x16], #0x8\n"
+      "ldr d11, [x17], #0x8\n"
       "mov x25, #0x38\n"
       "ldr d15, [x24], #0x8\n"
       "ldr d19, [x23], #0x8\n"
@@ -2700,7 +2699,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr d27, [x21], #0x8\n"
       "ldr d31, [x20], #0x8\n"
       "tbz x8, #0, 180f\n"
-      "ld1 { v11.s }[2], [x16]\n"
+      "ld1 { v11.s }[2], [x17]\n"
       "ld1 { v15.s }[2], [x24]\n"
       "ld1 { v19.s }[2], [x23]\n"
       "ld1 { v23.s }[2], [x22]\n"
@@ -2710,7 +2709,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "173:"  // Height 6: Partial accumulate: partial_1_12
       "mov x25, #0x30\n"
       "tbz x8, #0, 180f\n"
-      "ldr s11, [x16, #0x0]\n"
+      "ldr s11, [x17, #0x0]\n"
       "ldr s15, [x24, #0x0]\n"
       "ldr s19, [x23, #0x0]\n"
       "ldr s23, [x22, #0x0]\n"
@@ -2719,7 +2718,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 180f\n"
       "174:"  // Height 6: Partial accumulate: partial_2_8
       "tbz x8, #1, 175f\n"
-      "ldr d10, [x16], #0x8\n"
+      "ldr d10, [x17], #0x8\n"
       "mov x25, #0x28\n"
       "ldr d14, [x24], #0x8\n"
       "ldr d18, [x23], #0x8\n"
@@ -2727,7 +2726,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr d26, [x21], #0x8\n"
       "ldr d30, [x20], #0x8\n"
       "tbz x8, #0, 180f\n"
-      "ld1 { v10.s }[2], [x16]\n"
+      "ld1 { v10.s }[2], [x17]\n"
       "ld1 { v14.s }[2], [x24]\n"
       "ld1 { v18.s }[2], [x23]\n"
       "ld1 { v22.s }[2], [x22]\n"
@@ -2737,7 +2736,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "175:"  // Height 6: Partial accumulate: partial_1_8
       "mov x25, #0x20\n"
       "tbz x8, #0, 180f\n"
-      "ldr s10, [x16, #0x0]\n"
+      "ldr s10, [x17, #0x0]\n"
       "ldr s14, [x24, #0x0]\n"
       "ldr s18, [x23, #0x0]\n"
       "ldr s22, [x22, #0x0]\n"
@@ -2746,14 +2745,14 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 180f\n"
       "176:"  // Height 6: Partial accumulate: partial_4_0
       "tbz x8, #2, 178f\n"
-      "ld1 { v8.4s }, [x16], #0x10\n"
+      "ld1 { v8.4s }, [x17], #0x10\n"
       "ld1 { v12.4s }, [x24], #0x10\n"
       "ld1 { v16.4s }, [x23], #0x10\n"
       "ld1 { v20.4s }, [x22], #0x10\n"
       "ld1 { v24.4s }, [x21], #0x10\n"
       "ld1 { v28.4s }, [x20], #0x10\n"
       "tbz x8, #1, 177f\n"
-      "ldr d9, [x16], #0x8\n"
+      "ldr d9, [x17], #0x8\n"
       "mov x25, #0x18\n"
       "ldr d13, [x24], #0x8\n"
       "ldr d17, [x23], #0x8\n"
@@ -2761,7 +2760,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr d25, [x21], #0x8\n"
       "ldr d29, [x20], #0x8\n"
       "tbz x8, #0, 180f\n"
-      "ld1 { v9.s }[2], [x16]\n"
+      "ld1 { v9.s }[2], [x17]\n"
       "ld1 { v13.s }[2], [x24]\n"
       "ld1 { v17.s }[2], [x23]\n"
       "ld1 { v21.s }[2], [x22]\n"
@@ -2771,7 +2770,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "177:"  // Height 6: Partial accumulate: partial_1_4
       "mov x25, #0x10\n"
       "tbz x8, #0, 180f\n"
-      "ldr s9, [x16, #0x0]\n"
+      "ldr s9, [x17, #0x0]\n"
       "ldr s13, [x24, #0x0]\n"
       "ldr s17, [x23, #0x0]\n"
       "ldr s21, [x22, #0x0]\n"
@@ -2780,7 +2779,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 180f\n"
       "178:"  // Height 6: Partial accumulate: partial_2_0
       "tbz x8, #1, 179f\n"
-      "ldr d8, [x16], #0x8\n"
+      "ldr d8, [x17], #0x8\n"
       "mov x25, #0x8\n"
       "ldr d12, [x24], #0x8\n"
       "ldr d16, [x23], #0x8\n"
@@ -2788,7 +2787,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr d24, [x21], #0x8\n"
       "ldr d28, [x20], #0x8\n"
       "tbz x8, #0, 180f\n"
-      "ld1 { v8.s }[2], [x16]\n"
+      "ld1 { v8.s }[2], [x17]\n"
       "ld1 { v12.s }[2], [x24]\n"
       "ld1 { v16.s }[2], [x23]\n"
       "ld1 { v20.s }[2], [x22]\n"
@@ -2796,7 +2795,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ld1 { v28.s }[2], [x20]\n"
       "b 180f\n"
       "179:"  // Height 6: Partial accumulate: partial_1_0
-      "ldr s8, [x16, #0x0]\n"
+      "ldr s8, [x17, #0x0]\n"
       "mov x25, #0x0\n"
       "ldr s12, [x24, #0x0]\n"
       "ldr s16, [x23, #0x0]\n"
@@ -2804,13 +2803,13 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr s24, [x21, #0x0]\n"
       "ldr s28, [x20, #0x0]\n"
       "180:"  // Height 6: Partial accumulate: Done
-      "sub x16, x16, x25\n"
+      "sub x17, x17, x25\n"
       "b 183f\n"
       "181:"  // Height 6: full accumulate
-      "ldr q8, [x16, #0x0]\n"
-      "ldr q9, [x16, #0x10]\n"
-      "ldr q10, [x16, #0x20]\n"
-      "ldr q11, [x16, #0x30]\n"
+      "ldr q8, [x17, #0x0]\n"
+      "ldr q9, [x17, #0x10]\n"
+      "ldr q10, [x17, #0x20]\n"
+      "ldr q11, [x17, #0x30]\n"
       "ldr q12, [x24, #0x0]\n"
       "ldr q13, [x24, #0x10]\n"
       "ldr q14, [x24, #0x20]\n"
@@ -2861,8 +2860,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "mov x15, #0x0\n"
       "184:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w14, [x20, x15, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 185f\n"
       "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2898,14 +2897,14 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr q3, [x10, #0x0]\n"
       "ldr q4, [x9, #0x0]\n"
       "ldr q5, [x28, #0x0]\n"
-      "ldr q6, [x17, #0x0]\n"
-      "ldr q7, [x17, #0x10]\n"
+      "ldr q6, [x16, #0x0]\n"
+      "ldr q7, [x16, #0x10]\n"
       "blt 188f\n"
       "187:"  // Height 6: Multiply loop: Main loop head
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
-      "ldr x21, [x17, #0x28]\n"
+      "ldr x21, [x16, #0x28]\n"
       ".inst 0x6f81e0cc  // udot v12.4s, v6.16b, v1.4b[0]\n"
-      "ldr x20, [x17, #0x38]\n"
+      "ldr x20, [x16, #0x38]\n"
       ".inst 0x6f82e0d0  // udot v16.4s, v6.16b, v2.4b[0]\n"
       "add x13, x13, #0x10\n"
       ".inst 0x6f83e0d4  // udot v20.4s, v6.16b, v3.4b[0]\n"
@@ -2913,151 +2912,151 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f84e0d8  // udot v24.4s, v6.16b, v4.4b[0]\n"
       "add x11, x11, #0x10\n"
       ".inst 0x6f85e0dc  // udot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr d6, [x17, #0x20]\n"
+      "ldr d6, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
-      "add x10, x10, #0x10\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
-      "add x9, x9, #0x10\n"
+      "ldr x21, [x16, #0x48]\n"
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
-      "mov v6.d[1], x21\n"
+      "add x10, x10, #0x10\n"
       ".inst 0x6f83e0f5  // udot v21.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x17, #0x48]\n"
+      "add x9, x9, #0x10\n"
       ".inst 0x6f84e0f9  // udot v25.4s, v7.16b, v4.4b[0]\n"
       "add x28, x28, #0x10\n"
       ".inst 0x6f85e0fd  // udot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr d7, [x17, #0x30]\n"
+      "ldr d7, [x16, #0x30]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x6f80e0ca  // udot v10.4s, v6.16b, v0.4b[0]\n"
-      "ldr x27, [x13, #0x8]\n"
       ".inst 0x6f81e0ce  // udot v14.4s, v6.16b, v1.4b[0]\n"
-      "ldr x26, [x12, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0x58]\n"
       ".inst 0x6f82e0d2  // udot v18.4s, v6.16b, v2.4b[0]\n"
+      "ldr x27, [x13, #0x8]\n"
       ".inst 0x6f83e0d6  // udot v22.4s, v6.16b, v3.4b[0]\n"
-      "ldr x20, [x17, #0x58]\n"
+      "ldr x26, [x12, #0x8]\n"
       ".inst 0x6f84e0da  // udot v26.4s, v6.16b, v4.4b[0]\n"
       "ldr x25, [x11, #0x8]\n"
       ".inst 0x6f85e0de  // udot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr d6, [x17, #0x40]\n"
+      "ldr d6, [x16, #0x40]\n"
       ".inst 0x6f80e0eb  // udot v11.4s, v7.16b, v0.4b[0]\n"
-      "ldr x24, [x10, #0x8]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x6f81e0ef  // udot v15.4s, v7.16b, v1.4b[0]\n"
-      "ldr x23, [x9, #0x8]\n"
+      "ldr x21, [x16, #0x68]\n"
       ".inst 0x6f82e0f3  // udot v19.4s, v7.16b, v2.4b[0]\n"
-      "mov v6.d[1], x21\n"
+      "ldr x24, [x10, #0x8]\n"
       ".inst 0x6f83e0f7  // udot v23.4s, v7.16b, v3.4b[0]\n"
-      "ldr x21, [x17, #0x68]\n"
+      "ldr x23, [x9, #0x8]\n"
       ".inst 0x6f84e0fb  // udot v27.4s, v7.16b, v4.4b[0]\n"
       "ldr x22, [x28, #0x8]\n"
       ".inst 0x6f85e0ff  // udot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr d7, [x17, #0x50]\n"
+      "ldr d7, [x16, #0x50]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x6fa0e0c8  // udot v8.4s, v6.16b, v0.4b[1]\n"
-      "sub x14, x14, #0x10\n"
       ".inst 0x6fa1e0cc  // udot v12.4s, v6.16b, v1.4b[1]\n"
-      "cmp x14, #0x20\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0x78]\n"
       ".inst 0x6fa2e0d0  // udot v16.4s, v6.16b, v2.4b[1]\n"
+      "sub x14, x14, #0x10\n"
       ".inst 0x6fa3e0d4  // udot v20.4s, v6.16b, v3.4b[1]\n"
-      "ldr x20, [x17, #0x78]\n"
+      "cmp x14, #0x20\n"
       ".inst 0x6fa4e0d8  // udot v24.4s, v6.16b, v4.4b[1]\n"
       "prfm pldl1keep, [x13, #0x80]\n"
       ".inst 0x6fa5e0dc  // udot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr d6, [x17, #0x60]\n"
+      "ldr d6, [x16, #0x60]\n"
       ".inst 0x6fa0e0e9  // udot v9.4s, v7.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x12, #0x80]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x6fa1e0ed  // udot v13.4s, v7.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x11, #0x80]\n"
+      "ldr x21, [x16, #0x88]\n"
       ".inst 0x6fa2e0f1  // udot v17.4s, v7.16b, v2.4b[1]\n"
-      "mov v6.d[1], x21\n"
+      "prfm pldl1keep, [x12, #0x80]\n"
       ".inst 0x6fa3e0f5  // udot v21.4s, v7.16b, v3.4b[1]\n"
-      "ldr x21, [x17, #0x88]\n"
+      "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x6fa4e0f9  // udot v25.4s, v7.16b, v4.4b[1]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x6fa5e0fd  // udot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr d7, [x17, #0x70]\n"
+      "ldr d7, [x16, #0x70]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x6fa0e0ca  // udot v10.4s, v6.16b, v0.4b[1]\n"
-      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x6fa1e0ce  // udot v14.4s, v6.16b, v1.4b[1]\n"
-      "prfm pldl1keep, [x28, #0x80]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0x98]\n"
       ".inst 0x6fa2e0d2  // udot v18.4s, v6.16b, v2.4b[1]\n"
+      "prfm pldl1keep, [x9, #0x80]\n"
       ".inst 0x6fa3e0d6  // udot v22.4s, v6.16b, v3.4b[1]\n"
-      "ldr x20, [x17, #0x98]\n"
+      "prfm pldl1keep, [x28, #0x80]\n"
       ".inst 0x6fa4e0da  // udot v26.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x6fa5e0de  // udot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr d6, [x17, #0x80]\n"
+      "ldr d6, [x16, #0x80]\n"
       ".inst 0x6fa0e0eb  // udot v11.4s, v7.16b, v0.4b[1]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x6fa1e0ef  // udot v15.4s, v7.16b, v1.4b[1]\n"
+      "ldr x21, [x16, #0xa8]\n"
       ".inst 0x6fa2e0f3  // udot v19.4s, v7.16b, v2.4b[1]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x6fa3e0f7  // udot v23.4s, v7.16b, v3.4b[1]\n"
-      "ldr x21, [x17, #0xa8]\n"
       ".inst 0x6fa4e0fb  // udot v27.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x6fa5e0ff  // udot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr d7, [x17, #0x90]\n"
+      "ldr d7, [x16, #0x90]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x6f80e8c8  // udot v8.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x6f81e8cc  // udot v12.4s, v6.16b, v1.4b[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0xb8]\n"
       ".inst 0x6f82e8d0  // udot v16.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x6f83e8d4  // udot v20.4s, v6.16b, v3.4b[2]\n"
-      "ldr x20, [x17, #0xb8]\n"
       ".inst 0x6f84e8d8  // udot v24.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x6f85e8dc  // udot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr d6, [x17, #0xa0]\n"
+      "ldr d6, [x16, #0xa0]\n"
       ".inst 0x6f80e8e9  // udot v9.4s, v7.16b, v0.4b[2]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x6f81e8ed  // udot v13.4s, v7.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xc8]\n"
       ".inst 0x6f82e8f1  // udot v17.4s, v7.16b, v2.4b[2]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x6f83e8f5  // udot v21.4s, v7.16b, v3.4b[2]\n"
-      "ldr x21, [x17, #0xc8]\n"
       ".inst 0x6f84e8f9  // udot v25.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x6f85e8fd  // udot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr d7, [x17, #0xb0]\n"
+      "ldr d7, [x16, #0xb0]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x6f80e8ca  // udot v10.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x6f81e8ce  // udot v14.4s, v6.16b, v1.4b[2]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0xd8]\n"
       ".inst 0x6f82e8d2  // udot v18.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x6f83e8d6  // udot v22.4s, v6.16b, v3.4b[2]\n"
-      "ldr x20, [x17, #0xd8]\n"
       ".inst 0x6f84e8da  // udot v26.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x6f85e8de  // udot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr d6, [x17, #0xc0]\n"
+      "ldr d6, [x16, #0xc0]\n"
       ".inst 0x6f80e8eb  // udot v11.4s, v7.16b, v0.4b[2]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x6f81e8ef  // udot v15.4s, v7.16b, v1.4b[2]\n"
+      "ldr x21, [x16, #0xe8]\n"
       ".inst 0x6f82e8f3  // udot v19.4s, v7.16b, v2.4b[2]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x6f83e8f7  // udot v23.4s, v7.16b, v3.4b[2]\n"
-      "ldr x21, [x17, #0xe8]\n"
       ".inst 0x6f84e8fb  // udot v27.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x6f85e8ff  // udot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr d7, [x17, #0xd0]\n"
+      "ldr d7, [x16, #0xd0]\n"
+      "mov v7.d[1], x20\n"
       ".inst 0x6fa0e8c8  // udot v8.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x6fa1e8cc  // udot v12.4s, v6.16b, v1.4b[3]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0xf8]\n"
       ".inst 0x6fa2e8d0  // udot v16.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x6fa3e8d4  // udot v20.4s, v6.16b, v3.4b[3]\n"
-      "ldr x20, [x17, #0xf8]\n"
       ".inst 0x6fa4e8d8  // udot v24.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x6fa5e8dc  // udot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr d6, [x17, #0xe0]\n"
+      "ldr d6, [x16, #0xe0]\n"
       ".inst 0x6fa0e8e9  // udot v9.4s, v7.16b, v0.4b[3]\n"
+      "mov v6.d[1], x21\n"
       ".inst 0x6fa1e8ed  // udot v13.4s, v7.16b, v1.4b[3]\n"
       ".inst 0x6fa2e8f1  // udot v17.4s, v7.16b, v2.4b[3]\n"
-      "mov v6.d[1], x21\n"
       ".inst 0x6fa3e8f5  // udot v21.4s, v7.16b, v3.4b[3]\n"
       ".inst 0x6fa4e8f9  // udot v25.4s, v7.16b, v4.4b[3]\n"
       ".inst 0x6fa5e8fd  // udot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr d7, [x17, #0xf0]\n"
-      "add x17, x17, #0x100\n"
+      "ldr d7, [x16, #0xf0]\n"
+      "mov v7.d[1], x20\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa0e8ca  // udot v10.4s, v6.16b, v0.4b[3]\n"
+      "ldr x21, [x16, #0x8]\n"
       ".inst 0x6fa1e8ce  // udot v14.4s, v6.16b, v1.4b[3]\n"
-      "ldr x21, [x17, #0x8]\n"
-      "mov v7.d[1], x20\n"
+      "ldr x20, [x16, #0x18]\n"
       ".inst 0x6fa2e8d2  // udot v18.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x6fa3e8d6  // udot v22.4s, v6.16b, v3.4b[3]\n"
-      "ldr x20, [x17, #0x18]\n"
       ".inst 0x6fa4e8da  // udot v26.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x6fa5e8de  // udot v30.4s, v6.16b, v5.4b[3]\n"
-      "ldr d6, [x17, #0x0]\n"
+      "ldr d6, [x16, #0x0]\n"
       ".inst 0x6fa0e8eb  // udot v11.4s, v7.16b, v0.4b[3]\n"
       "ldr d0, [x13, #0x0]\n"
       ".inst 0x6fa1e8ef  // udot v15.4s, v7.16b, v1.4b[3]\n"
@@ -3070,7 +3069,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr d4, [x9, #0x0]\n"
       ".inst 0x6fa5e8ff  // udot v31.4s, v7.16b, v5.4b[3]\n"
       "ldr d5, [x28, #0x0]\n"
-      "ldr d7, [x17, #0x10]\n"
+      "ldr d7, [x16, #0x10]\n"
       "mov v6.d[1], x21\n"
       "mov v0.d[1], x27\n"
       "mov v1.d[1], x26\n"
@@ -3092,7 +3091,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f84e0d8  // udot v24.4s, v6.16b, v4.4b[0]\n"
       "add x9, x9, #0x10\n"
       ".inst 0x6f85e0dc  // udot v28.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x17, #0x20]\n"
+      "ldr q6, [x16, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       "add x28, x28, #0x10\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
@@ -3104,7 +3103,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f84e0f9  // udot v25.4s, v7.16b, v4.4b[0]\n"
       "prfm pldl1keep, [x11, #0x80]\n"
       ".inst 0x6f85e0fd  // udot v29.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x17, #0x30]\n"
+      "ldr q7, [x16, #0x30]\n"
       ".inst 0x6f80e0ca  // udot v10.4s, v6.16b, v0.4b[0]\n"
       "prfm pldl1keep, [x10, #0x80]\n"
       ".inst 0x6f81e0ce  // udot v14.4s, v6.16b, v1.4b[0]\n"
@@ -3114,86 +3113,86 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       ".inst 0x6f83e0d6  // udot v22.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x6f84e0da  // udot v26.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x6f85e0de  // udot v30.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x17, #0x40]\n"
+      "ldr q6, [x16, #0x40]\n"
       ".inst 0x6f80e0eb  // udot v11.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x6f81e0ef  // udot v15.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x6f82e0f3  // udot v19.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x6f83e0f7  // udot v23.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x6f84e0fb  // udot v27.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x6f85e0ff  // udot v31.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x17, #0x50]\n"
+      "ldr q7, [x16, #0x50]\n"
       ".inst 0x6fa0e0c8  // udot v8.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x6fa1e0cc  // udot v12.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x6fa2e0d0  // udot v16.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x6fa3e0d4  // udot v20.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x6fa4e0d8  // udot v24.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x6fa5e0dc  // udot v28.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x17, #0x60]\n"
+      "ldr q6, [x16, #0x60]\n"
       ".inst 0x6fa0e0e9  // udot v9.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x6fa1e0ed  // udot v13.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x6fa2e0f1  // udot v17.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x6fa3e0f5  // udot v21.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x6fa4e0f9  // udot v25.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x6fa5e0fd  // udot v29.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x17, #0x70]\n"
+      "ldr q7, [x16, #0x70]\n"
       ".inst 0x6fa0e0ca  // udot v10.4s, v6.16b, v0.4b[1]\n"
       ".inst 0x6fa1e0ce  // udot v14.4s, v6.16b, v1.4b[1]\n"
       ".inst 0x6fa2e0d2  // udot v18.4s, v6.16b, v2.4b[1]\n"
       ".inst 0x6fa3e0d6  // udot v22.4s, v6.16b, v3.4b[1]\n"
       ".inst 0x6fa4e0da  // udot v26.4s, v6.16b, v4.4b[1]\n"
       ".inst 0x6fa5e0de  // udot v30.4s, v6.16b, v5.4b[1]\n"
-      "ldr q6, [x17, #0x80]\n"
+      "ldr q6, [x16, #0x80]\n"
       ".inst 0x6fa0e0eb  // udot v11.4s, v7.16b, v0.4b[1]\n"
       ".inst 0x6fa1e0ef  // udot v15.4s, v7.16b, v1.4b[1]\n"
       ".inst 0x6fa2e0f3  // udot v19.4s, v7.16b, v2.4b[1]\n"
       ".inst 0x6fa3e0f7  // udot v23.4s, v7.16b, v3.4b[1]\n"
       ".inst 0x6fa4e0fb  // udot v27.4s, v7.16b, v4.4b[1]\n"
       ".inst 0x6fa5e0ff  // udot v31.4s, v7.16b, v5.4b[1]\n"
-      "ldr q7, [x17, #0x90]\n"
+      "ldr q7, [x16, #0x90]\n"
       ".inst 0x6f80e8c8  // udot v8.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x6f81e8cc  // udot v12.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x6f82e8d0  // udot v16.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x6f83e8d4  // udot v20.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x6f84e8d8  // udot v24.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x6f85e8dc  // udot v28.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x17, #0xa0]\n"
+      "ldr q6, [x16, #0xa0]\n"
       ".inst 0x6f80e8e9  // udot v9.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x6f81e8ed  // udot v13.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x6f82e8f1  // udot v17.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x6f83e8f5  // udot v21.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x6f84e8f9  // udot v25.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x6f85e8fd  // udot v29.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x17, #0xb0]\n"
+      "ldr q7, [x16, #0xb0]\n"
       ".inst 0x6f80e8ca  // udot v10.4s, v6.16b, v0.4b[2]\n"
       ".inst 0x6f81e8ce  // udot v14.4s, v6.16b, v1.4b[2]\n"
       ".inst 0x6f82e8d2  // udot v18.4s, v6.16b, v2.4b[2]\n"
       ".inst 0x6f83e8d6  // udot v22.4s, v6.16b, v3.4b[2]\n"
       ".inst 0x6f84e8da  // udot v26.4s, v6.16b, v4.4b[2]\n"
       ".inst 0x6f85e8de  // udot v30.4s, v6.16b, v5.4b[2]\n"
-      "ldr q6, [x17, #0xc0]\n"
+      "ldr q6, [x16, #0xc0]\n"
       ".inst 0x6f80e8eb  // udot v11.4s, v7.16b, v0.4b[2]\n"
       ".inst 0x6f81e8ef  // udot v15.4s, v7.16b, v1.4b[2]\n"
       ".inst 0x6f82e8f3  // udot v19.4s, v7.16b, v2.4b[2]\n"
       ".inst 0x6f83e8f7  // udot v23.4s, v7.16b, v3.4b[2]\n"
       ".inst 0x6f84e8fb  // udot v27.4s, v7.16b, v4.4b[2]\n"
       ".inst 0x6f85e8ff  // udot v31.4s, v7.16b, v5.4b[2]\n"
-      "ldr q7, [x17, #0xd0]\n"
+      "ldr q7, [x16, #0xd0]\n"
       ".inst 0x6fa0e8c8  // udot v8.4s, v6.16b, v0.4b[3]\n"
       ".inst 0x6fa1e8cc  // udot v12.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x6fa2e8d0  // udot v16.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x6fa3e8d4  // udot v20.4s, v6.16b, v3.4b[3]\n"
       ".inst 0x6fa4e8d8  // udot v24.4s, v6.16b, v4.4b[3]\n"
       ".inst 0x6fa5e8dc  // udot v28.4s, v6.16b, v5.4b[3]\n"
-      "ldr q6, [x17, #0xe0]\n"
+      "ldr q6, [x16, #0xe0]\n"
       ".inst 0x6fa0e8e9  // udot v9.4s, v7.16b, v0.4b[3]\n"
       ".inst 0x6fa1e8ed  // udot v13.4s, v7.16b, v1.4b[3]\n"
       ".inst 0x6fa2e8f1  // udot v17.4s, v7.16b, v2.4b[3]\n"
       ".inst 0x6fa3e8f5  // udot v21.4s, v7.16b, v3.4b[3]\n"
       ".inst 0x6fa4e8f9  // udot v25.4s, v7.16b, v4.4b[3]\n"
       ".inst 0x6fa5e8fd  // udot v29.4s, v7.16b, v5.4b[3]\n"
-      "ldr q7, [x17, #0xf0]\n"
+      "ldr q7, [x16, #0xf0]\n"
       ".inst 0x6fa0e8ca  // udot v10.4s, v6.16b, v0.4b[3]\n"
-      "add x17, x17, #0x100\n"
+      "add x16, x16, #0x100\n"
       ".inst 0x6fa1e8ce  // udot v14.4s, v6.16b, v1.4b[3]\n"
       ".inst 0x6fa2e8d2  // udot v18.4s, v6.16b, v2.4b[3]\n"
       ".inst 0x6fa3e8d6  // udot v22.4s, v6.16b, v3.4b[3]\n"
@@ -3218,24 +3217,24 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr s4, [x10], #0x4\n"
       "ldr s3, [x9], #0x4\n"
       "ldr s2, [x28], #0x4\n"
-      "ldr q1, [x17, #0x0]\n"
-      "ldr q0, [x17, #0x10]\n"
+      "ldr q1, [x16, #0x0]\n"
       ".inst 0x6f87e028  // udot v8.4s, v1.16b, v7.4b[0]\n"
+      "ldr q0, [x16, #0x10]\n"
       ".inst 0x6f86e02c  // udot v12.4s, v1.16b, v6.4b[0]\n"
       ".inst 0x6f85e030  // udot v16.4s, v1.16b, v5.4b[0]\n"
       ".inst 0x6f84e034  // udot v20.4s, v1.16b, v4.4b[0]\n"
       ".inst 0x6f83e038  // udot v24.4s, v1.16b, v3.4b[0]\n"
       ".inst 0x6f82e03c  // udot v28.4s, v1.16b, v2.4b[0]\n"
-      "ldr q1, [x17, #0x20]\n"
+      "ldr q1, [x16, #0x20]\n"
       ".inst 0x6f87e009  // udot v9.4s, v0.16b, v7.4b[0]\n"
       ".inst 0x6f86e00d  // udot v13.4s, v0.16b, v6.4b[0]\n"
       ".inst 0x6f85e011  // udot v17.4s, v0.16b, v5.4b[0]\n"
       ".inst 0x6f84e015  // udot v21.4s, v0.16b, v4.4b[0]\n"
       ".inst 0x6f83e019  // udot v25.4s, v0.16b, v3.4b[0]\n"
       ".inst 0x6f82e01d  // udot v29.4s, v0.16b, v2.4b[0]\n"
-      "ldr q0, [x17, #0x30]\n"
+      "ldr q0, [x16, #0x30]\n"
       ".inst 0x6f87e02a  // udot v10.4s, v1.16b, v7.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f86e02e  // udot v14.4s, v1.16b, v6.4b[0]\n"
       ".inst 0x6f85e032  // udot v18.4s, v1.16b, v5.4b[0]\n"
       ".inst 0x6f84e036  // udot v22.4s, v1.16b, v4.4b[0]\n"
@@ -3273,24 +3272,24 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "ldr b4, [x9, #0x0]\n"
       "ldr b5, [x28, #0x0]\n"
       "193:"  // Height 6: Multiply loop: Ragged operand read: Done
-      "ldr q7, [x17, #0x0]\n"
-      "ldr q6, [x17, #0x10]\n"
+      "ldr q7, [x16, #0x0]\n"
       ".inst 0x6f80e0e8  // udot v8.4s, v7.16b, v0.4b[0]\n"
+      "ldr q6, [x16, #0x10]\n"
       ".inst 0x6f81e0ec  // udot v12.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x6f82e0f0  // udot v16.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x6f83e0f4  // udot v20.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x6f84e0f8  // udot v24.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x6f85e0fc  // udot v28.4s, v7.16b, v5.4b[0]\n"
-      "ldr q7, [x17, #0x20]\n"
+      "ldr q7, [x16, #0x20]\n"
       ".inst 0x6f80e0c9  // udot v9.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x6f81e0cd  // udot v13.4s, v6.16b, v1.4b[0]\n"
       ".inst 0x6f82e0d1  // udot v17.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x6f83e0d5  // udot v21.4s, v6.16b, v3.4b[0]\n"
       ".inst 0x6f84e0d9  // udot v25.4s, v6.16b, v4.4b[0]\n"
       ".inst 0x6f85e0dd  // udot v29.4s, v6.16b, v5.4b[0]\n"
-      "ldr q6, [x17, #0x30]\n"
+      "ldr q6, [x16, #0x30]\n"
       ".inst 0x6f80e0ea  // udot v10.4s, v7.16b, v0.4b[0]\n"
-      "add x17, x17, #0x40\n"
+      "add x16, x16, #0x40\n"
       ".inst 0x6f81e0ee  // udot v14.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x6f82e0f2  // udot v18.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x6f83e0f6  // udot v22.4s, v7.16b, v3.4b[0]\n"
@@ -3308,22 +3307,22 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "cmp x15, x20\n"
       "bne 184b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x8, #0x10\n"
-      "prfm pstl1keep, [x16, #0x0]\n"
-      "add x24, x16, x20, LSL #2\n"
+      "add x24, x17, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "add x21, x22, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "add x20, x21, x20, LSL #2\n"
+      "cmp x8, #0x10\n"
+      "prfm pstl1keep, [x17, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "prfm pstl1keep, [x21, #0x0]\n"
       "prfm pstl1keep, [x20, #0x0]\n"
       "bge 203f\n"
       "tbz x8, #3, 198f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
-      "st1 { v9.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
+      "st1 { v9.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v13.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
@@ -3335,21 +3334,21 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "st1 { v28.4s }, [x20], #0x10\n"
       "st1 { v29.4s }, [x20], #0x10\n"
       "tbz x8, #2, 196f\n"
-      "st1 { v10.4s }, [x16], #0x10\n"
+      "st1 { v10.4s }, [x17], #0x10\n"
       "st1 { v14.4s }, [x24], #0x10\n"
       "st1 { v18.4s }, [x23], #0x10\n"
       "st1 { v22.4s }, [x22], #0x10\n"
       "st1 { v26.4s }, [x21], #0x10\n"
       "st1 { v30.4s }, [x20], #0x10\n"
       "tbz x8, #1, 195f\n"
-      "str d11, [x16], #0x8\n"
+      "str d11, [x17], #0x8\n"
       "str d15, [x24], #0x8\n"
       "str d19, [x23], #0x8\n"
       "str d23, [x22], #0x8\n"
       "str d27, [x21], #0x8\n"
       "str d31, [x20], #0x8\n"
       "tbz x8, #0, 202f\n"
-      "st1 { v11.s }[2], [x16]\n"
+      "st1 { v11.s }[2], [x17]\n"
       "st1 { v15.s }[2], [x24]\n"
       "st1 { v19.s }[2], [x23]\n"
       "st1 { v23.s }[2], [x22]\n"
@@ -3358,7 +3357,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 202f\n"
       "195:"  // Height 6: Partial direct writeback: partial_1_12
       "tbz x8, #0, 202f\n"
-      "str s11, [x16, #0x0]\n"
+      "str s11, [x17, #0x0]\n"
       "str s15, [x24, #0x0]\n"
       "str s19, [x23, #0x0]\n"
       "str s23, [x22, #0x0]\n"
@@ -3367,14 +3366,14 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 202f\n"
       "196:"  // Height 6: Partial direct writeback: partial_2_8
       "tbz x8, #1, 197f\n"
-      "str d10, [x16], #0x8\n"
+      "str d10, [x17], #0x8\n"
       "str d14, [x24], #0x8\n"
       "str d18, [x23], #0x8\n"
       "str d22, [x22], #0x8\n"
       "str d26, [x21], #0x8\n"
       "str d30, [x20], #0x8\n"
       "tbz x8, #0, 202f\n"
-      "st1 { v10.s }[2], [x16]\n"
+      "st1 { v10.s }[2], [x17]\n"
       "st1 { v14.s }[2], [x24]\n"
       "st1 { v18.s }[2], [x23]\n"
       "st1 { v22.s }[2], [x22]\n"
@@ -3383,7 +3382,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 202f\n"
       "197:"  // Height 6: Partial direct writeback: partial_1_8
       "tbz x8, #0, 202f\n"
-      "str s10, [x16, #0x0]\n"
+      "str s10, [x17, #0x0]\n"
       "str s14, [x24, #0x0]\n"
       "str s18, [x23, #0x0]\n"
       "str s22, [x22, #0x0]\n"
@@ -3392,21 +3391,21 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 202f\n"
       "198:"  // Height 6: Partial direct writeback: partial_4_0
       "tbz x8, #2, 200f\n"
-      "st1 { v8.4s }, [x16], #0x10\n"
+      "st1 { v8.4s }, [x17], #0x10\n"
       "st1 { v12.4s }, [x24], #0x10\n"
       "st1 { v16.4s }, [x23], #0x10\n"
       "st1 { v20.4s }, [x22], #0x10\n"
       "st1 { v24.4s }, [x21], #0x10\n"
       "st1 { v28.4s }, [x20], #0x10\n"
       "tbz x8, #1, 199f\n"
-      "str d9, [x16], #0x8\n"
+      "str d9, [x17], #0x8\n"
       "str d13, [x24], #0x8\n"
       "str d17, [x23], #0x8\n"
       "str d21, [x22], #0x8\n"
       "str d25, [x21], #0x8\n"
       "str d29, [x20], #0x8\n"
       "tbz x8, #0, 202f\n"
-      "st1 { v9.s }[2], [x16]\n"
+      "st1 { v9.s }[2], [x17]\n"
       "st1 { v13.s }[2], [x24]\n"
       "st1 { v17.s }[2], [x23]\n"
       "st1 { v21.s }[2], [x22]\n"
@@ -3415,7 +3414,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 202f\n"
       "199:"  // Height 6: Partial direct writeback: partial_1_4
       "tbz x8, #0, 202f\n"
-      "str s9, [x16, #0x0]\n"
+      "str s9, [x17, #0x0]\n"
       "str s13, [x24, #0x0]\n"
       "str s17, [x23, #0x0]\n"
       "str s21, [x22, #0x0]\n"
@@ -3424,14 +3423,14 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "b 202f\n"
       "200:"  // Height 6: Partial direct writeback: partial_2_0
       "tbz x8, #1, 201f\n"
-      "str d8, [x16], #0x8\n"
+      "str d8, [x17], #0x8\n"
       "str d12, [x24], #0x8\n"
       "str d16, [x23], #0x8\n"
       "str d20, [x22], #0x8\n"
       "str d24, [x21], #0x8\n"
       "str d28, [x20], #0x8\n"
       "tbz x8, #0, 202f\n"
-      "st1 { v8.s }[2], [x16]\n"
+      "st1 { v8.s }[2], [x17]\n"
       "st1 { v12.s }[2], [x24]\n"
       "st1 { v16.s }[2], [x23]\n"
       "st1 { v20.s }[2], [x22]\n"
@@ -3439,7 +3438,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "st1 { v28.s }[2], [x20]\n"
       "b 202f\n"
       "201:"  // Height 6: Partial direct writeback: partial_1_0
-      "str s8, [x16, #0x0]\n"
+      "str s8, [x17, #0x0]\n"
       "str s12, [x24, #0x0]\n"
       "str s16, [x23, #0x0]\n"
       "str s20, [x22, #0x0]\n"
@@ -3448,11 +3447,11 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "202:"  // Height 6: Partial direct writeback: Done
       "b 204f\n"
       "203:"  // Height 6: Full writeback
-      "str q8, [x16, #0x0]\n"
-      "str q9, [x16, #0x10]\n"
-      "str q10, [x16, #0x20]\n"
-      "str q11, [x16, #0x30]\n"
-      "add x16, x16, #0x40\n"
+      "str q8, [x17, #0x0]\n"
+      "str q9, [x17, #0x10]\n"
+      "str q10, [x17, #0x20]\n"
+      "str q11, [x17, #0x30]\n"
+      "add x17, x17, #0x40\n"
       "str q12, [x24, #0x0]\n"
       "str q13, [x24, #0x10]\n"
       "str q14, [x24, #0x20]\n"
@@ -3488,8 +3487,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "206:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp
index 8b7f5afb7e..d3367b959a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp
@@ -44,18 +44,18 @@ void a64_hybrid_u8u32_dot_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -88,7 +88,7 @@ void a64_hybrid_u8u32_dot_6x16 (
       "beq 35f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "tbz %x[flags], #0, 12f\n"
       "cmp x11, #0x10\n"
@@ -163,8 +163,8 @@ void a64_hybrid_u8u32_dot_6x16 (
       "mov x28, #0x0\n"
       "14:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 15f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -188,10 +188,6 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
       "ldr q17, [x10, #0x40]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
@@ -216,21 +212,22 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr q17, [x10, #0xe0]\n"
       ".inst 0x6fa0ea09  // udot v9.4s, v16.16b, v0.4b[3]\n"
       "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6fa0ea2a  // udot v10.4s, v17.16b, v0.4b[3]\n"
-      "ldr q6, [x10, #0x0]\n"
       ".inst 0x6fa0ea0b  // udot v11.4s, v16.16b, v0.4b[3]\n"
       "ldr q0, [x26, #0x0]\n"
+      "cmp x27, #0x20\n"
+      "add x10, x10, #0x100\n"
+      "ldr q6, [x10, #0x0]\n"
       "ldr q7, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 17b\n"
       "18:"  // Height 1: Multiply loop: Single iteration only
       ".inst 0x6f80e0c8  // udot v8.4s, v6.16b, v0.4b[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
       "ldr q17, [x10, #0x40]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
@@ -255,26 +252,29 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr q17, [x10, #0xe0]\n"
       ".inst 0x6fa0ea09  // udot v9.4s, v16.16b, v0.4b[3]\n"
       "ldr q16, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6fa0ea2a  // udot v10.4s, v17.16b, v0.4b[3]\n"
       ".inst 0x6fa0ea0b  // udot v11.4s, v16.16b, v0.4b[3]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "19:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 24f\n"
       "cmp x27, #0x4\n"
       "blt 21f\n"
       "20:"  // Height 1: Multiply loop: Odd block loop
       "ldr s18, [x26], #0x4\n"
-      "ldr q17, [x10, #0x0]\n"
+      "ldr q16, [x10, #0x0]\n"
+      ".inst 0x6f92e208  // udot v8.4s, v16.16b, v18.4b[0]\n"
       "sub x27, x27, #0x4\n"
       "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
-      ".inst 0x6f92e228  // udot v8.4s, v17.16b, v18.4b[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x6f92e209  // udot v9.4s, v16.16b, v18.4b[0]\n"
+      "cmp x27, #0x4\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6f92e22a  // udot v10.4s, v17.16b, v18.4b[0]\n"
       ".inst 0x6f92e20b  // udot v11.4s, v16.16b, v18.4b[0]\n"
+      "add x10, x10, #0x40\n"
       "bge 20b\n"
       "21:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 24f\n"
@@ -289,12 +289,12 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr q17, [x10, #0x0]\n"
       "ldr q16, [x10, #0x10]\n"
       ".inst 0x6f80e228  // udot v8.4s, v17.16b, v0.4b[0]\n"
-      "ldr q17, [x10, #0x20]\n"
       ".inst 0x6f80e209  // udot v9.4s, v16.16b, v0.4b[0]\n"
+      "ldr q17, [x10, #0x20]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
+      "add x10, x10, #0x40\n"
       "24:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -362,7 +362,7 @@ void a64_hybrid_u8u32_dot_6x16 (
       "35:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "36:"  // Height 2: Column loop
       "tbz %x[flags], #0, 46f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
@@ -463,8 +463,8 @@ void a64_hybrid_u8u32_dot_6x16 (
       "mov x28, #0x0\n"
       "48:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 49f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -496,22 +496,22 @@ void a64_hybrid_u8u32_dot_6x16 (
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x6f81e22e  // udot v14.4s, v17.16b, v1.4b[0]\n"
       "ldr q17, [x10, #0x40]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x6f81e20f  // udot v15.4s, v16.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "cmp x27, #0x20\n"
       ".inst 0x6fa0e228  // udot v8.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x6fa1e22c  // udot v12.4s, v17.16b, v1.4b[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6fa0e209  // udot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x6fa1e20d  // udot v13.4s, v16.16b, v1.4b[1]\n"
       "ldr q16, [x10, #0x70]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6fa0e22a  // udot v10.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x6fa1e22e  // udot v14.4s, v17.16b, v1.4b[1]\n"
       "ldr q17, [x10, #0x80]\n"
@@ -555,18 +555,18 @@ void a64_hybrid_u8u32_dot_6x16 (
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
       "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x6f81e22e  // udot v14.4s, v17.16b, v1.4b[0]\n"
       "ldr q17, [x10, #0x40]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x6f81e20f  // udot v15.4s, v16.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x50]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6fa0e228  // udot v8.4s, v17.16b, v0.4b[1]\n"
       ".inst 0x6fa1e22c  // udot v12.4s, v17.16b, v1.4b[1]\n"
       "ldr q17, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6fa0e209  // udot v9.4s, v16.16b, v0.4b[1]\n"
       ".inst 0x6fa1e20d  // udot v13.4s, v16.16b, v1.4b[1]\n"
       "ldr q16, [x10, #0x70]\n"
@@ -607,18 +607,18 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr s19, [x26], #0x4\n"
       "ldr s18, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr q17, [x10, #0x0]\n"
       "ldr q16, [x10, #0x10]\n"
-      "cmp x27, #0x4\n"
       ".inst 0x6f93e228  // udot v8.4s, v17.16b, v19.4b[0]\n"
       ".inst 0x6f92e22c  // udot v12.4s, v17.16b, v18.4b[0]\n"
       "ldr q17, [x10, #0x20]\n"
       ".inst 0x6f93e209  // udot v9.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x6f92e20d  // udot v13.4s, v16.16b, v18.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6f93e22a  // udot v10.4s, v17.16b, v19.4b[0]\n"
       ".inst 0x6f92e22e  // udot v14.4s, v17.16b, v18.4b[0]\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6f93e20b  // udot v11.4s, v16.16b, v19.4b[0]\n"
       ".inst 0x6f92e20f  // udot v15.4s, v16.16b, v18.4b[0]\n"
       "bge 54b\n"
@@ -643,9 +643,9 @@ void a64_hybrid_u8u32_dot_6x16 (
       ".inst 0x6f80e209  // udot v9.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x6f81e20d  // udot v13.4s, v16.16b, v1.4b[0]\n"
       "ldr q16, [x10, #0x30]\n"
-      "add x10, x10, #0x40\n"
       ".inst 0x6f80e22a  // udot v10.4s, v17.16b, v0.4b[0]\n"
       ".inst 0x6f81e22e  // udot v14.4s, v17.16b, v1.4b[0]\n"
+      "add x10, x10, #0x40\n"
       ".inst 0x6f80e20b  // udot v11.4s, v16.16b, v0.4b[0]\n"
       ".inst 0x6f81e20f  // udot v15.4s, v16.16b, v1.4b[0]\n"
       "58:"  // Height 2: Multiply loop: No odd multiplies
@@ -654,9 +654,9 @@ void a64_hybrid_u8u32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 48b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
       "cmp x11, #0x10\n"
       "prfm pstl1keep, [x9, #0x0]\n"
-      "add x24, x9, x20, LSL #2\n"
       "prfm pstl1keep, [x24, #0x0]\n"
       "bge 67f\n"
       "tbz x11, #3, 62f\n"
@@ -738,12 +738,12 @@ void a64_hybrid_u8u32_dot_6x16 (
       "69:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "70:"  // Height 3: Column loop
       "tbz %x[flags], #0, 80f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x23, x24, x20, LSL #2\n"
       "bge 79f\n"
       "tbz x11, #3, 74f\n"
@@ -864,8 +864,8 @@ void a64_hybrid_u8u32_dot_6x16 (
       "mov x28, #0x0\n"
       "82:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 83f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -905,18 +905,18 @@ void a64_hybrid_u8u32_dot_6x16 (
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "add x24, x24, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6f80e2aa  // udot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x6f81e2ae  // udot v14.4s, v21.16b, v1.4b[0]\n"
+      "cmp x27, #0x20\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6f82e2b2  // udot v18.4s, v21.16b, v2.4b[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6f80e28b  // udot v11.4s, v20.16b, v0.4b[0]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6f81e28f  // udot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x6f82e293  // udot v19.4s, v20.16b, v2.4b[0]\n"
       "ldr q20, [x10, #0x50]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6fa0e2a8  // udot v8.4s, v21.16b, v0.4b[1]\n"
       ".inst 0x6fa1e2ac  // udot v12.4s, v21.16b, v1.4b[1]\n"
       ".inst 0x6fa2e2b0  // udot v16.4s, v21.16b, v2.4b[1]\n"
@@ -983,14 +983,14 @@ void a64_hybrid_u8u32_dot_6x16 (
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
       "ldr q20, [x10, #0x30]\n"
       "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6f80e2aa  // udot v10.4s, v21.16b, v0.4b[0]\n"
       ".inst 0x6f81e2ae  // udot v14.4s, v21.16b, v1.4b[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6f82e2b2  // udot v18.4s, v21.16b, v2.4b[0]\n"
       "ldr q21, [x10, #0x40]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6f80e28b  // udot v11.4s, v20.16b, v0.4b[0]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6f81e28f  // udot v15.4s, v20.16b, v1.4b[0]\n"
       ".inst 0x6f82e293  // udot v19.4s, v20.16b, v2.4b[0]\n"
       "ldr q20, [x10, #0x50]\n"
@@ -1049,12 +1049,12 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr s24, [x26], #0x4\n"
       "ldr s23, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s22, [x24], #0x4\n"
       "ldr q21, [x10, #0x0]\n"
-      "cmp x27, #0x4\n"
-      "ldr q20, [x10, #0x10]\n"
       ".inst 0x6f98e2a8  // udot v8.4s, v21.16b, v24.4b[0]\n"
       ".inst 0x6f97e2ac  // udot v12.4s, v21.16b, v23.4b[0]\n"
+      "ldr q20, [x10, #0x10]\n"
       ".inst 0x6f96e2b0  // udot v16.4s, v21.16b, v22.4b[0]\n"
       "ldr q21, [x10, #0x20]\n"
       ".inst 0x6f98e289  // udot v9.4s, v20.16b, v24.4b[0]\n"
@@ -1108,11 +1108,11 @@ void a64_hybrid_u8u32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 82b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "cmp x11, #0x10\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "prfm pstl1keep, [x23, #0x0]\n"
       "bge 101f\n"
       "tbz x11, #3, 96f\n"
@@ -1214,13 +1214,13 @@ void a64_hybrid_u8u32_dot_6x16 (
       "103:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "104:"  // Height 4: Column loop
       "tbz %x[flags], #0, 114f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x22, x23, x20, LSL #2\n"
       "bge 113f\n"
       "tbz x11, #3, 108f\n"
@@ -1365,8 +1365,8 @@ void a64_hybrid_u8u32_dot_6x16 (
       "mov x28, #0x0\n"
       "116:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 117f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1506,14 +1506,14 @@ void a64_hybrid_u8u32_dot_6x16 (
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
       "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x6f83e0f5  // udot v21.4s, v7.16b, v3.4b[0]\n"
       "ldr q24, [x10, #0x30]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6f80e32a  // udot v10.4s, v25.16b, v0.4b[0]\n"
       ".inst 0x6f81e32e  // udot v14.4s, v25.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6f82e332  // udot v18.4s, v25.16b, v2.4b[0]\n"
       ".inst 0x6f83e336  // udot v22.4s, v25.16b, v3.4b[0]\n"
@@ -1591,9 +1591,9 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr s29, [x26], #0x4\n"
       "ldr s28, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s27, [x24], #0x4\n"
       "ldr s26, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
       "ldr q25, [x10, #0x0]\n"
       "ldr q24, [x10, #0x10]\n"
       ".inst 0x6f9de328  // udot v8.4s, v25.16b, v29.4b[0]\n"
@@ -1662,13 +1662,13 @@ void a64_hybrid_u8u32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 116b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "prfm pstl1keep, [x22, #0x0]\n"
       "bge 135f\n"
       "tbz x11, #3, 130f\n"
@@ -1790,14 +1790,14 @@ void a64_hybrid_u8u32_dot_6x16 (
       "137:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "138:"  // Height 5: Column loop
       "tbz %x[flags], #0, 148f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x21, x22, x20, LSL #2\n"
       "bge 147f\n"
       "tbz x11, #3, 142f\n"
@@ -1966,8 +1966,8 @@ void a64_hybrid_u8u32_dot_6x16 (
       "mov x28, #0x0\n"
       "150:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 151f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2133,12 +2133,12 @@ void a64_hybrid_u8u32_dot_6x16 (
       "add x22, x22, #0x10\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
+      "sub x27, x27, #0x10\n"
       "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6f83e0f5  // udot v21.4s, v7.16b, v3.4b[0]\n"
       ".inst 0x6f84e0f9  // udot v25.4s, v7.16b, v4.4b[0]\n"
       "ldr q28, [x10, #0x30]\n"
-      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6f80e3aa  // udot v10.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x6f81e3ae  // udot v14.4s, v29.16b, v1.4b[0]\n"
       "prfm pldl1keep, [x24, #0x80]\n"
@@ -2233,14 +2233,14 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr s2, [x26], #0x4\n"
       "ldr s1, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s0, [x24], #0x4\n"
       "ldr s31, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
       "ldr s30, [x22], #0x4\n"
       "ldr q29, [x10, #0x0]\n"
-      "ldr q28, [x10, #0x10]\n"
       ".inst 0x6f82e3a8  // udot v8.4s, v29.16b, v2.4b[0]\n"
       ".inst 0x6f81e3ac  // udot v12.4s, v29.16b, v1.4b[0]\n"
+      "ldr q28, [x10, #0x10]\n"
       ".inst 0x6f80e3b0  // udot v16.4s, v29.16b, v0.4b[0]\n"
       ".inst 0x6f9fe3b4  // udot v20.4s, v29.16b, v31.4b[0]\n"
       ".inst 0x6f9ee3b8  // udot v24.4s, v29.16b, v30.4b[0]\n"
@@ -2316,15 +2316,15 @@ void a64_hybrid_u8u32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 150b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "add x21, x22, x20, LSL #2\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "cmp x11, #0x10\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "prfm pstl1keep, [x21, #0x0]\n"
       "bge 169f\n"
       "tbz x11, #3, 164f\n"
@@ -2465,20 +2465,19 @@ void a64_hybrid_u8u32_dot_6x16 (
       "b 206f\n"
       "171:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "172:"  // Height 6: Column loop
       "tbz %x[flags], #0, 182f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x20, x21, x20, LSL #2\n"
       "bge 181f\n"
       "tbz x11, #3, 176f\n"
@@ -2671,8 +2670,8 @@ void a64_hybrid_u8u32_dot_6x16 (
       "mov x28, #0x0\n"
       "184:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 185f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2862,18 +2861,18 @@ void a64_hybrid_u8u32_dot_6x16 (
       ".inst 0x6f80e0e9  // udot v9.4s, v7.16b, v0.4b[0]\n"
       ".inst 0x6f81e0ed  // udot v13.4s, v7.16b, v1.4b[0]\n"
       "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6f82e0f1  // udot v17.4s, v7.16b, v2.4b[0]\n"
       ".inst 0x6f83e0f5  // udot v21.4s, v7.16b, v3.4b[0]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6f84e0f9  // udot v25.4s, v7.16b, v4.4b[0]\n"
       ".inst 0x6f85e0fd  // udot v29.4s, v7.16b, v5.4b[0]\n"
       "ldr q7, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6f80e0ca  // udot v10.4s, v6.16b, v0.4b[0]\n"
       ".inst 0x6f81e0ce  // udot v14.4s, v6.16b, v1.4b[0]\n"
-      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6f82e0d2  // udot v18.4s, v6.16b, v2.4b[0]\n"
       ".inst 0x6f83e0d6  // udot v22.4s, v6.16b, v3.4b[0]\n"
@@ -2979,9 +2978,9 @@ void a64_hybrid_u8u32_dot_6x16 (
       "ldr s7, [x26], #0x4\n"
       "ldr s6, [x25], #0x4\n"
       "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "ldr s5, [x24], #0x4\n"
       "ldr s4, [x23], #0x4\n"
-      "cmp x27, #0x4\n"
       "ldr s3, [x22], #0x4\n"
       "ldr s2, [x21], #0x4\n"
       "ldr q1, [x10, #0x0]\n"
@@ -3074,16 +3073,16 @@ void a64_hybrid_u8u32_dot_6x16 (
       "cmp x28, x20\n"
       "bne 184b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "add x24, x9, x20, LSL #2\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "add x23, x24, x20, LSL #2\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "add x22, x23, x20, LSL #2\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "add x21, x22, x20, LSL #2\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "add x20, x21, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "prfm pstl1keep, [x21, #0x0]\n"
       "prfm pstl1keep, [x20, #0x0]\n"
       "bge 203f\n"
@@ -3254,8 +3253,8 @@ void a64_hybrid_u8u32_dot_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "206:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16.hpp
index baa4e28e88..09fba7e253 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16.hpp
@@ -70,7 +70,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 6, 16, 8> transforms = {};
+    StdTransformsFixed<rhs_operand_type, result_type, 6, 16, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp
index 790a350838..8c6fbd4c83 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp
@@ -44,18 +44,18 @@ void a64_hybrid_u8u32_mmla_6x16 (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -88,7 +88,7 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "beq 38f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "tbz %x[flags], #0, 13f\n"
       "cmp x11, #0x10\n"
@@ -176,8 +176,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "mov x28, #0x0\n"
       "15:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 16f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -197,12 +197,7 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q6, [x10, #0x10]\n"
       "blt 19f\n"
       "18:"  // Height 1: Multiply loop: Main loop head
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
       "trn1 v19.2d, v1.2d, v20.2d\n"
-      "trn2 v1.2d, v1.2d, v20.2d\n"
       ".inst 0x6e87a668  // ummla v8.4s, v19.16b, v7.16b\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e86a66c  // ummla v12.4s, v19.16b, v6.16b\n"
@@ -215,6 +210,7 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e91a66e  // ummla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v20.2d\n"
       ".inst 0x6e92a66b  // ummla v11.4s, v19.16b, v18.16b\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x6e91a66f  // ummla v15.4s, v19.16b, v17.16b\n"
@@ -231,38 +227,39 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x6e91a42e  // ummla v14.4s, v1.16b, v17.16b\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "cmp x27, #0x20\n"
       ".inst 0x6e92a42b  // ummla v11.4s, v1.16b, v18.16b\n"
-      "ldr q7, [x10, #0x0]\n"
       ".inst 0x6e91a42f  // ummla v15.4s, v1.16b, v17.16b\n"
       "ldr q1, [x26, #0x0]\n"
+      "add x10, x10, #0x100\n"
+      "ldr q7, [x10, #0x0]\n"
       "ldr q6, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       "bge 18b\n"
       "19:"  // Height 1: Multiply loop: Single iteration only
-      "add x26, x26, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "trn1 v19.2d, v1.2d, v17.2d\n"
-      "trn2 v1.2d, v1.2d, v17.2d\n"
-      ".inst 0x6e87a668  // ummla v8.4s, v19.16b, v7.16b\n"
+      "trn1 v20.2d, v1.2d, v21.2d\n"
+      ".inst 0x6e87a688  // ummla v8.4s, v20.16b, v7.16b\n"
       "ldr q18, [x10, #0x20]\n"
-      ".inst 0x6e86a66c  // ummla v12.4s, v19.16b, v6.16b\n"
+      ".inst 0x6e86a68c  // ummla v12.4s, v20.16b, v6.16b\n"
       "ldr q17, [x10, #0x30]\n"
-      ".inst 0x6e92a669  // ummla v9.4s, v19.16b, v18.16b\n"
+      ".inst 0x6e92a689  // ummla v9.4s, v20.16b, v18.16b\n"
       "ldr q18, [x10, #0x40]\n"
-      ".inst 0x6e91a66d  // ummla v13.4s, v19.16b, v17.16b\n"
+      ".inst 0x6e91a68d  // ummla v13.4s, v20.16b, v17.16b\n"
       "ldr q17, [x10, #0x50]\n"
-      ".inst 0x6e92a66a  // ummla v10.4s, v19.16b, v18.16b\n"
-      "ldr q20, [x10, #0x60]\n"
-      ".inst 0x6e91a66e  // ummla v14.4s, v19.16b, v17.16b\n"
+      ".inst 0x6e92a68a  // ummla v10.4s, v20.16b, v18.16b\n"
+      "ldr q19, [x10, #0x60]\n"
+      ".inst 0x6e91a68e  // ummla v14.4s, v20.16b, v17.16b\n"
       "ldr q18, [x10, #0x70]\n"
-      ".inst 0x6e94a66b  // ummla v11.4s, v19.16b, v20.16b\n"
+      "trn2 v1.2d, v1.2d, v21.2d\n"
+      ".inst 0x6e93a68b  // ummla v11.4s, v20.16b, v19.16b\n"
       "ldr q17, [x10, #0x80]\n"
-      ".inst 0x6e92a66f  // ummla v15.4s, v19.16b, v18.16b\n"
-      "ldr q20, [x10, #0x90]\n"
+      ".inst 0x6e92a68f  // ummla v15.4s, v20.16b, v18.16b\n"
+      "ldr q19, [x10, #0x90]\n"
       ".inst 0x6e91a428  // ummla v8.4s, v1.16b, v17.16b\n"
       "ldr q18, [x10, #0xa0]\n"
-      ".inst 0x6e94a42c  // ummla v12.4s, v1.16b, v20.16b\n"
+      ".inst 0x6e93a42c  // ummla v12.4s, v1.16b, v19.16b\n"
       "ldr q17, [x10, #0xb0]\n"
       ".inst 0x6e92a429  // ummla v9.4s, v1.16b, v18.16b\n"
       "ldr q18, [x10, #0xc0]\n"
@@ -272,21 +269,22 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x6e91a42e  // ummla v14.4s, v1.16b, v17.16b\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6e92a42b  // ummla v11.4s, v1.16b, v18.16b\n"
       ".inst 0x6e91a42f  // ummla v15.4s, v1.16b, v17.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "20:"  // Height 1: Multiply loop: Main loop skip
       "cbz x27, 27f\n"
       "cmp x27, #0x8\n"
       "blt 22f\n"
       "21:"  // Height 1: Multiply loop: Odd block loop
       "ldr d19, [x26], #0x8\n"
-      "ldr q20, [x10, #0x0]\n"
-      "sub x27, x27, #0x8\n"
+      "ldr q18, [x10, #0x0]\n"
+      "trn1 v19.2d, v19.2d, v17.2d\n"
       "ldr q17, [x10, #0x10]\n"
-      "cmp x27, #0x8\n"
-      "trn1 v19.2d, v19.2d, v18.2d\n"
-      ".inst 0x6e94a668  // ummla v8.4s, v19.16b, v20.16b\n"
+      ".inst 0x6e92a668  // ummla v8.4s, v19.16b, v18.16b\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e91a66c  // ummla v12.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x30]\n"
@@ -298,9 +296,11 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e91a66e  // ummla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6e92a66b  // ummla v11.4s, v19.16b, v18.16b\n"
       ".inst 0x6e91a66f  // ummla v15.4s, v19.16b, v17.16b\n"
+      "add x10, x10, #0x80\n"
       "bge 21b\n"
       "22:"  // Height 1: Multiply loop: Skip odd blocks
       "cbz x27, 27f\n"
@@ -324,24 +324,24 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "25:"  // Height 1: Multiply loop: Ragged operand read: partial_1_0
       "ldr b1, [x26, #0x0]\n"
       "26:"  // Height 1: Multiply loop: Ragged operand read: Done
-      "ldr q24, [x10, #0x0]\n"
-      "ldr q20, [x10, #0x10]\n"
+      "ldr q23, [x10, #0x0]\n"
+      "ldr q18, [x10, #0x10]\n"
       "trn1 v19.2d, v1.2d, v17.2d\n"
-      ".inst 0x6e98a668  // ummla v8.4s, v19.16b, v24.16b\n"
+      ".inst 0x6e97a668  // ummla v8.4s, v19.16b, v23.16b\n"
       "ldr q17, [x10, #0x20]\n"
-      ".inst 0x6e94a66c  // ummla v12.4s, v19.16b, v20.16b\n"
-      "ldr q0, [x10, #0x30]\n"
+      ".inst 0x6e92a66c  // ummla v12.4s, v19.16b, v18.16b\n"
+      "ldr q31, [x10, #0x30]\n"
       ".inst 0x6e91a669  // ummla v9.4s, v19.16b, v17.16b\n"
       "ldr q20, [x10, #0x40]\n"
-      ".inst 0x6e80a66d  // ummla v13.4s, v19.16b, v0.16b\n"
+      ".inst 0x6e9fa66d  // ummla v13.4s, v19.16b, v31.16b\n"
       "ldr q17, [x10, #0x50]\n"
       ".inst 0x6e94a66a  // ummla v10.4s, v19.16b, v20.16b\n"
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e91a66e  // ummla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e92a66b  // ummla v11.4s, v19.16b, v18.16b\n"
       ".inst 0x6e91a66f  // ummla v15.4s, v19.16b, v17.16b\n"
+      "add x10, x10, #0x80\n"
       "27:"  // Height 1: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -413,7 +413,7 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "38:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "39:"  // Height 2: Column loop
       "tbz %x[flags], #0, 50f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
@@ -523,8 +523,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "mov x28, #0x0\n"
       "52:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 53f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -549,14 +549,6 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "blt 56f\n"
       "55:"  // Height 2: Multiply loop: Main loop head
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "ldr q2, [x25, #0x0]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e87a668  // ummla v8.4s, v19.16b, v7.16b\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e86a66c  // ummla v12.4s, v19.16b, v6.16b\n"
@@ -569,6 +561,7 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e91a66e  // ummla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x6e92a66b  // ummla v11.4s, v19.16b, v18.16b\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x6e91a66f  // ummla v15.4s, v19.16b, v17.16b\n"
@@ -585,21 +578,22 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x6e91a42e  // ummla v14.4s, v1.16b, v17.16b\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "sub x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "ldr q2, [x25, #0x0]\n"
+      "cmp x27, #0x20\n"
       ".inst 0x6e92a42b  // ummla v11.4s, v1.16b, v18.16b\n"
+      "add x10, x10, #0x100\n"
       "ldr q7, [x10, #0x0]\n"
       ".inst 0x6e91a42f  // ummla v15.4s, v1.16b, v17.16b\n"
       "ldr q1, [x26, #0x0]\n"
       "ldr q6, [x10, #0x10]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       "bge 55b\n"
       "56:"  // Height 2: Multiply loop: Single iteration only
       "trn1 v19.2d, v1.2d, v2.2d\n"
-      "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e87a668  // ummla v8.4s, v19.16b, v7.16b\n"
       "ldr q18, [x10, #0x20]\n"
       ".inst 0x6e86a66c  // ummla v12.4s, v19.16b, v6.16b\n"
@@ -612,6 +606,7 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e91a66e  // ummla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
+      "trn2 v1.2d, v1.2d, v2.2d\n"
       ".inst 0x6e92a66b  // ummla v11.4s, v19.16b, v18.16b\n"
       "ldr q18, [x10, #0x80]\n"
       ".inst 0x6e91a66f  // ummla v15.4s, v19.16b, v17.16b\n"
@@ -628,36 +623,41 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0xe0]\n"
       ".inst 0x6e91a42e  // ummla v14.4s, v1.16b, v17.16b\n"
       "ldr q17, [x10, #0xf0]\n"
-      "add x10, x10, #0x100\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e92a42b  // ummla v11.4s, v1.16b, v18.16b\n"
       ".inst 0x6e91a42f  // ummla v15.4s, v1.16b, v17.16b\n"
+      "sub x27, x27, #0x10\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "add x10, x10, #0x100\n"
       "57:"  // Height 2: Multiply loop: Main loop skip
       "cbz x27, 64f\n"
       "cmp x27, #0x8\n"
       "blt 59f\n"
       "58:"  // Height 2: Multiply loop: Odd block loop
-      "ldr d20, [x26], #0x8\n"
-      "ldr d19, [x25], #0x8\n"
+      "ldr d18, [x26], #0x8\n"
+      "ldr d17, [x25], #0x8\n"
+      "trn1 v19.2d, v18.2d, v17.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr q18, [x10, #0x0]\n"
-      "ldr q17, [x10, #0x10]\n"
-      "cmp x27, #0x8\n"
-      "trn1 v22.2d, v20.2d, v19.2d\n"
-      ".inst 0x6e92a6c8  // ummla v8.4s, v22.16b, v18.16b\n"
-      "ldr q2, [x10, #0x20]\n"
-      ".inst 0x6e91a6cc  // ummla v12.4s, v22.16b, v17.16b\n"
+      "ldr q17, [x10, #0x0]\n"
+      "ldr q22, [x10, #0x10]\n"
+      ".inst 0x6e91a668  // ummla v8.4s, v19.16b, v17.16b\n"
+      ".inst 0x6e96a66c  // ummla v12.4s, v19.16b, v22.16b\n"
+      "ldr q1, [x10, #0x20]\n"
       "ldr q17, [x10, #0x30]\n"
-      ".inst 0x6e82a6c9  // ummla v9.4s, v22.16b, v2.16b\n"
+      ".inst 0x6e81a669  // ummla v9.4s, v19.16b, v1.16b\n"
+      ".inst 0x6e91a66d  // ummla v13.4s, v19.16b, v17.16b\n"
       "ldr q18, [x10, #0x40]\n"
-      ".inst 0x6e91a6cd  // ummla v13.4s, v22.16b, v17.16b\n"
       "ldr q17, [x10, #0x50]\n"
-      ".inst 0x6e92a6ca  // ummla v10.4s, v22.16b, v18.16b\n"
+      ".inst 0x6e92a66a  // ummla v10.4s, v19.16b, v18.16b\n"
+      ".inst 0x6e91a66e  // ummla v14.4s, v19.16b, v17.16b\n"
       "ldr q18, [x10, #0x60]\n"
-      ".inst 0x6e91a6ce  // ummla v14.4s, v22.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x6e92a66b  // ummla v11.4s, v19.16b, v18.16b\n"
+      ".inst 0x6e91a66f  // ummla v15.4s, v19.16b, v17.16b\n"
       "add x10, x10, #0x80\n"
-      ".inst 0x6e92a6cb  // ummla v11.4s, v22.16b, v18.16b\n"
-      ".inst 0x6e91a6cf  // ummla v15.4s, v22.16b, v17.16b\n"
       "bge 58b\n"
       "59:"  // Height 2: Multiply loop: Skip odd blocks
       "cbz x27, 64f\n"
@@ -703,27 +703,27 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q18, [x10, #0x60]\n"
       ".inst 0x6e91a66e  // ummla v14.4s, v19.16b, v17.16b\n"
       "ldr q17, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e92a66b  // ummla v11.4s, v19.16b, v18.16b\n"
       ".inst 0x6e91a66f  // ummla v15.4s, v19.16b, v17.16b\n"
+      "add x10, x10, #0x80\n"
       "64:"  // Height 2: Multiply loop: No odd multiplies
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 52b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
       "cmp x11, #0x10\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
       "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "bge 73f\n"
       "tbz x11, #3, 68f\n"
       "st1 { v7.4s }, [x9], #0x10\n"
@@ -804,12 +804,12 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "75:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "76:"  // Height 3: Column loop
       "tbz %x[flags], #0, 87f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x23, x24, x20, LSL #2\n"
       "bge 85f\n"
       "tbz x11, #3, 80f\n"
@@ -951,8 +951,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "mov x28, #0x0\n"
       "89:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 90f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -982,38 +982,35 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "92:"  // Height 3: Multiply loop: Main loop head
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e87a788  // ummla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x6e86a78c  // ummla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
+      "trn1 v27.2d, v3.2d, v29.2d\n"
       ".inst 0x6e87a770  // ummla v16.4s, v27.16b, v7.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e86a78c  // ummla v12.4s, v28.16b, v6.16b\n"
       ".inst 0x6e86a774  // ummla v20.4s, v27.16b, v6.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e9aa789  // ummla v9.4s, v28.16b, v26.16b\n"
+      "trn2 v3.2d, v3.2d, v29.2d\n"
       ".inst 0x6e9aa771  // ummla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e99a78d  // ummla v13.4s, v28.16b, v25.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6e99a775  // ummla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e9aa78a  // ummla v10.4s, v28.16b, v26.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e9aa772  // ummla v18.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e99a78e  // ummla v14.4s, v28.16b, v25.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e99a776  // ummla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x6e9aa78b  // ummla v11.4s, v28.16b, v26.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e9aa773  // ummla v19.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x6e99a78f  // ummla v15.4s, v28.16b, v25.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x6e99a777  // ummla v23.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x90]\n"
       "ldr q2, [x25, #0x0]\n"
@@ -1021,12 +1018,15 @@ void a64_hybrid_u8u32_mmla_6x16 (
       ".inst 0x6e9aa470  // ummla v16.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x6e99a42c  // ummla v12.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e99a474  // ummla v20.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x6e9aa429  // ummla v9.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e9aa471  // ummla v17.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x6e99a42d  // ummla v13.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e99a475  // ummla v21.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xd0]\n"
       ".inst 0x6e9aa42a  // ummla v10.4s, v1.16b, v26.16b\n"
@@ -1048,43 +1048,43 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "93:"  // Height 3: Multiply loop: Single iteration only
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e87a788  // ummla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x6e86a78c  // ummla v12.4s, v28.16b, v6.16b\n"
-      "trn2 v3.2d, v3.2d, v25.2d\n"
+      "trn1 v27.2d, v3.2d, v29.2d\n"
       ".inst 0x6e87a770  // ummla v16.4s, v27.16b, v7.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e86a78c  // ummla v12.4s, v28.16b, v6.16b\n"
       ".inst 0x6e86a774  // ummla v20.4s, v27.16b, v6.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e9aa789  // ummla v9.4s, v28.16b, v26.16b\n"
+      "trn2 v3.2d, v3.2d, v29.2d\n"
       ".inst 0x6e9aa771  // ummla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e99a78d  // ummla v13.4s, v28.16b, v25.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e99a775  // ummla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e9aa78a  // ummla v10.4s, v28.16b, v26.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e9aa772  // ummla v18.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e99a78e  // ummla v14.4s, v28.16b, v25.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e99a776  // ummla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x6e9aa78b  // ummla v11.4s, v28.16b, v26.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6e9aa773  // ummla v19.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x6e99a78f  // ummla v15.4s, v28.16b, v25.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e99a777  // ummla v23.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x90]\n"
       ".inst 0x6e9aa428  // ummla v8.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e9aa470  // ummla v16.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x6e99a42c  // ummla v12.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e99a474  // ummla v20.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x6e9aa429  // ummla v9.4s, v1.16b, v26.16b\n"
@@ -1109,25 +1109,25 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 96f\n"
       "95:"  // Height 3: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
-      "sub x27, x27, #0x8\n"
-      "ldr d27, [x24], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "trn1 v28.2d, v26.2d, v25.2d\n"
+      "ldr d25, [x24], #0x8\n"
       "ldr q26, [x10, #0x0]\n"
-      "cmp x27, #0x8\n"
-      "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v27.2d, v29.2d\n"
+      "trn1 v27.2d, v25.2d, v27.2d\n"
       ".inst 0x6e9aa788  // ummla v8.4s, v28.16b, v26.16b\n"
+      "ldr q25, [x10, #0x10]\n"
       ".inst 0x6e9aa770  // ummla v16.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x20]\n"
       ".inst 0x6e99a78c  // ummla v12.4s, v28.16b, v25.16b\n"
       ".inst 0x6e99a774  // ummla v20.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e9aa789  // ummla v9.4s, v28.16b, v26.16b\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6e9aa771  // ummla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e99a78d  // ummla v13.4s, v28.16b, v25.16b\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6e99a775  // ummla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e9aa78a  // ummla v10.4s, v28.16b, v26.16b\n"
@@ -1136,8 +1136,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       ".inst 0x6e99a78e  // ummla v14.4s, v28.16b, v25.16b\n"
       ".inst 0x6e99a776  // ummla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e9aa78b  // ummla v11.4s, v28.16b, v26.16b\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x6e9aa773  // ummla v19.4s, v27.16b, v26.16b\n"
       ".inst 0x6e99a78f  // ummla v15.4s, v28.16b, v25.16b\n"
       ".inst 0x6e99a777  // ummla v23.4s, v27.16b, v25.16b\n"
@@ -1183,9 +1183,9 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn1 v27.2d, v3.2d, v25.2d\n"
       ".inst 0x6e9aa788  // ummla v8.4s, v28.16b, v26.16b\n"
-      ".inst 0x6e9da78c  // ummla v12.4s, v28.16b, v29.16b\n"
       ".inst 0x6e9aa770  // ummla v16.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e9da78c  // ummla v12.4s, v28.16b, v29.16b\n"
       ".inst 0x6e9da774  // ummla v20.4s, v27.16b, v29.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e9aa789  // ummla v9.4s, v28.16b, v26.16b\n"
@@ -1211,20 +1211,20 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 89b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "cmp x11, #0x10\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
-      "uzp2 v9.2d, v9.2d, v13.2d\n"
       "prfm pstl1keep, [x9, #0x0]\n"
+      "uzp2 v9.2d, v9.2d, v13.2d\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v18.2d, v18.2d, v22.2d\n"
@@ -1329,13 +1329,13 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "112:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "113:"  // Height 4: Column loop
       "tbz %x[flags], #0, 124f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x22, x23, x20, LSL #2\n"
       "bge 122f\n"
       "tbz x11, #3, 117f\n"
@@ -1497,8 +1497,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "mov x28, #0x0\n"
       "126:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 127f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1532,38 +1532,33 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "129:"  // Height 4: Multiply loop: Main loop head
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e87a788  // ummla v8.4s, v28.16b, v7.16b\n"
       "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "ldr q4, [x23, #0x0]\n"
-      ".inst 0x6e87a788  // ummla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x6e86a78c  // ummla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e87a770  // ummla v16.4s, v27.16b, v7.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e86a78c  // ummla v12.4s, v28.16b, v6.16b\n"
       ".inst 0x6e86a774  // ummla v20.4s, v27.16b, v6.16b\n"
       "ldr q25, [x10, #0x30]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e9aa789  // ummla v9.4s, v28.16b, v26.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x6e9aa771  // ummla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e99a78d  // ummla v13.4s, v28.16b, v25.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e99a775  // ummla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e9aa78a  // ummla v10.4s, v28.16b, v26.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e9aa772  // ummla v18.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e99a78e  // ummla v14.4s, v28.16b, v25.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e99a776  // ummla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x6e9aa78b  // ummla v11.4s, v28.16b, v26.16b\n"
+      "add x23, x23, #0x10\n"
+      "ldr q4, [x23, #0x0]\n"
       ".inst 0x6e9aa773  // ummla v19.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x6e99a78f  // ummla v15.4s, v28.16b, v25.16b\n"
@@ -1574,18 +1569,23 @@ void a64_hybrid_u8u32_mmla_6x16 (
       ".inst 0x6e9aa470  // ummla v16.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x6e99a42c  // ummla v12.4s, v1.16b, v25.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x6e99a474  // ummla v20.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x6e9aa429  // ummla v9.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e9aa471  // ummla v17.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x6e99a42d  // ummla v13.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e99a475  // ummla v21.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xd0]\n"
       ".inst 0x6e9aa42a  // ummla v10.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e9aa472  // ummla v18.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xe0]\n"
       ".inst 0x6e99a42e  // ummla v14.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e99a476  // ummla v22.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xf0]\n"
       "add x10, x10, #0x100\n"
@@ -1601,48 +1601,48 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "130:"  // Height 4: Multiply loop: Single iteration only
       "trn1 v28.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e87a788  // ummla v8.4s, v28.16b, v7.16b\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v27.2d, v3.2d, v4.2d\n"
-      "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      ".inst 0x6e87a788  // ummla v8.4s, v28.16b, v7.16b\n"
-      ".inst 0x6e86a78c  // ummla v12.4s, v28.16b, v6.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e87a770  // ummla v16.4s, v27.16b, v7.16b\n"
       "ldr q26, [x10, #0x20]\n"
+      ".inst 0x6e86a78c  // ummla v12.4s, v28.16b, v6.16b\n"
       ".inst 0x6e86a774  // ummla v20.4s, v27.16b, v6.16b\n"
       "ldr q25, [x10, #0x30]\n"
       ".inst 0x6e9aa789  // ummla v9.4s, v28.16b, v26.16b\n"
+      "trn2 v3.2d, v3.2d, v4.2d\n"
       ".inst 0x6e9aa771  // ummla v17.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x40]\n"
       ".inst 0x6e99a78d  // ummla v13.4s, v28.16b, v25.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e99a775  // ummla v21.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x50]\n"
       ".inst 0x6e9aa78a  // ummla v10.4s, v28.16b, v26.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e9aa772  // ummla v18.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x60]\n"
       ".inst 0x6e99a78e  // ummla v14.4s, v28.16b, v25.16b\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e99a776  // ummla v22.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x70]\n"
       ".inst 0x6e9aa78b  // ummla v11.4s, v28.16b, v26.16b\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6e9aa773  // ummla v19.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x80]\n"
       ".inst 0x6e99a78f  // ummla v15.4s, v28.16b, v25.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e99a777  // ummla v23.4s, v27.16b, v25.16b\n"
       "ldr q25, [x10, #0x90]\n"
       ".inst 0x6e9aa428  // ummla v8.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e9aa470  // ummla v16.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xa0]\n"
       ".inst 0x6e99a42c  // ummla v12.4s, v1.16b, v25.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e99a474  // ummla v20.4s, v3.16b, v25.16b\n"
       "ldr q25, [x10, #0xb0]\n"
       ".inst 0x6e9aa429  // ummla v9.4s, v1.16b, v26.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e9aa471  // ummla v17.4s, v3.16b, v26.16b\n"
       "ldr q26, [x10, #0xc0]\n"
       ".inst 0x6e99a42d  // ummla v13.4s, v1.16b, v25.16b\n"
@@ -1664,16 +1664,16 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 133f\n"
       "132:"  // Height 4: Multiply loop: Odd block loop
-      "ldr d30, [x26], #0x8\n"
-      "ldr d28, [x25], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d25, [x25], #0x8\n"
+      "trn1 v28.2d, v26.2d, v25.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d29, [x24], #0x8\n"
-      "ldr d27, [x23], #0x8\n"
+      "ldr d26, [x24], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "trn1 v27.2d, v26.2d, v25.2d\n"
       "cmp x27, #0x8\n"
       "ldr q26, [x10, #0x0]\n"
       "ldr q25, [x10, #0x10]\n"
-      "trn1 v28.2d, v30.2d, v28.2d\n"
-      "trn1 v27.2d, v29.2d, v27.2d\n"
       ".inst 0x6e9aa788  // ummla v8.4s, v28.16b, v26.16b\n"
       ".inst 0x6e9aa770  // ummla v16.4s, v27.16b, v26.16b\n"
       "ldr q26, [x10, #0x20]\n"
@@ -1774,24 +1774,24 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 126b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
+      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp2 v11.2d, v11.2d, v15.2d\n"
-      "prfm pstl1keep, [x24, #0x0]\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
@@ -1918,14 +1918,14 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "149:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "150:"  // Height 5: Column loop
       "tbz %x[flags], #0, 161f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x21, x22, x20, LSL #2\n"
       "bge 159f\n"
       "tbz x11, #3, 154f\n"
@@ -2123,8 +2123,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "mov x28, #0x0\n"
       "163:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 164f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2161,51 +2161,51 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "166:"  // Height 5: Multiply loop: Main loop head
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "sub x27, x27, #0x10\n"
+      ".inst 0x6e87a4c8  // ummla v8.4s, v6.16b, v7.16b\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x6e87a450  // ummla v16.4s, v2.16b, v7.16b\n"
+      "sub x27, x27, #0x10\n"
       "trn1 v4.2d, v5.2d, v0.2d\n"
       "trn2 v5.2d, v5.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
-      ".inst 0x6e87a4c8  // ummla v8.4s, v6.16b, v7.16b\n"
-      ".inst 0x6e87a450  // ummla v16.4s, v2.16b, v7.16b\n"
-      "add x22, x22, #0x10\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e87a498  // ummla v24.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x6e80a4cc  // ummla v12.4s, v6.16b, v0.16b\n"
       ".inst 0x6e80a454  // ummla v20.4s, v2.16b, v0.16b\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6e80a49c  // ummla v28.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x6e87a4c9  // ummla v9.4s, v6.16b, v7.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e87a451  // ummla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a499  // ummla v25.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e80a4cd  // ummla v13.4s, v6.16b, v0.16b\n"
       ".inst 0x6e80a455  // ummla v21.4s, v2.16b, v0.16b\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e80a49d  // ummla v29.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x50]\n"
       ".inst 0x6e87a4ca  // ummla v10.4s, v6.16b, v7.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x6e87a452  // ummla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a49a  // ummla v26.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e80a4ce  // ummla v14.4s, v6.16b, v0.16b\n"
       ".inst 0x6e80a456  // ummla v22.4s, v2.16b, v0.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e80a49e  // ummla v30.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x70]\n"
       ".inst 0x6e87a4cb  // ummla v11.4s, v6.16b, v7.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e87a453  // ummla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a49b  // ummla v27.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e80a4cf  // ummla v15.4s, v6.16b, v0.16b\n"
       ".inst 0x6e80a457  // ummla v23.4s, v2.16b, v0.16b\n"
       "ldr q2, [x25, #0x0]\n"
@@ -2251,47 +2251,47 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "167:"  // Height 5: Multiply loop: Single iteration only
       "trn1 v6.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
-      "add x26, x26, #0x10\n"
+      ".inst 0x6e87a4c8  // ummla v8.4s, v6.16b, v7.16b\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
+      ".inst 0x6e87a450  // ummla v16.4s, v2.16b, v7.16b\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v0.2d\n"
       "trn2 v5.2d, v5.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
-      ".inst 0x6e87a4c8  // ummla v8.4s, v6.16b, v7.16b\n"
-      ".inst 0x6e87a450  // ummla v16.4s, v2.16b, v7.16b\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e87a498  // ummla v24.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x6e80a4cc  // ummla v12.4s, v6.16b, v0.16b\n"
       ".inst 0x6e80a454  // ummla v20.4s, v2.16b, v0.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e80a49c  // ummla v28.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x6e87a4c9  // ummla v9.4s, v6.16b, v7.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e87a451  // ummla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a499  // ummla v25.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e80a4cd  // ummla v13.4s, v6.16b, v0.16b\n"
       ".inst 0x6e80a455  // ummla v21.4s, v2.16b, v0.16b\n"
+      "add x22, x22, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6e80a49d  // ummla v29.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x50]\n"
       ".inst 0x6e87a4ca  // ummla v10.4s, v6.16b, v7.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e87a452  // ummla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a49a  // ummla v26.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e80a4ce  // ummla v14.4s, v6.16b, v0.16b\n"
       ".inst 0x6e80a456  // ummla v22.4s, v2.16b, v0.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e80a49e  // ummla v30.4s, v4.16b, v0.16b\n"
       "ldr q0, [x10, #0x70]\n"
       ".inst 0x6e87a4cb  // ummla v11.4s, v6.16b, v7.16b\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e87a453  // ummla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a49b  // ummla v27.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x80]\n"
@@ -2335,24 +2335,24 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 170f\n"
       "169:"  // Height 5: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d3, [x24], #0x8\n"
-      "ldr d2, [x23], #0x8\n"
-      "cmp x27, #0x8\n"
       "ldr d0, [x22], #0x8\n"
       "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v3.2d, v2.2d\n"
-      "trn1 v2.2d, v0.2d, v5.2d\n"
-      "ldr q0, [x10, #0x10]\n"
+      "trn1 v2.2d, v0.2d, v2.2d\n"
       ".inst 0x6e81a488  // ummla v8.4s, v4.16b, v1.16b\n"
+      "ldr q0, [x10, #0x10]\n"
       ".inst 0x6e81a470  // ummla v16.4s, v3.16b, v1.16b\n"
       ".inst 0x6e81a458  // ummla v24.4s, v2.16b, v1.16b\n"
       "ldr q1, [x10, #0x20]\n"
       ".inst 0x6e80a48c  // ummla v12.4s, v4.16b, v0.16b\n"
       ".inst 0x6e80a474  // ummla v20.4s, v3.16b, v0.16b\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6e80a45c  // ummla v28.4s, v2.16b, v0.16b\n"
       "ldr q0, [x10, #0x30]\n"
       ".inst 0x6e81a489  // ummla v9.4s, v4.16b, v1.16b\n"
@@ -2371,8 +2371,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       ".inst 0x6e80a476  // ummla v22.4s, v3.16b, v0.16b\n"
       ".inst 0x6e80a45e  // ummla v30.4s, v2.16b, v0.16b\n"
       "ldr q0, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e86a48b  // ummla v11.4s, v4.16b, v6.16b\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x6e86a473  // ummla v19.4s, v3.16b, v6.16b\n"
       ".inst 0x6e86a45b  // ummla v27.4s, v2.16b, v6.16b\n"
       ".inst 0x6e80a48f  // ummla v15.4s, v4.16b, v0.16b\n"
@@ -2471,28 +2471,28 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 163b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "cmp x11, #0x10\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
       "prfm pstl1keep, [x24, #0x0]\n"
-      "add x21, x22, x20, LSL #2\n"
+      "prfm pstl1keep, [x23, #0x0]\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp2 v16.2d, v16.2d, v20.2d\n"
-      "prfm pstl1keep, [x23, #0x0]\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
-      "prfm pstl1keep, [x21, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
@@ -2640,20 +2640,19 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "b 224f\n"
       "186:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "187:"  // Height 6: Column loop
       "tbz %x[flags], #0, 198f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "add x20, x21, x20, LSL #2\n"
       "bge 196f\n"
       "tbz x11, #3, 191f\n"
@@ -2871,8 +2870,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "mov x28, #0x0\n"
       "200:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 201f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2913,56 +2912,56 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "203:"  // Height 6: Multiply loop: Main loop head
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e87a408  // ummla v8.4s, v0.16b, v7.16b\n"
       "sub x27, x27, #0x10\n"
-      "add x26, x26, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x25, x25, #0x10\n"
-      "add x24, x24, #0x10\n"
+      ".inst 0x6e87a450  // ummla v16.4s, v2.16b, v7.16b\n"
+      "add x26, x26, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
       "ldr q6, [x10, #0x10]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0x6e87a408  // ummla v8.4s, v0.16b, v7.16b\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      ".inst 0x6e87a450  // ummla v16.4s, v2.16b, v7.16b\n"
-      "cmp x27, #0x20\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e87a498  // ummla v24.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x6e86a40c  // ummla v12.4s, v0.16b, v6.16b\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e86a454  // ummla v20.4s, v2.16b, v6.16b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6e86a49c  // ummla v28.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e87a409  // ummla v9.4s, v0.16b, v7.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e87a451  // ummla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a499  // ummla v25.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e86a40d  // ummla v13.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a455  // ummla v21.4s, v2.16b, v6.16b\n"
+      "add x22, x22, #0x10\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x6e86a49d  // ummla v29.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x50]\n"
       ".inst 0x6e87a40a  // ummla v10.4s, v0.16b, v7.16b\n"
+      "cmp x27, #0x20\n"
       ".inst 0x6e87a452  // ummla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a49a  // ummla v26.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e86a40e  // ummla v14.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a456  // ummla v22.4s, v2.16b, v6.16b\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
       ".inst 0x6e86a49e  // ummla v30.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x70]\n"
       ".inst 0x6e87a40b  // ummla v11.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e87a453  // ummla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a49b  // ummla v27.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e86a40f  // ummla v15.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a457  // ummla v23.4s, v2.16b, v6.16b\n"
       "ldr q2, [x25, #0x0]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e86a49f  // ummla v31.4s, v4.16b, v6.16b\n"
       "ldr q0, [x10, #0x90]\n"
       "ldr q4, [x23, #0x0]\n"
@@ -3006,52 +3005,52 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "204:"  // Height 6: Multiply loop: Single iteration only
       "trn1 v0.2d, v1.2d, v2.2d\n"
       "trn2 v1.2d, v1.2d, v2.2d\n"
+      ".inst 0x6e87a408  // ummla v8.4s, v0.16b, v7.16b\n"
       "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
       "trn1 v2.2d, v3.2d, v4.2d\n"
       "trn2 v3.2d, v3.2d, v4.2d\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      ".inst 0x6e87a450  // ummla v16.4s, v2.16b, v7.16b\n"
+      "add x25, x25, #0x10\n"
       "trn1 v4.2d, v5.2d, v6.2d\n"
       "trn2 v5.2d, v5.2d, v6.2d\n"
       "ldr q6, [x10, #0x10]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x6e87a408  // ummla v8.4s, v0.16b, v7.16b\n"
-      "add x21, x21, #0x10\n"
-      "prfm pldl1keep, [x26, #0x80]\n"
-      "prfm pldl1keep, [x25, #0x80]\n"
-      ".inst 0x6e87a450  // ummla v16.4s, v2.16b, v7.16b\n"
-      "prfm pldl1keep, [x24, #0x80]\n"
-      "sub x27, x27, #0x10\n"
-      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e87a498  // ummla v24.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x20]\n"
       ".inst 0x6e86a40c  // ummla v12.4s, v0.16b, v6.16b\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e86a454  // ummla v20.4s, v2.16b, v6.16b\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6e86a49c  // ummla v28.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x30]\n"
-      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e87a409  // ummla v9.4s, v0.16b, v7.16b\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6e87a451  // ummla v17.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a499  // ummla v25.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x40]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6e86a40d  // ummla v13.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a455  // ummla v21.4s, v2.16b, v6.16b\n"
+      "add x21, x21, #0x10\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x6e86a49d  // ummla v29.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x50]\n"
       ".inst 0x6e87a40a  // ummla v10.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x26, #0x80]\n"
       ".inst 0x6e87a452  // ummla v18.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a49a  // ummla v26.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x60]\n"
+      "prfm pldl1keep, [x25, #0x80]\n"
       ".inst 0x6e86a40e  // ummla v14.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a456  // ummla v22.4s, v2.16b, v6.16b\n"
+      "prfm pldl1keep, [x24, #0x80]\n"
+      "prfm pldl1keep, [x23, #0x80]\n"
       ".inst 0x6e86a49e  // ummla v30.4s, v4.16b, v6.16b\n"
       "ldr q6, [x10, #0x70]\n"
       ".inst 0x6e87a40b  // ummla v11.4s, v0.16b, v7.16b\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       ".inst 0x6e87a453  // ummla v19.4s, v2.16b, v7.16b\n"
       ".inst 0x6e87a49b  // ummla v27.4s, v4.16b, v7.16b\n"
       "ldr q7, [x10, #0x80]\n"
+      "prfm pldl1keep, [x21, #0x80]\n"
       ".inst 0x6e86a40f  // ummla v15.4s, v0.16b, v6.16b\n"
       ".inst 0x6e86a457  // ummla v23.4s, v2.16b, v6.16b\n"
       ".inst 0x6e86a49f  // ummla v31.4s, v4.16b, v6.16b\n"
@@ -3092,18 +3091,18 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "cmp x27, #0x8\n"
       "blt 207f\n"
       "206:"  // Height 6: Multiply loop: Odd block loop
-      "ldr d6, [x26], #0x8\n"
-      "ldr d4, [x25], #0x8\n"
+      "ldr d1, [x26], #0x8\n"
+      "ldr d0, [x25], #0x8\n"
+      "trn1 v4.2d, v1.2d, v0.2d\n"
       "sub x27, x27, #0x8\n"
-      "ldr d5, [x24], #0x8\n"
-      "ldr d3, [x23], #0x8\n"
+      "ldr d1, [x24], #0x8\n"
+      "ldr d0, [x23], #0x8\n"
+      "trn1 v3.2d, v1.2d, v0.2d\n"
       "cmp x27, #0x8\n"
-      "ldr d2, [x22], #0x8\n"
+      "ldr d1, [x22], #0x8\n"
       "ldr d0, [x21], #0x8\n"
+      "trn1 v2.2d, v1.2d, v0.2d\n"
       "ldr q1, [x10, #0x0]\n"
-      "trn1 v4.2d, v6.2d, v4.2d\n"
-      "trn1 v3.2d, v5.2d, v3.2d\n"
-      "trn1 v2.2d, v2.2d, v0.2d\n"
       "ldr q0, [x10, #0x10]\n"
       ".inst 0x6e81a488  // ummla v8.4s, v4.16b, v1.16b\n"
       ".inst 0x6e81a470  // ummla v16.4s, v3.16b, v1.16b\n"
@@ -3197,9 +3196,9 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "ldr q0, [x10, #0x0]\n"
       "trn1 v7.2d, v1.2d, v2.2d\n"
       "trn1 v3.2d, v3.2d, v4.2d\n"
+      ".inst 0x6e80a4e8  // ummla v8.4s, v7.16b, v0.16b\n"
       "trn1 v2.2d, v5.2d, v6.2d\n"
       "ldr q1, [x10, #0x10]\n"
-      ".inst 0x6e80a4e8  // ummla v8.4s, v7.16b, v0.16b\n"
       ".inst 0x6e80a470  // ummla v16.4s, v3.16b, v0.16b\n"
       ".inst 0x6e80a458  // ummla v24.4s, v2.16b, v0.16b\n"
       "ldr q0, [x10, #0x20]\n"
@@ -3223,8 +3222,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       ".inst 0x6e81a476  // ummla v22.4s, v3.16b, v1.16b\n"
       ".inst 0x6e81a45e  // ummla v30.4s, v2.16b, v1.16b\n"
       "ldr q6, [x10, #0x70]\n"
-      "add x10, x10, #0x80\n"
       ".inst 0x6e80a4eb  // ummla v11.4s, v7.16b, v0.16b\n"
+      "add x10, x10, #0x80\n"
       ".inst 0x6e80a473  // ummla v19.4s, v3.16b, v0.16b\n"
       ".inst 0x6e80a45b  // ummla v27.4s, v2.16b, v0.16b\n"
       ".inst 0x6e86a4ef  // ummla v15.4s, v7.16b, v6.16b\n"
@@ -3236,32 +3235,32 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "cmp x28, x20\n"
       "bne 200b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "cmp x11, #0x10\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 v7.2d, v8.2d, v12.2d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 v8.2d, v8.2d, v12.2d\n"
       "uzp1 v12.2d, v9.2d, v13.2d\n"
+      "add x20, x21, x20, LSL #2\n"
+      "cmp x11, #0x10\n"
       "uzp2 v9.2d, v9.2d, v13.2d\n"
-      "prfm pstl1keep, [x9, #0x0]\n"
       "uzp1 v13.2d, v10.2d, v14.2d\n"
       "uzp2 v10.2d, v10.2d, v14.2d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 v14.2d, v11.2d, v15.2d\n"
-      "uzp2 v11.2d, v11.2d, v15.2d\n"
+      "prfm pstl1keep, [x9, #0x0]\n"
       "prfm pstl1keep, [x24, #0x0]\n"
-      "add x21, x22, x20, LSL #2\n"
+      "uzp2 v11.2d, v11.2d, v15.2d\n"
       "uzp1 v15.2d, v16.2d, v20.2d\n"
-      "uzp2 v16.2d, v16.2d, v20.2d\n"
       "prfm pstl1keep, [x23, #0x0]\n"
-      "add x20, x21, x20, LSL #2\n"
+      "prfm pstl1keep, [x22, #0x0]\n"
+      "uzp2 v16.2d, v16.2d, v20.2d\n"
       "uzp1 v20.2d, v17.2d, v21.2d\n"
+      "prfm pstl1keep, [x21, #0x0]\n"
+      "prfm pstl1keep, [x20, #0x0]\n"
       "uzp2 v17.2d, v17.2d, v21.2d\n"
-      "prfm pstl1keep, [x22, #0x0]\n"
       "uzp1 v21.2d, v18.2d, v22.2d\n"
       "uzp2 v18.2d, v18.2d, v22.2d\n"
-      "prfm pstl1keep, [x21, #0x0]\n"
-      "prfm pstl1keep, [x20, #0x0]\n"
       "uzp1 v22.2d, v19.2d, v23.2d\n"
       "uzp2 v19.2d, v19.2d, v23.2d\n"
       "uzp1 v23.2d, v24.2d, v28.2d\n"
@@ -3440,8 +3439,8 @@ void a64_hybrid_u8u32_mmla_6x16 (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "224:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12.hpp
index 13188f0e4d..0a97c405ac 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12.hpp
@@ -40,8 +40,7 @@ void a64_interleaved_bf16fp32_dot_8x12( ARGLIST );
 class cls_a64_interleaved_bf16fp32_dot_8x12
 {
 public:
-    typedef bfloat16 lhs_operand_type;
-    typedef bfloat16 rhs_operand_type;
+    typedef bfloat16 operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 2> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 2, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 2> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 2, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp
index ccfd19db36..7ab854a3fe 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp
@@ -55,30 +55,30 @@ void a64_interleaved_bf16fp32_dot_8x12(
       "ldr q4, [x22, #0x0]\n"
       "ldr q5, [x22, #0x10]\n"
       "mov %x[Apanel], x21\n"
+      "ldr q0, [%x[Apanel], #0x0]\n"
+      "ldr q1, [%x[Apanel], #0x10]\n"
+      "movi v8.16b, #0x0\n"
       "ldr q6, [x22, #0x20]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
-      "movi v8.16b, #0x0\n"
+      "cmp x20, #0x2\n"
       "movi v9.16b, #0x0\n"
+      "prfm pldl1keep, [%x[Apanel], #0x0]\n"
       "movi v10.16b, #0x0\n"
-      "ldr q0, [%x[Apanel], #0x0]\n"
-      "ldr q1, [%x[Apanel], #0x10]\n"
       "movi v11.16b, #0x0\n"
-      "cmp x20, #0x2\n"
+      "prfm pldl1keep, [x22, #0x0]\n"
       "movi v12.16b, #0x0\n"
       "movi v13.16b, #0x0\n"
-      "prfm pldl1keep, [%x[Apanel], #0x0]\n"
+      "prfm pldl1keep, [x22, #0x40]\n"
       "movi v14.16b, #0x0\n"
       "movi v15.16b, #0x0\n"
-      "prfm pldl1keep, [x22, #0x0]\n"
+      "prfm pldl1keep, [%x[Apanel], #0x40]\n"
       "movi v16.16b, #0x0\n"
       "movi v17.16b, #0x0\n"
-      "prfm pldl1keep, [x22, #0x40]\n"
+      "prfm pldl1keep, [x22, #0x80]\n"
       "movi v18.16b, #0x0\n"
       "movi v19.16b, #0x0\n"
-      "prfm pldl1keep, [%x[Apanel], #0x40]\n"
       "movi v20.16b, #0x0\n"
       "movi v21.16b, #0x0\n"
-      "prfm pldl1keep, [x22, #0x80]\n"
       "movi v22.16b, #0x0\n"
       "movi v23.16b, #0x0\n"
       "movi v24.16b, #0x0\n"
@@ -159,9 +159,9 @@ void a64_interleaved_bf16fp32_dot_8x12(
       "bge 3b\n"
       "4:"  // main loop skip
       "add %x[Apanel], %x[Apanel], #0x20\n"
-      "add x22, x22, #0x30\n"
       ".inst 0x4f40f088  // bfdot v8.4s, v4.8h, v0.h[0]\n"
       ".inst 0x4f60f08b  // bfdot v11.4s, v4.8h, v0.h[1]\n"
+      "add x22, x22, #0x30\n"
       ".inst 0x4f40f88e  // bfdot v14.4s, v4.8h, v0.h[2]\n"
       ".inst 0x4f60f891  // bfdot v17.4s, v4.8h, v0.h[3]\n"
       ".inst 0x4f41f094  // bfdot v20.4s, v4.8h, v1.h[0]\n"
@@ -190,13 +190,13 @@ void a64_interleaved_bf16fp32_dot_8x12(
       "add %x[Apanel], %x[Apanel], #0x20\n"
       "ldr q2, [x22, #0x0]\n"
       "ldr q1, [x22, #0x10]\n"
-      "ldr q0, [x22, #0x20]\n"
-      "add x22, x22, #0x30\n"
       ".inst 0x4f44f048  // bfdot v8.4s, v2.8h, v4.h[0]\n"
+      "ldr q0, [x22, #0x20]\n"
       ".inst 0x4f64f04b  // bfdot v11.4s, v2.8h, v4.h[1]\n"
       ".inst 0x4f44f84e  // bfdot v14.4s, v2.8h, v4.h[2]\n"
       ".inst 0x4f64f851  // bfdot v17.4s, v2.8h, v4.h[3]\n"
       ".inst 0x4f43f054  // bfdot v20.4s, v2.8h, v3.h[0]\n"
+      "add x22, x22, #0x30\n"
       ".inst 0x4f63f057  // bfdot v23.4s, v2.8h, v3.h[1]\n"
       ".inst 0x4f43f85a  // bfdot v26.4s, v2.8h, v3.h[2]\n"
       ".inst 0x4f63f85d  // bfdot v29.4s, v2.8h, v3.h[3]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp
index 7a270b6082..0a46f26c55 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp
@@ -41,8 +41,7 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510( ARGLIST );
 class cls_a64_interleaved_bf16fp32_mmla_8x12
 {
 public:
-    typedef bfloat16 lhs_operand_type;
-    typedef bfloat16 rhs_operand_type;
+    typedef bfloat16 operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -64,8 +63,8 @@ public:
     }
 
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 4> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp
index 1513d378ca..6eaac71e5f 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp
@@ -54,18 +54,18 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
       "2:"  // Width loop
       "ldp q4, q5, [x22], #0x20\n"
       "mov %x[Apanel], x21\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "ld1 { v0.8h }, [%x[Apanel]], #0x10\n"
+      "ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
       "movi v8.16b, #0x0\n"
+      "ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v9.16b, #0x0\n"
       "movi v10.16b, #0x0\n"
       "movi v11.16b, #0x0\n"
-      "ld1 { v0.8h }, [%x[Apanel]], #0x10\n"
       "movi v12.16b, #0x0\n"
       "movi v13.16b, #0x0\n"
-      "ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
-      "cmp x20, #0x2\n"
       "movi v14.16b, #0x0\n"
-      "ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
       "movi v15.16b, #0x0\n"
       "movi v16.16b, #0x0\n"
       "movi v17.16b, #0x0\n"
@@ -97,7 +97,7 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
       ".inst 0x6e44ecda  // bfmmla v26.4s, v6.8h, v4.8h\n"
       "cmp x20, #0x2\n"
       ".inst 0x6e45ecdd  // bfmmla v29.4s, v6.8h, v5.8h\n"
-      "ldp q5, q4, [x22], #0x20\n"
+      "ldp q4, q5, [x22], #0x20\n"
       ".inst 0x6e43ec09  // bfmmla v9.4s, v0.8h, v3.8h\n"
       ".inst 0x6e47ec0c  // bfmmla v12.4s, v0.8h, v7.8h\n"
       ".inst 0x6e43ec2f  // bfmmla v15.4s, v1.8h, v3.8h\n"
@@ -106,28 +106,28 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
       ".inst 0x6e47ec58  // bfmmla v24.4s, v2.8h, v7.8h\n"
       ".inst 0x6e43ecdb  // bfmmla v27.4s, v6.8h, v3.8h\n"
       ".inst 0x6e47ecde  // bfmmla v30.4s, v6.8h, v7.8h\n"
-      "ldp q3, q7, [x22], #0x20\n"
-      ".inst 0x6e45ec0a  // bfmmla v10.4s, v0.8h, v5.8h\n"
-      ".inst 0x6e44ec0d  // bfmmla v13.4s, v0.8h, v4.8h\n"
+      "ldp q7, q3, [x22], #0x20\n"
+      ".inst 0x6e44ec0a  // bfmmla v10.4s, v0.8h, v4.8h\n"
+      ".inst 0x6e45ec0d  // bfmmla v13.4s, v0.8h, v5.8h\n"
       "ld1 { v0.8h }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e45ec30  // bfmmla v16.4s, v1.8h, v5.8h\n"
-      ".inst 0x6e44ec33  // bfmmla v19.4s, v1.8h, v4.8h\n"
-      ".inst 0x6e45ec56  // bfmmla v22.4s, v2.8h, v5.8h\n"
+      ".inst 0x6e44ec30  // bfmmla v16.4s, v1.8h, v4.8h\n"
+      ".inst 0x6e45ec33  // bfmmla v19.4s, v1.8h, v5.8h\n"
       "ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e44ec59  // bfmmla v25.4s, v2.8h, v4.8h\n"
-      ".inst 0x6e45ecdc  // bfmmla v28.4s, v6.8h, v5.8h\n"
+      ".inst 0x6e44ec56  // bfmmla v22.4s, v2.8h, v4.8h\n"
+      ".inst 0x6e45ec59  // bfmmla v25.4s, v2.8h, v5.8h\n"
       "ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e44ecdf  // bfmmla v31.4s, v6.8h, v4.8h\n"
-      ".inst 0x6e43ec08  // bfmmla v8.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e44ecdc  // bfmmla v28.4s, v6.8h, v4.8h\n"
+      ".inst 0x6e45ecdf  // bfmmla v31.4s, v6.8h, v5.8h\n"
       "ld1 { v6.8h }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e47ec0b  // bfmmla v11.4s, v0.8h, v7.8h\n"
+      ".inst 0x6e47ec08  // bfmmla v8.4s, v0.8h, v7.8h\n"
       "ldp q4, q5, [x22], #0x20\n"
-      ".inst 0x6e43ec2e  // bfmmla v14.4s, v1.8h, v3.8h\n"
-      ".inst 0x6e47ec31  // bfmmla v17.4s, v1.8h, v7.8h\n"
-      ".inst 0x6e43ec54  // bfmmla v20.4s, v2.8h, v3.8h\n"
-      ".inst 0x6e47ec57  // bfmmla v23.4s, v2.8h, v7.8h\n"
-      ".inst 0x6e43ecda  // bfmmla v26.4s, v6.8h, v3.8h\n"
-      ".inst 0x6e47ecdd  // bfmmla v29.4s, v6.8h, v7.8h\n"
+      ".inst 0x6e43ec0b  // bfmmla v11.4s, v0.8h, v3.8h\n"
+      ".inst 0x6e47ec2e  // bfmmla v14.4s, v1.8h, v7.8h\n"
+      ".inst 0x6e43ec31  // bfmmla v17.4s, v1.8h, v3.8h\n"
+      ".inst 0x6e47ec54  // bfmmla v20.4s, v2.8h, v7.8h\n"
+      ".inst 0x6e43ec57  // bfmmla v23.4s, v2.8h, v3.8h\n"
+      ".inst 0x6e47ecda  // bfmmla v26.4s, v6.8h, v7.8h\n"
+      ".inst 0x6e43ecdd  // bfmmla v29.4s, v6.8h, v3.8h\n"
       "ldp q7, q3, [x22], #0x20\n"
       ".inst 0x6e44ec09  // bfmmla v9.4s, v0.8h, v4.8h\n"
       ".inst 0x6e45ec0c  // bfmmla v12.4s, v0.8h, v5.8h\n"
@@ -143,11 +143,11 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
       "ld1 { v0.8h }, [%x[Apanel]], #0x10\n"
       ".inst 0x6e47ec30  // bfmmla v16.4s, v1.8h, v7.8h\n"
       ".inst 0x6e43ec33  // bfmmla v19.4s, v1.8h, v3.8h\n"
-      ".inst 0x6e47ec56  // bfmmla v22.4s, v2.8h, v7.8h\n"
       "ld1 { v1.8h }, [%x[Apanel]], #0x10\n"
+      ".inst 0x6e47ec56  // bfmmla v22.4s, v2.8h, v7.8h\n"
       ".inst 0x6e43ec59  // bfmmla v25.4s, v2.8h, v3.8h\n"
-      ".inst 0x6e47ecdc  // bfmmla v28.4s, v6.8h, v7.8h\n"
       "ld1 { v2.8h }, [%x[Apanel]], #0x10\n"
+      ".inst 0x6e47ecdc  // bfmmla v28.4s, v6.8h, v7.8h\n"
       ".inst 0x6e43ecdf  // bfmmla v31.4s, v6.8h, v3.8h\n"
       "bge 3b\n"
       "4:"  // main loop skip
@@ -182,9 +182,9 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
       "ldp q1, q0, [x22], #0x20\n"
       "ld1 { v7.8h }, [%x[Apanel]], #0x10\n"
       "ld1 { v6.8h }, [%x[Apanel]], #0x10\n"
+      ".inst 0x6e41ece8  // bfmmla v8.4s, v7.8h, v1.8h\n"
       "ld1 { v5.8h }, [%x[Apanel]], #0x10\n"
       "ld1 { v4.8h }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e41ece8  // bfmmla v8.4s, v7.8h, v1.8h\n"
       ".inst 0x6e40eceb  // bfmmla v11.4s, v7.8h, v0.8h\n"
       "ldp q3, q2, [x22], #0x20\n"
       ".inst 0x6e41ecce  // bfmmla v14.4s, v6.8h, v1.8h\n"
@@ -212,41 +212,41 @@ void a64_interleaved_bf16fp32_mmla_8x12_a510(
       ".inst 0x6e40ec9f  // bfmmla v31.4s, v4.8h, v0.8h\n"
       "5:"  // multiply loop done
       "subs x23, x23, #0x1\n"
-      "uzp1 v2.2d, v8.2d, v11.2d\n"
+      "uzp1 v0.2d, v8.2d, v11.2d\n"
       "uzp2 v8.2d, v8.2d, v11.2d\n"
       "uzp1 v1.2d, v9.2d, v12.2d\n"
       "uzp2 v9.2d, v9.2d, v12.2d\n"
+      "str q0, [%x[Cpanel], #0x0]\n"
       "uzp1 v0.2d, v10.2d, v13.2d\n"
       "uzp2 v10.2d, v10.2d, v13.2d\n"
-      "str q2, [%x[Cpanel], #0x0]\n"
-      "uzp1 v3.2d, v14.2d, v17.2d\n"
-      "uzp2 v14.2d, v14.2d, v17.2d\n"
       "str q1, [%x[Cpanel], #0x10]\n"
+      "str q0, [%x[Cpanel], #0x20]\n"
+      "uzp1 v0.2d, v14.2d, v17.2d\n"
+      "uzp2 v14.2d, v14.2d, v17.2d\n"
+      "str q8, [%x[Cpanel], #0x30]\n"
       "uzp1 v2.2d, v15.2d, v18.2d\n"
       "uzp2 v15.2d, v15.2d, v18.2d\n"
-      "str q0, [%x[Cpanel], #0x20]\n"
+      "str q9, [%x[Cpanel], #0x40]\n"
       "uzp1 v17.2d, v16.2d, v19.2d\n"
       "uzp2 v16.2d, v16.2d, v19.2d\n"
-      "str q8, [%x[Cpanel], #0x30]\n"
+      "str q10, [%x[Cpanel], #0x50]\n"
       "uzp1 v1.2d, v20.2d, v23.2d\n"
       "uzp2 v20.2d, v20.2d, v23.2d\n"
-      "str q9, [%x[Cpanel], #0x40]\n"
+      "str q0, [%x[Cpanel], #0x60]\n"
       "uzp1 v0.2d, v21.2d, v24.2d\n"
       "uzp2 v21.2d, v21.2d, v24.2d\n"
-      "str q10, [%x[Cpanel], #0x50]\n"
+      "str q2, [%x[Cpanel], #0x70]\n"
       "uzp1 v23.2d, v22.2d, v25.2d\n"
       "uzp2 v22.2d, v22.2d, v25.2d\n"
-      "str q3, [%x[Cpanel], #0x60]\n"
+      "str q17, [%x[Cpanel], #0x80]\n"
       "uzp1 v19.2d, v26.2d, v29.2d\n"
       "uzp2 v26.2d, v26.2d, v29.2d\n"
-      "str q2, [%x[Cpanel], #0x70]\n"
+      "str q14, [%x[Cpanel], #0x90]\n"
       "uzp1 v18.2d, v27.2d, v30.2d\n"
       "uzp2 v27.2d, v27.2d, v30.2d\n"
-      "str q17, [%x[Cpanel], #0x80]\n"
+      "str q15, [%x[Cpanel], #0xa0]\n"
       "uzp1 v17.2d, v28.2d, v31.2d\n"
       "uzp2 v28.2d, v28.2d, v31.2d\n"
-      "str q14, [%x[Cpanel], #0x90]\n"
-      "str q15, [%x[Cpanel], #0xa0]\n"
       "str q16, [%x[Cpanel], #0xb0]\n"
       "str q1, [%x[Cpanel], #0xc0]\n"
       "str q0, [%x[Cpanel], #0xd0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp
index f4493c6855..8360c9691b 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp
@@ -55,21 +55,21 @@ void a64_interleaved_bf16fp32_mmla_8x12(
       "ldr q4, [x22, #0x0]\n"
       "ldr q5, [x22, #0x10]\n"
       "mov %x[Apanel], x21\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "ldr q0, [%x[Apanel], #0x0]\n"
+      "ldr q1, [%x[Apanel], #0x10]\n"
       "movi v8.16b, #0x0\n"
+      "ldr q2, [%x[Apanel], #0x20]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v9.16b, #0x0\n"
       "movi v10.16b, #0x0\n"
-      "movi v11.16b, #0x0\n"
       "add x22, x22, #0x20\n"
-      "ldr q0, [%x[Apanel], #0x0]\n"
-      "ldr q1, [%x[Apanel], #0x10]\n"
+      "movi v11.16b, #0x0\n"
       "movi v12.16b, #0x0\n"
-      "ldr q2, [%x[Apanel], #0x20]\n"
-      "cmp x20, #0x2\n"
+      "add %x[Apanel], %x[Apanel], #0x30\n"
       "movi v13.16b, #0x0\n"
       "movi v14.16b, #0x0\n"
       "movi v15.16b, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
       "movi v16.16b, #0x0\n"
       "movi v17.16b, #0x0\n"
       "movi v18.16b, #0x0\n"
@@ -196,19 +196,19 @@ void a64_interleaved_bf16fp32_mmla_8x12(
       "cbz x20, 5f\n"
       "ldr q1, [x22, #0x0]\n"
       "ldr q7, [%x[Apanel], #0x0]\n"
+      ".inst 0x6e41ece8  // bfmmla v8.4s, v7.8h, v1.8h\n"
       "ldr q6, [%x[Apanel], #0x10]\n"
       "ldr q0, [x22, #0x10]\n"
+      ".inst 0x6e40eceb  // bfmmla v11.4s, v7.8h, v0.8h\n"
       "ldr q5, [%x[Apanel], #0x20]\n"
       "ldr q4, [%x[Apanel], #0x30]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
+      ".inst 0x6e41ecce  // bfmmla v14.4s, v6.8h, v1.8h\n"
       "ldr q3, [x22, #0x20]\n"
       "ldr q2, [x22, #0x30]\n"
-      ".inst 0x6e41ece8  // bfmmla v8.4s, v7.8h, v1.8h\n"
-      ".inst 0x6e40eceb  // bfmmla v11.4s, v7.8h, v0.8h\n"
-      ".inst 0x6e41ecce  // bfmmla v14.4s, v6.8h, v1.8h\n"
       ".inst 0x6e40ecd1  // bfmmla v17.4s, v6.8h, v0.8h\n"
       ".inst 0x6e41ecb4  // bfmmla v20.4s, v5.8h, v1.8h\n"
       ".inst 0x6e40ecb7  // bfmmla v23.4s, v5.8h, v0.8h\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
       ".inst 0x6e41ec9a  // bfmmla v26.4s, v4.8h, v1.8h\n"
       "ldr q1, [x22, #0x40]\n"
       ".inst 0x6e40ec9d  // bfmmla v29.4s, v4.8h, v0.8h\n"
@@ -216,8 +216,8 @@ void a64_interleaved_bf16fp32_mmla_8x12(
       ".inst 0x6e43ece9  // bfmmla v9.4s, v7.8h, v3.8h\n"
       ".inst 0x6e42ecec  // bfmmla v12.4s, v7.8h, v2.8h\n"
       ".inst 0x6e43eccf  // bfmmla v15.4s, v6.8h, v3.8h\n"
-      "add x22, x22, #0x60\n"
       ".inst 0x6e42ecd2  // bfmmla v18.4s, v6.8h, v2.8h\n"
+      "add x22, x22, #0x60\n"
       ".inst 0x6e43ecb5  // bfmmla v21.4s, v5.8h, v3.8h\n"
       ".inst 0x6e42ecb8  // bfmmla v24.4s, v5.8h, v2.8h\n"
       ".inst 0x6e43ec9b  // bfmmla v27.4s, v4.8h, v3.8h\n"
@@ -232,41 +232,41 @@ void a64_interleaved_bf16fp32_mmla_8x12(
       ".inst 0x6e40ec9f  // bfmmla v31.4s, v4.8h, v0.8h\n"
       "5:"  // multiply loop done
       "subs x23, x23, #0x1\n"
-      "uzp1 v2.2d, v8.2d, v11.2d\n"
+      "uzp1 v0.2d, v8.2d, v11.2d\n"
       "uzp2 v8.2d, v8.2d, v11.2d\n"
       "uzp1 v1.2d, v9.2d, v12.2d\n"
       "uzp2 v9.2d, v9.2d, v12.2d\n"
+      "str q0, [%x[Cpanel], #0x0]\n"
       "uzp1 v0.2d, v10.2d, v13.2d\n"
       "uzp2 v10.2d, v10.2d, v13.2d\n"
-      "str q2, [%x[Cpanel], #0x0]\n"
-      "uzp1 v3.2d, v14.2d, v17.2d\n"
-      "uzp2 v14.2d, v14.2d, v17.2d\n"
       "str q1, [%x[Cpanel], #0x10]\n"
+      "str q0, [%x[Cpanel], #0x20]\n"
+      "uzp1 v0.2d, v14.2d, v17.2d\n"
+      "uzp2 v14.2d, v14.2d, v17.2d\n"
+      "str q8, [%x[Cpanel], #0x30]\n"
       "uzp1 v2.2d, v15.2d, v18.2d\n"
       "uzp2 v15.2d, v15.2d, v18.2d\n"
-      "str q0, [%x[Cpanel], #0x20]\n"
+      "str q9, [%x[Cpanel], #0x40]\n"
       "uzp1 v17.2d, v16.2d, v19.2d\n"
       "uzp2 v16.2d, v16.2d, v19.2d\n"
-      "str q8, [%x[Cpanel], #0x30]\n"
+      "str q10, [%x[Cpanel], #0x50]\n"
       "uzp1 v1.2d, v20.2d, v23.2d\n"
       "uzp2 v20.2d, v20.2d, v23.2d\n"
-      "str q9, [%x[Cpanel], #0x40]\n"
+      "str q0, [%x[Cpanel], #0x60]\n"
       "uzp1 v0.2d, v21.2d, v24.2d\n"
       "uzp2 v21.2d, v21.2d, v24.2d\n"
-      "str q10, [%x[Cpanel], #0x50]\n"
+      "str q2, [%x[Cpanel], #0x70]\n"
       "uzp1 v23.2d, v22.2d, v25.2d\n"
       "uzp2 v22.2d, v22.2d, v25.2d\n"
-      "str q3, [%x[Cpanel], #0x60]\n"
+      "str q17, [%x[Cpanel], #0x80]\n"
       "uzp1 v19.2d, v26.2d, v29.2d\n"
       "uzp2 v26.2d, v26.2d, v29.2d\n"
-      "str q2, [%x[Cpanel], #0x70]\n"
+      "str q14, [%x[Cpanel], #0x90]\n"
       "uzp1 v18.2d, v27.2d, v30.2d\n"
       "uzp2 v27.2d, v27.2d, v30.2d\n"
-      "str q17, [%x[Cpanel], #0x80]\n"
+      "str q15, [%x[Cpanel], #0xa0]\n"
       "uzp1 v17.2d, v28.2d, v31.2d\n"
       "uzp2 v28.2d, v28.2d, v31.2d\n"
-      "str q14, [%x[Cpanel], #0x90]\n"
-      "str q15, [%x[Cpanel], #0xa0]\n"
       "str q16, [%x[Cpanel], #0xb0]\n"
       "str q1, [%x[Cpanel], #0xc0]\n"
       "str q0, [%x[Cpanel], #0xd0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12.hpp
index 7c26bfa682..94b5bcc0a8 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12.hpp
@@ -40,8 +40,7 @@ void a64_interleaved_s8s32_mmla_8x12_a510( ARGLIST );
 class cls_a64_interleaved_s8s32_mmla_8x12
 {
 public:
-    typedef int8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
+    typedef int8_t operand_type;
     typedef int32_t result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 8> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 8, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 8> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 8, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp
index ba169540c2..2cd659d033 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp
@@ -54,18 +54,18 @@ void a64_interleaved_s8s32_mmla_8x12_a510(
       "2:"  // Width loop
       "ldp q4, q5, [x22], #0x20\n"
       "mov %x[Apanel], x21\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "ld1 { v0.16b }, [%x[Apanel]], #0x10\n"
+      "ld1 { v1.16b }, [%x[Apanel]], #0x10\n"
       "movi v8.4s, #0x0\n"
+      "ld1 { v2.16b }, [%x[Apanel]], #0x10\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v9.4s, #0x0\n"
       "movi v10.4s, #0x0\n"
       "movi v11.4s, #0x0\n"
-      "ld1 { v0.16b }, [%x[Apanel]], #0x10\n"
       "movi v12.4s, #0x0\n"
       "movi v13.4s, #0x0\n"
-      "ld1 { v1.16b }, [%x[Apanel]], #0x10\n"
-      "cmp x20, #0x2\n"
       "movi v14.4s, #0x0\n"
-      "ld1 { v2.16b }, [%x[Apanel]], #0x10\n"
       "movi v15.4s, #0x0\n"
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -97,7 +97,7 @@ void a64_interleaved_s8s32_mmla_8x12_a510(
       ".inst 0x4e84a4da  // smmla v26.4s, v6.16b, v4.16b\n"
       "cmp x20, #0x2\n"
       ".inst 0x4e85a4dd  // smmla v29.4s, v6.16b, v5.16b\n"
-      "ldp q5, q4, [x22], #0x20\n"
+      "ldp q4, q5, [x22], #0x20\n"
       ".inst 0x4e83a409  // smmla v9.4s, v0.16b, v3.16b\n"
       ".inst 0x4e87a40c  // smmla v12.4s, v0.16b, v7.16b\n"
       ".inst 0x4e83a42f  // smmla v15.4s, v1.16b, v3.16b\n"
@@ -106,28 +106,28 @@ void a64_interleaved_s8s32_mmla_8x12_a510(
       ".inst 0x4e87a458  // smmla v24.4s, v2.16b, v7.16b\n"
       ".inst 0x4e83a4db  // smmla v27.4s, v6.16b, v3.16b\n"
       ".inst 0x4e87a4de  // smmla v30.4s, v6.16b, v7.16b\n"
-      "ldp q3, q7, [x22], #0x20\n"
-      ".inst 0x4e85a40a  // smmla v10.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e84a40d  // smmla v13.4s, v0.16b, v4.16b\n"
+      "ldp q7, q3, [x22], #0x20\n"
+      ".inst 0x4e84a40a  // smmla v10.4s, v0.16b, v4.16b\n"
+      ".inst 0x4e85a40d  // smmla v13.4s, v0.16b, v5.16b\n"
       "ld1 { v0.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x4e85a430  // smmla v16.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e84a433  // smmla v19.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e85a456  // smmla v22.4s, v2.16b, v5.16b\n"
+      ".inst 0x4e84a430  // smmla v16.4s, v1.16b, v4.16b\n"
+      ".inst 0x4e85a433  // smmla v19.4s, v1.16b, v5.16b\n"
       "ld1 { v1.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x4e84a459  // smmla v25.4s, v2.16b, v4.16b\n"
-      ".inst 0x4e85a4dc  // smmla v28.4s, v6.16b, v5.16b\n"
+      ".inst 0x4e84a456  // smmla v22.4s, v2.16b, v4.16b\n"
+      ".inst 0x4e85a459  // smmla v25.4s, v2.16b, v5.16b\n"
       "ld1 { v2.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x4e84a4df  // smmla v31.4s, v6.16b, v4.16b\n"
-      ".inst 0x4e83a408  // smmla v8.4s, v0.16b, v3.16b\n"
+      ".inst 0x4e84a4dc  // smmla v28.4s, v6.16b, v4.16b\n"
+      ".inst 0x4e85a4df  // smmla v31.4s, v6.16b, v5.16b\n"
       "ld1 { v6.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x4e87a40b  // smmla v11.4s, v0.16b, v7.16b\n"
+      ".inst 0x4e87a408  // smmla v8.4s, v0.16b, v7.16b\n"
       "ldp q4, q5, [x22], #0x20\n"
-      ".inst 0x4e83a42e  // smmla v14.4s, v1.16b, v3.16b\n"
-      ".inst 0x4e87a431  // smmla v17.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e83a454  // smmla v20.4s, v2.16b, v3.16b\n"
-      ".inst 0x4e87a457  // smmla v23.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e83a4da  // smmla v26.4s, v6.16b, v3.16b\n"
-      ".inst 0x4e87a4dd  // smmla v29.4s, v6.16b, v7.16b\n"
+      ".inst 0x4e83a40b  // smmla v11.4s, v0.16b, v3.16b\n"
+      ".inst 0x4e87a42e  // smmla v14.4s, v1.16b, v7.16b\n"
+      ".inst 0x4e83a431  // smmla v17.4s, v1.16b, v3.16b\n"
+      ".inst 0x4e87a454  // smmla v20.4s, v2.16b, v7.16b\n"
+      ".inst 0x4e83a457  // smmla v23.4s, v2.16b, v3.16b\n"
+      ".inst 0x4e87a4da  // smmla v26.4s, v6.16b, v7.16b\n"
+      ".inst 0x4e83a4dd  // smmla v29.4s, v6.16b, v3.16b\n"
       "ldp q7, q3, [x22], #0x20\n"
       ".inst 0x4e84a409  // smmla v9.4s, v0.16b, v4.16b\n"
       ".inst 0x4e85a40c  // smmla v12.4s, v0.16b, v5.16b\n"
@@ -143,11 +143,11 @@ void a64_interleaved_s8s32_mmla_8x12_a510(
       "ld1 { v0.16b }, [%x[Apanel]], #0x10\n"
       ".inst 0x4e87a430  // smmla v16.4s, v1.16b, v7.16b\n"
       ".inst 0x4e83a433  // smmla v19.4s, v1.16b, v3.16b\n"
-      ".inst 0x4e87a456  // smmla v22.4s, v2.16b, v7.16b\n"
       "ld1 { v1.16b }, [%x[Apanel]], #0x10\n"
+      ".inst 0x4e87a456  // smmla v22.4s, v2.16b, v7.16b\n"
       ".inst 0x4e83a459  // smmla v25.4s, v2.16b, v3.16b\n"
-      ".inst 0x4e87a4dc  // smmla v28.4s, v6.16b, v7.16b\n"
       "ld1 { v2.16b }, [%x[Apanel]], #0x10\n"
+      ".inst 0x4e87a4dc  // smmla v28.4s, v6.16b, v7.16b\n"
       ".inst 0x4e83a4df  // smmla v31.4s, v6.16b, v3.16b\n"
       "bge 3b\n"
       "4:"  // main loop skip
@@ -182,9 +182,9 @@ void a64_interleaved_s8s32_mmla_8x12_a510(
       "ldp q1, q0, [x22], #0x20\n"
       "ld1 { v7.16b }, [%x[Apanel]], #0x10\n"
       "ld1 { v6.16b }, [%x[Apanel]], #0x10\n"
+      ".inst 0x4e81a4e8  // smmla v8.4s, v7.16b, v1.16b\n"
       "ld1 { v5.16b }, [%x[Apanel]], #0x10\n"
       "ld1 { v4.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x4e81a4e8  // smmla v8.4s, v7.16b, v1.16b\n"
       ".inst 0x4e80a4eb  // smmla v11.4s, v7.16b, v0.16b\n"
       "ldp q3, q2, [x22], #0x20\n"
       ".inst 0x4e81a4ce  // smmla v14.4s, v6.16b, v1.16b\n"
@@ -212,41 +212,41 @@ void a64_interleaved_s8s32_mmla_8x12_a510(
       ".inst 0x4e80a49f  // smmla v31.4s, v4.16b, v0.16b\n"
       "5:"  // multiply loop done
       "subs x23, x23, #0x1\n"
-      "uzp1 v2.2d, v8.2d, v11.2d\n"
+      "uzp1 v0.2d, v8.2d, v11.2d\n"
       "uzp2 v8.2d, v8.2d, v11.2d\n"
       "uzp1 v1.2d, v9.2d, v12.2d\n"
       "uzp2 v9.2d, v9.2d, v12.2d\n"
+      "str q0, [%x[Cpanel], #0x0]\n"
       "uzp1 v0.2d, v10.2d, v13.2d\n"
       "uzp2 v10.2d, v10.2d, v13.2d\n"
-      "str q2, [%x[Cpanel], #0x0]\n"
-      "uzp1 v3.2d, v14.2d, v17.2d\n"
-      "uzp2 v14.2d, v14.2d, v17.2d\n"
       "str q1, [%x[Cpanel], #0x10]\n"
+      "str q0, [%x[Cpanel], #0x20]\n"
+      "uzp1 v0.2d, v14.2d, v17.2d\n"
+      "uzp2 v14.2d, v14.2d, v17.2d\n"
+      "str q8, [%x[Cpanel], #0x30]\n"
       "uzp1 v2.2d, v15.2d, v18.2d\n"
       "uzp2 v15.2d, v15.2d, v18.2d\n"
-      "str q0, [%x[Cpanel], #0x20]\n"
+      "str q9, [%x[Cpanel], #0x40]\n"
       "uzp1 v17.2d, v16.2d, v19.2d\n"
       "uzp2 v16.2d, v16.2d, v19.2d\n"
-      "str q8, [%x[Cpanel], #0x30]\n"
+      "str q10, [%x[Cpanel], #0x50]\n"
       "uzp1 v1.2d, v20.2d, v23.2d\n"
       "uzp2 v20.2d, v20.2d, v23.2d\n"
-      "str q9, [%x[Cpanel], #0x40]\n"
+      "str q0, [%x[Cpanel], #0x60]\n"
       "uzp1 v0.2d, v21.2d, v24.2d\n"
       "uzp2 v21.2d, v21.2d, v24.2d\n"
-      "str q10, [%x[Cpanel], #0x50]\n"
+      "str q2, [%x[Cpanel], #0x70]\n"
       "uzp1 v23.2d, v22.2d, v25.2d\n"
       "uzp2 v22.2d, v22.2d, v25.2d\n"
-      "str q3, [%x[Cpanel], #0x60]\n"
+      "str q17, [%x[Cpanel], #0x80]\n"
       "uzp1 v19.2d, v26.2d, v29.2d\n"
       "uzp2 v26.2d, v26.2d, v29.2d\n"
-      "str q2, [%x[Cpanel], #0x70]\n"
+      "str q14, [%x[Cpanel], #0x90]\n"
       "uzp1 v18.2d, v27.2d, v30.2d\n"
       "uzp2 v27.2d, v27.2d, v30.2d\n"
-      "str q17, [%x[Cpanel], #0x80]\n"
+      "str q15, [%x[Cpanel], #0xa0]\n"
       "uzp1 v17.2d, v28.2d, v31.2d\n"
       "uzp2 v28.2d, v28.2d, v31.2d\n"
-      "str q14, [%x[Cpanel], #0x90]\n"
-      "str q15, [%x[Cpanel], #0xa0]\n"
       "str q16, [%x[Cpanel], #0xb0]\n"
       "str q1, [%x[Cpanel], #0xc0]\n"
       "str q0, [%x[Cpanel], #0xd0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp
index 63c6277719..a0ada9f949 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp
@@ -55,21 +55,21 @@ void a64_interleaved_s8s32_mmla_8x12(
       "ldr q4, [x22, #0x0]\n"
       "ldr q5, [x22, #0x10]\n"
       "mov %x[Apanel], x21\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "ldr q0, [%x[Apanel], #0x0]\n"
+      "ldr q1, [%x[Apanel], #0x10]\n"
       "movi v8.4s, #0x0\n"
+      "ldr q2, [%x[Apanel], #0x20]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v9.4s, #0x0\n"
       "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
       "add x22, x22, #0x20\n"
-      "ldr q0, [%x[Apanel], #0x0]\n"
-      "ldr q1, [%x[Apanel], #0x10]\n"
+      "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
-      "ldr q2, [%x[Apanel], #0x20]\n"
-      "cmp x20, #0x2\n"
+      "add %x[Apanel], %x[Apanel], #0x30\n"
       "movi v13.4s, #0x0\n"
       "movi v14.4s, #0x0\n"
       "movi v15.4s, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
       "movi v18.4s, #0x0\n"
@@ -196,19 +196,19 @@ void a64_interleaved_s8s32_mmla_8x12(
       "cbz x20, 5f\n"
       "ldr q1, [x22, #0x0]\n"
       "ldr q7, [%x[Apanel], #0x0]\n"
+      ".inst 0x4e81a4e8  // smmla v8.4s, v7.16b, v1.16b\n"
       "ldr q6, [%x[Apanel], #0x10]\n"
       "ldr q0, [x22, #0x10]\n"
+      ".inst 0x4e80a4eb  // smmla v11.4s, v7.16b, v0.16b\n"
       "ldr q5, [%x[Apanel], #0x20]\n"
       "ldr q4, [%x[Apanel], #0x30]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
+      ".inst 0x4e81a4ce  // smmla v14.4s, v6.16b, v1.16b\n"
       "ldr q3, [x22, #0x20]\n"
       "ldr q2, [x22, #0x30]\n"
-      ".inst 0x4e81a4e8  // smmla v8.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e80a4eb  // smmla v11.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e81a4ce  // smmla v14.4s, v6.16b, v1.16b\n"
       ".inst 0x4e80a4d1  // smmla v17.4s, v6.16b, v0.16b\n"
       ".inst 0x4e81a4b4  // smmla v20.4s, v5.16b, v1.16b\n"
       ".inst 0x4e80a4b7  // smmla v23.4s, v5.16b, v0.16b\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
       ".inst 0x4e81a49a  // smmla v26.4s, v4.16b, v1.16b\n"
       "ldr q1, [x22, #0x40]\n"
       ".inst 0x4e80a49d  // smmla v29.4s, v4.16b, v0.16b\n"
@@ -216,8 +216,8 @@ void a64_interleaved_s8s32_mmla_8x12(
       ".inst 0x4e83a4e9  // smmla v9.4s, v7.16b, v3.16b\n"
       ".inst 0x4e82a4ec  // smmla v12.4s, v7.16b, v2.16b\n"
       ".inst 0x4e83a4cf  // smmla v15.4s, v6.16b, v3.16b\n"
-      "add x22, x22, #0x60\n"
       ".inst 0x4e82a4d2  // smmla v18.4s, v6.16b, v2.16b\n"
+      "add x22, x22, #0x60\n"
       ".inst 0x4e83a4b5  // smmla v21.4s, v5.16b, v3.16b\n"
       ".inst 0x4e82a4b8  // smmla v24.4s, v5.16b, v2.16b\n"
       ".inst 0x4e83a49b  // smmla v27.4s, v4.16b, v3.16b\n"
@@ -232,41 +232,41 @@ void a64_interleaved_s8s32_mmla_8x12(
       ".inst 0x4e80a49f  // smmla v31.4s, v4.16b, v0.16b\n"
       "5:"  // multiply loop done
       "subs x23, x23, #0x1\n"
-      "uzp1 v2.2d, v8.2d, v11.2d\n"
+      "uzp1 v0.2d, v8.2d, v11.2d\n"
       "uzp2 v8.2d, v8.2d, v11.2d\n"
       "uzp1 v1.2d, v9.2d, v12.2d\n"
       "uzp2 v9.2d, v9.2d, v12.2d\n"
+      "str q0, [%x[Cpanel], #0x0]\n"
       "uzp1 v0.2d, v10.2d, v13.2d\n"
       "uzp2 v10.2d, v10.2d, v13.2d\n"
-      "str q2, [%x[Cpanel], #0x0]\n"
-      "uzp1 v3.2d, v14.2d, v17.2d\n"
-      "uzp2 v14.2d, v14.2d, v17.2d\n"
       "str q1, [%x[Cpanel], #0x10]\n"
+      "str q0, [%x[Cpanel], #0x20]\n"
+      "uzp1 v0.2d, v14.2d, v17.2d\n"
+      "uzp2 v14.2d, v14.2d, v17.2d\n"
+      "str q8, [%x[Cpanel], #0x30]\n"
       "uzp1 v2.2d, v15.2d, v18.2d\n"
       "uzp2 v15.2d, v15.2d, v18.2d\n"
-      "str q0, [%x[Cpanel], #0x20]\n"
+      "str q9, [%x[Cpanel], #0x40]\n"
       "uzp1 v17.2d, v16.2d, v19.2d\n"
       "uzp2 v16.2d, v16.2d, v19.2d\n"
-      "str q8, [%x[Cpanel], #0x30]\n"
+      "str q10, [%x[Cpanel], #0x50]\n"
       "uzp1 v1.2d, v20.2d, v23.2d\n"
       "uzp2 v20.2d, v20.2d, v23.2d\n"
-      "str q9, [%x[Cpanel], #0x40]\n"
+      "str q0, [%x[Cpanel], #0x60]\n"
       "uzp1 v0.2d, v21.2d, v24.2d\n"
       "uzp2 v21.2d, v21.2d, v24.2d\n"
-      "str q10, [%x[Cpanel], #0x50]\n"
+      "str q2, [%x[Cpanel], #0x70]\n"
       "uzp1 v23.2d, v22.2d, v25.2d\n"
       "uzp2 v22.2d, v22.2d, v25.2d\n"
-      "str q3, [%x[Cpanel], #0x60]\n"
+      "str q17, [%x[Cpanel], #0x80]\n"
       "uzp1 v19.2d, v26.2d, v29.2d\n"
       "uzp2 v26.2d, v26.2d, v29.2d\n"
-      "str q2, [%x[Cpanel], #0x70]\n"
+      "str q14, [%x[Cpanel], #0x90]\n"
       "uzp1 v18.2d, v27.2d, v30.2d\n"
       "uzp2 v27.2d, v27.2d, v30.2d\n"
-      "str q17, [%x[Cpanel], #0x80]\n"
+      "str q15, [%x[Cpanel], #0xa0]\n"
       "uzp1 v17.2d, v28.2d, v31.2d\n"
       "uzp2 v28.2d, v28.2d, v31.2d\n"
-      "str q14, [%x[Cpanel], #0x90]\n"
-      "str q15, [%x[Cpanel], #0xa0]\n"
       "str q16, [%x[Cpanel], #0xb0]\n"
       "str q1, [%x[Cpanel], #0xc0]\n"
       "str q0, [%x[Cpanel], #0xd0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8s8s32_mmla_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8s8s32_mmla_8x12.hpp
deleted file mode 100644
index c22303ce06..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8s8s32_mmla_8x12.hpp
+++ /dev/null
@@ -1,107 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef __aarch64__
-
-#include "../std_transforms_fixed.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    const uint8_t *, const int8_t *, \
-    int32_t *, int, int, int
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void a64_interleaved_u8s8s32_mmla_8x12( ARGLIST );
-
-class cls_a64_interleaved_u8s8s32_mmla_8x12
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef int32_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 8;
-    }
-
-    static unsigned int out_width()
-    {
-        return 12;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 8;
-    }
-
-
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 8> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 8, true> transforms_quantized = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-
-        if (std::is_same<T, uint32_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 62.58, 4.06, 8.02 };
-                case CPUModel::A510:
-                    return { 47.83, 3.59, 3.72 };
-                case CPUModel::V1:
-                    return { 111.52, 4.97, 10.80 };
-            }
-        }
-
-
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 62.57, 4.10, 0.51 };
-                case CPUModel::A510:
-                    return { 47.66, 2.47, 0.29 };
-                case CPUModel::V1:
-                    return { 75.54, 8.06, 0.63 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=a64_interleaved_u8s8s32_mmla_8x12;
-    cls_a64_interleaved_u8s8s32_mmla_8x12(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8s8s32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8s8s32_mmla_8x12/generic.cpp
deleted file mode 100644
index b7f2e0c04b..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8s8s32_mmla_8x12/generic.cpp
+++ /dev/null
@@ -1,294 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef __aarch64__
-
-#include <cstddef>
-#include <cstdint>
-
-namespace arm_gemm {
-
-void a64_interleaved_u8s8s32_mmla_8x12(
-    const uint8_t *Apanel,
-    const int8_t *Bpanel,
-    int32_t *Cpanel,
-    int ablocks,
-    int bblocks,
-    int K) {
-
-    struct KernelArgs {
-        size_t K = {};
-        const int8_t *Bpanel = {};
-        size_t bblocks = {};
-    } ka;
-
-    ka.K = (K/8) - 1;
-    ka.Bpanel = Bpanel;
-    ka.bblocks = bblocks;
-
-    __asm__ __volatile__(
-      "1:"  // Height loop
-      "ldr x23, [%x[args_ptr], %[offsetof_bblocks]]\n"
-      "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n"
-      "mov x21, %x[Apanel]\n"
-      "2:"  // Width loop
-      "ldr q4, [x22, #0x0]\n"
-      "ldr q5, [x22, #0x10]\n"
-      "mov %x[Apanel], x21\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
-      "movi v8.4s, #0x0\n"
-      "movi v9.4s, #0x0\n"
-      "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
-      "add x22, x22, #0x20\n"
-      "ldr q0, [%x[Apanel], #0x0]\n"
-      "ldr q1, [%x[Apanel], #0x10]\n"
-      "movi v12.4s, #0x0\n"
-      "ldr q2, [%x[Apanel], #0x20]\n"
-      "cmp x20, #0x2\n"
-      "movi v13.4s, #0x0\n"
-      "movi v14.4s, #0x0\n"
-      "movi v15.4s, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
-      "movi v16.4s, #0x0\n"
-      "movi v17.4s, #0x0\n"
-      "movi v18.4s, #0x0\n"
-      "movi v19.4s, #0x0\n"
-      "movi v20.4s, #0x0\n"
-      "movi v21.4s, #0x0\n"
-      "movi v22.4s, #0x0\n"
-      "movi v23.4s, #0x0\n"
-      "movi v24.4s, #0x0\n"
-      "movi v25.4s, #0x0\n"
-      "movi v26.4s, #0x0\n"
-      "movi v27.4s, #0x0\n"
-      "movi v28.4s, #0x0\n"
-      "movi v29.4s, #0x0\n"
-      "movi v30.4s, #0x0\n"
-      "movi v31.4s, #0x0\n"
-      "blt 4f\n"
-      "3:"  // main loop head
-      "ldr q6, [%x[Apanel], #0x0]\n"
-      "ldr q7, [x22, #0x0]\n"
-      ".inst 0x4e84ac08  // usmmla v8.4s, v0.16b, v4.16b\n"
-      "ldr q3, [x22, #0x10]\n"
-      ".inst 0x4e85ac0b  // usmmla v11.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e84ac2e  // usmmla v14.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e85ac31  // usmmla v17.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e84ac54  // usmmla v20.4s, v2.16b, v4.16b\n"
-      "sub x20, x20, #0x2\n"
-      ".inst 0x4e85ac57  // usmmla v23.4s, v2.16b, v5.16b\n"
-      ".inst 0x4e84acda  // usmmla v26.4s, v6.16b, v4.16b\n"
-      "ldr q4, [x22, #0x20]\n"
-      ".inst 0x4e85acdd  // usmmla v29.4s, v6.16b, v5.16b\n"
-      "ldr q5, [x22, #0x30]\n"
-      ".inst 0x4e87ac09  // usmmla v9.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e83ac0c  // usmmla v12.4s, v0.16b, v3.16b\n"
-      ".inst 0x4e87ac2f  // usmmla v15.4s, v1.16b, v7.16b\n"
-      "cmp x20, #0x2\n"
-      ".inst 0x4e83ac32  // usmmla v18.4s, v1.16b, v3.16b\n"
-      ".inst 0x4e87ac55  // usmmla v21.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e83ac58  // usmmla v24.4s, v2.16b, v3.16b\n"
-      ".inst 0x4e87acdb  // usmmla v27.4s, v6.16b, v7.16b\n"
-      "ldr q7, [x22, #0x40]\n"
-      ".inst 0x4e83acde  // usmmla v30.4s, v6.16b, v3.16b\n"
-      "ldr q3, [x22, #0x50]\n"
-      ".inst 0x4e84ac0a  // usmmla v10.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e85ac0d  // usmmla v13.4s, v0.16b, v5.16b\n"
-      "ldr q0, [%x[Apanel], #0x10]\n"
-      ".inst 0x4e84ac30  // usmmla v16.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e85ac33  // usmmla v19.4s, v1.16b, v5.16b\n"
-      "ldr q1, [%x[Apanel], #0x20]\n"
-      ".inst 0x4e84ac56  // usmmla v22.4s, v2.16b, v4.16b\n"
-      ".inst 0x4e85ac59  // usmmla v25.4s, v2.16b, v5.16b\n"
-      "ldr q2, [%x[Apanel], #0x30]\n"
-      ".inst 0x4e84acdc  // usmmla v28.4s, v6.16b, v4.16b\n"
-      "ldr q4, [x22, #0x60]\n"
-      ".inst 0x4e85acdf  // usmmla v31.4s, v6.16b, v5.16b\n"
-      "ldr q6, [%x[Apanel], #0x40]\n"
-      "ldr q5, [x22, #0x70]\n"
-      ".inst 0x4e87ac08  // usmmla v8.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e83ac0b  // usmmla v11.4s, v0.16b, v3.16b\n"
-      ".inst 0x4e87ac2e  // usmmla v14.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e83ac31  // usmmla v17.4s, v1.16b, v3.16b\n"
-      ".inst 0x4e87ac54  // usmmla v20.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e83ac57  // usmmla v23.4s, v2.16b, v3.16b\n"
-      ".inst 0x4e87acda  // usmmla v26.4s, v6.16b, v7.16b\n"
-      "ldr q7, [x22, #0x80]\n"
-      ".inst 0x4e83acdd  // usmmla v29.4s, v6.16b, v3.16b\n"
-      "ldr q3, [x22, #0x90]\n"
-      ".inst 0x4e84ac09  // usmmla v9.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e85ac0c  // usmmla v12.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e84ac2f  // usmmla v15.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e85ac32  // usmmla v18.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e84ac55  // usmmla v21.4s, v2.16b, v4.16b\n"
-      ".inst 0x4e85ac58  // usmmla v24.4s, v2.16b, v5.16b\n"
-      ".inst 0x4e84acdb  // usmmla v27.4s, v6.16b, v4.16b\n"
-      "ldr q4, [x22, #0xa0]\n"
-      ".inst 0x4e85acde  // usmmla v30.4s, v6.16b, v5.16b\n"
-      "ldr q5, [x22, #0xb0]\n"
-      ".inst 0x4e87ac0a  // usmmla v10.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e83ac0d  // usmmla v13.4s, v0.16b, v3.16b\n"
-      "ldr q0, [%x[Apanel], #0x50]\n"
-      ".inst 0x4e87ac30  // usmmla v16.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e83ac33  // usmmla v19.4s, v1.16b, v3.16b\n"
-      "ldr q1, [%x[Apanel], #0x60]\n"
-      ".inst 0x4e87ac56  // usmmla v22.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e83ac59  // usmmla v25.4s, v2.16b, v3.16b\n"
-      "ldr q2, [%x[Apanel], #0x70]\n"
-      ".inst 0x4e87acdc  // usmmla v28.4s, v6.16b, v7.16b\n"
-      ".inst 0x4e83acdf  // usmmla v31.4s, v6.16b, v3.16b\n"
-      "add %x[Apanel], %x[Apanel], #0x80\n"
-      "add x22, x22, #0xc0\n"
-      "bge 3b\n"
-      "4:"  // main loop skip
-      "ldr q3, [%x[Apanel], #0x0]\n"
-      "ldr q6, [x22, #0x0]\n"
-      ".inst 0x4e84ac08  // usmmla v8.4s, v0.16b, v4.16b\n"
-      "ldr q7, [x22, #0x10]\n"
-      ".inst 0x4e85ac0b  // usmmla v11.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e84ac2e  // usmmla v14.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e85ac31  // usmmla v17.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e84ac54  // usmmla v20.4s, v2.16b, v4.16b\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      ".inst 0x4e85ac57  // usmmla v23.4s, v2.16b, v5.16b\n"
-      ".inst 0x4e84ac7a  // usmmla v26.4s, v3.16b, v4.16b\n"
-      "ldr q4, [x22, #0x20]\n"
-      ".inst 0x4e85ac7d  // usmmla v29.4s, v3.16b, v5.16b\n"
-      "ldr q5, [x22, #0x30]\n"
-      ".inst 0x4e86ac09  // usmmla v9.4s, v0.16b, v6.16b\n"
-      ".inst 0x4e87ac0c  // usmmla v12.4s, v0.16b, v7.16b\n"
-      ".inst 0x4e86ac2f  // usmmla v15.4s, v1.16b, v6.16b\n"
-      "add x22, x22, #0x40\n"
-      ".inst 0x4e87ac32  // usmmla v18.4s, v1.16b, v7.16b\n"
-      ".inst 0x4e86ac55  // usmmla v21.4s, v2.16b, v6.16b\n"
-      ".inst 0x4e87ac58  // usmmla v24.4s, v2.16b, v7.16b\n"
-      ".inst 0x4e86ac7b  // usmmla v27.4s, v3.16b, v6.16b\n"
-      ".inst 0x4e87ac7e  // usmmla v30.4s, v3.16b, v7.16b\n"
-      ".inst 0x4e84ac0a  // usmmla v10.4s, v0.16b, v4.16b\n"
-      ".inst 0x4e85ac0d  // usmmla v13.4s, v0.16b, v5.16b\n"
-      ".inst 0x4e84ac30  // usmmla v16.4s, v1.16b, v4.16b\n"
-      ".inst 0x4e85ac33  // usmmla v19.4s, v1.16b, v5.16b\n"
-      ".inst 0x4e84ac56  // usmmla v22.4s, v2.16b, v4.16b\n"
-      ".inst 0x4e85ac59  // usmmla v25.4s, v2.16b, v5.16b\n"
-      ".inst 0x4e84ac7c  // usmmla v28.4s, v3.16b, v4.16b\n"
-      ".inst 0x4e85ac7f  // usmmla v31.4s, v3.16b, v5.16b\n"
-      "cbz x20, 5f\n"
-      "ldr q1, [x22, #0x0]\n"
-      "ldr q7, [%x[Apanel], #0x0]\n"
-      "ldr q6, [%x[Apanel], #0x10]\n"
-      "ldr q0, [x22, #0x10]\n"
-      "ldr q5, [%x[Apanel], #0x20]\n"
-      "ldr q4, [%x[Apanel], #0x30]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
-      "ldr q3, [x22, #0x20]\n"
-      "ldr q2, [x22, #0x30]\n"
-      ".inst 0x4e81ace8  // usmmla v8.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e80aceb  // usmmla v11.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e81acce  // usmmla v14.4s, v6.16b, v1.16b\n"
-      ".inst 0x4e80acd1  // usmmla v17.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e81acb4  // usmmla v20.4s, v5.16b, v1.16b\n"
-      ".inst 0x4e80acb7  // usmmla v23.4s, v5.16b, v0.16b\n"
-      ".inst 0x4e81ac9a  // usmmla v26.4s, v4.16b, v1.16b\n"
-      "ldr q1, [x22, #0x40]\n"
-      ".inst 0x4e80ac9d  // usmmla v29.4s, v4.16b, v0.16b\n"
-      "ldr q0, [x22, #0x50]\n"
-      ".inst 0x4e83ace9  // usmmla v9.4s, v7.16b, v3.16b\n"
-      ".inst 0x4e82acec  // usmmla v12.4s, v7.16b, v2.16b\n"
-      ".inst 0x4e83accf  // usmmla v15.4s, v6.16b, v3.16b\n"
-      "add x22, x22, #0x60\n"
-      ".inst 0x4e82acd2  // usmmla v18.4s, v6.16b, v2.16b\n"
-      ".inst 0x4e83acb5  // usmmla v21.4s, v5.16b, v3.16b\n"
-      ".inst 0x4e82acb8  // usmmla v24.4s, v5.16b, v2.16b\n"
-      ".inst 0x4e83ac9b  // usmmla v27.4s, v4.16b, v3.16b\n"
-      ".inst 0x4e82ac9e  // usmmla v30.4s, v4.16b, v2.16b\n"
-      ".inst 0x4e81acea  // usmmla v10.4s, v7.16b, v1.16b\n"
-      ".inst 0x4e80aced  // usmmla v13.4s, v7.16b, v0.16b\n"
-      ".inst 0x4e81acd0  // usmmla v16.4s, v6.16b, v1.16b\n"
-      ".inst 0x4e80acd3  // usmmla v19.4s, v6.16b, v0.16b\n"
-      ".inst 0x4e81acb6  // usmmla v22.4s, v5.16b, v1.16b\n"
-      ".inst 0x4e80acb9  // usmmla v25.4s, v5.16b, v0.16b\n"
-      ".inst 0x4e81ac9c  // usmmla v28.4s, v4.16b, v1.16b\n"
-      ".inst 0x4e80ac9f  // usmmla v31.4s, v4.16b, v0.16b\n"
-      "5:"  // multiply loop done
-      "subs x23, x23, #0x1\n"
-      "uzp1 v2.2d, v8.2d, v11.2d\n"
-      "uzp2 v8.2d, v8.2d, v11.2d\n"
-      "uzp1 v1.2d, v9.2d, v12.2d\n"
-      "uzp2 v9.2d, v9.2d, v12.2d\n"
-      "uzp1 v0.2d, v10.2d, v13.2d\n"
-      "uzp2 v10.2d, v10.2d, v13.2d\n"
-      "str q2, [%x[Cpanel], #0x0]\n"
-      "uzp1 v3.2d, v14.2d, v17.2d\n"
-      "uzp2 v14.2d, v14.2d, v17.2d\n"
-      "str q1, [%x[Cpanel], #0x10]\n"
-      "uzp1 v2.2d, v15.2d, v18.2d\n"
-      "uzp2 v15.2d, v15.2d, v18.2d\n"
-      "str q0, [%x[Cpanel], #0x20]\n"
-      "uzp1 v17.2d, v16.2d, v19.2d\n"
-      "uzp2 v16.2d, v16.2d, v19.2d\n"
-      "str q8, [%x[Cpanel], #0x30]\n"
-      "uzp1 v1.2d, v20.2d, v23.2d\n"
-      "uzp2 v20.2d, v20.2d, v23.2d\n"
-      "str q9, [%x[Cpanel], #0x40]\n"
-      "uzp1 v0.2d, v21.2d, v24.2d\n"
-      "uzp2 v21.2d, v21.2d, v24.2d\n"
-      "str q10, [%x[Cpanel], #0x50]\n"
-      "uzp1 v23.2d, v22.2d, v25.2d\n"
-      "uzp2 v22.2d, v22.2d, v25.2d\n"
-      "str q3, [%x[Cpanel], #0x60]\n"
-      "uzp1 v19.2d, v26.2d, v29.2d\n"
-      "uzp2 v26.2d, v26.2d, v29.2d\n"
-      "str q2, [%x[Cpanel], #0x70]\n"
-      "uzp1 v18.2d, v27.2d, v30.2d\n"
-      "uzp2 v27.2d, v27.2d, v30.2d\n"
-      "str q17, [%x[Cpanel], #0x80]\n"
-      "uzp1 v17.2d, v28.2d, v31.2d\n"
-      "uzp2 v28.2d, v28.2d, v31.2d\n"
-      "str q14, [%x[Cpanel], #0x90]\n"
-      "str q15, [%x[Cpanel], #0xa0]\n"
-      "str q16, [%x[Cpanel], #0xb0]\n"
-      "str q1, [%x[Cpanel], #0xc0]\n"
-      "str q0, [%x[Cpanel], #0xd0]\n"
-      "str q23, [%x[Cpanel], #0xe0]\n"
-      "str q20, [%x[Cpanel], #0xf0]\n"
-      "str q21, [%x[Cpanel], #0x100]\n"
-      "str q22, [%x[Cpanel], #0x110]\n"
-      "str q19, [%x[Cpanel], #0x120]\n"
-      "str q18, [%x[Cpanel], #0x130]\n"
-      "str q17, [%x[Cpanel], #0x140]\n"
-      "str q26, [%x[Cpanel], #0x150]\n"
-      "str q27, [%x[Cpanel], #0x160]\n"
-      "str q28, [%x[Cpanel], #0x170]\n"
-      "add %x[Cpanel], %x[Cpanel], #0x180\n"
-      "bgt 2b\n"
-      "subs %x[ablocks], %x[ablocks], #0x1\n"
-      "bne 1b\n"
-      : [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks)
-      : [args_ptr] "r" (&ka), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_bblocks] "I" (offsetof(KernelArgs, bblocks))
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23"
-    );
-}
-
-} // namespace arm_gemm
-#endif // __aarch64__
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12.hpp
index 922438a255..1073d15f01 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12.hpp
@@ -40,8 +40,7 @@ void a64_interleaved_u8u32_mmla_8x12_a510( ARGLIST );
 class cls_a64_interleaved_u8u32_mmla_8x12
 {
 public:
-    typedef uint8_t lhs_operand_type;
-    typedef uint8_t rhs_operand_type;
+    typedef uint8_t operand_type;
     typedef uint32_t result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 8> transforms = {};
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 12, 8, true> transforms_quantized = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 8> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 8, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp
index 8affb0ea86..741fa6ac08 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp
@@ -54,18 +54,18 @@ void a64_interleaved_u8u32_mmla_8x12_a510(
       "2:"  // Width loop
       "ldp q4, q5, [x22], #0x20\n"
       "mov %x[Apanel], x21\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "ld1 { v0.16b }, [%x[Apanel]], #0x10\n"
+      "ld1 { v1.16b }, [%x[Apanel]], #0x10\n"
       "movi v8.4s, #0x0\n"
+      "ld1 { v2.16b }, [%x[Apanel]], #0x10\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v9.4s, #0x0\n"
       "movi v10.4s, #0x0\n"
       "movi v11.4s, #0x0\n"
-      "ld1 { v0.16b }, [%x[Apanel]], #0x10\n"
       "movi v12.4s, #0x0\n"
       "movi v13.4s, #0x0\n"
-      "ld1 { v1.16b }, [%x[Apanel]], #0x10\n"
-      "cmp x20, #0x2\n"
       "movi v14.4s, #0x0\n"
-      "ld1 { v2.16b }, [%x[Apanel]], #0x10\n"
       "movi v15.4s, #0x0\n"
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
@@ -97,7 +97,7 @@ void a64_interleaved_u8u32_mmla_8x12_a510(
       ".inst 0x6e84a4da  // ummla v26.4s, v6.16b, v4.16b\n"
       "cmp x20, #0x2\n"
       ".inst 0x6e85a4dd  // ummla v29.4s, v6.16b, v5.16b\n"
-      "ldp q5, q4, [x22], #0x20\n"
+      "ldp q4, q5, [x22], #0x20\n"
       ".inst 0x6e83a409  // ummla v9.4s, v0.16b, v3.16b\n"
       ".inst 0x6e87a40c  // ummla v12.4s, v0.16b, v7.16b\n"
       ".inst 0x6e83a42f  // ummla v15.4s, v1.16b, v3.16b\n"
@@ -106,28 +106,28 @@ void a64_interleaved_u8u32_mmla_8x12_a510(
       ".inst 0x6e87a458  // ummla v24.4s, v2.16b, v7.16b\n"
       ".inst 0x6e83a4db  // ummla v27.4s, v6.16b, v3.16b\n"
       ".inst 0x6e87a4de  // ummla v30.4s, v6.16b, v7.16b\n"
-      "ldp q3, q7, [x22], #0x20\n"
-      ".inst 0x6e85a40a  // ummla v10.4s, v0.16b, v5.16b\n"
-      ".inst 0x6e84a40d  // ummla v13.4s, v0.16b, v4.16b\n"
+      "ldp q7, q3, [x22], #0x20\n"
+      ".inst 0x6e84a40a  // ummla v10.4s, v0.16b, v4.16b\n"
+      ".inst 0x6e85a40d  // ummla v13.4s, v0.16b, v5.16b\n"
       "ld1 { v0.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e85a430  // ummla v16.4s, v1.16b, v5.16b\n"
-      ".inst 0x6e84a433  // ummla v19.4s, v1.16b, v4.16b\n"
-      ".inst 0x6e85a456  // ummla v22.4s, v2.16b, v5.16b\n"
+      ".inst 0x6e84a430  // ummla v16.4s, v1.16b, v4.16b\n"
+      ".inst 0x6e85a433  // ummla v19.4s, v1.16b, v5.16b\n"
       "ld1 { v1.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e84a459  // ummla v25.4s, v2.16b, v4.16b\n"
-      ".inst 0x6e85a4dc  // ummla v28.4s, v6.16b, v5.16b\n"
+      ".inst 0x6e84a456  // ummla v22.4s, v2.16b, v4.16b\n"
+      ".inst 0x6e85a459  // ummla v25.4s, v2.16b, v5.16b\n"
       "ld1 { v2.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e84a4df  // ummla v31.4s, v6.16b, v4.16b\n"
-      ".inst 0x6e83a408  // ummla v8.4s, v0.16b, v3.16b\n"
+      ".inst 0x6e84a4dc  // ummla v28.4s, v6.16b, v4.16b\n"
+      ".inst 0x6e85a4df  // ummla v31.4s, v6.16b, v5.16b\n"
       "ld1 { v6.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e87a40b  // ummla v11.4s, v0.16b, v7.16b\n"
+      ".inst 0x6e87a408  // ummla v8.4s, v0.16b, v7.16b\n"
       "ldp q4, q5, [x22], #0x20\n"
-      ".inst 0x6e83a42e  // ummla v14.4s, v1.16b, v3.16b\n"
-      ".inst 0x6e87a431  // ummla v17.4s, v1.16b, v7.16b\n"
-      ".inst 0x6e83a454  // ummla v20.4s, v2.16b, v3.16b\n"
-      ".inst 0x6e87a457  // ummla v23.4s, v2.16b, v7.16b\n"
-      ".inst 0x6e83a4da  // ummla v26.4s, v6.16b, v3.16b\n"
-      ".inst 0x6e87a4dd  // ummla v29.4s, v6.16b, v7.16b\n"
+      ".inst 0x6e83a40b  // ummla v11.4s, v0.16b, v3.16b\n"
+      ".inst 0x6e87a42e  // ummla v14.4s, v1.16b, v7.16b\n"
+      ".inst 0x6e83a431  // ummla v17.4s, v1.16b, v3.16b\n"
+      ".inst 0x6e87a454  // ummla v20.4s, v2.16b, v7.16b\n"
+      ".inst 0x6e83a457  // ummla v23.4s, v2.16b, v3.16b\n"
+      ".inst 0x6e87a4da  // ummla v26.4s, v6.16b, v7.16b\n"
+      ".inst 0x6e83a4dd  // ummla v29.4s, v6.16b, v3.16b\n"
       "ldp q7, q3, [x22], #0x20\n"
       ".inst 0x6e84a409  // ummla v9.4s, v0.16b, v4.16b\n"
       ".inst 0x6e85a40c  // ummla v12.4s, v0.16b, v5.16b\n"
@@ -143,11 +143,11 @@ void a64_interleaved_u8u32_mmla_8x12_a510(
       "ld1 { v0.16b }, [%x[Apanel]], #0x10\n"
       ".inst 0x6e87a430  // ummla v16.4s, v1.16b, v7.16b\n"
       ".inst 0x6e83a433  // ummla v19.4s, v1.16b, v3.16b\n"
-      ".inst 0x6e87a456  // ummla v22.4s, v2.16b, v7.16b\n"
       "ld1 { v1.16b }, [%x[Apanel]], #0x10\n"
+      ".inst 0x6e87a456  // ummla v22.4s, v2.16b, v7.16b\n"
       ".inst 0x6e83a459  // ummla v25.4s, v2.16b, v3.16b\n"
-      ".inst 0x6e87a4dc  // ummla v28.4s, v6.16b, v7.16b\n"
       "ld1 { v2.16b }, [%x[Apanel]], #0x10\n"
+      ".inst 0x6e87a4dc  // ummla v28.4s, v6.16b, v7.16b\n"
       ".inst 0x6e83a4df  // ummla v31.4s, v6.16b, v3.16b\n"
       "bge 3b\n"
       "4:"  // main loop skip
@@ -182,9 +182,9 @@ void a64_interleaved_u8u32_mmla_8x12_a510(
       "ldp q1, q0, [x22], #0x20\n"
       "ld1 { v7.16b }, [%x[Apanel]], #0x10\n"
       "ld1 { v6.16b }, [%x[Apanel]], #0x10\n"
+      ".inst 0x6e81a4e8  // ummla v8.4s, v7.16b, v1.16b\n"
       "ld1 { v5.16b }, [%x[Apanel]], #0x10\n"
       "ld1 { v4.16b }, [%x[Apanel]], #0x10\n"
-      ".inst 0x6e81a4e8  // ummla v8.4s, v7.16b, v1.16b\n"
       ".inst 0x6e80a4eb  // ummla v11.4s, v7.16b, v0.16b\n"
       "ldp q3, q2, [x22], #0x20\n"
       ".inst 0x6e81a4ce  // ummla v14.4s, v6.16b, v1.16b\n"
@@ -212,41 +212,41 @@ void a64_interleaved_u8u32_mmla_8x12_a510(
       ".inst 0x6e80a49f  // ummla v31.4s, v4.16b, v0.16b\n"
       "5:"  // multiply loop done
       "subs x23, x23, #0x1\n"
-      "uzp1 v2.2d, v8.2d, v11.2d\n"
+      "uzp1 v0.2d, v8.2d, v11.2d\n"
       "uzp2 v8.2d, v8.2d, v11.2d\n"
       "uzp1 v1.2d, v9.2d, v12.2d\n"
       "uzp2 v9.2d, v9.2d, v12.2d\n"
+      "str q0, [%x[Cpanel], #0x0]\n"
       "uzp1 v0.2d, v10.2d, v13.2d\n"
       "uzp2 v10.2d, v10.2d, v13.2d\n"
-      "str q2, [%x[Cpanel], #0x0]\n"
-      "uzp1 v3.2d, v14.2d, v17.2d\n"
-      "uzp2 v14.2d, v14.2d, v17.2d\n"
       "str q1, [%x[Cpanel], #0x10]\n"
+      "str q0, [%x[Cpanel], #0x20]\n"
+      "uzp1 v0.2d, v14.2d, v17.2d\n"
+      "uzp2 v14.2d, v14.2d, v17.2d\n"
+      "str q8, [%x[Cpanel], #0x30]\n"
       "uzp1 v2.2d, v15.2d, v18.2d\n"
       "uzp2 v15.2d, v15.2d, v18.2d\n"
-      "str q0, [%x[Cpanel], #0x20]\n"
+      "str q9, [%x[Cpanel], #0x40]\n"
       "uzp1 v17.2d, v16.2d, v19.2d\n"
       "uzp2 v16.2d, v16.2d, v19.2d\n"
-      "str q8, [%x[Cpanel], #0x30]\n"
+      "str q10, [%x[Cpanel], #0x50]\n"
       "uzp1 v1.2d, v20.2d, v23.2d\n"
       "uzp2 v20.2d, v20.2d, v23.2d\n"
-      "str q9, [%x[Cpanel], #0x40]\n"
+      "str q0, [%x[Cpanel], #0x60]\n"
       "uzp1 v0.2d, v21.2d, v24.2d\n"
       "uzp2 v21.2d, v21.2d, v24.2d\n"
-      "str q10, [%x[Cpanel], #0x50]\n"
+      "str q2, [%x[Cpanel], #0x70]\n"
       "uzp1 v23.2d, v22.2d, v25.2d\n"
       "uzp2 v22.2d, v22.2d, v25.2d\n"
-      "str q3, [%x[Cpanel], #0x60]\n"
+      "str q17, [%x[Cpanel], #0x80]\n"
       "uzp1 v19.2d, v26.2d, v29.2d\n"
       "uzp2 v26.2d, v26.2d, v29.2d\n"
-      "str q2, [%x[Cpanel], #0x70]\n"
+      "str q14, [%x[Cpanel], #0x90]\n"
       "uzp1 v18.2d, v27.2d, v30.2d\n"
       "uzp2 v27.2d, v27.2d, v30.2d\n"
-      "str q17, [%x[Cpanel], #0x80]\n"
+      "str q15, [%x[Cpanel], #0xa0]\n"
       "uzp1 v17.2d, v28.2d, v31.2d\n"
       "uzp2 v28.2d, v28.2d, v31.2d\n"
-      "str q14, [%x[Cpanel], #0x90]\n"
-      "str q15, [%x[Cpanel], #0xa0]\n"
       "str q16, [%x[Cpanel], #0xb0]\n"
       "str q1, [%x[Cpanel], #0xc0]\n"
       "str q0, [%x[Cpanel], #0xd0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp
index 747572ef84..613c3f09e5 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp
@@ -55,21 +55,21 @@ void a64_interleaved_u8u32_mmla_8x12(
       "ldr q4, [x22, #0x0]\n"
       "ldr q5, [x22, #0x10]\n"
       "mov %x[Apanel], x21\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "ldr q0, [%x[Apanel], #0x0]\n"
+      "ldr q1, [%x[Apanel], #0x10]\n"
       "movi v8.4s, #0x0\n"
+      "ldr q2, [%x[Apanel], #0x20]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "movi v9.4s, #0x0\n"
       "movi v10.4s, #0x0\n"
-      "movi v11.4s, #0x0\n"
       "add x22, x22, #0x20\n"
-      "ldr q0, [%x[Apanel], #0x0]\n"
-      "ldr q1, [%x[Apanel], #0x10]\n"
+      "movi v11.4s, #0x0\n"
       "movi v12.4s, #0x0\n"
-      "ldr q2, [%x[Apanel], #0x20]\n"
-      "cmp x20, #0x2\n"
+      "add %x[Apanel], %x[Apanel], #0x30\n"
       "movi v13.4s, #0x0\n"
       "movi v14.4s, #0x0\n"
       "movi v15.4s, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
       "movi v16.4s, #0x0\n"
       "movi v17.4s, #0x0\n"
       "movi v18.4s, #0x0\n"
@@ -196,19 +196,19 @@ void a64_interleaved_u8u32_mmla_8x12(
       "cbz x20, 5f\n"
       "ldr q1, [x22, #0x0]\n"
       "ldr q7, [%x[Apanel], #0x0]\n"
+      ".inst 0x6e81a4e8  // ummla v8.4s, v7.16b, v1.16b\n"
       "ldr q6, [%x[Apanel], #0x10]\n"
       "ldr q0, [x22, #0x10]\n"
+      ".inst 0x6e80a4eb  // ummla v11.4s, v7.16b, v0.16b\n"
       "ldr q5, [%x[Apanel], #0x20]\n"
       "ldr q4, [%x[Apanel], #0x30]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
+      ".inst 0x6e81a4ce  // ummla v14.4s, v6.16b, v1.16b\n"
       "ldr q3, [x22, #0x20]\n"
       "ldr q2, [x22, #0x30]\n"
-      ".inst 0x6e81a4e8  // ummla v8.4s, v7.16b, v1.16b\n"
-      ".inst 0x6e80a4eb  // ummla v11.4s, v7.16b, v0.16b\n"
-      ".inst 0x6e81a4ce  // ummla v14.4s, v6.16b, v1.16b\n"
       ".inst 0x6e80a4d1  // ummla v17.4s, v6.16b, v0.16b\n"
       ".inst 0x6e81a4b4  // ummla v20.4s, v5.16b, v1.16b\n"
       ".inst 0x6e80a4b7  // ummla v23.4s, v5.16b, v0.16b\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
       ".inst 0x6e81a49a  // ummla v26.4s, v4.16b, v1.16b\n"
       "ldr q1, [x22, #0x40]\n"
       ".inst 0x6e80a49d  // ummla v29.4s, v4.16b, v0.16b\n"
@@ -216,8 +216,8 @@ void a64_interleaved_u8u32_mmla_8x12(
       ".inst 0x6e83a4e9  // ummla v9.4s, v7.16b, v3.16b\n"
       ".inst 0x6e82a4ec  // ummla v12.4s, v7.16b, v2.16b\n"
       ".inst 0x6e83a4cf  // ummla v15.4s, v6.16b, v3.16b\n"
-      "add x22, x22, #0x60\n"
       ".inst 0x6e82a4d2  // ummla v18.4s, v6.16b, v2.16b\n"
+      "add x22, x22, #0x60\n"
       ".inst 0x6e83a4b5  // ummla v21.4s, v5.16b, v3.16b\n"
       ".inst 0x6e82a4b8  // ummla v24.4s, v5.16b, v2.16b\n"
       ".inst 0x6e83a49b  // ummla v27.4s, v4.16b, v3.16b\n"
@@ -232,41 +232,41 @@ void a64_interleaved_u8u32_mmla_8x12(
       ".inst 0x6e80a49f  // ummla v31.4s, v4.16b, v0.16b\n"
       "5:"  // multiply loop done
       "subs x23, x23, #0x1\n"
-      "uzp1 v2.2d, v8.2d, v11.2d\n"
+      "uzp1 v0.2d, v8.2d, v11.2d\n"
       "uzp2 v8.2d, v8.2d, v11.2d\n"
       "uzp1 v1.2d, v9.2d, v12.2d\n"
       "uzp2 v9.2d, v9.2d, v12.2d\n"
+      "str q0, [%x[Cpanel], #0x0]\n"
       "uzp1 v0.2d, v10.2d, v13.2d\n"
       "uzp2 v10.2d, v10.2d, v13.2d\n"
-      "str q2, [%x[Cpanel], #0x0]\n"
-      "uzp1 v3.2d, v14.2d, v17.2d\n"
-      "uzp2 v14.2d, v14.2d, v17.2d\n"
       "str q1, [%x[Cpanel], #0x10]\n"
+      "str q0, [%x[Cpanel], #0x20]\n"
+      "uzp1 v0.2d, v14.2d, v17.2d\n"
+      "uzp2 v14.2d, v14.2d, v17.2d\n"
+      "str q8, [%x[Cpanel], #0x30]\n"
       "uzp1 v2.2d, v15.2d, v18.2d\n"
       "uzp2 v15.2d, v15.2d, v18.2d\n"
-      "str q0, [%x[Cpanel], #0x20]\n"
+      "str q9, [%x[Cpanel], #0x40]\n"
       "uzp1 v17.2d, v16.2d, v19.2d\n"
       "uzp2 v16.2d, v16.2d, v19.2d\n"
-      "str q8, [%x[Cpanel], #0x30]\n"
+      "str q10, [%x[Cpanel], #0x50]\n"
       "uzp1 v1.2d, v20.2d, v23.2d\n"
       "uzp2 v20.2d, v20.2d, v23.2d\n"
-      "str q9, [%x[Cpanel], #0x40]\n"
+      "str q0, [%x[Cpanel], #0x60]\n"
       "uzp1 v0.2d, v21.2d, v24.2d\n"
       "uzp2 v21.2d, v21.2d, v24.2d\n"
-      "str q10, [%x[Cpanel], #0x50]\n"
+      "str q2, [%x[Cpanel], #0x70]\n"
       "uzp1 v23.2d, v22.2d, v25.2d\n"
       "uzp2 v22.2d, v22.2d, v25.2d\n"
-      "str q3, [%x[Cpanel], #0x60]\n"
+      "str q17, [%x[Cpanel], #0x80]\n"
       "uzp1 v19.2d, v26.2d, v29.2d\n"
       "uzp2 v26.2d, v26.2d, v29.2d\n"
-      "str q2, [%x[Cpanel], #0x70]\n"
+      "str q14, [%x[Cpanel], #0x90]\n"
       "uzp1 v18.2d, v27.2d, v30.2d\n"
       "uzp2 v27.2d, v27.2d, v30.2d\n"
-      "str q17, [%x[Cpanel], #0x80]\n"
+      "str q15, [%x[Cpanel], #0xa0]\n"
       "uzp1 v17.2d, v28.2d, v31.2d\n"
       "uzp2 v28.2d, v28.2d, v31.2d\n"
-      "str q14, [%x[Cpanel], #0x90]\n"
-      "str q15, [%x[Cpanel], #0xa0]\n"
       "str q16, [%x[Cpanel], #0xb0]\n"
       "str q1, [%x[Cpanel], #0xc0]\n"
       "str q0, [%x[Cpanel], #0xd0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12.hpp
index 4707a17adb..19acfe8ae9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12.hpp
@@ -49,8 +49,7 @@ void a64_sgemm_asimd_8x12_x1(const float *, const float *, float *, int, int, in
 // structure.
 class cls_a64_sgemm_8x12 {
 public:
-    typedef float lhs_operand_type;
-    typedef float rhs_operand_type;
+    typedef float operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
@@ -69,7 +68,7 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixedTRB<lhs_operand_type, rhs_operand_type, result_type, 8, 12> transforms = {};
+    StdTransformsFixedTRB<operand_type, result_type, 8, 12> transforms = {};
 
     template<typename T>
     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp
index b9a2a3a3ef..00ec904e51 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp
@@ -54,312 +54,312 @@ void a64_sgemm_asimd_8x12_a53(const float *Apanel, const float *Bpanel, float *C
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.4s, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.4s, #0x0\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "movi    v11.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.4s, #0x0\n"
+                "movi	v8.4s, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.4s, #0x0\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "movi	v11.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v13.4s, #0x0\n"
+                "movi	v13.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v14.4s, #0x0\n"
+                "movi	v14.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.4s, #0x0\n"
+                "movi	v15.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v16.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v17.4s, #0x0\n"
+                "movi	v17.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v18.4s, #0x0\n"
+                "movi	v18.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v19.4s, #0x0\n"
+                "movi	v19.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.4s, #0x0\n"
+                "movi	v20.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "movi    v21.4s, #0x0\n"
+                "movi	v21.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "movi    v22.4s, #0x0\n"
-                "movi    v23.4s, #0x0\n"
-                "movi    v24.4s, #0x0\n"
-                "movi    v25.4s, #0x0\n"
-                "movi    v26.4s, #0x0\n"
-                "movi    v27.4s, #0x0\n"
-                "movi    v28.4s, #0x0\n"
-                "movi    v29.4s, #0x0\n"
-                "movi    v30.4s, #0x0\n"
-                "movi    v31.4s, #0x0\n"
+                "movi	v22.4s, #0x0\n"
+                "movi	v23.4s, #0x0\n"
+                "movi	v24.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 "1:\n"
                 // Unroll 0
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
                 "nop\n"
-                "fmla    v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "ldr    x20, [%[a_ptr], #40]\n"
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
-                "ins    %[a0a].d[1], x20\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "ldr    x20, [%[a_ptr], #56]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
-                "ins    %[a1a].d[1], x20\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[b_ptr], #56]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "ldr	x20, [%[a_ptr], #40]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
+                "ins	%[a0a].d[1], x20\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "ldr	x20, [%[a_ptr], #56]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
+                "ins	%[a1a].d[1], x20\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[b_ptr], #56]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
 
                 ASM_PREFETCH("[%[a_ptr], #320]")
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
 
                 ASM_PREFETCH("[%[b_ptr], #448]")
                 "nop\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
 
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
                 "nop\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "ldr    x20, [%[b_ptr], #72]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "ldr	x20, [%[b_ptr], #72]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
 
                 ASM_PREFETCH("[%[b_ptr], #512]")
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
 
                 // Unroll 1
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
                 "nop\n"
-                "fmla    v8.4s , %[b0].4s, %[a0a].s[0]\n"
-                "ldr    x20, [%[b_ptr], #88]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0a].s[1]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0a].s[2]\n"
-
-                "ldr    %d[a0], [%[a_ptr], #64]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v11.4s, %[b0].4s, %[a0a].s[3]\n"
-                "ldr    x20, [%[a_ptr], #72]\n"
-                "fmla    v12.4s, %[b0].4s, %[a1a].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1a].s[1]\n"
-
-                "ldr    %d[a1], [%[a_ptr], #80]\n"
-                "ins    %[a0].d[1], x20\n"
-                "fmla    v14.4s, %[b0].4s, %[a1a].s[2]\n"
-                "ldr    x20, [%[a_ptr], #88]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1a].s[3]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0a].s[0]\n"
-
-                "ldr    %d[b0], [%[b_ptr], #96]\n"
-                "ins    %[a1].d[1], x20\n"
-                "fmla    v17.4s, %[b1].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #104]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0a].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0a].s[3]\n"
+                "fmla	v8.4s , %[b0].4s, %[a0a].s[0]\n"
+                "ldr	x20, [%[b_ptr], #88]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0a].s[1]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0a].s[2]\n"
+
+                "ldr	%d[a0], [%[a_ptr], #64]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v11.4s, %[b0].4s, %[a0a].s[3]\n"
+                "ldr	x20, [%[a_ptr], #72]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1a].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1a].s[1]\n"
+
+                "ldr	%d[a1], [%[a_ptr], #80]\n"
+                "ins	%[a0].d[1], x20\n"
+                "fmla	v14.4s, %[b0].4s, %[a1a].s[2]\n"
+                "ldr	x20, [%[a_ptr], #88]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1a].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0a].s[0]\n"
+
+                "ldr	%d[b0], [%[b_ptr], #96]\n"
+                "ins	%[a1].d[1], x20\n"
+                "fmla	v17.4s, %[b1].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #104]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0a].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0a].s[3]\n"
 
                 "nop\n"
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v20.4s, %[b1].4s, %[a1a].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1a].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1a].s[2]\n"
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v20.4s, %[b1].4s, %[a1a].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1a].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1a].s[2]\n"
 
                 "nop\n"
                 "nop\n"
-                "fmla    v23.4s, %[b1].4s, %[a1a].s[3]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0a].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0a].s[1]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1a].s[3]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0a].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0a].s[1]\n"
 
-                "ldr    %d[b1], [%[b_ptr], #112]\n"
+                "ldr	%d[b1], [%[b_ptr], #112]\n"
                 "nop\n"
-                "fmla    v26.4s, %[b2].4s, %[a0a].s[2]\n"
-                "ldr    x20, [%[b_ptr], #120]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0a].s[3]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "fmla    v28.4s, %[b2].4s, %[a1a].s[0]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "fmla	v26.4s, %[b2].4s, %[a0a].s[2]\n"
+                "ldr	x20, [%[b_ptr], #120]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0a].s[3]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "fmla	v28.4s, %[b2].4s, %[a1a].s[0]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
 
                 "nop\n"
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v29.4s, %[b2].4s, %[a1a].s[1]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1a].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1a].s[3]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v29.4s, %[b2].4s, %[a1a].s[1]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1a].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1a].s[3]\n"
 
-                "bne    1b\n"
+                "bne	1b\n"
 
                 // Branch here if K=1 or 2.  Do the right thing for odd/even at the end.
                 "4:\n"
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration. (even K)
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
                 "nop\n"
-                "fmla    v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "ldr    x20, [%[a_ptr], #40]\n"
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
-                "ins    %[a0a].d[1], x20\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "ldr    x20, [%[a_ptr], #56]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
-                "ins    %[a1a].d[1], x20\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[b_ptr], #56]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "ldr	x20, [%[a_ptr], #40]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
+                "ins	%[a0a].d[1], x20\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "ldr	x20, [%[a_ptr], #56]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
+                "ins	%[a1a].d[1], x20\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[b_ptr], #56]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
 
                 "nop\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
 
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
                 "nop\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "ldr    x20, [%[b_ptr], #72]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "ldr	x20, [%[b_ptr], #72]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
 
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
 
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
                 "nop\n"
-                "fmla    v8.4s , %[b0].4s, %[a0a].s[0]\n"
-                "ldr    x20, [%[b_ptr], #88]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0a].s[1]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0a].s[2]\n"
-
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v11.4s, %[b0].4s, %[a0a].s[3]\n"
-                "fmla    v12.4s, %[b0].4s, %[a1a].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1a].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1a].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1a].s[3]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0a].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0a].s[1]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0a].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0a].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1a].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1a].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1a].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1a].s[3]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0a].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0a].s[1]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0a].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0a].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1a].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1a].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "fmla    v30.4s, %[b2].4s, %[a1a].s[2]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v31.4s, %[b2].4s, %[a1a].s[3]\n"
-                "b    3f\n"
+                "fmla	v8.4s , %[b0].4s, %[a0a].s[0]\n"
+                "ldr	x20, [%[b_ptr], #88]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0a].s[1]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0a].s[2]\n"
+
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v11.4s, %[b0].4s, %[a0a].s[3]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1a].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1a].s[1]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1a].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1a].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0a].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0a].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0a].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0a].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1a].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1a].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1a].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1a].s[3]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0a].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0a].s[1]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0a].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0a].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1a].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1a].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "fmla	v30.4s, %[b2].4s, %[a1a].s[2]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v31.4s, %[b2].4s, %[a1a].s[3]\n"
+                "b	3f\n"
 
                 // Detached final iteration. (odd K)
                 "2:\n"
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
                 "nop\n"
-                "fmla    v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "fmla	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
 
                 // Common tail
                 "3:\n"
-                "str    q8,  [%[c_ptr]]\n"
-                "str    q16,  [%[c_ptr], #16]\n"
-                "str    q24,  [%[c_ptr], #32]\n"
-                "str    q9,  [%[c_ptr], #48]\n"
-                "str    q17,  [%[c_ptr], #64]\n"
-                "str    q25,  [%[c_ptr], #80]\n"
-                "str    q10,  [%[c_ptr], #96]\n"
-                "str    q18,  [%[c_ptr], #112]\n"
-                "str    q26,  [%[c_ptr], #128]\n"
-                "str    q11,  [%[c_ptr], #144]\n"
-                "str    q19,  [%[c_ptr], #160]\n"
-                "str    q27,  [%[c_ptr], #176]\n"
-                "str    q12,  [%[c_ptr], #192]\n"
-                "str    q20,  [%[c_ptr], #208]\n"
-                "str    q28,  [%[c_ptr], #224]\n"
-                "str    q13,  [%[c_ptr], #240]\n"
-                "str    q21,  [%[c_ptr], #256]\n"
-                "str    q29,  [%[c_ptr], #272]\n"
-                "str    q14,  [%[c_ptr], #288]\n"
-                "str    q22,  [%[c_ptr], #304]\n"
-                "str    q30,  [%[c_ptr], #320]\n"
-                "str    q15,  [%[c_ptr], #336]\n"
-                "str    q23,  [%[c_ptr], #352]\n"
-                "str    q31,  [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q8,  [%[c_ptr]]\n"
+                "str	q16,  [%[c_ptr], #16]\n"
+                "str	q24,  [%[c_ptr], #32]\n"
+                "str	q9,  [%[c_ptr], #48]\n"
+                "str	q17,  [%[c_ptr], #64]\n"
+                "str	q25,  [%[c_ptr], #80]\n"
+                "str	q10,  [%[c_ptr], #96]\n"
+                "str	q18,  [%[c_ptr], #112]\n"
+                "str	q26,  [%[c_ptr], #128]\n"
+                "str	q11,  [%[c_ptr], #144]\n"
+                "str	q19,  [%[c_ptr], #160]\n"
+                "str	q27,  [%[c_ptr], #176]\n"
+                "str	q12,  [%[c_ptr], #192]\n"
+                "str	q20,  [%[c_ptr], #208]\n"
+                "str	q28,  [%[c_ptr], #224]\n"
+                "str	q13,  [%[c_ptr], #240]\n"
+                "str	q21,  [%[c_ptr], #256]\n"
+                "str	q29,  [%[c_ptr], #272]\n"
+                "str	q14,  [%[c_ptr], #288]\n"
+                "str	q22,  [%[c_ptr], #304]\n"
+                "str	q30,  [%[c_ptr], #320]\n"
+                "str	q15,  [%[c_ptr], #336]\n"
+                "str	q23,  [%[c_ptr], #352]\n"
+                "str	q31,  [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "+w" (a0), [a1] "+w" (a1), [a0a] "+w" (a0a), [a1a] "+w" (a1a),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp
index 4303d1346c..de85605561 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp
@@ -54,313 +54,313 @@ void a64_sgemm_asimd_8x12_a55(const float *Apanel, const float *Bpanel, float *C
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.4s, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.4s, #0x0\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "movi    v11.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.4s, #0x0\n"
+                "movi	v8.4s, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.4s, #0x0\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "movi	v11.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v13.4s, #0x0\n"
+                "movi	v13.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v14.4s, #0x0\n"
+                "movi	v14.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.4s, #0x0\n"
+                "movi	v15.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v16.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v17.4s, #0x0\n"
+                "movi	v17.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v18.4s, #0x0\n"
+                "movi	v18.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v19.4s, #0x0\n"
+                "movi	v19.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.4s, #0x0\n"
+                "movi	v20.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "movi    v21.4s, #0x0\n"
+                "movi	v21.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "movi    v22.4s, #0x0\n"
-                "movi    v23.4s, #0x0\n"
-                "movi    v24.4s, #0x0\n"
-                "movi    v25.4s, #0x0\n"
-                "movi    v26.4s, #0x0\n"
-                "movi    v27.4s, #0x0\n"
-                "movi    v28.4s, #0x0\n"
-                "movi    v29.4s, #0x0\n"
-                "movi    v30.4s, #0x0\n"
-                "movi    v31.4s, #0x0\n"
+                "movi	v22.4s, #0x0\n"
+                "movi	v23.4s, #0x0\n"
+                "movi	v24.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 "1:\n"
                 // Unroll 0
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
-                "fmla    v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
+                "fmla	v8.4s , %[b0].4s, %[a0].s[0]\n"
 
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "subs    %w[k], %w[k], #1\n"
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "subs	%w[k], %w[k], #1\n"
 
 
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
-                "ins    %[b2].d[1], x20\n"
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
+                "ins	%[b2].d[1], x20\n"
 
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[a_ptr], #40]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[a_ptr], #40]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
 
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
-                "ins    %[a0a].d[1], x20\n"
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
+                "ins	%[a0a].d[1], x20\n"
 
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[a_ptr], #56]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[a_ptr], #56]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
 
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
-                "ins    %[a1a].d[1], x20\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
+                "ins	%[a1a].d[1], x20\n"
                 ASM_PREFETCH("[%[a_ptr], #320]")
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[b_ptr], #56]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
-                "ins    %[b0].d[1], x20\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[b_ptr], #72]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[b_ptr], #56]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
+                "ins	%[b0].d[1], x20\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[b_ptr], #72]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
                 ASM_PREFETCH("[%[b_ptr], #448]")
 
 
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #512]")
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
 
                 // Unroll 1
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
-                "ins    %[b1].d[1], x20\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
+                "ins	%[b1].d[1], x20\n"
 
-                "fmla    v8.4s , %[b0].4s, %[a0a].s[0]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #88]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0a].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0a].s[3]\n"
+                "fmla	v8.4s , %[b0].4s, %[a0a].s[0]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #88]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0a].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0a].s[3]\n"
 
-                "ldr    %d[a0], [%[a_ptr], #64]\n"
-                "ins    %[b2].d[1], x20\n"
+                "ldr	%d[a0], [%[a_ptr], #64]\n"
+                "ins	%[b2].d[1], x20\n"
 
-                "fmla    v12.4s, %[b0].4s, %[a1a].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1a].s[1]\n"
-                "ldr    x20, [%[a_ptr], #72]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1a].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1a].s[3]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1a].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1a].s[1]\n"
+                "ldr	x20, [%[a_ptr], #72]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1a].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1a].s[3]\n"
 
-                "ldr    %d[a1], [%[a_ptr], #80]\n"
-                "ins    %[a0].d[1], x20\n"
+                "ldr	%d[a1], [%[a_ptr], #80]\n"
+                "ins	%[a0].d[1], x20\n"
 
-                "fmla    v16.4s, %[b1].4s, %[a0a].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[a_ptr], #88]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0a].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0a].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0a].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[a_ptr], #88]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0a].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0a].s[3]\n"
 
 
-                "ldr    %d[b0], [%[b_ptr], #96]\n"
-                "ins    %[a1].d[1], x20\n"
+                "ldr	%d[b0], [%[b_ptr], #96]\n"
+                "ins	%[a1].d[1], x20\n"
 
-                "fmla    v20.4s, %[b1].4s, %[a1a].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #104]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1a].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1a].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1a].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #104]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1a].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1a].s[3]\n"
 
-                "ldr    %d[b1], [%[b_ptr], #112]\n"
-                "ins    %[b0].d[1], x20\n"
+                "ldr	%d[b1], [%[b_ptr], #112]\n"
+                "ins	%[b0].d[1], x20\n"
 
-                "fmla    v24.4s, %[b2].4s, %[a0a].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #120]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0a].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0a].s[3]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
+                "fmla	v24.4s, %[b2].4s, %[a0a].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #120]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0a].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0a].s[3]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
 
-                "fmla    v28.4s, %[b2].4s, %[a1a].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1a].s[1]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1a].s[2]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v31.4s, %[b2].4s, %[a1a].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1a].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1a].s[1]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1a].s[2]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v31.4s, %[b2].4s, %[a1a].s[3]\n"
 
 
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
-                "ins    %[b1].d[1], x20\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
+                "ins	%[b1].d[1], x20\n"
 
 
-                "bne    1b\n"
+                "bne	1b\n"
 
                 // Branch here if K=1 or 2.  Do the right thing for odd/even at the end.
                 "4:\n"
-                "cbnz    %w[oddk], 2f\n"
-                "fmla    v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "cbnz	%w[oddk], 2f\n"
+                "fmla	v8.4s , %[b0].4s, %[a0].s[0]\n"
 
                 // Detached final iteration. (even K)
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
-                "ins    %[b2].d[1], x20\n"
-
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[a_ptr], #40]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
-                "ins    %[a0a].d[1], x20\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[a_ptr], #56]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
-                "ins    %[a1a].d[1], x20\n"
-
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[b_ptr], #56]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
-                "ins    %[b0].d[1], x20\n"
-
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[b_ptr], #72]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
-                "ins    %[b1].d[1], x20\n"
-
-                "fmla    v8.4s , %[b0].4s, %[a0a].s[0]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #88]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0a].s[2]\n"
-
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v11.4s, %[b0].4s, %[a0a].s[3]\n"
-                "fmla    v12.4s, %[b0].4s, %[a1a].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1a].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1a].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1a].s[3]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0a].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0a].s[1]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0a].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0a].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1a].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1a].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1a].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1a].s[3]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0a].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0a].s[1]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0a].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0a].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1a].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1a].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "fmla    v30.4s, %[b2].4s, %[a1a].s[2]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v31.4s, %[b2].4s, %[a1a].s[3]\n"
-                "b    3f\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
+                "ins	%[b2].d[1], x20\n"
+
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[a_ptr], #40]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
+                "ins	%[a0a].d[1], x20\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[a_ptr], #56]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
+                "ins	%[a1a].d[1], x20\n"
+
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[b_ptr], #56]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
+                "ins	%[b0].d[1], x20\n"
+
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[b_ptr], #72]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
+                "ins	%[b1].d[1], x20\n"
+
+                "fmla	v8.4s , %[b0].4s, %[a0a].s[0]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #88]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0a].s[2]\n"
+
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v11.4s, %[b0].4s, %[a0a].s[3]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1a].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1a].s[1]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1a].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1a].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0a].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0a].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0a].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0a].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1a].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1a].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1a].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1a].s[3]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0a].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0a].s[1]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0a].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0a].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1a].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1a].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "fmla	v30.4s, %[b2].4s, %[a1a].s[2]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v31.4s, %[b2].4s, %[a1a].s[3]\n"
+                "b	3f\n"
 
                 // Detached final iteration. (odd K)
                 "2:\n"
 
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
-                "fmla    v8.4s , %[b0].4s, %[a0].s[0]\n"
-
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
+                "fmla	v8.4s , %[b0].4s, %[a0].s[0]\n"
+
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
 
                 // Common tail
                 "3:\n"
-                "str    q8,  [%[c_ptr]]\n"
-                "str    q16,  [%[c_ptr], #16]\n"
-                "str    q24,  [%[c_ptr], #32]\n"
-                "str    q9,  [%[c_ptr], #48]\n"
-                "str    q17,  [%[c_ptr], #64]\n"
-                "str    q25,  [%[c_ptr], #80]\n"
-                "str    q10,  [%[c_ptr], #96]\n"
-                "str    q18,  [%[c_ptr], #112]\n"
-                "str    q26,  [%[c_ptr], #128]\n"
-                "str    q11,  [%[c_ptr], #144]\n"
-                "str    q19,  [%[c_ptr], #160]\n"
-                "str    q27,  [%[c_ptr], #176]\n"
-                "str    q12,  [%[c_ptr], #192]\n"
-                "str    q20,  [%[c_ptr], #208]\n"
-                "str    q28,  [%[c_ptr], #224]\n"
-                "str    q13,  [%[c_ptr], #240]\n"
-                "str    q21,  [%[c_ptr], #256]\n"
-                "str    q29,  [%[c_ptr], #272]\n"
-                "str    q14,  [%[c_ptr], #288]\n"
-                "str    q22,  [%[c_ptr], #304]\n"
-                "str    q30,  [%[c_ptr], #320]\n"
-                "str    q15,  [%[c_ptr], #336]\n"
-                "str    q23,  [%[c_ptr], #352]\n"
-                "str    q31,  [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q8,  [%[c_ptr]]\n"
+                "str	q16,  [%[c_ptr], #16]\n"
+                "str	q24,  [%[c_ptr], #32]\n"
+                "str	q9,  [%[c_ptr], #48]\n"
+                "str	q17,  [%[c_ptr], #64]\n"
+                "str	q25,  [%[c_ptr], #80]\n"
+                "str	q10,  [%[c_ptr], #96]\n"
+                "str	q18,  [%[c_ptr], #112]\n"
+                "str	q26,  [%[c_ptr], #128]\n"
+                "str	q11,  [%[c_ptr], #144]\n"
+                "str	q19,  [%[c_ptr], #160]\n"
+                "str	q27,  [%[c_ptr], #176]\n"
+                "str	q12,  [%[c_ptr], #192]\n"
+                "str	q20,  [%[c_ptr], #208]\n"
+                "str	q28,  [%[c_ptr], #224]\n"
+                "str	q13,  [%[c_ptr], #240]\n"
+                "str	q21,  [%[c_ptr], #256]\n"
+                "str	q29,  [%[c_ptr], #272]\n"
+                "str	q14,  [%[c_ptr], #288]\n"
+                "str	q22,  [%[c_ptr], #304]\n"
+                "str	q30,  [%[c_ptr], #320]\n"
+                "str	q15,  [%[c_ptr], #336]\n"
+                "str	q23,  [%[c_ptr], #352]\n"
+                "str	q31,  [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "+w" (a0), [a1] "+w" (a1), [a0a] "+w" (a0a), [a1a] "+w" (a1a),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp
index fdbbe6b749..928b22a190 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp
@@ -100,275 +100,275 @@ void a64_sgemm_asimd_8x12_a55r1(const float *Apanel, const float *Bpanel, float
 
                 // The loop is offset by these two instructions which must
                 // always be executed.
-                "fmla    v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
+                "fmla	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 "1:\n"
                 // Unroll 0
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
-
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[a_ptr], #40]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "ins    %[a0a].d[1], x20\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[a_ptr], #56]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "ins    %[a1a].d[1], x20\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[b_ptr], #56]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[b_ptr], #72]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
+
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[a_ptr], #40]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "ins	%[a0a].d[1], x20\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[a_ptr], #56]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "ins	%[a1a].d[1], x20\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[b_ptr], #56]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[b_ptr], #72]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
                 ASM_PREFETCH("[%[a_ptr], #448]")
 
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #576]")
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
 
                 // Unroll 1
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
-
-                "fmla    v8.4s , %[b0].4s, %[a0a].s[0]\n"
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v9.4s , %[b0].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #88]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0a].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0a].s[3]\n"
-                "ldr    %d[a0], [%[a_ptr], #64]\n"
-
-                "fmla    v12.4s, %[b0].4s, %[a1a].s[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.4s, %[b0].4s, %[a1a].s[1]\n"
-                "ldr    x20, [%[a_ptr], #72]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1a].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1a].s[3]\n"
-                "ldr    %d[a1], [%[a_ptr], #80]\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0a].s[0]\n"
-                "ins    %[a0].d[1], x20\n"
-                "fmla    v17.4s, %[b1].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[a_ptr], #88]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0a].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0a].s[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #96]\n"
-
-                "fmla    v20.4s, %[b1].4s, %[a1a].s[0]\n"
-                "ins    %[a1].d[1], x20\n"
-                "fmla    v21.4s, %[b1].4s, %[a1a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #104]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1a].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1a].s[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #112]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0a].s[0]\n"
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v25.4s, %[b2].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #120]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0a].s[2]\n"
-
-                "fmla    v27.4s, %[b2].4s, %[a0a].s[3]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-
-                "fmla    v28.4s, %[b2].4s, %[a1a].s[0]\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
+
+                "fmla	v8.4s , %[b0].4s, %[a0a].s[0]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v9.4s , %[b0].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #88]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0a].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0a].s[3]\n"
+                "ldr	%d[a0], [%[a_ptr], #64]\n"
+
+                "fmla	v12.4s, %[b0].4s, %[a1a].s[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.4s, %[b0].4s, %[a1a].s[1]\n"
+                "ldr	x20, [%[a_ptr], #72]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1a].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1a].s[3]\n"
+                "ldr	%d[a1], [%[a_ptr], #80]\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0a].s[0]\n"
+                "ins	%[a0].d[1], x20\n"
+                "fmla	v17.4s, %[b1].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[a_ptr], #88]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0a].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0a].s[3]\n"
+                "ldr	%d[b0], [%[b_ptr], #96]\n"
+
+                "fmla	v20.4s, %[b1].4s, %[a1a].s[0]\n"
+                "ins	%[a1].d[1], x20\n"
+                "fmla	v21.4s, %[b1].4s, %[a1a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #104]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1a].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1a].s[3]\n"
+                "ldr	%d[b1], [%[b_ptr], #112]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0a].s[0]\n"
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v25.4s, %[b2].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #120]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0a].s[2]\n"
+
+                "fmla	v27.4s, %[b2].4s, %[a0a].s[3]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+
+                "fmla	v28.4s, %[b2].4s, %[a1a].s[0]\n"
                 ASM_PREFETCH("[%[b_ptr], #640]")
-                "fmla    v29.4s, %[b2].4s, %[a1a].s[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v30.4s, %[b2].4s, %[a1a].s[2]\n"
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v31.4s, %[b2].4s, %[a1a].s[3]\n"
-                "ldr    %d[b2], [%[b_ptr], #32]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1a].s[1]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v30.4s, %[b2].4s, %[a1a].s[2]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v31.4s, %[b2].4s, %[a1a].s[3]\n"
+                "ldr	%d[b2], [%[b_ptr], #32]\n"
 
-                "fmla    v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "b.ne    1b\n"
+                "fmla	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "b.ne	1b\n"
 
                 // Branch here if K=1 or 2.  Do the right thing for odd/even at the end.
                 "4:\n"
 
-        // Start final iteration - branch off to "odd" code before we load a0a.
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[b_ptr], #40]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "cbnz    %w[oddk], 2f\n"
+		// Start final iteration - branch off to "odd" code before we load a0a.
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[b_ptr], #40]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Even K continuation
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "ldr    %d[a0a], [%[a_ptr], #32]\n"
-
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[a_ptr], #40]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "ldr	%d[a0a], [%[a_ptr], #32]\n"
+
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[a_ptr], #40]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
                 ASM_PREFETCHW("[%[c_ptr]]")
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "ldr    %d[a1a], [%[a_ptr], #48]\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "ins    %[a0a].d[1], x20\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "ldr    x20, [%[a_ptr], #56]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "ldr    %d[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "ins    %[a1a].d[1], x20\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[b_ptr], #56]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "ldr	%d[a1a], [%[a_ptr], #48]\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "ins	%[a0a].d[1], x20\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "ldr	x20, [%[a_ptr], #56]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "ldr	%d[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "ins	%[a1a].d[1], x20\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[b_ptr], #56]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
                 ASM_PREFETCHW("[%[c_ptr], #64]")
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
 
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #128]")
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "ldr    %d[b1], [%[b_ptr], #64]\n"
-
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "ins    %[b0].d[1], x20\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "ldr    x20, [%[b_ptr], #72]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "ldr	%d[b1], [%[b_ptr], #64]\n"
+
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "ins	%[b0].d[1], x20\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "ldr	x20, [%[b_ptr], #72]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
                 ASM_PREFETCHW("[%[c_ptr], #192]")
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "ldr    %d[b2], [%[b_ptr], #80]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ldr	%d[b2], [%[b_ptr], #80]\n"
 
-                "fmla    v8.4s , %[b0].4s, %[a0a].s[0]\n"
-                "ins    %[b1].d[1], x20\n"
-                "fmla    v9.4s , %[b0].4s, %[a0a].s[1]\n"
-                "ldr    x20, [%[b_ptr], #88]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0a].s[2]\n"
-                "ins    %[b2].d[1], x20\n"
+                "fmla	v8.4s , %[b0].4s, %[a0a].s[0]\n"
+                "ins	%[b1].d[1], x20\n"
+                "fmla	v9.4s , %[b0].4s, %[a0a].s[1]\n"
+                "ldr	x20, [%[b_ptr], #88]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0a].s[2]\n"
+                "ins	%[b2].d[1], x20\n"
 
-                "fmla    v11.4s, %[b0].4s, %[a0a].s[3]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0a].s[3]\n"
                 ASM_PREFETCHW("[%[c_ptr], #256]")
-                "fmla    v12.4s, %[b0].4s, %[a1a].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1a].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1a].s[2]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1a].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1a].s[1]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1a].s[2]\n"
                 ASM_PREFETCHW("[%[c_ptr], #320]")
-                "fmla    v15.4s, %[b0].4s, %[a1a].s[3]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0a].s[0]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1a].s[3]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0a].s[0]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #384]")
-                "fmla    v17.4s, %[b1].4s, %[a0a].s[1]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0a].s[2]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0a].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0a].s[2]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #448]")
-                "fmla    v19.4s, %[b1].4s, %[a0a].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1a].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1a].s[1]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0a].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1a].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1a].s[1]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #512]")
-                "fmla    v22.4s, %[b1].4s, %[a1a].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1a].s[3]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1a].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1a].s[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #576]")
-                "fmla    v24.4s, %[b2].4s, %[a0a].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0a].s[1]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0a].s[2]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0a].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0a].s[1]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0a].s[2]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #640]")
-                "fmla    v27.4s, %[b2].4s, %[a0a].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1a].s[0]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0a].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1a].s[0]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #704]")
-                "fmla    v29.4s, %[b2].4s, %[a1a].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "fmla    v30.4s, %[b2].4s, %[a1a].s[2]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v31.4s, %[b2].4s, %[a1a].s[3]\n"
-                "b    3f\n"
+                "fmla	v29.4s, %[b2].4s, %[a1a].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "fmla	v30.4s, %[b2].4s, %[a1a].s[2]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v31.4s, %[b2].4s, %[a1a].s[3]\n"
+                "b	3f\n"
 
                 // Odd K continuation
                 "2:\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
                 ASM_PREFETCHW("[%[c_ptr]]")
-                "fmla    v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "ins    %[b2].d[1], x20\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "fmla	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "ins	%[b2].d[1], x20\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #64]")
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
                 ASM_PREFETCHW("[%[c_ptr], #128]")
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #192]")
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
                 ASM_PREFETCHW("[%[c_ptr], #256]")
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
                 ASM_PREFETCHW("[%[c_ptr], #320]")
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #384]")
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #448]")
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #512]")
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #576]")
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #640]")
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
                 ASM_PREFETCHWL2("[%[c_ptr], #704]")
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
 
                 // Common tail
                 "3:\n"
-                "str    q8,   [%[c_ptr]]\n"
-                "str    q16,  [%[c_ptr], #16]\n"
-                "str    q24,  [%[c_ptr], #32]\n"
-                "str    q9,   [%[c_ptr], #48]\n"
-                "str    q17,  [%[c_ptr], #64]\n"
-                "str    q25,  [%[c_ptr], #80]\n"
-                "str    q10,  [%[c_ptr], #96]\n"
-                "str    q18,  [%[c_ptr], #112]\n"
-                "str    q26,  [%[c_ptr], #128]\n"
-                "str    q11,  [%[c_ptr], #144]\n"
-                "str    q19,  [%[c_ptr], #160]\n"
-                "str    q27,  [%[c_ptr], #176]\n"
-                "str    q12,  [%[c_ptr], #192]\n"
-                "str    q20,  [%[c_ptr], #208]\n"
-                "str    q28,  [%[c_ptr], #224]\n"
-                "str    q13,  [%[c_ptr], #240]\n"
-                "str    q21,  [%[c_ptr], #256]\n"
-                "str    q29,  [%[c_ptr], #272]\n"
-                "str    q14,  [%[c_ptr], #288]\n"
-                "str    q22,  [%[c_ptr], #304]\n"
-                "str    q30,  [%[c_ptr], #320]\n"
-                "str    q15,  [%[c_ptr], #336]\n"
-                "str    q23,  [%[c_ptr], #352]\n"
-                "str    q31,  [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q8,   [%[c_ptr]]\n"
+                "str	q16,  [%[c_ptr], #16]\n"
+                "str	q24,  [%[c_ptr], #32]\n"
+                "str	q9,   [%[c_ptr], #48]\n"
+                "str	q17,  [%[c_ptr], #64]\n"
+                "str	q25,  [%[c_ptr], #80]\n"
+                "str	q10,  [%[c_ptr], #96]\n"
+                "str	q18,  [%[c_ptr], #112]\n"
+                "str	q26,  [%[c_ptr], #128]\n"
+                "str	q11,  [%[c_ptr], #144]\n"
+                "str	q19,  [%[c_ptr], #160]\n"
+                "str	q27,  [%[c_ptr], #176]\n"
+                "str	q12,  [%[c_ptr], #192]\n"
+                "str	q20,  [%[c_ptr], #208]\n"
+                "str	q28,  [%[c_ptr], #224]\n"
+                "str	q13,  [%[c_ptr], #240]\n"
+                "str	q21,  [%[c_ptr], #256]\n"
+                "str	q29,  [%[c_ptr], #272]\n"
+                "str	q14,  [%[c_ptr], #288]\n"
+                "str	q22,  [%[c_ptr], #304]\n"
+                "str	q30,  [%[c_ptr], #320]\n"
+                "str	q15,  [%[c_ptr], #336]\n"
+                "str	q23,  [%[c_ptr], #352]\n"
+                "str	q31,  [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "+w" (a0), [a1] "+w" (a1), [a0a] "+w" (a0a), [a1a] "+w" (a1a),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp
index 5e1cce3233..711fc77d9f 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp
@@ -64,281 +64,281 @@ void a64_sgemm_asimd_8x12(const float *Apanel, const float *Bpanel, float *Cpane
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.4s, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.4s, #0x0\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "movi    v11.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.4s, #0x0\n"
+                "movi	v8.4s, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.4s, #0x0\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "movi	v11.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v13.4s, #0x0\n"
+                "movi	v13.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v14.4s, #0x0\n"
+                "movi	v14.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.4s, #0x0\n"
+                "movi	v15.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v16.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v17.4s, #0x0\n"
+                "movi	v17.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v18.4s, #0x0\n"
+                "movi	v18.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v19.4s, #0x0\n"
+                "movi	v19.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.4s, #0x0\n"
+                "movi	v20.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "movi    v21.4s, #0x0\n"
+                "movi	v21.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "movi    v22.4s, #0x0\n"
-                "movi    v23.4s, #0x0\n"
-                "movi    v24.4s, #0x0\n"
-                "movi    v25.4s, #0x0\n"
-                "movi    v26.4s, #0x0\n"
-                "movi    v27.4s, #0x0\n"
-                "movi    v28.4s, #0x0\n"
-                "movi    v29.4s, #0x0\n"
-                "movi    v30.4s, #0x0\n"
-                "movi    v31.4s, #0x0\n"
+                "movi	v22.4s, #0x0\n"
+                "movi	v23.4s, #0x0\n"
+                "movi	v24.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 // Loop proper
                 "1:\n"
-                "fmla     v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "fmla      v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "ldr    %q[a0a], [%[a_ptr], #32]\n"
-                "fmla     v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "ldr    %q[a1a], [%[a_ptr], #48]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "fmla 	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "fmla  	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "ldr	%q[a0a], [%[a_ptr], #32]\n"
+                "fmla 	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "ldr	%q[a1a], [%[a_ptr], #48]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #320]")
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #448]")
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
-
-                "fmla     v8.4s , %[b0].4s, %[a0a].s[0]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0a].s[1]\n"
-                "ldr    %q[a0], [%[a_ptr], #64]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0a].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0a].s[3]\n"
-                "fmla     v12.4s, %[b0].4s, %[a1a].s[0]\n"
-                "ldr    %q[a1], [%[a_ptr], #80]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
+
+                "fmla 	v8.4s , %[b0].4s, %[a0a].s[0]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0a].s[1]\n"
+                "ldr	%q[a0], [%[a_ptr], #64]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0a].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0a].s[3]\n"
+                "fmla 	v12.4s, %[b0].4s, %[a1a].s[0]\n"
+                "ldr	%q[a1], [%[a_ptr], #80]\n"
                 "fmla   v13.4s, %[b0].4s, %[a1a].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1a].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1a].s[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #96]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1a].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1a].s[3]\n"
+                "ldr	%q[b0], [%[b_ptr], #96]\n"
 
-                "fmla    v16.4s, %[b1].4s, %[a0a].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0a].s[1]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0a].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0a].s[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #512]")
-                "fmla    v18.4s, %[b1].4s, %[a0a].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0a].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1a].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1a].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1a].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1a].s[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #112]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0a].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0a].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "fmla    v26.4s, %[b2].4s, %[a0a].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0a].s[3]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v28.4s, %[b2].4s, %[a1a].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1a].s[1]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v30.4s, %[b2].4s, %[a1a].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1a].s[3]\n"
-                "bne    1b\n"
+                "fmla	v18.4s, %[b1].4s, %[a0a].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0a].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1a].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1a].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1a].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1a].s[3]\n"
+                "ldr	%q[b1], [%[b_ptr], #112]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0a].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0a].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "fmla	v26.4s, %[b2].4s, %[a0a].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0a].s[3]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v28.4s, %[b2].4s, %[a1a].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1a].s[1]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v30.4s, %[b2].4s, %[a1a].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1a].s[3]\n"
+                "bne	1b\n"
 
                 // Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
                 "4:\n"
 
                 // Branch to alternative tail for odd K
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration (even K)
-                "fmla     v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "fmla 	v8.4s , %[b0].4s, %[a0].s[0]\n"
                 "fmla   v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "ldr    %q[a0a], [%[a_ptr], #32]\n"
-                "fmla     v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "ldr	%q[a0a], [%[a_ptr], #32]\n"
+                "fmla 	v12.4s, %[b0].4s, %[a1].s[0]\n"
                 "fmla   v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "ldr    %q[a1a], [%[a_ptr], #48]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
-
-                "fmla     v8.4s , %[b0].4s, %[a0a].s[0]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0a].s[0]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
+                "ldr	%q[a1a], [%[a_ptr], #48]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
+
+                "fmla 	v8.4s , %[b0].4s, %[a0a].s[0]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0a].s[0]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
                 "fmla   v9.4s , %[b0].4s, %[a0a].s[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0a].s[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0a].s[0]\n"
-                "str    q24, [%[c_ptr], #32]\n"
-
-                "fmla    v25.4s, %[b2].4s, %[a0a].s[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0a].s[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0a].s[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0a].s[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-
-                "fmla    v11.4s, %[b0].4s, %[a0a].s[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0a].s[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0a].s[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-
-                "fmla     v12.4s, %[b0].4s, %[a1a].s[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1a].s[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1a].s[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q8, [%[c_ptr], #0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0a].s[1]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0a].s[0]\n"
+                "str	q24, [%[c_ptr], #32]\n"
+
+                "fmla	v25.4s, %[b2].4s, %[a0a].s[1]\n"
+                "str	q9, [%[c_ptr], #48]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0a].s[2]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0a].s[2]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0a].s[2]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+
+                "fmla	v11.4s, %[b0].4s, %[a0a].s[3]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0a].s[3]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0a].s[3]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+
+                "fmla 	v12.4s, %[b0].4s, %[a1a].s[0]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1a].s[0]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1a].s[0]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 "fmla   v13.4s, %[b0].4s, %[a1a].s[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1a].s[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1a].s[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-
-                "fmla    v14.4s, %[b0].4s, %[a1a].s[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1a].s[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1a].s[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-
-                "fmla    v15.4s, %[b0].4s, %[a1a].s[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1a].s[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1a].s[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
-
-                "b    3f\n"
+                "str	q20, [%[c_ptr], #208]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1a].s[1]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1a].s[1]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+
+                "fmla	v14.4s, %[b0].4s, %[a1a].s[2]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1a].s[2]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1a].s[2]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+
+                "fmla	v15.4s, %[b0].4s, %[a1a].s[3]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1a].s[3]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1a].s[3]\n"
+                "str	q15, [%[c_ptr], #336]\n"
+
+                "b	3f\n"
 
                 // Detached final iteration (odd K)
                 "2:\n"
-                "fmla     v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla 	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
                 "fmla   v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "str    q24, [%[c_ptr], #32]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
-
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-
-                "fmla     v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q8, [%[c_ptr], #0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "str	q24, [%[c_ptr], #32]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "str	q9, [%[c_ptr], #48]\n"
+
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+
+                "fmla 	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 "fmla   v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q20, [%[c_ptr], #208]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
                 // Common tail
                 "3:\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "+w" (a0), [a1] "+w" (a1), [a0a] "+w" (a0a), [a1a] "+w" (a1a),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp
index 1567b05d3e..a348b4b67e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp
@@ -62,281 +62,281 @@ void a64_sgemm_asimd_8x12_x1(const float *Apanel, const float *Bpanel, float *Cp
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.4s, #0x0\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "movi    v9.4s, #0x0\n"
-                "ldr    %q[b0], [%[b_ptr]]\n"
-                "movi    v10.4s, #0x0\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "movi    v11.4s, #0x0\n"
-                "ldr    %q[b1], [%[b_ptr], #16]\n"
-                "movi    v12.4s, #0x0\n"
+                "movi	v8.4s, #0x0\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "movi	v9.4s, #0x0\n"
+                "ldr	%q[b0], [%[b_ptr]]\n"
+                "movi	v10.4s, #0x0\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "movi	v11.4s, #0x0\n"
+                "ldr	%q[b1], [%[b_ptr], #16]\n"
+                "movi	v12.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
-                "movi    v13.4s, #0x0\n"
+                "movi	v13.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #64]")
-                "movi    v14.4s, #0x0\n"
+                "movi	v14.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.4s, #0x0\n"
+                "movi	v15.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #128]")
-                "movi    v16.4s, #0x0\n"
+                "movi	v16.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #192]")
-                "movi    v17.4s, #0x0\n"
+                "movi	v17.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #256]")
-                "movi    v18.4s, #0x0\n"
+                "movi	v18.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #192]")
-                "movi    v19.4s, #0x0\n"
+                "movi	v19.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #320]")
-                "movi    v20.4s, #0x0\n"
+                "movi	v20.4s, #0x0\n"
                 ASM_PREFETCH("[%[a_ptr], #256]")
-                "movi    v21.4s, #0x0\n"
+                "movi	v21.4s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #384]")
-                "movi    v22.4s, #0x0\n"
-                "movi    v23.4s, #0x0\n"
-                "movi    v24.4s, #0x0\n"
-                "movi    v25.4s, #0x0\n"
-                "movi    v26.4s, #0x0\n"
-                "movi    v27.4s, #0x0\n"
-                "movi    v28.4s, #0x0\n"
-                "movi    v29.4s, #0x0\n"
-                "movi    v30.4s, #0x0\n"
-                "movi    v31.4s, #0x0\n"
+                "movi	v22.4s, #0x0\n"
+                "movi	v23.4s, #0x0\n"
+                "movi	v24.4s, #0x0\n"
+                "movi	v25.4s, #0x0\n"
+                "movi	v26.4s, #0x0\n"
+                "movi	v27.4s, #0x0\n"
+                "movi	v28.4s, #0x0\n"
+                "movi	v29.4s, #0x0\n"
+                "movi	v30.4s, #0x0\n"
+                "movi	v31.4s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 // Loop proper
                 "1:\n"
-                "fmla     v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "fmla      v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "fmla     v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "fmla    v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "fmla 	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "fmla  	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "fmla 	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v13.4s, %[b0].4s, %[a1].s[1]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
                 ASM_PREFETCH("[%[a_ptr], #320]")
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
                 ASM_PREFETCH("[%[b_ptr], #448]")
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "ldr    %q[a0], [%[a_ptr], #32]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #48]\n"
-
-                "fmla     v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "fmla    v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #80]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "fmla     v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "ldr	%q[a0], [%[a_ptr], #32]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ldr	%q[a1], [%[a_ptr], #48]\n"
+
+                "fmla 	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "fmla	v9.4s , %[b0].4s, %[a0].s[1]\n"
+                "ldr	%q[b2], [%[b_ptr], #80]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "fmla 	v12.4s, %[b0].4s, %[a1].s[0]\n"
                 "fmla   v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #96]\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "ldr	%q[b0], [%[b_ptr], #96]\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
                 ASM_PREFETCH("[%[b_ptr], #512]")
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #112]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "ldr    %q[a0], [%[a_ptr]]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #16]\n"
-                "bne    1b\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "ldr	%q[b1], [%[b_ptr], #112]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "ldr	%q[a0], [%[a_ptr]]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ldr	%q[a1], [%[a_ptr], #16]\n"
+                "bne	1b\n"
 
                 // Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
                 "4:\n"
 
                 // Branch to alternative tail for odd K
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration (even K)
-                "fmla     v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "fmla 	v8.4s , %[b0].4s, %[a0].s[0]\n"
                 "fmla   v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "fmla     v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "fmla 	v12.4s, %[b0].4s, %[a1].s[0]\n"
                 "fmla   v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "ldr    %q[b0], [%[b_ptr], #48]\n"
-
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "ldr    %q[b1], [%[b_ptr], #64]\n"
-
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #64\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "ldr    %q[a0], [%[a_ptr], #-32]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "add    %[b_ptr], %[b_ptr], #96\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "ldr    %q[a1], [%[a_ptr], #-16]\n"
-
-                "fmla     v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #-16]\n"
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "ldr	%q[b0], [%[b_ptr], #48]\n"
+
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "ldr	%q[b1], [%[b_ptr], #64]\n"
+
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #64\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "ldr	%q[a0], [%[a_ptr], #-32]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "add	%[b_ptr], %[b_ptr], #96\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "ldr	%q[a1], [%[a_ptr], #-16]\n"
+
+                "fmla 	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "ldr	%q[b2], [%[b_ptr], #-16]\n"
                 "fmla   v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "str    q24, [%[c_ptr], #32]\n"
-
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-
-                "fmla     v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q8, [%[c_ptr], #0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "str	q24, [%[c_ptr], #32]\n"
+
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "str	q9, [%[c_ptr], #48]\n"
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+
+                "fmla 	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 "fmla   v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
-
-                "b    3f\n"
+                "str	q20, [%[c_ptr], #208]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "str	q15, [%[c_ptr], #336]\n"
+
+                "b	3f\n"
 
                 // Detached final iteration (odd K)
                 "2:\n"
-                "fmla     v8.4s , %[b0].4s, %[a0].s[0]\n"
-                "ldr    %q[b2], [%[b_ptr], #32]\n"
-                "fmla    v16.4s, %[b1].4s, %[a0].s[0]\n"
+                "fmla 	v8.4s , %[b0].4s, %[a0].s[0]\n"
+                "ldr	%q[b2], [%[b_ptr], #32]\n"
+                "fmla	v16.4s, %[b1].4s, %[a0].s[0]\n"
                 "fmla   v9.4s , %[b0].4s, %[a0].s[1]\n"
-                "str    q8, [%[c_ptr], #0]\n"
-                "fmla    v17.4s, %[b1].4s, %[a0].s[1]\n"
-                "str    q16, [%[c_ptr], #16]\n"
-                "fmla    v24.4s, %[b2].4s, %[a0].s[0]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
-                "add    %[a_ptr], %[a_ptr], #32\n"
-                "str    q24, [%[c_ptr], #32]\n"
-                "fmla    v25.4s, %[b2].4s, %[a0].s[1]\n"
-                "str    q9, [%[c_ptr], #48]\n"
-
-                "fmla    v10.4s, %[b0].4s, %[a0].s[2]\n"
-                "str    q17, [%[c_ptr], #64]\n"
-                "fmla    v18.4s, %[b1].4s, %[a0].s[2]\n"
-                "str    q25, [%[c_ptr], #80]\n"
-                "fmla    v26.4s, %[b2].4s, %[a0].s[2]\n"
-                "str    q10, [%[c_ptr], #96]\n"
-
-                "fmla    v11.4s, %[b0].4s, %[a0].s[3]\n"
-                "str    q18, [%[c_ptr], #112]\n"
-                "fmla    v19.4s, %[b1].4s, %[a0].s[3]\n"
-                "str    q26, [%[c_ptr], #128]\n"
-                "fmla    v27.4s, %[b2].4s, %[a0].s[3]\n"
-                "str    q11, [%[c_ptr], #144]\n"
-
-                "fmla     v12.4s, %[b0].4s, %[a1].s[0]\n"
-                "str    q19, [%[c_ptr], #160]\n"
-                "fmla    v20.4s, %[b1].4s, %[a1].s[0]\n"
-                "str    q27, [%[c_ptr], #176]\n"
-                "fmla    v28.4s, %[b2].4s, %[a1].s[0]\n"
-                "str    q12, [%[c_ptr], #192]\n"
+                "str	q8, [%[c_ptr], #0]\n"
+                "fmla	v17.4s, %[b1].4s, %[a0].s[1]\n"
+                "str	q16, [%[c_ptr], #16]\n"
+                "fmla	v24.4s, %[b2].4s, %[a0].s[0]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
+                "add	%[a_ptr], %[a_ptr], #32\n"
+                "str	q24, [%[c_ptr], #32]\n"
+                "fmla	v25.4s, %[b2].4s, %[a0].s[1]\n"
+                "str	q9, [%[c_ptr], #48]\n"
+
+                "fmla	v10.4s, %[b0].4s, %[a0].s[2]\n"
+                "str	q17, [%[c_ptr], #64]\n"
+                "fmla	v18.4s, %[b1].4s, %[a0].s[2]\n"
+                "str	q25, [%[c_ptr], #80]\n"
+                "fmla	v26.4s, %[b2].4s, %[a0].s[2]\n"
+                "str	q10, [%[c_ptr], #96]\n"
+
+                "fmla	v11.4s, %[b0].4s, %[a0].s[3]\n"
+                "str	q18, [%[c_ptr], #112]\n"
+                "fmla	v19.4s, %[b1].4s, %[a0].s[3]\n"
+                "str	q26, [%[c_ptr], #128]\n"
+                "fmla	v27.4s, %[b2].4s, %[a0].s[3]\n"
+                "str	q11, [%[c_ptr], #144]\n"
+
+                "fmla 	v12.4s, %[b0].4s, %[a1].s[0]\n"
+                "str	q19, [%[c_ptr], #160]\n"
+                "fmla	v20.4s, %[b1].4s, %[a1].s[0]\n"
+                "str	q27, [%[c_ptr], #176]\n"
+                "fmla	v28.4s, %[b2].4s, %[a1].s[0]\n"
+                "str	q12, [%[c_ptr], #192]\n"
 
                 "fmla   v13.4s, %[b0].4s, %[a1].s[1]\n"
-                "str    q20, [%[c_ptr], #208]\n"
-                "fmla    v21.4s, %[b1].4s, %[a1].s[1]\n"
-                "str    q28, [%[c_ptr], #224]\n"
-                "fmla    v29.4s, %[b2].4s, %[a1].s[1]\n"
-                "str    q13, [%[c_ptr], #240]\n"
-
-                "fmla    v14.4s, %[b0].4s, %[a1].s[2]\n"
-                "str    q21, [%[c_ptr], #256]\n"
-                "fmla    v22.4s, %[b1].4s, %[a1].s[2]\n"
-                "str    q29, [%[c_ptr], #272]\n"
-                "fmla    v30.4s, %[b2].4s, %[a1].s[2]\n"
-                "str    q14, [%[c_ptr], #288]\n"
-
-                "fmla    v15.4s, %[b0].4s, %[a1].s[3]\n"
-                "str    q22, [%[c_ptr], #304]\n"
-                "fmla    v23.4s, %[b1].4s, %[a1].s[3]\n"
-                "str    q30, [%[c_ptr], #320]\n"
-                "fmla    v31.4s, %[b2].4s, %[a1].s[3]\n"
-                "str    q15, [%[c_ptr], #336]\n"
+                "str	q20, [%[c_ptr], #208]\n"
+                "fmla	v21.4s, %[b1].4s, %[a1].s[1]\n"
+                "str	q28, [%[c_ptr], #224]\n"
+                "fmla	v29.4s, %[b2].4s, %[a1].s[1]\n"
+                "str	q13, [%[c_ptr], #240]\n"
+
+                "fmla	v14.4s, %[b0].4s, %[a1].s[2]\n"
+                "str	q21, [%[c_ptr], #256]\n"
+                "fmla	v22.4s, %[b1].4s, %[a1].s[2]\n"
+                "str	q29, [%[c_ptr], #272]\n"
+                "fmla	v30.4s, %[b2].4s, %[a1].s[2]\n"
+                "str	q14, [%[c_ptr], #288]\n"
+
+                "fmla	v15.4s, %[b0].4s, %[a1].s[3]\n"
+                "str	q22, [%[c_ptr], #304]\n"
+                "fmla	v23.4s, %[b1].4s, %[a1].s[3]\n"
+                "str	q30, [%[c_ptr], #320]\n"
+                "fmla	v31.4s, %[b2].4s, %[a1].s[3]\n"
+                "str	q15, [%[c_ptr], #336]\n"
 
                 // Common tail
                 "3:\n"
-                "str    q23, [%[c_ptr], #352]\n"
-                "str    q31, [%[c_ptr], #368]\n"
-                "add    %[c_ptr], %[c_ptr], #384\n"
+                "str	q23, [%[c_ptr], #352]\n"
+                "str	q31, [%[c_ptr], #368]\n"
+                "add	%[c_ptr], %[c_ptr], #384\n"
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
               [a0] "+w" (a0), [a1] "+w" (a1),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp
index 2fed3264ab..b35cc91a5d 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6.hpp
@@ -42,8 +42,7 @@ void a64_sgemm_asimd_8x6(const float *, const float *, float *, int, int, int);
 // structure.
 class cls_a64_sgemm_8x6 {
 public:
-    typedef float lhs_operand_type;
-    typedef float rhs_operand_type;
+    typedef float operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
@@ -62,7 +61,7 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsFixed<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 1> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 6, 1> transforms = {};
 
     kern_type kernel=a64_sgemm_asimd_8x6;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp
index fb5044684f..a968105af1 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp
@@ -64,305 +64,305 @@ void a64_sgemm_asimd_8x6(const float *Apanel, const float *Bpanel, float *Cpanel
 
             __asm __volatile (
                 // Initialize result registers, load initial operands, prime prefetches.
-                "movi    v8.2s, #0x0\n"
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "movi    v9.2s, #0x0\n"
-                "movi    v10.2s, #0x0\n"
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
-                "movi    v11.2s, #0x0\n"
-                "movi    v12.2s, #0x0\n"
-                "movi    v13.2s, #0x0\n"
-                "movi    v14.2s, #0x0\n"
+                "movi	v8.2s, #0x0\n"
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "movi	v9.2s, #0x0\n"
+                "movi	v10.2s, #0x0\n"
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "movi	v11.2s, #0x0\n"
+                "movi	v12.2s, #0x0\n"
+                "movi	v13.2s, #0x0\n"
+                "movi	v14.2s, #0x0\n"
                 ASM_PREFETCH("[%[b_ptr], #64]")
                 ASM_PREFETCHU("[%[a_ptr], #52]")
                 ASM_PREFETCHU("[%[a_ptr], #116]")
                 ASM_PREFETCH("[%[b_ptr], #128]")
-                "movi    v15.2s, #0x0\n"
-                "movi    v16.2s, #0x0\n"
-                "movi    v17.2s, #0x0\n"
-                "movi    v18.2s, #0x0\n"
-                "movi    v19.2s, #0x0\n"
-                "movi    v20.2s, #0x0\n"
-                "movi    v21.2s, #0x0\n"
-                "movi    v22.2s, #0x0\n"
-                "movi    v23.2s, #0x0\n"
-                "movi    v24.2s, #0x0\n"
-                "movi    v25.2s, #0x0\n"
-                "movi    v26.2s, #0x0\n"
-                "movi    v27.2s, #0x0\n"
-                "movi    v28.2s, #0x0\n"
-                "movi    v29.2s, #0x0\n"
-                "movi    v30.2s, #0x0\n"
-                "movi    v31.2s, #0x0\n"
+                "movi	v15.2s, #0x0\n"
+                "movi	v16.2s, #0x0\n"
+                "movi	v17.2s, #0x0\n"
+                "movi	v18.2s, #0x0\n"
+                "movi	v19.2s, #0x0\n"
+                "movi	v20.2s, #0x0\n"
+                "movi	v21.2s, #0x0\n"
+                "movi	v22.2s, #0x0\n"
+                "movi	v23.2s, #0x0\n"
+                "movi	v24.2s, #0x0\n"
+                "movi	v25.2s, #0x0\n"
+                "movi	v26.2s, #0x0\n"
+                "movi	v27.2s, #0x0\n"
+                "movi	v28.2s, #0x0\n"
+                "movi	v29.2s, #0x0\n"
+                "movi	v30.2s, #0x0\n"
+                "movi	v31.2s, #0x0\n"
 
                 // Skip loop if we are doing zero iterations of it.
-                "cbz    %w[k], 4f\n"
+                "cbz	%w[k], 4f\n"
 
                 // Loop proper
                 "1:\n"
-                "ldr    %d[b0], [%[b_ptr], #0]\n"
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
-                "ldr    %d[b1], [%[b_ptr], #8]\n"
-                "fmla     v8.2s , %[b0].2s, %[a0].2s\n"
-                "fmla      v9.2s , %[b0].2s, %[a1].2s\n"
-                "fmla    v10.2s, %[b0].2s, %[a2].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v16.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v17.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v11.2s, %[b0].2s, %[a3].2s\n"
-
-                "ldr    %d[b2], [%[b_ptr], #16]\n"
-                "fmla    v18.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v19.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v24.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "fmla    v25.2s, %[b2].2s, %[a1].2s\n"
-                "fmla    v26.2s, %[b2].2s, %[a2].2s\n"
-                "fmla    v27.2s, %[b2].2s, %[a3].2s\n"
-
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
-                "fmla     v12.2s, %[b0].2s, %[a0].2s\n"
-                "fmla    v20.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v28.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
-                "fmla    v13.2s, %[b0].2s, %[a1].2s\n"
-                "fmla    v21.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v29.2s, %[b2].2s, %[a1].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v14.2s, %[b0].2s, %[a2].2s\n"
-                "fmla    v22.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v30.2s, %[b2].2s, %[a2].2s\n"
-
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "fmla    v15.2s, %[b0].2s, %[a3].2s\n"
-                "fmla    v23.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v31.2s, %[b2].2s, %[a3].2s\n"
-
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
+                "ldr	%d[b0], [%[b_ptr], #0]\n"
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "ldr	%d[b1], [%[b_ptr], #8]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #16]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
                  ASM_PREFETCH("[%[b_ptr], #128]")
-                "subs    %w[k], %w[k], #1\n"
+                "subs	%w[k], %w[k], #1\n"
                  ASM_PREFETCHU("[%[a_ptr], #156]")
-                "ldr    %d[b0], [%[b_ptr], #24]\n"
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
-
-                "ldr    %d[b1], [%[b_ptr], #32]\n"
-                "fmla     v8.2s , %[b0].2s, %[a0].2s\n"
-                "fmla      v9.2s , %[b0].2s, %[a1].2s\n"
-                "fmla    v10.2s, %[b0].2s, %[a2].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v16.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v17.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v11.2s, %[b0].2s, %[a3].2s\n"
-
-                "ldr    %d[b2], [%[b_ptr], #40]\n"
-                "fmla    v18.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v19.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v24.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "fmla    v25.2s, %[b2].2s, %[a1].2s\n"
-                "fmla    v26.2s, %[b2].2s, %[a2].2s\n"
-                "fmla    v27.2s, %[b2].2s, %[a3].2s\n"
-
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
-                "fmla     v12.2s, %[b0].2s, %[a0].2s\n"
-                "fmla    v20.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v28.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
-                "fmla    v13.2s, %[b0].2s, %[a1].2s\n"
-                "fmla    v21.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v29.2s, %[b2].2s, %[a1].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v14.2s, %[b0].2s, %[a2].2s\n"
-                "fmla    v22.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v30.2s, %[b2].2s, %[a2].2s\n"
-
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "fmla    v15.2s, %[b0].2s, %[a3].2s\n"
-                "fmla    v23.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v31.2s, %[b2].2s, %[a3].2s\n"
-
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
+                "ldr	%d[b0], [%[b_ptr], #24]\n"
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+
+                "ldr	%d[b1], [%[b_ptr], #32]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #40]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
                  ASM_PREFETCHU("[%[a_ptr], #188]")
-                "bne    1b\n"
+                "bne	1b\n"
 
                 // Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
                 "4:\n"
                 ASM_PREFETCH("[%[c_ptr]]")
                 ASM_PREFETCH("[%[c_ptr], #64]")
 
-                "ldr    %d[b0], [%[b_ptr]]\n"
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
+                "ldr	%d[b0], [%[b_ptr]]\n"
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
 
                 // Branch to alternative tail for odd K
-                "cbnz    %w[oddk], 2f\n"
+                "cbnz	%w[oddk], 2f\n"
 
                 // Detached final iteration (even K)
-                "ldr    %d[b1], [%[b_ptr], #8]\n"
-                "fmla     v8.2s , %[b0].2s, %[a0].2s\n"
-                "fmla      v9.2s , %[b0].2s, %[a1].2s\n"
-                "fmla    v10.2s, %[b0].2s, %[a2].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v16.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v17.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v11.2s, %[b0].2s, %[a3].2s\n"
-
-                "ldr    %d[b2], [%[b_ptr], #16]\n"
-                "fmla    v18.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v19.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v24.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "fmla    v25.2s, %[b2].2s, %[a1].2s\n"
-                "fmla    v26.2s, %[b2].2s, %[a2].2s\n"
-                "fmla    v27.2s, %[b2].2s, %[a3].2s\n"
-
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
-                "fmla     v12.2s, %[b0].2s, %[a0].2s\n"
-                "fmla    v20.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v28.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
-                "fmla    v13.2s, %[b0].2s, %[a1].2s\n"
-                "fmla    v21.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v29.2s, %[b2].2s, %[a1].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v14.2s, %[b0].2s, %[a2].2s\n"
-                "fmla    v22.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v30.2s, %[b2].2s, %[a2].2s\n"
-
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "fmla    v15.2s, %[b0].2s, %[a3].2s\n"
-                "fmla    v23.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v31.2s, %[b2].2s, %[a3].2s\n"
-
-                "ldr    %d[b0], [%[b_ptr], #24]\n"
-                "add    %[b_ptr], %[b_ptr], #48\n"
+                "ldr	%d[b1], [%[b_ptr], #8]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #16]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "ldr	%d[b0], [%[b_ptr], #24]\n"
+                "add	%[b_ptr], %[b_ptr], #48\n"
                  ASM_PREFETCH("[%[b_ptr], #128]")
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
-
-                "ldr    %d[b1], [%[b_ptr], #-16]\n"
-                "fmla     v8.2s , %[b0].2s, %[a0].2s\n"
-                "fmla      v9.2s , %[b0].2s, %[a1].2s\n"
-                "fmla    v10.2s, %[b0].2s, %[a2].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v16.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v17.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v11.2s, %[b0].2s, %[a3].2s\n"
-
-                "ldr    %d[b2], [%[b_ptr], #-8]\n"
-                "fmla    v18.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v19.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v24.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "fmla    v25.2s, %[b2].2s, %[a1].2s\n"
-                "fmla    v26.2s, %[b2].2s, %[a2].2s\n"
-                "fmla    v27.2s, %[b2].2s, %[a3].2s\n"
-
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
-                "fmla     v12.2s, %[b0].2s, %[a0].2s\n"
-                "fmla    v20.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v28.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
-                "fmla    v13.2s, %[b0].2s, %[a1].2s\n"
-                "fmla    v21.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v29.2s, %[b2].2s, %[a1].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v14.2s, %[b0].2s, %[a2].2s\n"
-                "fmla    v22.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v30.2s, %[b2].2s, %[a2].2s\n"
-
-                "fmla    v15.2s, %[b0].2s, %[a3].2s\n"
-                "fmla    v23.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v31.2s, %[b2].2s, %[a3].2s\n"
-
-                "b    3f\n"
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+
+                "ldr	%d[b1], [%[b_ptr], #-16]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #-8]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "b	3f\n"
 
                 // Detached final iteration (odd K)
                 "2:\n"
-                "ldr    %d[b1], [%[b_ptr], #8]\n"
-                "fmla     v8.2s , %[b0].2s, %[a0].2s\n"
-                "fmla      v9.2s , %[b0].2s, %[a1].2s\n"
-                "fmla    v10.2s, %[b0].2s, %[a2].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v16.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v17.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v11.2s, %[b0].2s, %[a3].2s\n"
-
-                "ldr    %d[b2], [%[b_ptr], #16]\n"
-                "fmla    v18.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v19.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v24.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a0].2s }, [%[a_ptr]], #4\n"
-                "fmla    v25.2s, %[b2].2s, %[a1].2s\n"
-                "fmla    v26.2s, %[b2].2s, %[a2].2s\n"
-                "fmla    v27.2s, %[b2].2s, %[a3].2s\n"
-
-                "ld1r    { %[a1].2s }, [%[a_ptr]], #4\n"
-                "fmla     v12.2s, %[b0].2s, %[a0].2s\n"
-                "fmla    v20.2s, %[b1].2s, %[a0].2s\n"
-                "fmla    v28.2s, %[b2].2s, %[a0].2s\n"
-
-                "ld1r    { %[a2].2s }, [%[a_ptr]], #4\n"
-                "fmla    v13.2s, %[b0].2s, %[a1].2s\n"
-                "fmla    v21.2s, %[b1].2s, %[a1].2s\n"
-                "fmla    v29.2s, %[b2].2s, %[a1].2s\n"
-
-                "ld1r    { %[a3].2s }, [%[a_ptr]], #4\n"
-                "fmla    v14.2s, %[b0].2s, %[a2].2s\n"
-                "fmla    v22.2s, %[b1].2s, %[a2].2s\n"
-                "fmla    v30.2s, %[b2].2s, %[a2].2s\n"
-
-                "fmla    v15.2s, %[b0].2s, %[a3].2s\n"
-                "fmla    v23.2s, %[b1].2s, %[a3].2s\n"
-                "fmla    v31.2s, %[b2].2s, %[a3].2s\n"
-
-                "add    %[b_ptr], %[b_ptr], #24\n"
+                "ldr	%d[b1], [%[b_ptr], #8]\n"
+                "fmla 	v8.2s , %[b0].2s, %[a0].2s\n"
+                "fmla  	v9.2s , %[b0].2s, %[a1].2s\n"
+                "fmla	v10.2s, %[b0].2s, %[a2].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v16.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v17.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v11.2s, %[b0].2s, %[a3].2s\n"
+
+                "ldr	%d[b2], [%[b_ptr], #16]\n"
+                "fmla	v18.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v19.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v24.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a0].2s }, [%[a_ptr]], #4\n"
+                "fmla	v25.2s, %[b2].2s, %[a1].2s\n"
+                "fmla	v26.2s, %[b2].2s, %[a2].2s\n"
+                "fmla	v27.2s, %[b2].2s, %[a3].2s\n"
+
+                "ld1r	{ %[a1].2s }, [%[a_ptr]], #4\n"
+                "fmla 	v12.2s, %[b0].2s, %[a0].2s\n"
+                "fmla	v20.2s, %[b1].2s, %[a0].2s\n"
+                "fmla	v28.2s, %[b2].2s, %[a0].2s\n"
+
+                "ld1r	{ %[a2].2s }, [%[a_ptr]], #4\n"
+                "fmla	v13.2s, %[b0].2s, %[a1].2s\n"
+                "fmla	v21.2s, %[b1].2s, %[a1].2s\n"
+                "fmla	v29.2s, %[b2].2s, %[a1].2s\n"
+
+                "ld1r	{ %[a3].2s }, [%[a_ptr]], #4\n"
+                "fmla	v14.2s, %[b0].2s, %[a2].2s\n"
+                "fmla	v22.2s, %[b1].2s, %[a2].2s\n"
+                "fmla	v30.2s, %[b2].2s, %[a2].2s\n"
+
+                "fmla	v15.2s, %[b0].2s, %[a3].2s\n"
+                "fmla	v23.2s, %[b1].2s, %[a3].2s\n"
+                "fmla	v31.2s, %[b2].2s, %[a3].2s\n"
+
+                "add	%[b_ptr], %[b_ptr], #24\n"
 
                 // Common tail
                 "3:\n"
-                "str    d8, [%[c_ptr], #0]\n"
-                "str    d16, [%[c_ptr], #8]\n"
-                "str    d24, [%[c_ptr], #16]\n"
-                "str    d9, [%[c_ptr], #24]\n"
-                "str    d17, [%[c_ptr], #32]\n"
-                "str    d25, [%[c_ptr], #40]\n"
-                "str    d10, [%[c_ptr], #48]\n"
-                "str    d18, [%[c_ptr], #56]\n"
-                "str    d26, [%[c_ptr], #64]\n"
-                "str    d11, [%[c_ptr], #72]\n"
-                "str    d19, [%[c_ptr], #80]\n"
-                "str    d27, [%[c_ptr], #88]\n"
-                "str    d12, [%[c_ptr], #96]\n"
-                "str    d20, [%[c_ptr], #104]\n"
-                "str    d28, [%[c_ptr], #112]\n"
-                "str    d13, [%[c_ptr], #120]\n"
-                "str    d21, [%[c_ptr], #128]\n"
-                "str    d29, [%[c_ptr], #136]\n"
-                "str    d14, [%[c_ptr], #144]\n"
-                "str    d22, [%[c_ptr], #152]\n"
-                "str    d30, [%[c_ptr], #160]\n"
-                "str    d15, [%[c_ptr], #168]\n"
-                "str    d23, [%[c_ptr], #176]\n"
-                "str    d31, [%[c_ptr], #184]\n"
-                "add    %[c_ptr], %[c_ptr], #192\n"
+                "str	d8, [%[c_ptr], #0]\n"
+                "str	d16, [%[c_ptr], #8]\n"
+                "str	d24, [%[c_ptr], #16]\n"
+                "str	d9, [%[c_ptr], #24]\n"
+                "str	d17, [%[c_ptr], #32]\n"
+                "str	d25, [%[c_ptr], #40]\n"
+                "str	d10, [%[c_ptr], #48]\n"
+                "str	d18, [%[c_ptr], #56]\n"
+                "str	d26, [%[c_ptr], #64]\n"
+                "str	d11, [%[c_ptr], #72]\n"
+                "str	d19, [%[c_ptr], #80]\n"
+                "str	d27, [%[c_ptr], #88]\n"
+                "str	d12, [%[c_ptr], #96]\n"
+                "str	d20, [%[c_ptr], #104]\n"
+                "str	d28, [%[c_ptr], #112]\n"
+                "str	d13, [%[c_ptr], #120]\n"
+                "str	d21, [%[c_ptr], #128]\n"
+                "str	d29, [%[c_ptr], #136]\n"
+                "str	d14, [%[c_ptr], #144]\n"
+                "str	d22, [%[c_ptr], #152]\n"
+                "str	d30, [%[c_ptr], #160]\n"
+                "str	d15, [%[c_ptr], #168]\n"
+                "str	d23, [%[c_ptr], #176]\n"
+                "str	d31, [%[c_ptr], #184]\n"
+                "add	%[c_ptr], %[c_ptr], #192\n"
 
             :
               [a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp
index 3616f39f2a..702b5f69ff 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp
@@ -170,28 +170,28 @@ void a64_sgemv_pretransposed(const float *A, int lda, const float *X, float *Y,
             x0 = vld1q_f32(x_ptr);
 
             __asm __volatile (
-                "ldr    q2, [%[a_ptr], #0]\n"
-                "ldr    q3, [%[a_ptr], #16]\n"
-                "ldr    q4, [%[a_ptr], #32]\n"
-                "ldr    q5, [%[a_ptr], #48]\n"
-                "ldr    q6, [%[a_ptr], #64]\n"
-                "ldr    q7, [%[a_ptr], #80]\n"
-                "ldr    q8, [%[a_ptr], #96]\n"
-                "ldr    q9, [%[a_ptr], #112]\n"
-                "ldr    q10, [%[a_ptr], #128]\n"
-                "ldr    q11, [%[a_ptr], #144]\n"
-                "ldr    q12, [%[a_ptr], #160]\n"
-                "ldr    q13, [%[a_ptr], #176]\n"
-                "ldr    q14, [%[a_ptr], #192]\n"
-                "ldr    q15, [%[a_ptr], #208]\n"
-                "ldr    q16, [%[a_ptr], #224]\n"
-                "ldr    q17, [%[a_ptr], #240]\n"
-                "ldr    q18, [%[a_ptr], #256]\n"
-                "ldr    q19, [%[a_ptr], #272]\n"
-                "ldr    q20, [%[a_ptr], #288]\n"
-                "ldr    q21, [%[a_ptr], #304]\n"
-                "ldr    q22, [%[a_ptr], #320]\n"
-                "ldr    q23, [%[a_ptr], #336]\n"
+                "ldr	q2, [%[a_ptr], #0]\n"
+                "ldr	q3, [%[a_ptr], #16]\n"
+                "ldr	q4, [%[a_ptr], #32]\n"
+                "ldr	q5, [%[a_ptr], #48]\n"
+                "ldr	q6, [%[a_ptr], #64]\n"
+                "ldr	q7, [%[a_ptr], #80]\n"
+                "ldr	q8, [%[a_ptr], #96]\n"
+                "ldr	q9, [%[a_ptr], #112]\n"
+                "ldr	q10, [%[a_ptr], #128]\n"
+                "ldr	q11, [%[a_ptr], #144]\n"
+                "ldr	q12, [%[a_ptr], #160]\n"
+                "ldr	q13, [%[a_ptr], #176]\n"
+                "ldr	q14, [%[a_ptr], #192]\n"
+                "ldr	q15, [%[a_ptr], #208]\n"
+                "ldr	q16, [%[a_ptr], #224]\n"
+                "ldr	q17, [%[a_ptr], #240]\n"
+                "ldr	q18, [%[a_ptr], #256]\n"
+                "ldr	q19, [%[a_ptr], #272]\n"
+                "ldr	q20, [%[a_ptr], #288]\n"
+                "ldr	q21, [%[a_ptr], #304]\n"
+                "ldr	q22, [%[a_ptr], #320]\n"
+                "ldr	q23, [%[a_ptr], #336]\n"
                 ASM_PREFETCH("[%[a_ptr], #384]")
                 ASM_PREFETCH("[%[a_ptr], #448]")
                 ASM_PREFETCH("[%[a_ptr], #512]")
@@ -218,305 +218,305 @@ void a64_sgemv_pretransposed(const float *A, int lda, const float *X, float *Y,
                 ASM_PREFETCH("[%[a_ptr], #1856]")
                 ASM_PREFETCH("[%[a_ptr], #1920]")
                 ASM_PREFETCH("[%[a_ptr], #1984]")
-                "add    %[a_ptr], %[a_ptr], #352\n"
+                "add	%[a_ptr], %[a_ptr], #352\n"
 
-                "cbz    %w[k], 2f\n"
+                "cbz	%w[k], 2f\n"
 
                 "1:\n"
                 // Unroll 0
-                "fmla    %[r0].4s, v2.4s, %[x0].s[0]\n"
-                "ldr    %q[x0a], [%[x_ptr], #16]\n"
-                "fmla    %[r1].4s, v3.4s, %[x0].s[0]\n"
-                "ldr    q3, [%[a_ptr], #0]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    %[r2].4s, v4.4s, %[x0].s[0]\n"
-                "ldr    q4, [%[a_ptr], #16]\n"
-                "fmla    %[r3].4s, v5.4s, %[x0].s[0]\n"
-                "ldr    q5, [%[a_ptr], #32]\n"
-                "add    %[x_ptr], %[x_ptr], #32\n"
+                "fmla	%[r0].4s, v2.4s, %[x0].s[0]\n"
+                "ldr	%q[x0a], [%[x_ptr], #16]\n"
+                "fmla	%[r1].4s, v3.4s, %[x0].s[0]\n"
+                "ldr	q3, [%[a_ptr], #0]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	%[r2].4s, v4.4s, %[x0].s[0]\n"
+                "ldr	q4, [%[a_ptr], #16]\n"
+                "fmla	%[r3].4s, v5.4s, %[x0].s[0]\n"
+                "ldr	q5, [%[a_ptr], #32]\n"
+                "add	%[x_ptr], %[x_ptr], #32\n"
                 ASM_PREFETCH("[%[a_ptr], #1664]")
-                "fmla    %[r4].4s, v6.4s, %[x0].s[0]\n"
-                "ldr    q6, [%[a_ptr], #48]\n"
-                "fmla    %[r5].4s, v7.4s, %[x0].s[0]\n"
-                "ldr    q7, [%[a_ptr], #64]\n"
-                "fmla    %[r6].4s, v8.4s, %[x0].s[0]\n"
-                "ldr    q8, [%[a_ptr], #80]\n"
-                "fmla    %[r7].4s, v9.4s, %[x0].s[0]\n"
-                "ldr    q9, [%[a_ptr], #96]\n"
+                "fmla	%[r4].4s, v6.4s, %[x0].s[0]\n"
+                "ldr	q6, [%[a_ptr], #48]\n"
+                "fmla	%[r5].4s, v7.4s, %[x0].s[0]\n"
+                "ldr	q7, [%[a_ptr], #64]\n"
+                "fmla	%[r6].4s, v8.4s, %[x0].s[0]\n"
+                "ldr	q8, [%[a_ptr], #80]\n"
+                "fmla	%[r7].4s, v9.4s, %[x0].s[0]\n"
+                "ldr	q9, [%[a_ptr], #96]\n"
                 ASM_PREFETCH("[%[a_ptr], #1728]")
 
                 // Unroll 1
-                "fmla    %[r0].4s, v10.4s, %[x0].s[1]\n"
-                "ldr    q10, [%[a_ptr], #112]\n"
-                "fmla    %[r1].4s, v11.4s, %[x0].s[1]\n"
-                "ldr    q11, [%[a_ptr], #128]\n"
-                "fmla    %[r2].4s, v12.4s, %[x0].s[1]\n"
-                "ldr    q12, [%[a_ptr], #144]\n"
-                "fmla    %[r3].4s, v13.4s, %[x0].s[1]\n"
-                "ldr    q13, [%[a_ptr], #160]\n"
+                "fmla	%[r0].4s, v10.4s, %[x0].s[1]\n"
+                "ldr	q10, [%[a_ptr], #112]\n"
+                "fmla	%[r1].4s, v11.4s, %[x0].s[1]\n"
+                "ldr	q11, [%[a_ptr], #128]\n"
+                "fmla	%[r2].4s, v12.4s, %[x0].s[1]\n"
+                "ldr	q12, [%[a_ptr], #144]\n"
+                "fmla	%[r3].4s, v13.4s, %[x0].s[1]\n"
+                "ldr	q13, [%[a_ptr], #160]\n"
                 ASM_PREFETCH("[%[a_ptr], #1792]")
-                "fmla    %[r4].4s, v14.4s, %[x0].s[1]\n"
-                "ldr    q14, [%[a_ptr], #176]\n"
-                "fmla    %[r5].4s, v15.4s, %[x0].s[1]\n"
-                "ldr    q15, [%[a_ptr], #192]\n"
-                "fmla    %[r6].4s, v16.4s, %[x0].s[1]\n"
-                "ldr    q16, [%[a_ptr], #208]\n"
-                "fmla    %[r7].4s, v17.4s, %[x0].s[1]\n"
-                "ldr    q17, [%[a_ptr], #224]\n"
+                "fmla	%[r4].4s, v14.4s, %[x0].s[1]\n"
+                "ldr	q14, [%[a_ptr], #176]\n"
+                "fmla	%[r5].4s, v15.4s, %[x0].s[1]\n"
+                "ldr	q15, [%[a_ptr], #192]\n"
+                "fmla	%[r6].4s, v16.4s, %[x0].s[1]\n"
+                "ldr	q16, [%[a_ptr], #208]\n"
+                "fmla	%[r7].4s, v17.4s, %[x0].s[1]\n"
+                "ldr	q17, [%[a_ptr], #224]\n"
                 ASM_PREFETCH("[%[a_ptr], #1856]")
 
                 // Unroll 2
-                "fmla    %[r0].4s, v18.4s, %[x0].s[2]\n"
-                "ldr    q18, [%[a_ptr], #240]\n"
-                "fmla    %[r1].4s, v19.4s, %[x0].s[2]\n"
-                "ldr    q19, [%[a_ptr], #256]\n"
-                "fmla    %[r2].4s, v20.4s, %[x0].s[2]\n"
-                "ldr    q20, [%[a_ptr], #272]\n"
-                "fmla    %[r3].4s, v21.4s, %[x0].s[2]\n"
-                "ldr    q21, [%[a_ptr], #288]\n"
+                "fmla	%[r0].4s, v18.4s, %[x0].s[2]\n"
+                "ldr	q18, [%[a_ptr], #240]\n"
+                "fmla	%[r1].4s, v19.4s, %[x0].s[2]\n"
+                "ldr	q19, [%[a_ptr], #256]\n"
+                "fmla	%[r2].4s, v20.4s, %[x0].s[2]\n"
+                "ldr	q20, [%[a_ptr], #272]\n"
+                "fmla	%[r3].4s, v21.4s, %[x0].s[2]\n"
+                "ldr	q21, [%[a_ptr], #288]\n"
                 ASM_PREFETCH("[%[a_ptr], #1920]")
-                "fmla    %[r4].4s, v22.4s, %[x0].s[2]\n"
-                "ldr    q22, [%[a_ptr], #304]\n"
-                "fmla    %[r5].4s, v23.4s, %[x0].s[2]\n"
-                "ldr    q23, [%[a_ptr], #320]\n"
-                "fmla    %[r6].4s, v3.4s, %[x0].s[2]\n"
-                "ldr    q2, [%[a_ptr], #336]\n"
-                "ldr    q3, [%[a_ptr], #352]\n"
-                "fmla    %[r7].4s, v4.4s, %[x0].s[2]\n"
-                "ldr    q4, [%[a_ptr], #368]\n"
+                "fmla	%[r4].4s, v22.4s, %[x0].s[2]\n"
+                "ldr	q22, [%[a_ptr], #304]\n"
+                "fmla	%[r5].4s, v23.4s, %[x0].s[2]\n"
+                "ldr	q23, [%[a_ptr], #320]\n"
+                "fmla	%[r6].4s, v3.4s, %[x0].s[2]\n"
+                "ldr	q2, [%[a_ptr], #336]\n"
+                "ldr	q3, [%[a_ptr], #352]\n"
+                "fmla	%[r7].4s, v4.4s, %[x0].s[2]\n"
+                "ldr	q4, [%[a_ptr], #368]\n"
                 ASM_PREFETCH("[%[a_ptr], #1984]")
 
                 // Unroll 3
-                "fmla    %[r0].4s, v5.4s, %[x0].s[3]\n"
-                "ldr    q5, [%[a_ptr], #384]\n"
-                "fmla    %[r1].4s, v6.4s, %[x0].s[3]\n"
-                "ldr    q6, [%[a_ptr], #400]\n"
-                "fmla    %[r2].4s, v7.4s, %[x0].s[3]\n"
-                "ldr    q7, [%[a_ptr], #416]\n"
-                "fmla    %[r3].4s, v8.4s, %[x0].s[3]\n"
+                "fmla	%[r0].4s, v5.4s, %[x0].s[3]\n"
+                "ldr	q5, [%[a_ptr], #384]\n"
+                "fmla	%[r1].4s, v6.4s, %[x0].s[3]\n"
+                "ldr	q6, [%[a_ptr], #400]\n"
+                "fmla	%[r2].4s, v7.4s, %[x0].s[3]\n"
+                "ldr	q7, [%[a_ptr], #416]\n"
+                "fmla	%[r3].4s, v8.4s, %[x0].s[3]\n"
                 ASM_PREFETCH("[%[a_ptr], #2048]")
-                "ldr    q8, [%[a_ptr], #432]\n"
-                "fmla    %[r4].4s, v9.4s, %[x0].s[3]\n"
-                "ldr    q9, [%[a_ptr], #448]\n"
-                "fmla    %[r5].4s, v10.4s, %[x0].s[3]\n"
-                "ldr    q10, [%[a_ptr], #464]\n"
-                "fmla    %[r6].4s, v11.4s, %[x0].s[3]\n"
-                "ldr    q11, [%[a_ptr], #480]\n"
-                "fmla    %[r7].4s, v12.4s, %[x0].s[3]\n"
-                "ldr    q12, [%[a_ptr], #496]\n"
+                "ldr	q8, [%[a_ptr], #432]\n"
+                "fmla	%[r4].4s, v9.4s, %[x0].s[3]\n"
+                "ldr	q9, [%[a_ptr], #448]\n"
+                "fmla	%[r5].4s, v10.4s, %[x0].s[3]\n"
+                "ldr	q10, [%[a_ptr], #464]\n"
+                "fmla	%[r6].4s, v11.4s, %[x0].s[3]\n"
+                "ldr	q11, [%[a_ptr], #480]\n"
+                "fmla	%[r7].4s, v12.4s, %[x0].s[3]\n"
+                "ldr	q12, [%[a_ptr], #496]\n"
                 ASM_PREFETCH("[%[a_ptr], #2112]")
 
                 // Unroll 4
-                "fmla    %[r0].4s, v13.4s, %[x0a].s[0]\n"
-                "ldr    %q[x0], [%[x_ptr]]\n"
-                "fmla    %[r1].4s, v14.4s, %[x0a].s[0]\n"
-                "ldr    q14, [%[a_ptr], #512]\n"
-                "fmla    %[r2].4s, v15.4s, %[x0a].s[0]\n"
-                "ldr    q15, [%[a_ptr], #528]\n"
-                "fmla    %[r3].4s, v16.4s, %[x0a].s[0]\n"
+                "fmla	%[r0].4s, v13.4s, %[x0a].s[0]\n"
+                "ldr	%q[x0], [%[x_ptr]]\n"
+                "fmla	%[r1].4s, v14.4s, %[x0a].s[0]\n"
+                "ldr	q14, [%[a_ptr], #512]\n"
+                "fmla	%[r2].4s, v15.4s, %[x0a].s[0]\n"
+                "ldr	q15, [%[a_ptr], #528]\n"
+                "fmla	%[r3].4s, v16.4s, %[x0a].s[0]\n"
                 ASM_PREFETCH("[%[a_ptr], #2176]")
-                "ldr    q16, [%[a_ptr], #544]\n"
-                "fmla    %[r4].4s, v17.4s, %[x0a].s[0]\n"
-                "ldr    q17, [%[a_ptr], #560]\n"
-                "fmla    %[r5].4s, v18.4s, %[x0a].s[0]\n"
-                "ldr    q18, [%[a_ptr], #576]\n"
-                "fmla    %[r6].4s, v19.4s, %[x0a].s[0]\n"
-                "ldr    q19, [%[a_ptr], #592]\n"
-                "fmla    %[r7].4s, v20.4s, %[x0a].s[0]\n"
-                "ldr    q20, [%[a_ptr], #608]\n"
+                "ldr	q16, [%[a_ptr], #544]\n"
+                "fmla	%[r4].4s, v17.4s, %[x0a].s[0]\n"
+                "ldr	q17, [%[a_ptr], #560]\n"
+                "fmla	%[r5].4s, v18.4s, %[x0a].s[0]\n"
+                "ldr	q18, [%[a_ptr], #576]\n"
+                "fmla	%[r6].4s, v19.4s, %[x0a].s[0]\n"
+                "ldr	q19, [%[a_ptr], #592]\n"
+                "fmla	%[r7].4s, v20.4s, %[x0a].s[0]\n"
+                "ldr	q20, [%[a_ptr], #608]\n"
                 ASM_PREFETCH("[%[a_ptr], #2240]")
 
                 // Unroll 5
-                "fmla    %[r0].4s, v21.4s, %[x0a].s[1]\n"
-                "ldr    q21, [%[a_ptr], #624]\n"
-                "fmla    %[r1].4s, v22.4s, %[x0a].s[1]\n"
-                "ldr    q22, [%[a_ptr], #640]\n"
-                "fmla    %[r2].4s, v23.4s, %[x0a].s[1]\n"
-                "ldr    q23, [%[a_ptr], #656]\n"
-                "fmla    %[r3].4s, v2.4s, %[x0a].s[1]\n"
-                "ldr    q2, [%[a_ptr], #672]\n"
+                "fmla	%[r0].4s, v21.4s, %[x0a].s[1]\n"
+                "ldr	q21, [%[a_ptr], #624]\n"
+                "fmla	%[r1].4s, v22.4s, %[x0a].s[1]\n"
+                "ldr	q22, [%[a_ptr], #640]\n"
+                "fmla	%[r2].4s, v23.4s, %[x0a].s[1]\n"
+                "ldr	q23, [%[a_ptr], #656]\n"
+                "fmla	%[r3].4s, v2.4s, %[x0a].s[1]\n"
+                "ldr	q2, [%[a_ptr], #672]\n"
                 ASM_PREFETCH("[%[a_ptr], #2304]")
-                "fmla    %[r4].4s, v3.4s, %[x0a].s[1]\n"
-                "ldr    q3, [%[a_ptr], #688]\n"
-                "fmla    %[r5].4s, v4.4s, %[x0a].s[1]\n"
-                "ldr    q4, [%[a_ptr], #704]\n"
-                "fmla    %[r6].4s, v5.4s, %[x0a].s[1]\n"
-                "ldr    q5, [%[a_ptr], #720]\n"
-                "fmla    %[r7].4s, v6.4s, %[x0a].s[1]\n"
-                "ldr    q6, [%[a_ptr], #736]\n"
+                "fmla	%[r4].4s, v3.4s, %[x0a].s[1]\n"
+                "ldr	q3, [%[a_ptr], #688]\n"
+                "fmla	%[r5].4s, v4.4s, %[x0a].s[1]\n"
+                "ldr	q4, [%[a_ptr], #704]\n"
+                "fmla	%[r6].4s, v5.4s, %[x0a].s[1]\n"
+                "ldr	q5, [%[a_ptr], #720]\n"
+                "fmla	%[r7].4s, v6.4s, %[x0a].s[1]\n"
+                "ldr	q6, [%[a_ptr], #736]\n"
                 ASM_PREFETCH("[%[a_ptr], #2368]")
 
                 // Unroll 6
-                "fmla    %[r0].4s, v7.4s, %[x0a].s[2]\n"
-                "ldr    q7, [%[a_ptr], #752]\n"
-                "fmla    %[r1].4s, v8.4s, %[x0a].s[2]\n"
-                "ldr    q8, [%[a_ptr], #768]\n"
-                "fmla    %[r2].4s, v9.4s, %[x0a].s[2]\n"
-                "ldr    q9, [%[a_ptr], #784]\n"
-                "fmla    %[r3].4s, v10.4s, %[x0a].s[2]\n"
-                "ldr    q10, [%[a_ptr], #800]\n"
+                "fmla	%[r0].4s, v7.4s, %[x0a].s[2]\n"
+                "ldr	q7, [%[a_ptr], #752]\n"
+                "fmla	%[r1].4s, v8.4s, %[x0a].s[2]\n"
+                "ldr	q8, [%[a_ptr], #768]\n"
+                "fmla	%[r2].4s, v9.4s, %[x0a].s[2]\n"
+                "ldr	q9, [%[a_ptr], #784]\n"
+                "fmla	%[r3].4s, v10.4s, %[x0a].s[2]\n"
+                "ldr	q10, [%[a_ptr], #800]\n"
                 ASM_PREFETCH("[%[a_ptr], #2432]")
-                "fmla    %[r4].4s, v11.4s, %[x0a].s[2]\n"
-                "ldr    q11, [%[a_ptr], #816]\n"
-                "fmla    %[r5].4s, v12.4s, %[x0a].s[2]\n"
-                "ldr    q12, [%[a_ptr], #832]\n"
-                "fmla    %[r6].4s, v14.4s, %[x0a].s[2]\n"
-                "ldr    q13, [%[a_ptr], #848]\n"
-                "ldr    q14, [%[a_ptr], #864]\n"
-                "fmla    %[r7].4s, v15.4s, %[x0a].s[2]\n"
-                "ldr    q15, [%[a_ptr], #880]\n"
+                "fmla	%[r4].4s, v11.4s, %[x0a].s[2]\n"
+                "ldr	q11, [%[a_ptr], #816]\n"
+                "fmla	%[r5].4s, v12.4s, %[x0a].s[2]\n"
+                "ldr	q12, [%[a_ptr], #832]\n"
+                "fmla	%[r6].4s, v14.4s, %[x0a].s[2]\n"
+                "ldr	q13, [%[a_ptr], #848]\n"
+                "ldr	q14, [%[a_ptr], #864]\n"
+                "fmla	%[r7].4s, v15.4s, %[x0a].s[2]\n"
+                "ldr	q15, [%[a_ptr], #880]\n"
                 ASM_PREFETCH("[%[a_ptr], #2496]")
 
                 // Unroll 7
-                "fmla    %[r0].4s, v16.4s, %[x0a].s[3]\n"
-                "ldr    q16, [%[a_ptr], #896]\n"
-                "fmla    %[r1].4s, v17.4s, %[x0a].s[3]\n"
-                "ldr    q17, [%[a_ptr], #912]\n"
-                "fmla    %[r2].4s, v18.4s, %[x0a].s[3]\n"
-                "ldr    q18, [%[a_ptr], #928]\n"
-                "fmla    %[r3].4s, v19.4s, %[x0a].s[3]\n"
+                "fmla	%[r0].4s, v16.4s, %[x0a].s[3]\n"
+                "ldr	q16, [%[a_ptr], #896]\n"
+                "fmla	%[r1].4s, v17.4s, %[x0a].s[3]\n"
+                "ldr	q17, [%[a_ptr], #912]\n"
+                "fmla	%[r2].4s, v18.4s, %[x0a].s[3]\n"
+                "ldr	q18, [%[a_ptr], #928]\n"
+                "fmla	%[r3].4s, v19.4s, %[x0a].s[3]\n"
                 ASM_PREFETCH("[%[a_ptr], #2560]")
-                "ldr    q19, [%[a_ptr], #944]\n"
-                "fmla    %[r4].4s, v20.4s, %[x0a].s[3]\n"
-                "ldr    q20, [%[a_ptr], #960]\n"
-                "fmla    %[r5].4s, v21.4s, %[x0a].s[3]\n"
-                "ldr    q21, [%[a_ptr], #976]\n"
-                "add    %[a_ptr], %[a_ptr], #1024\n"
-                "fmla    %[r6].4s, v22.4s, %[x0a].s[3]\n"
-                "ldr    q22, [%[a_ptr], #-32]\n"
-                "fmla    %[r7].4s, v23.4s, %[x0a].s[3]\n"
-                "ldr    q23, [%[a_ptr], #-16]\n"
+                "ldr	q19, [%[a_ptr], #944]\n"
+                "fmla	%[r4].4s, v20.4s, %[x0a].s[3]\n"
+                "ldr	q20, [%[a_ptr], #960]\n"
+                "fmla	%[r5].4s, v21.4s, %[x0a].s[3]\n"
+                "ldr	q21, [%[a_ptr], #976]\n"
+                "add	%[a_ptr], %[a_ptr], #1024\n"
+                "fmla	%[r6].4s, v22.4s, %[x0a].s[3]\n"
+                "ldr	q22, [%[a_ptr], #-32]\n"
+                "fmla	%[r7].4s, v23.4s, %[x0a].s[3]\n"
+                "ldr	q23, [%[a_ptr], #-16]\n"
                 ASM_PREFETCH("[%[a_ptr], #1600]")
-                "bne    1b\n"
+                "bne	1b\n"
 
                 // Detached final iteration
                 "2:\n"
 
                 // Unroll 0
-                "fmla    %[r0].4s, v2.4s, %[x0].s[0]\n"
-                "ldr    %q[x0a], [%[x_ptr], #16]\n"
-                "fmla    %[r1].4s, v3.4s, %[x0].s[0]\n"
-                "ldr    q3, [%[a_ptr], #0]\n"
-                "subs    %w[k], %w[k], #1\n"
-                "fmla    %[r2].4s, v4.4s, %[x0].s[0]\n"
-                "ldr    q4, [%[a_ptr], #16]\n"
-                "fmla    %[r3].4s, v5.4s, %[x0].s[0]\n"
-                "ldr    q5, [%[a_ptr], #32]\n"
-                "add    %[x_ptr], %[x_ptr], #32\n"
-                "fmla    %[r4].4s, v6.4s, %[x0].s[0]\n"
-                "ldr    q6, [%[a_ptr], #48]\n"
-                "fmla    %[r5].4s, v7.4s, %[x0].s[0]\n"
-                "ldr    q7, [%[a_ptr], #64]\n"
-                "fmla    %[r6].4s, v8.4s, %[x0].s[0]\n"
-                "ldr    q8, [%[a_ptr], #80]\n"
-                "fmla    %[r7].4s, v9.4s, %[x0].s[0]\n"
-                "ldr    q9, [%[a_ptr], #96]\n"
+                "fmla	%[r0].4s, v2.4s, %[x0].s[0]\n"
+                "ldr	%q[x0a], [%[x_ptr], #16]\n"
+                "fmla	%[r1].4s, v3.4s, %[x0].s[0]\n"
+                "ldr	q3, [%[a_ptr], #0]\n"
+                "subs	%w[k], %w[k], #1\n"
+                "fmla	%[r2].4s, v4.4s, %[x0].s[0]\n"
+                "ldr	q4, [%[a_ptr], #16]\n"
+                "fmla	%[r3].4s, v5.4s, %[x0].s[0]\n"
+                "ldr	q5, [%[a_ptr], #32]\n"
+                "add	%[x_ptr], %[x_ptr], #32\n"
+                "fmla	%[r4].4s, v6.4s, %[x0].s[0]\n"
+                "ldr	q6, [%[a_ptr], #48]\n"
+                "fmla	%[r5].4s, v7.4s, %[x0].s[0]\n"
+                "ldr	q7, [%[a_ptr], #64]\n"
+                "fmla	%[r6].4s, v8.4s, %[x0].s[0]\n"
+                "ldr	q8, [%[a_ptr], #80]\n"
+                "fmla	%[r7].4s, v9.4s, %[x0].s[0]\n"
+                "ldr	q9, [%[a_ptr], #96]\n"
 
                 // Unroll 1
-                "fmla    %[r0].4s, v10.4s, %[x0].s[1]\n"
-                "ldr    q10, [%[a_ptr], #112]\n"
-                "fmla    %[r1].4s, v11.4s, %[x0].s[1]\n"
-                "ldr    q11, [%[a_ptr], #128]\n"
-                "fmla    %[r2].4s, v12.4s, %[x0].s[1]\n"
-                "ldr    q12, [%[a_ptr], #144]\n"
-                "fmla    %[r3].4s, v13.4s, %[x0].s[1]\n"
-                "ldr    q13, [%[a_ptr], #160]\n"
-                "fmla    %[r4].4s, v14.4s, %[x0].s[1]\n"
-                "ldr    q14, [%[a_ptr], #176]\n"
-                "fmla    %[r5].4s, v15.4s, %[x0].s[1]\n"
-                "ldr    q15, [%[a_ptr], #192]\n"
-                "fmla    %[r6].4s, v16.4s, %[x0].s[1]\n"
-                "ldr    q16, [%[a_ptr], #208]\n"
-                "fmla    %[r7].4s, v17.4s, %[x0].s[1]\n"
-                "ldr    q17, [%[a_ptr], #224]\n"
+                "fmla	%[r0].4s, v10.4s, %[x0].s[1]\n"
+                "ldr	q10, [%[a_ptr], #112]\n"
+                "fmla	%[r1].4s, v11.4s, %[x0].s[1]\n"
+                "ldr	q11, [%[a_ptr], #128]\n"
+                "fmla	%[r2].4s, v12.4s, %[x0].s[1]\n"
+                "ldr	q12, [%[a_ptr], #144]\n"
+                "fmla	%[r3].4s, v13.4s, %[x0].s[1]\n"
+                "ldr	q13, [%[a_ptr], #160]\n"
+                "fmla	%[r4].4s, v14.4s, %[x0].s[1]\n"
+                "ldr	q14, [%[a_ptr], #176]\n"
+                "fmla	%[r5].4s, v15.4s, %[x0].s[1]\n"
+                "ldr	q15, [%[a_ptr], #192]\n"
+                "fmla	%[r6].4s, v16.4s, %[x0].s[1]\n"
+                "ldr	q16, [%[a_ptr], #208]\n"
+                "fmla	%[r7].4s, v17.4s, %[x0].s[1]\n"
+                "ldr	q17, [%[a_ptr], #224]\n"
 
                 // Unroll 2
-                "fmla    %[r0].4s, v18.4s, %[x0].s[2]\n"
-                "ldr    q18, [%[a_ptr], #240]\n"
-                "fmla    %[r1].4s, v19.4s, %[x0].s[2]\n"
-                "ldr    q19, [%[a_ptr], #256]\n"
-                "fmla    %[r2].4s, v20.4s, %[x0].s[2]\n"
-                "ldr    q20, [%[a_ptr], #272]\n"
-                "fmla    %[r3].4s, v21.4s, %[x0].s[2]\n"
-                "ldr    q21, [%[a_ptr], #288]\n"
-                "fmla    %[r4].4s, v22.4s, %[x0].s[2]\n"
-                "ldr    q22, [%[a_ptr], #304]\n"
-                "fmla    %[r5].4s, v23.4s, %[x0].s[2]\n"
-                "ldr    q23, [%[a_ptr], #320]\n"
-                "fmla    %[r6].4s, v3.4s, %[x0].s[2]\n"
-                "ldr    q2, [%[a_ptr], #336]\n"
-                "ldr    q3, [%[a_ptr], #352]\n"
-                "fmla    %[r7].4s, v4.4s, %[x0].s[2]\n"
-                "ldr    q4, [%[a_ptr], #368]\n"
+                "fmla	%[r0].4s, v18.4s, %[x0].s[2]\n"
+                "ldr	q18, [%[a_ptr], #240]\n"
+                "fmla	%[r1].4s, v19.4s, %[x0].s[2]\n"
+                "ldr	q19, [%[a_ptr], #256]\n"
+                "fmla	%[r2].4s, v20.4s, %[x0].s[2]\n"
+                "ldr	q20, [%[a_ptr], #272]\n"
+                "fmla	%[r3].4s, v21.4s, %[x0].s[2]\n"
+                "ldr	q21, [%[a_ptr], #288]\n"
+                "fmla	%[r4].4s, v22.4s, %[x0].s[2]\n"
+                "ldr	q22, [%[a_ptr], #304]\n"
+                "fmla	%[r5].4s, v23.4s, %[x0].s[2]\n"
+                "ldr	q23, [%[a_ptr], #320]\n"
+                "fmla	%[r6].4s, v3.4s, %[x0].s[2]\n"
+                "ldr	q2, [%[a_ptr], #336]\n"
+                "ldr	q3, [%[a_ptr], #352]\n"
+                "fmla	%[r7].4s, v4.4s, %[x0].s[2]\n"
+                "ldr	q4, [%[a_ptr], #368]\n"
 
                 // Unroll 3
-                "fmla    %[r0].4s, v5.4s, %[x0].s[3]\n"
-                "ldr    q5, [%[a_ptr], #384]\n"
-                "fmla    %[r1].4s, v6.4s, %[x0].s[3]\n"
-                "ldr    q6, [%[a_ptr], #400]\n"
-                "fmla    %[r2].4s, v7.4s, %[x0].s[3]\n"
-                "ldr    q7, [%[a_ptr], #416]\n"
-                "fmla    %[r3].4s, v8.4s, %[x0].s[3]\n"
-                "ldr    q8, [%[a_ptr], #432]\n"
-                "fmla    %[r4].4s, v9.4s, %[x0].s[3]\n"
-                "ldr    q9, [%[a_ptr], #448]\n"
-                "fmla    %[r5].4s, v10.4s, %[x0].s[3]\n"
-                "ldr    q10, [%[a_ptr], #464]\n"
-                "fmla    %[r6].4s, v11.4s, %[x0].s[3]\n"
-                "ldr    q11, [%[a_ptr], #480]\n"
-                "fmla    %[r7].4s, v12.4s, %[x0].s[3]\n"
-                "ldr    q12, [%[a_ptr], #496]\n"
+                "fmla	%[r0].4s, v5.4s, %[x0].s[3]\n"
+                "ldr	q5, [%[a_ptr], #384]\n"
+                "fmla	%[r1].4s, v6.4s, %[x0].s[3]\n"
+                "ldr	q6, [%[a_ptr], #400]\n"
+                "fmla	%[r2].4s, v7.4s, %[x0].s[3]\n"
+                "ldr	q7, [%[a_ptr], #416]\n"
+                "fmla	%[r3].4s, v8.4s, %[x0].s[3]\n"
+                "ldr	q8, [%[a_ptr], #432]\n"
+                "fmla	%[r4].4s, v9.4s, %[x0].s[3]\n"
+                "ldr	q9, [%[a_ptr], #448]\n"
+                "fmla	%[r5].4s, v10.4s, %[x0].s[3]\n"
+                "ldr	q10, [%[a_ptr], #464]\n"
+                "fmla	%[r6].4s, v11.4s, %[x0].s[3]\n"
+                "ldr	q11, [%[a_ptr], #480]\n"
+                "fmla	%[r7].4s, v12.4s, %[x0].s[3]\n"
+                "ldr	q12, [%[a_ptr], #496]\n"
 
                 // Unroll 4
-                "fmla    %[r0].4s, v13.4s, %[x0a].s[0]\n"
-                "fmla    %[r1].4s, v14.4s, %[x0a].s[0]\n"
-                "ldr    q14, [%[a_ptr], #512]\n"
-                "fmla    %[r2].4s, v15.4s, %[x0a].s[0]\n"
-                "ldr    q15, [%[a_ptr], #528]\n"
-                "fmla    %[r3].4s, v16.4s, %[x0a].s[0]\n"
-                "ldr    q16, [%[a_ptr], #544]\n"
-                "fmla    %[r4].4s, v17.4s, %[x0a].s[0]\n"
-                "ldr    q17, [%[a_ptr], #560]\n"
-                "fmla    %[r5].4s, v18.4s, %[x0a].s[0]\n"
-                "ldr    q18, [%[a_ptr], #576]\n"
-                "fmla    %[r6].4s, v19.4s, %[x0a].s[0]\n"
-                "ldr    q19, [%[a_ptr], #592]\n"
-                "fmla    %[r7].4s, v20.4s, %[x0a].s[0]\n"
-                "ldr    q20, [%[a_ptr], #608]\n"
+                "fmla	%[r0].4s, v13.4s, %[x0a].s[0]\n"
+                "fmla	%[r1].4s, v14.4s, %[x0a].s[0]\n"
+                "ldr	q14, [%[a_ptr], #512]\n"
+                "fmla	%[r2].4s, v15.4s, %[x0a].s[0]\n"
+                "ldr	q15, [%[a_ptr], #528]\n"
+                "fmla	%[r3].4s, v16.4s, %[x0a].s[0]\n"
+                "ldr	q16, [%[a_ptr], #544]\n"
+                "fmla	%[r4].4s, v17.4s, %[x0a].s[0]\n"
+                "ldr	q17, [%[a_ptr], #560]\n"
+                "fmla	%[r5].4s, v18.4s, %[x0a].s[0]\n"
+                "ldr	q18, [%[a_ptr], #576]\n"
+                "fmla	%[r6].4s, v19.4s, %[x0a].s[0]\n"
+                "ldr	q19, [%[a_ptr], #592]\n"
+                "fmla	%[r7].4s, v20.4s, %[x0a].s[0]\n"
+                "ldr	q20, [%[a_ptr], #608]\n"
 
                 // Unroll 5
-                "fmla    %[r0].4s, v21.4s, %[x0a].s[1]\n"
-                "ldr    q21, [%[a_ptr], #624]\n"
-                "fmla    %[r1].4s, v22.4s, %[x0a].s[1]\n"
-                "ldr    q22, [%[a_ptr], #640]\n"
-                "fmla    %[r2].4s, v23.4s, %[x0a].s[1]\n"
-                "ldr    q23, [%[a_ptr], #656]\n"
-                "fmla    %[r3].4s, v2.4s, %[x0a].s[1]\n"
-                "add    %[a_ptr], %[a_ptr], #672\n"
-                "fmla    %[r4].4s, v3.4s, %[x0a].s[1]\n"
-                "fmla    %[r5].4s, v4.4s, %[x0a].s[1]\n"
-                "fmla    %[r6].4s, v5.4s, %[x0a].s[1]\n"
-                "fmla    %[r7].4s, v6.4s, %[x0a].s[1]\n"
+                "fmla	%[r0].4s, v21.4s, %[x0a].s[1]\n"
+                "ldr	q21, [%[a_ptr], #624]\n"
+                "fmla	%[r1].4s, v22.4s, %[x0a].s[1]\n"
+                "ldr	q22, [%[a_ptr], #640]\n"
+                "fmla	%[r2].4s, v23.4s, %[x0a].s[1]\n"
+                "ldr	q23, [%[a_ptr], #656]\n"
+                "fmla	%[r3].4s, v2.4s, %[x0a].s[1]\n"
+                "add	%[a_ptr], %[a_ptr], #672\n"
+                "fmla	%[r4].4s, v3.4s, %[x0a].s[1]\n"
+                "fmla	%[r5].4s, v4.4s, %[x0a].s[1]\n"
+                "fmla	%[r6].4s, v5.4s, %[x0a].s[1]\n"
+                "fmla	%[r7].4s, v6.4s, %[x0a].s[1]\n"
 
                 // Unroll 6
-                "fmla    %[r0].4s, v7.4s, %[x0a].s[2]\n"
-                "fmla    %[r1].4s, v8.4s, %[x0a].s[2]\n"
-                "fmla    %[r2].4s, v9.4s, %[x0a].s[2]\n"
-                "fmla    %[r3].4s, v10.4s, %[x0a].s[2]\n"
-                "fmla    %[r4].4s, v11.4s, %[x0a].s[2]\n"
-                "fmla    %[r5].4s, v12.4s, %[x0a].s[2]\n"
-                "fmla    %[r6].4s, v14.4s, %[x0a].s[2]\n"
-                "fmla    %[r7].4s, v15.4s, %[x0a].s[2]\n"
+                "fmla	%[r0].4s, v7.4s, %[x0a].s[2]\n"
+                "fmla	%[r1].4s, v8.4s, %[x0a].s[2]\n"
+                "fmla	%[r2].4s, v9.4s, %[x0a].s[2]\n"
+                "fmla	%[r3].4s, v10.4s, %[x0a].s[2]\n"
+                "fmla	%[r4].4s, v11.4s, %[x0a].s[2]\n"
+                "fmla	%[r5].4s, v12.4s, %[x0a].s[2]\n"
+                "fmla	%[r6].4s, v14.4s, %[x0a].s[2]\n"
+                "fmla	%[r7].4s, v15.4s, %[x0a].s[2]\n"
 
                 // Unroll 7
-                "fmla    %[r0].4s, v16.4s, %[x0a].s[3]\n"
-                "fmla    %[r1].4s, v17.4s, %[x0a].s[3]\n"
-                "fmla    %[r2].4s, v18.4s, %[x0a].s[3]\n"
-                "fmla    %[r3].4s, v19.4s, %[x0a].s[3]\n"
-                "fmla    %[r4].4s, v20.4s, %[x0a].s[3]\n"
-                "fmla    %[r5].4s, v21.4s, %[x0a].s[3]\n"
-                "fmla    %[r6].4s, v22.4s, %[x0a].s[3]\n"
-                "fmla    %[r7].4s, v23.4s, %[x0a].s[3]\n"
+                "fmla	%[r0].4s, v16.4s, %[x0a].s[3]\n"
+                "fmla	%[r1].4s, v17.4s, %[x0a].s[3]\n"
+                "fmla	%[r2].4s, v18.4s, %[x0a].s[3]\n"
+                "fmla	%[r3].4s, v19.4s, %[x0a].s[3]\n"
+                "fmla	%[r4].4s, v20.4s, %[x0a].s[3]\n"
+                "fmla	%[r5].4s, v21.4s, %[x0a].s[3]\n"
+                "fmla	%[r6].4s, v22.4s, %[x0a].s[3]\n"
+                "fmla	%[r7].4s, v23.4s, %[x0a].s[3]\n"
             :
               [a_ptr] "+r" (a_ptr), [x_ptr] "+r" (x_ptr),
               [x0] "+w" (x0), [x0a] "+w" (x0a), [k] "+r" (k),
@@ -532,53 +532,53 @@ void a64_sgemv_pretransposed(const float *A, int lda, const float *X, float *Y,
             int l=(M%8)-1;
 
             __asm __volatile (
-                "ldr    q2, [%[a_ptr], #0]\n"
-                "ldr    q3, [%[a_ptr], #16]\n"
-                "ldr    q4, [%[a_ptr], #32]\n"
-                "ldr    q5, [%[a_ptr], #48]\n"
-                "ldr    q6, [%[a_ptr], #64]\n"
-                "ldr    q7, [%[a_ptr], #80]\n"
-                "ldr    q8, [%[a_ptr], #96]\n"
-                "ldr    q9, [%[a_ptr], #112]\n"
-                "ldr    %s[x0], [%[x_ptr]]\n"
-                "add    %[a_ptr], %[a_ptr], #128\n"
-                "add    %[x_ptr], %[x_ptr], #4\n"
-
-                "cbz    %w[l], 2f\n"
+                "ldr	q2, [%[a_ptr], #0]\n"
+                "ldr	q3, [%[a_ptr], #16]\n"
+                "ldr	q4, [%[a_ptr], #32]\n"
+                "ldr	q5, [%[a_ptr], #48]\n"
+                "ldr	q6, [%[a_ptr], #64]\n"
+                "ldr	q7, [%[a_ptr], #80]\n"
+                "ldr	q8, [%[a_ptr], #96]\n"
+                "ldr	q9, [%[a_ptr], #112]\n"
+                "ldr	%s[x0], [%[x_ptr]]\n"
+                "add	%[a_ptr], %[a_ptr], #128\n"
+                "add	%[x_ptr], %[x_ptr], #4\n"
+
+                "cbz	%w[l], 2f\n"
 
                 "1:\n"
-                "fmla    %[r0].4s, v2.4s, %[x0].s[0]\n"
-                "ldr    q2, [%[a_ptr], #0]\n"
-                "subs    %w[l], %w[l], #1\n"
-                "fmla    %[r1].4s, v3.4s, %[x0].s[0]\n"
-                "ldr    q3, [%[a_ptr], #16]\n"
-                "fmla    %[r2].4s, v4.4s, %[x0].s[0]\n"
-                "ldr    q4, [%[a_ptr], #32]\n"
-                "fmla    %[r3].4s, v5.4s, %[x0].s[0]\n"
-                "ldr    q5, [%[a_ptr], #48]\n"
-                "fmla    %[r4].4s, v6.4s, %[x0].s[0]\n"
-                "ldr    q6, [%[a_ptr], #64]\n"
-                "fmla    %[r5].4s, v7.4s, %[x0].s[0]\n"
-                "ldr    q7, [%[a_ptr], #80]\n"
-                "fmla    %[r6].4s, v8.4s, %[x0].s[0]\n"
-                "ldr    q8, [%[a_ptr], #96]\n"
-                "fmla    %[r7].4s, v9.4s, %[x0].s[0]\n"
-                "ldr    q9, [%[a_ptr], #112]\n"
-                "ldr    %s[x0], [%[x_ptr]]\n"
-                "add    %[a_ptr], %[a_ptr], #128\n"
-                "add    %[x_ptr], %[x_ptr], #4\n"
-                "bne    1b\n"
+                "fmla	%[r0].4s, v2.4s, %[x0].s[0]\n"
+                "ldr	q2, [%[a_ptr], #0]\n"
+                "subs	%w[l], %w[l], #1\n"
+                "fmla	%[r1].4s, v3.4s, %[x0].s[0]\n"
+                "ldr	q3, [%[a_ptr], #16]\n"
+                "fmla	%[r2].4s, v4.4s, %[x0].s[0]\n"
+                "ldr	q4, [%[a_ptr], #32]\n"
+                "fmla	%[r3].4s, v5.4s, %[x0].s[0]\n"
+                "ldr	q5, [%[a_ptr], #48]\n"
+                "fmla	%[r4].4s, v6.4s, %[x0].s[0]\n"
+                "ldr	q6, [%[a_ptr], #64]\n"
+                "fmla	%[r5].4s, v7.4s, %[x0].s[0]\n"
+                "ldr	q7, [%[a_ptr], #80]\n"
+                "fmla	%[r6].4s, v8.4s, %[x0].s[0]\n"
+                "ldr	q8, [%[a_ptr], #96]\n"
+                "fmla	%[r7].4s, v9.4s, %[x0].s[0]\n"
+                "ldr	q9, [%[a_ptr], #112]\n"
+                "ldr	%s[x0], [%[x_ptr]]\n"
+                "add	%[a_ptr], %[a_ptr], #128\n"
+                "add	%[x_ptr], %[x_ptr], #4\n"
+                "bne	1b\n"
 
                 "2:\n"
 
-                "fmla    %[r0].4s, v2.4s, %[x0].s[0]\n"
-                "fmla    %[r1].4s, v3.4s, %[x0].s[0]\n"
-                "fmla    %[r2].4s, v4.4s, %[x0].s[0]\n"
-                "fmla    %[r3].4s, v5.4s, %[x0].s[0]\n"
-                "fmla    %[r4].4s, v6.4s, %[x0].s[0]\n"
-                "fmla    %[r5].4s, v7.4s, %[x0].s[0]\n"
-                "fmla    %[r6].4s, v8.4s, %[x0].s[0]\n"
-                "fmla    %[r7].4s, v9.4s, %[x0].s[0]\n"
+                "fmla	%[r0].4s, v2.4s, %[x0].s[0]\n"
+                "fmla	%[r1].4s, v3.4s, %[x0].s[0]\n"
+                "fmla	%[r2].4s, v4.4s, %[x0].s[0]\n"
+                "fmla	%[r3].4s, v5.4s, %[x0].s[0]\n"
+                "fmla	%[r4].4s, v6.4s, %[x0].s[0]\n"
+                "fmla	%[r5].4s, v7.4s, %[x0].s[0]\n"
+                "fmla	%[r6].4s, v8.4s, %[x0].s[0]\n"
+                "fmla	%[r7].4s, v9.4s, %[x0].s[0]\n"
             :
               [a_ptr] "+r" (a_ptr), [x_ptr] "+r" (x_ptr),
               [x0] "+w" (x0), [l] "+r" (l),
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4.hpp
index b8b93bf31f..d072470939 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4.hpp
@@ -72,7 +72,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<operand_type, operand_type, result_type, 6, 4, 1> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 6, 4, 1> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel=a64_smallK_hybrid_fp32_mla_6x4;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4.hpp
index 72f517fe35..94312be08a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4.hpp
@@ -72,7 +72,7 @@ public:
         return true;
     }
 
-    StdTransformsFixed<operand_type, operand_type, result_type, 8, 4, 1> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 4, 1> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel=a64_smallK_hybrid_fp32_mla_8x4;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4.hpp
index 6fdca066b1..d244aecc70 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4.hpp
@@ -73,7 +73,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<operand_type, operand_type, result_type, 6, 4, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 6, 4, 4> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel=a64_smallK_hybrid_s8s32_dot_6x4;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4.hpp
index 27d3e2310c..85583c46cf 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4.hpp
@@ -73,7 +73,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<operand_type, operand_type, result_type, 8, 4, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 4, 4> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel=a64_smallK_hybrid_s8s32_dot_8x4;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4.hpp
index 082ca66e01..c474b9c1f9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4.hpp
@@ -73,7 +73,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<operand_type, operand_type, result_type, 6, 4, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 6, 4, 4> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel=a64_smallK_hybrid_u8u32_dot_6x4;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4.hpp
index 866b97a316..65a2281638 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4.hpp
@@ -73,7 +73,7 @@ public:
         return false;
     }
 
-    StdTransformsFixed<operand_type, operand_type, result_type, 8, 4, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 4, 4> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel=a64_smallK_hybrid_u8u32_dot_8x4;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp
index b102e1dea4..5c9a465817 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp
@@ -64,21 +64,21 @@ void sme2_gemv_bf16fp32_dot_16VL (
     __asm__ __volatile__(
       "ptrue p8.b\n"
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x9, #0x0\n"
       "cntw x28, ALL, MUL #4\n"
-      "mov x27, %x[B_ptr]\n"
-      "add x26, %x[N], x28\n"
-      "mov x25, %x[output_ptr]\n"
-      "sub x26, x26, #0x1\n"
-      "ptrue p1.b\n"
-      "udiv x26, x26, x28\n"
-      ".inst 0x25207811  // ptrue pn9.b\n"
-      "add x22, x26, #0x3\n"
-      "mov x21, #0x1\n"
+      "add x27, %x[N], x28\n"
+      "sub x27, x27, #0x1\n"
+      "udiv x27, x27, x28\n"
+      "add x22, x27, #0x3\n"
       "and x22, x22, #0xfffffffffffffffc\n"
       "mul x22, x22, x28\n"
       "mul x22, x22, %x[K]\n"
+      "mov x9, #0x0\n"
+      "mov x26, %x[B_ptr]\n"
+      "mov x25, %x[output_ptr]\n"
+      "ptrue p1.b\n"
+      ".inst 0x25207811  // ptrue pn9.b\n"
       "lsl x22, x22, #0x1\n"
+      "mov x21, #0x1\n"
       "1:"  // RHS size check loop
       "cmp x22, #0x200000\n"
       "blt 2f\n"
@@ -92,13 +92,13 @@ void sme2_gemv_bf16fp32_dot_16VL (
       "lsl x21, x21, #0x16\n"
       "orr x22, x22, x20\n"
       "orr x22, x22, x21\n"
-      ".inst 0xf8b64b7a  // rprfm pldonce, x22, [x27]\n"
+      ".inst 0xf8b64b5a  // rprfm pldonce, x22, [x26]\n"
       "3:"  // RHS prefetch exit
       "mov x24, %x[bias]\n"
       "4:"  // Column loop
-      "cmp x26, #0x4\n"
+      "cmp x27, #0x4\n"
       "bge 28f\n"
-      "cmp x26, #0x2\n"
+      "cmp x27, #0x2\n"
       "bgt 20f\n"
       "beq 12f\n"
       "mov x23, %x[A_ptr]\n"
@@ -108,8 +108,8 @@ void sme2_gemv_bf16fp32_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 5f\n"
-      ".inst 0xa040c714  // ld1w { z20.s-z23.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xc0042e80  // mova za.d[x9, #0], { z20.d-z23.d }\n"
+      ".inst 0xa040c718  // ld1w { z24.s-z27.s }, pn9.b/Z, [x24]\n"
+      ".inst 0xc0042f00  // mova za.d[x9, #0], { z24.d-z27.d }\n"
       "b 6f\n"
       "5:"  // Width 1: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -118,59 +118,59 @@ void sme2_gemv_bf16fp32_dot_16VL (
       "ble 8f\n"
       "7:"  // Width 1: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      "ld1rqh { z2.h }, p0/Z, [x23]\n"
+      "ld1rqh { z8.h }, p0/Z, [x23]\n"
       "sub x22, x22, #0x8\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
+      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158b298  // bfdot za.s[x9, 0], { z20.h-z23.h }, z8.h[0]\n"
+      "addvl x26, x26, #16\n"
       "cmp x22, #0x8\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc152b398  // bfdot za.s[x9, 0], { z28.h-z31.h }, z2.h[0]\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc152b618  // bfdot za.s[x9, 0], { z16.h-z19.h }, z2.h[1]\n"
-      ".inst 0xc152bb18  // bfdot za.s[x9, 0], { z24.h-z27.h }, z2.h[2]\n"
-      ".inst 0xc152bd98  // bfdot za.s[x9, 0], { z12.h-z15.h }, z2.h[3]\n"
+      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158b498  // bfdot za.s[x9, 0], { z4.h-z7.h }, z8.h[1]\n"
+      "addvl x26, x26, #16\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158bb98  // bfdot za.s[x9, 0], { z28.h-z31.h }, z8.h[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158bf18  // bfdot za.s[x9, 0], { z24.h-z27.h }, z8.h[3]\n"
+      "addvl x26, x26, #16\n"
       "bgt 7b\n"
       "8:"  // Width 1: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
+      "ld1rqh { z11.h }, p0/Z, [x23]\n"
       "subs x22, x22, #0x2\n"
-      "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b198  // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[0]\n"
+      ".inst 0xc15bb398  // bfdot za.s[x9, 0], { z28.h-z31.h }, z11.h[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 9f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b498  // bfdot za.s[x9, 0], { z4.h-z7.h }, z3.h[1]\n"
+      ".inst 0xc15bb598  // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 9f\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b998  // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[2]\n"
+      ".inst 0xc15bba18  // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 9f\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153bd18  // bfdot za.s[x9, 0], { z8.h-z11.h }, z3.h[3]\n"
+      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15bbc18  // bfdot za.s[x9, 0], { z0.h-z3.h }, z11.h[3]\n"
+      "addvl x26, x26, #16\n"
       "9:"  // Width 1: Multiply loop: multiply skip
       "tbz %x[flags], #1, 10f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
-      "ld1rw { z23.s }, p1/Z, [x21]\n"
-      "ld1rw { z22.s }, p1/Z, [x20]\n"
-      ".inst 0xc1b6cae0  // fclamp { z0.s-z3.s }, z23.s, z22.s\n"
-      ".inst 0xa060c320  // st1w { z0.s-z3.s }, p8, [x25]\n"
+      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
+      "ld1rw { z3.s }, p1/Z, [x21]\n"
+      "ld1rw { z29.s }, p1/Z, [x20]\n"
+      ".inst 0xc1bdc868  // fclamp { z8.s-z11.s }, z3.s, z29.s\n"
+      ".inst 0xa060c328  // st1w { z8.s-z11.s }, p8, [x25]\n"
       "addvl x25, x25, #4\n"
       "b 11f\n"
       "10:"  // Width 1: No activation
-      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
-      ".inst 0xa060c320  // st1w { z0.s-z3.s }, p8, [x25]\n"
+      ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c32c  // st1w { z12.s-z15.s }, p8, [x25]\n"
       "addvl x25, x25, #4\n"
       "11:"  // Width 1: Output done
       "b 36f\n"
@@ -182,10 +182,10 @@ void sme2_gemv_bf16fp32_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 13f\n"
-      ".inst 0xa040c71c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xc0042f80  // mova za.d[x9, #0], { z28.d-z31.d }\n"
-      ".inst 0xc0042c81  // mova za.d[x9, #1], { z4.d-z7.d }\n"
+      ".inst 0xa040c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24]\n"
+      ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
+      ".inst 0xa041c714  // ld1w { z20.s-z23.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042e81  // mova za.d[x9, #1], { z20.d-z23.d }\n"
       "b 14f\n"
       "13:"  // Width 2: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -194,80 +194,80 @@ void sme2_gemv_bf16fp32_dot_16VL (
       "ble 16f\n"
       "15:"  // Width 2: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
+      "ld1rqh { z9.h }, p0/Z, [x23]\n"
       "sub x22, x22, #0x8\n"
-      "ld1rqh { z1.h }, p0/Z, [x23]\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc159b198  // bfdot za.s[x9, 0], { z12.h-z15.h }, z9.h[0]\n"
       "cmp x22, #0x8\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc151b398  // bfdot za.s[x9, 0], { z28.h-z31.h }, z1.h[0]\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc151b199  // bfdot za.s[x9, 1], { z12.h-z15.h }, z1.h[0]\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc151b618  // bfdot za.s[x9, 0], { z16.h-z19.h }, z1.h[1]\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc151b719  // bfdot za.s[x9, 1], { z24.h-z27.h }, z1.h[1]\n"
-      ".inst 0xc151bb98  // bfdot za.s[x9, 0], { z28.h-z31.h }, z1.h[2]\n"
-      ".inst 0xc151b999  // bfdot za.s[x9, 1], { z12.h-z15.h }, z1.h[2]\n"
-      ".inst 0xc151bd18  // bfdot za.s[x9, 0], { z8.h-z11.h }, z1.h[3]\n"
-      ".inst 0xc151be99  // bfdot za.s[x9, 1], { z20.h-z23.h }, z1.h[3]\n"
+      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc159b099  // bfdot za.s[x9, 1], { z4.h-z7.h }, z9.h[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc159b598  // bfdot za.s[x9, 0], { z12.h-z15.h }, z9.h[1]\n"
+      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc159b499  // bfdot za.s[x9, 1], { z4.h-z7.h }, z9.h[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc159bb18  // bfdot za.s[x9, 0], { z24.h-z27.h }, z9.h[2]\n"
+      ".inst 0xa041a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc159b819  // bfdot za.s[x9, 1], { z0.h-z3.h }, z9.h[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc159bc18  // bfdot za.s[x9, 0], { z0.h-z3.h }, z9.h[3]\n"
+      ".inst 0xa041a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc159bf99  // bfdot za.s[x9, 1], { z28.h-z31.h }, z9.h[3]\n"
+      "addvl x26, x26, #16\n"
       "bgt 15b\n"
       "16:"  // Width 2: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
+      "ld1rqh { z11.h }, p0/Z, [x23]\n"
       "subs x22, x22, #0x2\n"
-      "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b218  // bfdot za.s[x9, 0], { z16.h-z19.h }, z3.h[0]\n"
-      ".inst 0xc153b399  // bfdot za.s[x9, 1], { z28.h-z31.h }, z3.h[0]\n"
+      ".inst 0xc15bb198  // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[0]\n"
+      ".inst 0xa041a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bb019  // bfdot za.s[x9, 1], { z0.h-z3.h }, z11.h[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 17f\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b698  // bfdot za.s[x9, 0], { z20.h-z23.h }, z3.h[1]\n"
-      ".inst 0xc153b619  // bfdot za.s[x9, 1], { z16.h-z19.h }, z3.h[1]\n"
+      ".inst 0xc15bb718  // bfdot za.s[x9, 0], { z24.h-z27.h }, z11.h[1]\n"
+      ".inst 0xa041a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bb419  // bfdot za.s[x9, 1], { z0.h-z3.h }, z11.h[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 17f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b898  // bfdot za.s[x9, 0], { z4.h-z7.h }, z3.h[2]\n"
-      ".inst 0xc153ba19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z3.h[2]\n"
+      ".inst 0xc15bb998  // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[2]\n"
+      ".inst 0xa041a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bbb99  // bfdot za.s[x9, 1], { z28.h-z31.h }, z11.h[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 17f\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153bf98  // bfdot za.s[x9, 0], { z28.h-z31.h }, z3.h[3]\n"
-      ".inst 0xc153bd99  // bfdot za.s[x9, 1], { z12.h-z15.h }, z3.h[3]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15bbe18  // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[3]\n"
+      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bbe99  // bfdot za.s[x9, 1], { z20.h-z23.h }, z11.h[3]\n"
+      "addvl x26, x26, #16\n"
       "17:"  // Width 2: Multiply loop: multiply skip
       "tbz %x[flags], #1, 18f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c28  // mova { z8.d-z11.d }, za.d[x9, #1]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
-      "ld1rw { z23.s }, p1/Z, [x20]\n"
-      ".inst 0xc1b7ca24  // fclamp { z4.s-z7.s }, z17.s, z23.s\n"
-      ".inst 0xc1b7ca28  // fclamp { z8.s-z11.s }, z17.s, z23.s\n"
-      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c328  // st1w { z8.s-z11.s }, p8, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
+      "ld1rw { z9.s }, p1/Z, [x21]\n"
+      ".inst 0xc0062c24  // mova { z4.d-z7.d }, za.d[x9, #1]\n"
+      "ld1rw { z8.s }, p1/Z, [x20]\n"
+      ".inst 0xc1a8c920  // fclamp { z0.s-z3.s }, z9.s, z8.s\n"
+      ".inst 0xa060c720  // st1w { z0.s-z3.s }, pn9.b, [x25]\n"
+      ".inst 0xc1a8c924  // fclamp { z4.s-z7.s }, z9.s, z8.s\n"
+      ".inst 0xa061c324  // st1w { z4.s-z7.s }, p8, [x25, #0x4, MUL VL]\n"
       "addvl x25, x25, #8\n"
       "b 19f\n"
       "18:"  // Width 2: No activation
-      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c30  // mova { z16.d-z19.d }, za.d[x9, #1]\n"
-      ".inst 0xa060c728  // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c330  // st1w { z16.s-z19.s }, p8, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c10  // mova { z16.d-z19.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c730  // st1w { z16.s-z19.s }, pn9.b, [x25]\n"
+      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
+      ".inst 0xa061c32c  // st1w { z12.s-z15.s }, p8, [x25, #0x4, MUL VL]\n"
       "addvl x25, x25, #8\n"
       "19:"  // Width 2: Output done
       "b 36f\n"
@@ -280,12 +280,12 @@ void sme2_gemv_bf16fp32_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 21f\n"
-      ".inst 0xa040c708  // ld1w { z8.s-z11.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c700  // ld1w { z0.s-z3.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xc0042d00  // mova za.d[x9, #0], { z8.d-z11.d }\n"
-      ".inst 0xc0042c01  // mova za.d[x9, #1], { z0.d-z3.d }\n"
-      ".inst 0xc0042c82  // mova za.d[x9, #2], { z4.d-z7.d }\n"
+      ".inst 0xa040c718  // ld1w { z24.s-z27.s }, pn9.b/Z, [x24]\n"
+      ".inst 0xc0042f00  // mova za.d[x9, #0], { z24.d-z27.d }\n"
+      ".inst 0xa041c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042c81  // mova za.d[x9, #1], { z4.d-z7.d }\n"
+      ".inst 0xa042c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
+      ".inst 0xc0042e02  // mova za.d[x9, #2], { z16.d-z19.d }\n"
       "b 22f\n"
       "21:"  // Width 3: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -294,101 +294,101 @@ void sme2_gemv_bf16fp32_dot_16VL (
       "ble 24f\n"
       "23:"  // Width 3: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
+      "ld1rqh { z15.h }, p0/Z, [x23]\n"
       "sub x22, x22, #0x8\n"
-      "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fb018  // bfdot za.s[x9, 0], { z0.h-z3.h }, z15.h[0]\n"
       "cmp x22, #0x8\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b198  // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[0]\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153b119  // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[0]\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153b09a  // bfdot za.s[x9, 2], { z4.h-z7.h }, z3.h[0]\n"
-      ".inst 0xa042a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153b618  // bfdot za.s[x9, 0], { z16.h-z19.h }, z3.h[1]\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153b699  // bfdot za.s[x9, 1], { z20.h-z23.h }, z3.h[1]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b71a  // bfdot za.s[x9, 2], { z24.h-z27.h }, z3.h[1]\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153b998  // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[2]\n"
-      ".inst 0xa042a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153bb99  // bfdot za.s[x9, 1], { z28.h-z31.h }, z3.h[2]\n"
-      ".inst 0xc153b91a  // bfdot za.s[x9, 2], { z8.h-z11.h }, z3.h[2]\n"
-      ".inst 0xc153bc98  // bfdot za.s[x9, 0], { z4.h-z7.h }, z3.h[3]\n"
-      ".inst 0xc153be99  // bfdot za.s[x9, 1], { z20.h-z23.h }, z3.h[3]\n"
-      ".inst 0xc153bf1a  // bfdot za.s[x9, 2], { z24.h-z27.h }, z3.h[3]\n"
+      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb099  // bfdot za.s[x9, 1], { z4.h-z7.h }, z15.h[0]\n"
+      ".inst 0xa042a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fb01a  // bfdot za.s[x9, 2], { z0.h-z3.h }, z15.h[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fb698  // bfdot za.s[x9, 0], { z20.h-z23.h }, z15.h[1]\n"
+      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb699  // bfdot za.s[x9, 1], { z20.h-z23.h }, z15.h[1]\n"
+      ".inst 0xa042a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fb51a  // bfdot za.s[x9, 2], { z8.h-z11.h }, z15.h[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fbb18  // bfdot za.s[x9, 0], { z24.h-z27.h }, z15.h[2]\n"
+      ".inst 0xa041a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb919  // bfdot za.s[x9, 1], { z8.h-z11.h }, z15.h[2]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fba1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z15.h[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fbe18  // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[3]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fbe19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[3]\n"
+      ".inst 0xa042a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fbd1a  // bfdot za.s[x9, 2], { z8.h-z11.h }, z15.h[3]\n"
+      "addvl x26, x26, #16\n"
       "bgt 23b\n"
       "24:"  // Width 3: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
+      "ld1rqh { z11.h }, p0/Z, [x23]\n"
       "subs x22, x22, #0x2\n"
-      "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b298  // bfdot za.s[x9, 0], { z20.h-z23.h }, z3.h[0]\n"
-      ".inst 0xc153b199  // bfdot za.s[x9, 1], { z12.h-z15.h }, z3.h[0]\n"
-      ".inst 0xc153b09a  // bfdot za.s[x9, 2], { z4.h-z7.h }, z3.h[0]\n"
+      ".inst 0xc15bb398  // bfdot za.s[x9, 0], { z28.h-z31.h }, z11.h[0]\n"
+      ".inst 0xa041a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bb019  // bfdot za.s[x9, 1], { z0.h-z3.h }, z11.h[0]\n"
+      ".inst 0xa042a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bb29a  // bfdot za.s[x9, 2], { z20.h-z23.h }, z11.h[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 25f\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b698  // bfdot za.s[x9, 0], { z20.h-z23.h }, z3.h[1]\n"
-      ".inst 0xc153b519  // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[1]\n"
-      ".inst 0xc153b61a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z3.h[1]\n"
+      ".inst 0xc15bb598  // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[1]\n"
+      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bb499  // bfdot za.s[x9, 1], { z4.h-z7.h }, z11.h[1]\n"
+      ".inst 0xa042a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bb79a  // bfdot za.s[x9, 2], { z28.h-z31.h }, z11.h[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 25f\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153bb98  // bfdot za.s[x9, 0], { z28.h-z31.h }, z3.h[2]\n"
-      ".inst 0xc153bb19  // bfdot za.s[x9, 1], { z24.h-z27.h }, z3.h[2]\n"
-      ".inst 0xc153b99a  // bfdot za.s[x9, 2], { z12.h-z15.h }, z3.h[2]\n"
+      ".inst 0xc15bb898  // bfdot za.s[x9, 0], { z4.h-z7.h }, z11.h[2]\n"
+      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bba99  // bfdot za.s[x9, 1], { z20.h-z23.h }, z11.h[2]\n"
+      ".inst 0xa042a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bb99a  // bfdot za.s[x9, 2], { z12.h-z15.h }, z11.h[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 25f\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153bd18  // bfdot za.s[x9, 0], { z8.h-z11.h }, z3.h[3]\n"
-      ".inst 0xc153bf99  // bfdot za.s[x9, 1], { z28.h-z31.h }, z3.h[3]\n"
-      ".inst 0xc153bd9a  // bfdot za.s[x9, 2], { z12.h-z15.h }, z3.h[3]\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15bbd98  // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[3]\n"
+      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bbe99  // bfdot za.s[x9, 1], { z20.h-z23.h }, z11.h[3]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bbe1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z11.h[3]\n"
+      "addvl x26, x26, #16\n"
       "25:"  // Width 3: Multiply loop: multiply skip
       "tbz %x[flags], #1, 26f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      "ld1rw { z21.s }, p1/Z, [x21]\n"
-      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
-      "ld1rw { z20.s }, p1/Z, [x20]\n"
-      ".inst 0xc1b4caa8  // fclamp { z8.s-z11.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xa060c728  // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c72c  // st1w { z12.s-z15.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c330  // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n"
+      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
+      "ld1rw { z17.s }, p1/Z, [x21]\n"
+      ".inst 0xc0062c28  // mova { z8.d-z11.d }, za.d[x9, #1]\n"
+      "ld1rw { z16.s }, p1/Z, [x20]\n"
+      ".inst 0xc1b0ca24  // fclamp { z4.s-z7.s }, z17.s, z16.s\n"
+      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
+      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
+      ".inst 0xc1b0ca28  // fclamp { z8.s-z11.s }, z17.s, z16.s\n"
+      ".inst 0xa061c728  // st1w { z8.s-z11.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc1b0ca2c  // fclamp { z12.s-z15.s }, z17.s, z16.s\n"
+      ".inst 0xa062c32c  // st1w { z12.s-z15.s }, p8, [x25, #0x8, MUL VL]\n"
       "addvl x25, x25, #12\n"
       "b 27f\n"
       "26:"  // Width 3: No activation
-      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
+      ".inst 0xc0062c14  // mova { z20.d-z23.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c734  // st1w { z20.s-z23.s }, pn9.b, [x25]\n"
       ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c5c  // mova { z28.d-z31.d }, za.d[x9, #2]\n"
-      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
       ".inst 0xa061c72c  // st1w { z12.s-z15.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c33c  // st1w { z28.s-z31.s }, p8, [x25, #0x8, MUL VL]\n"
+      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
+      ".inst 0xa062c32c  // st1w { z12.s-z15.s }, p8, [x25, #0x8, MUL VL]\n"
       "addvl x25, x25, #12\n"
       "27:"  // Width 3: Output done
       "b 36f\n"
@@ -402,14 +402,14 @@ void sme2_gemv_bf16fp32_dot_16VL (
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 29f\n"
       ".inst 0xa040c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c70c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042c71c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xa043c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n"
       ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
-      "addvl x24, x24, #16\n"
-      ".inst 0xc0042d81  // mova za.d[x9, #1], { z12.d-z15.d }\n"
-      ".inst 0xc0042f82  // mova za.d[x9, #2], { z28.d-z31.d }\n"
+      ".inst 0xa041c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042e01  // mova za.d[x9, #1], { z16.d-z19.d }\n"
+      ".inst 0xa042c70c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
+      ".inst 0xc0042d82  // mova za.d[x9, #2], { z12.d-z15.d }\n"
+      ".inst 0xa043c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n"
       ".inst 0xc0042e03  // mova za.d[x9, #3], { z16.d-z19.d }\n"
+      "addvl x24, x24, #16\n"
       "b 30f\n"
       "29:"  // Width 4: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -418,125 +418,125 @@ void sme2_gemv_bf16fp32_dot_16VL (
       "ble 32f\n"
       "31:"  // Width 4: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
+      "ld1rqh { z8.h }, p0/Z, [x23]\n"
       "sub x22, x22, #0x8\n"
-      "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158b218  // bfdot za.s[x9, 0], { z16.h-z19.h }, z8.h[0]\n"
       "cmp x22, #0x8\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153b198  // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b299  // bfdot za.s[x9, 1], { z20.h-z23.h }, z3.h[0]\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153b21a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z3.h[0]\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153b11b  // bfdot za.s[x9, 3], { z8.h-z11.h }, z3.h[0]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153b718  // bfdot za.s[x9, 0], { z24.h-z27.h }, z3.h[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b599  // bfdot za.s[x9, 1], { z12.h-z15.h }, z3.h[1]\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153b51a  // bfdot za.s[x9, 2], { z8.h-z11.h }, z3.h[1]\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153b49b  // bfdot za.s[x9, 3], { z4.h-z7.h }, z3.h[1]\n"
-      ".inst 0xa042a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153bb18  // bfdot za.s[x9, 0], { z24.h-z27.h }, z3.h[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b919  // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[2]\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153ba1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z3.h[2]\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153b89b  // bfdot za.s[x9, 3], { z4.h-z7.h }, z3.h[2]\n"
-      ".inst 0xa042a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153bd18  // bfdot za.s[x9, 0], { z8.h-z11.h }, z3.h[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153bf99  // bfdot za.s[x9, 1], { z28.h-z31.h }, z3.h[3]\n"
-      ".inst 0xc153bd9a  // bfdot za.s[x9, 2], { z12.h-z15.h }, z3.h[3]\n"
-      ".inst 0xc153be9b  // bfdot za.s[x9, 3], { z20.h-z23.h }, z3.h[3]\n"
+      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc158b199  // bfdot za.s[x9, 1], { z12.h-z15.h }, z8.h[0]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc158b21a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z8.h[0]\n"
+      ".inst 0xa043a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc158b19b  // bfdot za.s[x9, 3], { z12.h-z15.h }, z8.h[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158b598  // bfdot za.s[x9, 0], { z12.h-z15.h }, z8.h[1]\n"
+      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc158b699  // bfdot za.s[x9, 1], { z20.h-z23.h }, z8.h[1]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc158b61a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z8.h[1]\n"
+      ".inst 0xa043a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc158b69b  // bfdot za.s[x9, 3], { z20.h-z23.h }, z8.h[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158b898  // bfdot za.s[x9, 0], { z4.h-z7.h }, z8.h[2]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc158ba19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z8.h[2]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc158ba1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z8.h[2]\n"
+      ".inst 0xa043a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc158b81b  // bfdot za.s[x9, 3], { z0.h-z3.h }, z8.h[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158be98  // bfdot za.s[x9, 0], { z20.h-z23.h }, z8.h[3]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc158be19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z8.h[3]\n"
+      ".inst 0xa042a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc158bc9a  // bfdot za.s[x9, 2], { z4.h-z7.h }, z8.h[3]\n"
+      ".inst 0xa043a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc158be9b  // bfdot za.s[x9, 3], { z20.h-z23.h }, z8.h[3]\n"
+      "addvl x26, x26, #16\n"
       "bgt 31b\n"
       "32:"  // Width 4: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
+      "ld1rqh { z11.h }, p0/Z, [x23]\n"
       "subs x22, x22, #0x2\n"
-      "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153b218  // bfdot za.s[x9, 0], { z16.h-z19.h }, z3.h[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b199  // bfdot za.s[x9, 1], { z12.h-z15.h }, z3.h[0]\n"
-      ".inst 0xc153b39a  // bfdot za.s[x9, 2], { z28.h-z31.h }, z3.h[0]\n"
-      ".inst 0xc153b29b  // bfdot za.s[x9, 3], { z20.h-z23.h }, z3.h[0]\n"
+      ".inst 0xc15bb218  // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[0]\n"
+      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bb299  // bfdot za.s[x9, 1], { z20.h-z23.h }, z11.h[0]\n"
+      ".inst 0xa042a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bb39a  // bfdot za.s[x9, 2], { z28.h-z31.h }, z11.h[0]\n"
+      ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15bb21b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z11.h[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 33f\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153b598  // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b499  // bfdot za.s[x9, 1], { z4.h-z7.h }, z3.h[1]\n"
-      ".inst 0xc153b71a  // bfdot za.s[x9, 2], { z24.h-z27.h }, z3.h[1]\n"
-      ".inst 0xc153b69b  // bfdot za.s[x9, 3], { z20.h-z23.h }, z3.h[1]\n"
+      ".inst 0xc15bb418  // bfdot za.s[x9, 0], { z0.h-z3.h }, z11.h[1]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bb619  // bfdot za.s[x9, 1], { z16.h-z19.h }, z11.h[1]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bb61a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z11.h[1]\n"
+      ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15bb61b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z11.h[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 33f\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153b998  // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153b919  // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[2]\n"
-      ".inst 0xc153ba9a  // bfdot za.s[x9, 2], { z20.h-z23.h }, z3.h[2]\n"
-      ".inst 0xc153ba1b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z3.h[2]\n"
+      ".inst 0xc15bba18  // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[2]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bba19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z11.h[2]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bba1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z11.h[2]\n"
+      ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15bba1b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z11.h[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 33f\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153bd98  // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153bd19  // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[3]\n"
-      ".inst 0xc153be9a  // bfdot za.s[x9, 2], { z20.h-z23.h }, z3.h[3]\n"
-      ".inst 0xc153be1b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z3.h[3]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15bbe18  // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[3]\n"
+      ".inst 0xa041a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bbf19  // bfdot za.s[x9, 1], { z24.h-z27.h }, z11.h[3]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bbe1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z11.h[3]\n"
+      ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15bbe1b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z11.h[3]\n"
+      "addvl x26, x26, #16\n"
       "33:"  // Width 4: Multiply loop: multiply skip
       "tbz %x[flags], #1, 34f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
+      ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
       "ld1rw { z21.s }, p1/Z, [x21]\n"
-      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
+      ".inst 0xc0062c38  // mova { z24.d-z27.d }, za.d[x9, #1]\n"
       "ld1rw { z20.s }, p1/Z, [x20]\n"
+      ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
+      ".inst 0xc0062c40  // mova { z0.d-z3.d }, za.d[x9, #2]\n"
+      ".inst 0xa060c72c  // st1w { z12.s-z15.s }, pn9.b, [x25]\n"
+      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
       ".inst 0xc0062c70  // mova { z16.d-z19.d }, za.d[x9, #3]\n"
-      ".inst 0xc1b4caa4  // fclamp { z4.s-z7.s }, z21.s, z20.s\n"
+      ".inst 0xa061c738  // st1w { z24.s-z27.s }, pn9.b, [x25, #0x4, MUL VL]\n"
       ".inst 0xc1b4caa0  // fclamp { z0.s-z3.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
+      ".inst 0xa062c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x8, MUL VL]\n"
       ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c72c  // st1w { z12.s-z15.s }, pn9.b, [x25, #0x8, MUL VL]\n"
       ".inst 0xa063c330  // st1w { z16.s-z19.s }, p8, [x25, #0xc, MUL VL]\n"
       "addvl x25, x25, #16\n"
       "b 35f\n"
       "34:"  // Width 4: No activation
       ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
-      ".inst 0xc0062c64  // mova { z4.d-z7.d }, za.d[x9, #3]\n"
       ".inst 0xa060c72c  // st1w { z12.s-z15.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c30  // mova { z16.d-z19.d }, za.d[x9, #1]\n"
+      ".inst 0xa061c730  // st1w { z16.s-z19.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
       ".inst 0xa062c730  // st1w { z16.s-z19.s }, pn9.b, [x25, #0x8, MUL VL]\n"
+      ".inst 0xc0062c64  // mova { z4.d-z7.d }, za.d[x9, #3]\n"
       ".inst 0xa063c324  // st1w { z4.s-z7.s }, p8, [x25, #0xc, MUL VL]\n"
       "addvl x25, x25, #16\n"
       "35:"  // Width 4: Output done
-      "subs x26, x26, #0x4\n"
+      "subs x27, x27, #0x4\n"
       "sub %x[N], %x[N], x28, LSL #2\n"
       "bgt 4b\n"
       "36:"  // Exit
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16_mla_16VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16_mla_16VL.hpp
deleted file mode 100644
index a473be77f1..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16_mla_16VL.hpp
+++ /dev/null
@@ -1,85 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#if defined(ARM_COMPUTE_ENABLE_SME2)
-#include "../std_transforms_sme.hpp"
-
-#define ARGLIST  \
-    const __fp16 *, const __fp16 *, \
-    __fp16 *, size_t, size_t, \
-    const __fp16 *, Activation, bool
-
-namespace arm_gemm
-{
-void sme2_gemv_fp16_mla_16VL( ARGLIST );
-
-class cls_sme2_gemv_fp16_mla_16VL
-{
-public:
-    typedef __fp16 operand_type;
-    typedef __fp16 result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    static unsigned int out_width()
-    {
-        return sme::get_vector_length<__fp16>() * 16;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 1;
-    }
-
-    static constexpr bool supports_accumulate()
-    {
-        return false;
-    }
-
-    static constexpr bool supports_bias()
-    {
-        return true;
-    }
-
-    static constexpr bool supports_activation()
-    {
-        return true;
-    }
-
-
-    StdTransformsSME<operand_type, result_type, 1, 16, 1> transforms = {};
-
-
-    // Default to the generic kernel
-    kern_type kernel=sme2_gemv_fp16_mla_16VL;
-    cls_sme2_gemv_fp16_mla_16VL(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-
-#endif  // defined(ARM_COMPUTE_ENABLE_SME2)
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16_mla_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16_mla_16VL/generic.cpp
deleted file mode 100644
index 4d18cc4670..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp16_mla_16VL/generic.cpp
+++ /dev/null
@@ -1,776 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#if defined(ARM_COMPUTE_ENABLE_SME2)
-
-#include "arm_gemm.hpp"
-#include "../../utils.hpp"
-
-#include <cassert>
-#include <limits>
-
-namespace arm_gemm {
-
-void sme2_gemv_fp16_mla_16VL (
-    const __fp16 *A_ptr, const __fp16 *B_ptr, __fp16 *output_ptr,
-    size_t N, size_t K,
-    const __fp16 *bias, Activation act, bool
-)
-{
-    struct KernelArgs {
-        __fp16 maxval = static_cast<__fp16>(std::numeric_limits<float>::infinity());
-        __fp16 minval = - static_cast<__fp16>(std::numeric_limits<float>::infinity());
-        const __fp16 *B_ptr = {};
-        size_t output_offset = {};
-        unsigned int input_initial_col = {};
-    } ka;
-
-    unsigned long flags=0;
-    ka.B_ptr = B_ptr;
-    switch(act.type) {
-        default:
-        case Activation::Type::None:
-            break;
-        case Activation::Type::BoundedReLU:
-            ka.maxval = static_cast<__fp16>(act.param1);
-            /* fall through */
-        case Activation::Type::ReLU:
-            ka.minval = 0;
-            flags |= 0x2;
-            break;
-    }
-    __asm__ __volatile__(
-      "ptrue p8.b\n"
-      ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x9, #0x0\n"
-      "cnth x28, ALL, MUL #4\n"
-      "mov x27, %x[B_ptr]\n"
-      "add x26, %x[N], x28\n"
-      "mov x25, %x[output_ptr]\n"
-      "sub x26, x26, #0x1\n"
-      "ptrue p1.b\n"
-      "udiv x26, x26, x28\n"
-      ".inst 0x25207811  // ptrue pn9.b\n"
-      "add x22, x26, #0x3\n"
-      "mov x21, #0x1\n"
-      "and x22, x22, #0xfffffffffffffffc\n"
-      "mul x22, x22, x28\n"
-      "mul x22, x22, %x[K]\n"
-      "lsl x22, x22, #0x1\n"
-      "1:"  // RHS size check loop
-      "cmp x22, #0x200000\n"
-      "blt 2f\n"
-      "tbnz x22, #0, 3f\n"
-      "lsr x22, x22, #0x1\n"
-      "lsl x21, x21, #0x1\n"
-      "b 1b\n"
-      "2:"  // RHS do prefetch
-      "lsl x20, x22, #0x26\n"
-      "sub x21, x21, #0x1\n"
-      "lsl x21, x21, #0x16\n"
-      "orr x22, x22, x20\n"
-      "orr x22, x22, x21\n"
-      ".inst 0xf8b64b7a  // rprfm pldonce, x22, [x27]\n"
-      "3:"  // RHS prefetch exit
-      "mov x24, %x[bias]\n"
-      "4:"  // Column loop
-      "cmp x26, #0x4\n"
-      "bge 28f\n"
-      "cmp x26, #0x2\n"
-      "bgt 20f\n"
-      "beq 12f\n"
-      "mov x23, %x[A_ptr]\n"
-      "lsl x21, %x[K], #0x1\n"
-      "mov x20, %x[N]\n"
-      "mov x22, %x[K]\n"
-      ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
-      ".inst 0x257467f0  // whilelt p8.h, XZR, x20, VLx4\n"
-      "cbz x24, 5f\n"
-      ".inst 0xa040a708  // ld1h { z8.h-z11.h }, pn9.b/Z, [x24]\n"
-      ".inst 0xc0042d00  // mova za.d[x9, #0], { z8.d-z11.d }\n"
-      "b 6f\n"
-      "5:"  // Width 1: no bias
-      ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
-      "6:"  // Width 1: setup done
-      "cmp x22, #0x8\n"
-      "ble 8f\n"
-      "7:"  // Width 1: Multiply loop: Main loop head
-      "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      "ld1rqh { z3.h }, p0/Z, [x23]\n"
-      "sub x22, x22, #0x8\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      "cmp x22, #0x8\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc113b280  // fmla za.h[x9, 0], { z20.h-z23.h }, z3.h[0]\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc113b388  // fmla za.h[x9, 0], { z28.h-z31.h }, z3.h[1]\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc113b500  // fmla za.h[x9, 0], { z8.h-z11.h }, z3.h[2]\n"
-      ".inst 0xc113b488  // fmla za.h[x9, 0], { z4.h-z7.h }, z3.h[3]\n"
-      ".inst 0xc113ba80  // fmla za.h[x9, 0], { z20.h-z23.h }, z3.h[4]\n"
-      ".inst 0xc113b988  // fmla za.h[x9, 0], { z12.h-z15.h }, z3.h[5]\n"
-      ".inst 0xc113bf80  // fmla za.h[x9, 0], { z28.h-z31.h }, z3.h[6]\n"
-      ".inst 0xc113be08  // fmla za.h[x9, 0], { z16.h-z19.h }, z3.h[7]\n"
-      "bgt 7b\n"
-      "8:"  // Width 1: Multiply loop: Single iteration only
-      "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "ld1rqh { z15.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb280  // fmla za.h[x9, 0], { z20.h-z23.h }, z15.h[0]\n"
-      "ble 9f\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb208  // fmla za.h[x9, 0], { z16.h-z19.h }, z15.h[1]\n"
-      "ble 9f\n"
-      ".inst 0xa040a761  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb400  // fmla za.h[x9, 0], { z0.h-z3.h }, z15.h[2]\n"
-      "ble 9f\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb788  // fmla za.h[x9, 0], { z28.h-z31.h }, z15.h[3]\n"
-      "ble 9f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb880  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[4]\n"
-      "ble 9f\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fba08  // fmla za.h[x9, 0], { z16.h-z19.h }, z15.h[5]\n"
-      "ble 9f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbc80  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[6]\n"
-      "ble 9f\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbf88  // fmla za.h[x9, 0], { z28.h-z31.h }, z15.h[7]\n"
-      "9:"  // Width 1: Multiply loop: multiply skip
-      "tbz %x[flags], #1, 10f\n"
-      "add x21, %x[args_ptr], %[offset_min]\n"
-      "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
-      "ld1rh { z4.h }, p1/Z, [x21]\n"
-      "ld1rh { z21.h }, p1/Z, [x20]\n"
-      ".inst 0xc175c888  // fclamp { z8.h-z11.h }, z4.h, z21.h\n"
-      ".inst 0xa060a328  // st1h { z8.h-z11.h }, p8, [x25]\n"
-      "addvl x25, x25, #4\n"
-      "b 11f\n"
-      "10:"  // Width 1: No activation
-      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
-      ".inst 0xa060a320  // st1h { z0.h-z3.h }, p8, [x25]\n"
-      "addvl x25, x25, #4\n"
-      "11:"  // Width 1: Output done
-      "b 36f\n"
-      "12:"  // Width 2
-      "mov x23, %x[A_ptr]\n"
-      "lsl x21, %x[K], #0x1\n"
-      "sub x20, %x[N], x28\n"
-      "mov x22, %x[K]\n"
-      ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
-      ".inst 0x257467f0  // whilelt p8.h, XZR, x20, VLx4\n"
-      "cbz x24, 13f\n"
-      ".inst 0xa040a71c  // ld1h { z28.h-z31.h }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041a70c  // ld1h { z12.h-z15.h }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xc0042f80  // mova za.d[x9, #0], { z28.d-z31.d }\n"
-      ".inst 0xc0042d81  // mova za.d[x9, #1], { z12.d-z15.d }\n"
-      "b 14f\n"
-      "13:"  // Width 2: no bias
-      ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
-      "14:"  // Width 2: setup done
-      "cmp x22, #0x8\n"
-      "ble 16f\n"
-      "15:"  // Width 2: Multiply loop: Main loop head
-      "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      "sub x22, x22, #0x8\n"
-      "ld1rqh { z3.h }, p0/Z, [x23]\n"
-      "cmp x22, #0x8\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc113b100  // fmla za.h[x9, 0], { z8.h-z11.h }, z3.h[0]\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc113b301  // fmla za.h[x9, 1], { z24.h-z27.h }, z3.h[0]\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc113b288  // fmla za.h[x9, 0], { z20.h-z23.h }, z3.h[1]\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc113b209  // fmla za.h[x9, 1], { z16.h-z19.h }, z3.h[1]\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc113b500  // fmla za.h[x9, 0], { z8.h-z11.h }, z3.h[2]\n"
-      ".inst 0xa041a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc113b581  // fmla za.h[x9, 1], { z12.h-z15.h }, z3.h[2]\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc113b688  // fmla za.h[x9, 0], { z20.h-z23.h }, z3.h[3]\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc113b709  // fmla za.h[x9, 1], { z24.h-z27.h }, z3.h[3]\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc113bb80  // fmla za.h[x9, 0], { z28.h-z31.h }, z3.h[4]\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc113b881  // fmla za.h[x9, 1], { z4.h-z7.h }, z3.h[4]\n"
-      ".inst 0xc113b988  // fmla za.h[x9, 0], { z12.h-z15.h }, z3.h[5]\n"
-      ".inst 0xc113ba09  // fmla za.h[x9, 1], { z16.h-z19.h }, z3.h[5]\n"
-      ".inst 0xc113bd00  // fmla za.h[x9, 0], { z8.h-z11.h }, z3.h[6]\n"
-      ".inst 0xc113be81  // fmla za.h[x9, 1], { z20.h-z23.h }, z3.h[6]\n"
-      ".inst 0xc113bf08  // fmla za.h[x9, 0], { z24.h-z27.h }, z3.h[7]\n"
-      ".inst 0xc113bf89  // fmla za.h[x9, 1], { z28.h-z31.h }, z3.h[7]\n"
-      "bgt 15b\n"
-      "16:"  // Width 2: Multiply loop: Single iteration only
-      "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "ld1rqh { z15.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb380  // fmla za.h[x9, 0], { z28.h-z31.h }, z15.h[0]\n"
-      ".inst 0xc11fb101  // fmla za.h[x9, 1], { z8.h-z11.h }, z15.h[0]\n"
-      "ble 17f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb088  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[1]\n"
-      ".inst 0xc11fb289  // fmla za.h[x9, 1], { z20.h-z23.h }, z15.h[1]\n"
-      "ble 17f\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb600  // fmla za.h[x9, 0], { z16.h-z19.h }, z15.h[2]\n"
-      ".inst 0xc11fb501  // fmla za.h[x9, 1], { z8.h-z11.h }, z15.h[2]\n"
-      "ble 17f\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb508  // fmla za.h[x9, 0], { z8.h-z11.h }, z15.h[3]\n"
-      ".inst 0xc11fb789  // fmla za.h[x9, 1], { z28.h-z31.h }, z15.h[3]\n"
-      "ble 17f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb880  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[4]\n"
-      ".inst 0xc11fba01  // fmla za.h[x9, 1], { z16.h-z19.h }, z15.h[4]\n"
-      "ble 17f\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbb08  // fmla za.h[x9, 0], { z24.h-z27.h }, z15.h[5]\n"
-      ".inst 0xc11fb909  // fmla za.h[x9, 1], { z8.h-z11.h }, z15.h[5]\n"
-      "ble 17f\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbf80  // fmla za.h[x9, 0], { z28.h-z31.h }, z15.h[6]\n"
-      ".inst 0xc11fbd01  // fmla za.h[x9, 1], { z8.h-z11.h }, z15.h[6]\n"
-      "ble 17f\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbf08  // fmla za.h[x9, 0], { z24.h-z27.h }, z15.h[7]\n"
-      ".inst 0xc11fbf89  // fmla za.h[x9, 1], { z28.h-z31.h }, z15.h[7]\n"
-      "17:"  // Width 2: Multiply loop: multiply skip
-      "tbz %x[flags], #1, 18f\n"
-      "add x21, %x[args_ptr], %[offset_min]\n"
-      "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c10  // mova { z16.d-z19.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c3c  // mova { z28.d-z31.d }, za.d[x9, #1]\n"
-      "ld1rh { z15.h }, p1/Z, [x21]\n"
-      "ld1rh { z27.h }, p1/Z, [x20]\n"
-      ".inst 0xc17bc9f0  // fclamp { z16.h-z19.h }, z15.h, z27.h\n"
-      ".inst 0xc17bc9fc  // fclamp { z28.h-z31.h }, z15.h, z27.h\n"
-      ".inst 0xa060a730  // st1h { z16.h-z19.h }, pn9.b, [x25]\n"
-      ".inst 0xa061a33c  // st1h { z28.h-z31.h }, p8, [x25, #0x4, MUL VL]\n"
-      "addvl x25, x25, #8\n"
-      "b 19f\n"
-      "18:"  // Width 2: No activation
-      ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c3c  // mova { z28.d-z31.d }, za.d[x9, #1]\n"
-      ".inst 0xa060a72c  // st1h { z12.h-z15.h }, pn9.b, [x25]\n"
-      ".inst 0xa061a33c  // st1h { z28.h-z31.h }, p8, [x25, #0x4, MUL VL]\n"
-      "addvl x25, x25, #8\n"
-      "19:"  // Width 2: Output done
-      "b 36f\n"
-      "20:"  // Width 3
-      "mov x20, #0x2\n"
-      "mov x23, %x[A_ptr]\n"
-      "lsl x21, %x[K], #0x1\n"
-      "msub x20, x28, x20, %x[N]\n"
-      "mov x22, %x[K]\n"
-      ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
-      ".inst 0x257467f0  // whilelt p8.h, XZR, x20, VLx4\n"
-      "cbz x24, 21f\n"
-      ".inst 0xa040a700  // ld1h { z0.h-z3.h }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041a710  // ld1h { z16.h-z19.h }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042a708  // ld1h { z8.h-z11.h }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xc0042c00  // mova za.d[x9, #0], { z0.d-z3.d }\n"
-      ".inst 0xc0042e01  // mova za.d[x9, #1], { z16.d-z19.d }\n"
-      ".inst 0xc0042d02  // mova za.d[x9, #2], { z8.d-z11.d }\n"
-      "b 22f\n"
-      "21:"  // Width 3: no bias
-      ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
-      "22:"  // Width 3: setup done
-      "cmp x22, #0x8\n"
-      "ble 24f\n"
-      "23:"  // Width 3: Multiply loop: Main loop head
-      "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
-      "sub x22, x22, #0x8\n"
-      "ld1rqh { z0.h }, p0/Z, [x23]\n"
-      "cmp x22, #0x8\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc110b180  // fmla za.h[x9, 0], { z12.h-z15.h }, z0.h[0]\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc110b281  // fmla za.h[x9, 1], { z20.h-z23.h }, z0.h[0]\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc110b102  // fmla za.h[x9, 2], { z8.h-z11.h }, z0.h[0]\n"
-      ".inst 0xa042a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc110b088  // fmla za.h[x9, 0], { z4.h-z7.h }, z0.h[1]\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc110b309  // fmla za.h[x9, 1], { z24.h-z27.h }, z0.h[1]\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc110b18a  // fmla za.h[x9, 2], { z12.h-z15.h }, z0.h[1]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc110b680  // fmla za.h[x9, 0], { z20.h-z23.h }, z0.h[2]\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc110b781  // fmla za.h[x9, 1], { z28.h-z31.h }, z0.h[2]\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc110b502  // fmla za.h[x9, 2], { z8.h-z11.h }, z0.h[2]\n"
-      ".inst 0xa042a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc110b608  // fmla za.h[x9, 0], { z16.h-z19.h }, z0.h[3]\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc110b709  // fmla za.h[x9, 1], { z24.h-z27.h }, z0.h[3]\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc110b48a  // fmla za.h[x9, 2], { z4.h-z7.h }, z0.h[3]\n"
-      ".inst 0xa042a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc110bb80  // fmla za.h[x9, 0], { z28.h-z31.h }, z0.h[4]\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc110b901  // fmla za.h[x9, 1], { z8.h-z11.h }, z0.h[4]\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc110b882  // fmla za.h[x9, 2], { z4.h-z7.h }, z0.h[4]\n"
-      ".inst 0xa042a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc110b988  // fmla za.h[x9, 0], { z12.h-z15.h }, z0.h[5]\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc110b909  // fmla za.h[x9, 1], { z8.h-z11.h }, z0.h[5]\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc110bb0a  // fmla za.h[x9, 2], { z24.h-z27.h }, z0.h[5]\n"
-      ".inst 0xa042a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc110be80  // fmla za.h[x9, 0], { z20.h-z23.h }, z0.h[6]\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc110be01  // fmla za.h[x9, 1], { z16.h-z19.h }, z0.h[6]\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc110bd82  // fmla za.h[x9, 2], { z12.h-z15.h }, z0.h[6]\n"
-      ".inst 0xa042a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc110bc88  // fmla za.h[x9, 0], { z4.h-z7.h }, z0.h[7]\n"
-      ".inst 0xc110be09  // fmla za.h[x9, 1], { z16.h-z19.h }, z0.h[7]\n"
-      ".inst 0xc110be8a  // fmla za.h[x9, 2], { z20.h-z23.h }, z0.h[7]\n"
-      "bgt 23b\n"
-      "24:"  // Width 3: Multiply loop: Single iteration only
-      "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "ld1rqh { z15.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb080  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[0]\n"
-      ".inst 0xc11fb301  // fmla za.h[x9, 1], { z24.h-z27.h }, z15.h[0]\n"
-      ".inst 0xc11fb382  // fmla za.h[x9, 2], { z28.h-z31.h }, z15.h[0]\n"
-      "ble 25f\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb308  // fmla za.h[x9, 0], { z24.h-z27.h }, z15.h[1]\n"
-      ".inst 0xc11fb089  // fmla za.h[x9, 1], { z4.h-z7.h }, z15.h[1]\n"
-      ".inst 0xc11fb10a  // fmla za.h[x9, 2], { z8.h-z11.h }, z15.h[1]\n"
-      "ble 25f\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb780  // fmla za.h[x9, 0], { z28.h-z31.h }, z15.h[2]\n"
-      ".inst 0xc11fb601  // fmla za.h[x9, 1], { z16.h-z19.h }, z15.h[2]\n"
-      ".inst 0xc11fb502  // fmla za.h[x9, 2], { z8.h-z11.h }, z15.h[2]\n"
-      "ble 25f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb488  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[3]\n"
-      ".inst 0xc11fb709  // fmla za.h[x9, 1], { z24.h-z27.h }, z15.h[3]\n"
-      ".inst 0xc11fb60a  // fmla za.h[x9, 2], { z16.h-z19.h }, z15.h[3]\n"
-      "ble 25f\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb900  // fmla za.h[x9, 0], { z8.h-z11.h }, z15.h[4]\n"
-      ".inst 0xc11fbb01  // fmla za.h[x9, 1], { z24.h-z27.h }, z15.h[4]\n"
-      ".inst 0xc11fba02  // fmla za.h[x9, 2], { z16.h-z19.h }, z15.h[4]\n"
-      "ble 25f\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb908  // fmla za.h[x9, 0], { z8.h-z11.h }, z15.h[5]\n"
-      ".inst 0xc11fba09  // fmla za.h[x9, 1], { z16.h-z19.h }, z15.h[5]\n"
-      ".inst 0xc11fbb0a  // fmla za.h[x9, 2], { z24.h-z27.h }, z15.h[5]\n"
-      "ble 25f\n"
-      ".inst 0xa040a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbd00  // fmla za.h[x9, 0], { z8.h-z11.h }, z15.h[6]\n"
-      ".inst 0xc11fbe81  // fmla za.h[x9, 1], { z20.h-z23.h }, z15.h[6]\n"
-      ".inst 0xc11fbf02  // fmla za.h[x9, 2], { z24.h-z27.h }, z15.h[6]\n"
-      "ble 25f\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbf88  // fmla za.h[x9, 0], { z28.h-z31.h }, z15.h[7]\n"
-      ".inst 0xc11fbd09  // fmla za.h[x9, 1], { z8.h-z11.h }, z15.h[7]\n"
-      ".inst 0xc11fbe8a  // fmla za.h[x9, 2], { z20.h-z23.h }, z15.h[7]\n"
-      "25:"  // Width 3: Multiply loop: multiply skip
-      "tbz %x[flags], #1, 26f\n"
-      "add x21, %x[args_ptr], %[offset_min]\n"
-      "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
-      "ld1rh { z17.h }, p1/Z, [x21]\n"
-      ".inst 0xc0062c44  // mova { z4.d-z7.d }, za.d[x9, #2]\n"
-      "ld1rh { z16.h }, p1/Z, [x20]\n"
-      ".inst 0xc170ca2c  // fclamp { z12.h-z15.h }, z17.h, z16.h\n"
-      ".inst 0xc170ca20  // fclamp { z0.h-z3.h }, z17.h, z16.h\n"
-      ".inst 0xa060a72c  // st1h { z12.h-z15.h }, pn9.b, [x25]\n"
-      ".inst 0xc170ca24  // fclamp { z4.h-z7.h }, z17.h, z16.h\n"
-      ".inst 0xa061a720  // st1h { z0.h-z3.h }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062a324  // st1h { z4.h-z7.h }, p8, [x25, #0x8, MUL VL]\n"
-      "addvl x25, x25, #12\n"
-      "b 27f\n"
-      "26:"  // Width 3: No activation
-      ".inst 0xc0062c10  // mova { z16.d-z19.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c38  // mova { z24.d-z27.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c54  // mova { z20.d-z23.d }, za.d[x9, #2]\n"
-      ".inst 0xa060a730  // st1h { z16.h-z19.h }, pn9.b, [x25]\n"
-      ".inst 0xa061a738  // st1h { z24.h-z27.h }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062a334  // st1h { z20.h-z23.h }, p8, [x25, #0x8, MUL VL]\n"
-      "addvl x25, x25, #12\n"
-      "27:"  // Width 3: Output done
-      "b 36f\n"
-      "28:"  // Width 4
-      "mov x20, #0x3\n"
-      "mov x23, %x[A_ptr]\n"
-      "lsl x21, %x[K], #0x1\n"
-      "msub x20, x28, x20, %x[N]\n"
-      "mov x22, %x[K]\n"
-      ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
-      ".inst 0x257467f0  // whilelt p8.h, XZR, x20, VLx4\n"
-      "cbz x24, 29f\n"
-      ".inst 0xa040a714  // ld1h { z20.h-z23.h }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041a704  // ld1h { z4.h-z7.h }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042a708  // ld1h { z8.h-z11.h }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xa043a700  // ld1h { z0.h-z3.h }, pn9.b/Z, [x24, #0xc, MUL VL]\n"
-      ".inst 0xc0042e80  // mova za.d[x9, #0], { z20.d-z23.d }\n"
-      "addvl x24, x24, #16\n"
-      ".inst 0xc0042c81  // mova za.d[x9, #1], { z4.d-z7.d }\n"
-      ".inst 0xc0042d02  // mova za.d[x9, #2], { z8.d-z11.d }\n"
-      ".inst 0xc0042c03  // mova za.d[x9, #3], { z0.d-z3.d }\n"
-      "b 30f\n"
-      "29:"  // Width 4: no bias
-      ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
-      "30:"  // Width 4: setup done
-      "cmp x22, #0x8\n"
-      "ble 32f\n"
-      "31:"  // Width 4: Multiply loop: Main loop head
-      "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      "sub x22, x22, #0x8\n"
-      "ld1rqh { z4.h }, p0/Z, [x23]\n"
-      "cmp x22, #0x8\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc114b300  // fmla za.h[x9, 0], { z24.h-z27.h }, z4.h[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc114b181  // fmla za.h[x9, 1], { z12.h-z15.h }, z4.h[0]\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc114b382  // fmla za.h[x9, 2], { z28.h-z31.h }, z4.h[0]\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc114b103  // fmla za.h[x9, 3], { z8.h-z11.h }, z4.h[0]\n"
-      ".inst 0xa042a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc114b208  // fmla za.h[x9, 0], { z16.h-z19.h }, z4.h[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc114b189  // fmla za.h[x9, 1], { z12.h-z15.h }, z4.h[1]\n"
-      ".inst 0xa040a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc114b30a  // fmla za.h[x9, 2], { z24.h-z27.h }, z4.h[1]\n"
-      ".inst 0xa041a761  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc114b10b  // fmla za.h[x9, 3], { z8.h-z11.h }, z4.h[1]\n"
-      ".inst 0xa042a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc114b580  // fmla za.h[x9, 0], { z12.h-z15.h }, z4.h[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc114b401  // fmla za.h[x9, 1], { z0.h-z3.h }, z4.h[2]\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc114b702  // fmla za.h[x9, 2], { z24.h-z27.h }, z4.h[2]\n"
-      ".inst 0xa041a761  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc114b503  // fmla za.h[x9, 3], { z8.h-z11.h }, z4.h[2]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc114b788  // fmla za.h[x9, 0], { z28.h-z31.h }, z4.h[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc114b409  // fmla za.h[x9, 1], { z0.h-z3.h }, z4.h[3]\n"
-      ".inst 0xa040a761  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc114b50a  // fmla za.h[x9, 2], { z8.h-z11.h }, z4.h[3]\n"
-      ".inst 0xa041a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc114b58b  // fmla za.h[x9, 3], { z12.h-z15.h }, z4.h[3]\n"
-      ".inst 0xa042a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc114b800  // fmla za.h[x9, 0], { z0.h-z3.h }, z4.h[4]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc114bb01  // fmla za.h[x9, 1], { z24.h-z27.h }, z4.h[4]\n"
-      ".inst 0xa040a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc114b982  // fmla za.h[x9, 2], { z12.h-z15.h }, z4.h[4]\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc114b903  // fmla za.h[x9, 3], { z8.h-z11.h }, z4.h[4]\n"
-      ".inst 0xa042a761  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc114ba88  // fmla za.h[x9, 0], { z20.h-z23.h }, z4.h[5]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc114ba09  // fmla za.h[x9, 1], { z16.h-z19.h }, z4.h[5]\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc114b80a  // fmla za.h[x9, 2], { z0.h-z3.h }, z4.h[5]\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc114b98b  // fmla za.h[x9, 3], { z12.h-z15.h }, z4.h[5]\n"
-      ".inst 0xa042a761  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc114bf00  // fmla za.h[x9, 0], { z24.h-z27.h }, z4.h[6]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc114be01  // fmla za.h[x9, 1], { z16.h-z19.h }, z4.h[6]\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xc114bc02  // fmla za.h[x9, 2], { z0.h-z3.h }, z4.h[6]\n"
-      ".inst 0xa041a76d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc114bd03  // fmla za.h[x9, 3], { z8.h-z11.h }, z4.h[6]\n"
-      ".inst 0xa042a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc114be08  // fmla za.h[x9, 0], { z16.h-z19.h }, z4.h[7]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc114bd89  // fmla za.h[x9, 1], { z12.h-z15.h }, z4.h[7]\n"
-      ".inst 0xc114be8a  // fmla za.h[x9, 2], { z20.h-z23.h }, z4.h[7]\n"
-      ".inst 0xc114bd0b  // fmla za.h[x9, 3], { z8.h-z11.h }, z4.h[7]\n"
-      "bgt 31b\n"
-      "32:"  // Width 4: Multiply loop: Single iteration only
-      "whilelt p0.h, XZR, x22\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      "ld1rqh { z15.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa041a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a761  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc11fb080  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb201  // fmla za.h[x9, 1], { z16.h-z19.h }, z15.h[0]\n"
-      ".inst 0xc11fb102  // fmla za.h[x9, 2], { z8.h-z11.h }, z15.h[0]\n"
-      ".inst 0xc11fb003  // fmla za.h[x9, 3], { z0.h-z3.h }, z15.h[0]\n"
-      "ble 33f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc11fb088  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb389  // fmla za.h[x9, 1], { z28.h-z31.h }, z15.h[1]\n"
-      ".inst 0xc11fb10a  // fmla za.h[x9, 2], { z8.h-z11.h }, z15.h[1]\n"
-      ".inst 0xc11fb20b  // fmla za.h[x9, 3], { z16.h-z19.h }, z15.h[1]\n"
-      "ble 33f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc11fb480  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb501  // fmla za.h[x9, 1], { z8.h-z11.h }, z15.h[2]\n"
-      ".inst 0xc11fb682  // fmla za.h[x9, 2], { z20.h-z23.h }, z15.h[2]\n"
-      ".inst 0xc11fb603  // fmla za.h[x9, 3], { z16.h-z19.h }, z15.h[2]\n"
-      "ble 33f\n"
-      ".inst 0xa040a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a761  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc11fb788  // fmla za.h[x9, 0], { z28.h-z31.h }, z15.h[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fb509  // fmla za.h[x9, 1], { z8.h-z11.h }, z15.h[3]\n"
-      ".inst 0xc11fb48a  // fmla za.h[x9, 2], { z4.h-z7.h }, z15.h[3]\n"
-      ".inst 0xc11fb40b  // fmla za.h[x9, 3], { z0.h-z3.h }, z15.h[3]\n"
-      "ble 33f\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc11fbb00  // fmla za.h[x9, 0], { z24.h-z27.h }, z15.h[4]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fba81  // fmla za.h[x9, 1], { z20.h-z23.h }, z15.h[4]\n"
-      ".inst 0xc11fba02  // fmla za.h[x9, 2], { z16.h-z19.h }, z15.h[4]\n"
-      ".inst 0xc11fb883  // fmla za.h[x9, 3], { z4.h-z7.h }, z15.h[4]\n"
-      "ble 33f\n"
-      ".inst 0xa040a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc11fba08  // fmla za.h[x9, 0], { z16.h-z19.h }, z15.h[5]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbb89  // fmla za.h[x9, 1], { z28.h-z31.h }, z15.h[5]\n"
-      ".inst 0xc11fba8a  // fmla za.h[x9, 2], { z20.h-z23.h }, z15.h[5]\n"
-      ".inst 0xc11fb88b  // fmla za.h[x9, 3], { z4.h-z7.h }, z15.h[5]\n"
-      "ble 33f\n"
-      ".inst 0xa040a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n"
-      "subs x22, x22, #0x1\n"
-      ".inst 0xa041a769  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a771  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc11fbc80  // fmla za.h[x9, 0], { z4.h-z7.h }, z15.h[6]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbd01  // fmla za.h[x9, 1], { z8.h-z11.h }, z15.h[6]\n"
-      ".inst 0xc11fbe82  // fmla za.h[x9, 2], { z20.h-z23.h }, z15.h[6]\n"
-      ".inst 0xc11fbe03  // fmla za.h[x9, 3], { z16.h-z19.h }, z15.h[6]\n"
-      "ble 33f\n"
-      ".inst 0xa040a779  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041a77d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042a765  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043a775  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc11fbf08  // fmla za.h[x9, 0], { z24.h-z27.h }, z15.h[7]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc11fbf89  // fmla za.h[x9, 1], { z28.h-z31.h }, z15.h[7]\n"
-      ".inst 0xc11fbc8a  // fmla za.h[x9, 2], { z4.h-z7.h }, z15.h[7]\n"
-      ".inst 0xc11fbe8b  // fmla za.h[x9, 3], { z20.h-z23.h }, z15.h[7]\n"
-      "33:"  // Width 4: Multiply loop: multiply skip
-      "tbz %x[flags], #1, 34f\n"
-      "add x21, %x[args_ptr], %[offset_min]\n"
-      "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      "ld1rh { z17.h }, p1/Z, [x21]\n"
-      ".inst 0xc0062c40  // mova { z0.d-z3.d }, za.d[x9, #2]\n"
-      "ld1rh { z16.h }, p1/Z, [x20]\n"
-      ".inst 0xc0062c68  // mova { z8.d-z11.d }, za.d[x9, #3]\n"
-      ".inst 0xc170ca24  // fclamp { z4.h-z7.h }, z17.h, z16.h\n"
-      ".inst 0xc170ca2c  // fclamp { z12.h-z15.h }, z17.h, z16.h\n"
-      ".inst 0xc170ca20  // fclamp { z0.h-z3.h }, z17.h, z16.h\n"
-      ".inst 0xa060a724  // st1h { z4.h-z7.h }, pn9.b, [x25]\n"
-      ".inst 0xa061a72c  // st1h { z12.h-z15.h }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xc170ca28  // fclamp { z8.h-z11.h }, z17.h, z16.h\n"
-      ".inst 0xa062a720  // st1h { z0.h-z3.h }, pn9.b, [x25, #0x8, MUL VL]\n"
-      ".inst 0xa063a328  // st1h { z8.h-z11.h }, p8, [x25, #0xc, MUL VL]\n"
-      "addvl x25, x25, #16\n"
-      "b 35f\n"
-      "34:"  // Width 4: No activation
-      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
-      ".inst 0xc0062c60  // mova { z0.d-z3.d }, za.d[x9, #3]\n"
-      ".inst 0xa060a728  // st1h { z8.h-z11.h }, pn9.b, [x25]\n"
-      ".inst 0xa061a72c  // st1h { z12.h-z15.h }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062a730  // st1h { z16.h-z19.h }, pn9.b, [x25, #0x8, MUL VL]\n"
-      ".inst 0xa063a320  // st1h { z0.h-z3.h }, p8, [x25, #0xc, MUL VL]\n"
-      "addvl x25, x25, #16\n"
-      "35:"  // Width 4: Output done
-      "subs x26, x26, #0x4\n"
-      "sub %x[N], %x[N], x28, LSL #2\n"
-      "bgt 4b\n"
-      "36:"  // Exit
-      ".inst 0xd503467f  // SMSTOP\n"
-      "ptrue p8.b\n"
-      : [N] "+&r" (N)
-      : [A_ptr] "r" (A_ptr), [B_ptr] "r" (B_ptr), [K] "r" (K), [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [output_ptr] "r" (output_ptr)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
-    );
-}
-
-} // namespace arm_gemm
-
-#endif  // defined(ARM_COMPUTE_ENABLE_SME2)
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp
index aec02fa337..1cce355583 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32_mla_16VL/generic.cpp
@@ -63,21 +63,21 @@ void sme2_gemv_fp32_mla_16VL (
     __asm__ __volatile__(
       "ptrue p8.b\n"
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x9, #0x0\n"
       "cntw x28, ALL, MUL #4\n"
-      "mov x27, %x[B_ptr]\n"
-      "add x26, %x[N], x28\n"
-      "mov x25, %x[output_ptr]\n"
-      "sub x26, x26, #0x1\n"
-      "ptrue p1.b\n"
-      "udiv x26, x26, x28\n"
-      ".inst 0x25207811  // ptrue pn9.b\n"
-      "add x22, x26, #0x3\n"
-      "mov x21, #0x1\n"
+      "add x27, %x[N], x28\n"
+      "sub x27, x27, #0x1\n"
+      "udiv x27, x27, x28\n"
+      "add x22, x27, #0x3\n"
       "and x22, x22, #0xfffffffffffffffc\n"
       "mul x22, x22, x28\n"
       "mul x22, x22, %x[K]\n"
+      "mov x9, #0x0\n"
+      "mov x26, %x[B_ptr]\n"
+      "mov x25, %x[output_ptr]\n"
+      "ptrue p1.b\n"
+      ".inst 0x25207811  // ptrue pn9.b\n"
       "lsl x22, x22, #0x2\n"
+      "mov x21, #0x1\n"
       "1:"  // RHS size check loop
       "cmp x22, #0x200000\n"
       "blt 2f\n"
@@ -91,13 +91,13 @@ void sme2_gemv_fp32_mla_16VL (
       "lsl x21, x21, #0x16\n"
       "orr x22, x22, x20\n"
       "orr x22, x22, x21\n"
-      ".inst 0xf8b64b7a  // rprfm pldonce, x22, [x27]\n"
+      ".inst 0xf8b64b5a  // rprfm pldonce, x22, [x26]\n"
       "3:"  // RHS prefetch exit
       "mov x24, %x[bias]\n"
       "4:"  // Column loop
-      "cmp x26, #0x4\n"
+      "cmp x27, #0x4\n"
       "bge 28f\n"
-      "cmp x26, #0x2\n"
+      "cmp x27, #0x2\n"
       "bgt 20f\n"
       "beq 12f\n"
       "mov x23, %x[A_ptr]\n"
@@ -107,8 +107,8 @@ void sme2_gemv_fp32_mla_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 5f\n"
-      ".inst 0xa040c714  // ld1w { z20.s-z23.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xc0042e80  // mova za.d[x9, #0], { z20.d-z23.d }\n"
+      ".inst 0xa040c718  // ld1w { z24.s-z27.s }, pn9.b/Z, [x24]\n"
+      ".inst 0xc0042f00  // mova za.d[x9, #0], { z24.d-z27.d }\n"
       "b 6f\n"
       "5:"  // Width 1: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -117,59 +117,59 @@ void sme2_gemv_fp32_mla_16VL (
       "ble 8f\n"
       "7:"  // Width 1: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x22\n"
-      ".inst 0xa040c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      "ld1rqw { z2.s }, p0/Z, [x23]\n"
+      "ld1rqw { z8.s }, p0/Z, [x23]\n"
       "sub x22, x22, #0x4\n"
-      "add x23, x23, #0x10\n"
-      ".inst 0xa040c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
+      ".inst 0xa040c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158a280  // fmla za.s[x9, 0], { z20.s-z23.s }, z8.s[0]\n"
+      "addvl x26, x26, #16\n"
       "cmp x22, #0x4\n"
-      ".inst 0xa040c779  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc152a380  // fmla za.s[x9, 0], { z28.s-z31.s }, z2.s[0]\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc152a600  // fmla za.s[x9, 0], { z16.s-z19.s }, z2.s[1]\n"
-      ".inst 0xc152ab00  // fmla za.s[x9, 0], { z24.s-z27.s }, z2.s[2]\n"
-      ".inst 0xc152ad80  // fmla za.s[x9, 0], { z12.s-z15.s }, z2.s[3]\n"
+      ".inst 0xa040c745  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158a480  // fmla za.s[x9, 0], { z4.s-z7.s }, z8.s[1]\n"
+      "addvl x26, x26, #16\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0xa040c75d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158ab80  // fmla za.s[x9, 0], { z28.s-z31.s }, z8.s[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c759  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158af00  // fmla za.s[x9, 0], { z24.s-z27.s }, z8.s[3]\n"
+      "addvl x26, x26, #16\n"
       "bgt 7b\n"
       "8:"  // Width 1: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x22\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
+      "ld1rqw { z11.s }, p0/Z, [x23]\n"
       "subs x22, x22, #0x1\n"
-      "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      ".inst 0xa040c75d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a180  // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[0]\n"
+      ".inst 0xc15ba380  // fmla za.s[x9, 0], { z28.s-z31.s }, z11.s[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 9f\n"
-      ".inst 0xa040c765  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x1\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a480  // fmla za.s[x9, 0], { z4.s-z7.s }, z3.s[1]\n"
+      ".inst 0xc15ba580  // fmla za.s[x9, 0], { z12.s-z15.s }, z11.s[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 9f\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x1\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a980  // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[2]\n"
+      ".inst 0xc15baa00  // fmla za.s[x9, 0], { z16.s-z19.s }, z11.s[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 9f\n"
-      ".inst 0xa040c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153ad00  // fmla za.s[x9, 0], { z8.s-z11.s }, z3.s[3]\n"
+      ".inst 0xa040c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15bac00  // fmla za.s[x9, 0], { z0.s-z3.s }, z11.s[3]\n"
+      "addvl x26, x26, #16\n"
       "9:"  // Width 1: Multiply loop: multiply skip
       "tbz %x[flags], #1, 10f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
-      "ld1rw { z23.s }, p1/Z, [x21]\n"
-      "ld1rw { z22.s }, p1/Z, [x20]\n"
-      ".inst 0xc1b6cae0  // fclamp { z0.s-z3.s }, z23.s, z22.s\n"
-      ".inst 0xa060c320  // st1w { z0.s-z3.s }, p8, [x25]\n"
+      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
+      "ld1rw { z3.s }, p1/Z, [x21]\n"
+      "ld1rw { z29.s }, p1/Z, [x20]\n"
+      ".inst 0xc1bdc868  // fclamp { z8.s-z11.s }, z3.s, z29.s\n"
+      ".inst 0xa060c328  // st1w { z8.s-z11.s }, p8, [x25]\n"
       "addvl x25, x25, #4\n"
       "b 11f\n"
       "10:"  // Width 1: No activation
-      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
-      ".inst 0xa060c320  // st1w { z0.s-z3.s }, p8, [x25]\n"
+      ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c32c  // st1w { z12.s-z15.s }, p8, [x25]\n"
       "addvl x25, x25, #4\n"
       "11:"  // Width 1: Output done
       "b 36f\n"
@@ -181,10 +181,10 @@ void sme2_gemv_fp32_mla_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 13f\n"
-      ".inst 0xa040c71c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xc0042f80  // mova za.d[x9, #0], { z28.d-z31.d }\n"
-      ".inst 0xc0042c81  // mova za.d[x9, #1], { z4.d-z7.d }\n"
+      ".inst 0xa040c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24]\n"
+      ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
+      ".inst 0xa041c714  // ld1w { z20.s-z23.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042e81  // mova za.d[x9, #1], { z20.d-z23.d }\n"
       "b 14f\n"
       "13:"  // Width 2: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -193,80 +193,80 @@ void sme2_gemv_fp32_mla_16VL (
       "ble 16f\n"
       "15:"  // Width 2: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x22\n"
-      ".inst 0xa040c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n"
+      "ld1rqw { z9.s }, p0/Z, [x23]\n"
       "sub x22, x22, #0x4\n"
-      "ld1rqw { z1.s }, p0/Z, [x23]\n"
+      ".inst 0xa040c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc159a180  // fmla za.s[x9, 0], { z12.s-z15.s }, z9.s[0]\n"
       "cmp x22, #0x4\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xc151a380  // fmla za.s[x9, 0], { z28.s-z31.s }, z1.s[0]\n"
-      ".inst 0xa041c779  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc151a181  // fmla za.s[x9, 1], { z12.s-z15.s }, z1.s[0]\n"
-      ".inst 0xa040c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xc151a600  // fmla za.s[x9, 0], { z16.s-z19.s }, z1.s[1]\n"
-      ".inst 0xa041c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc151a701  // fmla za.s[x9, 1], { z24.s-z27.s }, z1.s[1]\n"
-      ".inst 0xc151ab80  // fmla za.s[x9, 0], { z28.s-z31.s }, z1.s[2]\n"
-      ".inst 0xc151a981  // fmla za.s[x9, 1], { z12.s-z15.s }, z1.s[2]\n"
-      ".inst 0xc151ad00  // fmla za.s[x9, 0], { z8.s-z11.s }, z1.s[3]\n"
-      ".inst 0xc151ae81  // fmla za.s[x9, 1], { z20.s-z23.s }, z1.s[3]\n"
+      ".inst 0xa041c745  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc159a081  // fmla za.s[x9, 1], { z4.s-z7.s }, z9.s[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc159a580  // fmla za.s[x9, 0], { z12.s-z15.s }, z9.s[1]\n"
+      ".inst 0xa041c745  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc159a481  // fmla za.s[x9, 1], { z4.s-z7.s }, z9.s[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c759  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc159ab00  // fmla za.s[x9, 0], { z24.s-z27.s }, z9.s[2]\n"
+      ".inst 0xa041c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc159a801  // fmla za.s[x9, 1], { z0.s-z3.s }, z9.s[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc159ac00  // fmla za.s[x9, 0], { z0.s-z3.s }, z9.s[3]\n"
+      ".inst 0xa041c75d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc159af81  // fmla za.s[x9, 1], { z28.s-z31.s }, z9.s[3]\n"
+      "addvl x26, x26, #16\n"
       "bgt 15b\n"
       "16:"  // Width 2: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x22\n"
-      ".inst 0xa040c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n"
+      "ld1rqw { z11.s }, p0/Z, [x23]\n"
       "subs x22, x22, #0x1\n"
-      "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      ".inst 0xa040c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a200  // fmla za.s[x9, 0], { z16.s-z19.s }, z3.s[0]\n"
-      ".inst 0xc153a381  // fmla za.s[x9, 1], { z28.s-z31.s }, z3.s[0]\n"
+      ".inst 0xc15ba180  // fmla za.s[x9, 0], { z12.s-z15.s }, z11.s[0]\n"
+      ".inst 0xa041c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15ba001  // fmla za.s[x9, 1], { z0.s-z3.s }, z11.s[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 17f\n"
-      ".inst 0xa040c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040c759  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x1\n"
-      ".inst 0xa041c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a680  // fmla za.s[x9, 0], { z20.s-z23.s }, z3.s[1]\n"
-      ".inst 0xc153a601  // fmla za.s[x9, 1], { z16.s-z19.s }, z3.s[1]\n"
+      ".inst 0xc15ba700  // fmla za.s[x9, 0], { z24.s-z27.s }, z11.s[1]\n"
+      ".inst 0xa041c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15ba401  // fmla za.s[x9, 1], { z0.s-z3.s }, z11.s[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 17f\n"
-      ".inst 0xa040c765  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x1\n"
-      ".inst 0xa041c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a880  // fmla za.s[x9, 0], { z4.s-z7.s }, z3.s[2]\n"
-      ".inst 0xc153aa01  // fmla za.s[x9, 1], { z16.s-z19.s }, z3.s[2]\n"
+      ".inst 0xc15ba980  // fmla za.s[x9, 0], { z12.s-z15.s }, z11.s[2]\n"
+      ".inst 0xa041c75d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bab81  // fmla za.s[x9, 1], { z28.s-z31.s }, z11.s[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 17f\n"
-      ".inst 0xa040c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153af80  // fmla za.s[x9, 0], { z28.s-z31.s }, z3.s[3]\n"
-      ".inst 0xc153ad81  // fmla za.s[x9, 1], { z12.s-z15.s }, z3.s[3]\n"
+      ".inst 0xa040c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15bae00  // fmla za.s[x9, 0], { z16.s-z19.s }, z11.s[3]\n"
+      ".inst 0xa041c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bae81  // fmla za.s[x9, 1], { z20.s-z23.s }, z11.s[3]\n"
+      "addvl x26, x26, #16\n"
       "17:"  // Width 2: Multiply loop: multiply skip
       "tbz %x[flags], #1, 18f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c28  // mova { z8.d-z11.d }, za.d[x9, #1]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
-      "ld1rw { z23.s }, p1/Z, [x20]\n"
-      ".inst 0xc1b7ca24  // fclamp { z4.s-z7.s }, z17.s, z23.s\n"
-      ".inst 0xc1b7ca28  // fclamp { z8.s-z11.s }, z17.s, z23.s\n"
-      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c328  // st1w { z8.s-z11.s }, p8, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
+      "ld1rw { z9.s }, p1/Z, [x21]\n"
+      ".inst 0xc0062c24  // mova { z4.d-z7.d }, za.d[x9, #1]\n"
+      "ld1rw { z8.s }, p1/Z, [x20]\n"
+      ".inst 0xc1a8c920  // fclamp { z0.s-z3.s }, z9.s, z8.s\n"
+      ".inst 0xa060c720  // st1w { z0.s-z3.s }, pn9.b, [x25]\n"
+      ".inst 0xc1a8c924  // fclamp { z4.s-z7.s }, z9.s, z8.s\n"
+      ".inst 0xa061c324  // st1w { z4.s-z7.s }, p8, [x25, #0x4, MUL VL]\n"
       "addvl x25, x25, #8\n"
       "b 19f\n"
       "18:"  // Width 2: No activation
-      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c30  // mova { z16.d-z19.d }, za.d[x9, #1]\n"
-      ".inst 0xa060c728  // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c330  // st1w { z16.s-z19.s }, p8, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c10  // mova { z16.d-z19.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c730  // st1w { z16.s-z19.s }, pn9.b, [x25]\n"
+      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
+      ".inst 0xa061c32c  // st1w { z12.s-z15.s }, p8, [x25, #0x4, MUL VL]\n"
       "addvl x25, x25, #8\n"
       "19:"  // Width 2: Output done
       "b 36f\n"
@@ -279,12 +279,12 @@ void sme2_gemv_fp32_mla_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 21f\n"
-      ".inst 0xa040c708  // ld1w { z8.s-z11.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c700  // ld1w { z0.s-z3.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xc0042d00  // mova za.d[x9, #0], { z8.d-z11.d }\n"
-      ".inst 0xc0042c01  // mova za.d[x9, #1], { z0.d-z3.d }\n"
-      ".inst 0xc0042c82  // mova za.d[x9, #2], { z4.d-z7.d }\n"
+      ".inst 0xa040c718  // ld1w { z24.s-z27.s }, pn9.b/Z, [x24]\n"
+      ".inst 0xc0042f00  // mova za.d[x9, #0], { z24.d-z27.d }\n"
+      ".inst 0xa041c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042c81  // mova za.d[x9, #1], { z4.d-z7.d }\n"
+      ".inst 0xa042c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
+      ".inst 0xc0042e02  // mova za.d[x9, #2], { z16.d-z19.d }\n"
       "b 22f\n"
       "21:"  // Width 3: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -293,101 +293,101 @@ void sme2_gemv_fp32_mla_16VL (
       "ble 24f\n"
       "23:"  // Width 3: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x22\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
+      "ld1rqw { z15.s }, p0/Z, [x23]\n"
       "sub x22, x22, #0x4\n"
-      "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      ".inst 0xa040c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fa000  // fmla za.s[x9, 0], { z0.s-z3.s }, z15.s[0]\n"
       "cmp x22, #0x4\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c765  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a180  // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[0]\n"
-      ".inst 0xa040c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153a101  // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[0]\n"
-      ".inst 0xa041c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153a082  // fmla za.s[x9, 2], { z4.s-z7.s }, z3.s[0]\n"
-      ".inst 0xa042c779  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153a600  // fmla za.s[x9, 0], { z16.s-z19.s }, z3.s[1]\n"
-      ".inst 0xa041c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153a681  // fmla za.s[x9, 1], { z20.s-z23.s }, z3.s[1]\n"
-      ".inst 0xa042c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a702  // fmla za.s[x9, 2], { z24.s-z27.s }, z3.s[1]\n"
-      ".inst 0xa040c765  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153a980  // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[2]\n"
-      ".inst 0xa042c779  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153ab81  // fmla za.s[x9, 1], { z28.s-z31.s }, z3.s[2]\n"
-      ".inst 0xc153a902  // fmla za.s[x9, 2], { z8.s-z11.s }, z3.s[2]\n"
-      ".inst 0xc153ac80  // fmla za.s[x9, 0], { z4.s-z7.s }, z3.s[3]\n"
-      ".inst 0xc153ae81  // fmla za.s[x9, 1], { z20.s-z23.s }, z3.s[3]\n"
-      ".inst 0xc153af02  // fmla za.s[x9, 2], { z24.s-z27.s }, z3.s[3]\n"
+      ".inst 0xa041c745  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fa081  // fmla za.s[x9, 1], { z4.s-z7.s }, z15.s[0]\n"
+      ".inst 0xa042c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fa002  // fmla za.s[x9, 2], { z0.s-z3.s }, z15.s[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fa680  // fmla za.s[x9, 0], { z20.s-z23.s }, z15.s[1]\n"
+      ".inst 0xa041c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fa681  // fmla za.s[x9, 1], { z20.s-z23.s }, z15.s[1]\n"
+      ".inst 0xa042c749  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fa502  // fmla za.s[x9, 2], { z8.s-z11.s }, z15.s[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c759  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fab00  // fmla za.s[x9, 0], { z24.s-z27.s }, z15.s[2]\n"
+      ".inst 0xa041c749  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fa901  // fmla za.s[x9, 1], { z8.s-z11.s }, z15.s[2]\n"
+      ".inst 0xa042c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15faa02  // fmla za.s[x9, 2], { z16.s-z19.s }, z15.s[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fae00  // fmla za.s[x9, 0], { z16.s-z19.s }, z15.s[3]\n"
+      ".inst 0xa041c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fae01  // fmla za.s[x9, 1], { z16.s-z19.s }, z15.s[3]\n"
+      ".inst 0xa042c749  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fad02  // fmla za.s[x9, 2], { z8.s-z11.s }, z15.s[3]\n"
+      "addvl x26, x26, #16\n"
       "bgt 23b\n"
       "24:"  // Width 3: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x22\n"
-      ".inst 0xa040c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27]\n"
+      "ld1rqw { z11.s }, p0/Z, [x23]\n"
       "subs x22, x22, #0x1\n"
-      "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      ".inst 0xa040c75d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c765  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a280  // fmla za.s[x9, 0], { z20.s-z23.s }, z3.s[0]\n"
-      ".inst 0xc153a181  // fmla za.s[x9, 1], { z12.s-z15.s }, z3.s[0]\n"
-      ".inst 0xc153a082  // fmla za.s[x9, 2], { z4.s-z7.s }, z3.s[0]\n"
+      ".inst 0xc15ba380  // fmla za.s[x9, 0], { z28.s-z31.s }, z11.s[0]\n"
+      ".inst 0xa041c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15ba001  // fmla za.s[x9, 1], { z0.s-z3.s }, z11.s[0]\n"
+      ".inst 0xa042c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15ba282  // fmla za.s[x9, 2], { z20.s-z23.s }, z11.s[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 25f\n"
-      ".inst 0xa040c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x1\n"
-      ".inst 0xa041c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a680  // fmla za.s[x9, 0], { z20.s-z23.s }, z3.s[1]\n"
-      ".inst 0xc153a501  // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[1]\n"
-      ".inst 0xc153a602  // fmla za.s[x9, 2], { z16.s-z19.s }, z3.s[1]\n"
+      ".inst 0xc15ba580  // fmla za.s[x9, 0], { z12.s-z15.s }, z11.s[1]\n"
+      ".inst 0xa041c745  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15ba481  // fmla za.s[x9, 1], { z4.s-z7.s }, z11.s[1]\n"
+      ".inst 0xa042c75d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15ba782  // fmla za.s[x9, 2], { z28.s-z31.s }, z11.s[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 25f\n"
-      ".inst 0xa040c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040c745  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x1\n"
-      ".inst 0xa041c779  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153ab80  // fmla za.s[x9, 0], { z28.s-z31.s }, z3.s[2]\n"
-      ".inst 0xc153ab01  // fmla za.s[x9, 1], { z24.s-z27.s }, z3.s[2]\n"
-      ".inst 0xc153a982  // fmla za.s[x9, 2], { z12.s-z15.s }, z3.s[2]\n"
+      ".inst 0xc15ba880  // fmla za.s[x9, 0], { z4.s-z7.s }, z11.s[2]\n"
+      ".inst 0xa041c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15baa81  // fmla za.s[x9, 1], { z20.s-z23.s }, z11.s[2]\n"
+      ".inst 0xa042c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15ba982  // fmla za.s[x9, 2], { z12.s-z15.s }, z11.s[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 25f\n"
-      ".inst 0xa040c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153ad00  // fmla za.s[x9, 0], { z8.s-z11.s }, z3.s[3]\n"
-      ".inst 0xc153af81  // fmla za.s[x9, 1], { z28.s-z31.s }, z3.s[3]\n"
-      ".inst 0xc153ad82  // fmla za.s[x9, 2], { z12.s-z15.s }, z3.s[3]\n"
+      ".inst 0xa040c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15bad80  // fmla za.s[x9, 0], { z12.s-z15.s }, z11.s[3]\n"
+      ".inst 0xa041c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15bae81  // fmla za.s[x9, 1], { z20.s-z23.s }, z11.s[3]\n"
+      ".inst 0xa042c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bae02  // fmla za.s[x9, 2], { z16.s-z19.s }, z11.s[3]\n"
+      "addvl x26, x26, #16\n"
       "25:"  // Width 3: Multiply loop: multiply skip
       "tbz %x[flags], #1, 26f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      "ld1rw { z21.s }, p1/Z, [x21]\n"
-      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
-      "ld1rw { z20.s }, p1/Z, [x20]\n"
-      ".inst 0xc1b4caa8  // fclamp { z8.s-z11.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xa060c728  // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c72c  // st1w { z12.s-z15.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c330  // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n"
+      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
+      "ld1rw { z17.s }, p1/Z, [x21]\n"
+      ".inst 0xc0062c28  // mova { z8.d-z11.d }, za.d[x9, #1]\n"
+      "ld1rw { z16.s }, p1/Z, [x20]\n"
+      ".inst 0xc1b0ca24  // fclamp { z4.s-z7.s }, z17.s, z16.s\n"
+      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
+      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
+      ".inst 0xc1b0ca28  // fclamp { z8.s-z11.s }, z17.s, z16.s\n"
+      ".inst 0xa061c728  // st1w { z8.s-z11.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc1b0ca2c  // fclamp { z12.s-z15.s }, z17.s, z16.s\n"
+      ".inst 0xa062c32c  // st1w { z12.s-z15.s }, p8, [x25, #0x8, MUL VL]\n"
       "addvl x25, x25, #12\n"
       "b 27f\n"
       "26:"  // Width 3: No activation
-      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
+      ".inst 0xc0062c14  // mova { z20.d-z23.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c734  // st1w { z20.s-z23.s }, pn9.b, [x25]\n"
       ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c5c  // mova { z28.d-z31.d }, za.d[x9, #2]\n"
-      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
       ".inst 0xa061c72c  // st1w { z12.s-z15.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c33c  // st1w { z28.s-z31.s }, p8, [x25, #0x8, MUL VL]\n"
+      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
+      ".inst 0xa062c32c  // st1w { z12.s-z15.s }, p8, [x25, #0x8, MUL VL]\n"
       "addvl x25, x25, #12\n"
       "27:"  // Width 3: Output done
       "b 36f\n"
@@ -401,14 +401,14 @@ void sme2_gemv_fp32_mla_16VL (
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 29f\n"
       ".inst 0xa040c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c70c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042c71c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xa043c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n"
       ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
-      "addvl x24, x24, #16\n"
-      ".inst 0xc0042d81  // mova za.d[x9, #1], { z12.d-z15.d }\n"
-      ".inst 0xc0042f82  // mova za.d[x9, #2], { z28.d-z31.d }\n"
+      ".inst 0xa041c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042e01  // mova za.d[x9, #1], { z16.d-z19.d }\n"
+      ".inst 0xa042c70c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
+      ".inst 0xc0042d82  // mova za.d[x9, #2], { z12.d-z15.d }\n"
+      ".inst 0xa043c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n"
       ".inst 0xc0042e03  // mova za.d[x9, #3], { z16.d-z19.d }\n"
+      "addvl x24, x24, #16\n"
       "b 30f\n"
       "29:"  // Width 4: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -417,125 +417,125 @@ void sme2_gemv_fp32_mla_16VL (
       "ble 32f\n"
       "31:"  // Width 4: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x22\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
+      "ld1rqw { z8.s }, p0/Z, [x23]\n"
       "sub x22, x22, #0x4\n"
-      "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      ".inst 0xa040c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158a200  // fmla za.s[x9, 0], { z16.s-z19.s }, z8.s[0]\n"
       "cmp x22, #0x4\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153a180  // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a281  // fmla za.s[x9, 1], { z20.s-z23.s }, z3.s[0]\n"
-      ".inst 0xa040c779  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153a202  // fmla za.s[x9, 2], { z16.s-z19.s }, z3.s[0]\n"
-      ".inst 0xa041c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153a103  // fmla za.s[x9, 3], { z8.s-z11.s }, z3.s[0]\n"
-      ".inst 0xa042c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043c765  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153a700  // fmla za.s[x9, 0], { z24.s-z27.s }, z3.s[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a581  // fmla za.s[x9, 1], { z12.s-z15.s }, z3.s[1]\n"
-      ".inst 0xa040c779  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153a502  // fmla za.s[x9, 2], { z8.s-z11.s }, z3.s[1]\n"
-      ".inst 0xa041c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153a483  // fmla za.s[x9, 3], { z4.s-z7.s }, z3.s[1]\n"
-      ".inst 0xa042c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043c765  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153ab00  // fmla za.s[x9, 0], { z24.s-z27.s }, z3.s[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a901  // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[2]\n"
-      ".inst 0xa040c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xc153aa02  // fmla za.s[x9, 2], { z16.s-z19.s }, z3.s[2]\n"
-      ".inst 0xa041c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc153a883  // fmla za.s[x9, 3], { z4.s-z7.s }, z3.s[2]\n"
-      ".inst 0xa042c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153ad00  // fmla za.s[x9, 0], { z8.s-z11.s }, z3.s[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153af81  // fmla za.s[x9, 1], { z28.s-z31.s }, z3.s[3]\n"
-      ".inst 0xc153ad82  // fmla za.s[x9, 2], { z12.s-z15.s }, z3.s[3]\n"
-      ".inst 0xc153ae83  // fmla za.s[x9, 3], { z20.s-z23.s }, z3.s[3]\n"
+      ".inst 0xa041c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc158a181  // fmla za.s[x9, 1], { z12.s-z15.s }, z8.s[0]\n"
+      ".inst 0xa042c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc158a202  // fmla za.s[x9, 2], { z16.s-z19.s }, z8.s[0]\n"
+      ".inst 0xa043c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc158a183  // fmla za.s[x9, 3], { z12.s-z15.s }, z8.s[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c74d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158a580  // fmla za.s[x9, 0], { z12.s-z15.s }, z8.s[1]\n"
+      ".inst 0xa041c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc158a681  // fmla za.s[x9, 1], { z20.s-z23.s }, z8.s[1]\n"
+      ".inst 0xa042c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc158a602  // fmla za.s[x9, 2], { z16.s-z19.s }, z8.s[1]\n"
+      ".inst 0xa043c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc158a683  // fmla za.s[x9, 3], { z20.s-z23.s }, z8.s[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c745  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158a880  // fmla za.s[x9, 0], { z4.s-z7.s }, z8.s[2]\n"
+      ".inst 0xa041c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc158aa01  // fmla za.s[x9, 1], { z16.s-z19.s }, z8.s[2]\n"
+      ".inst 0xa042c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc158aa02  // fmla za.s[x9, 2], { z16.s-z19.s }, z8.s[2]\n"
+      ".inst 0xa043c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc158a803  // fmla za.s[x9, 3], { z0.s-z3.s }, z8.s[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc158ae80  // fmla za.s[x9, 0], { z20.s-z23.s }, z8.s[3]\n"
+      ".inst 0xa041c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc158ae01  // fmla za.s[x9, 1], { z16.s-z19.s }, z8.s[3]\n"
+      ".inst 0xa042c745  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc158ac82  // fmla za.s[x9, 2], { z4.s-z7.s }, z8.s[3]\n"
+      ".inst 0xa043c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc158ae83  // fmla za.s[x9, 3], { z20.s-z23.s }, z8.s[3]\n"
+      "addvl x26, x26, #16\n"
       "bgt 31b\n"
       "32:"  // Width 4: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x22\n"
-      ".inst 0xa040c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n"
+      "ld1rqw { z11.s }, p0/Z, [x23]\n"
       "subs x22, x22, #0x1\n"
-      "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      ".inst 0xa040c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa041c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c77d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153a200  // fmla za.s[x9, 0], { z16.s-z19.s }, z3.s[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a181  // fmla za.s[x9, 1], { z12.s-z15.s }, z3.s[0]\n"
-      ".inst 0xc153a382  // fmla za.s[x9, 2], { z28.s-z31.s }, z3.s[0]\n"
-      ".inst 0xc153a283  // fmla za.s[x9, 3], { z20.s-z23.s }, z3.s[0]\n"
+      ".inst 0xc15ba200  // fmla za.s[x9, 0], { z16.s-z19.s }, z11.s[0]\n"
+      ".inst 0xa041c755  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15ba281  // fmla za.s[x9, 1], { z20.s-z23.s }, z11.s[0]\n"
+      ".inst 0xa042c75d  // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15ba382  // fmla za.s[x9, 2], { z28.s-z31.s }, z11.s[0]\n"
+      ".inst 0xa043c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15ba203  // fmla za.s[x9, 3], { z16.s-z19.s }, z11.s[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 33f\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040c741  // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x1\n"
-      ".inst 0xa041c765  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c779  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153a580  // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a481  // fmla za.s[x9, 1], { z4.s-z7.s }, z3.s[1]\n"
-      ".inst 0xc153a702  // fmla za.s[x9, 2], { z24.s-z27.s }, z3.s[1]\n"
-      ".inst 0xc153a683  // fmla za.s[x9, 3], { z20.s-z23.s }, z3.s[1]\n"
+      ".inst 0xc15ba400  // fmla za.s[x9, 0], { z0.s-z3.s }, z11.s[1]\n"
+      ".inst 0xa041c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15ba601  // fmla za.s[x9, 1], { z16.s-z19.s }, z11.s[1]\n"
+      ".inst 0xa042c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15ba602  // fmla za.s[x9, 2], { z16.s-z19.s }, z11.s[1]\n"
+      ".inst 0xa043c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15ba603  // fmla za.s[x9, 3], { z16.s-z19.s }, z11.s[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 33f\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x1\n"
-      ".inst 0xa041c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153a980  // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153a901  // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[2]\n"
-      ".inst 0xc153aa82  // fmla za.s[x9, 2], { z20.s-z23.s }, z3.s[2]\n"
-      ".inst 0xc153aa03  // fmla za.s[x9, 3], { z16.s-z19.s }, z3.s[2]\n"
+      ".inst 0xc15baa00  // fmla za.s[x9, 0], { z16.s-z19.s }, z11.s[2]\n"
+      ".inst 0xa041c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15baa01  // fmla za.s[x9, 1], { z16.s-z19.s }, z11.s[2]\n"
+      ".inst 0xa042c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15baa02  // fmla za.s[x9, 2], { z16.s-z19.s }, z11.s[2]\n"
+      ".inst 0xa043c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15baa03  // fmla za.s[x9, 3], { z16.s-z19.s }, z11.s[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 33f\n"
-      ".inst 0xa040c76d  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xa041c769  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042c775  // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043c771  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc153ad80  // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc153ad01  // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[3]\n"
-      ".inst 0xc153ae82  // fmla za.s[x9, 2], { z20.s-z23.s }, z3.s[3]\n"
-      ".inst 0xc153ae03  // fmla za.s[x9, 3], { z16.s-z19.s }, z3.s[3]\n"
+      ".inst 0xa040c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15bae00  // fmla za.s[x9, 0], { z16.s-z19.s }, z11.s[3]\n"
+      ".inst 0xa041c759  // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15baf01  // fmla za.s[x9, 1], { z24.s-z27.s }, z11.s[3]\n"
+      ".inst 0xa042c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15bae02  // fmla za.s[x9, 2], { z16.s-z19.s }, z11.s[3]\n"
+      ".inst 0xa043c751  // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15bae03  // fmla za.s[x9, 3], { z16.s-z19.s }, z11.s[3]\n"
+      "addvl x26, x26, #16\n"
       "33:"  // Width 4: Multiply loop: multiply skip
       "tbz %x[flags], #1, 34f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
+      ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
       "ld1rw { z21.s }, p1/Z, [x21]\n"
-      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
+      ".inst 0xc0062c38  // mova { z24.d-z27.d }, za.d[x9, #1]\n"
       "ld1rw { z20.s }, p1/Z, [x20]\n"
+      ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
+      ".inst 0xc0062c40  // mova { z0.d-z3.d }, za.d[x9, #2]\n"
+      ".inst 0xa060c72c  // st1w { z12.s-z15.s }, pn9.b, [x25]\n"
+      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
       ".inst 0xc0062c70  // mova { z16.d-z19.d }, za.d[x9, #3]\n"
-      ".inst 0xc1b4caa4  // fclamp { z4.s-z7.s }, z21.s, z20.s\n"
+      ".inst 0xa061c738  // st1w { z24.s-z27.s }, pn9.b, [x25, #0x4, MUL VL]\n"
       ".inst 0xc1b4caa0  // fclamp { z0.s-z3.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
+      ".inst 0xa062c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x8, MUL VL]\n"
       ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c72c  // st1w { z12.s-z15.s }, pn9.b, [x25, #0x8, MUL VL]\n"
       ".inst 0xa063c330  // st1w { z16.s-z19.s }, p8, [x25, #0xc, MUL VL]\n"
       "addvl x25, x25, #16\n"
       "b 35f\n"
       "34:"  // Width 4: No activation
       ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
-      ".inst 0xc0062c64  // mova { z4.d-z7.d }, za.d[x9, #3]\n"
       ".inst 0xa060c72c  // st1w { z12.s-z15.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c30  // mova { z16.d-z19.d }, za.d[x9, #1]\n"
+      ".inst 0xa061c730  // st1w { z16.s-z19.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
       ".inst 0xa062c730  // st1w { z16.s-z19.s }, pn9.b, [x25, #0x8, MUL VL]\n"
+      ".inst 0xc0062c64  // mova { z4.d-z7.d }, za.d[x9, #3]\n"
       ".inst 0xa063c324  // st1w { z4.s-z7.s }, p8, [x25, #0xc, MUL VL]\n"
       "addvl x25, x25, #16\n"
       "35:"  // Width 4: Output done
-      "subs x26, x26, #0x4\n"
+      "subs x27, x27, #0x4\n"
       "sub %x[N], %x[N], x28, LSL #2\n"
       "bgt 4b\n"
       "36:"  // Exit
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp
index 7ed04da9df..9747587495 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_fp32bf16fp32_dot_16VL/generic.cpp
@@ -64,22 +64,22 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
     __asm__ __volatile__(
       "ptrue p8.b\n"
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x10, #0x0\n"
-      "cntw x9, ALL, MUL #4\n"
-      "mov x28, #0x4\n"
-      "add x27, %x[N], x9\n"
+      "cntw x10, ALL, MUL #4\n"
+      "add x28, %x[N], x10\n"
+      "sub x28, x28, #0x1\n"
+      "udiv x28, x28, x10\n"
+      "add x22, x28, #0x3\n"
+      "and x22, x22, #0xfffffffffffffffc\n"
+      "mul x22, x22, x10\n"
+      "mul x22, x22, %x[K]\n"
+      "mov x9, #0x0\n"
+      "mov x27, #0x4\n"
       "mov x26, %x[B_ptr]\n"
-      "sub x27, x27, #0x1\n"
       "mov x25, %x[output_ptr]\n"
-      "udiv x27, x27, x9\n"
       "ptrue p2.b\n"
-      "add x22, x27, #0x3\n"
       ".inst 0x25207811  // ptrue pn9.b\n"
-      "and x22, x22, #0xfffffffffffffffc\n"
-      "mov x21, #0x1\n"
-      "mul x22, x22, x9\n"
-      "mul x22, x22, %x[K]\n"
       "lsl x22, x22, #0x1\n"
+      "mov x21, #0x1\n"
       "1:"  // RHS size check loop
       "cmp x22, #0x200000\n"
       "blt 2f\n"
@@ -97,9 +97,9 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       "3:"  // RHS prefetch exit
       "mov x24, %x[bias]\n"
       "4:"  // Column loop
-      "cmp x27, #0x4\n"
+      "cmp x28, #0x4\n"
       "bge 28f\n"
-      "cmp x27, #0x2\n"
+      "cmp x28, #0x2\n"
       "bgt 20f\n"
       "beq 12f\n"
       "mov x23, %x[A_ptr]\n"
@@ -110,7 +110,7 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 5f\n"
       ".inst 0xa040c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xc0044e00  // mova za.d[x10, #0], { z16.d-z19.d }\n"
+      ".inst 0xc0042e00  // mova za.d[x9, #0], { z16.d-z19.d }\n"
       "b 6f\n"
       "5:"  // Width 1: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -119,72 +119,72 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       "ble 8f\n"
       "7:"  // Width 1: Multiply loop: Main loop head
       "whilelt p1.s, XZR, x22\n"
-      "whilelt p0.s, x28, x22\n"
-      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
-      "ld1rqw { z8.s }, p1/Z, [x23]\n"
-      "addvl x26, x26, #16\n"
+      "whilelt p0.s, x27, x22\n"
+      "ld1rqw { z10.s }, p1/Z, [x23]\n"
+      ".inst 0x658aa94a  // bfcvt z10.h, p2/M, z10.s\n"
+      "ld1rqw { z16.s }, p0/Z, [x23, #16]\n"
+      ".inst 0x658aaa10  // bfcvt z16.h, p2/M, z16.s\n"
+      "uzp1 z10.h, z10.h, z10.h\n"
       "sub x22, x22, #0x8\n"
-      "ld1rqw { z11.s }, p0/Z, [x23, #16]\n"
-      "cmp x22, #0x8\n"
-      "add x23, x23, #0x20\n"
-      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
-      "addvl x26, x26, #16\n"
-      ".inst 0x658aa908  // bfcvt z8.h, p2/M, z8.s\n"
-      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
+      "uzp1 z16.h, z16.h, z16.h\n"
+      "trn1 z10.d, z10.d, z16.d\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0x658aa96b  // bfcvt z11.h, p2/M, z11.s\n"
+      ".inst 0xc15ab198  // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[0]\n"
       ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
       "addvl x26, x26, #16\n"
-      "uzp1 z8.h, z8.h, z8.h\n"
-      "uzp1 z11.h, z11.h, z11.h\n"
-      "trn1 z8.d, z8.d, z11.d\n"
-      ".inst 0xc158d098  // bfdot za.s[x10, 0], { z4.h-z7.h }, z8.h[0]\n"
-      ".inst 0xc158d698  // bfdot za.s[x10, 0], { z20.h-z23.h }, z8.h[1]\n"
-      ".inst 0xc158da18  // bfdot za.s[x10, 0], { z16.h-z19.h }, z8.h[2]\n"
-      ".inst 0xc158dd98  // bfdot za.s[x10, 0], { z12.h-z15.h }, z8.h[3]\n"
+      "cmp x22, #0x8\n"
+      ".inst 0xc15ab598  // bfdot za.s[x9, 0], { z12.h-z15.h }, z10.h[1]\n"
+      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
+      "addvl x26, x26, #16\n"
+      "add x23, x23, #0x20\n"
+      ".inst 0xc15ab818  // bfdot za.s[x9, 0], { z0.h-z3.h }, z10.h[2]\n"
+      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xc15abf98  // bfdot za.s[x9, 0], { z28.h-z31.h }, z10.h[3]\n"
       "bgt 7b\n"
       "8:"  // Width 1: Multiply loop: Single iteration only
       "whilelt p1.s, XZR, x22\n"
-      "whilelt p0.s, x28, x22\n"
-      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
-      "ld1rqw { z3.s }, p1/Z, [x23]\n"
+      "whilelt p0.s, x27, x22\n"
+      "ld1rqw { z15.s }, p1/Z, [x23]\n"
+      ".inst 0x658aa9ef  // bfcvt z15.h, p2/M, z15.s\n"
+      "ld1rqw { z17.s }, p0/Z, [x23, #16]\n"
+      ".inst 0x658aaa31  // bfcvt z17.h, p2/M, z17.s\n"
+      "uzp1 z15.h, z15.h, z15.h\n"
       "subs x22, x22, #0x2\n"
-      "addvl x26, x26, #16\n"
-      "ld1rqw { z24.s }, p0/Z, [x23, #16]\n"
+      "uzp1 z17.h, z17.h, z17.h\n"
+      "trn1 z15.d, z15.d, z17.d\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x20\n"
-      ".inst 0x658aa863  // bfcvt z3.h, p2/M, z3.s\n"
-      ".inst 0x658aab18  // bfcvt z24.h, p2/M, z24.s\n"
-      "uzp1 z3.h, z3.h, z3.h\n"
-      "uzp1 z24.h, z24.h, z24.h\n"
-      "trn1 z3.d, z3.d, z24.d\n"
-      ".inst 0xc153d198  // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[0]\n"
+      ".inst 0xc15fb218  // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 9f\n"
-      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
+      ".inst 0xc15fb418  // bfdot za.s[x9, 0], { z0.h-z3.h }, z15.h[1]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153d598  // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[1]\n"
       "ble 9f\n"
-      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
+      ".inst 0xc15fb898  // bfdot za.s[x9, 0], { z4.h-z7.h }, z15.h[2]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153d998  // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[2]\n"
       "ble 9f\n"
-      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa040a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fbd18  // bfdot za.s[x9, 0], { z8.h-z11.h }, z15.h[3]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153df98  // bfdot za.s[x10, 0], { z28.h-z31.h }, z3.h[3]\n"
       "9:"  // Width 1: Multiply loop: multiply skip
       "tbz %x[flags], #1, 10f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0064c1c  // mova { z28.d-z31.d }, za.d[x10, #0]\n"
-      "ld1rw { z4.s }, p2/Z, [x21]\n"
-      "ld1rw { z18.s }, p2/Z, [x20]\n"
-      ".inst 0xc1b2c89c  // fclamp { z28.s-z31.s }, z4.s, z18.s\n"
-      ".inst 0xa060c33c  // st1w { z28.s-z31.s }, p8, [x25]\n"
+      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
+      "ld1rw { z8.s }, p2/Z, [x21]\n"
+      "ld1rw { z26.s }, p2/Z, [x20]\n"
+      ".inst 0xc1bac900  // fclamp { z0.s-z3.s }, z8.s, z26.s\n"
+      ".inst 0xa060c320  // st1w { z0.s-z3.s }, p8, [x25]\n"
       "addvl x25, x25, #4\n"
       "b 11f\n"
       "10:"  // Width 1: No activation
-      ".inst 0xc0064c04  // mova { z4.d-z7.d }, za.d[x10, #0]\n"
+      ".inst 0xc0062c04  // mova { z4.d-z7.d }, za.d[x9, #0]\n"
       ".inst 0xa060c324  // st1w { z4.s-z7.s }, p8, [x25]\n"
       "addvl x25, x25, #4\n"
       "11:"  // Width 1: Output done
@@ -192,15 +192,15 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       "12:"  // Width 2
       "mov x23, %x[A_ptr]\n"
       "lsl x21, %x[K], #0x2\n"
-      "sub x20, %x[N], x9\n"
+      "sub x20, %x[N], x10\n"
       "mov x22, %x[K]\n"
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 13f\n"
-      ".inst 0xa040c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c700  // ld1w { z0.s-z3.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xc0044e00  // mova za.d[x10, #0], { z16.d-z19.d }\n"
-      ".inst 0xc0044c01  // mova za.d[x10, #1], { z0.d-z3.d }\n"
+      ".inst 0xa040c718  // ld1w { z24.s-z27.s }, pn9.b/Z, [x24]\n"
+      ".inst 0xc0042f00  // mova za.d[x9, #0], { z24.d-z27.d }\n"
+      ".inst 0xa041c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042e01  // mova za.d[x9, #1], { z16.d-z19.d }\n"
       "b 14f\n"
       "13:"  // Width 2: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -209,94 +209,94 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       "ble 16f\n"
       "15:"  // Width 2: Multiply loop: Main loop head
       "whilelt p1.s, XZR, x22\n"
-      "whilelt p0.s, x28, x22\n"
-      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
-      "ld1rqw { z8.s }, p1/Z, [x23]\n"
+      "whilelt p0.s, x27, x22\n"
+      "ld1rqw { z13.s }, p1/Z, [x23]\n"
+      ".inst 0x658aa9ad  // bfcvt z13.h, p2/M, z13.s\n"
+      "ld1rqw { z27.s }, p0/Z, [x23, #16]\n"
+      ".inst 0x658aab7b  // bfcvt z27.h, p2/M, z27.s\n"
+      "uzp1 z13.h, z13.h, z13.h\n"
       "sub x22, x22, #0x8\n"
-      "ld1rqw { z9.s }, p0/Z, [x23, #16]\n"
+      "uzp1 z27.h, z27.h, z27.h\n"
+      "trn1 z13.d, z13.d, z27.d\n"
+      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
       "cmp x22, #0x8\n"
+      ".inst 0xa041a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15db298  // bfdot za.s[x9, 0], { z20.h-z23.h }, z13.h[0]\n"
+      "addvl x26, x26, #16\n"
       "add x23, x23, #0x20\n"
+      ".inst 0xc15db019  // bfdot za.s[x9, 1], { z0.h-z3.h }, z13.h[0]\n"
+      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
       ".inst 0xa041a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15db698  // bfdot za.s[x9, 0], { z20.h-z23.h }, z13.h[1]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0x658aa908  // bfcvt z8.h, p2/M, z8.s\n"
-      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
-      ".inst 0x658aa929  // bfcvt z9.h, p2/M, z9.s\n"
+      ".inst 0xc15db719  // bfdot za.s[x9, 1], { z24.h-z27.h }, z13.h[1]\n"
+      ".inst 0xa040a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n"
       ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15db918  // bfdot za.s[x9, 0], { z8.h-z11.h }, z13.h[2]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
-      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      "addvl x26, x26, #16\n"
-      "uzp1 z8.h, z8.h, z8.h\n"
-      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
-      "uzp1 z9.h, z9.h, z9.h\n"
-      "trn1 z8.d, z8.d, z9.d\n"
-      ".inst 0xc158d298  // bfdot za.s[x10, 0], { z20.h-z23.h }, z8.h[0]\n"
-      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15dba19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z13.h[2]\n"
+      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15dbc18  // bfdot za.s[x9, 0], { z0.h-z3.h }, z13.h[3]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc158d319  // bfdot za.s[x10, 1], { z24.h-z27.h }, z8.h[0]\n"
-      ".inst 0xc158d418  // bfdot za.s[x10, 0], { z0.h-z3.h }, z8.h[1]\n"
-      ".inst 0xc158d619  // bfdot za.s[x10, 1], { z16.h-z19.h }, z8.h[1]\n"
-      ".inst 0xc158d898  // bfdot za.s[x10, 0], { z4.h-z7.h }, z8.h[2]\n"
-      ".inst 0xc158d999  // bfdot za.s[x10, 1], { z12.h-z15.h }, z8.h[2]\n"
-      ".inst 0xc158df98  // bfdot za.s[x10, 0], { z28.h-z31.h }, z8.h[3]\n"
-      ".inst 0xc158de99  // bfdot za.s[x10, 1], { z20.h-z23.h }, z8.h[3]\n"
+      ".inst 0xc15dbc99  // bfdot za.s[x9, 1], { z4.h-z7.h }, z13.h[3]\n"
       "bgt 15b\n"
       "16:"  // Width 2: Multiply loop: Single iteration only
       "whilelt p1.s, XZR, x22\n"
-      "whilelt p0.s, x28, x22\n"
-      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
-      "ld1rqw { z3.s }, p1/Z, [x23]\n"
+      "whilelt p0.s, x27, x22\n"
+      "ld1rqw { z15.s }, p1/Z, [x23]\n"
+      ".inst 0x658aa9ef  // bfcvt z15.h, p2/M, z15.s\n"
+      "ld1rqw { z5.s }, p0/Z, [x23, #16]\n"
+      ".inst 0x658aa8a5  // bfcvt z5.h, p2/M, z5.s\n"
+      "uzp1 z15.h, z15.h, z15.h\n"
       "subs x22, x22, #0x2\n"
-      "ld1rqw { z16.s }, p0/Z, [x23, #16]\n"
+      "uzp1 z5.h, z5.h, z5.h\n"
+      "trn1 z15.d, z15.d, z5.d\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x20\n"
-      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xa041a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb218  // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[0]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0x658aa863  // bfcvt z3.h, p2/M, z3.s\n"
-      ".inst 0x658aaa10  // bfcvt z16.h, p2/M, z16.s\n"
-      "uzp1 z3.h, z3.h, z3.h\n"
-      "uzp1 z16.h, z16.h, z16.h\n"
-      "trn1 z3.d, z3.d, z16.d\n"
-      ".inst 0xc153d318  // bfdot za.s[x10, 0], { z24.h-z27.h }, z3.h[0]\n"
-      ".inst 0xc153d099  // bfdot za.s[x10, 1], { z4.h-z7.h }, z3.h[0]\n"
+      ".inst 0xc15fb319  // bfdot za.s[x9, 1], { z24.h-z27.h }, z15.h[0]\n"
       "ble 17f\n"
-      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb798  // bfdot za.s[x9, 0], { z28.h-z31.h }, z15.h[1]\n"
+      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb499  // bfdot za.s[x9, 1], { z4.h-z7.h }, z15.h[1]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153d498  // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[1]\n"
-      ".inst 0xc153d719  // bfdot za.s[x10, 1], { z24.h-z27.h }, z3.h[1]\n"
       "ble 17f\n"
-      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fba18  // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[2]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fba19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[2]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153d898  // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[2]\n"
-      ".inst 0xc153d999  // bfdot za.s[x10, 1], { z12.h-z15.h }, z3.h[2]\n"
       "ble 17f\n"
-      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
-      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fbf98  // bfdot za.s[x9, 0], { z28.h-z31.h }, z15.h[3]\n"
+      ".inst 0xa041a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fbd19  // bfdot za.s[x9, 1], { z8.h-z11.h }, z15.h[3]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153dc98  // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[3]\n"
-      ".inst 0xc153de99  // bfdot za.s[x10, 1], { z20.h-z23.h }, z3.h[3]\n"
       "17:"  // Width 2: Multiply loop: multiply skip
       "tbz %x[flags], #1, 18f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0064c08  // mova { z8.d-z11.d }, za.d[x10, #0]\n"
-      ".inst 0xc0064c2c  // mova { z12.d-z15.d }, za.d[x10, #1]\n"
-      "ld1rw { z3.s }, p2/Z, [x21]\n"
+      ".inst 0xc0062c14  // mova { z20.d-z23.d }, za.d[x9, #0]\n"
+      "ld1rw { z11.s }, p2/Z, [x21]\n"
+      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
-      ".inst 0xc1bcc868  // fclamp { z8.s-z11.s }, z3.s, z28.s\n"
-      ".inst 0xc1bcc86c  // fclamp { z12.s-z15.s }, z3.s, z28.s\n"
-      ".inst 0xa060c728  // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
+      ".inst 0xc1bcc974  // fclamp { z20.s-z23.s }, z11.s, z28.s\n"
+      ".inst 0xa060c734  // st1w { z20.s-z23.s }, pn9.b, [x25]\n"
+      ".inst 0xc1bcc96c  // fclamp { z12.s-z15.s }, z11.s, z28.s\n"
       ".inst 0xa061c32c  // st1w { z12.s-z15.s }, p8, [x25, #0x4, MUL VL]\n"
       "addvl x25, x25, #8\n"
       "b 19f\n"
       "18:"  // Width 2: No activation
-      ".inst 0xc0064c04  // mova { z4.d-z7.d }, za.d[x10, #0]\n"
-      ".inst 0xc0064c38  // mova { z24.d-z27.d }, za.d[x10, #1]\n"
-      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c338  // st1w { z24.s-z27.s }, p8, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c720  // st1w { z0.s-z3.s }, pn9.b, [x25]\n"
+      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
+      ".inst 0xa061c320  // st1w { z0.s-z3.s }, p8, [x25, #0x4, MUL VL]\n"
       "addvl x25, x25, #8\n"
       "19:"  // Width 2: Output done
       "b 36f\n"
@@ -304,17 +304,17 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       "mov x20, #0x2\n"
       "mov x23, %x[A_ptr]\n"
       "lsl x21, %x[K], #0x2\n"
-      "msub x20, x9, x20, %x[N]\n"
+      "msub x20, x10, x20, %x[N]\n"
       "mov x22, %x[K]\n"
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 21f\n"
-      ".inst 0xa040c700  // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c718  // ld1w { z24.s-z27.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042c71c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xc0044c00  // mova za.d[x10, #0], { z0.d-z3.d }\n"
-      ".inst 0xc0044f01  // mova za.d[x10, #1], { z24.d-z27.d }\n"
-      ".inst 0xc0044f82  // mova za.d[x10, #2], { z28.d-z31.d }\n"
+      ".inst 0xa040c71c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x24]\n"
+      ".inst 0xc0042f80  // mova za.d[x9, #0], { z28.d-z31.d }\n"
+      ".inst 0xa041c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042c81  // mova za.d[x9, #1], { z4.d-z7.d }\n"
+      ".inst 0xa042c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
+      ".inst 0xc0042c82  // mova za.d[x9, #2], { z4.d-z7.d }\n"
       "b 22f\n"
       "21:"  // Width 3: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -323,114 +323,114 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       "ble 24f\n"
       "23:"  // Width 3: Multiply loop: Main loop head
       "whilelt p1.s, XZR, x22\n"
-      "whilelt p0.s, x28, x22\n"
-      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
-      "ld1rqw { z3.s }, p1/Z, [x23]\n"
+      "whilelt p0.s, x27, x22\n"
+      "ld1rqw { z14.s }, p1/Z, [x23]\n"
+      ".inst 0x658aa9ce  // bfcvt z14.h, p2/M, z14.s\n"
+      "ld1rqw { z16.s }, p0/Z, [x23, #16]\n"
+      ".inst 0x658aaa10  // bfcvt z16.h, p2/M, z16.s\n"
+      "uzp1 z14.h, z14.h, z14.h\n"
       "sub x22, x22, #0x8\n"
-      "ld1rqw { z0.s }, p0/Z, [x23, #16]\n"
+      "uzp1 z16.h, z16.h, z16.h\n"
+      "trn1 z14.d, z14.d, z16.d\n"
+      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
       "cmp x22, #0x8\n"
-      "add x23, x23, #0x20\n"
       ".inst 0xa041a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0x658aa863  // bfcvt z3.h, p2/M, z3.s\n"
-      ".inst 0xa042a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15eb098  // bfdot za.s[x9, 0], { z4.h-z7.h }, z14.h[0]\n"
+      "add x23, x23, #0x20\n"
+      ".inst 0xa042a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15eb319  // bfdot za.s[x9, 1], { z24.h-z27.h }, z14.h[0]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0x658aa800  // bfcvt z0.h, p2/M, z0.s\n"
-      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15eb01a  // bfdot za.s[x9, 2], { z0.h-z3.h }, z14.h[0]\n"
+      ".inst 0xa040a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15eb518  // bfdot za.s[x9, 0], { z8.h-z11.h }, z14.h[1]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15eb499  // bfdot za.s[x9, 1], { z4.h-z7.h }, z14.h[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xc15eb61a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z14.h[1]\n"
+      ".inst 0xa040a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n"
       ".inst 0xa041a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa042a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15eb818  // bfdot za.s[x9, 0], { z0.h-z3.h }, z14.h[2]\n"
+      ".inst 0xa042a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15ebb99  // bfdot za.s[x9, 1], { z28.h-z31.h }, z14.h[2]\n"
       "addvl x26, x26, #16\n"
-      "uzp1 z3.h, z3.h, z3.h\n"
-      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
-      "trn1 z3.d, z3.d, z0.d\n"
-      ".inst 0xc153d198  // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[0]\n"
-      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xc153d319  // bfdot za.s[x10, 1], { z24.h-z27.h }, z3.h[0]\n"
-      ".inst 0xa042a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15eb81a  // bfdot za.s[x9, 2], { z0.h-z3.h }, z14.h[2]\n"
+      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa041a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15ebf18  // bfdot za.s[x9, 0], { z24.h-z27.h }, z14.h[3]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15ebf99  // bfdot za.s[x9, 1], { z28.h-z31.h }, z14.h[3]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153d09a  // bfdot za.s[x10, 2], { z4.h-z7.h }, z3.h[0]\n"
-      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
-      ".inst 0xc153d618  // bfdot za.s[x10, 0], { z16.h-z19.h }, z3.h[1]\n"
-      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xc153d799  // bfdot za.s[x10, 1], { z28.h-z31.h }, z3.h[1]\n"
-      ".inst 0xa042a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
-      "addvl x26, x26, #16\n"
-      ".inst 0xc153d51a  // bfdot za.s[x10, 2], { z8.h-z11.h }, z3.h[1]\n"
-      ".inst 0xc153da98  // bfdot za.s[x10, 0], { z20.h-z23.h }, z3.h[2]\n"
-      ".inst 0xc153d999  // bfdot za.s[x10, 1], { z12.h-z15.h }, z3.h[2]\n"
-      ".inst 0xc153db1a  // bfdot za.s[x10, 2], { z24.h-z27.h }, z3.h[2]\n"
-      ".inst 0xc153dc98  // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[3]\n"
-      ".inst 0xc153de19  // bfdot za.s[x10, 1], { z16.h-z19.h }, z3.h[3]\n"
-      ".inst 0xc153df9a  // bfdot za.s[x10, 2], { z28.h-z31.h }, z3.h[3]\n"
+      ".inst 0xc15ebe1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z14.h[3]\n"
       "bgt 23b\n"
       "24:"  // Width 3: Multiply loop: Single iteration only
       "whilelt p1.s, XZR, x22\n"
-      "whilelt p0.s, x28, x22\n"
-      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
-      "ld1rqw { z3.s }, p1/Z, [x23]\n"
+      "whilelt p0.s, x27, x22\n"
+      "ld1rqw { z15.s }, p1/Z, [x23]\n"
+      ".inst 0x658aa9ef  // bfcvt z15.h, p2/M, z15.s\n"
+      "ld1rqw { z31.s }, p0/Z, [x23, #16]\n"
+      ".inst 0x658aabff  // bfcvt z31.h, p2/M, z31.s\n"
+      "uzp1 z15.h, z15.h, z15.h\n"
       "subs x22, x22, #0x2\n"
-      "ld1rqw { z20.s }, p0/Z, [x23, #16]\n"
+      "uzp1 z31.h, z31.h, z31.h\n"
+      "trn1 z15.d, z15.d, z31.d\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
       "add x23, x23, #0x20\n"
-      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0x658aa863  // bfcvt z3.h, p2/M, z3.s\n"
-      ".inst 0xa042a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xa041a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb218  // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[0]\n"
+      ".inst 0xa042a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fb019  // bfdot za.s[x9, 1], { z0.h-z3.h }, z15.h[0]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0x658aaa94  // bfcvt z20.h, p2/M, z20.s\n"
-      "uzp1 z3.h, z3.h, z3.h\n"
-      "uzp1 z20.h, z20.h, z20.h\n"
-      "trn1 z3.d, z3.d, z20.d\n"
-      ".inst 0xc153d098  // bfdot za.s[x10, 0], { z4.h-z7.h }, z3.h[0]\n"
-      ".inst 0xc153d199  // bfdot za.s[x10, 1], { z12.h-z15.h }, z3.h[0]\n"
-      ".inst 0xc153d31a  // bfdot za.s[x10, 2], { z24.h-z27.h }, z3.h[0]\n"
+      ".inst 0xc15fb09a  // bfdot za.s[x9, 2], { z4.h-z7.h }, z15.h[0]\n"
       "ble 25f\n"
       ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa042a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fb698  // bfdot za.s[x9, 0], { z20.h-z23.h }, z15.h[1]\n"
+      ".inst 0xa041a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb699  // bfdot za.s[x9, 1], { z20.h-z23.h }, z15.h[1]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fb61a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z15.h[1]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153d698  // bfdot za.s[x10, 0], { z20.h-z23.h }, z3.h[1]\n"
-      ".inst 0xc153d619  // bfdot za.s[x10, 1], { z16.h-z19.h }, z3.h[1]\n"
-      ".inst 0xc153d79a  // bfdot za.s[x10, 2], { z28.h-z31.h }, z3.h[1]\n"
       "ble 25f\n"
-      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa040a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa042a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fb898  // bfdot za.s[x9, 0], { z4.h-z7.h }, z15.h[2]\n"
+      ".inst 0xa041a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb819  // bfdot za.s[x9, 1], { z0.h-z3.h }, z15.h[2]\n"
+      ".inst 0xa042a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fbb1a  // bfdot za.s[x9, 2], { z24.h-z27.h }, z15.h[2]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153db98  // bfdot za.s[x10, 0], { z28.h-z31.h }, z3.h[2]\n"
-      ".inst 0xc153d919  // bfdot za.s[x10, 1], { z8.h-z11.h }, z3.h[2]\n"
-      ".inst 0xc153d99a  // bfdot za.s[x10, 2], { z12.h-z15.h }, z3.h[2]\n"
       "ble 25f\n"
-      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
-      ".inst 0xa041a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa042a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fbf98  // bfdot za.s[x9, 0], { z28.h-z31.h }, z15.h[3]\n"
+      ".inst 0xa041a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fbd19  // bfdot za.s[x9, 1], { z8.h-z11.h }, z15.h[3]\n"
+      ".inst 0xa042a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fbc9a  // bfdot za.s[x9, 2], { z4.h-z7.h }, z15.h[3]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153dd98  // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[3]\n"
-      ".inst 0xc153df19  // bfdot za.s[x10, 1], { z24.h-z27.h }, z3.h[3]\n"
-      ".inst 0xc153dd1a  // bfdot za.s[x10, 2], { z8.h-z11.h }, z3.h[3]\n"
       "25:"  // Width 3: Multiply loop: multiply skip
       "tbz %x[flags], #1, 26f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0064c08  // mova { z8.d-z11.d }, za.d[x10, #0]\n"
-      ".inst 0xc0064c20  // mova { z0.d-z3.d }, za.d[x10, #1]\n"
+      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
       "ld1rw { z17.s }, p2/Z, [x21]\n"
-      ".inst 0xc0064c44  // mova { z4.d-z7.d }, za.d[x10, #2]\n"
+      ".inst 0xc0062c24  // mova { z4.d-z7.d }, za.d[x9, #1]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
-      ".inst 0xc1b0ca28  // fclamp { z8.s-z11.s }, z17.s, z16.s\n"
-      ".inst 0xc1b0ca20  // fclamp { z0.s-z3.s }, z17.s, z16.s\n"
+      ".inst 0xc1b0ca3c  // fclamp { z28.s-z31.s }, z17.s, z16.s\n"
+      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
+      ".inst 0xa060c73c  // st1w { z28.s-z31.s }, pn9.b, [x25]\n"
       ".inst 0xc1b0ca24  // fclamp { z4.s-z7.s }, z17.s, z16.s\n"
-      ".inst 0xa060c728  // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c324  // st1w { z4.s-z7.s }, p8, [x25, #0x8, MUL VL]\n"
+      ".inst 0xa061c724  // st1w { z4.s-z7.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc1b0ca2c  // fclamp { z12.s-z15.s }, z17.s, z16.s\n"
+      ".inst 0xa062c32c  // st1w { z12.s-z15.s }, p8, [x25, #0x8, MUL VL]\n"
       "addvl x25, x25, #12\n"
       "b 27f\n"
       "26:"  // Width 3: No activation
-      ".inst 0xc0064c08  // mova { z8.d-z11.d }, za.d[x10, #0]\n"
-      ".inst 0xc0064c20  // mova { z0.d-z3.d }, za.d[x10, #1]\n"
-      ".inst 0xc0064c50  // mova { z16.d-z19.d }, za.d[x10, #2]\n"
-      ".inst 0xa060c728  // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c00  // mova { z0.d-z3.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c720  // st1w { z0.s-z3.s }, pn9.b, [x25]\n"
+      ".inst 0xc0062c30  // mova { z16.d-z19.d }, za.d[x9, #1]\n"
+      ".inst 0xa061c730  // st1w { z16.s-z19.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
       ".inst 0xa062c330  // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n"
       "addvl x25, x25, #12\n"
       "27:"  // Width 3: Output done
@@ -439,20 +439,20 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       "mov x20, #0x3\n"
       "mov x23, %x[A_ptr]\n"
       "lsl x21, %x[K], #0x2\n"
-      "msub x20, x9, x20, %x[N]\n"
+      "msub x20, x10, x20, %x[N]\n"
       "mov x22, %x[K]\n"
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       ".inst 0x25b467f0  // whilelt p8.s, XZR, x20, VLx4\n"
       "cbz x24, 29f\n"
       ".inst 0xa040c70c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x24]\n"
-      ".inst 0xa041c704  // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042c714  // ld1w { z20.s-z23.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xa043c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n"
-      ".inst 0xc0044d80  // mova za.d[x10, #0], { z12.d-z15.d }\n"
+      ".inst 0xc0042d80  // mova za.d[x9, #0], { z12.d-z15.d }\n"
+      ".inst 0xa041c70c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042d81  // mova za.d[x9, #1], { z12.d-z15.d }\n"
+      ".inst 0xa042c710  // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
+      ".inst 0xc0042e02  // mova za.d[x9, #2], { z16.d-z19.d }\n"
+      ".inst 0xa043c714  // ld1w { z20.s-z23.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n"
+      ".inst 0xc0042e83  // mova za.d[x9, #3], { z20.d-z23.d }\n"
       "addvl x24, x24, #16\n"
-      ".inst 0xc0044c81  // mova za.d[x10, #1], { z4.d-z7.d }\n"
-      ".inst 0xc0044e82  // mova za.d[x10, #2], { z20.d-z23.d }\n"
-      ".inst 0xc0044e03  // mova za.d[x10, #3], { z16.d-z19.d }\n"
       "b 30f\n"
       "29:"  // Width 4: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -461,140 +461,140 @@ void sme2_gemv_fp32bf16fp32_dot_16VL (
       "ble 32f\n"
       "31:"  // Width 4: Multiply loop: Main loop head
       "whilelt p1.s, XZR, x22\n"
-      "whilelt p0.s, x28, x22\n"
-      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
-      "ld1rqw { z7.s }, p1/Z, [x23]\n"
+      "whilelt p0.s, x27, x22\n"
+      "ld1rqw { z6.s }, p1/Z, [x23]\n"
+      ".inst 0x658aa8c6  // bfcvt z6.h, p2/M, z6.s\n"
+      "ld1rqw { z16.s }, p0/Z, [x23, #16]\n"
+      ".inst 0x658aaa10  // bfcvt z16.h, p2/M, z16.s\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
       "sub x22, x22, #0x8\n"
-      "ld1rqw { z4.s }, p0/Z, [x23, #16]\n"
+      "uzp1 z16.h, z16.h, z16.h\n"
+      "trn1 z6.d, z6.d, z16.d\n"
+      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
       "cmp x22, #0x8\n"
-      "add x23, x23, #0x20\n"
-      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0x658aa8e7  // bfcvt z7.h, p2/M, z7.s\n"
-      ".inst 0xa042a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
-      ".inst 0x658aa884  // bfcvt z4.h, p2/M, z4.s\n"
-      ".inst 0xa043a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
-      "addvl x26, x26, #16\n"
-      ".inst 0xa040a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n"
       ".inst 0xa041a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      "uzp1 z7.h, z7.h, z7.h\n"
-      ".inst 0xa042a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
-      "uzp1 z4.h, z4.h, z4.h\n"
-      "trn1 z7.d, z7.d, z4.d\n"
-      ".inst 0xc157d218  // bfdot za.s[x10, 0], { z16.h-z19.h }, z7.h[0]\n"
+      ".inst 0xc156b198  // bfdot za.s[x9, 0], { z12.h-z15.h }, z6.h[0]\n"
+      "add x23, x23, #0x20\n"
+      ".inst 0xa042a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc156b119  // bfdot za.s[x9, 1], { z8.h-z11.h }, z6.h[0]\n"
       ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc156b19a  // bfdot za.s[x9, 2], { z12.h-z15.h }, z6.h[0]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc157d199  // bfdot za.s[x10, 1], { z12.h-z15.h }, z7.h[0]\n"
-      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
-      ".inst 0xc157d39a  // bfdot za.s[x10, 2], { z28.h-z31.h }, z7.h[0]\n"
-      ".inst 0xa041a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xc157d31b  // bfdot za.s[x10, 3], { z24.h-z27.h }, z7.h[0]\n"
-      ".inst 0xa042a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
-      ".inst 0xc157d698  // bfdot za.s[x10, 0], { z20.h-z23.h }, z7.h[1]\n"
+      ".inst 0xc156b21b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z6.h[0]\n"
+      ".inst 0xa040a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc156b518  // bfdot za.s[x9, 0], { z8.h-z11.h }, z6.h[1]\n"
+      ".inst 0xa042a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc156b599  // bfdot za.s[x9, 1], { z12.h-z15.h }, z6.h[1]\n"
       ".inst 0xa043a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc156b41a  // bfdot za.s[x9, 2], { z0.h-z3.h }, z6.h[1]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc157d519  // bfdot za.s[x10, 1], { z8.h-z11.h }, z7.h[1]\n"
+      ".inst 0xc156b69b  // bfdot za.s[x9, 3], { z20.h-z23.h }, z6.h[1]\n"
       ".inst 0xa040a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n"
-      ".inst 0xc157d41a  // bfdot za.s[x10, 2], { z0.h-z3.h }, z7.h[1]\n"
-      ".inst 0xa041a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xc157d61b  // bfdot za.s[x10, 3], { z16.h-z19.h }, z7.h[1]\n"
-      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
-      ".inst 0xc157d998  // bfdot za.s[x10, 0], { z12.h-z15.h }, z7.h[2]\n"
-      ".inst 0xa043a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
-      "addvl x26, x26, #16\n"
-      ".inst 0xc157db99  // bfdot za.s[x10, 1], { z28.h-z31.h }, z7.h[2]\n"
-      ".inst 0xc157db1a  // bfdot za.s[x10, 2], { z24.h-z27.h }, z7.h[2]\n"
-      ".inst 0xc157da9b  // bfdot za.s[x10, 3], { z20.h-z23.h }, z7.h[2]\n"
-      ".inst 0xc157dd18  // bfdot za.s[x10, 0], { z8.h-z11.h }, z7.h[3]\n"
-      ".inst 0xc157dc19  // bfdot za.s[x10, 1], { z0.h-z3.h }, z7.h[3]\n"
-      ".inst 0xc157de1a  // bfdot za.s[x10, 2], { z16.h-z19.h }, z7.h[3]\n"
-      ".inst 0xc157dd9b  // bfdot za.s[x10, 3], { z12.h-z15.h }, z7.h[3]\n"
+      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc156b918  // bfdot za.s[x9, 0], { z8.h-z11.h }, z6.h[2]\n"
+      ".inst 0xa042a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc156b999  // bfdot za.s[x9, 1], { z12.h-z15.h }, z6.h[2]\n"
+      ".inst 0xa043a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc156b91a  // bfdot za.s[x9, 2], { z8.h-z11.h }, z6.h[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xc156ba9b  // bfdot za.s[x9, 3], { z20.h-z23.h }, z6.h[2]\n"
+      ".inst 0xa040a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc156bf98  // bfdot za.s[x9, 0], { z28.h-z31.h }, z6.h[3]\n"
+      ".inst 0xa042a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc156bd99  // bfdot za.s[x9, 1], { z12.h-z15.h }, z6.h[3]\n"
+      ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc156bf1a  // bfdot za.s[x9, 2], { z24.h-z27.h }, z6.h[3]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xc156be1b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z6.h[3]\n"
       "bgt 31b\n"
       "32:"  // Width 4: Multiply loop: Single iteration only
       "whilelt p1.s, XZR, x22\n"
-      "whilelt p0.s, x28, x22\n"
-      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
-      "ld1rqw { z3.s }, p1/Z, [x23]\n"
-      "subs x22, x22, #0x2\n"
+      "whilelt p0.s, x27, x22\n"
+      "ld1rqw { z15.s }, p1/Z, [x23]\n"
+      ".inst 0x658aa9ef  // bfcvt z15.h, p2/M, z15.s\n"
       "ld1rqw { z16.s }, p0/Z, [x23, #16]\n"
-      "add x23, x23, #0x20\n"
-      ".inst 0xa041a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0x658aa863  // bfcvt z3.h, p2/M, z3.s\n"
-      ".inst 0xa042a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
       ".inst 0x658aaa10  // bfcvt z16.h, p2/M, z16.s\n"
-      ".inst 0xa043a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
-      "addvl x26, x26, #16\n"
-      "uzp1 z3.h, z3.h, z3.h\n"
+      "uzp1 z15.h, z15.h, z15.h\n"
+      "subs x22, x22, #0x2\n"
       "uzp1 z16.h, z16.h, z16.h\n"
-      "trn1 z3.d, z3.d, z16.d\n"
-      ".inst 0xc153d318  // bfdot za.s[x10, 0], { z24.h-z27.h }, z3.h[0]\n"
-      ".inst 0xc153d399  // bfdot za.s[x10, 1], { z28.h-z31.h }, z3.h[0]\n"
-      ".inst 0xc153d11a  // bfdot za.s[x10, 2], { z8.h-z11.h }, z3.h[0]\n"
-      ".inst 0xc153d19b  // bfdot za.s[x10, 3], { z12.h-z15.h }, z3.h[0]\n"
+      "trn1 z15.d, z15.d, z16.d\n"
+      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
+      "add x23, x23, #0x20\n"
+      ".inst 0xa041a745  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb318  // bfdot za.s[x9, 0], { z24.h-z27.h }, z15.h[0]\n"
+      ".inst 0xa042a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fb099  // bfdot za.s[x9, 1], { z4.h-z7.h }, z15.h[0]\n"
+      ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15fb01a  // bfdot za.s[x9, 2], { z0.h-z3.h }, z15.h[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xc15fb21b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z15.h[0]\n"
       "ble 33f\n"
-      ".inst 0xa040a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa042a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
-      ".inst 0xa043a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
-      ".inst 0xc153d598  // bfdot za.s[x10, 0], { z12.h-z15.h }, z3.h[1]\n"
+      ".inst 0xc15fb718  // bfdot za.s[x9, 0], { z24.h-z27.h }, z15.h[1]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fb619  // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[1]\n"
+      ".inst 0xa042a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fb69a  // bfdot za.s[x9, 2], { z20.h-z23.h }, z15.h[1]\n"
+      ".inst 0xa043a741  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15fb41b  // bfdot za.s[x9, 3], { z0.h-z3.h }, z15.h[1]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153d799  // bfdot za.s[x10, 1], { z28.h-z31.h }, z3.h[1]\n"
-      ".inst 0xc153d51a  // bfdot za.s[x10, 2], { z8.h-z11.h }, z3.h[1]\n"
-      ".inst 0xc153d71b  // bfdot za.s[x10, 3], { z24.h-z27.h }, z3.h[1]\n"
       "ble 33f\n"
-      ".inst 0xa040a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
       "subs x22, x22, #0x2\n"
-      ".inst 0xa041a74d  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa042a75d  // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
-      ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
-      ".inst 0xc153db18  // bfdot za.s[x10, 0], { z24.h-z27.h }, z3.h[2]\n"
+      ".inst 0xc15fba18  // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[2]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fba19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[2]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fba1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z15.h[2]\n"
+      ".inst 0xa043a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc15fba9b  // bfdot za.s[x9, 3], { z20.h-z23.h }, z15.h[2]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153d999  // bfdot za.s[x10, 1], { z12.h-z15.h }, z3.h[2]\n"
-      ".inst 0xc153db9a  // bfdot za.s[x10, 2], { z28.h-z31.h }, z3.h[2]\n"
-      ".inst 0xc153da1b  // bfdot za.s[x10, 3], { z16.h-z19.h }, z3.h[2]\n"
       "ble 33f\n"
-      ".inst 0xa040a749  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26]\n"
-      ".inst 0xa041a759  // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa042a755  // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xa040a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n"
+      ".inst 0xc15fbe18  // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[3]\n"
+      ".inst 0xa041a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc15fbe19  // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[3]\n"
+      ".inst 0xa042a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc15fbe1a  // bfdot za.s[x9, 2], { z16.h-z19.h }, z15.h[3]\n"
       ".inst 0xa043a751  // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
-      ".inst 0xc153dd18  // bfdot za.s[x10, 0], { z8.h-z11.h }, z3.h[3]\n"
+      ".inst 0xc15fbe1b  // bfdot za.s[x9, 3], { z16.h-z19.h }, z15.h[3]\n"
       "addvl x26, x26, #16\n"
-      ".inst 0xc153df19  // bfdot za.s[x10, 1], { z24.h-z27.h }, z3.h[3]\n"
-      ".inst 0xc153de9a  // bfdot za.s[x10, 2], { z20.h-z23.h }, z3.h[3]\n"
-      ".inst 0xc153de1b  // bfdot za.s[x10, 3], { z16.h-z19.h }, z3.h[3]\n"
       "33:"  // Width 4: Multiply loop: multiply skip
       "tbz %x[flags], #1, 34f\n"
       "add x21, %x[args_ptr], %[offset_min]\n"
       "add x20, %x[args_ptr], %[offset_max]\n"
-      ".inst 0xc0064c04  // mova { z4.d-z7.d }, za.d[x10, #0]\n"
-      ".inst 0xc0064c28  // mova { z8.d-z11.d }, za.d[x10, #1]\n"
+      ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
       "ld1rw { z21.s }, p2/Z, [x21]\n"
-      ".inst 0xc0064c4c  // mova { z12.d-z15.d }, za.d[x10, #2]\n"
+      ".inst 0xc0062c38  // mova { z24.d-z27.d }, za.d[x9, #1]\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      ".inst 0xc0064c70  // mova { z16.d-z19.d }, za.d[x10, #3]\n"
-      ".inst 0xc1b4caa4  // fclamp { z4.s-z7.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4caa8  // fclamp { z8.s-z11.s }, z21.s, z20.s\n"
       ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
+      ".inst 0xc0062c40  // mova { z0.d-z3.d }, za.d[x9, #2]\n"
+      ".inst 0xa060c72c  // st1w { z12.s-z15.s }, pn9.b, [x25]\n"
+      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
+      ".inst 0xc0062c70  // mova { z16.d-z19.d }, za.d[x9, #3]\n"
+      ".inst 0xa061c738  // st1w { z24.s-z27.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc1b4caa0  // fclamp { z0.s-z3.s }, z21.s, z20.s\n"
+      ".inst 0xa062c720  // st1w { z0.s-z3.s }, pn9.b, [x25, #0x8, MUL VL]\n"
       ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xa060c724  // st1w { z4.s-z7.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c728  // st1w { z8.s-z11.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c72c  // st1w { z12.s-z15.s }, pn9.b, [x25, #0x8, MUL VL]\n"
       ".inst 0xa063c330  // st1w { z16.s-z19.s }, p8, [x25, #0xc, MUL VL]\n"
       "addvl x25, x25, #16\n"
       "b 35f\n"
       "34:"  // Width 4: No activation
-      ".inst 0xc0064c00  // mova { z0.d-z3.d }, za.d[x10, #0]\n"
-      ".inst 0xc0064c24  // mova { z4.d-z7.d }, za.d[x10, #1]\n"
-      ".inst 0xc0064c4c  // mova { z12.d-z15.d }, za.d[x10, #2]\n"
-      ".inst 0xc0064c78  // mova { z24.d-z27.d }, za.d[x10, #3]\n"
-      ".inst 0xa060c720  // st1w { z0.s-z3.s }, pn9.b, [x25]\n"
-      ".inst 0xa061c724  // st1w { z4.s-z7.s }, pn9.b, [x25, #0x4, MUL VL]\n"
-      ".inst 0xa062c72c  // st1w { z12.s-z15.s }, pn9.b, [x25, #0x8, MUL VL]\n"
+      ".inst 0xc0062c10  // mova { z16.d-z19.d }, za.d[x9, #0]\n"
+      ".inst 0xa060c730  // st1w { z16.s-z19.s }, pn9.b, [x25]\n"
+      ".inst 0xc0062c30  // mova { z16.d-z19.d }, za.d[x9, #1]\n"
+      ".inst 0xa061c730  // st1w { z16.s-z19.s }, pn9.b, [x25, #0x4, MUL VL]\n"
+      ".inst 0xc0062c54  // mova { z20.d-z23.d }, za.d[x9, #2]\n"
+      ".inst 0xa062c734  // st1w { z20.s-z23.s }, pn9.b, [x25, #0x8, MUL VL]\n"
+      ".inst 0xc0062c78  // mova { z24.d-z27.d }, za.d[x9, #3]\n"
       ".inst 0xa063c338  // st1w { z24.s-z27.s }, p8, [x25, #0xc, MUL VL]\n"
       "addvl x25, x25, #16\n"
       "35:"  // Width 4: Output done
-      "subs x27, x27, #0x4\n"
-      "sub %x[N], %x[N], x9, LSL #2\n"
+      "subs x28, x28, #0x4\n"
+      "sub %x[N], %x[N], x10, LSL #2\n"
       "bgt 4b\n"
       "36:"  // Exit
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_s8qa_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_s8qa_dot_16VL/generic.cpp
index 1a7cc1e70e..a1c441555e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_s8qa_dot_16VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_s8qa_dot_16VL/generic.cpp
@@ -52,20 +52,20 @@ void sme2_gemv_s8qa_dot_16VL (
     __asm__ __volatile__(
       "ptrue p8.b\n"
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x9, #0x0\n"
       "cntw x28, ALL, MUL #4\n"
-      "mov x27, %x[B_ptr]\n"
-      "add x26, %x[N], x28\n"
+      "add x27, %x[N], x28\n"
+      "sub x27, x27, #0x1\n"
+      "udiv x27, x27, x28\n"
+      "add x22, x27, #0x3\n"
+      "and x22, x22, #0xfffffffffffffffc\n"
+      "mul x22, x22, x28\n"
+      "mov x9, #0x0\n"
+      "mov x26, %x[B_ptr]\n"
       "mov x25, %x[output_ptr]\n"
-      "sub x26, x26, #0x1\n"
       "ptrue p2.b\n"
-      "udiv x26, x26, x28\n"
       ".inst 0x25207810  // ptrue pn8.b\n"
-      "add x22, x26, #0x3\n"
-      "mov x21, #0x1\n"
-      "and x22, x22, #0xfffffffffffffffc\n"
-      "mul x22, x22, x28\n"
       "mul x22, x22, %x[K]\n"
+      "mov x21, #0x1\n"
       "1:"  // RHS size check loop
       "cmp x22, #0x200000\n"
       "blt 2f\n"
@@ -79,16 +79,16 @@ void sme2_gemv_s8qa_dot_16VL (
       "lsl x21, x21, #0x16\n"
       "orr x22, x22, x20\n"
       "orr x22, x22, x21\n"
-      ".inst 0xf8b64b7a  // rprfm pldonce, x22, [x27]\n"
+      ".inst 0xf8b64b5a  // rprfm pldonce, x22, [x26]\n"
       "3:"  // RHS prefetch exit
       "mov x24, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z8.b, #0x1\n"
+      "mov z28.s, #0x0\n"
+      "mov z29.b, #0x1\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "4:"  // Column loop
-      "cmp x26, #0x4\n"
+      "cmp x27, #0x4\n"
       "bge 34f\n"
-      "cmp x26, #0x2\n"
+      "cmp x27, #0x2\n"
       "bgt 24f\n"
       "beq 14f\n"
       "mov x23, %x[A_ptr]\n"
@@ -98,8 +98,8 @@ void sme2_gemv_s8qa_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       "whilelt p1.b, XZR, x20\n"
       "cbz x24, 5f\n"
-      ".inst 0xa040c30c  // ld1w { z12.s-z15.s }, pn8.b/Z, [x24]\n"
-      ".inst 0xc0042d80  // mova za.d[x9, #0], { z12.d-z15.d }\n"
+      ".inst 0xa040c300  // ld1w { z0.s-z3.s }, pn8.b/Z, [x24]\n"
+      ".inst 0xc0042c00  // mova za.d[x9, #0], { z0.d-z3.d }\n"
       "b 6f\n"
       "5:"  // Width 1: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -108,82 +108,82 @@ void sme2_gemv_s8qa_dot_16VL (
       "ble 9f\n"
       "7:"  // Width 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b020  // sdot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b720  // sdot za.s[x9, 0], { z24.b-z27.b }, z9.b[1]\n"
-      ".inst 0xc159baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z9.b[2]\n"
-      ".inst 0xc159bfa0  // sdot za.s[x9, 0], { z28.b-z31.b }, z9.b[3]\n"
+      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b2a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b5a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b9a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bda0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "tbnz %x[flags], #31, 8f\n"
-      "sdot z11.s, z9.b, z8.b\n"
+      "sdot z28.s, z1.b, z29.b\n"
       "8:"  // Width 1: Multiply loop: unique 1: skip row sum
       "sub x22, x22, #0x10\n"
       "cmp x22, #0x10\n"
       "bgt 7b\n"
       "9:"  // Width 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x22, x22, #0x4\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b320  // sdot za.s[x9, 0], { z24.b-z27.b }, z9.b[0]\n"
+      ".inst 0xc151b1a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 10f\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b420  // sdot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n"
+      ".inst 0xc151b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 10f\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bba0  // sdot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n"
+      ".inst 0xc151b920  // sdot za.s[x9, 0], { z8.b-z11.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 10f\n"
-      ".inst 0xa0408371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159be20  // sdot za.s[x9, 0], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa0408349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bd20  // sdot za.s[x9, 0], { z8.b-z11.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "10:"  // Width 1: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 11f\n"
-      "sdot z11.s, z9.b, z8.b\n"
+      "sdot z28.s, z1.b, z29.b\n"
       "11:"  // Width 1: Multiply loop: unique 2: skip row sum
       "tbnz %x[flags], #31, 12f\n"
       "add x21, %x[qp], %[b_offset]\n"
       "mov x20, #0x4\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
+      "ld1rw { z26.s }, p2/Z, [x21]\n"
+      "neg z26.s, p2/M, z26.s\n"
       "whilelt p0.s, XZR, x20\n"
+      "saddv d28, p0, z28.s\n"
+      "mov z28.s, z28.s[0]\n"
+      "mul z28.s, p2/M, z28.s, z26.s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
-      "saddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "neg z17.s, p2/M, z17.s\n"
-      "mul z11.s, p2/M, z11.s, z17.s\n"
       "12:"  // Width 1: skip row sum fixup
-      ".inst 0xc0904960  // addha za0.s, p2/M, p2/M, z11.s\n"
-      "add x21, %x[qp], %[per_layer_mul]\n"
+      ".inst 0xc0904b80  // addha za0.s, p2/M, p2/M, z28.s\n"
+      "add x20, %x[qp], %[per_layer_mul]\n"
+      "ld1rw { z1.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      ".inst 0xc0904961  // addha za1.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z10.s }, p2/Z, [x21]\n"
-      "add x22, %x[qp], %[c_offset]\n"
+      ".inst 0xc0904b81  // addha za1.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z7.s }, p2/Z, [x20]\n"
+      "add x20, %x[qp], %[c_offset]\n"
       "add x21, %x[qp], %[minval]\n"
-      ".inst 0xc0904962  // addha za2.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z6.s }, p2/Z, [x20]\n"
+      ".inst 0xc0904b82  // addha za2.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z2.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[maxval]\n"
-      ".inst 0xc0904963  // addha za3.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z7.s }, p2/Z, [x22]\n"
-      "ld1rw { z30.s }, p2/Z, [x21]\n"
-      "ld1rw { z23.s }, p2/Z, [x20]\n"
+      ".inst 0xc0904b83  // addha za3.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z16.s }, p2/Z, [x21]\n"
       ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
-      ".inst 0xc1aaac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z10.s\n"
-      ".inst 0xc1a6aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n"
-      ".inst 0xc1a7ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n"
-      ".inst 0xc1b7cfcc  // sclamp { z12.s-z15.s }, z30.s, z23.s\n"
+      ".inst 0xc1a1ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n"
+      ".inst 0xc1a7aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n"
+      "ld1rw { z30.s }, p2/Z, [x20]\n"
+      ".inst 0xc1a2ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
+      ".inst 0xc1bece0c  // sclamp { z12.s-z15.s }, z16.s, z30.s\n"
       "uzp1 z12.h, z12.h, z13.h\n"
       "uzp1 z19.h, z14.h, z15.h\n"
       "uzp1 z12.b, z12.b, z19.b\n"
@@ -199,10 +199,10 @@ void sme2_gemv_s8qa_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       "whilelt p1.b, XZR, x20\n"
       "cbz x24, 15f\n"
-      ".inst 0xa040c304  // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n"
-      ".inst 0xa041c300  // ld1w { z0.s-z3.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
-      ".inst 0xc0042c01  // mova za.d[x9, #1], { z0.d-z3.d }\n"
+      ".inst 0xa040c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24]\n"
+      ".inst 0xc0042e00  // mova za.d[x9, #0], { z16.d-z19.d }\n"
+      ".inst 0xa041c318  // ld1w { z24.s-z27.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042f01  // mova za.d[x9, #1], { z24.d-z27.d }\n"
       "b 16f\n"
       "15:"  // Width 2: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -211,111 +211,111 @@ void sme2_gemv_s8qa_dot_16VL (
       "ble 19f\n"
       "17:"  // Width 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa0408371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b2a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z9.b[0]\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b321  // sdot za.s[x9, 1], { z24.b-z27.b }, z9.b[0]\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b620  // sdot za.s[x9, 0], { z16.b-z19.b }, z9.b[1]\n"
-      ".inst 0xa0418371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b6a1  // sdot za.s[x9, 1], { z20.b-z23.b }, z9.b[1]\n"
-      ".inst 0xc159bba0  // sdot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xc159b821  // sdot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xc159bda0  // sdot za.s[x9, 0], { z12.b-z15.b }, z9.b[3]\n"
-      ".inst 0xc159be21  // sdot za.s[x9, 1], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b1a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n"
+      ".inst 0xa0418359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b321  // sdot za.s[x9, 1], { z24.b-z27.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b620  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa0418355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b6a1  // sdot za.s[x9, 1], { z20.b-z23.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b9a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b9a1  // sdot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bca0  // sdot za.s[x9, 0], { z4.b-z7.b }, z1.b[3]\n"
+      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151bd21  // sdot za.s[x9, 1], { z8.b-z11.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "tbnz %x[flags], #31, 18f\n"
-      "sdot z11.s, z9.b, z8.b\n"
+      "sdot z28.s, z1.b, z29.b\n"
       "18:"  // Width 2: Multiply loop: unique 3: skip row sum
       "sub x22, x22, #0x10\n"
       "cmp x22, #0x10\n"
       "bgt 17b\n"
       "19:"  // Width 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x22, x22, #0x4\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      ".inst 0xa0408359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b020  // sdot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n"
-      ".inst 0xc159b321  // sdot za.s[x9, 1], { z24.b-z27.b }, z9.b[0]\n"
+      ".inst 0xc151b320  // sdot za.s[x9, 0], { z24.b-z27.b }, z1.b[0]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b221  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 20f\n"
-      ".inst 0xa0408371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b620  // sdot za.s[x9, 0], { z16.b-z19.b }, z9.b[1]\n"
-      ".inst 0xc159b721  // sdot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n"
+      ".inst 0xc151b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z1.b[1]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b5a1  // sdot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 20f\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bba0  // sdot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xc159b821  // sdot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
+      ".inst 0xc151b9a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b9a1  // sdot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 20f\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159bf21  // sdot za.s[x9, 1], { z24.b-z27.b }, z9.b[3]\n"
+      ".inst 0xa0408359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bf20  // sdot za.s[x9, 0], { z24.b-z27.b }, z1.b[3]\n"
+      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151bd21  // sdot za.s[x9, 1], { z8.b-z11.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "20:"  // Width 2: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 21f\n"
-      "sdot z11.s, z9.b, z8.b\n"
+      "sdot z28.s, z1.b, z29.b\n"
       "21:"  // Width 2: Multiply loop: unique 4: skip row sum
       "tbnz %x[flags], #31, 22f\n"
       "add x21, %x[qp], %[b_offset]\n"
       "mov x20, #0x4\n"
-      "ld1rw { z1.s }, p2/Z, [x21]\n"
+      "ld1rw { z16.s }, p2/Z, [x21]\n"
+      "neg z16.s, p2/M, z16.s\n"
       "whilelt p0.s, XZR, x20\n"
+      "saddv d28, p0, z28.s\n"
+      "mov z28.s, z28.s[0]\n"
+      "mul z28.s, p2/M, z28.s, z16.s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
-      "saddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "neg z1.s, p2/M, z1.s\n"
-      "mul z11.s, p2/M, z11.s, z1.s\n"
       "22:"  // Width 2: skip row sum fixup
-      ".inst 0xc0904960  // addha za0.s, p2/M, p2/M, z11.s\n"
-      "add x21, %x[qp], %[per_layer_mul]\n"
+      ".inst 0xc0904b80  // addha za0.s, p2/M, p2/M, z28.s\n"
+      "add x20, %x[qp], %[per_layer_mul]\n"
+      "ld1rw { z6.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      ".inst 0xc0904961  // addha za1.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z2.s }, p2/Z, [x21]\n"
-      "add x22, %x[qp], %[c_offset]\n"
+      ".inst 0xc0904b81  // addha za1.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z5.s }, p2/Z, [x20]\n"
+      "add x20, %x[qp], %[c_offset]\n"
       "add x21, %x[qp], %[minval]\n"
-      ".inst 0xc0904962  // addha za2.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z3.s }, p2/Z, [x20]\n"
+      ".inst 0xc0904b82  // addha za2.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z9.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[maxval]\n"
-      ".inst 0xc0904963  // addha za3.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z9.s }, p2/Z, [x22]\n"
-      "ld1rw { z6.s }, p2/Z, [x21]\n"
-      "ld1rw { z29.s }, p2/Z, [x20]\n"
-      ".inst 0xc0062c14  // mova { z20.d-z23.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      ".inst 0xc1a2ac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z2.s\n"
-      ".inst 0xc1a2ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
-      ".inst 0xc1a3aa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z3.s\n"
-      ".inst 0xc1a3aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z3.s\n"
-      ".inst 0xc1a9ab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z9.s\n"
-      ".inst 0xc1a9ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z9.s\n"
-      ".inst 0xc1bdccd4  // sclamp { z20.s-z23.s }, z6.s, z29.s\n"
-      ".inst 0xc1bdcccc  // sclamp { z12.s-z15.s }, z6.s, z29.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z16.h, z22.h, z23.h\n"
-      "uzp1 z12.h, z12.h, z13.h\n"
-      "uzp1 z24.h, z14.h, z15.h\n"
-      "uzp1 z20.b, z20.b, z16.b\n"
-      "uzp1 z12.b, z12.b, z24.b\n"
-      "st1b { z20.b }, p2, [x25]\n"
-      "st1b { z12.b }, p1, [x25, #1, MUL VL]\n"
+      ".inst 0xc0904b83  // addha za3.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z16.s }, p2/Z, [x21]\n"
+      ".inst 0xc0062c18  // mova { z24.d-z27.d }, za.d[x9, #0]\n"
+      ".inst 0xc1a6ac18  // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n"
+      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
+      ".inst 0xc1a6ac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n"
+      ".inst 0xc1a5aa38  // srshl { z24.s-z27.s }, { z24.s-z27.s }, z5.s\n"
+      "ld1rw { z21.s }, p2/Z, [x20]\n"
+      ".inst 0xc1a5aa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z5.s\n"
+      ".inst 0xc1a9ab18  // add { z24.s-z27.s }, { z24.s-z27.s }, z9.s\n"
+      ".inst 0xc1a9ab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z9.s\n"
+      ".inst 0xc1b5ce18  // sclamp { z24.s-z27.s }, z16.s, z21.s\n"
+      ".inst 0xc1b5ce00  // sclamp { z0.s-z3.s }, z16.s, z21.s\n"
+      "uzp1 z24.h, z24.h, z25.h\n"
+      "uzp1 z9.h, z26.h, z27.h\n"
+      "uzp1 z0.h, z0.h, z1.h\n"
+      "uzp1 z26.h, z2.h, z3.h\n"
+      "uzp1 z24.b, z24.b, z9.b\n"
+      "st1b { z24.b }, p2, [x25]\n"
+      "uzp1 z0.b, z0.b, z26.b\n"
+      "st1b { z0.b }, p1, [x25, #1, MUL VL]\n"
       "addvl x25, x25, #2\n"
       "23:"  // Width 2: Output done
       "b 44f\n"
@@ -328,12 +328,12 @@ void sme2_gemv_s8qa_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       "whilelt p1.b, XZR, x20\n"
       "cbz x24, 25f\n"
-      ".inst 0xa040c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24]\n"
-      ".inst 0xa041c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042c30c  // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xc0042e80  // mova za.d[x9, #0], { z20.d-z23.d }\n"
-      ".inst 0xc0042e01  // mova za.d[x9, #1], { z16.d-z19.d }\n"
-      ".inst 0xc0042d82  // mova za.d[x9, #2], { z12.d-z15.d }\n"
+      ".inst 0xa040c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24]\n"
+      ".inst 0xc0042e00  // mova za.d[x9, #0], { z16.d-z19.d }\n"
+      ".inst 0xa041c30c  // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042d81  // mova za.d[x9, #1], { z12.d-z15.d }\n"
+      ".inst 0xa042c318  // ld1w { z24.s-z27.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n"
+      ".inst 0xc0042f02  // mova za.d[x9, #2], { z24.d-z27.d }\n"
       "b 26f\n"
       "25:"  // Width 3: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -342,136 +342,136 @@ void sme2_gemv_s8qa_dot_16VL (
       "ble 29f\n"
       "27:"  // Width 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b3a0  // sdot za.s[x9, 0], { z28.b-z31.b }, z9.b[0]\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b221  // sdot za.s[x9, 1], { z16.b-z19.b }, z9.b[0]\n"
-      ".inst 0xa041837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b1a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z9.b[0]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b720  // sdot za.s[x9, 0], { z24.b-z27.b }, z9.b[1]\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b7a1  // sdot za.s[x9, 1], { z28.b-z31.b }, z9.b[1]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b9a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z9.b[2]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b821  // sdot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xc159bba2  // sdot za.s[x9, 2], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xc159bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159bf21  // sdot za.s[x9, 1], { z24.b-z27.b }, z9.b[3]\n"
-      ".inst 0xc159be22  // sdot za.s[x9, 2], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b221  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b1a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b5a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b5a1  // sdot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa0428355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b6a2  // sdot za.s[x9, 2], { z20.b-z23.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b920  // sdot za.s[x9, 0], { z8.b-z11.b }, z1.b[2]\n"
+      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b8a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z1.b[2]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151ba22  // sdot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bf20  // sdot za.s[x9, 0], { z24.b-z27.b }, z1.b[3]\n"
+      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151bca1  // sdot za.s[x9, 1], { z4.b-z7.b }, z1.b[3]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151be22  // sdot za.s[x9, 2], { z16.b-z19.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "tbnz %x[flags], #31, 28f\n"
-      "sdot z11.s, z9.b, z8.b\n"
+      "sdot z28.s, z1.b, z29.b\n"
       "28:"  // Width 3: Multiply loop: unique 5: skip row sum
       "sub x22, x22, #0x10\n"
       "cmp x22, #0x10\n"
       "bgt 27b\n"
       "29:"  // Width 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa040836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x22, x22, #0x4\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b1a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z9.b[0]\n"
-      ".inst 0xc159b021  // sdot za.s[x9, 1], { z0.b-z3.b }, z9.b[0]\n"
-      ".inst 0xc159b3a2  // sdot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n"
+      ".inst 0xc151b2a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z1.b[0]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b221  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b222  // sdot za.s[x9, 2], { z16.b-z19.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 30f\n"
-      ".inst 0xa0408365  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b4a0  // sdot za.s[x9, 0], { z4.b-z7.b }, z9.b[1]\n"
-      ".inst 0xc159b721  // sdot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n"
-      ".inst 0xc159b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n"
+      ".inst 0xc151b720  // sdot za.s[x9, 0], { z24.b-z27.b }, z1.b[1]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b5a1  // sdot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 30f\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bba0  // sdot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xc159b821  // sdot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xc159bb22  // sdot za.s[x9, 2], { z24.b-z27.b }, z9.b[2]\n"
+      ".inst 0xc151ba20  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151ba21  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0428355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151baa2  // sdot za.s[x9, 2], { z20.b-z23.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 30f\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bc20  // sdot za.s[x9, 0], { z0.b-z3.b }, z9.b[3]\n"
-      ".inst 0xc159bea1  // sdot za.s[x9, 1], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159be22  // sdot za.s[x9, 2], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bda0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151be21  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[3]\n"
+      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151bda2  // sdot za.s[x9, 2], { z12.b-z15.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "30:"  // Width 3: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 31f\n"
-      "sdot z11.s, z9.b, z8.b\n"
+      "sdot z28.s, z1.b, z29.b\n"
       "31:"  // Width 3: Multiply loop: unique 6: skip row sum
       "tbnz %x[flags], #31, 32f\n"
       "add x21, %x[qp], %[b_offset]\n"
       "mov x20, #0x4\n"
       "ld1rw { z16.s }, p2/Z, [x21]\n"
+      "neg z16.s, p2/M, z16.s\n"
       "whilelt p0.s, XZR, x20\n"
+      "saddv d28, p0, z28.s\n"
+      "mov z28.s, z28.s[0]\n"
+      "mul z28.s, p2/M, z28.s, z16.s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
-      "saddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "neg z16.s, p2/M, z16.s\n"
-      "mul z11.s, p2/M, z11.s, z16.s\n"
       "32:"  // Width 3: skip row sum fixup
-      ".inst 0xc0904960  // addha za0.s, p2/M, p2/M, z11.s\n"
-      "add x21, %x[qp], %[per_layer_mul]\n"
+      ".inst 0xc0904b80  // addha za0.s, p2/M, p2/M, z28.s\n"
+      "add x20, %x[qp], %[per_layer_mul]\n"
+      "ld1rw { z2.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      ".inst 0xc0904961  // addha za1.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z2.s }, p2/Z, [x21]\n"
-      "add x22, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[minval]\n"
-      ".inst 0xc0904962  // addha za2.s, p2/M, p2/M, z11.s\n"
+      ".inst 0xc0904b81  // addha za1.s, p2/M, p2/M, z28.s\n"
       "ld1rw { z1.s }, p2/Z, [x20]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "add x21, %x[qp], %[minval]\n"
+      ".inst 0xc0904b82  // addha za2.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z3.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[maxval]\n"
-      ".inst 0xc0904963  // addha za3.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z0.s }, p2/Z, [x22]\n"
-      "ld1rw { z21.s }, p2/Z, [x21]\n"
-      "ld1rw { z20.s }, p2/Z, [x20]\n"
-      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
-      ".inst 0xc1a2ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n"
+      ".inst 0xc0904b83  // addha za3.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z16.s }, p2/Z, [x21]\n"
+      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
+      ".inst 0xc1a2ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n"
+      ".inst 0xc0062c24  // mova { z4.d-z7.d }, za.d[x9, #1]\n"
+      ".inst 0xc1a2ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
       ".inst 0xc1a2ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
-      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
-      ".inst 0xc1a1aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n"
+      ".inst 0xc1a1aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n"
+      "ld1rw { z0.s }, p2/Z, [x20]\n"
+      ".inst 0xc1a1aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
       ".inst 0xc1a1aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n"
-      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
-      ".inst 0xc1a0ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n"
-      ".inst 0xc1a0ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n"
-      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4ceac  // sclamp { z12.s-z15.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4ceb0  // sclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "uzp1 z28.h, z28.h, z29.h\n"
-      "uzp1 z21.h, z30.h, z31.h\n"
+      ".inst 0xc1a3ab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z3.s\n"
+      ".inst 0xc1a3ab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n"
+      ".inst 0xc1a3ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z3.s\n"
+      ".inst 0xc1a0ce08  // sclamp { z8.s-z11.s }, z16.s, z0.s\n"
+      ".inst 0xc1a0ce04  // sclamp { z4.s-z7.s }, z16.s, z0.s\n"
+      "uzp1 z8.h, z8.h, z9.h\n"
+      ".inst 0xc1a0ce0c  // sclamp { z12.s-z15.s }, z16.s, z0.s\n"
+      "uzp1 z18.h, z10.h, z11.h\n"
+      "uzp1 z4.h, z4.h, z5.h\n"
+      "uzp1 z17.h, z6.h, z7.h\n"
       "uzp1 z12.h, z12.h, z13.h\n"
-      "uzp1 z20.h, z14.h, z15.h\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
-      "uzp1 z28.b, z28.b, z21.b\n"
-      "uzp1 z12.b, z12.b, z20.b\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z28.b }, p2, [x25]\n"
-      "st1b { z12.b }, p2, [x25, #1, MUL VL]\n"
-      "st1b { z16.b }, p1, [x25, #2, MUL VL]\n"
+      "uzp1 z16.h, z14.h, z15.h\n"
+      "uzp1 z8.b, z8.b, z18.b\n"
+      "st1b { z8.b }, p2, [x25]\n"
+      "uzp1 z4.b, z4.b, z17.b\n"
+      "st1b { z4.b }, p2, [x25, #1, MUL VL]\n"
+      "uzp1 z12.b, z12.b, z16.b\n"
+      "st1b { z12.b }, p1, [x25, #2, MUL VL]\n"
       "addvl x25, x25, #3\n"
       "33:"  // Width 3: Output done
       "b 44f\n"
@@ -484,15 +484,15 @@ void sme2_gemv_s8qa_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       "whilelt p1.b, XZR, x20\n"
       "cbz x24, 35f\n"
-      ".inst 0xa040c300  // ld1w { z0.s-z3.s }, pn8.b/Z, [x24]\n"
-      ".inst 0xa041c30c  // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xa040c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24]\n"
+      ".inst 0xc0042e80  // mova za.d[x9, #0], { z20.d-z23.d }\n"
+      ".inst 0xa041c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042e01  // mova za.d[x9, #1], { z16.d-z19.d }\n"
       ".inst 0xa042c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xa043c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0xc, MUL VL]\n"
-      ".inst 0xc0042c00  // mova za.d[x9, #0], { z0.d-z3.d }\n"
-      "addvl x24, x24, #16\n"
-      ".inst 0xc0042d81  // mova za.d[x9, #1], { z12.d-z15.d }\n"
       ".inst 0xc0042e02  // mova za.d[x9, #2], { z16.d-z19.d }\n"
-      ".inst 0xc0042e83  // mova za.d[x9, #3], { z20.d-z23.d }\n"
+      ".inst 0xa043c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0xc, MUL VL]\n"
+      ".inst 0xc0042e03  // mova za.d[x9, #3], { z16.d-z19.d }\n"
+      "addvl x24, x24, #16\n"
       "b 36f\n"
       "35:"  // Width 4: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -501,164 +501,164 @@ void sme2_gemv_s8qa_dot_16VL (
       "ble 39f\n"
       "37:"  // Width 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159b020  // sdot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b2a1  // sdot za.s[x9, 1], { z20.b-z23.b }, z9.b[0]\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b3a2  // sdot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n"
-      ".inst 0xa041837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b323  // sdot za.s[x9, 3], { z24.b-z27.b }, z9.b[0]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159b420  // sdot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b7a1  // sdot za.s[x9, 1], { z28.b-z31.b }, z9.b[1]\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n"
-      ".inst 0xa041837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b5a3  // sdot za.s[x9, 3], { z12.b-z15.b }, z9.b[1]\n"
-      ".inst 0xa0428361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z9.b[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bba1  // sdot za.s[x9, 1], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b822  // sdot za.s[x9, 2], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b9a3  // sdot za.s[x9, 3], { z12.b-z15.b }, z9.b[2]\n"
-      ".inst 0xa0428361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159bf20  // sdot za.s[x9, 0], { z24.b-z27.b }, z9.b[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bea1  // sdot za.s[x9, 1], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159bc22  // sdot za.s[x9, 2], { z0.b-z3.b }, z9.b[3]\n"
-      ".inst 0xc159be23  // sdot za.s[x9, 3], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b221  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b1a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z1.b[0]\n"
+      ".inst 0xa043834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151b1a3  // sdot za.s[x9, 3], { z12.b-z15.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b620  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b5a1  // sdot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151b623  // sdot za.s[x9, 3], { z16.b-z19.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151ba20  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b9a1  // sdot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151ba22  // sdot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151ba23  // sdot za.s[x9, 3], { z16.b-z19.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bda0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151bda1  // sdot za.s[x9, 1], { z12.b-z15.b }, z1.b[3]\n"
+      ".inst 0xa0428359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151bf22  // sdot za.s[x9, 2], { z24.b-z27.b }, z1.b[3]\n"
+      ".inst 0xa0438345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151bca3  // sdot za.s[x9, 3], { z4.b-z7.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "tbnz %x[flags], #31, 38f\n"
-      "sdot z11.s, z9.b, z8.b\n"
+      "sdot z28.s, z1.b, z29.b\n"
       "38:"  // Width 4: Multiply loop: unique 7: skip row sum
       "sub x22, x22, #0x10\n"
       "cmp x22, #0x10\n"
       "bgt 37b\n"
       "39:"  // Width 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x22, x22, #0x4\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159b020  // sdot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b2a1  // sdot za.s[x9, 1], { z20.b-z23.b }, z9.b[0]\n"
-      ".inst 0xc159b3a2  // sdot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n"
-      ".inst 0xc159b223  // sdot za.s[x9, 3], { z16.b-z19.b }, z9.b[0]\n"
+      ".inst 0xc151b1a0  // sdot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n"
+      ".inst 0xa0418359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b321  // sdot za.s[x9, 1], { z24.b-z27.b }, z1.b[0]\n"
+      ".inst 0xa0428349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b122  // sdot za.s[x9, 2], { z8.b-z11.b }, z1.b[0]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151b223  // sdot za.s[x9, 3], { z16.b-z19.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 40f\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159b420  // sdot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b721  // sdot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n"
-      ".inst 0xc159b7a2  // sdot za.s[x9, 2], { z28.b-z31.b }, z9.b[1]\n"
-      ".inst 0xc159b623  // sdot za.s[x9, 3], { z16.b-z19.b }, z9.b[1]\n"
+      ".inst 0xc151b620  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b621  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b5a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa0438355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151b6a3  // sdot za.s[x9, 3], { z20.b-z23.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 40f\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159bb20  // sdot za.s[x9, 0], { z24.b-z27.b }, z9.b[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b821  // sdot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xc159baa2  // sdot za.s[x9, 2], { z20.b-z23.b }, z9.b[2]\n"
-      ".inst 0xc159b9a3  // sdot za.s[x9, 3], { z12.b-z15.b }, z9.b[2]\n"
+      ".inst 0xc151ba20  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151ba21  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151ba22  // sdot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151ba23  // sdot za.s[x9, 3], { z16.b-z19.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 40f\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa041837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159bf20  // sdot za.s[x9, 0], { z24.b-z27.b }, z9.b[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bfa1  // sdot za.s[x9, 1], { z28.b-z31.b }, z9.b[3]\n"
-      ".inst 0xc159bea2  // sdot za.s[x9, 2], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159be23  // sdot za.s[x9, 3], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151be20  // sdot za.s[x9, 0], { z16.b-z19.b }, z1.b[3]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151be21  // sdot za.s[x9, 1], { z16.b-z19.b }, z1.b[3]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151be22  // sdot za.s[x9, 2], { z16.b-z19.b }, z1.b[3]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151be23  // sdot za.s[x9, 3], { z16.b-z19.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "40:"  // Width 4: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 41f\n"
-      "sdot z11.s, z9.b, z8.b\n"
+      "sdot z28.s, z1.b, z29.b\n"
       "41:"  // Width 4: Multiply loop: unique 8: skip row sum
       "tbnz %x[flags], #31, 42f\n"
       "add x21, %x[qp], %[b_offset]\n"
       "mov x20, #0x4\n"
       "ld1rw { z16.s }, p2/Z, [x21]\n"
+      "neg z16.s, p2/M, z16.s\n"
       "whilelt p0.s, XZR, x20\n"
+      "saddv d28, p0, z28.s\n"
+      "mov z28.s, z28.s[0]\n"
+      "mul z28.s, p2/M, z28.s, z16.s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
-      "saddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "neg z16.s, p2/M, z16.s\n"
-      "mul z11.s, p2/M, z11.s, z16.s\n"
       "42:"  // Width 4: skip row sum fixup
-      ".inst 0xc0904960  // addha za0.s, p2/M, p2/M, z11.s\n"
-      "add x21, %x[qp], %[per_layer_mul]\n"
+      ".inst 0xc0904b80  // addha za0.s, p2/M, p2/M, z28.s\n"
+      "add x20, %x[qp], %[per_layer_mul]\n"
+      "ld1rw { z11.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      ".inst 0xc0904961  // addha za1.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z10.s }, p2/Z, [x21]\n"
-      "add x22, %x[qp], %[c_offset]\n"
+      ".inst 0xc0904b81  // addha za1.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z7.s }, p2/Z, [x20]\n"
+      "add x20, %x[qp], %[c_offset]\n"
       "add x21, %x[qp], %[minval]\n"
-      ".inst 0xc0904962  // addha za2.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z12.s }, p2/Z, [x20]\n"
+      ".inst 0xc0904b82  // addha za2.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z6.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[maxval]\n"
-      ".inst 0xc0904963  // addha za3.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z13.s }, p2/Z, [x22]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
-      "ld1rw { z18.s }, p2/Z, [x20]\n"
-      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c58  // mova { z24.d-z27.d }, za.d[x9, #2]\n"
-      ".inst 0xc0062c74  // mova { z20.d-z23.d }, za.d[x9, #3]\n"
-      ".inst 0xc1aaac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z10.s\n"
-      ".inst 0xc1aaac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z10.s\n"
-      ".inst 0xc1aaac18  // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z10.s\n"
-      ".inst 0xc1aaac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z10.s\n"
-      ".inst 0xc1acaa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z12.s\n"
-      ".inst 0xc1acaa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z12.s\n"
-      ".inst 0xc1acaa38  // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n"
-      ".inst 0xc1acaa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z12.s\n"
-      ".inst 0xc1adab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z13.s\n"
-      ".inst 0xc1adab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z13.s\n"
-      ".inst 0xc1adab18  // add { z24.s-z27.s }, { z24.s-z27.s }, z13.s\n"
-      ".inst 0xc1adab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z13.s\n"
-      ".inst 0xc1b2ce3c  // sclamp { z28.s-z31.s }, z17.s, z18.s\n"
-      ".inst 0xc1b2ce20  // sclamp { z0.s-z3.s }, z17.s, z18.s\n"
-      ".inst 0xc1b2ce38  // sclamp { z24.s-z27.s }, z17.s, z18.s\n"
-      ".inst 0xc1b2ce34  // sclamp { z20.s-z23.s }, z17.s, z18.s\n"
-      "uzp1 z28.h, z28.h, z29.h\n"
-      "uzp1 z14.h, z30.h, z31.h\n"
-      "uzp1 z0.h, z0.h, z1.h\n"
-      "uzp1 z18.h, z2.h, z3.h\n"
+      ".inst 0xc0904b83  // addha za3.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z3.s }, p2/Z, [x21]\n"
+      ".inst 0xc0062c18  // mova { z24.d-z27.d }, za.d[x9, #0]\n"
+      ".inst 0xc1abac18  // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z11.s\n"
+      ".inst 0xc0062c30  // mova { z16.d-z19.d }, za.d[x9, #1]\n"
+      ".inst 0xc1abac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z11.s\n"
+      ".inst 0xc0062c54  // mova { z20.d-z23.d }, za.d[x9, #2]\n"
+      ".inst 0xc1abac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z11.s\n"
+      ".inst 0xc0062c6c  // mova { z12.d-z15.d }, za.d[x9, #3]\n"
+      ".inst 0xc1abac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z11.s\n"
+      ".inst 0xc1a7aa38  // srshl { z24.s-z27.s }, { z24.s-z27.s }, z7.s\n"
+      "ld1rw { z31.s }, p2/Z, [x20]\n"
+      ".inst 0xc1a7aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z7.s\n"
+      ".inst 0xc1a7aa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z7.s\n"
+      ".inst 0xc1a7aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n"
+      ".inst 0xc1a6ab18  // add { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n"
+      ".inst 0xc1a6ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n"
+      ".inst 0xc1a6ab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z6.s\n"
+      ".inst 0xc1a6ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n"
+      ".inst 0xc1bfcc78  // sclamp { z24.s-z27.s }, z3.s, z31.s\n"
+      ".inst 0xc1bfcc70  // sclamp { z16.s-z19.s }, z3.s, z31.s\n"
       "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
+      ".inst 0xc1bfcc74  // sclamp { z20.s-z23.s }, z3.s, z31.s\n"
+      ".inst 0xc1bfcc6c  // sclamp { z12.s-z15.s }, z3.s, z31.s\n"
+      "uzp1 z25.h, z26.h, z27.h\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
+      "uzp1 z18.h, z18.h, z19.h\n"
       "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z16.h, z22.h, z23.h\n"
-      "uzp1 z28.b, z28.b, z14.b\n"
-      "uzp1 z0.b, z0.b, z18.b\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "uzp1 z20.b, z20.b, z16.b\n"
-      "st1b { z28.b }, p2, [x25]\n"
-      "st1b { z0.b }, p2, [x25, #1, MUL VL]\n"
-      "st1b { z24.b }, p2, [x25, #2, MUL VL]\n"
-      "st1b { z20.b }, p1, [x25, #3, MUL VL]\n"
+      "uzp1 z17.h, z22.h, z23.h\n"
+      "uzp1 z12.h, z12.h, z13.h\n"
+      "uzp1 z30.h, z14.h, z15.h\n"
+      "uzp1 z24.b, z24.b, z25.b\n"
+      "st1b { z24.b }, p2, [x25]\n"
+      "uzp1 z16.b, z16.b, z18.b\n"
+      "st1b { z16.b }, p2, [x25, #1, MUL VL]\n"
+      "uzp1 z20.b, z20.b, z17.b\n"
+      "uzp1 z12.b, z12.b, z30.b\n"
+      "st1b { z20.b }, p2, [x25, #2, MUL VL]\n"
+      "st1b { z12.b }, p1, [x25, #3, MUL VL]\n"
       "addvl x25, x25, #4\n"
       "43:"  // Width 4: Output done
-      "subs x26, x26, #0x4\n"
+      "subs x27, x27, #0x4\n"
       "sub %x[N], %x[N], x28, LSL #2\n"
       "bgt 4b\n"
       "44:"  // Exit
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp
index 1cbaf00052..9bf699462a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp
@@ -52,20 +52,20 @@ void sme2_gemv_u8qa_dot_16VL (
     __asm__ __volatile__(
       "ptrue p8.b\n"
       ".inst 0xd503477f  // SMSTART ZA\n"
-      "mov x9, #0x0\n"
       "cntw x28, ALL, MUL #4\n"
-      "mov x27, %x[B_ptr]\n"
-      "add x26, %x[N], x28\n"
+      "add x27, %x[N], x28\n"
+      "sub x27, x27, #0x1\n"
+      "udiv x27, x27, x28\n"
+      "add x22, x27, #0x3\n"
+      "and x22, x22, #0xfffffffffffffffc\n"
+      "mul x22, x22, x28\n"
+      "mov x9, #0x0\n"
+      "mov x26, %x[B_ptr]\n"
       "mov x25, %x[output_ptr]\n"
-      "sub x26, x26, #0x1\n"
       "ptrue p2.b\n"
-      "udiv x26, x26, x28\n"
       ".inst 0x25207810  // ptrue pn8.b\n"
-      "add x22, x26, #0x3\n"
-      "mov x21, #0x1\n"
-      "and x22, x22, #0xfffffffffffffffc\n"
-      "mul x22, x22, x28\n"
       "mul x22, x22, %x[K]\n"
+      "mov x21, #0x1\n"
       "1:"  // RHS size check loop
       "cmp x22, #0x200000\n"
       "blt 2f\n"
@@ -79,16 +79,16 @@ void sme2_gemv_u8qa_dot_16VL (
       "lsl x21, x21, #0x16\n"
       "orr x22, x22, x20\n"
       "orr x22, x22, x21\n"
-      ".inst 0xf8b64b7a  // rprfm pldonce, x22, [x27]\n"
+      ".inst 0xf8b64b5a  // rprfm pldonce, x22, [x26]\n"
       "3:"  // RHS prefetch exit
       "mov x24, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z8.b, #0x1\n"
+      "mov z28.s, #0x0\n"
+      "mov z29.b, #0x1\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "4:"  // Column loop
-      "cmp x26, #0x4\n"
+      "cmp x27, #0x4\n"
       "bge 34f\n"
-      "cmp x26, #0x2\n"
+      "cmp x27, #0x2\n"
       "bgt 24f\n"
       "beq 14f\n"
       "mov x23, %x[A_ptr]\n"
@@ -98,8 +98,8 @@ void sme2_gemv_u8qa_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       "whilelt p1.b, XZR, x20\n"
       "cbz x24, 5f\n"
-      ".inst 0xa040c30c  // ld1w { z12.s-z15.s }, pn8.b/Z, [x24]\n"
-      ".inst 0xc0042d80  // mova za.d[x9, #0], { z12.d-z15.d }\n"
+      ".inst 0xa040c300  // ld1w { z0.s-z3.s }, pn8.b/Z, [x24]\n"
+      ".inst 0xc0042c00  // mova za.d[x9, #0], { z0.d-z3.d }\n"
       "b 6f\n"
       "5:"  // Width 1: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -108,82 +108,82 @@ void sme2_gemv_u8qa_dot_16VL (
       "ble 9f\n"
       "7:"  // Width 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b030  // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b730  // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[1]\n"
-      ".inst 0xc159bab0  // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[2]\n"
-      ".inst 0xc159bfb0  // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[3]\n"
+      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b2b0  // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b5b0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b9b0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bdb0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "tbnz %x[flags], #31, 8f\n"
-      "udot z11.s, z9.b, z8.b\n"
+      "udot z28.s, z1.b, z29.b\n"
       "8:"  // Width 1: Multiply loop: unique 1: skip row sum
       "sub x22, x22, #0x10\n"
       "cmp x22, #0x10\n"
       "bgt 7b\n"
       "9:"  // Width 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x22, x22, #0x4\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b330  // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[0]\n"
+      ".inst 0xc151b1b0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 10f\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b430  // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n"
+      ".inst 0xc151b6b0  // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 10f\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bbb0  // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n"
+      ".inst 0xc151b930  // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 10f\n"
-      ".inst 0xa0408371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159be30  // udot za.s[x9, 0], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa0408349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bd30  // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "10:"  // Width 1: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 11f\n"
-      "udot z11.s, z9.b, z8.b\n"
+      "udot z28.s, z1.b, z29.b\n"
       "11:"  // Width 1: Multiply loop: unique 2: skip row sum
       "tbnz %x[flags], #31, 12f\n"
       "add x21, %x[qp], %[b_offset]\n"
       "mov x20, #0x4\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
+      "ld1rw { z26.s }, p2/Z, [x21]\n"
+      "neg z26.s, p2/M, z26.s\n"
       "whilelt p0.s, XZR, x20\n"
+      "uaddv d28, p0, z28.s\n"
+      "mov z28.s, z28.s[0]\n"
+      "mul z28.s, p2/M, z28.s, z26.s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
-      "uaddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "neg z17.s, p2/M, z17.s\n"
-      "mul z11.s, p2/M, z11.s, z17.s\n"
       "12:"  // Width 1: skip row sum fixup
-      ".inst 0xc0904960  // addha za0.s, p2/M, p2/M, z11.s\n"
-      "add x21, %x[qp], %[per_layer_mul]\n"
+      ".inst 0xc0904b80  // addha za0.s, p2/M, p2/M, z28.s\n"
+      "add x20, %x[qp], %[per_layer_mul]\n"
+      "ld1rw { z1.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      ".inst 0xc0904961  // addha za1.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z10.s }, p2/Z, [x21]\n"
-      "add x22, %x[qp], %[c_offset]\n"
+      ".inst 0xc0904b81  // addha za1.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z7.s }, p2/Z, [x20]\n"
+      "add x20, %x[qp], %[c_offset]\n"
       "add x21, %x[qp], %[minval]\n"
-      ".inst 0xc0904962  // addha za2.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z6.s }, p2/Z, [x20]\n"
+      ".inst 0xc0904b82  // addha za2.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z2.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[maxval]\n"
-      ".inst 0xc0904963  // addha za3.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z7.s }, p2/Z, [x22]\n"
-      "ld1rw { z30.s }, p2/Z, [x21]\n"
-      "ld1rw { z23.s }, p2/Z, [x20]\n"
+      ".inst 0xc0904b83  // addha za3.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z16.s }, p2/Z, [x21]\n"
       ".inst 0xc0062c0c  // mova { z12.d-z15.d }, za.d[x9, #0]\n"
-      ".inst 0xc1aaac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z10.s\n"
-      ".inst 0xc1a6aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n"
-      ".inst 0xc1a7ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n"
-      ".inst 0xc1b7cfcc  // sclamp { z12.s-z15.s }, z30.s, z23.s\n"
+      ".inst 0xc1a1ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n"
+      ".inst 0xc1a7aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n"
+      "ld1rw { z30.s }, p2/Z, [x20]\n"
+      ".inst 0xc1a2ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
+      ".inst 0xc1bece0c  // sclamp { z12.s-z15.s }, z16.s, z30.s\n"
       "uzp1 z12.h, z12.h, z13.h\n"
       "uzp1 z19.h, z14.h, z15.h\n"
       "uzp1 z12.b, z12.b, z19.b\n"
@@ -199,10 +199,10 @@ void sme2_gemv_u8qa_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       "whilelt p1.b, XZR, x20\n"
       "cbz x24, 15f\n"
-      ".inst 0xa040c304  // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n"
-      ".inst 0xa041c300  // ld1w { z0.s-z3.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
-      ".inst 0xc0042c01  // mova za.d[x9, #1], { z0.d-z3.d }\n"
+      ".inst 0xa040c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24]\n"
+      ".inst 0xc0042e00  // mova za.d[x9, #0], { z16.d-z19.d }\n"
+      ".inst 0xa041c318  // ld1w { z24.s-z27.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042f01  // mova za.d[x9, #1], { z24.d-z27.d }\n"
       "b 16f\n"
       "15:"  // Width 2: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -211,111 +211,111 @@ void sme2_gemv_u8qa_dot_16VL (
       "ble 19f\n"
       "17:"  // Width 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa0408371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b2b0  // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[0]\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b331  // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[0]\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b630  // udot za.s[x9, 0], { z16.b-z19.b }, z9.b[1]\n"
-      ".inst 0xa0418371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b6b1  // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[1]\n"
-      ".inst 0xc159bbb0  // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xc159b831  // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xc159bdb0  // udot za.s[x9, 0], { z12.b-z15.b }, z9.b[3]\n"
-      ".inst 0xc159be31  // udot za.s[x9, 1], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b1b0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n"
+      ".inst 0xa0418359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b331  // udot za.s[x9, 1], { z24.b-z27.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b630  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa0418355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b6b1  // udot za.s[x9, 1], { z20.b-z23.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b9b0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b9b1  // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bcb0  // udot za.s[x9, 0], { z4.b-z7.b }, z1.b[3]\n"
+      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151bd31  // udot za.s[x9, 1], { z8.b-z11.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "tbnz %x[flags], #31, 18f\n"
-      "udot z11.s, z9.b, z8.b\n"
+      "udot z28.s, z1.b, z29.b\n"
       "18:"  // Width 2: Multiply loop: unique 3: skip row sum
       "sub x22, x22, #0x10\n"
       "cmp x22, #0x10\n"
       "bgt 17b\n"
       "19:"  // Width 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x22, x22, #0x4\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      ".inst 0xa0408359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b030  // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n"
-      ".inst 0xc159b331  // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[0]\n"
+      ".inst 0xc151b330  // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[0]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b231  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 20f\n"
-      ".inst 0xa0408371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b630  // udot za.s[x9, 0], { z16.b-z19.b }, z9.b[1]\n"
-      ".inst 0xc159b731  // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n"
+      ".inst 0xc151b6b0  // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[1]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b5b1  // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 20f\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bbb0  // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xc159b831  // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
+      ".inst 0xc151b9b0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b9b1  // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 20f\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159beb0  // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159bf31  // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[3]\n"
+      ".inst 0xa0408359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bf30  // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[3]\n"
+      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151bd31  // udot za.s[x9, 1], { z8.b-z11.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "20:"  // Width 2: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 21f\n"
-      "udot z11.s, z9.b, z8.b\n"
+      "udot z28.s, z1.b, z29.b\n"
       "21:"  // Width 2: Multiply loop: unique 4: skip row sum
       "tbnz %x[flags], #31, 22f\n"
       "add x21, %x[qp], %[b_offset]\n"
       "mov x20, #0x4\n"
-      "ld1rw { z1.s }, p2/Z, [x21]\n"
+      "ld1rw { z16.s }, p2/Z, [x21]\n"
+      "neg z16.s, p2/M, z16.s\n"
       "whilelt p0.s, XZR, x20\n"
+      "uaddv d28, p0, z28.s\n"
+      "mov z28.s, z28.s[0]\n"
+      "mul z28.s, p2/M, z28.s, z16.s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
-      "uaddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "neg z1.s, p2/M, z1.s\n"
-      "mul z11.s, p2/M, z11.s, z1.s\n"
       "22:"  // Width 2: skip row sum fixup
-      ".inst 0xc0904960  // addha za0.s, p2/M, p2/M, z11.s\n"
-      "add x21, %x[qp], %[per_layer_mul]\n"
+      ".inst 0xc0904b80  // addha za0.s, p2/M, p2/M, z28.s\n"
+      "add x20, %x[qp], %[per_layer_mul]\n"
+      "ld1rw { z6.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      ".inst 0xc0904961  // addha za1.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z2.s }, p2/Z, [x21]\n"
-      "add x22, %x[qp], %[c_offset]\n"
+      ".inst 0xc0904b81  // addha za1.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z5.s }, p2/Z, [x20]\n"
+      "add x20, %x[qp], %[c_offset]\n"
       "add x21, %x[qp], %[minval]\n"
-      ".inst 0xc0904962  // addha za2.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z3.s }, p2/Z, [x20]\n"
+      ".inst 0xc0904b82  // addha za2.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z9.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[maxval]\n"
-      ".inst 0xc0904963  // addha za3.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z9.s }, p2/Z, [x22]\n"
-      "ld1rw { z6.s }, p2/Z, [x21]\n"
-      "ld1rw { z29.s }, p2/Z, [x20]\n"
-      ".inst 0xc0062c14  // mova { z20.d-z23.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      ".inst 0xc1a2ac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z2.s\n"
-      ".inst 0xc1a2ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
-      ".inst 0xc1a3aa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z3.s\n"
-      ".inst 0xc1a3aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z3.s\n"
-      ".inst 0xc1a9ab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z9.s\n"
-      ".inst 0xc1a9ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z9.s\n"
-      ".inst 0xc1bdccd4  // sclamp { z20.s-z23.s }, z6.s, z29.s\n"
-      ".inst 0xc1bdcccc  // sclamp { z12.s-z15.s }, z6.s, z29.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z16.h, z22.h, z23.h\n"
-      "uzp1 z12.h, z12.h, z13.h\n"
-      "uzp1 z24.h, z14.h, z15.h\n"
-      "uzp1 z20.b, z20.b, z16.b\n"
-      "uzp1 z12.b, z12.b, z24.b\n"
-      "st1b { z20.b }, p2, [x25]\n"
-      "st1b { z12.b }, p1, [x25, #1, MUL VL]\n"
+      ".inst 0xc0904b83  // addha za3.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z16.s }, p2/Z, [x21]\n"
+      ".inst 0xc0062c18  // mova { z24.d-z27.d }, za.d[x9, #0]\n"
+      ".inst 0xc1a6ac18  // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n"
+      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
+      ".inst 0xc1a6ac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n"
+      ".inst 0xc1a5aa38  // srshl { z24.s-z27.s }, { z24.s-z27.s }, z5.s\n"
+      "ld1rw { z21.s }, p2/Z, [x20]\n"
+      ".inst 0xc1a5aa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z5.s\n"
+      ".inst 0xc1a9ab18  // add { z24.s-z27.s }, { z24.s-z27.s }, z9.s\n"
+      ".inst 0xc1a9ab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z9.s\n"
+      ".inst 0xc1b5ce18  // sclamp { z24.s-z27.s }, z16.s, z21.s\n"
+      ".inst 0xc1b5ce00  // sclamp { z0.s-z3.s }, z16.s, z21.s\n"
+      "uzp1 z24.h, z24.h, z25.h\n"
+      "uzp1 z9.h, z26.h, z27.h\n"
+      "uzp1 z0.h, z0.h, z1.h\n"
+      "uzp1 z26.h, z2.h, z3.h\n"
+      "uzp1 z24.b, z24.b, z9.b\n"
+      "st1b { z24.b }, p2, [x25]\n"
+      "uzp1 z0.b, z0.b, z26.b\n"
+      "st1b { z0.b }, p1, [x25, #1, MUL VL]\n"
       "addvl x25, x25, #2\n"
       "23:"  // Width 2: Output done
       "b 44f\n"
@@ -328,12 +328,12 @@ void sme2_gemv_u8qa_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       "whilelt p1.b, XZR, x20\n"
       "cbz x24, 25f\n"
-      ".inst 0xa040c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24]\n"
-      ".inst 0xa041c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
-      ".inst 0xa042c30c  // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xc0042e80  // mova za.d[x9, #0], { z20.d-z23.d }\n"
-      ".inst 0xc0042e01  // mova za.d[x9, #1], { z16.d-z19.d }\n"
-      ".inst 0xc0042d82  // mova za.d[x9, #2], { z12.d-z15.d }\n"
+      ".inst 0xa040c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24]\n"
+      ".inst 0xc0042e00  // mova za.d[x9, #0], { z16.d-z19.d }\n"
+      ".inst 0xa041c30c  // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042d81  // mova za.d[x9, #1], { z12.d-z15.d }\n"
+      ".inst 0xa042c318  // ld1w { z24.s-z27.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n"
+      ".inst 0xc0042f02  // mova za.d[x9, #2], { z24.d-z27.d }\n"
       "b 26f\n"
       "25:"  // Width 3: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -342,136 +342,136 @@ void sme2_gemv_u8qa_dot_16VL (
       "ble 29f\n"
       "27:"  // Width 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b3b0  // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[0]\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b231  // udot za.s[x9, 1], { z16.b-z19.b }, z9.b[0]\n"
-      ".inst 0xa041837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b1b2  // udot za.s[x9, 2], { z12.b-z15.b }, z9.b[0]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xa040836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b730  // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[1]\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b7b1  // udot za.s[x9, 1], { z28.b-z31.b }, z9.b[1]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b632  // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b9b0  // udot za.s[x9, 0], { z12.b-z15.b }, z9.b[2]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b831  // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xc159bbb2  // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xc159beb0  // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159bf31  // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[3]\n"
-      ".inst 0xc159be32  // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b230  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b231  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b1b2  // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b5b0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b5b1  // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa0428355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b6b2  // udot za.s[x9, 2], { z20.b-z23.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b930  // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[2]\n"
+      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b8b1  // udot za.s[x9, 1], { z4.b-z7.b }, z1.b[2]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151ba32  // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bf30  // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[3]\n"
+      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151bcb1  // udot za.s[x9, 1], { z4.b-z7.b }, z1.b[3]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151be32  // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "tbnz %x[flags], #31, 28f\n"
-      "udot z11.s, z9.b, z8.b\n"
+      "udot z28.s, z1.b, z29.b\n"
       "28:"  // Width 3: Multiply loop: unique 5: skip row sum
       "sub x22, x22, #0x10\n"
       "cmp x22, #0x10\n"
       "bgt 27b\n"
       "29:"  // Width 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa040836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x22, x22, #0x4\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b1b0  // udot za.s[x9, 0], { z12.b-z15.b }, z9.b[0]\n"
-      ".inst 0xc159b031  // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[0]\n"
-      ".inst 0xc159b3b2  // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n"
+      ".inst 0xc151b2b0  // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[0]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b231  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b232  // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 30f\n"
-      ".inst 0xa0408365  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b4b0  // udot za.s[x9, 0], { z4.b-z7.b }, z9.b[1]\n"
-      ".inst 0xc159b731  // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n"
-      ".inst 0xc159b632  // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n"
+      ".inst 0xc151b730  // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[1]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b5b1  // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b632  // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 30f\n"
-      ".inst 0xa040837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bbb0  // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xc159b831  // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xc159bb32  // udot za.s[x9, 2], { z24.b-z27.b }, z9.b[2]\n"
+      ".inst 0xc151ba30  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151ba31  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0428355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151bab2  // udot za.s[x9, 2], { z20.b-z23.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 30f\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bc30  // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[3]\n"
-      ".inst 0xc159beb1  // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159be32  // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bdb0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151be31  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[3]\n"
+      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151bdb2  // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "30:"  // Width 3: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 31f\n"
-      "udot z11.s, z9.b, z8.b\n"
+      "udot z28.s, z1.b, z29.b\n"
       "31:"  // Width 3: Multiply loop: unique 6: skip row sum
       "tbnz %x[flags], #31, 32f\n"
       "add x21, %x[qp], %[b_offset]\n"
       "mov x20, #0x4\n"
       "ld1rw { z16.s }, p2/Z, [x21]\n"
+      "neg z16.s, p2/M, z16.s\n"
       "whilelt p0.s, XZR, x20\n"
+      "uaddv d28, p0, z28.s\n"
+      "mov z28.s, z28.s[0]\n"
+      "mul z28.s, p2/M, z28.s, z16.s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
-      "uaddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "neg z16.s, p2/M, z16.s\n"
-      "mul z11.s, p2/M, z11.s, z16.s\n"
       "32:"  // Width 3: skip row sum fixup
-      ".inst 0xc0904960  // addha za0.s, p2/M, p2/M, z11.s\n"
-      "add x21, %x[qp], %[per_layer_mul]\n"
+      ".inst 0xc0904b80  // addha za0.s, p2/M, p2/M, z28.s\n"
+      "add x20, %x[qp], %[per_layer_mul]\n"
+      "ld1rw { z2.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      ".inst 0xc0904961  // addha za1.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z2.s }, p2/Z, [x21]\n"
-      "add x22, %x[qp], %[c_offset]\n"
-      "add x21, %x[qp], %[minval]\n"
-      ".inst 0xc0904962  // addha za2.s, p2/M, p2/M, z11.s\n"
+      ".inst 0xc0904b81  // addha za1.s, p2/M, p2/M, z28.s\n"
       "ld1rw { z1.s }, p2/Z, [x20]\n"
+      "add x20, %x[qp], %[c_offset]\n"
+      "add x21, %x[qp], %[minval]\n"
+      ".inst 0xc0904b82  // addha za2.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z3.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[maxval]\n"
-      ".inst 0xc0904963  // addha za3.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z0.s }, p2/Z, [x22]\n"
-      "ld1rw { z21.s }, p2/Z, [x21]\n"
-      "ld1rw { z20.s }, p2/Z, [x20]\n"
-      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c50  // mova { z16.d-z19.d }, za.d[x9, #2]\n"
-      ".inst 0xc1a2ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n"
+      ".inst 0xc0904b83  // addha za3.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z16.s }, p2/Z, [x21]\n"
+      ".inst 0xc0062c08  // mova { z8.d-z11.d }, za.d[x9, #0]\n"
+      ".inst 0xc1a2ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n"
+      ".inst 0xc0062c24  // mova { z4.d-z7.d }, za.d[x9, #1]\n"
+      ".inst 0xc1a2ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      ".inst 0xc0062c4c  // mova { z12.d-z15.d }, za.d[x9, #2]\n"
       ".inst 0xc1a2ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
-      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
-      ".inst 0xc1a1aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n"
+      ".inst 0xc1a1aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n"
+      "ld1rw { z0.s }, p2/Z, [x20]\n"
+      ".inst 0xc1a1aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
       ".inst 0xc1a1aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n"
-      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
-      ".inst 0xc1a0ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n"
-      ".inst 0xc1a0ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n"
-      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4ceac  // sclamp { z12.s-z15.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4ceb0  // sclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "uzp1 z28.h, z28.h, z29.h\n"
-      "uzp1 z21.h, z30.h, z31.h\n"
+      ".inst 0xc1a3ab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z3.s\n"
+      ".inst 0xc1a3ab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n"
+      ".inst 0xc1a3ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z3.s\n"
+      ".inst 0xc1a0ce08  // sclamp { z8.s-z11.s }, z16.s, z0.s\n"
+      ".inst 0xc1a0ce04  // sclamp { z4.s-z7.s }, z16.s, z0.s\n"
+      "uzp1 z8.h, z8.h, z9.h\n"
+      ".inst 0xc1a0ce0c  // sclamp { z12.s-z15.s }, z16.s, z0.s\n"
+      "uzp1 z18.h, z10.h, z11.h\n"
+      "uzp1 z4.h, z4.h, z5.h\n"
+      "uzp1 z17.h, z6.h, z7.h\n"
       "uzp1 z12.h, z12.h, z13.h\n"
-      "uzp1 z20.h, z14.h, z15.h\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
-      "uzp1 z28.b, z28.b, z21.b\n"
-      "uzp1 z12.b, z12.b, z20.b\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z28.b }, p2, [x25]\n"
-      "st1b { z12.b }, p2, [x25, #1, MUL VL]\n"
-      "st1b { z16.b }, p1, [x25, #2, MUL VL]\n"
+      "uzp1 z16.h, z14.h, z15.h\n"
+      "uzp1 z8.b, z8.b, z18.b\n"
+      "st1b { z8.b }, p2, [x25]\n"
+      "uzp1 z4.b, z4.b, z17.b\n"
+      "st1b { z4.b }, p2, [x25, #1, MUL VL]\n"
+      "uzp1 z12.b, z12.b, z16.b\n"
+      "st1b { z12.b }, p1, [x25, #2, MUL VL]\n"
       "addvl x25, x25, #3\n"
       "33:"  // Width 3: Output done
       "b 44f\n"
@@ -484,15 +484,15 @@ void sme2_gemv_u8qa_dot_16VL (
       ".inst 0xf8b54af8  // rprfm pldmany, x21, [x23]\n"
       "whilelt p1.b, XZR, x20\n"
       "cbz x24, 35f\n"
-      ".inst 0xa040c300  // ld1w { z0.s-z3.s }, pn8.b/Z, [x24]\n"
-      ".inst 0xa041c30c  // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xa040c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24]\n"
+      ".inst 0xc0042e80  // mova za.d[x9, #0], { z20.d-z23.d }\n"
+      ".inst 0xa041c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
+      ".inst 0xc0042e01  // mova za.d[x9, #1], { z16.d-z19.d }\n"
       ".inst 0xa042c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n"
-      ".inst 0xa043c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0xc, MUL VL]\n"
-      ".inst 0xc0042c00  // mova za.d[x9, #0], { z0.d-z3.d }\n"
-      "addvl x24, x24, #16\n"
-      ".inst 0xc0042d81  // mova za.d[x9, #1], { z12.d-z15.d }\n"
       ".inst 0xc0042e02  // mova za.d[x9, #2], { z16.d-z19.d }\n"
-      ".inst 0xc0042e83  // mova za.d[x9, #3], { z20.d-z23.d }\n"
+      ".inst 0xa043c310  // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0xc, MUL VL]\n"
+      ".inst 0xc0042e03  // mova za.d[x9, #3], { z16.d-z19.d }\n"
+      "addvl x24, x24, #16\n"
       "b 36f\n"
       "35:"  // Width 4: no bias
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
@@ -501,164 +501,164 @@ void sme2_gemv_u8qa_dot_16VL (
       "ble 39f\n"
       "37:"  // Width 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159b030  // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b2b1  // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[0]\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b3b2  // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n"
-      ".inst 0xa041837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b333  // udot za.s[x9, 3], { z24.b-z27.b }, z9.b[0]\n"
-      ".inst 0xa0428371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159b430  // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b7b1  // udot za.s[x9, 1], { z28.b-z31.b }, z9.b[1]\n"
-      ".inst 0xa0408375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b632  // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n"
-      ".inst 0xa041837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b5b3  // udot za.s[x9, 3], { z12.b-z15.b }, z9.b[1]\n"
-      ".inst 0xa0428361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159bab0  // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bbb1  // udot za.s[x9, 1], { z28.b-z31.b }, z9.b[2]\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xc159b832  // udot za.s[x9, 2], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xc159b9b3  // udot za.s[x9, 3], { z12.b-z15.b }, z9.b[2]\n"
-      ".inst 0xa0428361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159bf30  // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159beb1  // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159bc32  // udot za.s[x9, 2], { z0.b-z3.b }, z9.b[3]\n"
-      ".inst 0xc159be33  // udot za.s[x9, 3], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b230  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b231  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n"
+      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b1b2  // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[0]\n"
+      ".inst 0xa043834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151b1b3  // udot za.s[x9, 3], { z12.b-z15.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151b630  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b5b1  // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b632  // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151b633  // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151ba30  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b9b1  // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151ba32  // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151ba33  // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151bdb0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n"
+      ".inst 0xa041834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151bdb1  // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[3]\n"
+      ".inst 0xa0428359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151bf32  // udot za.s[x9, 2], { z24.b-z27.b }, z1.b[3]\n"
+      ".inst 0xa0438345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151bcb3  // udot za.s[x9, 3], { z4.b-z7.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "tbnz %x[flags], #31, 38f\n"
-      "udot z11.s, z9.b, z8.b\n"
+      "udot z28.s, z1.b, z29.b\n"
       "38:"  // Width 4: Multiply loop: unique 7: skip row sum
       "sub x22, x22, #0x10\n"
       "cmp x22, #0x10\n"
       "bgt 37b\n"
       "39:"  // Width 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x22\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x22, x22, #0x4\n"
-      "ld1rqb { z9.b }, p0/Z, [x23]\n"
+      ".inst 0xa040834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n"
       "add x23, x23, #0x10\n"
-      ".inst 0xa0418375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159b030  // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b2b1  // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[0]\n"
-      ".inst 0xc159b3b2  // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n"
-      ".inst 0xc159b233  // udot za.s[x9, 3], { z16.b-z19.b }, z9.b[0]\n"
+      ".inst 0xc151b1b0  // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n"
+      ".inst 0xa0418359  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b331  // udot za.s[x9, 1], { z24.b-z27.b }, z1.b[0]\n"
+      ".inst 0xa0428349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b132  // udot za.s[x9, 2], { z8.b-z11.b }, z1.b[0]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151b233  // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[0]\n"
+      "addvl x26, x26, #16\n"
       "ble 40f\n"
-      ".inst 0xa0408361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa042837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159b430  // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b731  // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n"
-      ".inst 0xc159b7b2  // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[1]\n"
-      ".inst 0xc159b633  // udot za.s[x9, 3], { z16.b-z19.b }, z9.b[1]\n"
+      ".inst 0xc151b630  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151b631  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[1]\n"
+      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151b5b2  // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[1]\n"
+      ".inst 0xa0438355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151b6b3  // udot za.s[x9, 3], { z20.b-z23.b }, z1.b[1]\n"
+      "addvl x26, x26, #16\n"
       "ble 40f\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
       "subs x22, x22, #0x4\n"
-      ".inst 0xa0418361  // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa043836d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159bb30  // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[2]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159b831  // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n"
-      ".inst 0xc159bab2  // udot za.s[x9, 2], { z20.b-z23.b }, z9.b[2]\n"
-      ".inst 0xc159b9b3  // udot za.s[x9, 3], { z12.b-z15.b }, z9.b[2]\n"
+      ".inst 0xc151ba30  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151ba31  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151ba32  // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151ba33  // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[2]\n"
+      "addvl x26, x26, #16\n"
       "ble 40f\n"
-      ".inst 0xa0408379  // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa041837d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0428375  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      ".inst 0xa0438371  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
-      ".inst 0xc159bf30  // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[3]\n"
-      "addvl x27, x27, #16\n"
-      ".inst 0xc159bfb1  // udot za.s[x9, 1], { z28.b-z31.b }, z9.b[3]\n"
-      ".inst 0xc159beb2  // udot za.s[x9, 2], { z20.b-z23.b }, z9.b[3]\n"
-      ".inst 0xc159be33  // udot za.s[x9, 3], { z16.b-z19.b }, z9.b[3]\n"
+      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
+      ".inst 0xc151be30  // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[3]\n"
+      ".inst 0xa0418351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
+      ".inst 0xc151be31  // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[3]\n"
+      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
+      ".inst 0xc151be32  // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[3]\n"
+      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
+      ".inst 0xc151be33  // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[3]\n"
+      "addvl x26, x26, #16\n"
       "40:"  // Width 4: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 41f\n"
-      "udot z11.s, z9.b, z8.b\n"
+      "udot z28.s, z1.b, z29.b\n"
       "41:"  // Width 4: Multiply loop: unique 8: skip row sum
       "tbnz %x[flags], #31, 42f\n"
       "add x21, %x[qp], %[b_offset]\n"
       "mov x20, #0x4\n"
       "ld1rw { z16.s }, p2/Z, [x21]\n"
+      "neg z16.s, p2/M, z16.s\n"
       "whilelt p0.s, XZR, x20\n"
+      "uaddv d28, p0, z28.s\n"
+      "mov z28.s, z28.s[0]\n"
+      "mul z28.s, p2/M, z28.s, z16.s\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
-      "uaddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "neg z16.s, p2/M, z16.s\n"
-      "mul z11.s, p2/M, z11.s, z16.s\n"
       "42:"  // Width 4: skip row sum fixup
-      ".inst 0xc0904960  // addha za0.s, p2/M, p2/M, z11.s\n"
-      "add x21, %x[qp], %[per_layer_mul]\n"
+      ".inst 0xc0904b80  // addha za0.s, p2/M, p2/M, z28.s\n"
+      "add x20, %x[qp], %[per_layer_mul]\n"
+      "ld1rw { z11.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      ".inst 0xc0904961  // addha za1.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z10.s }, p2/Z, [x21]\n"
-      "add x22, %x[qp], %[c_offset]\n"
+      ".inst 0xc0904b81  // addha za1.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z7.s }, p2/Z, [x20]\n"
+      "add x20, %x[qp], %[c_offset]\n"
       "add x21, %x[qp], %[minval]\n"
-      ".inst 0xc0904962  // addha za2.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z12.s }, p2/Z, [x20]\n"
+      ".inst 0xc0904b82  // addha za2.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z6.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[maxval]\n"
-      ".inst 0xc0904963  // addha za3.s, p2/M, p2/M, z11.s\n"
-      "ld1rw { z13.s }, p2/Z, [x22]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
-      "ld1rw { z18.s }, p2/Z, [x20]\n"
-      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
-      ".inst 0xc0062c20  // mova { z0.d-z3.d }, za.d[x9, #1]\n"
-      ".inst 0xc0062c58  // mova { z24.d-z27.d }, za.d[x9, #2]\n"
-      ".inst 0xc0062c74  // mova { z20.d-z23.d }, za.d[x9, #3]\n"
-      ".inst 0xc1aaac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z10.s\n"
-      ".inst 0xc1aaac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z10.s\n"
-      ".inst 0xc1aaac18  // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z10.s\n"
-      ".inst 0xc1aaac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z10.s\n"
-      ".inst 0xc1acaa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z12.s\n"
-      ".inst 0xc1acaa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z12.s\n"
-      ".inst 0xc1acaa38  // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n"
-      ".inst 0xc1acaa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z12.s\n"
-      ".inst 0xc1adab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z13.s\n"
-      ".inst 0xc1adab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z13.s\n"
-      ".inst 0xc1adab18  // add { z24.s-z27.s }, { z24.s-z27.s }, z13.s\n"
-      ".inst 0xc1adab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z13.s\n"
-      ".inst 0xc1b2ce3c  // sclamp { z28.s-z31.s }, z17.s, z18.s\n"
-      ".inst 0xc1b2ce20  // sclamp { z0.s-z3.s }, z17.s, z18.s\n"
-      ".inst 0xc1b2ce38  // sclamp { z24.s-z27.s }, z17.s, z18.s\n"
-      ".inst 0xc1b2ce34  // sclamp { z20.s-z23.s }, z17.s, z18.s\n"
-      "uzp1 z28.h, z28.h, z29.h\n"
-      "uzp1 z14.h, z30.h, z31.h\n"
-      "uzp1 z0.h, z0.h, z1.h\n"
-      "uzp1 z18.h, z2.h, z3.h\n"
+      ".inst 0xc0904b83  // addha za3.s, p2/M, p2/M, z28.s\n"
+      "ld1rw { z3.s }, p2/Z, [x21]\n"
+      ".inst 0xc0062c18  // mova { z24.d-z27.d }, za.d[x9, #0]\n"
+      ".inst 0xc1abac18  // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z11.s\n"
+      ".inst 0xc0062c30  // mova { z16.d-z19.d }, za.d[x9, #1]\n"
+      ".inst 0xc1abac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z11.s\n"
+      ".inst 0xc0062c54  // mova { z20.d-z23.d }, za.d[x9, #2]\n"
+      ".inst 0xc1abac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z11.s\n"
+      ".inst 0xc0062c6c  // mova { z12.d-z15.d }, za.d[x9, #3]\n"
+      ".inst 0xc1abac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z11.s\n"
+      ".inst 0xc1a7aa38  // srshl { z24.s-z27.s }, { z24.s-z27.s }, z7.s\n"
+      "ld1rw { z31.s }, p2/Z, [x20]\n"
+      ".inst 0xc1a7aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z7.s\n"
+      ".inst 0xc1a7aa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z7.s\n"
+      ".inst 0xc1a7aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n"
+      ".inst 0xc1a6ab18  // add { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n"
+      ".inst 0xc1a6ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n"
+      ".inst 0xc1a6ab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z6.s\n"
+      ".inst 0xc1a6ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n"
+      ".inst 0xc1bfcc78  // sclamp { z24.s-z27.s }, z3.s, z31.s\n"
+      ".inst 0xc1bfcc70  // sclamp { z16.s-z19.s }, z3.s, z31.s\n"
       "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
+      ".inst 0xc1bfcc74  // sclamp { z20.s-z23.s }, z3.s, z31.s\n"
+      ".inst 0xc1bfcc6c  // sclamp { z12.s-z15.s }, z3.s, z31.s\n"
+      "uzp1 z25.h, z26.h, z27.h\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
+      "uzp1 z18.h, z18.h, z19.h\n"
       "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z16.h, z22.h, z23.h\n"
-      "uzp1 z28.b, z28.b, z14.b\n"
-      "uzp1 z0.b, z0.b, z18.b\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "uzp1 z20.b, z20.b, z16.b\n"
-      "st1b { z28.b }, p2, [x25]\n"
-      "st1b { z0.b }, p2, [x25, #1, MUL VL]\n"
-      "st1b { z24.b }, p2, [x25, #2, MUL VL]\n"
-      "st1b { z20.b }, p1, [x25, #3, MUL VL]\n"
+      "uzp1 z17.h, z22.h, z23.h\n"
+      "uzp1 z12.h, z12.h, z13.h\n"
+      "uzp1 z30.h, z14.h, z15.h\n"
+      "uzp1 z24.b, z24.b, z25.b\n"
+      "st1b { z24.b }, p2, [x25]\n"
+      "uzp1 z16.b, z16.b, z18.b\n"
+      "st1b { z16.b }, p2, [x25, #1, MUL VL]\n"
+      "uzp1 z20.b, z20.b, z17.b\n"
+      "uzp1 z12.b, z12.b, z30.b\n"
+      "st1b { z20.b }, p2, [x25, #2, MUL VL]\n"
+      "st1b { z12.b }, p1, [x25, #3, MUL VL]\n"
       "addvl x25, x25, #4\n"
       "43:"  // Width 4: Output done
-      "subs x26, x26, #0x4\n"
+      "subs x27, x27, #0x4\n"
       "sub %x[N], %x[N], x28, LSL #2\n"
       "bgt 4b\n"
       "44:"  // Exit
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL.hpp
index db4f25bbfa..fcce7a1424 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
 class cls_sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL
 {
 public:
-  typedef bfloat16 lhs_operand_type;
-  typedef bfloat16 rhs_operand_type;
+  typedef bfloat16 operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const bfloat16 *const A, const bfloat16 *const B, float *const C, int ldc, const int M, const int N, const int K, const float *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 1, 4, 2> transforms = {};
+  StdTransformsSME<operand_type, result_type, 1, 4, 2> transforms = {};
 
   cls_sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp
index 87ba6d4819..36114c5060 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
         B(B), kstride_bytes(roundup(K, 2) * sizeof(bfloat16)),
         C(C), ldcb(ldc * sizeof(float)),
         M(M), N(N), K(K),
+        n_loops(((K / 2) - 1) / 2), n_tail_iters(((K / 2) - 1) % 2),
         min(-std::numeric_limits<float>::infinity()),
         max(std::numeric_limits<float>::infinity()),
         bias(bias),
@@ -87,13 +88,12 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
     const long kstride_bytes;
     float *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
     float min = -std::numeric_limits<float>::infinity();
     float max = std::numeric_limits<float>::infinity();
 
     const float *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -112,17 +112,17 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5dc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x14]\n"
+      ".inst 0xa040c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14]\n"
+      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
       ".inst 0xa041c5d8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c5d4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
-      "addvl x14, x14, #16\n"
       ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa042c5d4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
       ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
-      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa043c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
+      ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x14, x14, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w11, [%x[args], %[offsetof_M]]\n"
@@ -137,103 +137,103 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "fmov z15.s, #1.0\n"
-      ".inst 0xa109c280  // ld1w { z0.s, z4.s, z8.s, z12.s }, p8/Z, [x20, x9, LSL #2]\n"
-      ".inst 0x808001e0  // fmopa za0.s, p0/M, p0/M, z15.s, z0.s\n"
-      ".inst 0x808401e1  // fmopa za1.s, p0/M, p0/M, z15.s, z4.s\n"
-      ".inst 0x808801e2  // fmopa za2.s, p0/M, p0/M, z15.s, z8.s\n"
-      ".inst 0x808c01e3  // fmopa za3.s, p0/M, p0/M, z15.s, z12.s\n"
+      "fmov z6.s, #1.0\n"
+      ".inst 0xa009c29d  // ldnt1w { z28.s-z31.s }, p8/Z, [x20, x9, LSL #2]\n"
+      ".inst 0x809c00c0  // fmopa za0.s, p0/M, p0/M, z6.s, z28.s\n"
+      ".inst 0x809d00c1  // fmopa za1.s, p0/M, p0/M, z6.s, z29.s\n"
+      ".inst 0x809e00c2  // fmopa za2.s, p0/M, p0/M, z6.s, z30.s\n"
+      ".inst 0x809f00c3  // fmopa za3.s, p0/M, p0/M, z6.s, z31.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x9\n"
       "mov x21, x10\n"
       "incw x20, ALL, MUL #4\n"
       "incw x21\n"
       "cmp x20, x28\n"
-      "mov x20, x15\n"
       "csel x21, x10, x21, LT\n"
+      "mov x20, x15\n"
       "bfm x15, XZR, #0x0, #0x0  // bfc x15, #0x0, #0x1\n"
       "cmp x21, x11\n"
       "csel x15, x20, x15, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x1\n"
       "lsr x20, x20, #0x1\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x9, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      "ld1h { z20.h }, p0/Z, [x26]\n"
-      ".inst 0xa140a6f3  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x23]\n"
-      "ld1h { z4.h }, p0/Z, [x26, #1, MUL VL]\n"
-      ".inst 0xa041a6ec  // ld1h { z12.h-z15.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      "ld1h { z29.h }, p0/Z, [x26, #2, MUL VL]\n"
-      ".inst 0xa142a6f2  // ld1h { z18.h, z22.h, z26.h, z30.h }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      "ld1h { z2.h }, p0/Z, [x26, #3, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x9, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      "ld1h { z28.h }, p0/Z, [x26]\n"
+      ".inst 0xa040a6e9  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x23]\n"
+      "ld1h { z22.h }, p0/Z, [x26, #1, MUL VL]\n"
+      ".inst 0xa041a6ed  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      "ld1h { z30.h }, p0/Z, [x26, #2, MUL VL]\n"
+      ".inst 0xa042a6e5  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      "ld1h { z20.h }, p0/Z, [x26, #3, MUL VL]\n"
       "addvl x26, x26, #4\n"
-      ".inst 0xa043a6e8  // ld1h { z8.h-z11.h }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      ".inst 0xa143a6fb  // ldnt1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "ble 7f\n"
       "6:"  // K loop
+      ".inst 0x81880380  // bfmopa za0.s, p0/M, p0/M, z28.h, z8.h\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0x81890381  // bfmopa za1.s, p0/M, p0/M, z28.h, z9.h\n"
+      ".inst 0x818a0382  // bfmopa za2.s, p0/M, p0/M, z28.h, z10.h\n"
+      ".inst 0x818b0383  // bfmopa za3.s, p0/M, p0/M, z28.h, z11.h\n"
+      "ld1h { z28.h }, p0/Z, [x26]\n"
+      ".inst 0x818c02c0  // bfmopa za0.s, p0/M, p0/M, z22.h, z12.h\n"
+      ".inst 0xa040a6e9  // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x23]\n"
+      ".inst 0x818d02c1  // bfmopa za1.s, p0/M, p0/M, z22.h, z13.h\n"
+      ".inst 0x818e02c2  // bfmopa za2.s, p0/M, p0/M, z22.h, z14.h\n"
+      ".inst 0x818f02c3  // bfmopa za3.s, p0/M, p0/M, z22.h, z15.h\n"
+      "ld1h { z22.h }, p0/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x818403c0  // bfmopa za0.s, p0/M, p0/M, z30.h, z4.h\n"
+      ".inst 0xa041a6ed  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0x818503c1  // bfmopa za1.s, p0/M, p0/M, z30.h, z5.h\n"
+      ".inst 0x818603c2  // bfmopa za2.s, p0/M, p0/M, z30.h, z6.h\n"
+      ".inst 0x818703c3  // bfmopa za3.s, p0/M, p0/M, z30.h, z7.h\n"
+      "ld1h { z30.h }, p0/Z, [x26, #2, MUL VL]\n"
+      ".inst 0xa042a6e5  // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
       ".inst 0x81930280  // bfmopa za0.s, p0/M, p0/M, z20.h, z19.h\n"
-      "subs x21, x21, #0x1\n"
       ".inst 0x81970281  // bfmopa za1.s, p0/M, p0/M, z20.h, z23.h\n"
       ".inst 0x819b0282  // bfmopa za2.s, p0/M, p0/M, z20.h, z27.h\n"
       ".inst 0x819f0283  // bfmopa za3.s, p0/M, p0/M, z20.h, z31.h\n"
-      "ld1h { z20.h }, p0/Z, [x26]\n"
-      ".inst 0x818c0080  // bfmopa za0.s, p0/M, p0/M, z4.h, z12.h\n"
-      ".inst 0xa140a6f3  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x23]\n"
-      ".inst 0x818d0081  // bfmopa za1.s, p0/M, p0/M, z4.h, z13.h\n"
-      ".inst 0x818e0082  // bfmopa za2.s, p0/M, p0/M, z4.h, z14.h\n"
-      ".inst 0x818f0083  // bfmopa za3.s, p0/M, p0/M, z4.h, z15.h\n"
-      "ld1h { z4.h }, p0/Z, [x26, #1, MUL VL]\n"
-      ".inst 0x819203a0  // bfmopa za0.s, p0/M, p0/M, z29.h, z18.h\n"
-      ".inst 0xa041a6ec  // ld1h { z12.h-z15.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0x819603a1  // bfmopa za1.s, p0/M, p0/M, z29.h, z22.h\n"
-      ".inst 0x819a03a2  // bfmopa za2.s, p0/M, p0/M, z29.h, z26.h\n"
-      ".inst 0x819e03a3  // bfmopa za3.s, p0/M, p0/M, z29.h, z30.h\n"
-      "ld1h { z29.h }, p0/Z, [x26, #2, MUL VL]\n"
-      ".inst 0xa142a6f2  // ld1h { z18.h, z22.h, z26.h, z30.h }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      ".inst 0x81880040  // bfmopa za0.s, p0/M, p0/M, z2.h, z8.h\n"
-      ".inst 0x81890041  // bfmopa za1.s, p0/M, p0/M, z2.h, z9.h\n"
-      ".inst 0x818a0042  // bfmopa za2.s, p0/M, p0/M, z2.h, z10.h\n"
-      ".inst 0x818b0043  // bfmopa za3.s, p0/M, p0/M, z2.h, z11.h\n"
-      "ld1h { z2.h }, p0/Z, [x26, #3, MUL VL]\n"
+      "ld1h { z20.h }, p0/Z, [x26, #3, MUL VL]\n"
       "addvl x26, x26, #4\n"
-      ".inst 0xa043a6e8  // ld1h { z8.h-z11.h }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      ".inst 0xa143a6fb  // ldnt1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "bgt 6b\n"
       "7:"  // K loop tail
+      ".inst 0x81880380  // bfmopa za0.s, p0/M, p0/M, z28.h, z8.h\n"
+      ".inst 0x81890381  // bfmopa za1.s, p0/M, p0/M, z28.h, z9.h\n"
+      ".inst 0x818a0382  // bfmopa za2.s, p0/M, p0/M, z28.h, z10.h\n"
+      ".inst 0x818b0383  // bfmopa za3.s, p0/M, p0/M, z28.h, z11.h\n"
+      ".inst 0x818c02c0  // bfmopa za0.s, p0/M, p0/M, z22.h, z12.h\n"
+      ".inst 0x818d02c1  // bfmopa za1.s, p0/M, p0/M, z22.h, z13.h\n"
+      ".inst 0x818e02c2  // bfmopa za2.s, p0/M, p0/M, z22.h, z14.h\n"
+      ".inst 0x818f02c3  // bfmopa za3.s, p0/M, p0/M, z22.h, z15.h\n"
+      ".inst 0x818403c0  // bfmopa za0.s, p0/M, p0/M, z30.h, z4.h\n"
+      ".inst 0x818503c1  // bfmopa za1.s, p0/M, p0/M, z30.h, z5.h\n"
+      ".inst 0x818603c2  // bfmopa za2.s, p0/M, p0/M, z30.h, z6.h\n"
+      ".inst 0x818703c3  // bfmopa za3.s, p0/M, p0/M, z30.h, z7.h\n"
       ".inst 0x81930280  // bfmopa za0.s, p0/M, p0/M, z20.h, z19.h\n"
       ".inst 0x81970281  // bfmopa za1.s, p0/M, p0/M, z20.h, z23.h\n"
       ".inst 0x819b0282  // bfmopa za2.s, p0/M, p0/M, z20.h, z27.h\n"
       ".inst 0x819f0283  // bfmopa za3.s, p0/M, p0/M, z20.h, z31.h\n"
-      ".inst 0x818c0080  // bfmopa za0.s, p0/M, p0/M, z4.h, z12.h\n"
-      ".inst 0x818d0081  // bfmopa za1.s, p0/M, p0/M, z4.h, z13.h\n"
-      ".inst 0x818e0082  // bfmopa za2.s, p0/M, p0/M, z4.h, z14.h\n"
-      ".inst 0x818f0083  // bfmopa za3.s, p0/M, p0/M, z4.h, z15.h\n"
-      ".inst 0x819203a0  // bfmopa za0.s, p0/M, p0/M, z29.h, z18.h\n"
-      ".inst 0x819603a1  // bfmopa za1.s, p0/M, p0/M, z29.h, z22.h\n"
-      ".inst 0x819a03a2  // bfmopa za2.s, p0/M, p0/M, z29.h, z26.h\n"
-      ".inst 0x819e03a3  // bfmopa za3.s, p0/M, p0/M, z29.h, z30.h\n"
-      ".inst 0x81880040  // bfmopa za0.s, p0/M, p0/M, z2.h, z8.h\n"
-      ".inst 0x81890041  // bfmopa za1.s, p0/M, p0/M, z2.h, z9.h\n"
-      ".inst 0x818a0042  // bfmopa za2.s, p0/M, p0/M, z2.h, z10.h\n"
-      ".inst 0x818b0043  // bfmopa za3.s, p0/M, p0/M, z2.h, z11.h\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      "ld1h { z26.h }, p0/Z, [x26]\n"
-      "subs x20, x20, #0x1\n"
+      "ld1h { z8.h }, p0/Z, [x26]\n"
+      "subs x21, x21, #0x1\n"
       "addvl x26, x26, #1\n"
       ".inst 0xa140a6e3  // ld1h { z3.h, z7.h, z11.h, z15.h }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #4\n"
-      ".inst 0x81830340  // bfmopa za0.s, p0/M, p0/M, z26.h, z3.h\n"
-      ".inst 0x81870341  // bfmopa za1.s, p0/M, p0/M, z26.h, z7.h\n"
-      ".inst 0x818b0342  // bfmopa za2.s, p0/M, p0/M, z26.h, z11.h\n"
-      ".inst 0x818f0343  // bfmopa za3.s, p0/M, p0/M, z26.h, z15.h\n"
+      ".inst 0x81830100  // bfmopa za0.s, p0/M, p0/M, z8.h, z3.h\n"
+      ".inst 0x81870101  // bfmopa za1.s, p0/M, p0/M, z8.h, z7.h\n"
+      ".inst 0x818b0102  // bfmopa za2.s, p0/M, p0/M, z8.h, z11.h\n"
+      ".inst 0x818f0103  // bfmopa za3.s, p0/M, p0/M, z8.h, z15.h\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x15, #1, 14f\n"
@@ -241,25 +241,25 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5d8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x14]\n"
+      ".inst 0xa040c5d4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x14]\n"
       ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
-      ".inst 0xa042c5d4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xa060c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13]\n"
-      ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
-      ".inst 0xa061c5b0  // st1w { z16.s-z19.s }, pn9.b, [x13, #0x4, MUL VL]\n"
-      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
+      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
+      ".inst 0xa042c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5a0  // st1w { z0.s-z3.s }, pn9.b, [x13, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c5bc  // st1w { z28.s-z31.s }, pn9.b, [x13, #0xc, MUL VL]\n"
+      ".inst 0xa060c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13]\n"
+      "addvl x14, x14, #16\n"
+      ".inst 0xa061c5a4  // st1w { z4.s-z7.s }, pn9.b, [x13, #0x4, MUL VL]\n"
+      ".inst 0xa062c5ac  // st1w { z12.s-z15.s }, pn9.b, [x13, #0x8, MUL VL]\n"
+      ".inst 0xa063c5a0  // st1w { z0.s-z3.s }, pn9.b, [x13, #0xc, MUL VL]\n"
       "addvl x13, x13, #16\n"
       "blt 11b\n"
       "b 24f\n"
@@ -267,31 +267,31 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
+      ".inst 0xa060c5b0  // st1w { z16.s-z19.s }, pn9.b, [x13]\n"
       ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5ac  // st1w { z12.s-z15.s }, pn9.b, [x13]\n"
+      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
+      ".inst 0xa061c5b8  // st1w { z24.s-z27.s }, pn9.b, [x13, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5a4  // st1w { z4.s-z7.s }, pn9.b, [x13, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
       ".inst 0xa062c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13, #0x8, MUL VL]\n"
-      ".inst 0xa063c5b0  // st1w { z16.s-z19.s }, pn9.b, [x13, #0xc, MUL VL]\n"
+      ".inst 0xa063c5a4  // st1w { z4.s-z7.s }, pn9.b, [x13, #0xc, MUL VL]\n"
       "addvl x13, x13, #16\n"
       "blt 13b\n"
       "b 24f\n"
       "14:"  // Store to output array
       "ldr x25, [%x[args], %[offsetof_C]]\n"
+      "add x25, x25, x9, LSL #2\n"  // C += n
       "sub x24, x11, x10\n"
       "ldr x23, [%x[args], %[offsetof_ldcb]]\n"
-      "add x25, x25, x9, LSL #2\n"  // C += n
       "madd x25, x10, x23, x25\n"  // C += m * ldc
       "tbz x15, #2, 18f\n"
       "cntw x20\n"
-      "mov x12, #0x0\n"
       "cmp x24, x20\n"
       "csel x22, x24, x20, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Skip activation: Accumulator row 0 loop
@@ -301,30 +301,30 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
       ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
       ".inst 0xa160c320  // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xa160c321  // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa160c322  // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa160c323  // st1w { z3.s, z7.s, z11.s, z15.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "blt 15b\n"
       "16:"  // Store to output array: Skip activation: Accumulator row 0 oddments
       "cbz x20, 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
-      ".inst 0xa160c320  // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x25]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xa160c330  // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "beq 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xa160c321  // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x25]\n"
+      ".inst 0xa160c331  // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "beq 17f\n"
-      ".inst 0xa160c322  // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x25]\n"
+      ".inst 0xa160c332  // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "17:"  // Store to output array: Skip activation: Accumulator row 0 oddments: End
       "subs x24, x24, x22\n"
@@ -332,29 +332,29 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
       "b 22f\n"
       "18:"  // Store to output array: Skip activation: End
       "cntw x20\n"
-      "ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
-      "mov x12, #0x0\n"
       "cmp x24, x20\n"
-      "ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
       "csel x20, x24, x20, LT\n"
       "lsr x21, x20, #0x2\n"
+      "ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 20f\n"
       "19:"  // Store to output array: Accumulator row 0 loop
       ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
       ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
-      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
       ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
       ".inst 0xc1a0c834  // fclamp { z20.s-z23.s }, z1.s, z0.s\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
       ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
       ".inst 0xc1a0c83c  // fclamp { z28.s-z31.s }, z1.s, z0.s\n"
-      "add x12, x12, #0x4\n"
-      "cmp x12, x21, LSL #2\n"
       ".inst 0xa160c330  // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa160c331  // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa160c332  // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       ".inst 0xa160c333  // st1w { z19.s, z23.s, z27.s, z31.s }, p8, [x25]\n"
@@ -364,13 +364,13 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
       "cbz x20, 21f\n"
       ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
       ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
-      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
       ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
       ".inst 0xc1a0c834  // fclamp { z20.s-z23.s }, z1.s, z0.s\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
       ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
       ".inst 0xc1a0c83c  // fclamp { z28.s-z31.s }, z1.s, z0.s\n"
+      "subs x20, x20, #0x1\n"
       ".inst 0xa160c330  // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "beq 21f\n"
@@ -385,25 +385,25 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "23:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5c0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x14]\n"
-      ".inst 0xa041c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c5dc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa040c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa041c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
       ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x14, x14, #16\n"
       "blt 23b\n"
       "24:"  // End block
       "incw x9, ALL, MUL #4\n"
       "cmp x9, x28\n"
       "blt 3b\n"
       "incw x10\n"
-      "mov x9, #0x0\n"
       "cmp x10, x11\n"
+      "mov x9, #0x0\n"
       "mov x27, x26\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL.hpp
index 7f681b2734..8b0f5b013f 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
 class cls_sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL
 {
 public:
-  typedef bfloat16 lhs_operand_type;
-  typedef bfloat16 rhs_operand_type;
+  typedef bfloat16 operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const bfloat16 *const A, const bfloat16 *const B, float *const C, int ldc, const int M, const int N, const int K, const float *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 2, 2, 2> transforms = {};
+  StdTransformsSME<operand_type, result_type, 2, 2, 2> transforms = {};
 
   cls_sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp
index 8b3c6d7fec..cc44c9a537 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
         B(B), kstride_bytes(roundup(K, 2) * sizeof(bfloat16)),
         C(C), ldcb(ldc * sizeof(float)),
         M(M), N(N), K(K),
+        n_loops(((K / 2) - 1) / 2), n_tail_iters(((K / 2) - 1) % 2),
         min(-std::numeric_limits<float>::infinity()),
         max(std::numeric_limits<float>::infinity()),
         bias(bias),
@@ -87,13 +88,12 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
     const long kstride_bytes;
     float *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
     float min = -std::numeric_limits<float>::infinity();
     float max = std::numeric_limits<float>::infinity();
 
     const float *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -112,17 +112,17 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa040c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa041c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa042c5f4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xa043c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -137,103 +137,103 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "fmov z17.s, #1.0\n"
-      ".inst 0xa00a428a  // ld1w { z10.s-z11.s }, p8/Z, [x20, x10, LSL #2]\n"
-      ".inst 0x808a0220  // fmopa za0.s, p0/M, p0/M, z17.s, z10.s\n"
-      ".inst 0x808b0221  // fmopa za1.s, p0/M, p0/M, z17.s, z11.s\n"
-      ".inst 0x808a0222  // fmopa za2.s, p0/M, p0/M, z17.s, z10.s\n"
-      ".inst 0x808b0223  // fmopa za3.s, p0/M, p0/M, z17.s, z11.s\n"
+      "fmov z12.s, #1.0\n"
+      ".inst 0xa10a4289  // ldnt1w { z1.s, z9.s }, p8/Z, [x20, x10, LSL #2]\n"
+      ".inst 0x80810180  // fmopa za0.s, p0/M, p0/M, z12.s, z1.s\n"
+      ".inst 0x80890181  // fmopa za1.s, p0/M, p0/M, z12.s, z9.s\n"
+      ".inst 0x80810182  // fmopa za2.s, p0/M, p0/M, z12.s, z1.s\n"
+      ".inst 0x80890183  // fmopa za3.s, p0/M, p0/M, z12.s, z9.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20, ALL, MUL #2\n"
       "incw x21, ALL, MUL #2\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x1\n"
       "lsr x20, x20, #0x1\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa0402776  // ld1h { z22.h-z23.h }, pn9.b/Z, [x27]\n"
-      ".inst 0xa14026e7  // ld1h { z7.h, z15.h }, pn9.b/Z, [x23]\n"
-      ".inst 0xa1412766  // ld1h { z6.h, z14.h }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0xa04126f4  // ld1h { z20.h-z21.h }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa1422762  // ld1h { z2.h, z10.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14226e3  // ld1h { z3.h, z11.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa1432761  // ld1h { z1.h, z9.h }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x10, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa0402772  // ld1h { z18.h-z19.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa04026e3  // ldnt1h { z2.h-z3.h }, pn9.b/Z, [x23]\n"
+      ".inst 0xa0412764  // ld1h { z4.h-z5.h }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa04126fb  // ldnt1h { z26.h-z27.h }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0xa042276a  // ld1h { z10.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa04226f5  // ldnt1h { z20.h-z21.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa0432766  // ld1h { z6.h-z7.h }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa04326e4  // ld1h { z4.h-z5.h }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa04326e9  // ldnt1h { z8.h-z9.h }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0x818702c0  // bfmopa za0.s, p0/M, p0/M, z22.h, z7.h\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0x818f02c1  // bfmopa za1.s, p0/M, p0/M, z22.h, z15.h\n"
-      ".inst 0x818702e2  // bfmopa za2.s, p0/M, p0/M, z23.h, z7.h\n"
-      ".inst 0x818f02e3  // bfmopa za3.s, p0/M, p0/M, z23.h, z15.h\n"
-      ".inst 0xa0402776  // ld1h { z22.h-z23.h }, pn9.b/Z, [x27]\n"
-      ".inst 0x819400c0  // bfmopa za0.s, p0/M, p0/M, z6.h, z20.h\n"
-      ".inst 0xa14026e7  // ld1h { z7.h, z15.h }, pn9.b/Z, [x23]\n"
-      ".inst 0x819500c1  // bfmopa za1.s, p0/M, p0/M, z6.h, z21.h\n"
-      ".inst 0x819401c2  // bfmopa za2.s, p0/M, p0/M, z14.h, z20.h\n"
-      ".inst 0x819501c3  // bfmopa za3.s, p0/M, p0/M, z14.h, z21.h\n"
-      ".inst 0xa1412766  // ld1h { z6.h, z14.h }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0x81830040  // bfmopa za0.s, p0/M, p0/M, z2.h, z3.h\n"
-      ".inst 0xa04126f4  // ld1h { z20.h-z21.h }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0x818b0041  // bfmopa za1.s, p0/M, p0/M, z2.h, z11.h\n"
-      ".inst 0x81830142  // bfmopa za2.s, p0/M, p0/M, z10.h, z3.h\n"
-      ".inst 0x818b0143  // bfmopa za3.s, p0/M, p0/M, z10.h, z11.h\n"
-      ".inst 0xa1422762  // ld1h { z2.h, z10.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14226e3  // ld1h { z3.h, z11.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0x81840020  // bfmopa za0.s, p0/M, p0/M, z1.h, z4.h\n"
-      ".inst 0x81850021  // bfmopa za1.s, p0/M, p0/M, z1.h, z5.h\n"
-      ".inst 0x81840122  // bfmopa za2.s, p0/M, p0/M, z9.h, z4.h\n"
-      ".inst 0x81850123  // bfmopa za3.s, p0/M, p0/M, z9.h, z5.h\n"
-      ".inst 0xa1432761  // ld1h { z1.h, z9.h }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      ".inst 0x81820240  // bfmopa za0.s, p0/M, p0/M, z18.h, z2.h\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0x81830241  // bfmopa za1.s, p0/M, p0/M, z18.h, z3.h\n"
+      ".inst 0x81820262  // bfmopa za2.s, p0/M, p0/M, z19.h, z2.h\n"
+      ".inst 0x81830263  // bfmopa za3.s, p0/M, p0/M, z19.h, z3.h\n"
+      ".inst 0xa0402772  // ld1h { z18.h-z19.h }, pn9.b/Z, [x27]\n"
+      ".inst 0x819a0080  // bfmopa za0.s, p0/M, p0/M, z4.h, z26.h\n"
+      ".inst 0xa04026e3  // ldnt1h { z2.h-z3.h }, pn9.b/Z, [x23]\n"
+      ".inst 0x819b0081  // bfmopa za1.s, p0/M, p0/M, z4.h, z27.h\n"
+      ".inst 0x819a00a2  // bfmopa za2.s, p0/M, p0/M, z5.h, z26.h\n"
+      ".inst 0x819b00a3  // bfmopa za3.s, p0/M, p0/M, z5.h, z27.h\n"
+      ".inst 0xa0412764  // ld1h { z4.h-z5.h }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0x81940140  // bfmopa za0.s, p0/M, p0/M, z10.h, z20.h\n"
+      ".inst 0xa04126fb  // ldnt1h { z26.h-z27.h }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0x81950141  // bfmopa za1.s, p0/M, p0/M, z10.h, z21.h\n"
+      ".inst 0x81940162  // bfmopa za2.s, p0/M, p0/M, z11.h, z20.h\n"
+      ".inst 0x81950163  // bfmopa za3.s, p0/M, p0/M, z11.h, z21.h\n"
+      ".inst 0xa042276a  // ld1h { z10.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa04226f5  // ldnt1h { z20.h-z21.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0x818800c0  // bfmopa za0.s, p0/M, p0/M, z6.h, z8.h\n"
+      ".inst 0x818900c1  // bfmopa za1.s, p0/M, p0/M, z6.h, z9.h\n"
+      ".inst 0x818800e2  // bfmopa za2.s, p0/M, p0/M, z7.h, z8.h\n"
+      ".inst 0x818900e3  // bfmopa za3.s, p0/M, p0/M, z7.h, z9.h\n"
+      ".inst 0xa0432766  // ld1h { z6.h-z7.h }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa04326e4  // ld1h { z4.h-z5.h }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa04326e9  // ldnt1h { z8.h-z9.h }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0x818702c0  // bfmopa za0.s, p0/M, p0/M, z22.h, z7.h\n"
-      ".inst 0x818f02c1  // bfmopa za1.s, p0/M, p0/M, z22.h, z15.h\n"
-      ".inst 0x818702e2  // bfmopa za2.s, p0/M, p0/M, z23.h, z7.h\n"
-      ".inst 0x818f02e3  // bfmopa za3.s, p0/M, p0/M, z23.h, z15.h\n"
-      ".inst 0x819400c0  // bfmopa za0.s, p0/M, p0/M, z6.h, z20.h\n"
-      ".inst 0x819500c1  // bfmopa za1.s, p0/M, p0/M, z6.h, z21.h\n"
-      ".inst 0x819401c2  // bfmopa za2.s, p0/M, p0/M, z14.h, z20.h\n"
-      ".inst 0x819501c3  // bfmopa za3.s, p0/M, p0/M, z14.h, z21.h\n"
-      ".inst 0x81830040  // bfmopa za0.s, p0/M, p0/M, z2.h, z3.h\n"
-      ".inst 0x818b0041  // bfmopa za1.s, p0/M, p0/M, z2.h, z11.h\n"
-      ".inst 0x81830142  // bfmopa za2.s, p0/M, p0/M, z10.h, z3.h\n"
-      ".inst 0x818b0143  // bfmopa za3.s, p0/M, p0/M, z10.h, z11.h\n"
-      ".inst 0x81840020  // bfmopa za0.s, p0/M, p0/M, z1.h, z4.h\n"
-      ".inst 0x81850021  // bfmopa za1.s, p0/M, p0/M, z1.h, z5.h\n"
-      ".inst 0x81840122  // bfmopa za2.s, p0/M, p0/M, z9.h, z4.h\n"
-      ".inst 0x81850123  // bfmopa za3.s, p0/M, p0/M, z9.h, z5.h\n"
+      ".inst 0x81820240  // bfmopa za0.s, p0/M, p0/M, z18.h, z2.h\n"
+      ".inst 0x81830241  // bfmopa za1.s, p0/M, p0/M, z18.h, z3.h\n"
+      ".inst 0x81820262  // bfmopa za2.s, p0/M, p0/M, z19.h, z2.h\n"
+      ".inst 0x81830263  // bfmopa za3.s, p0/M, p0/M, z19.h, z3.h\n"
+      ".inst 0x819a0080  // bfmopa za0.s, p0/M, p0/M, z4.h, z26.h\n"
+      ".inst 0x819b0081  // bfmopa za1.s, p0/M, p0/M, z4.h, z27.h\n"
+      ".inst 0x819a00a2  // bfmopa za2.s, p0/M, p0/M, z5.h, z26.h\n"
+      ".inst 0x819b00a3  // bfmopa za3.s, p0/M, p0/M, z5.h, z27.h\n"
+      ".inst 0x81940140  // bfmopa za0.s, p0/M, p0/M, z10.h, z20.h\n"
+      ".inst 0x81950141  // bfmopa za1.s, p0/M, p0/M, z10.h, z21.h\n"
+      ".inst 0x81940162  // bfmopa za2.s, p0/M, p0/M, z11.h, z20.h\n"
+      ".inst 0x81950163  // bfmopa za3.s, p0/M, p0/M, z11.h, z21.h\n"
+      ".inst 0x818800c0  // bfmopa za0.s, p0/M, p0/M, z6.h, z8.h\n"
+      ".inst 0x818900c1  // bfmopa za1.s, p0/M, p0/M, z6.h, z9.h\n"
+      ".inst 0x818800e2  // bfmopa za2.s, p0/M, p0/M, z7.h, z8.h\n"
+      ".inst 0x818900e3  // bfmopa za3.s, p0/M, p0/M, z7.h, z9.h\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa040276a  // ld1h { z10.h-z11.h }, pn9.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xa040277e  // ld1h { z30.h-z31.h }, pn9.b/Z, [x27]\n"
+      "subs x21, x21, #0x1\n"
       "addvl x27, x27, #2\n"
-      ".inst 0xa04026ee  // ld1h { z14.h-z15.h }, pn9.b/Z, [x23]\n"
+      ".inst 0xa14026e5  // ld1h { z5.h, z13.h }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #2\n"
-      ".inst 0x818e0140  // bfmopa za0.s, p0/M, p0/M, z10.h, z14.h\n"
-      ".inst 0x818f0141  // bfmopa za1.s, p0/M, p0/M, z10.h, z15.h\n"
-      ".inst 0x818e0162  // bfmopa za2.s, p0/M, p0/M, z11.h, z14.h\n"
-      ".inst 0x818f0163  // bfmopa za3.s, p0/M, p0/M, z11.h, z15.h\n"
+      ".inst 0x818503c0  // bfmopa za0.s, p0/M, p0/M, z30.h, z5.h\n"
+      ".inst 0x818d03c1  // bfmopa za1.s, p0/M, p0/M, z30.h, z13.h\n"
+      ".inst 0x818503e2  // bfmopa za2.s, p0/M, p0/M, z31.h, z5.h\n"
+      ".inst 0x818d03e3  // bfmopa za3.s, p0/M, p0/M, z31.h, z13.h\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x16, #1, 14f\n"
@@ -241,24 +241,24 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
-      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
+      ".inst 0xa040c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
       ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
       ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xa060c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14]\n"
       ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
-      ".inst 0xa061c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840703  // mova za3h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa043c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
+      ".inst 0xa060c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n"
       ".inst 0xa063c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
@@ -267,31 +267,31 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
-      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
+      ".inst 0xa060c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14]\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
+      ".inst 0xa061c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c5d8  // st1w { z24.s-z27.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 30f\n"
       "14:"  // Store to output array
       "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "add x26, x26, x10, LSL #2\n"  // C += n
       "sub x25, x13, x11\n"
       "ldr x24, [%x[args], %[offsetof_ldcb]]\n"
-      "add x26, x26, x10, LSL #2\n"  // C += n
       "madd x26, x11, x24, x26\n"  // C += m * ldc
       "tbz x16, #2, 21f\n"
       "cntw x23\n"
-      "mov x12, #0x0\n"
       "cmp x25, x23\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Skip activation: Accumulator row 0 loop
@@ -299,36 +299,36 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
       ".inst 0xa1604344  // st1w { z4.s, z12.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xa1604345  // st1w { z5.s, z13.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa1604346  // st1w { z6.s, z14.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa1604347  // st1w { z7.s, z15.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "blt 15b\n"
       "16:"  // Store to output array: Skip activation: Accumulator row 0 oddments
       "cbz x20, 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
-      ".inst 0xa1604340  // st1w { z0.s, z8.s }, p8, [x26]\n"
+      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xa1604354  // st1w { z20.s, z28.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xa1604341  // st1w { z1.s, z9.s }, p8, [x26]\n"
+      ".inst 0xa1604355  // st1w { z21.s, z29.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 17f\n"
-      ".inst 0xa1604342  // st1w { z2.s, z10.s }, p8, [x26]\n"
+      ".inst 0xa1604356  // st1w { z22.s, z30.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "17:"  // Store to output array: Skip activation: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 21f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 19f\n"
       "18:"  // Store to output array: Skip activation: Accumulator row 1 loop
@@ -336,28 +336,28 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
       ".inst 0xa1604344  // st1w { z4.s, z12.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xa1604345  // st1w { z5.s, z13.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa1604346  // st1w { z6.s, z14.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa1604347  // st1w { z7.s, z15.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "blt 18b\n"
       "19:"  // Store to output array: Skip activation: Accumulator row 1 oddments
       "cbz x20, 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
-      ".inst 0xa1604344  // st1w { z4.s, z12.s }, p8, [x26]\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xa1604345  // st1w { z5.s, z13.s }, p8, [x26]\n"
+      ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 20f\n"
-      ".inst 0xa1604346  // st1w { z6.s, z14.s }, p8, [x26]\n"
+      ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "20:"  // Store to output array: Skip activation: Accumulator row 1 oddments: End
       "subs x25, x25, x22\n"
@@ -365,37 +365,37 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       "b 28f\n"
       "21:"  // Store to output array: Skip activation: End
       "cntw x23\n"
-      "ld1rw { z21.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
-      "mov x12, #0x0\n"
       "cmp x25, x23\n"
-      "ld1rw { z20.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 23f\n"
       "22:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
-      "add x12, x12, #0x4\n"
-      "cmp x12, x21, LSL #2\n"
-      ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
+      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xc1a0c834  // fclamp { z20.s-z23.s }, z1.s, z0.s\n"
+      ".inst 0xc1a0c83c  // fclamp { z28.s-z31.s }, z1.s, z0.s\n"
+      ".inst 0xa1604354  // st1w { z20.s, z28.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
+      "add x12, x12, #0x4\n"
+      ".inst 0xa1604355  // st1w { z21.s, z29.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      ".inst 0xa1604356  // st1w { z22.s, z30.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604353  // st1w { z19.s, z27.s }, p8, [x26]\n"
+      ".inst 0xa1604357  // st1w { z23.s, z31.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "blt 22b\n"
       "23:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 24f\n"
       ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
       ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
+      ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
+      ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
       ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 24f\n"
@@ -409,34 +409,34 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       "subs x25, x25, x22\n"
       "beq 28f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x23, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 26f\n"
       "25:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
-      ".inst 0xc0860468  // mova { z8.s-z11.s }, za3h.s[x12]\n"
-      ".inst 0xc1b4caa0  // fclamp { z0.s-z3.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4caa8  // fclamp { z8.s-z11.s }, z21.s, z20.s\n"
-      "add x12, x12, #0x4\n"
-      "cmp x12, x21, LSL #2\n"
-      ".inst 0xa1604340  // st1w { z0.s, z8.s }, p8, [x26]\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
+      ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
+      ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604341  // st1w { z1.s, z9.s }, p8, [x26]\n"
+      "add x12, x12, #0x4\n"
+      ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604342  // st1w { z2.s, z10.s }, p8, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604343  // st1w { z3.s, z11.s }, p8, [x26]\n"
+      ".inst 0xa1604353  // st1w { z19.s, z27.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "blt 25b\n"
       "26:"  // Store to output array: Accumulator row 1 oddments
       "cbz x20, 27f\n"
       ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
       ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
+      ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
       ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 27f\n"
@@ -451,25 +451,25 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "29:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa041c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
       ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 29b\n"
       "30:"  // End block
       "incw x10, ALL, MUL #2\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #2\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL.hpp
index 3c1dff268f..f8812a1a71 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
 class cls_sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL
 {
 public:
-  typedef bfloat16 lhs_operand_type;
-  typedef bfloat16 rhs_operand_type;
+  typedef bfloat16 operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const bfloat16 *const A, const bfloat16 *const B, float *const C, int ldc, const int M, const int N, const int K, const float *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 4, 1, 2> transforms = {};
+  StdTransformsSME<operand_type, result_type, 4, 1, 2> transforms = {};
 
   cls_sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp
index b4b94f305e..a63cadc63a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
         B(B), kstride_bytes(roundup(K, 2) * sizeof(bfloat16)),
         C(C), ldcb(ldc * sizeof(float)),
         M(M), N(N), K(K),
+        n_loops(((K / 2) - 1) / 2), n_tail_iters(((K / 2) - 1) % 2),
         min(-std::numeric_limits<float>::infinity()),
         max(std::numeric_limits<float>::infinity()),
         bias(bias),
@@ -87,13 +88,12 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
     const long kstride_bytes;
     float *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
     float min = -std::numeric_limits<float>::infinity();
     float max = std::numeric_limits<float>::infinity();
 
     const float *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -112,17 +112,17 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840783  // mova za3h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa040c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa041c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa042c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xa043c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -137,103 +137,103 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "fmov z6.s, #1.0\n"
-      "ld1w { z26.s }, p0/Z, [x20, x10, LSL #2]\n"
-      ".inst 0x809a24c0  // fmopa za0.s, p1/M, p1/M, z6.s, z26.s\n"
-      ".inst 0x809a24c1  // fmopa za1.s, p1/M, p1/M, z6.s, z26.s\n"
-      ".inst 0x809a24c2  // fmopa za2.s, p1/M, p1/M, z6.s, z26.s\n"
-      ".inst 0x809a24c3  // fmopa za3.s, p1/M, p1/M, z6.s, z26.s\n"
+      "fmov z11.s, #1.0\n"
+      "ldnt1w { z13.s }, p0/Z, [x20, x10, LSL #2]\n"
+      ".inst 0x808d2560  // fmopa za0.s, p1/M, p1/M, z11.s, z13.s\n"
+      ".inst 0x808d2561  // fmopa za1.s, p1/M, p1/M, z11.s, z13.s\n"
+      ".inst 0x808d2562  // fmopa za2.s, p1/M, p1/M, z11.s, z13.s\n"
+      ".inst 0x808d2563  // fmopa za3.s, p1/M, p1/M, z11.s, z13.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20\n"
       "incw x21, ALL, MUL #4\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x1\n"
       "lsr x20, x20, #0x1\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa140a363  // ld1h { z3.h, z7.h, z11.h, z15.h }, pn8.b/Z, [x27]\n"
-      "ld1h { z13.h }, p1/Z, [x23]\n"
-      ".inst 0xa141a372  // ld1h { z18.h, z22.h, z26.h, z30.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "ld1h { z21.h }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa142a373  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1h { z17.h }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa143a370  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x10, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa140a360  // ld1h { z0.h, z4.h, z8.h, z12.h }, pn8.b/Z, [x27]\n"
+      "ldnt1h { z19.h }, p1/Z, [x23]\n"
+      ".inst 0xa141a371  // ld1h { z17.h, z21.h, z25.h, z29.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      "ldnt1h { z22.h }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0xa142a370  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1h { z23.h }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0xa143a363  // ld1h { z3.h, z7.h, z11.h, z15.h }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1h { z2.h }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1h { z2.h }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0x818d2460  // bfmopa za0.s, p1/M, p1/M, z3.h, z13.h\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0x818d24e1  // bfmopa za1.s, p1/M, p1/M, z7.h, z13.h\n"
-      ".inst 0x818d2562  // bfmopa za2.s, p1/M, p1/M, z11.h, z13.h\n"
-      ".inst 0x818d25e3  // bfmopa za3.s, p1/M, p1/M, z15.h, z13.h\n"
-      ".inst 0xa140a363  // ld1h { z3.h, z7.h, z11.h, z15.h }, pn8.b/Z, [x27]\n"
-      ".inst 0x81952640  // bfmopa za0.s, p1/M, p1/M, z18.h, z21.h\n"
-      "ld1h { z13.h }, p1/Z, [x23]\n"
-      ".inst 0x819526c1  // bfmopa za1.s, p1/M, p1/M, z22.h, z21.h\n"
-      ".inst 0x81952742  // bfmopa za2.s, p1/M, p1/M, z26.h, z21.h\n"
-      ".inst 0x819527c3  // bfmopa za3.s, p1/M, p1/M, z30.h, z21.h\n"
-      ".inst 0xa141a372  // ld1h { z18.h, z22.h, z26.h, z30.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0x81912660  // bfmopa za0.s, p1/M, p1/M, z19.h, z17.h\n"
-      "ld1h { z21.h }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0x819126e1  // bfmopa za1.s, p1/M, p1/M, z23.h, z17.h\n"
-      ".inst 0x81912762  // bfmopa za2.s, p1/M, p1/M, z27.h, z17.h\n"
-      ".inst 0x819127e3  // bfmopa za3.s, p1/M, p1/M, z31.h, z17.h\n"
-      ".inst 0xa142a373  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1h { z17.h }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0x81822600  // bfmopa za0.s, p1/M, p1/M, z16.h, z2.h\n"
-      ".inst 0x81822681  // bfmopa za1.s, p1/M, p1/M, z20.h, z2.h\n"
-      ".inst 0x81822702  // bfmopa za2.s, p1/M, p1/M, z24.h, z2.h\n"
-      ".inst 0x81822783  // bfmopa za3.s, p1/M, p1/M, z28.h, z2.h\n"
-      ".inst 0xa143a370  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      ".inst 0x81932400  // bfmopa za0.s, p1/M, p1/M, z0.h, z19.h\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0x81932481  // bfmopa za1.s, p1/M, p1/M, z4.h, z19.h\n"
+      ".inst 0x81932502  // bfmopa za2.s, p1/M, p1/M, z8.h, z19.h\n"
+      ".inst 0x81932583  // bfmopa za3.s, p1/M, p1/M, z12.h, z19.h\n"
+      ".inst 0xa140a360  // ld1h { z0.h, z4.h, z8.h, z12.h }, pn8.b/Z, [x27]\n"
+      ".inst 0x81962620  // bfmopa za0.s, p1/M, p1/M, z17.h, z22.h\n"
+      "ldnt1h { z19.h }, p1/Z, [x23]\n"
+      ".inst 0x819626a1  // bfmopa za1.s, p1/M, p1/M, z21.h, z22.h\n"
+      ".inst 0x81962722  // bfmopa za2.s, p1/M, p1/M, z25.h, z22.h\n"
+      ".inst 0x819627a3  // bfmopa za3.s, p1/M, p1/M, z29.h, z22.h\n"
+      ".inst 0xa141a371  // ld1h { z17.h, z21.h, z25.h, z29.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0x81972600  // bfmopa za0.s, p1/M, p1/M, z16.h, z23.h\n"
+      "ldnt1h { z22.h }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0x81972681  // bfmopa za1.s, p1/M, p1/M, z20.h, z23.h\n"
+      ".inst 0x81972702  // bfmopa za2.s, p1/M, p1/M, z24.h, z23.h\n"
+      ".inst 0x81972783  // bfmopa za3.s, p1/M, p1/M, z28.h, z23.h\n"
+      ".inst 0xa142a370  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1h { z23.h }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0x81822460  // bfmopa za0.s, p1/M, p1/M, z3.h, z2.h\n"
+      ".inst 0x818224e1  // bfmopa za1.s, p1/M, p1/M, z7.h, z2.h\n"
+      ".inst 0x81822562  // bfmopa za2.s, p1/M, p1/M, z11.h, z2.h\n"
+      ".inst 0x818225e3  // bfmopa za3.s, p1/M, p1/M, z15.h, z2.h\n"
+      ".inst 0xa143a363  // ld1h { z3.h, z7.h, z11.h, z15.h }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1h { z2.h }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1h { z2.h }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0x818d2460  // bfmopa za0.s, p1/M, p1/M, z3.h, z13.h\n"
-      ".inst 0x818d24e1  // bfmopa za1.s, p1/M, p1/M, z7.h, z13.h\n"
-      ".inst 0x818d2562  // bfmopa za2.s, p1/M, p1/M, z11.h, z13.h\n"
-      ".inst 0x818d25e3  // bfmopa za3.s, p1/M, p1/M, z15.h, z13.h\n"
-      ".inst 0x81952640  // bfmopa za0.s, p1/M, p1/M, z18.h, z21.h\n"
-      ".inst 0x819526c1  // bfmopa za1.s, p1/M, p1/M, z22.h, z21.h\n"
-      ".inst 0x81952742  // bfmopa za2.s, p1/M, p1/M, z26.h, z21.h\n"
-      ".inst 0x819527c3  // bfmopa za3.s, p1/M, p1/M, z30.h, z21.h\n"
-      ".inst 0x81912660  // bfmopa za0.s, p1/M, p1/M, z19.h, z17.h\n"
-      ".inst 0x819126e1  // bfmopa za1.s, p1/M, p1/M, z23.h, z17.h\n"
-      ".inst 0x81912762  // bfmopa za2.s, p1/M, p1/M, z27.h, z17.h\n"
-      ".inst 0x819127e3  // bfmopa za3.s, p1/M, p1/M, z31.h, z17.h\n"
-      ".inst 0x81822600  // bfmopa za0.s, p1/M, p1/M, z16.h, z2.h\n"
-      ".inst 0x81822681  // bfmopa za1.s, p1/M, p1/M, z20.h, z2.h\n"
-      ".inst 0x81822702  // bfmopa za2.s, p1/M, p1/M, z24.h, z2.h\n"
-      ".inst 0x81822783  // bfmopa za3.s, p1/M, p1/M, z28.h, z2.h\n"
+      ".inst 0x81932400  // bfmopa za0.s, p1/M, p1/M, z0.h, z19.h\n"
+      ".inst 0x81932481  // bfmopa za1.s, p1/M, p1/M, z4.h, z19.h\n"
+      ".inst 0x81932502  // bfmopa za2.s, p1/M, p1/M, z8.h, z19.h\n"
+      ".inst 0x81932583  // bfmopa za3.s, p1/M, p1/M, z12.h, z19.h\n"
+      ".inst 0x81962620  // bfmopa za0.s, p1/M, p1/M, z17.h, z22.h\n"
+      ".inst 0x819626a1  // bfmopa za1.s, p1/M, p1/M, z21.h, z22.h\n"
+      ".inst 0x81962722  // bfmopa za2.s, p1/M, p1/M, z25.h, z22.h\n"
+      ".inst 0x819627a3  // bfmopa za3.s, p1/M, p1/M, z29.h, z22.h\n"
+      ".inst 0x81972600  // bfmopa za0.s, p1/M, p1/M, z16.h, z23.h\n"
+      ".inst 0x81972681  // bfmopa za1.s, p1/M, p1/M, z20.h, z23.h\n"
+      ".inst 0x81972702  // bfmopa za2.s, p1/M, p1/M, z24.h, z23.h\n"
+      ".inst 0x81972783  // bfmopa za3.s, p1/M, p1/M, z28.h, z23.h\n"
+      ".inst 0x81822460  // bfmopa za0.s, p1/M, p1/M, z3.h, z2.h\n"
+      ".inst 0x818224e1  // bfmopa za1.s, p1/M, p1/M, z7.h, z2.h\n"
+      ".inst 0x81822562  // bfmopa za2.s, p1/M, p1/M, z11.h, z2.h\n"
+      ".inst 0x818225e3  // bfmopa za3.s, p1/M, p1/M, z15.h, z2.h\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa140a370  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn8.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xa140a373  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn8.b/Z, [x27]\n"
+      "subs x21, x21, #0x1\n"
       "addvl x27, x27, #4\n"
-      "ld1h { z2.h }, p1/Z, [x23]\n"
+      "ld1h { z11.h }, p1/Z, [x23]\n"
       "addvl x23, x23, #1\n"
-      ".inst 0x81822600  // bfmopa za0.s, p1/M, p1/M, z16.h, z2.h\n"
-      ".inst 0x81822681  // bfmopa za1.s, p1/M, p1/M, z20.h, z2.h\n"
-      ".inst 0x81822702  // bfmopa za2.s, p1/M, p1/M, z24.h, z2.h\n"
-      ".inst 0x81822783  // bfmopa za3.s, p1/M, p1/M, z28.h, z2.h\n"
+      ".inst 0x818b2660  // bfmopa za0.s, p1/M, p1/M, z19.h, z11.h\n"
+      ".inst 0x818b26e1  // bfmopa za1.s, p1/M, p1/M, z23.h, z11.h\n"
+      ".inst 0x818b2762  // bfmopa za2.s, p1/M, p1/M, z27.h, z11.h\n"
+      ".inst 0x818b27e3  // bfmopa za3.s, p1/M, p1/M, z31.h, z11.h\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x16, #1, 14f\n"
@@ -241,25 +241,25 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
-      ".inst 0xa041c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
-      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
-      ".inst 0xa042c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
-      ".inst 0xa060c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14]\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xa061c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840403  // mova za3h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa040c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0860418  // mova { z24.s-z27.s }, za0h.s[x12]\n"
+      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa041c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xa042c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa043c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa060c1d8  // st1w { z24.s-z27.s }, pn8.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c1d0  // st1w { z16.s-z19.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 42f\n"
@@ -267,148 +267,148 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
-      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
-      ".inst 0xc0860468  // mova { z8.s-z11.s }, za3h.s[x12]\n"
-      ".inst 0xa060c1d0  // st1w { z16.s-z19.s }, pn8.b, [x14]\n"
+      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xa060c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14]\n"
+      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
+      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      ".inst 0xa061c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 42f\n"
       "14:"  // Store to output array
       "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "add x26, x26, x10, LSL #2\n"  // C += n
       "sub x25, x13, x11\n"
       "ldr x24, [%x[args], %[offsetof_ldcb]]\n"
-      "add x26, x26, x10, LSL #2\n"  // C += n
       "madd x26, x11, x24, x26\n"  // C += m * ldc
       "tbz x16, #2, 27f\n"
       "cntw x23\n"
-      "mov x12, #0x0\n"
       "cmp x25, x23\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Skip activation: Accumulator row 0 loop
-      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z12.s }, p0, [x26]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z13.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z14.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z15.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z7.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 15b\n"
       "16:"  // Store to output array: Skip activation: Accumulator row 0 oddments
       "cbz x20, 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      "st1w { z0.s }, p0, [x26]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 17f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z1.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 17f\n"
-      "st1w { z2.s }, p0, [x26]\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "17:"  // Store to output array: Skip activation: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 19f\n"
       "18:"  // Store to output array: Skip activation: Accumulator row 1 loop
-      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z12.s }, p0, [x26]\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      "st1w { z8.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z13.s }, p0, [x26]\n"
+      "st1w { z9.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z14.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z10.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z15.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z11.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 18b\n"
       "19:"  // Store to output array: Skip activation: Accumulator row 1 oddments
       "cbz x20, 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
+      "st1w { z24.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 20f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z25.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 20f\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "st1w { z26.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "20:"  // Store to output array: Skip activation: Accumulator row 1 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 22f\n"
       "21:"  // Store to output array: Skip activation: Accumulator row 2 loop
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z8.s }, p0, [x26]\n"
+      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z9.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z10.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z11.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z7.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 21b\n"
       "22:"  // Store to output array: Skip activation: Accumulator row 2 oddments
       "cbz x20, 23f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      "st1w { z12.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 23f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z13.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 23f\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "st1w { z14.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "23:"  // Store to output array: Skip activation: Accumulator row 2 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 25f\n"
       "24:"  // Store to output array: Skip activation: Accumulator row 3 loop
       ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
       "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "add x12, x12, #0x4\n"
       "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       "st1w { z19.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 24b\n"
@@ -431,63 +431,63 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
       "b 40f\n"
       "27:"  // Store to output array: Skip activation: End
       "cntw x23\n"
-      "ld1rw { z21.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
-      "mov x12, #0x0\n"
       "cmp x25, x23\n"
-      "ld1rw { z20.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ld1rw { z21.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "ld1rw { z20.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 29f\n"
       "28:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
+      ".inst 0xc1b4cabc  // fclamp { z28.s-z31.s }, z21.s, z20.s\n"
+      "st1w { z28.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z29.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z30.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z19.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z31.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 28b\n"
       "29:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 30f\n"
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc1b4cabc  // fclamp { z28.s-z31.s }, z21.s, z20.s\n"
+      "st1w { z28.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 30f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z29.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 30f\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "st1w { z30.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "30:"  // Store to output array: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 40f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 32f\n"
       "31:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc1b4caa4  // fclamp { z4.s-z7.s }, z21.s, z20.s\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z19.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z7.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 31b\n"
       "32:"  // Store to output array: Accumulator row 1 oddments
@@ -508,100 +508,100 @@ void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, con
       "subs x25, x25, x22\n"
       "beq 40f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 35f\n"
       "34:"  // Store to output array: Accumulator row 2 loop
       ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "cmp x12, x21, LSL #2\n"
       "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "add x12, x12, #0x4\n"
       "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       "st1w { z19.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 34b\n"
       "35:"  // Store to output array: Accumulator row 2 oddments
       "cbz x20, 36f\n"
-      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4cabc  // fclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      "st1w { z28.s }, p0, [x26]\n"
+      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
+      "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 36f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z29.s }, p0, [x26]\n"
+      "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 36f\n"
-      "st1w { z30.s }, p0, [x26]\n"
+      "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "36:"  // Store to output array: Accumulator row 2 oddments: End
       "subs x25, x25, x22\n"
       "beq 40f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x23, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 38f\n"
       "37:"  // Store to output array: Accumulator row 3 loop
       ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "cmp x12, x21, LSL #2\n"
       "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "add x12, x12, #0x4\n"
       "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       "st1w { z19.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 37b\n"
       "38:"  // Store to output array: Accumulator row 3 oddments
       "cbz x20, 39f\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
-      "st1w { z12.s }, p0, [x26]\n"
+      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
+      "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 39f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z13.s }, p0, [x26]\n"
+      "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 39f\n"
-      "st1w { z14.s }, p0, [x26]\n"
+      "st1w { z18.s }, p0, [x26]\n"
       "39:"  // Store to output array: Accumulator row 3 oddments: End
       "40:"  // Store to output array: End
       "tbz x16, #0, 42f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "41:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa041c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840781  // mova za1h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
       ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 41b\n"
       "42:"  // End block
       "incw x10\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #4\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL.hpp
index 82aaa4da49..c7fbede54e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL.hpp
@@ -23,7 +23,7 @@
  */
 #pragma once
 
-#ifdef ARM_COMPUTE_ENABLE_SME2
+#ifdef ARM_COMPUTE_ENABLE_SVE
 
 
 #include "../std_transforms_sme.hpp"
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
 class cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL
 {
 public:
-  typedef __fp16 lhs_operand_type;
-  typedef __fp16 rhs_operand_type;
+  typedef __fp16 operand_type;
   typedef __fp16 result_type;
 
   typedef void (*kern_type)(const __fp16 *const A, const __fp16 *const B, __fp16 *const C, int ldc, const int M, const int N, const int K, const __fp16 *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 1, 4, 2> transforms = {};
+  StdTransformsSME<operand_type, result_type, 1, 4, 2> transforms = {};
 
   cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const CPUInfo *)
   {
@@ -91,4 +90,4 @@ public:
 
 } // namespace arm_gemm
 
-#endif // ARM_COMPUTE_ENABLE_SME2
+#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL/generic.cpp
index 832fd0998a..871b154ad1 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL/generic.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifdef ARM_COMPUTE_ENABLE_SME2
+#ifdef __ARM_FEATURE_SVE
 
 #include "arm_gemm.hpp"
 
@@ -89,7 +89,6 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
 
     const __fp16 *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -109,14 +108,14 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
       ".inst 0xa040c578  // ld1w { z24.s-z27.s }, pn9.b/Z, [x11]\n"
-      ".inst 0xa041c568  // ld1w { z8.s-z11.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
-      ".inst 0xa042c57c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
-      ".inst 0xa043c564  // ld1w { z4.s-z7.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
+      ".inst 0xa041c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
+      ".inst 0xa042c564  // ld1w { z4.s-z7.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
+      ".inst 0xa043c574  // ld1w { z20.s-z23.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
       ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
       "addvl x11, x11, #16\n"
-      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
       "blt 1b\n"
@@ -133,17 +132,17 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
       ".inst 0x257a4770  // whilelt pn8.h, x27, x26, VLx2\n"
-      "fmov z29.h, #0.0\n"
-      "fmov z2.h, #1.0\n"
-      ".inst 0xa01b229f  // ldnt1h { z30.h-z31.h }, p8/Z, [x20, x27, LSL #1]\n"
-      "zip1 z22.h, z30.h, z29.h\n"
-      "zip2 z30.h, z30.h, z29.h\n"
-      "zip1 z20.h, z31.h, z29.h\n"
-      "zip2 z19.h, z31.h, z29.h\n"
-      ".inst 0x81b60040  // fmopa za0.s, p0/M, p0/M, z2.h, z22.h\n"
-      ".inst 0x81be0041  // fmopa za1.s, p0/M, p0/M, z2.h, z30.h\n"
-      ".inst 0x81b40042  // fmopa za2.s, p0/M, p0/M, z2.h, z20.h\n"
-      ".inst 0x81b30043  // fmopa za3.s, p0/M, p0/M, z2.h, z19.h\n"
+      "fmov z6.h, #0.0\n"
+      "fmov z19.h, #1.0\n"
+      ".inst 0xa01b2295  // ldnt1h { z20.h-z21.h }, p8/Z, [x20, x27, LSL #1]\n"
+      "zip1 z23.h, z20.h, z6.h\n"
+      "zip2 z12.h, z20.h, z6.h\n"
+      "zip1 z16.h, z21.h, z6.h\n"
+      "zip2 z8.h, z21.h, z6.h\n"
+      ".inst 0x81b70260  // fmopa za0.s, p0/M, p0/M, z19.h, z23.h\n"
+      ".inst 0x81ac0261  // fmopa za1.s, p0/M, p0/M, z19.h, z12.h\n"
+      ".inst 0x81b00262  // fmopa za2.s, p0/M, p0/M, z19.h, z16.h\n"
+      ".inst 0x81a80263  // fmopa za3.s, p0/M, p0/M, z19.h, z8.h\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x27\n"
       "mov x21, x28\n"
@@ -162,79 +161,79 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
       "add x20, x20, #0x1\n"
       "lsr x20, x20, #0x1\n"
       "lsr x21, x20, #0x2\n"
-      "madd x23, x27, x22, x23\n"  // bptr = B + n * kstride_bytes
       "and x20, x20, #0x3\n"
+      "madd x23, x27, x22, x23\n"  // bptr = B + n * kstride_bytes
       "cbz x21, 8f\n"
       "subs x21, x21, #0x1\n"
-      "ld1h { z20.h }, p0/Z, [x24]\n"
-      ".inst 0xa040a6f0  // ld1h { z16.h-z19.h }, pn9.b/Z, [x23]\n"
-      "ld1h { z31.h }, p0/Z, [x24, #1, MUL VL]\n"
-      ".inst 0xa141a6e2  // ld1h { z2.h, z6.h, z10.h, z14.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      "ld1h { z28.h }, p0/Z, [x24, #2, MUL VL]\n"
-      ".inst 0xa042a6f8  // ld1h { z24.h-z27.h }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      "ld1h { z22.h }, p0/Z, [x24, #3, MUL VL]\n"
+      "ld1h { z21.h }, p0/Z, [x24]\n"
+      ".inst 0xa140a6f8  // ldnt1h { z16.h, z20.h, z24.h, z28.h }, pn9.b/Z, [x23]\n"
+      "ld1h { z29.h }, p0/Z, [x24, #1, MUL VL]\n"
+      ".inst 0xa041a6ed  // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      "ld1h { z4.h }, p0/Z, [x24, #2, MUL VL]\n"
+      ".inst 0xa042a6e1  // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      "ld1h { z25.h }, p0/Z, [x24, #3, MUL VL]\n"
       "addvl x24, x24, #4\n"
-      ".inst 0xa143a6e1  // ld1h { z1.h, z5.h, z9.h, z13.h }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      ".inst 0xa143a6fb  // ldnt1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0x81b00280  // fmopa za0.s, p0/M, p0/M, z20.h, z16.h\n"
+      ".inst 0x81b002a0  // fmopa za0.s, p0/M, p0/M, z21.h, z16.h\n"
       "subs x21, x21, #0x1\n"
-      ".inst 0x81b10281  // fmopa za1.s, p0/M, p0/M, z20.h, z17.h\n"
-      ".inst 0x81b20282  // fmopa za2.s, p0/M, p0/M, z20.h, z18.h\n"
-      ".inst 0x81b30283  // fmopa za3.s, p0/M, p0/M, z20.h, z19.h\n"
-      "ld1h { z20.h }, p0/Z, [x24]\n"
-      ".inst 0x81a203e0  // fmopa za0.s, p0/M, p0/M, z31.h, z2.h\n"
-      ".inst 0xa040a6f0  // ld1h { z16.h-z19.h }, pn9.b/Z, [x23]\n"
-      ".inst 0x81a603e1  // fmopa za1.s, p0/M, p0/M, z31.h, z6.h\n"
-      ".inst 0x81aa03e2  // fmopa za2.s, p0/M, p0/M, z31.h, z10.h\n"
-      ".inst 0x81ae03e3  // fmopa za3.s, p0/M, p0/M, z31.h, z14.h\n"
-      "ld1h { z31.h }, p0/Z, [x24, #1, MUL VL]\n"
-      ".inst 0x81b80380  // fmopa za0.s, p0/M, p0/M, z28.h, z24.h\n"
-      ".inst 0xa141a6e2  // ld1h { z2.h, z6.h, z10.h, z14.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0x81b90381  // fmopa za1.s, p0/M, p0/M, z28.h, z25.h\n"
-      ".inst 0x81ba0382  // fmopa za2.s, p0/M, p0/M, z28.h, z26.h\n"
-      ".inst 0x81bb0383  // fmopa za3.s, p0/M, p0/M, z28.h, z27.h\n"
-      "ld1h { z28.h }, p0/Z, [x24, #2, MUL VL]\n"
-      ".inst 0xa042a6f8  // ld1h { z24.h-z27.h }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      ".inst 0x81a102c0  // fmopa za0.s, p0/M, p0/M, z22.h, z1.h\n"
-      ".inst 0x81a502c1  // fmopa za1.s, p0/M, p0/M, z22.h, z5.h\n"
-      ".inst 0x81a902c2  // fmopa za2.s, p0/M, p0/M, z22.h, z9.h\n"
-      ".inst 0x81ad02c3  // fmopa za3.s, p0/M, p0/M, z22.h, z13.h\n"
-      "ld1h { z22.h }, p0/Z, [x24, #3, MUL VL]\n"
+      ".inst 0x81b402a1  // fmopa za1.s, p0/M, p0/M, z21.h, z20.h\n"
+      ".inst 0x81b802a2  // fmopa za2.s, p0/M, p0/M, z21.h, z24.h\n"
+      ".inst 0x81bc02a3  // fmopa za3.s, p0/M, p0/M, z21.h, z28.h\n"
+      "ld1h { z21.h }, p0/Z, [x24]\n"
+      ".inst 0x81ac03a0  // fmopa za0.s, p0/M, p0/M, z29.h, z12.h\n"
+      ".inst 0xa140a6f0  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn9.b/Z, [x23]\n"
+      ".inst 0x81ad03a1  // fmopa za1.s, p0/M, p0/M, z29.h, z13.h\n"
+      ".inst 0x81ae03a2  // fmopa za2.s, p0/M, p0/M, z29.h, z14.h\n"
+      ".inst 0x81af03a3  // fmopa za3.s, p0/M, p0/M, z29.h, z15.h\n"
+      "ld1h { z29.h }, p0/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x81a00080  // fmopa za0.s, p0/M, p0/M, z4.h, z0.h\n"
+      ".inst 0xa041a6ec  // ld1h { z12.h-z15.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0x81a10081  // fmopa za1.s, p0/M, p0/M, z4.h, z1.h\n"
+      ".inst 0x81a20082  // fmopa za2.s, p0/M, p0/M, z4.h, z2.h\n"
+      ".inst 0x81a30083  // fmopa za3.s, p0/M, p0/M, z4.h, z3.h\n"
+      "ld1h { z4.h }, p0/Z, [x24, #2, MUL VL]\n"
+      ".inst 0xa042a6e0  // ld1h { z0.h-z3.h }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      ".inst 0x81b30320  // fmopa za0.s, p0/M, p0/M, z25.h, z19.h\n"
+      ".inst 0x81b70321  // fmopa za1.s, p0/M, p0/M, z25.h, z23.h\n"
+      ".inst 0x81bb0322  // fmopa za2.s, p0/M, p0/M, z25.h, z27.h\n"
+      ".inst 0x81bf0323  // fmopa za3.s, p0/M, p0/M, z25.h, z31.h\n"
+      "ld1h { z25.h }, p0/Z, [x24, #3, MUL VL]\n"
       "addvl x24, x24, #4\n"
-      ".inst 0xa143a6e1  // ld1h { z1.h, z5.h, z9.h, z13.h }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      ".inst 0xa143a6f3  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0x81b00280  // fmopa za0.s, p0/M, p0/M, z20.h, z16.h\n"
-      ".inst 0x81b10281  // fmopa za1.s, p0/M, p0/M, z20.h, z17.h\n"
-      ".inst 0x81b20282  // fmopa za2.s, p0/M, p0/M, z20.h, z18.h\n"
-      ".inst 0x81b30283  // fmopa za3.s, p0/M, p0/M, z20.h, z19.h\n"
-      ".inst 0x81a203e0  // fmopa za0.s, p0/M, p0/M, z31.h, z2.h\n"
-      ".inst 0x81a603e1  // fmopa za1.s, p0/M, p0/M, z31.h, z6.h\n"
-      ".inst 0x81aa03e2  // fmopa za2.s, p0/M, p0/M, z31.h, z10.h\n"
-      ".inst 0x81ae03e3  // fmopa za3.s, p0/M, p0/M, z31.h, z14.h\n"
-      ".inst 0x81b80380  // fmopa za0.s, p0/M, p0/M, z28.h, z24.h\n"
-      ".inst 0x81b90381  // fmopa za1.s, p0/M, p0/M, z28.h, z25.h\n"
-      ".inst 0x81ba0382  // fmopa za2.s, p0/M, p0/M, z28.h, z26.h\n"
-      ".inst 0x81bb0383  // fmopa za3.s, p0/M, p0/M, z28.h, z27.h\n"
-      ".inst 0x81a102c0  // fmopa za0.s, p0/M, p0/M, z22.h, z1.h\n"
-      ".inst 0x81a502c1  // fmopa za1.s, p0/M, p0/M, z22.h, z5.h\n"
-      ".inst 0x81a902c2  // fmopa za2.s, p0/M, p0/M, z22.h, z9.h\n"
-      ".inst 0x81ad02c3  // fmopa za3.s, p0/M, p0/M, z22.h, z13.h\n"
+      ".inst 0x81b002a0  // fmopa za0.s, p0/M, p0/M, z21.h, z16.h\n"
+      ".inst 0x81b402a1  // fmopa za1.s, p0/M, p0/M, z21.h, z20.h\n"
+      ".inst 0x81b802a2  // fmopa za2.s, p0/M, p0/M, z21.h, z24.h\n"
+      ".inst 0x81bc02a3  // fmopa za3.s, p0/M, p0/M, z21.h, z28.h\n"
+      ".inst 0x81ac03a0  // fmopa za0.s, p0/M, p0/M, z29.h, z12.h\n"
+      ".inst 0x81ad03a1  // fmopa za1.s, p0/M, p0/M, z29.h, z13.h\n"
+      ".inst 0x81ae03a2  // fmopa za2.s, p0/M, p0/M, z29.h, z14.h\n"
+      ".inst 0x81af03a3  // fmopa za3.s, p0/M, p0/M, z29.h, z15.h\n"
+      ".inst 0x81a00080  // fmopa za0.s, p0/M, p0/M, z4.h, z0.h\n"
+      ".inst 0x81a10081  // fmopa za1.s, p0/M, p0/M, z4.h, z1.h\n"
+      ".inst 0x81a20082  // fmopa za2.s, p0/M, p0/M, z4.h, z2.h\n"
+      ".inst 0x81a30083  // fmopa za3.s, p0/M, p0/M, z4.h, z3.h\n"
+      ".inst 0x81b30320  // fmopa za0.s, p0/M, p0/M, z25.h, z19.h\n"
+      ".inst 0x81b70321  // fmopa za1.s, p0/M, p0/M, z25.h, z23.h\n"
+      ".inst 0x81bb0322  // fmopa za2.s, p0/M, p0/M, z25.h, z27.h\n"
+      ".inst 0x81bf0323  // fmopa za3.s, p0/M, p0/M, z25.h, z31.h\n"
       "8:"  // K oddments
       "cbz x20, 10f\n"
       "9:"  // K oddments: Loop
-      "ld1h { z10.h }, p0/Z, [x24]\n"
+      "ld1h { z21.h }, p0/Z, [x24]\n"
       "subs x20, x20, #0x1\n"
       "addvl x24, x24, #1\n"
-      ".inst 0xa140a6f3  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x23]\n"
+      ".inst 0xa140a6f0  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #4\n"
-      ".inst 0x81b30140  // fmopa za0.s, p0/M, p0/M, z10.h, z19.h\n"
-      ".inst 0x81b70141  // fmopa za1.s, p0/M, p0/M, z10.h, z23.h\n"
-      ".inst 0x81bb0142  // fmopa za2.s, p0/M, p0/M, z10.h, z27.h\n"
-      ".inst 0x81bf0143  // fmopa za3.s, p0/M, p0/M, z10.h, z31.h\n"
+      ".inst 0x81b002a0  // fmopa za0.s, p0/M, p0/M, z21.h, z16.h\n"
+      ".inst 0x81b402a1  // fmopa za1.s, p0/M, p0/M, z21.h, z20.h\n"
+      ".inst 0x81b802a2  // fmopa za2.s, p0/M, p0/M, z21.h, z24.h\n"
+      ".inst 0x81bc02a3  // fmopa za3.s, p0/M, p0/M, z21.h, z28.h\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x13, #1, 14f\n"
@@ -242,21 +241,21 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11]\n"
-      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
+      ".inst 0xa040c574  // ld1w { z20.s-z23.s }, pn9.b/Z, [x11]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
       ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
-      ".inst 0xa041c560  // ld1w { z0.s-z3.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
+      ".inst 0xa041c56c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
       ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
       ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
-      ".inst 0xa042c56c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
-      ".inst 0xa043c568  // ld1w { z8.s-z11.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
-      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c568  // ld1w { z8.s-z11.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
+      ".inst 0xa043c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
+      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
       "addvl x11, x11, #16\n"
-      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xa060c554  // st1w { z20.s-z23.s }, pn9.b, [x10]\n"
-      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa060c540  // st1w { z0.s-z3.s }, pn9.b, [x10]\n"
+      ".inst 0xc0840502  // mova za2h.s[x12], { z8.s-z11.s }\n"
       ".inst 0xa061c558  // st1w { z24.s-z27.s }, pn9.b, [x10, #0x4, MUL VL]\n"
-      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
       ".inst 0xa062c544  // st1w { z4.s-z7.s }, pn9.b, [x10, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
@@ -268,16 +267,16 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
       ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c548  // st1w { z8.s-z11.s }, pn9.b, [x10]\n"
+      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
+      ".inst 0xa060c540  // st1w { z0.s-z3.s }, pn9.b, [x10]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c544  // st1w { z4.s-z7.s }, pn9.b, [x10, #0x4, MUL VL]\n"
+      ".inst 0xa061c548  // st1w { z8.s-z11.s }, pn9.b, [x10, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
       ".inst 0xa062c54c  // st1w { z12.s-z15.s }, pn9.b, [x10, #0x8, MUL VL]\n"
-      ".inst 0xa063c550  // st1w { z16.s-z19.s }, pn9.b, [x10, #0xc, MUL VL]\n"
+      ".inst 0xa063c544  // st1w { z4.s-z7.s }, pn9.b, [x10, #0xc, MUL VL]\n"
       "addvl x10, x10, #16\n"
       "blt 13b\n"
       "b 18f\n"
@@ -285,22 +284,22 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
       "ldr x23, [%x[args], %[offsetof_C]]\n"
       "sub x22, x9, x28\n"
       "cntw x21\n"
-      "ld1rh { z21.h }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
+      "ld1rh { z17.h }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
       "ldr x20, [%x[args], %[offsetof_ldcb]]\n"
       ".inst 0x257a4770  // whilelt pn8.h, x27, x26, VLx2\n"
       "cmp x22, x21\n"
-      "ld1rh { z20.h }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ld1rh { z16.h }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
       "mov x12, #0x0\n"
       "csel x22, x22, x21, LT\n"
       "add x23, x23, x27, LSL #1\n"  // C += n
       "madd x23, x28, x20, x23\n"  // C += m * ldc
       "15:"  // Store to output array: Accumulator loop
-      ".inst 0xc0060410  // mova { z16.b-z19.b }, za0h.b[x12, 0:3]\n"
+      ".inst 0xc0060414  // mova { z20.b-z23.b }, za0h.b[x12, 0:3]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc120e20e  // fcvt z14.h, { z16.s-z17.s }\n"
-      ".inst 0xc120e24f  // fcvt z15.h, { z18.s-z19.s }\n"
+      ".inst 0xc120e28e  // fcvt z14.h, { z20.s-z21.s }\n"
+      ".inst 0xc120e2cf  // fcvt z15.h, { z22.s-z23.s }\n"
       "cmp x12, x22, LSL #2\n"
-      ".inst 0xc174c2ae  // fclamp { z14.h-z15.h }, z21.h, z20.h\n"
+      ".inst 0xc170c22e  // fclamp { z14.h-z15.h }, z17.h, z16.h\n"
       ".inst 0xa06022ee  // st1h { z14.h-z15.h }, p8, [x23]\n"
       "add x23, x23, x20\n"
       "blt 15b\n"
@@ -309,15 +308,15 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
       "mov x12, #0x0\n"
       "cntw x20\n"
       "17:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c574  // ld1w { z20.s-z23.s }, pn9.b/Z, [x11]\n"
-      ".inst 0xa041c564  // ld1w { z4.s-z7.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
-      ".inst 0xa042c57c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
-      ".inst 0xa043c56c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
-      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xa040c578  // ld1w { z24.s-z27.s }, pn9.b/Z, [x11]\n"
+      ".inst 0xa041c574  // ld1w { z20.s-z23.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
+      ".inst 0xa042c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
+      ".inst 0xa043c57c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
+      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
       "addvl x11, x11, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0840681  // mova za1h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc0840783  // mova za3h.s[x12], { z28.s-z31.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
       "blt 17b\n"
@@ -339,4 +338,4 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL(const __fp16 *const A, c
 
 }  // namespace arm_gemm
 
-#endif  // ARM_COMPUTE_ENABLE_SME2
+#endif  // __ARM_FEATURE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL.hpp
index 66d32acda7..929b0c97ef 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL.hpp
@@ -23,7 +23,7 @@
  */
 #pragma once
 
-#ifdef ARM_COMPUTE_ENABLE_SME2
+#ifdef ARM_COMPUTE_ENABLE_SVE
 
 
 #include "../std_transforms_sme.hpp"
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL(const __fp16 *const A, c
 class cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL
 {
 public:
-  typedef __fp16 lhs_operand_type;
-  typedef __fp16 rhs_operand_type;
+  typedef __fp16 operand_type;
   typedef __fp16 result_type;
 
   typedef void (*kern_type)(const __fp16 *const A, const __fp16 *const B, __fp16 *const C, int ldc, const int M, const int N, const int K, const __fp16 *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 2, 2, 2> transforms = {};
+  StdTransformsSME<operand_type, result_type, 2, 2, 2> transforms = {};
 
   cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL(const CPUInfo *)
   {
@@ -91,4 +90,4 @@ public:
 
 } // namespace arm_gemm
 
-#endif // ARM_COMPUTE_ENABLE_SME2
+#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL/generic.cpp
index 23e053c0f5..969fb41a92 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL/generic.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifdef ARM_COMPUTE_ENABLE_SME2
+#ifdef __ARM_FEATURE_SVE
 
 #include "arm_gemm.hpp"
 
@@ -89,7 +89,6 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL(const __fp16 *const A, c
 
     const __fp16 *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -98,61 +97,61 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL(const __fp16 *const A, c
   KernelArgs args(A, B, C, ldc, M, N, K, bias, act, accumulate, accumulator_buffer);
 
   __asm__ __volatile__(
-      "ldr x15, [%x[args], %[offsetof_flags]]\n"
+      "ldr x16, [%x[args], %[offsetof_flags]]\n"
       ".inst 0xd503477f  // SMSTART ZA\n"
       "ptrue p1.b\n"
       ".inst 0x25207810  // ptrue pn8.b\n"
+      "ldr x15, [%x[args], %[offsetof_accumulator_buffer]]\n"
       "ldr x14, [%x[args], %[offsetof_accumulator_buffer]]\n"
-      "ldr x13, [%x[args], %[offsetof_accumulator_buffer]]\n"
-      "tbz x15, #0, 2f\n"
+      "tbz x16, #0, 2f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c1d8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x14]\n"
-      ".inst 0xa041c1d4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c1cc  // ld1w { z12.s-z15.s }, pn8.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c1d0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840681  // mova za1h.s[x12], { z20.s-z23.s }\n"
-      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa040c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xa041c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xa042c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa043c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
-      "ldr w11, [%x[args], %[offsetof_M]]\n"
+      "ldr w13, [%x[args], %[offsetof_M]]\n"
+      "mov x11, #0x0\n"
       "mov x10, #0x0\n"
-      "mov x9, #0x0\n"
-      "ldr w28, [%x[args], %[offsetof_N]]\n"
-      "ldr x27, [%x[args], %[offsetof_A]]\n"
+      "ldr w9, [%x[args], %[offsetof_N]]\n"
+      "ldr x28, [%x[args], %[offsetof_A]]\n"
       "3:"  // M and N loop
-      "mov x26, x27\n"
-      "tbnz x15, #0, 4f\n"
+      "mov x27, x28\n"
+      "tbnz x16, #0, 4f\n"
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "whilelt p0.h, x9, x28\n"
-      "fmov z7.h, #0.0\n"
-      "fmov z19.h, #1.0\n"
-      "ld1h { z20.h }, p0/Z, [x20, x9, LSL #1]\n"
-      "zip1 z21.h, z20.h, z7.h\n"
-      "zip2 z30.h, z20.h, z7.h\n"
-      ".inst 0x81b52660  // fmopa za0.s, p1/M, p1/M, z19.h, z21.h\n"
-      ".inst 0x81be2661  // fmopa za1.s, p1/M, p1/M, z19.h, z30.h\n"
-      ".inst 0x81b52662  // fmopa za2.s, p1/M, p1/M, z19.h, z21.h\n"
-      ".inst 0x81be2663  // fmopa za3.s, p1/M, p1/M, z19.h, z30.h\n"
+      "whilelt p0.h, x10, x9\n"
+      "fmov z10.h, #0.0\n"
+      "fmov z11.h, #1.0\n"
+      "ld1h { z18.h }, p0/Z, [x20, x10, LSL #1]\n"
+      "zip1 z2.h, z18.h, z10.h\n"
+      "zip2 z19.h, z18.h, z10.h\n"
+      ".inst 0x81a22560  // fmopa za0.s, p1/M, p1/M, z11.h, z2.h\n"
+      ".inst 0x81b32561  // fmopa za1.s, p1/M, p1/M, z11.h, z19.h\n"
+      ".inst 0x81a22562  // fmopa za2.s, p1/M, p1/M, z11.h, z2.h\n"
+      ".inst 0x81b32563  // fmopa za3.s, p1/M, p1/M, z11.h, z19.h\n"
       "4:"  // Prepare accumulators: Test for last block
-      "mov x20, x9\n"
-      "mov x21, x10\n"
+      "mov x20, x10\n"
+      "mov x21, x11\n"
       "incw x20, ALL, MUL #2\n"
       "incw x21, ALL, MUL #2\n"
-      "cmp x20, x28\n"
-      "mov x20, x15\n"
-      "csel x21, x10, x21, LT\n"
-      "bfm x15, XZR, #0x0, #0x0  // bfc x15, #0x0, #0x1\n"
-      "cmp x21, x11\n"
-      "csel x15, x20, x15, LT\n"
+      "cmp x20, x9\n"
+      "mov x20, x16\n"
+      "csel x21, x11, x21, LT\n"
+      "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
+      "cmp x21, x13\n"
+      "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
       "ldr x23, [%x[args], %[offsetof_B]]\n"
@@ -160,185 +159,294 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL(const __fp16 *const A, c
       "add x20, x20, #0x1\n"
       "lsr x20, x20, #0x1\n"
       "lsr x21, x20, #0x2\n"
-      "madd x23, x9, x22, x23\n"  // bptr = B + n * kstride_bytes
       "and x20, x20, #0x3\n"
+      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
       "cbz x21, 8f\n"
       "subs x21, x21, #0x1\n"
-      ".inst 0xa0402344  // ld1h { z4.h-z5.h }, pn8.b/Z, [x26]\n"
-      ".inst 0xa14022f1  // ld1h { z17.h, z25.h }, pn8.b/Z, [x23]\n"
-      ".inst 0xa0412352  // ld1h { z18.h-z19.h }, pn8.b/Z, [x26, #0x2, MUL VL]\n"
-      ".inst 0xa14122e3  // ld1h { z3.h, z11.h }, pn8.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa042234c  // ld1h { z12.h-z13.h }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa04222fc  // ld1h { z28.h-z29.h }, pn8.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa1432347  // ld1h { z7.h, z15.h }, pn8.b/Z, [x26, #0x6, MUL VL]\n"
-      "addvl x26, x26, #8\n"
-      ".inst 0xa14322f7  // ld1h { z23.h, z31.h }, pn8.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa0402374  // ld1h { z20.h-z21.h }, pn8.b/Z, [x27]\n"
+      ".inst 0xa14022ed  // ldnt1h { z5.h, z13.h }, pn8.b/Z, [x23]\n"
+      ".inst 0xa041236a  // ld1h { z10.h-z11.h }, pn8.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa14122ec  // ldnt1h { z4.h, z12.h }, pn8.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0xa0422372  // ld1h { z18.h-z19.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa04222fb  // ldnt1h { z26.h-z27.h }, pn8.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa1432366  // ld1h { z6.h, z14.h }, pn8.b/Z, [x27, #0x6, MUL VL]\n"
+      "addvl x27, x27, #8\n"
+      ".inst 0xa04322f9  // ldnt1h { z24.h-z25.h }, pn8.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0x81b12480  // fmopa za0.s, p1/M, p1/M, z4.h, z17.h\n"
+      ".inst 0x81a52680  // fmopa za0.s, p1/M, p1/M, z20.h, z5.h\n"
       "subs x21, x21, #0x1\n"
-      ".inst 0x81b92481  // fmopa za1.s, p1/M, p1/M, z4.h, z25.h\n"
-      ".inst 0x81b124a2  // fmopa za2.s, p1/M, p1/M, z5.h, z17.h\n"
-      ".inst 0x81b924a3  // fmopa za3.s, p1/M, p1/M, z5.h, z25.h\n"
-      ".inst 0xa0402344  // ld1h { z4.h-z5.h }, pn8.b/Z, [x26]\n"
-      ".inst 0x81a32640  // fmopa za0.s, p1/M, p1/M, z18.h, z3.h\n"
-      ".inst 0xa14022f1  // ld1h { z17.h, z25.h }, pn8.b/Z, [x23]\n"
-      ".inst 0x81ab2641  // fmopa za1.s, p1/M, p1/M, z18.h, z11.h\n"
-      ".inst 0x81a32662  // fmopa za2.s, p1/M, p1/M, z19.h, z3.h\n"
-      ".inst 0x81ab2663  // fmopa za3.s, p1/M, p1/M, z19.h, z11.h\n"
-      ".inst 0xa0412352  // ld1h { z18.h-z19.h }, pn8.b/Z, [x26, #0x2, MUL VL]\n"
-      ".inst 0x81bc2580  // fmopa za0.s, p1/M, p1/M, z12.h, z28.h\n"
-      ".inst 0xa14122e3  // ld1h { z3.h, z11.h }, pn8.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0x81bd2581  // fmopa za1.s, p1/M, p1/M, z12.h, z29.h\n"
-      ".inst 0x81bc25a2  // fmopa za2.s, p1/M, p1/M, z13.h, z28.h\n"
-      ".inst 0x81bd25a3  // fmopa za3.s, p1/M, p1/M, z13.h, z29.h\n"
-      ".inst 0xa042234c  // ld1h { z12.h-z13.h }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
-      ".inst 0xa04222fc  // ld1h { z28.h-z29.h }, pn8.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0x81b724e0  // fmopa za0.s, p1/M, p1/M, z7.h, z23.h\n"
-      ".inst 0x81bf24e1  // fmopa za1.s, p1/M, p1/M, z7.h, z31.h\n"
-      ".inst 0x81b725e2  // fmopa za2.s, p1/M, p1/M, z15.h, z23.h\n"
-      ".inst 0x81bf25e3  // fmopa za3.s, p1/M, p1/M, z15.h, z31.h\n"
-      ".inst 0xa1432347  // ld1h { z7.h, z15.h }, pn8.b/Z, [x26, #0x6, MUL VL]\n"
-      "addvl x26, x26, #8\n"
-      ".inst 0xa14322f7  // ld1h { z23.h, z31.h }, pn8.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0x81ad2681  // fmopa za1.s, p1/M, p1/M, z20.h, z13.h\n"
+      ".inst 0x81a526a2  // fmopa za2.s, p1/M, p1/M, z21.h, z5.h\n"
+      ".inst 0x81ad26a3  // fmopa za3.s, p1/M, p1/M, z21.h, z13.h\n"
+      ".inst 0xa0402374  // ld1h { z20.h-z21.h }, pn8.b/Z, [x27]\n"
+      ".inst 0x81a42540  // fmopa za0.s, p1/M, p1/M, z10.h, z4.h\n"
+      ".inst 0xa14022e5  // ld1h { z5.h, z13.h }, pn8.b/Z, [x23]\n"
+      ".inst 0x81ac2541  // fmopa za1.s, p1/M, p1/M, z10.h, z12.h\n"
+      ".inst 0x81a42562  // fmopa za2.s, p1/M, p1/M, z11.h, z4.h\n"
+      ".inst 0x81ac2563  // fmopa za3.s, p1/M, p1/M, z11.h, z12.h\n"
+      ".inst 0xa041236a  // ld1h { z10.h-z11.h }, pn8.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0x81ba2640  // fmopa za0.s, p1/M, p1/M, z18.h, z26.h\n"
+      ".inst 0xa14122e4  // ld1h { z4.h, z12.h }, pn8.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0x81bb2641  // fmopa za1.s, p1/M, p1/M, z18.h, z27.h\n"
+      ".inst 0x81ba2662  // fmopa za2.s, p1/M, p1/M, z19.h, z26.h\n"
+      ".inst 0x81bb2663  // fmopa za3.s, p1/M, p1/M, z19.h, z27.h\n"
+      ".inst 0xa0422372  // ld1h { z18.h-z19.h }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa04222fa  // ld1h { z26.h-z27.h }, pn8.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0x81b824c0  // fmopa za0.s, p1/M, p1/M, z6.h, z24.h\n"
+      ".inst 0x81b924c1  // fmopa za1.s, p1/M, p1/M, z6.h, z25.h\n"
+      ".inst 0x81b825c2  // fmopa za2.s, p1/M, p1/M, z14.h, z24.h\n"
+      ".inst 0x81b925c3  // fmopa za3.s, p1/M, p1/M, z14.h, z25.h\n"
+      ".inst 0xa1432366  // ld1h { z6.h, z14.h }, pn8.b/Z, [x27, #0x6, MUL VL]\n"
+      "addvl x27, x27, #8\n"
+      ".inst 0xa04322f8  // ld1h { z24.h-z25.h }, pn8.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0x81b12480  // fmopa za0.s, p1/M, p1/M, z4.h, z17.h\n"
-      ".inst 0x81b92481  // fmopa za1.s, p1/M, p1/M, z4.h, z25.h\n"
-      ".inst 0x81b124a2  // fmopa za2.s, p1/M, p1/M, z5.h, z17.h\n"
-      ".inst 0x81b924a3  // fmopa za3.s, p1/M, p1/M, z5.h, z25.h\n"
-      ".inst 0x81a32640  // fmopa za0.s, p1/M, p1/M, z18.h, z3.h\n"
-      ".inst 0x81ab2641  // fmopa za1.s, p1/M, p1/M, z18.h, z11.h\n"
-      ".inst 0x81a32662  // fmopa za2.s, p1/M, p1/M, z19.h, z3.h\n"
-      ".inst 0x81ab2663  // fmopa za3.s, p1/M, p1/M, z19.h, z11.h\n"
-      ".inst 0x81bc2580  // fmopa za0.s, p1/M, p1/M, z12.h, z28.h\n"
-      ".inst 0x81bd2581  // fmopa za1.s, p1/M, p1/M, z12.h, z29.h\n"
-      ".inst 0x81bc25a2  // fmopa za2.s, p1/M, p1/M, z13.h, z28.h\n"
-      ".inst 0x81bd25a3  // fmopa za3.s, p1/M, p1/M, z13.h, z29.h\n"
-      ".inst 0x81b724e0  // fmopa za0.s, p1/M, p1/M, z7.h, z23.h\n"
-      ".inst 0x81bf24e1  // fmopa za1.s, p1/M, p1/M, z7.h, z31.h\n"
-      ".inst 0x81b725e2  // fmopa za2.s, p1/M, p1/M, z15.h, z23.h\n"
-      ".inst 0x81bf25e3  // fmopa za3.s, p1/M, p1/M, z15.h, z31.h\n"
+      ".inst 0x81a52680  // fmopa za0.s, p1/M, p1/M, z20.h, z5.h\n"
+      ".inst 0x81ad2681  // fmopa za1.s, p1/M, p1/M, z20.h, z13.h\n"
+      ".inst 0x81a526a2  // fmopa za2.s, p1/M, p1/M, z21.h, z5.h\n"
+      ".inst 0x81ad26a3  // fmopa za3.s, p1/M, p1/M, z21.h, z13.h\n"
+      ".inst 0x81a42540  // fmopa za0.s, p1/M, p1/M, z10.h, z4.h\n"
+      ".inst 0x81ac2541  // fmopa za1.s, p1/M, p1/M, z10.h, z12.h\n"
+      ".inst 0x81a42562  // fmopa za2.s, p1/M, p1/M, z11.h, z4.h\n"
+      ".inst 0x81ac2563  // fmopa za3.s, p1/M, p1/M, z11.h, z12.h\n"
+      ".inst 0x81ba2640  // fmopa za0.s, p1/M, p1/M, z18.h, z26.h\n"
+      ".inst 0x81bb2641  // fmopa za1.s, p1/M, p1/M, z18.h, z27.h\n"
+      ".inst 0x81ba2662  // fmopa za2.s, p1/M, p1/M, z19.h, z26.h\n"
+      ".inst 0x81bb2663  // fmopa za3.s, p1/M, p1/M, z19.h, z27.h\n"
+      ".inst 0x81b824c0  // fmopa za0.s, p1/M, p1/M, z6.h, z24.h\n"
+      ".inst 0x81b924c1  // fmopa za1.s, p1/M, p1/M, z6.h, z25.h\n"
+      ".inst 0x81b825c2  // fmopa za2.s, p1/M, p1/M, z14.h, z24.h\n"
+      ".inst 0x81b925c3  // fmopa za3.s, p1/M, p1/M, z14.h, z25.h\n"
       "8:"  // K oddments
       "cbz x20, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa0402346  // ld1h { z6.h-z7.h }, pn8.b/Z, [x26]\n"
+      ".inst 0xa0402374  // ld1h { z20.h-z21.h }, pn8.b/Z, [x27]\n"
       "subs x20, x20, #0x1\n"
-      "addvl x26, x26, #2\n"
-      ".inst 0xa04022e0  // ld1h { z0.h-z1.h }, pn8.b/Z, [x23]\n"
+      "addvl x27, x27, #2\n"
+      ".inst 0xa14022e5  // ld1h { z5.h, z13.h }, pn8.b/Z, [x23]\n"
       "addvl x23, x23, #2\n"
-      ".inst 0x81a024c0  // fmopa za0.s, p1/M, p1/M, z6.h, z0.h\n"
-      ".inst 0x81a124c1  // fmopa za1.s, p1/M, p1/M, z6.h, z1.h\n"
-      ".inst 0x81a024e2  // fmopa za2.s, p1/M, p1/M, z7.h, z0.h\n"
-      ".inst 0x81a124e3  // fmopa za3.s, p1/M, p1/M, z7.h, z1.h\n"
+      ".inst 0x81a52680  // fmopa za0.s, p1/M, p1/M, z20.h, z5.h\n"
+      ".inst 0x81ad2681  // fmopa za1.s, p1/M, p1/M, z20.h, z13.h\n"
+      ".inst 0x81a526a2  // fmopa za2.s, p1/M, p1/M, z21.h, z5.h\n"
+      ".inst 0x81ad26a3  // fmopa za3.s, p1/M, p1/M, z21.h, z13.h\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
-      "tbz x15, #1, 14f\n"
-      "tbz x15, #0, 12f\n"
+      "tbz x16, #1, 14f\n"
+      "tbz x16, #0, 12f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c1d4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x14]\n"
-      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      ".inst 0xa041c1c4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
-      ".inst 0xc0860468  // mova { z8.s-z11.s }, za3h.s[x12]\n"
-      ".inst 0xa042c1dc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c1c0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xa060c1ac  // st1w { z12.s-z15.s }, pn8.b, [x13]\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xa061c1b0  // st1w { z16.s-z19.s }, pn8.b, [x13, #0x4, MUL VL]\n"
-      ".inst 0xc0840403  // mova za3h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa040c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      ".inst 0xa041c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
+      ".inst 0xa042c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa043c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa060c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14]\n"
+      ".inst 0xc0840702  // mova za2h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa061c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xc0840783  // mova za3h.s[x12], { z28.s-z31.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c1b8  // st1w { z24.s-z27.s }, pn8.b, [x13, #0x8, MUL VL]\n"
+      ".inst 0xa062c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c1a8  // st1w { z8.s-z11.s }, pn8.b, [x13, #0xc, MUL VL]\n"
-      "addvl x13, x13, #16\n"
+      ".inst 0xa063c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      "addvl x14, x14, #16\n"
       "blt 11b\n"
-      "b 18f\n"
+      "b 23f\n"
       "12:"  // Store to partial result buffer: Store only
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
-      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c1a8  // st1w { z8.s-z11.s }, pn8.b, [x13]\n"
+      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
+      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xa060c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c1ac  // st1w { z12.s-z15.s }, pn8.b, [x13, #0x4, MUL VL]\n"
+      ".inst 0xa061c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c1bc  // st1w { z28.s-z31.s }, pn8.b, [x13, #0x8, MUL VL]\n"
-      ".inst 0xa063c1b0  // st1w { z16.s-z19.s }, pn8.b, [x13, #0xc, MUL VL]\n"
-      "addvl x13, x13, #16\n"
+      ".inst 0xa062c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1d8  // st1w { z24.s-z27.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      "addvl x14, x14, #16\n"
       "blt 13b\n"
-      "b 18f\n"
+      "b 23f\n"
       "14:"  // Store to output array
-      "ldr x25, [%x[args], %[offsetof_C]]\n"
-      "sub x24, x11, x10\n"
-      "cntw x23, ALL, MUL #2\n"
-      "ld1rh { z18.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
-      "ldr x22, [%x[args], %[offsetof_ldcb]]\n"
-      "whilelt p0.h, x9, x28\n"
-      "cmp x24, x23\n"
-      "ld1rh { z17.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "sub x25, x13, x11\n"
+      "cntw x24\n"
+      "ld1rh { z20.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
+      "ldr x23, [%x[args], %[offsetof_ldcb]]\n"
+      "whilelt p0.h, x10, x9\n"
+      "cmp x25, x24\n"
+      "ld1rh { z19.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "csel x22, x25, x24, LT\n"
       "mov x12, #0x0\n"
-      "mov x21, #0x0\n"
-      "add x25, x25, x9, LSL #1\n"  // C += n
-      "mov x20, #0x2\n"
-      "madd x25, x10, x22, x25\n"  // C += m * ldc
-      "csel x24, x24, x23, LT\n"
-      "15:"  // Store to output array: Accumulator loop
-      ".inst 0xc006000e  // mova { z14.b-z15.b }, za0h.b[x12, 0:1]\n"
+      "add x26, x26, x10, LSL #1\n"  // C += n
+      "lsr x21, x22, #0x2\n"
+      "madd x26, x11, x23, x26\n"  // C += m * ldc
+      "and x20, x22, #0x3\n"
+      "cbz x21, 16f\n"
+      "15:"  // Store to output array: Accumulator row 0 loop
+      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      "fcvt z8.h, p1/m, z8.s\n"
+      "fcvt z9.h, p1/m, z9.s\n"
+      "fcvt z10.h, p1/m, z10.s\n"
+      "fcvt z11.h, p1/m, z11.s\n"
       "add x12, x12, #0x4\n"
-      "cmp x12, x23, LSL #1\n"
-      "add x21, x21, #0x1\n"
-      ".inst 0xc120e1d0  // fcvt z16.h, { z14.s-z15.s }\n"
-      "csel x12, x12, x20, LT\n"
-      "cmp x21, x24\n"
-      ".inst 0x64712650  // fclamp z16.h, z18.h, z17.h\n"
-      "st1h { z16.h }, p0, [x25]\n"
-      "add x25, x25, x22\n"
+      "fcvt z28.h, p1/m, z28.s\n"
+      "fcvt z29.h, p1/m, z29.s\n"
+      "cmp x12, x21, LSL #2\n"
+      "fcvt z30.h, p1/m, z30.s\n"
+      "fcvt z31.h, p1/m, z31.s\n"
+      ".inst 0xc173ca88  // fclamp { z8.h-z11.h }, z20.h, z19.h\n"
+      ".inst 0xc173ca9c  // fclamp { z28.h-z31.h }, z20.h, z19.h\n"
+      "uzp1 z16.h, z8.h, z28.h\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "uzp1 z18.h, z9.h, z29.h\n"
+      "uzp1 z17.h, z10.h, z30.h\n"
+      "uzp1 z16.h, z11.h, z31.h\n"
+      "st1h { z18.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "st1h { z17.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
       "blt 15b\n"
-      "16:"  // Store to output array: End
-      "tbz x15, #0, 18f\n"
+      "16:"  // Store to output array: Accumulator row 0 oddments
+      "cbz x20, 17f\n"
+      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
+      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
+      "fcvt z8.h, p1/m, z8.s\n"
+      "fcvt z9.h, p1/m, z9.s\n"
+      "fcvt z10.h, p1/m, z10.s\n"
+      "fcvt z11.h, p1/m, z11.s\n"
+      "subs x20, x20, #0x1\n"
+      "fcvt z12.h, p1/m, z12.s\n"
+      "fcvt z13.h, p1/m, z13.s\n"
+      "fcvt z14.h, p1/m, z14.s\n"
+      "fcvt z15.h, p1/m, z15.s\n"
+      ".inst 0xc173ca88  // fclamp { z8.h-z11.h }, z20.h, z19.h\n"
+      ".inst 0xc173ca8c  // fclamp { z12.h-z15.h }, z20.h, z19.h\n"
+      "uzp1 z16.h, z8.h, z12.h\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "beq 17f\n"
+      "subs x20, x20, #0x1\n"
+      "uzp1 z16.h, z9.h, z13.h\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "beq 17f\n"
+      "uzp1 z16.h, z10.h, z14.h\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "17:"  // Store to output array: Accumulator row 0 oddments: End
+      "subs x25, x25, x22\n"
+      "beq 21f\n"
+      "whilelt p0.h, x10, x9\n"
+      "cmp x25, x24\n"
+      "csel x20, x25, x24, LT\n"
+      "mov x12, #0x0\n"
+      "lsr x21, x20, #0x2\n"
+      "and x20, x20, #0x3\n"
+      "cbz x21, 19f\n"
+      "18:"  // Store to output array: Accumulator row 1 loop
+      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      "fcvt z0.h, p1/m, z0.s\n"
+      "fcvt z1.h, p1/m, z1.s\n"
+      "fcvt z2.h, p1/m, z2.s\n"
+      "fcvt z3.h, p1/m, z3.s\n"
+      "add x12, x12, #0x4\n"
+      "fcvt z28.h, p1/m, z28.s\n"
+      "fcvt z29.h, p1/m, z29.s\n"
+      "cmp x12, x21, LSL #2\n"
+      "fcvt z30.h, p1/m, z30.s\n"
+      "fcvt z31.h, p1/m, z31.s\n"
+      ".inst 0xc173ca80  // fclamp { z0.h-z3.h }, z20.h, z19.h\n"
+      ".inst 0xc173ca9c  // fclamp { z28.h-z31.h }, z20.h, z19.h\n"
+      "uzp1 z16.h, z0.h, z28.h\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "uzp1 z18.h, z1.h, z29.h\n"
+      "uzp1 z17.h, z2.h, z30.h\n"
+      "uzp1 z16.h, z3.h, z31.h\n"
+      "st1h { z18.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "st1h { z17.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "blt 18b\n"
+      "19:"  // Store to output array: Accumulator row 1 oddments
+      "cbz x20, 20f\n"
+      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
+      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      "fcvt z28.h, p1/m, z28.s\n"
+      "fcvt z29.h, p1/m, z29.s\n"
+      "fcvt z30.h, p1/m, z30.s\n"
+      "fcvt z31.h, p1/m, z31.s\n"
+      "subs x20, x20, #0x1\n"
+      "fcvt z12.h, p1/m, z12.s\n"
+      "fcvt z13.h, p1/m, z13.s\n"
+      "fcvt z14.h, p1/m, z14.s\n"
+      "fcvt z15.h, p1/m, z15.s\n"
+      ".inst 0xc173ca9c  // fclamp { z28.h-z31.h }, z20.h, z19.h\n"
+      ".inst 0xc173ca8c  // fclamp { z12.h-z15.h }, z20.h, z19.h\n"
+      "uzp1 z16.h, z28.h, z12.h\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "beq 20f\n"
+      "subs x20, x20, #0x1\n"
+      "uzp1 z16.h, z29.h, z13.h\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "add x26, x26, x23\n"
+      "beq 20f\n"
+      "uzp1 z16.h, z30.h, z14.h\n"
+      "st1h { z16.h }, p0, [x26]\n"
+      "20:"  // Store to output array: Accumulator row 1 oddments: End
+      "21:"  // Store to output array: End
+      "tbz x16, #0, 23f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
-      "17:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c1cc  // ld1w { z12.s-z15.s }, pn8.b/Z, [x14]\n"
-      ".inst 0xa041c1c4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c1c8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c1dc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xc0840502  // mova za2h.s[x12], { z8.s-z11.s }\n"
+      "22:"  // Store to output array: Refill accumulators: Loop
+      ".inst 0xa040c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xa041c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xa042c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa043c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
       ".inst 0xc0840783  // mova za3h.s[x12], { z28.s-z31.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
-      "blt 17b\n"
-      "18:"  // End block
-      "incw x9, ALL, MUL #2\n"
-      "cmp x9, x28\n"
-      "blt 3b\n"
+      "blt 22b\n"
+      "23:"  // End block
       "incw x10, ALL, MUL #2\n"
-      "mov x9, #0x0\n"
-      "cmp x10, x11\n"
-      "mov x27, x26\n"
+      "cmp x10, x9\n"
+      "blt 3b\n"
+      "incw x11, ALL, MUL #2\n"
+      "mov x10, #0x0\n"
+      "cmp x11, x13\n"
+      "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
       :
       : [args] "r" (&args), [offsetof_A] "I" (offsetof(KernelArgs, A)), [offsetof_B] "I" (offsetof(KernelArgs, B)), [offsetof_C] "I" (offsetof(KernelArgs, C)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_KernelArgs_max] "I" (offsetof(KernelArgs, max)), [offsetof_KernelArgs_min] "I" (offsetof(KernelArgs, min)), [offsetof_M] "I" (offsetof(KernelArgs, M)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_accumulator_buffer] "I" (offsetof(KernelArgs, accumulator_buffer)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_flags] "I" (offsetof(KernelArgs, flags)), [offsetof_kstride_bytes] "I" (offsetof(KernelArgs, kstride_bytes)), [offsetof_ldcb] "I" (offsetof(KernelArgs, ldcb))
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
 }  // namespace arm_gemm
 
-#endif  // ARM_COMPUTE_ENABLE_SME2
+#endif  // __ARM_FEATURE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL.hpp
index f63eb30efd..dba440632d 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL.hpp
@@ -23,7 +23,7 @@
  */
 #pragma once
 
-#ifdef ARM_COMPUTE_ENABLE_SME2
+#ifdef ARM_COMPUTE_ENABLE_SVE
 
 
 #include "../std_transforms_sme.hpp"
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
 class cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL
 {
 public:
-  typedef __fp16 lhs_operand_type;
-  typedef __fp16 rhs_operand_type;
+  typedef __fp16 operand_type;
   typedef __fp16 result_type;
 
   typedef void (*kern_type)(const __fp16 *const A, const __fp16 *const B, __fp16 *const C, int ldc, const int M, const int N, const int K, const __fp16 *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 4, 1, 2> transforms = {};
+  StdTransformsSME<operand_type, result_type, 4, 1, 2> transforms = {};
 
   cls_sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const CPUInfo *)
   {
@@ -91,4 +90,4 @@ public:
 
 } // namespace arm_gemm
 
-#endif // ARM_COMPUTE_ENABLE_SME2
+#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL/generic.cpp
index 3d98d3fe16..ee66b4b95c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL/generic.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifdef ARM_COMPUTE_ENABLE_SME2
+#ifdef __ARM_FEATURE_SVE
 
 #include "arm_gemm.hpp"
 
@@ -89,7 +89,6 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
 
     const __fp16 *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -108,15 +107,15 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5f4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa040c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xa041c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xa042c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa043c5f4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
       "addvl x15, x15, #16\n"
-      ".inst 0xc0840681  // mova za1h.s[x12], { z20.s-z23.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xc0840502  // mova za2h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
       "blt 1b\n"
@@ -134,14 +133,14 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
       "whilelt p0.h, x10, x9\n"
-      "fmov z13.h, #0.0\n"
-      "fmov z27.h, #1.0\n"
-      "ld1h { z14.h }, p0/Z, [x20, x10, LSL #1]\n"
-      "zip1 z30.h, z14.h, z13.h\n"
-      ".inst 0x81be2760  // fmopa za0.s, p1/M, p1/M, z27.h, z30.h\n"
-      ".inst 0x81be2761  // fmopa za1.s, p1/M, p1/M, z27.h, z30.h\n"
-      ".inst 0x81be2762  // fmopa za2.s, p1/M, p1/M, z27.h, z30.h\n"
-      ".inst 0x81be2763  // fmopa za3.s, p1/M, p1/M, z27.h, z30.h\n"
+      "fmov z5.h, #0.0\n"
+      "fmov z18.h, #1.0\n"
+      "ld1h { z31.h }, p0/Z, [x20, x10, LSL #1]\n"
+      "zip1 z15.h, z31.h, z5.h\n"
+      ".inst 0x81af2640  // fmopa za0.s, p1/M, p1/M, z18.h, z15.h\n"
+      ".inst 0x81af2641  // fmopa za1.s, p1/M, p1/M, z18.h, z15.h\n"
+      ".inst 0x81af2642  // fmopa za2.s, p1/M, p1/M, z18.h, z15.h\n"
+      ".inst 0x81af2643  // fmopa za3.s, p1/M, p1/M, z18.h, z15.h\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
@@ -160,79 +159,79 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       "add x20, x20, #0x1\n"
       "lsr x20, x20, #0x1\n"
       "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
       "and x20, x20, #0x3\n"
+      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
       "cbz x21, 8f\n"
       "subs x21, x21, #0x1\n"
-      ".inst 0xa040a778  // ld1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      "ld1h { z1.h }, p1/Z, [x23]\n"
-      ".inst 0xa041a764  // ld1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      "ld1h { z23.h }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa042a77c  // ld1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1h { z19.h }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa043a76c  // ld1h { z12.h-z15.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
+      ".inst 0xa140a773  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x27]\n"
+      "ldnt1h { z17.h }, p1/Z, [x23]\n"
+      ".inst 0xa041a76c  // ld1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      "ldnt1h { z26.h }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0xa042a760  // ld1h { z0.h-z3.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1h { z30.h }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0xa143a770  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1h { z0.h }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1h { z18.h }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0x81a12700  // fmopa za0.s, p1/M, p1/M, z24.h, z1.h\n"
+      ".inst 0x81b12660  // fmopa za0.s, p1/M, p1/M, z19.h, z17.h\n"
       "subs x21, x21, #0x1\n"
-      ".inst 0x81a12721  // fmopa za1.s, p1/M, p1/M, z25.h, z1.h\n"
-      ".inst 0x81a12742  // fmopa za2.s, p1/M, p1/M, z26.h, z1.h\n"
-      ".inst 0x81a12763  // fmopa za3.s, p1/M, p1/M, z27.h, z1.h\n"
-      ".inst 0xa040a778  // ld1h { z24.h-z27.h }, pn9.b/Z, [x27]\n"
-      ".inst 0x81b72480  // fmopa za0.s, p1/M, p1/M, z4.h, z23.h\n"
-      "ld1h { z1.h }, p1/Z, [x23]\n"
-      ".inst 0x81b724a1  // fmopa za1.s, p1/M, p1/M, z5.h, z23.h\n"
-      ".inst 0x81b724c2  // fmopa za2.s, p1/M, p1/M, z6.h, z23.h\n"
-      ".inst 0x81b724e3  // fmopa za3.s, p1/M, p1/M, z7.h, z23.h\n"
-      ".inst 0xa041a764  // ld1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0x81b32780  // fmopa za0.s, p1/M, p1/M, z28.h, z19.h\n"
-      "ld1h { z23.h }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0x81b327a1  // fmopa za1.s, p1/M, p1/M, z29.h, z19.h\n"
-      ".inst 0x81b327c2  // fmopa za2.s, p1/M, p1/M, z30.h, z19.h\n"
-      ".inst 0x81b327e3  // fmopa za3.s, p1/M, p1/M, z31.h, z19.h\n"
-      ".inst 0xa042a77c  // ld1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1h { z19.h }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0x81a02580  // fmopa za0.s, p1/M, p1/M, z12.h, z0.h\n"
-      ".inst 0x81a025a1  // fmopa za1.s, p1/M, p1/M, z13.h, z0.h\n"
-      ".inst 0x81a025c2  // fmopa za2.s, p1/M, p1/M, z14.h, z0.h\n"
-      ".inst 0x81a025e3  // fmopa za3.s, p1/M, p1/M, z15.h, z0.h\n"
-      ".inst 0xa043a76c  // ld1h { z12.h-z15.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
+      ".inst 0x81b126e1  // fmopa za1.s, p1/M, p1/M, z23.h, z17.h\n"
+      ".inst 0x81b12762  // fmopa za2.s, p1/M, p1/M, z27.h, z17.h\n"
+      ".inst 0x81b127e3  // fmopa za3.s, p1/M, p1/M, z31.h, z17.h\n"
+      ".inst 0xa140a773  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x27]\n"
+      ".inst 0x81ba2580  // fmopa za0.s, p1/M, p1/M, z12.h, z26.h\n"
+      "ld1h { z17.h }, p1/Z, [x23]\n"
+      ".inst 0x81ba25a1  // fmopa za1.s, p1/M, p1/M, z13.h, z26.h\n"
+      ".inst 0x81ba25c2  // fmopa za2.s, p1/M, p1/M, z14.h, z26.h\n"
+      ".inst 0x81ba25e3  // fmopa za3.s, p1/M, p1/M, z15.h, z26.h\n"
+      ".inst 0xa041a76c  // ld1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0x81be2400  // fmopa za0.s, p1/M, p1/M, z0.h, z30.h\n"
+      "ld1h { z26.h }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0x81be2421  // fmopa za1.s, p1/M, p1/M, z1.h, z30.h\n"
+      ".inst 0x81be2442  // fmopa za2.s, p1/M, p1/M, z2.h, z30.h\n"
+      ".inst 0x81be2463  // fmopa za3.s, p1/M, p1/M, z3.h, z30.h\n"
+      ".inst 0xa042a760  // ld1h { z0.h-z3.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
+      "ld1h { z30.h }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0x81b22600  // fmopa za0.s, p1/M, p1/M, z16.h, z18.h\n"
+      ".inst 0x81b22681  // fmopa za1.s, p1/M, p1/M, z20.h, z18.h\n"
+      ".inst 0x81b22702  // fmopa za2.s, p1/M, p1/M, z24.h, z18.h\n"
+      ".inst 0x81b22783  // fmopa za3.s, p1/M, p1/M, z28.h, z18.h\n"
+      ".inst 0xa143a770  // ld1h { z16.h, z20.h, z24.h, z28.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1h { z0.h }, p1/Z, [x23, #3, MUL VL]\n"
+      "ld1h { z18.h }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0x81a12700  // fmopa za0.s, p1/M, p1/M, z24.h, z1.h\n"
-      ".inst 0x81a12721  // fmopa za1.s, p1/M, p1/M, z25.h, z1.h\n"
-      ".inst 0x81a12742  // fmopa za2.s, p1/M, p1/M, z26.h, z1.h\n"
-      ".inst 0x81a12763  // fmopa za3.s, p1/M, p1/M, z27.h, z1.h\n"
-      ".inst 0x81b72480  // fmopa za0.s, p1/M, p1/M, z4.h, z23.h\n"
-      ".inst 0x81b724a1  // fmopa za1.s, p1/M, p1/M, z5.h, z23.h\n"
-      ".inst 0x81b724c2  // fmopa za2.s, p1/M, p1/M, z6.h, z23.h\n"
-      ".inst 0x81b724e3  // fmopa za3.s, p1/M, p1/M, z7.h, z23.h\n"
-      ".inst 0x81b32780  // fmopa za0.s, p1/M, p1/M, z28.h, z19.h\n"
-      ".inst 0x81b327a1  // fmopa za1.s, p1/M, p1/M, z29.h, z19.h\n"
-      ".inst 0x81b327c2  // fmopa za2.s, p1/M, p1/M, z30.h, z19.h\n"
-      ".inst 0x81b327e3  // fmopa za3.s, p1/M, p1/M, z31.h, z19.h\n"
-      ".inst 0x81a02580  // fmopa za0.s, p1/M, p1/M, z12.h, z0.h\n"
-      ".inst 0x81a025a1  // fmopa za1.s, p1/M, p1/M, z13.h, z0.h\n"
-      ".inst 0x81a025c2  // fmopa za2.s, p1/M, p1/M, z14.h, z0.h\n"
-      ".inst 0x81a025e3  // fmopa za3.s, p1/M, p1/M, z15.h, z0.h\n"
+      ".inst 0x81b12660  // fmopa za0.s, p1/M, p1/M, z19.h, z17.h\n"
+      ".inst 0x81b126e1  // fmopa za1.s, p1/M, p1/M, z23.h, z17.h\n"
+      ".inst 0x81b12762  // fmopa za2.s, p1/M, p1/M, z27.h, z17.h\n"
+      ".inst 0x81b127e3  // fmopa za3.s, p1/M, p1/M, z31.h, z17.h\n"
+      ".inst 0x81ba2580  // fmopa za0.s, p1/M, p1/M, z12.h, z26.h\n"
+      ".inst 0x81ba25a1  // fmopa za1.s, p1/M, p1/M, z13.h, z26.h\n"
+      ".inst 0x81ba25c2  // fmopa za2.s, p1/M, p1/M, z14.h, z26.h\n"
+      ".inst 0x81ba25e3  // fmopa za3.s, p1/M, p1/M, z15.h, z26.h\n"
+      ".inst 0x81be2400  // fmopa za0.s, p1/M, p1/M, z0.h, z30.h\n"
+      ".inst 0x81be2421  // fmopa za1.s, p1/M, p1/M, z1.h, z30.h\n"
+      ".inst 0x81be2442  // fmopa za2.s, p1/M, p1/M, z2.h, z30.h\n"
+      ".inst 0x81be2463  // fmopa za3.s, p1/M, p1/M, z3.h, z30.h\n"
+      ".inst 0x81b22600  // fmopa za0.s, p1/M, p1/M, z16.h, z18.h\n"
+      ".inst 0x81b22681  // fmopa za1.s, p1/M, p1/M, z20.h, z18.h\n"
+      ".inst 0x81b22702  // fmopa za2.s, p1/M, p1/M, z24.h, z18.h\n"
+      ".inst 0x81b22783  // fmopa za3.s, p1/M, p1/M, z28.h, z18.h\n"
       "8:"  // K oddments
       "cbz x20, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa040a768  // ld1h { z8.h-z11.h }, pn9.b/Z, [x27]\n"
+      ".inst 0xa140a773  // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x27]\n"
       "subs x20, x20, #0x1\n"
       "addvl x27, x27, #4\n"
-      "ld1h { z12.h }, p1/Z, [x23]\n"
+      "ld1h { z17.h }, p1/Z, [x23]\n"
       "addvl x23, x23, #1\n"
-      ".inst 0x81ac2500  // fmopa za0.s, p1/M, p1/M, z8.h, z12.h\n"
-      ".inst 0x81ac2521  // fmopa za1.s, p1/M, p1/M, z9.h, z12.h\n"
-      ".inst 0x81ac2542  // fmopa za2.s, p1/M, p1/M, z10.h, z12.h\n"
-      ".inst 0x81ac2563  // fmopa za3.s, p1/M, p1/M, z11.h, z12.h\n"
+      ".inst 0x81b12660  // fmopa za0.s, p1/M, p1/M, z19.h, z17.h\n"
+      ".inst 0x81b126e1  // fmopa za1.s, p1/M, p1/M, z23.h, z17.h\n"
+      ".inst 0x81b12762  // fmopa za2.s, p1/M, p1/M, z27.h, z17.h\n"
+      ".inst 0x81b127e3  // fmopa za3.s, p1/M, p1/M, z31.h, z17.h\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x16, #1, 14f\n"
@@ -240,25 +239,25 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
+      ".inst 0xa040c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
       ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
-      ".inst 0xa042c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa041c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
+      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xa042c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
       ".inst 0xa043c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
       "addvl x15, x15, #16\n"
-      ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
-      ".inst 0xa060c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14]\n"
-      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa060c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14]\n"
+      ".inst 0xc0840502  // mova za2h.s[x12], { z8.s-z11.s }\n"
       ".inst 0xa061c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa062c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa063c5d8  // st1w { z24.s-z27.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 29f\n"
@@ -266,16 +265,16 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14]\n"
+      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
+      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
+      ".inst 0xa060c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa061c5d8  // st1w { z24.s-z27.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c5d8  // st1w { z24.s-z27.s }, pn9.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 29f\n"
@@ -283,11 +282,11 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       "ldr x26, [%x[args], %[offsetof_C]]\n"
       "sub x25, x13, x11\n"
       "cntw x24\n"
-      "ld1rh { z21.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
+      "ld1rh { z29.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
       "ldr x23, [%x[args], %[offsetof_ldcb]]\n"
       "whilelt p0.s, x10, x9\n"
       "cmp x25, x24\n"
-      "ld1rh { z20.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ld1rh { z28.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
       "csel x22, x25, x24, LT\n"
       "mov x12, #0x0\n"
       "add x26, x26, x10, LSL #1\n"  // C += n
@@ -296,21 +295,21 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
       "add x12, x12, #0x4\n"
-      "fcvt z28.h, p1/m, z28.s\n"
-      "fcvt z29.h, p1/m, z29.s\n"
+      "fcvt z0.h, p1/m, z0.s\n"
+      "fcvt z1.h, p1/m, z1.s\n"
+      "fcvt z2.h, p1/m, z2.s\n"
+      "fcvt z3.h, p1/m, z3.s\n"
       "cmp x12, x21, LSL #2\n"
-      "fcvt z30.h, p1/m, z30.s\n"
-      "fcvt z31.h, p1/m, z31.s\n"
-      ".inst 0xc174cabc  // fclamp { z28.h-z31.h }, z21.h, z20.h\n"
-      "st1h { z28.s }, p0, [x26]\n"
+      ".inst 0xc17ccba0  // fclamp { z0.h-z3.h }, z29.h, z28.h\n"
+      "st1h { z0.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z29.s }, p0, [x26]\n"
+      "st1h { z1.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z30.s }, p0, [x26]\n"
+      "st1h { z2.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z31.s }, p0, [x26]\n"
+      "st1h { z3.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "blt 15b\n"
       "16:"  // Store to output array: Accumulator row 0 oddments
@@ -321,7 +320,7 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       "fcvt z17.h, p1/m, z17.s\n"
       "fcvt z18.h, p1/m, z18.s\n"
       "fcvt z19.h, p1/m, z19.s\n"
-      ".inst 0xc174cab0  // fclamp { z16.h-z19.h }, z21.h, z20.h\n"
+      ".inst 0xc17ccbb0  // fclamp { z16.h-z19.h }, z29.h, z28.h\n"
       "st1h { z16.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 17f\n"
@@ -334,152 +333,155 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
       "17:"  // Store to output array: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x24, LT\n"
+      "mov x12, #0x0\n"
       "lsr x21, x22, #0x2\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 19f\n"
       "18:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
+      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
       "add x12, x12, #0x4\n"
-      "fcvt z0.h, p1/m, z0.s\n"
-      "fcvt z1.h, p1/m, z1.s\n"
+      "fcvt z24.h, p1/m, z24.s\n"
+      "fcvt z25.h, p1/m, z25.s\n"
+      "fcvt z26.h, p1/m, z26.s\n"
+      "fcvt z27.h, p1/m, z27.s\n"
       "cmp x12, x21, LSL #2\n"
-      "fcvt z2.h, p1/m, z2.s\n"
-      "fcvt z3.h, p1/m, z3.s\n"
-      ".inst 0xc174caa0  // fclamp { z0.h-z3.h }, z21.h, z20.h\n"
-      "st1h { z0.s }, p0, [x26]\n"
+      ".inst 0xc17ccbb8  // fclamp { z24.h-z27.h }, z29.h, z28.h\n"
+      "st1h { z24.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z1.s }, p0, [x26]\n"
+      "st1h { z25.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z2.s }, p0, [x26]\n"
+      "st1h { z26.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z3.s }, p0, [x26]\n"
+      "st1h { z27.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "blt 18b\n"
       "19:"  // Store to output array: Accumulator row 1 oddments
       "cbz x20, 20f\n"
-      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
+      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      "fcvt z24.h, p1/m, z24.s\n"
-      "fcvt z25.h, p1/m, z25.s\n"
-      "fcvt z26.h, p1/m, z26.s\n"
-      "fcvt z27.h, p1/m, z27.s\n"
-      ".inst 0xc174cab8  // fclamp { z24.h-z27.h }, z21.h, z20.h\n"
-      "st1h { z24.s }, p0, [x26]\n"
+      "fcvt z0.h, p1/m, z0.s\n"
+      "fcvt z1.h, p1/m, z1.s\n"
+      "fcvt z2.h, p1/m, z2.s\n"
+      "fcvt z3.h, p1/m, z3.s\n"
+      ".inst 0xc17ccba0  // fclamp { z0.h-z3.h }, z29.h, z28.h\n"
+      "st1h { z0.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 20f\n"
       "subs x20, x20, #0x1\n"
-      "st1h { z25.s }, p0, [x26]\n"
+      "st1h { z1.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 20f\n"
-      "st1h { z26.s }, p0, [x26]\n"
+      "st1h { z2.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "20:"  // Store to output array: Accumulator row 1 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x24, LT\n"
+      "mov x12, #0x0\n"
       "lsr x21, x22, #0x2\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 22f\n"
       "21:"  // Store to output array: Accumulator row 2 loop
-      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
       "add x12, x12, #0x4\n"
-      "fcvt z16.h, p1/m, z16.s\n"
-      "fcvt z17.h, p1/m, z17.s\n"
+      "fcvt z20.h, p1/m, z20.s\n"
+      "fcvt z21.h, p1/m, z21.s\n"
+      "fcvt z22.h, p1/m, z22.s\n"
+      "fcvt z23.h, p1/m, z23.s\n"
       "cmp x12, x21, LSL #2\n"
-      "fcvt z18.h, p1/m, z18.s\n"
-      "fcvt z19.h, p1/m, z19.s\n"
-      ".inst 0xc174cab0  // fclamp { z16.h-z19.h }, z21.h, z20.h\n"
-      "st1h { z16.s }, p0, [x26]\n"
+      ".inst 0xc17ccbb4  // fclamp { z20.h-z23.h }, z29.h, z28.h\n"
+      "st1h { z20.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z17.s }, p0, [x26]\n"
+      "st1h { z21.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z18.s }, p0, [x26]\n"
+      "st1h { z22.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z19.s }, p0, [x26]\n"
+      "st1h { z23.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "blt 21b\n"
       "22:"  // Store to output array: Accumulator row 2 oddments
       "cbz x20, 23f\n"
-      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      "fcvt z28.h, p1/m, z28.s\n"
-      "fcvt z29.h, p1/m, z29.s\n"
-      "fcvt z30.h, p1/m, z30.s\n"
-      "fcvt z31.h, p1/m, z31.s\n"
-      ".inst 0xc174cabc  // fclamp { z28.h-z31.h }, z21.h, z20.h\n"
-      "st1h { z28.s }, p0, [x26]\n"
+      "fcvt z12.h, p1/m, z12.s\n"
+      "fcvt z13.h, p1/m, z13.s\n"
+      "fcvt z14.h, p1/m, z14.s\n"
+      "fcvt z15.h, p1/m, z15.s\n"
+      ".inst 0xc17ccbac  // fclamp { z12.h-z15.h }, z29.h, z28.h\n"
+      "st1h { z12.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 23f\n"
       "subs x20, x20, #0x1\n"
-      "st1h { z29.s }, p0, [x26]\n"
+      "st1h { z13.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 23f\n"
-      "st1h { z30.s }, p0, [x26]\n"
+      "st1h { z14.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "23:"  // Store to output array: Accumulator row 2 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x24, LT\n"
+      "mov x12, #0x0\n"
       "lsr x21, x20, #0x2\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 25f\n"
       "24:"  // Store to output array: Accumulator row 3 loop
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
       "add x12, x12, #0x4\n"
-      "fcvt z28.h, p1/m, z28.s\n"
-      "fcvt z29.h, p1/m, z29.s\n"
+      "fcvt z4.h, p1/m, z4.s\n"
+      "fcvt z5.h, p1/m, z5.s\n"
+      "fcvt z6.h, p1/m, z6.s\n"
+      "fcvt z7.h, p1/m, z7.s\n"
       "cmp x12, x21, LSL #2\n"
-      "fcvt z30.h, p1/m, z30.s\n"
-      "fcvt z31.h, p1/m, z31.s\n"
-      ".inst 0xc174cabc  // fclamp { z28.h-z31.h }, z21.h, z20.h\n"
-      "st1h { z28.s }, p0, [x26]\n"
+      ".inst 0xc17ccba4  // fclamp { z4.h-z7.h }, z29.h, z28.h\n"
+      "st1h { z4.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z29.s }, p0, [x26]\n"
+      "st1h { z5.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z30.s }, p0, [x26]\n"
+      "st1h { z6.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1h { z31.s }, p0, [x26]\n"
+      "st1h { z7.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "blt 24b\n"
       "25:"  // Store to output array: Accumulator row 3 oddments
       "cbz x20, 26f\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      "fcvt z28.h, p1/m, z28.s\n"
-      "fcvt z29.h, p1/m, z29.s\n"
-      "fcvt z30.h, p1/m, z30.s\n"
-      "fcvt z31.h, p1/m, z31.s\n"
-      ".inst 0xc174cabc  // fclamp { z28.h-z31.h }, z21.h, z20.h\n"
-      "st1h { z28.s }, p0, [x26]\n"
+      "fcvt z4.h, p1/m, z4.s\n"
+      "fcvt z5.h, p1/m, z5.s\n"
+      "fcvt z6.h, p1/m, z6.s\n"
+      "fcvt z7.h, p1/m, z7.s\n"
+      ".inst 0xc17ccba4  // fclamp { z4.h-z7.h }, z29.h, z28.h\n"
+      "st1h { z4.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 26f\n"
       "subs x20, x20, #0x1\n"
-      "st1h { z29.s }, p0, [x26]\n"
+      "st1h { z5.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 26f\n"
-      "st1h { z30.s }, p0, [x26]\n"
+      "st1h { z6.s }, p0, [x26]\n"
       "26:"  // Store to output array: Accumulator row 3 oddments: End
       "27:"  // Store to output array: End
       "tbz x16, #0, 29f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "28:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xa040c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n"
       ".inst 0xa041c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa042c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa043c5f4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
       "addvl x15, x15, #16\n"
       ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
       "blt 28b\n"
@@ -501,4 +503,4 @@ void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, c
 
 }  // namespace arm_gemm
 
-#endif  // ARM_COMPUTE_ENABLE_SME2
+#endif  // __ARM_FEATURE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL.hpp
index e3e2a0639f..9486319cfb 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
 class cls_sme2_interleaved_nomerge_fp32_mopa_1VLx4VL
 {
 public:
-  typedef float lhs_operand_type;
-  typedef float rhs_operand_type;
+  typedef float operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const float *const A, const float *const B, float *const C, int ldc, const int M, const int N, const int K, const float *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_fp32_mopa_1VLx4VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 1, 4, 1> transforms = {};
+  StdTransformsSME<operand_type, result_type, 1, 4, 1> transforms = {};
 
   cls_sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp
index 1f9f08e401..19e5d52b53 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
         B(B), kstride_bytes(K * sizeof(float)),
         C(C), ldcb(ldc * sizeof(float)),
         M(M), N(N), K(K),
+        n_loops((K - 1) / 2), n_tail_iters((K - 1) % 2),
         min(-std::numeric_limits<float>::infinity()),
         max(std::numeric_limits<float>::infinity()),
         bias(bias),
@@ -87,13 +88,12 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
     const long kstride_bytes;
     float *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
     float min = -std::numeric_limits<float>::infinity();
     float max = std::numeric_limits<float>::infinity();
 
     const float *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -112,17 +112,17 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5dc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x14]\n"
+      ".inst 0xa040c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14]\n"
+      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
       ".inst 0xa041c5d8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c5d4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
-      "addvl x14, x14, #16\n"
       ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa042c5d4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
       ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
-      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa043c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
+      ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x14, x14, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w11, [%x[args], %[offsetof_M]]\n"
@@ -137,101 +137,101 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "fmov z15.s, #1.0\n"
-      ".inst 0xa109c280  // ld1w { z0.s, z4.s, z8.s, z12.s }, p8/Z, [x20, x9, LSL #2]\n"
-      ".inst 0x808001e0  // fmopa za0.s, p0/M, p0/M, z15.s, z0.s\n"
-      ".inst 0x808401e1  // fmopa za1.s, p0/M, p0/M, z15.s, z4.s\n"
-      ".inst 0x808801e2  // fmopa za2.s, p0/M, p0/M, z15.s, z8.s\n"
-      ".inst 0x808c01e3  // fmopa za3.s, p0/M, p0/M, z15.s, z12.s\n"
+      "fmov z6.s, #1.0\n"
+      ".inst 0xa009c29d  // ldnt1w { z28.s-z31.s }, p8/Z, [x20, x9, LSL #2]\n"
+      ".inst 0x809c00c0  // fmopa za0.s, p0/M, p0/M, z6.s, z28.s\n"
+      ".inst 0x809d00c1  // fmopa za1.s, p0/M, p0/M, z6.s, z29.s\n"
+      ".inst 0x809e00c2  // fmopa za2.s, p0/M, p0/M, z6.s, z30.s\n"
+      ".inst 0x809f00c3  // fmopa za3.s, p0/M, p0/M, z6.s, z31.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x9\n"
       "mov x21, x10\n"
       "incw x20, ALL, MUL #4\n"
       "incw x21\n"
       "cmp x20, x28\n"
-      "mov x20, x15\n"
       "csel x21, x10, x21, LT\n"
+      "mov x20, x15\n"
       "bfm x15, XZR, #0x0, #0x0  // bfc x15, #0x0, #0x1\n"
       "cmp x21, x11\n"
       "csel x15, x20, x15, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
-      "lsr x21, x20, #0x2\n"
-      "and x20, x20, #0x3\n"
-      "madd x23, x9, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      "ld1w { z20.s }, p0/Z, [x26]\n"
-      ".inst 0xa140c6f3  // ld1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x23]\n"
-      "ld1w { z4.s }, p0/Z, [x26, #1, MUL VL]\n"
-      ".inst 0xa041c6ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      "ld1w { z29.s }, p0/Z, [x26, #2, MUL VL]\n"
-      ".inst 0xa142c6f2  // ld1w { z18.s, z22.s, z26.s, z30.s }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      "ld1w { z2.s }, p0/Z, [x26, #3, MUL VL]\n"
+      "lsr x23, x20, #0x2\n"
+      "and x22, x20, #0x3\n"
+      "ldr x21, [%x[args], %[offsetof_B]]\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x21, x9, x20, x21\n"  // bptr = B + n * kstride_bytes
+      "cbz x23, 8f\n"
+      "subs x23, x23, #0x1\n"
+      "ld1w { z28.s }, p0/Z, [x26]\n"
+      ".inst 0xa040c6a9  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x21]\n"
+      "ld1w { z22.s }, p0/Z, [x26, #1, MUL VL]\n"
+      ".inst 0xa041c6ad  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n"
+      "ld1w { z30.s }, p0/Z, [x26, #2, MUL VL]\n"
+      ".inst 0xa042c6a5  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x21, #0x8, MUL VL]\n"
+      "ld1w { z20.s }, p0/Z, [x26, #3, MUL VL]\n"
       "addvl x26, x26, #4\n"
-      ".inst 0xa043c6e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
-      "addvl x23, x23, #16\n"
+      ".inst 0xa143c6bb  // ldnt1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x21, #0xc, MUL VL]\n"
+      "addvl x21, x21, #16\n"
       "ble 7f\n"
       "6:"  // K loop
+      ".inst 0x80880380  // fmopa za0.s, p0/M, p0/M, z28.s, z8.s\n"
+      "subs x23, x23, #0x1\n"
+      ".inst 0x80890381  // fmopa za1.s, p0/M, p0/M, z28.s, z9.s\n"
+      ".inst 0x808a0382  // fmopa za2.s, p0/M, p0/M, z28.s, z10.s\n"
+      ".inst 0x808b0383  // fmopa za3.s, p0/M, p0/M, z28.s, z11.s\n"
+      "ld1w { z28.s }, p0/Z, [x26]\n"
+      ".inst 0x808c02c0  // fmopa za0.s, p0/M, p0/M, z22.s, z12.s\n"
+      ".inst 0xa040c6a9  // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x21]\n"
+      ".inst 0x808d02c1  // fmopa za1.s, p0/M, p0/M, z22.s, z13.s\n"
+      ".inst 0x808e02c2  // fmopa za2.s, p0/M, p0/M, z22.s, z14.s\n"
+      ".inst 0x808f02c3  // fmopa za3.s, p0/M, p0/M, z22.s, z15.s\n"
+      "ld1w { z22.s }, p0/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x808403c0  // fmopa za0.s, p0/M, p0/M, z30.s, z4.s\n"
+      ".inst 0xa041c6ad  // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n"
+      ".inst 0x808503c1  // fmopa za1.s, p0/M, p0/M, z30.s, z5.s\n"
+      ".inst 0x808603c2  // fmopa za2.s, p0/M, p0/M, z30.s, z6.s\n"
+      ".inst 0x808703c3  // fmopa za3.s, p0/M, p0/M, z30.s, z7.s\n"
+      "ld1w { z30.s }, p0/Z, [x26, #2, MUL VL]\n"
+      ".inst 0xa042c6a5  // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x21, #0x8, MUL VL]\n"
       ".inst 0x80930280  // fmopa za0.s, p0/M, p0/M, z20.s, z19.s\n"
-      "subs x21, x21, #0x1\n"
       ".inst 0x80970281  // fmopa za1.s, p0/M, p0/M, z20.s, z23.s\n"
       ".inst 0x809b0282  // fmopa za2.s, p0/M, p0/M, z20.s, z27.s\n"
       ".inst 0x809f0283  // fmopa za3.s, p0/M, p0/M, z20.s, z31.s\n"
-      "ld1w { z20.s }, p0/Z, [x26]\n"
-      ".inst 0x808c0080  // fmopa za0.s, p0/M, p0/M, z4.s, z12.s\n"
-      ".inst 0xa140c6f3  // ld1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x23]\n"
-      ".inst 0x808d0081  // fmopa za1.s, p0/M, p0/M, z4.s, z13.s\n"
-      ".inst 0x808e0082  // fmopa za2.s, p0/M, p0/M, z4.s, z14.s\n"
-      ".inst 0x808f0083  // fmopa za3.s, p0/M, p0/M, z4.s, z15.s\n"
-      "ld1w { z4.s }, p0/Z, [x26, #1, MUL VL]\n"
-      ".inst 0x809203a0  // fmopa za0.s, p0/M, p0/M, z29.s, z18.s\n"
-      ".inst 0xa041c6ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0x809603a1  // fmopa za1.s, p0/M, p0/M, z29.s, z22.s\n"
-      ".inst 0x809a03a2  // fmopa za2.s, p0/M, p0/M, z29.s, z26.s\n"
-      ".inst 0x809e03a3  // fmopa za3.s, p0/M, p0/M, z29.s, z30.s\n"
-      "ld1w { z29.s }, p0/Z, [x26, #2, MUL VL]\n"
-      ".inst 0xa142c6f2  // ld1w { z18.s, z22.s, z26.s, z30.s }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      ".inst 0x80880040  // fmopa za0.s, p0/M, p0/M, z2.s, z8.s\n"
-      ".inst 0x80890041  // fmopa za1.s, p0/M, p0/M, z2.s, z9.s\n"
-      ".inst 0x808a0042  // fmopa za2.s, p0/M, p0/M, z2.s, z10.s\n"
-      ".inst 0x808b0043  // fmopa za3.s, p0/M, p0/M, z2.s, z11.s\n"
-      "ld1w { z2.s }, p0/Z, [x26, #3, MUL VL]\n"
+      "ld1w { z20.s }, p0/Z, [x26, #3, MUL VL]\n"
       "addvl x26, x26, #4\n"
-      ".inst 0xa043c6e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
-      "addvl x23, x23, #16\n"
+      ".inst 0xa143c6bb  // ldnt1w { z19.s, z23.s, z27.s, z31.s }, pn9.b/Z, [x21, #0xc, MUL VL]\n"
+      "addvl x21, x21, #16\n"
       "bgt 6b\n"
       "7:"  // K loop tail
+      ".inst 0x80880380  // fmopa za0.s, p0/M, p0/M, z28.s, z8.s\n"
+      ".inst 0x80890381  // fmopa za1.s, p0/M, p0/M, z28.s, z9.s\n"
+      ".inst 0x808a0382  // fmopa za2.s, p0/M, p0/M, z28.s, z10.s\n"
+      ".inst 0x808b0383  // fmopa za3.s, p0/M, p0/M, z28.s, z11.s\n"
+      ".inst 0x808c02c0  // fmopa za0.s, p0/M, p0/M, z22.s, z12.s\n"
+      ".inst 0x808d02c1  // fmopa za1.s, p0/M, p0/M, z22.s, z13.s\n"
+      ".inst 0x808e02c2  // fmopa za2.s, p0/M, p0/M, z22.s, z14.s\n"
+      ".inst 0x808f02c3  // fmopa za3.s, p0/M, p0/M, z22.s, z15.s\n"
+      ".inst 0x808403c0  // fmopa za0.s, p0/M, p0/M, z30.s, z4.s\n"
+      ".inst 0x808503c1  // fmopa za1.s, p0/M, p0/M, z30.s, z5.s\n"
+      ".inst 0x808603c2  // fmopa za2.s, p0/M, p0/M, z30.s, z6.s\n"
+      ".inst 0x808703c3  // fmopa za3.s, p0/M, p0/M, z30.s, z7.s\n"
       ".inst 0x80930280  // fmopa za0.s, p0/M, p0/M, z20.s, z19.s\n"
       ".inst 0x80970281  // fmopa za1.s, p0/M, p0/M, z20.s, z23.s\n"
       ".inst 0x809b0282  // fmopa za2.s, p0/M, p0/M, z20.s, z27.s\n"
       ".inst 0x809f0283  // fmopa za3.s, p0/M, p0/M, z20.s, z31.s\n"
-      ".inst 0x808c0080  // fmopa za0.s, p0/M, p0/M, z4.s, z12.s\n"
-      ".inst 0x808d0081  // fmopa za1.s, p0/M, p0/M, z4.s, z13.s\n"
-      ".inst 0x808e0082  // fmopa za2.s, p0/M, p0/M, z4.s, z14.s\n"
-      ".inst 0x808f0083  // fmopa za3.s, p0/M, p0/M, z4.s, z15.s\n"
-      ".inst 0x809203a0  // fmopa za0.s, p0/M, p0/M, z29.s, z18.s\n"
-      ".inst 0x809603a1  // fmopa za1.s, p0/M, p0/M, z29.s, z22.s\n"
-      ".inst 0x809a03a2  // fmopa za2.s, p0/M, p0/M, z29.s, z26.s\n"
-      ".inst 0x809e03a3  // fmopa za3.s, p0/M, p0/M, z29.s, z30.s\n"
-      ".inst 0x80880040  // fmopa za0.s, p0/M, p0/M, z2.s, z8.s\n"
-      ".inst 0x80890041  // fmopa za1.s, p0/M, p0/M, z2.s, z9.s\n"
-      ".inst 0x808a0042  // fmopa za2.s, p0/M, p0/M, z2.s, z10.s\n"
-      ".inst 0x808b0043  // fmopa za3.s, p0/M, p0/M, z2.s, z11.s\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x22, 10f\n"
       "9:"  // K oddments: Loop
-      "ld1w { z26.s }, p0/Z, [x26]\n"
-      "subs x20, x20, #0x1\n"
+      "ld1w { z8.s }, p0/Z, [x26]\n"
+      "subs x22, x22, #0x1\n"
       "addvl x26, x26, #1\n"
-      ".inst 0xa140c6e3  // ld1w { z3.s, z7.s, z11.s, z15.s }, pn9.b/Z, [x23]\n"
-      "addvl x23, x23, #4\n"
-      ".inst 0x80830340  // fmopa za0.s, p0/M, p0/M, z26.s, z3.s\n"
-      ".inst 0x80870341  // fmopa za1.s, p0/M, p0/M, z26.s, z7.s\n"
-      ".inst 0x808b0342  // fmopa za2.s, p0/M, p0/M, z26.s, z11.s\n"
-      ".inst 0x808f0343  // fmopa za3.s, p0/M, p0/M, z26.s, z15.s\n"
+      ".inst 0xa140c6a3  // ld1w { z3.s, z7.s, z11.s, z15.s }, pn9.b/Z, [x21]\n"
+      "addvl x21, x21, #4\n"
+      ".inst 0x80830100  // fmopa za0.s, p0/M, p0/M, z8.s, z3.s\n"
+      ".inst 0x80870101  // fmopa za1.s, p0/M, p0/M, z8.s, z7.s\n"
+      ".inst 0x808b0102  // fmopa za2.s, p0/M, p0/M, z8.s, z11.s\n"
+      ".inst 0x808f0103  // fmopa za3.s, p0/M, p0/M, z8.s, z15.s\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x15, #1, 14f\n"
@@ -239,25 +239,25 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5d8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x14]\n"
+      ".inst 0xa040c5d4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x14]\n"
       ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
-      ".inst 0xa042c5d4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xa060c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13]\n"
-      ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
-      ".inst 0xa061c5b0  // st1w { z16.s-z19.s }, pn9.b, [x13, #0x4, MUL VL]\n"
-      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
+      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
+      ".inst 0xa042c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5a0  // st1w { z0.s-z3.s }, pn9.b, [x13, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c5bc  // st1w { z28.s-z31.s }, pn9.b, [x13, #0xc, MUL VL]\n"
+      ".inst 0xa060c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13]\n"
+      "addvl x14, x14, #16\n"
+      ".inst 0xa061c5a4  // st1w { z4.s-z7.s }, pn9.b, [x13, #0x4, MUL VL]\n"
+      ".inst 0xa062c5ac  // st1w { z12.s-z15.s }, pn9.b, [x13, #0x8, MUL VL]\n"
+      ".inst 0xa063c5a0  // st1w { z0.s-z3.s }, pn9.b, [x13, #0xc, MUL VL]\n"
       "addvl x13, x13, #16\n"
       "blt 11b\n"
       "b 24f\n"
@@ -265,31 +265,31 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
+      ".inst 0xa060c5b0  // st1w { z16.s-z19.s }, pn9.b, [x13]\n"
       ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5ac  // st1w { z12.s-z15.s }, pn9.b, [x13]\n"
+      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
+      ".inst 0xa061c5b8  // st1w { z24.s-z27.s }, pn9.b, [x13, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5a4  // st1w { z4.s-z7.s }, pn9.b, [x13, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
       ".inst 0xa062c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13, #0x8, MUL VL]\n"
-      ".inst 0xa063c5b0  // st1w { z16.s-z19.s }, pn9.b, [x13, #0xc, MUL VL]\n"
+      ".inst 0xa063c5a4  // st1w { z4.s-z7.s }, pn9.b, [x13, #0xc, MUL VL]\n"
       "addvl x13, x13, #16\n"
       "blt 13b\n"
       "b 24f\n"
       "14:"  // Store to output array
       "ldr x25, [%x[args], %[offsetof_C]]\n"
+      "add x25, x25, x9, LSL #2\n"  // C += n
       "sub x24, x11, x10\n"
       "ldr x23, [%x[args], %[offsetof_ldcb]]\n"
-      "add x25, x25, x9, LSL #2\n"  // C += n
       "madd x25, x10, x23, x25\n"  // C += m * ldc
       "tbz x15, #2, 18f\n"
       "cntw x20\n"
-      "mov x12, #0x0\n"
       "cmp x24, x20\n"
       "csel x22, x24, x20, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Skip activation: Accumulator row 0 loop
@@ -299,30 +299,30 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
       ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
       ".inst 0xa160c320  // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xa160c321  // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa160c322  // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa160c323  // st1w { z3.s, z7.s, z11.s, z15.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "blt 15b\n"
       "16:"  // Store to output array: Skip activation: Accumulator row 0 oddments
       "cbz x20, 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
-      ".inst 0xa160c320  // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x25]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xa160c330  // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "beq 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xa160c321  // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x25]\n"
+      ".inst 0xa160c331  // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "beq 17f\n"
-      ".inst 0xa160c322  // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x25]\n"
+      ".inst 0xa160c332  // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "17:"  // Store to output array: Skip activation: Accumulator row 0 oddments: End
       "subs x24, x24, x22\n"
@@ -330,29 +330,29 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
       "b 22f\n"
       "18:"  // Store to output array: Skip activation: End
       "cntw x20\n"
-      "ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
-      "mov x12, #0x0\n"
       "cmp x24, x20\n"
-      "ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
       "csel x20, x24, x20, LT\n"
       "lsr x21, x20, #0x2\n"
+      "ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 20f\n"
       "19:"  // Store to output array: Accumulator row 0 loop
       ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
       ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
-      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
       ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
       ".inst 0xc1a0c834  // fclamp { z20.s-z23.s }, z1.s, z0.s\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
       ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
       ".inst 0xc1a0c83c  // fclamp { z28.s-z31.s }, z1.s, z0.s\n"
-      "add x12, x12, #0x4\n"
-      "cmp x12, x21, LSL #2\n"
       ".inst 0xa160c330  // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa160c331  // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa160c332  // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       ".inst 0xa160c333  // st1w { z19.s, z23.s, z27.s, z31.s }, p8, [x25]\n"
@@ -362,13 +362,13 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
       "cbz x20, 21f\n"
       ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
       ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
-      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
       ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
       ".inst 0xc1a0c834  // fclamp { z20.s-z23.s }, z1.s, z0.s\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
       ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
       ".inst 0xc1a0c83c  // fclamp { z28.s-z31.s }, z1.s, z0.s\n"
+      "subs x20, x20, #0x1\n"
       ".inst 0xa160c330  // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x25]\n"
       "add x25, x25, x23\n"
       "beq 21f\n"
@@ -383,25 +383,25 @@ void sme2_interleaved_nomerge_fp32_mopa_1VLx4VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "23:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5c0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x14]\n"
-      ".inst 0xa041c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c5dc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa040c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa041c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
       ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x14, x14, #16\n"
       "blt 23b\n"
       "24:"  // End block
       "incw x9, ALL, MUL #4\n"
       "cmp x9, x28\n"
       "blt 3b\n"
       "incw x10\n"
-      "mov x9, #0x0\n"
       "cmp x10, x11\n"
+      "mov x9, #0x0\n"
       "mov x27, x26\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL.hpp
index 275399748a..ed54e70e28 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
 class cls_sme2_interleaved_nomerge_fp32_mopa_2VLx2VL
 {
 public:
-  typedef float lhs_operand_type;
-  typedef float rhs_operand_type;
+  typedef float operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const float *const A, const float *const B, float *const C, int ldc, const int M, const int N, const int K, const float *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_fp32_mopa_2VLx2VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 2, 2, 1> transforms = {};
+  StdTransformsSME<operand_type, result_type, 2, 2, 1> transforms = {};
 
   cls_sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp
index 45fcc7a860..1e46aee27a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
         B(B), kstride_bytes(K * sizeof(float)),
         C(C), ldcb(ldc * sizeof(float)),
         M(M), N(N), K(K),
+        n_loops((K - 1) / 2), n_tail_iters((K - 1) % 2),
         min(-std::numeric_limits<float>::infinity()),
         max(std::numeric_limits<float>::infinity()),
         bias(bias),
@@ -87,13 +88,12 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
     const long kstride_bytes;
     float *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
     float min = -std::numeric_limits<float>::infinity();
     float max = std::numeric_limits<float>::infinity();
 
     const float *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -112,17 +112,17 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa040c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa041c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa042c5f4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xa043c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -137,101 +137,101 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "fmov z17.s, #1.0\n"
-      ".inst 0xa00a428a  // ld1w { z10.s-z11.s }, p8/Z, [x20, x10, LSL #2]\n"
-      ".inst 0x808a0220  // fmopa za0.s, p0/M, p0/M, z17.s, z10.s\n"
-      ".inst 0x808b0221  // fmopa za1.s, p0/M, p0/M, z17.s, z11.s\n"
-      ".inst 0x808a0222  // fmopa za2.s, p0/M, p0/M, z17.s, z10.s\n"
-      ".inst 0x808b0223  // fmopa za3.s, p0/M, p0/M, z17.s, z11.s\n"
+      "fmov z12.s, #1.0\n"
+      ".inst 0xa10a4289  // ldnt1w { z1.s, z9.s }, p8/Z, [x20, x10, LSL #2]\n"
+      ".inst 0x80810180  // fmopa za0.s, p0/M, p0/M, z12.s, z1.s\n"
+      ".inst 0x80890181  // fmopa za1.s, p0/M, p0/M, z12.s, z9.s\n"
+      ".inst 0x80810182  // fmopa za2.s, p0/M, p0/M, z12.s, z1.s\n"
+      ".inst 0x80890183  // fmopa za3.s, p0/M, p0/M, z12.s, z9.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20, ALL, MUL #2\n"
       "incw x21, ALL, MUL #2\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
-      "lsr x21, x20, #0x2\n"
-      "and x20, x20, #0x3\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa0404776  // ld1w { z22.s-z23.s }, pn9.b/Z, [x27]\n"
-      ".inst 0xa14046e7  // ld1w { z7.s, z15.s }, pn9.b/Z, [x23]\n"
-      ".inst 0xa1414766  // ld1w { z6.s, z14.s }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0xa04146f4  // ld1w { z20.s-z21.s }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa1424762  // ld1w { z2.s, z10.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14246e3  // ld1w { z3.s, z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa1434761  // ld1w { z1.s, z9.s }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      "lsr x23, x20, #0x2\n"
+      "and x22, x20, #0x3\n"
+      "ldr x21, [%x[args], %[offsetof_B]]\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x21, x10, x20, x21\n"  // bptr = B + n * kstride_bytes
+      "cbz x23, 8f\n"
+      "subs x23, x23, #0x1\n"
+      ".inst 0xa0404772  // ld1w { z18.s-z19.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa04046a3  // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x21]\n"
+      ".inst 0xa0414764  // ld1w { z4.s-z5.s }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa04146bb  // ldnt1w { z26.s-z27.s }, pn9.b/Z, [x21, #0x2, MUL VL]\n"
+      ".inst 0xa042476a  // ld1w { z10.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa04246b5  // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n"
+      ".inst 0xa0434766  // ld1w { z6.s-z7.s }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa04346e4  // ld1w { z4.s-z5.s }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
-      "addvl x23, x23, #8\n"
+      ".inst 0xa04346a9  // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x21, #0x6, MUL VL]\n"
+      "addvl x21, x21, #8\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0x808702c0  // fmopa za0.s, p0/M, p0/M, z22.s, z7.s\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0x808f02c1  // fmopa za1.s, p0/M, p0/M, z22.s, z15.s\n"
-      ".inst 0x808702e2  // fmopa za2.s, p0/M, p0/M, z23.s, z7.s\n"
-      ".inst 0x808f02e3  // fmopa za3.s, p0/M, p0/M, z23.s, z15.s\n"
-      ".inst 0xa0404776  // ld1w { z22.s-z23.s }, pn9.b/Z, [x27]\n"
-      ".inst 0x809400c0  // fmopa za0.s, p0/M, p0/M, z6.s, z20.s\n"
-      ".inst 0xa14046e7  // ld1w { z7.s, z15.s }, pn9.b/Z, [x23]\n"
-      ".inst 0x809500c1  // fmopa za1.s, p0/M, p0/M, z6.s, z21.s\n"
-      ".inst 0x809401c2  // fmopa za2.s, p0/M, p0/M, z14.s, z20.s\n"
-      ".inst 0x809501c3  // fmopa za3.s, p0/M, p0/M, z14.s, z21.s\n"
-      ".inst 0xa1414766  // ld1w { z6.s, z14.s }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0x80830040  // fmopa za0.s, p0/M, p0/M, z2.s, z3.s\n"
-      ".inst 0xa04146f4  // ld1w { z20.s-z21.s }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0x808b0041  // fmopa za1.s, p0/M, p0/M, z2.s, z11.s\n"
-      ".inst 0x80830142  // fmopa za2.s, p0/M, p0/M, z10.s, z3.s\n"
-      ".inst 0x808b0143  // fmopa za3.s, p0/M, p0/M, z10.s, z11.s\n"
-      ".inst 0xa1424762  // ld1w { z2.s, z10.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14246e3  // ld1w { z3.s, z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0x80840020  // fmopa za0.s, p0/M, p0/M, z1.s, z4.s\n"
-      ".inst 0x80850021  // fmopa za1.s, p0/M, p0/M, z1.s, z5.s\n"
-      ".inst 0x80840122  // fmopa za2.s, p0/M, p0/M, z9.s, z4.s\n"
-      ".inst 0x80850123  // fmopa za3.s, p0/M, p0/M, z9.s, z5.s\n"
-      ".inst 0xa1434761  // ld1w { z1.s, z9.s }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      ".inst 0x80820240  // fmopa za0.s, p0/M, p0/M, z18.s, z2.s\n"
+      "subs x23, x23, #0x1\n"
+      ".inst 0x80830241  // fmopa za1.s, p0/M, p0/M, z18.s, z3.s\n"
+      ".inst 0x80820262  // fmopa za2.s, p0/M, p0/M, z19.s, z2.s\n"
+      ".inst 0x80830263  // fmopa za3.s, p0/M, p0/M, z19.s, z3.s\n"
+      ".inst 0xa0404772  // ld1w { z18.s-z19.s }, pn9.b/Z, [x27]\n"
+      ".inst 0x809a0080  // fmopa za0.s, p0/M, p0/M, z4.s, z26.s\n"
+      ".inst 0xa04046a3  // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x21]\n"
+      ".inst 0x809b0081  // fmopa za1.s, p0/M, p0/M, z4.s, z27.s\n"
+      ".inst 0x809a00a2  // fmopa za2.s, p0/M, p0/M, z5.s, z26.s\n"
+      ".inst 0x809b00a3  // fmopa za3.s, p0/M, p0/M, z5.s, z27.s\n"
+      ".inst 0xa0414764  // ld1w { z4.s-z5.s }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0x80940140  // fmopa za0.s, p0/M, p0/M, z10.s, z20.s\n"
+      ".inst 0xa04146bb  // ldnt1w { z26.s-z27.s }, pn9.b/Z, [x21, #0x2, MUL VL]\n"
+      ".inst 0x80950141  // fmopa za1.s, p0/M, p0/M, z10.s, z21.s\n"
+      ".inst 0x80940162  // fmopa za2.s, p0/M, p0/M, z11.s, z20.s\n"
+      ".inst 0x80950163  // fmopa za3.s, p0/M, p0/M, z11.s, z21.s\n"
+      ".inst 0xa042476a  // ld1w { z10.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa04246b5  // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x21, #0x4, MUL VL]\n"
+      ".inst 0x808800c0  // fmopa za0.s, p0/M, p0/M, z6.s, z8.s\n"
+      ".inst 0x808900c1  // fmopa za1.s, p0/M, p0/M, z6.s, z9.s\n"
+      ".inst 0x808800e2  // fmopa za2.s, p0/M, p0/M, z7.s, z8.s\n"
+      ".inst 0x808900e3  // fmopa za3.s, p0/M, p0/M, z7.s, z9.s\n"
+      ".inst 0xa0434766  // ld1w { z6.s-z7.s }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa04346e4  // ld1w { z4.s-z5.s }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
-      "addvl x23, x23, #8\n"
+      ".inst 0xa04346a9  // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x21, #0x6, MUL VL]\n"
+      "addvl x21, x21, #8\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0x808702c0  // fmopa za0.s, p0/M, p0/M, z22.s, z7.s\n"
-      ".inst 0x808f02c1  // fmopa za1.s, p0/M, p0/M, z22.s, z15.s\n"
-      ".inst 0x808702e2  // fmopa za2.s, p0/M, p0/M, z23.s, z7.s\n"
-      ".inst 0x808f02e3  // fmopa za3.s, p0/M, p0/M, z23.s, z15.s\n"
-      ".inst 0x809400c0  // fmopa za0.s, p0/M, p0/M, z6.s, z20.s\n"
-      ".inst 0x809500c1  // fmopa za1.s, p0/M, p0/M, z6.s, z21.s\n"
-      ".inst 0x809401c2  // fmopa za2.s, p0/M, p0/M, z14.s, z20.s\n"
-      ".inst 0x809501c3  // fmopa za3.s, p0/M, p0/M, z14.s, z21.s\n"
-      ".inst 0x80830040  // fmopa za0.s, p0/M, p0/M, z2.s, z3.s\n"
-      ".inst 0x808b0041  // fmopa za1.s, p0/M, p0/M, z2.s, z11.s\n"
-      ".inst 0x80830142  // fmopa za2.s, p0/M, p0/M, z10.s, z3.s\n"
-      ".inst 0x808b0143  // fmopa za3.s, p0/M, p0/M, z10.s, z11.s\n"
-      ".inst 0x80840020  // fmopa za0.s, p0/M, p0/M, z1.s, z4.s\n"
-      ".inst 0x80850021  // fmopa za1.s, p0/M, p0/M, z1.s, z5.s\n"
-      ".inst 0x80840122  // fmopa za2.s, p0/M, p0/M, z9.s, z4.s\n"
-      ".inst 0x80850123  // fmopa za3.s, p0/M, p0/M, z9.s, z5.s\n"
+      ".inst 0x80820240  // fmopa za0.s, p0/M, p0/M, z18.s, z2.s\n"
+      ".inst 0x80830241  // fmopa za1.s, p0/M, p0/M, z18.s, z3.s\n"
+      ".inst 0x80820262  // fmopa za2.s, p0/M, p0/M, z19.s, z2.s\n"
+      ".inst 0x80830263  // fmopa za3.s, p0/M, p0/M, z19.s, z3.s\n"
+      ".inst 0x809a0080  // fmopa za0.s, p0/M, p0/M, z4.s, z26.s\n"
+      ".inst 0x809b0081  // fmopa za1.s, p0/M, p0/M, z4.s, z27.s\n"
+      ".inst 0x809a00a2  // fmopa za2.s, p0/M, p0/M, z5.s, z26.s\n"
+      ".inst 0x809b00a3  // fmopa za3.s, p0/M, p0/M, z5.s, z27.s\n"
+      ".inst 0x80940140  // fmopa za0.s, p0/M, p0/M, z10.s, z20.s\n"
+      ".inst 0x80950141  // fmopa za1.s, p0/M, p0/M, z10.s, z21.s\n"
+      ".inst 0x80940162  // fmopa za2.s, p0/M, p0/M, z11.s, z20.s\n"
+      ".inst 0x80950163  // fmopa za3.s, p0/M, p0/M, z11.s, z21.s\n"
+      ".inst 0x808800c0  // fmopa za0.s, p0/M, p0/M, z6.s, z8.s\n"
+      ".inst 0x808900c1  // fmopa za1.s, p0/M, p0/M, z6.s, z9.s\n"
+      ".inst 0x808800e2  // fmopa za2.s, p0/M, p0/M, z7.s, z8.s\n"
+      ".inst 0x808900e3  // fmopa za3.s, p0/M, p0/M, z7.s, z9.s\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x22, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa040476a  // ld1w { z10.s-z11.s }, pn9.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xa040477e  // ld1w { z30.s-z31.s }, pn9.b/Z, [x27]\n"
+      "subs x22, x22, #0x1\n"
       "addvl x27, x27, #2\n"
-      ".inst 0xa04046ee  // ld1w { z14.s-z15.s }, pn9.b/Z, [x23]\n"
-      "addvl x23, x23, #2\n"
-      ".inst 0x808e0140  // fmopa za0.s, p0/M, p0/M, z10.s, z14.s\n"
-      ".inst 0x808f0141  // fmopa za1.s, p0/M, p0/M, z10.s, z15.s\n"
-      ".inst 0x808e0162  // fmopa za2.s, p0/M, p0/M, z11.s, z14.s\n"
-      ".inst 0x808f0163  // fmopa za3.s, p0/M, p0/M, z11.s, z15.s\n"
+      ".inst 0xa14046a5  // ld1w { z5.s, z13.s }, pn9.b/Z, [x21]\n"
+      "addvl x21, x21, #2\n"
+      ".inst 0x808503c0  // fmopa za0.s, p0/M, p0/M, z30.s, z5.s\n"
+      ".inst 0x808d03c1  // fmopa za1.s, p0/M, p0/M, z30.s, z13.s\n"
+      ".inst 0x808503e2  // fmopa za2.s, p0/M, p0/M, z31.s, z5.s\n"
+      ".inst 0x808d03e3  // fmopa za3.s, p0/M, p0/M, z31.s, z13.s\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x16, #1, 14f\n"
@@ -239,24 +239,24 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
-      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
+      ".inst 0xa040c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
       ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
       ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xa060c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14]\n"
       ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
-      ".inst 0xa061c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840703  // mova za3h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa043c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
+      ".inst 0xa060c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n"
       ".inst 0xa063c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
@@ -265,31 +265,31 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
-      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
+      ".inst 0xa060c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14]\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
+      ".inst 0xa061c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c5d8  // st1w { z24.s-z27.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 30f\n"
       "14:"  // Store to output array
       "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "add x26, x26, x10, LSL #2\n"  // C += n
       "sub x25, x13, x11\n"
       "ldr x24, [%x[args], %[offsetof_ldcb]]\n"
-      "add x26, x26, x10, LSL #2\n"  // C += n
       "madd x26, x11, x24, x26\n"  // C += m * ldc
       "tbz x16, #2, 21f\n"
       "cntw x23\n"
-      "mov x12, #0x0\n"
       "cmp x25, x23\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Skip activation: Accumulator row 0 loop
@@ -297,36 +297,36 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
       ".inst 0xa1604344  // st1w { z4.s, z12.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xa1604345  // st1w { z5.s, z13.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa1604346  // st1w { z6.s, z14.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa1604347  // st1w { z7.s, z15.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "blt 15b\n"
       "16:"  // Store to output array: Skip activation: Accumulator row 0 oddments
       "cbz x20, 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
-      ".inst 0xa1604340  // st1w { z0.s, z8.s }, p8, [x26]\n"
+      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xa1604354  // st1w { z20.s, z28.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xa1604341  // st1w { z1.s, z9.s }, p8, [x26]\n"
+      ".inst 0xa1604355  // st1w { z21.s, z29.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 17f\n"
-      ".inst 0xa1604342  // st1w { z2.s, z10.s }, p8, [x26]\n"
+      ".inst 0xa1604356  // st1w { z22.s, z30.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "17:"  // Store to output array: Skip activation: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 21f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 19f\n"
       "18:"  // Store to output array: Skip activation: Accumulator row 1 loop
@@ -334,28 +334,28 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
       ".inst 0xa1604344  // st1w { z4.s, z12.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xa1604345  // st1w { z5.s, z13.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa1604346  // st1w { z6.s, z14.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa1604347  // st1w { z7.s, z15.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "blt 18b\n"
       "19:"  // Store to output array: Skip activation: Accumulator row 1 oddments
       "cbz x20, 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
-      ".inst 0xa1604344  // st1w { z4.s, z12.s }, p8, [x26]\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xa1604345  // st1w { z5.s, z13.s }, p8, [x26]\n"
+      ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 20f\n"
-      ".inst 0xa1604346  // st1w { z6.s, z14.s }, p8, [x26]\n"
+      ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "20:"  // Store to output array: Skip activation: Accumulator row 1 oddments: End
       "subs x25, x25, x22\n"
@@ -363,37 +363,37 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       "b 28f\n"
       "21:"  // Store to output array: Skip activation: End
       "cntw x23\n"
-      "ld1rw { z21.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
-      "mov x12, #0x0\n"
       "cmp x25, x23\n"
-      "ld1rw { z20.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 23f\n"
       "22:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
-      "add x12, x12, #0x4\n"
-      "cmp x12, x21, LSL #2\n"
-      ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
+      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xc1a0c834  // fclamp { z20.s-z23.s }, z1.s, z0.s\n"
+      ".inst 0xc1a0c83c  // fclamp { z28.s-z31.s }, z1.s, z0.s\n"
+      ".inst 0xa1604354  // st1w { z20.s, z28.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
+      "add x12, x12, #0x4\n"
+      ".inst 0xa1604355  // st1w { z21.s, z29.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      ".inst 0xa1604356  // st1w { z22.s, z30.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604353  // st1w { z19.s, z27.s }, p8, [x26]\n"
+      ".inst 0xa1604357  // st1w { z23.s, z31.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "blt 22b\n"
       "23:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 24f\n"
       ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
       ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
+      ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
+      ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
       ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 24f\n"
@@ -407,34 +407,34 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       "subs x25, x25, x22\n"
       "beq 28f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x23, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 26f\n"
       "25:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
-      ".inst 0xc0860468  // mova { z8.s-z11.s }, za3h.s[x12]\n"
-      ".inst 0xc1b4caa0  // fclamp { z0.s-z3.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4caa8  // fclamp { z8.s-z11.s }, z21.s, z20.s\n"
-      "add x12, x12, #0x4\n"
-      "cmp x12, x21, LSL #2\n"
-      ".inst 0xa1604340  // st1w { z0.s, z8.s }, p8, [x26]\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
+      ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
+      ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604341  // st1w { z1.s, z9.s }, p8, [x26]\n"
+      "add x12, x12, #0x4\n"
+      ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604342  // st1w { z2.s, z10.s }, p8, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
-      ".inst 0xa1604343  // st1w { z3.s, z11.s }, p8, [x26]\n"
+      ".inst 0xa1604353  // st1w { z19.s, z27.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "blt 25b\n"
       "26:"  // Store to output array: Accumulator row 1 oddments
       "cbz x20, 27f\n"
       ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
       ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xc1a0c830  // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
+      ".inst 0xc1a0c838  // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cab8  // fclamp { z24.s-z27.s }, z21.s, z20.s\n"
       ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x24\n"
       "beq 27f\n"
@@ -449,25 +449,25 @@ void sme2_interleaved_nomerge_fp32_mopa_2VLx2VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "29:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa041c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
       ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 29b\n"
       "30:"  // End block
       "incw x10, ALL, MUL #2\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #2\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL.hpp
index 453505a227..1348f00d37 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
 class cls_sme2_interleaved_nomerge_fp32_mopa_4VLx1VL
 {
 public:
-  typedef float lhs_operand_type;
-  typedef float rhs_operand_type;
+  typedef float operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const float *const A, const float *const B, float *const C, int ldc, const int M, const int N, const int K, const float *const bias, const Activation act, bool accumulate, float *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_fp32_mopa_4VLx1VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 4, 1, 1> transforms = {};
+  StdTransformsSME<operand_type, result_type, 4, 1, 1> transforms = {};
 
   cls_sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp
index 63f93b9b5b..a69e1f84e0 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
         B(B), kstride_bytes(K * sizeof(float)),
         C(C), ldcb(ldc * sizeof(float)),
         M(M), N(N), K(K),
+        n_loops((K - 1) / 2), n_tail_iters((K - 1) % 2),
         min(-std::numeric_limits<float>::infinity()),
         max(std::numeric_limits<float>::infinity()),
         bias(bias),
@@ -87,13 +88,12 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
     const long kstride_bytes;
     float *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
     float min = -std::numeric_limits<float>::infinity();
     float max = std::numeric_limits<float>::infinity();
 
     const float *const bias;
 
-
     float *const accumulator_buffer;
     uint64_t flags;
   };
@@ -112,17 +112,17 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840783  // mova za3h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa040c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa041c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa042c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xa043c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -137,101 +137,101 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "fmov z6.s, #1.0\n"
-      "ld1w { z26.s }, p0/Z, [x20, x10, LSL #2]\n"
-      ".inst 0x809a24c0  // fmopa za0.s, p1/M, p1/M, z6.s, z26.s\n"
-      ".inst 0x809a24c1  // fmopa za1.s, p1/M, p1/M, z6.s, z26.s\n"
-      ".inst 0x809a24c2  // fmopa za2.s, p1/M, p1/M, z6.s, z26.s\n"
-      ".inst 0x809a24c3  // fmopa za3.s, p1/M, p1/M, z6.s, z26.s\n"
+      "fmov z11.s, #1.0\n"
+      "ldnt1w { z13.s }, p0/Z, [x20, x10, LSL #2]\n"
+      ".inst 0x808d2560  // fmopa za0.s, p1/M, p1/M, z11.s, z13.s\n"
+      ".inst 0x808d2561  // fmopa za1.s, p1/M, p1/M, z11.s, z13.s\n"
+      ".inst 0x808d2562  // fmopa za2.s, p1/M, p1/M, z11.s, z13.s\n"
+      ".inst 0x808d2563  // fmopa za3.s, p1/M, p1/M, z11.s, z13.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20\n"
       "incw x21, ALL, MUL #4\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
-      "lsr x21, x20, #0x2\n"
-      "and x20, x20, #0x3\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa140c363  // ld1w { z3.s, z7.s, z11.s, z15.s }, pn8.b/Z, [x27]\n"
-      "ld1w { z13.s }, p1/Z, [x23]\n"
-      ".inst 0xa141c372  // ld1w { z18.s, z22.s, z26.s, z30.s }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "ld1w { z21.s }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa142c373  // ld1w { z19.s, z23.s, z27.s, z31.s }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1w { z17.s }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa143c370  // ld1w { z16.s, z20.s, z24.s, z28.s }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      "lsr x23, x20, #0x2\n"
+      "and x22, x20, #0x3\n"
+      "ldr x21, [%x[args], %[offsetof_B]]\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x21, x10, x20, x21\n"  // bptr = B + n * kstride_bytes
+      "cbz x23, 8f\n"
+      "subs x23, x23, #0x1\n"
+      ".inst 0xa140c360  // ld1w { z0.s, z4.s, z8.s, z12.s }, pn8.b/Z, [x27]\n"
+      "ldnt1w { z19.s }, p1/Z, [x21]\n"
+      ".inst 0xa141c371  // ld1w { z17.s, z21.s, z25.s, z29.s }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      "ldnt1w { z22.s }, p1/Z, [x21, #1, MUL VL]\n"
+      ".inst 0xa142c370  // ld1w { z16.s, z20.s, z24.s, z28.s }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1w { z23.s }, p1/Z, [x21, #2, MUL VL]\n"
+      ".inst 0xa143c363  // ld1w { z3.s, z7.s, z11.s, z15.s }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1w { z2.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "addvl x23, x23, #4\n"
+      "ldnt1w { z2.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "addvl x21, x21, #4\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0x808d2460  // fmopa za0.s, p1/M, p1/M, z3.s, z13.s\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0x808d24e1  // fmopa za1.s, p1/M, p1/M, z7.s, z13.s\n"
-      ".inst 0x808d2562  // fmopa za2.s, p1/M, p1/M, z11.s, z13.s\n"
-      ".inst 0x808d25e3  // fmopa za3.s, p1/M, p1/M, z15.s, z13.s\n"
-      ".inst 0xa140c363  // ld1w { z3.s, z7.s, z11.s, z15.s }, pn8.b/Z, [x27]\n"
-      ".inst 0x80952640  // fmopa za0.s, p1/M, p1/M, z18.s, z21.s\n"
-      "ld1w { z13.s }, p1/Z, [x23]\n"
-      ".inst 0x809526c1  // fmopa za1.s, p1/M, p1/M, z22.s, z21.s\n"
-      ".inst 0x80952742  // fmopa za2.s, p1/M, p1/M, z26.s, z21.s\n"
-      ".inst 0x809527c3  // fmopa za3.s, p1/M, p1/M, z30.s, z21.s\n"
-      ".inst 0xa141c372  // ld1w { z18.s, z22.s, z26.s, z30.s }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0x80912660  // fmopa za0.s, p1/M, p1/M, z19.s, z17.s\n"
-      "ld1w { z21.s }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0x809126e1  // fmopa za1.s, p1/M, p1/M, z23.s, z17.s\n"
-      ".inst 0x80912762  // fmopa za2.s, p1/M, p1/M, z27.s, z17.s\n"
-      ".inst 0x809127e3  // fmopa za3.s, p1/M, p1/M, z31.s, z17.s\n"
-      ".inst 0xa142c373  // ld1w { z19.s, z23.s, z27.s, z31.s }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1w { z17.s }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0x80822600  // fmopa za0.s, p1/M, p1/M, z16.s, z2.s\n"
-      ".inst 0x80822681  // fmopa za1.s, p1/M, p1/M, z20.s, z2.s\n"
-      ".inst 0x80822702  // fmopa za2.s, p1/M, p1/M, z24.s, z2.s\n"
-      ".inst 0x80822783  // fmopa za3.s, p1/M, p1/M, z28.s, z2.s\n"
-      ".inst 0xa143c370  // ld1w { z16.s, z20.s, z24.s, z28.s }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      ".inst 0x80932400  // fmopa za0.s, p1/M, p1/M, z0.s, z19.s\n"
+      "subs x23, x23, #0x1\n"
+      ".inst 0x80932481  // fmopa za1.s, p1/M, p1/M, z4.s, z19.s\n"
+      ".inst 0x80932502  // fmopa za2.s, p1/M, p1/M, z8.s, z19.s\n"
+      ".inst 0x80932583  // fmopa za3.s, p1/M, p1/M, z12.s, z19.s\n"
+      ".inst 0xa140c360  // ld1w { z0.s, z4.s, z8.s, z12.s }, pn8.b/Z, [x27]\n"
+      ".inst 0x80962620  // fmopa za0.s, p1/M, p1/M, z17.s, z22.s\n"
+      "ldnt1w { z19.s }, p1/Z, [x21]\n"
+      ".inst 0x809626a1  // fmopa za1.s, p1/M, p1/M, z21.s, z22.s\n"
+      ".inst 0x80962722  // fmopa za2.s, p1/M, p1/M, z25.s, z22.s\n"
+      ".inst 0x809627a3  // fmopa za3.s, p1/M, p1/M, z29.s, z22.s\n"
+      ".inst 0xa141c371  // ld1w { z17.s, z21.s, z25.s, z29.s }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0x80972600  // fmopa za0.s, p1/M, p1/M, z16.s, z23.s\n"
+      "ldnt1w { z22.s }, p1/Z, [x21, #1, MUL VL]\n"
+      ".inst 0x80972681  // fmopa za1.s, p1/M, p1/M, z20.s, z23.s\n"
+      ".inst 0x80972702  // fmopa za2.s, p1/M, p1/M, z24.s, z23.s\n"
+      ".inst 0x80972783  // fmopa za3.s, p1/M, p1/M, z28.s, z23.s\n"
+      ".inst 0xa142c370  // ld1w { z16.s, z20.s, z24.s, z28.s }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1w { z23.s }, p1/Z, [x21, #2, MUL VL]\n"
+      ".inst 0x80822460  // fmopa za0.s, p1/M, p1/M, z3.s, z2.s\n"
+      ".inst 0x808224e1  // fmopa za1.s, p1/M, p1/M, z7.s, z2.s\n"
+      ".inst 0x80822562  // fmopa za2.s, p1/M, p1/M, z11.s, z2.s\n"
+      ".inst 0x808225e3  // fmopa za3.s, p1/M, p1/M, z15.s, z2.s\n"
+      ".inst 0xa143c363  // ld1w { z3.s, z7.s, z11.s, z15.s }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1w { z2.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "addvl x23, x23, #4\n"
+      "ldnt1w { z2.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "addvl x21, x21, #4\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0x808d2460  // fmopa za0.s, p1/M, p1/M, z3.s, z13.s\n"
-      ".inst 0x808d24e1  // fmopa za1.s, p1/M, p1/M, z7.s, z13.s\n"
-      ".inst 0x808d2562  // fmopa za2.s, p1/M, p1/M, z11.s, z13.s\n"
-      ".inst 0x808d25e3  // fmopa za3.s, p1/M, p1/M, z15.s, z13.s\n"
-      ".inst 0x80952640  // fmopa za0.s, p1/M, p1/M, z18.s, z21.s\n"
-      ".inst 0x809526c1  // fmopa za1.s, p1/M, p1/M, z22.s, z21.s\n"
-      ".inst 0x80952742  // fmopa za2.s, p1/M, p1/M, z26.s, z21.s\n"
-      ".inst 0x809527c3  // fmopa za3.s, p1/M, p1/M, z30.s, z21.s\n"
-      ".inst 0x80912660  // fmopa za0.s, p1/M, p1/M, z19.s, z17.s\n"
-      ".inst 0x809126e1  // fmopa za1.s, p1/M, p1/M, z23.s, z17.s\n"
-      ".inst 0x80912762  // fmopa za2.s, p1/M, p1/M, z27.s, z17.s\n"
-      ".inst 0x809127e3  // fmopa za3.s, p1/M, p1/M, z31.s, z17.s\n"
-      ".inst 0x80822600  // fmopa za0.s, p1/M, p1/M, z16.s, z2.s\n"
-      ".inst 0x80822681  // fmopa za1.s, p1/M, p1/M, z20.s, z2.s\n"
-      ".inst 0x80822702  // fmopa za2.s, p1/M, p1/M, z24.s, z2.s\n"
-      ".inst 0x80822783  // fmopa za3.s, p1/M, p1/M, z28.s, z2.s\n"
+      ".inst 0x80932400  // fmopa za0.s, p1/M, p1/M, z0.s, z19.s\n"
+      ".inst 0x80932481  // fmopa za1.s, p1/M, p1/M, z4.s, z19.s\n"
+      ".inst 0x80932502  // fmopa za2.s, p1/M, p1/M, z8.s, z19.s\n"
+      ".inst 0x80932583  // fmopa za3.s, p1/M, p1/M, z12.s, z19.s\n"
+      ".inst 0x80962620  // fmopa za0.s, p1/M, p1/M, z17.s, z22.s\n"
+      ".inst 0x809626a1  // fmopa za1.s, p1/M, p1/M, z21.s, z22.s\n"
+      ".inst 0x80962722  // fmopa za2.s, p1/M, p1/M, z25.s, z22.s\n"
+      ".inst 0x809627a3  // fmopa za3.s, p1/M, p1/M, z29.s, z22.s\n"
+      ".inst 0x80972600  // fmopa za0.s, p1/M, p1/M, z16.s, z23.s\n"
+      ".inst 0x80972681  // fmopa za1.s, p1/M, p1/M, z20.s, z23.s\n"
+      ".inst 0x80972702  // fmopa za2.s, p1/M, p1/M, z24.s, z23.s\n"
+      ".inst 0x80972783  // fmopa za3.s, p1/M, p1/M, z28.s, z23.s\n"
+      ".inst 0x80822460  // fmopa za0.s, p1/M, p1/M, z3.s, z2.s\n"
+      ".inst 0x808224e1  // fmopa za1.s, p1/M, p1/M, z7.s, z2.s\n"
+      ".inst 0x80822562  // fmopa za2.s, p1/M, p1/M, z11.s, z2.s\n"
+      ".inst 0x808225e3  // fmopa za3.s, p1/M, p1/M, z15.s, z2.s\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x22, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa140c370  // ld1w { z16.s, z20.s, z24.s, z28.s }, pn8.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xa140c373  // ld1w { z19.s, z23.s, z27.s, z31.s }, pn8.b/Z, [x27]\n"
+      "subs x22, x22, #0x1\n"
       "addvl x27, x27, #4\n"
-      "ld1w { z2.s }, p1/Z, [x23]\n"
-      "addvl x23, x23, #1\n"
-      ".inst 0x80822600  // fmopa za0.s, p1/M, p1/M, z16.s, z2.s\n"
-      ".inst 0x80822681  // fmopa za1.s, p1/M, p1/M, z20.s, z2.s\n"
-      ".inst 0x80822702  // fmopa za2.s, p1/M, p1/M, z24.s, z2.s\n"
-      ".inst 0x80822783  // fmopa za3.s, p1/M, p1/M, z28.s, z2.s\n"
+      "ld1w { z11.s }, p1/Z, [x21]\n"
+      "addvl x21, x21, #1\n"
+      ".inst 0x808b2660  // fmopa za0.s, p1/M, p1/M, z19.s, z11.s\n"
+      ".inst 0x808b26e1  // fmopa za1.s, p1/M, p1/M, z23.s, z11.s\n"
+      ".inst 0x808b2762  // fmopa za2.s, p1/M, p1/M, z27.s, z11.s\n"
+      ".inst 0x808b27e3  // fmopa za3.s, p1/M, p1/M, z31.s, z11.s\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x16, #1, 14f\n"
@@ -239,25 +239,25 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
-      ".inst 0xa041c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
-      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
-      ".inst 0xa042c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
-      ".inst 0xa060c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14]\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xa061c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840403  // mova za3h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa040c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0860418  // mova { z24.s-z27.s }, za0h.s[x12]\n"
+      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa041c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xa042c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa043c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa060c1d8  // st1w { z24.s-z27.s }, pn8.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c1d0  // st1w { z16.s-z19.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 42f\n"
@@ -265,148 +265,148 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
-      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
-      ".inst 0xc0860468  // mova { z8.s-z11.s }, za3h.s[x12]\n"
-      ".inst 0xa060c1d0  // st1w { z16.s-z19.s }, pn8.b, [x14]\n"
+      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xa060c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14]\n"
+      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
+      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      ".inst 0xa061c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 42f\n"
       "14:"  // Store to output array
       "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "add x26, x26, x10, LSL #2\n"  // C += n
       "sub x25, x13, x11\n"
       "ldr x24, [%x[args], %[offsetof_ldcb]]\n"
-      "add x26, x26, x10, LSL #2\n"  // C += n
       "madd x26, x11, x24, x26\n"  // C += m * ldc
       "tbz x16, #2, 27f\n"
       "cntw x23\n"
-      "mov x12, #0x0\n"
       "cmp x25, x23\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Skip activation: Accumulator row 0 loop
-      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z12.s }, p0, [x26]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z13.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z14.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z15.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z7.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 15b\n"
       "16:"  // Store to output array: Skip activation: Accumulator row 0 oddments
       "cbz x20, 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      "st1w { z0.s }, p0, [x26]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 17f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z1.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 17f\n"
-      "st1w { z2.s }, p0, [x26]\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "17:"  // Store to output array: Skip activation: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 19f\n"
       "18:"  // Store to output array: Skip activation: Accumulator row 1 loop
-      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z12.s }, p0, [x26]\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      "st1w { z8.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z13.s }, p0, [x26]\n"
+      "st1w { z9.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z14.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z10.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z15.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z11.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 18b\n"
       "19:"  // Store to output array: Skip activation: Accumulator row 1 oddments
       "cbz x20, 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
+      "st1w { z24.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 20f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z25.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 20f\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "st1w { z26.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "20:"  // Store to output array: Skip activation: Accumulator row 1 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 22f\n"
       "21:"  // Store to output array: Skip activation: Accumulator row 2 loop
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z8.s }, p0, [x26]\n"
+      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z9.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z10.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z11.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z7.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 21b\n"
       "22:"  // Store to output array: Skip activation: Accumulator row 2 oddments
       "cbz x20, 23f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      "st1w { z12.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 23f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z13.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 23f\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "st1w { z14.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "23:"  // Store to output array: Skip activation: Accumulator row 2 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 25f\n"
       "24:"  // Store to output array: Skip activation: Accumulator row 3 loop
       ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "cmp x12, x21, LSL #2\n"
       "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "add x12, x12, #0x4\n"
       "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       "st1w { z19.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 24b\n"
@@ -429,63 +429,63 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
       "b 40f\n"
       "27:"  // Store to output array: Skip activation: End
       "cntw x23\n"
-      "ld1rw { z21.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
-      "mov x12, #0x0\n"
       "cmp x25, x23\n"
-      "ld1rw { z20.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "ld1rw { z21.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "ld1rw { z20.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 29f\n"
       "28:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
+      ".inst 0xc1b4cabc  // fclamp { z28.s-z31.s }, z21.s, z20.s\n"
+      "st1w { z28.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z29.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z30.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z19.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z31.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 28b\n"
       "29:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 30f\n"
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc1b4cabc  // fclamp { z28.s-z31.s }, z21.s, z20.s\n"
+      "st1w { z28.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 30f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z29.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 30f\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "st1w { z30.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "30:"  // Store to output array: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 40f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 32f\n"
       "31:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc1b4caa4  // fclamp { z4.s-z7.s }, z21.s, z20.s\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1w { z19.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z7.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 31b\n"
       "32:"  // Store to output array: Accumulator row 1 oddments
@@ -506,100 +506,100 @@ void sme2_interleaved_nomerge_fp32_mopa_4VLx1VL(const float *const A, const floa
       "subs x25, x25, x22\n"
       "beq 40f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 35f\n"
       "34:"  // Store to output array: Accumulator row 2 loop
       ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "cmp x12, x21, LSL #2\n"
       "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "add x12, x12, #0x4\n"
       "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       "st1w { z19.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 34b\n"
       "35:"  // Store to output array: Accumulator row 2 oddments
       "cbz x20, 36f\n"
-      ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4cabc  // fclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      "st1w { z28.s }, p0, [x26]\n"
+      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
+      "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 36f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z29.s }, p0, [x26]\n"
+      "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 36f\n"
-      "st1w { z30.s }, p0, [x26]\n"
+      "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "36:"  // Store to output array: Accumulator row 2 oddments: End
       "subs x25, x25, x22\n"
       "beq 40f\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x23, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 38f\n"
       "37:"  // Store to output array: Accumulator row 3 loop
       ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
-      "cmp x12, x21, LSL #2\n"
       "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "add x12, x12, #0x4\n"
       "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "cmp x12, x21, LSL #2\n"
       "st1w { z19.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 37b\n"
       "38:"  // Store to output array: Accumulator row 3 oddments
       "cbz x20, 39f\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1b4caac  // fclamp { z12.s-z15.s }, z21.s, z20.s\n"
-      "st1w { z12.s }, p0, [x26]\n"
+      ".inst 0xc1b4cab0  // fclamp { z16.s-z19.s }, z21.s, z20.s\n"
+      "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 39f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z13.s }, p0, [x26]\n"
+      "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 39f\n"
-      "st1w { z14.s }, p0, [x26]\n"
+      "st1w { z18.s }, p0, [x26]\n"
       "39:"  // Store to output array: Accumulator row 3 oddments: End
       "40:"  // Store to output array: End
       "tbz x16, #0, 42f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "41:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa041c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840781  // mova za1h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
       ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 41b\n"
       "42:"  // End block
       "incw x10\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #4\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL.hpp
index b70bef3bbe..a4b14325f2 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8q_mopa_1VLx4VL(const int8_t *const A, const int8
 class cls_sme2_interleaved_nomerge_s8q_mopa_1VLx4VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef int8_t result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, int8_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Requantize32 &rq, const int n_0, bool accumulate, int32_t *const accumulator_buffer);
@@ -61,7 +60,7 @@ public:
 
   static constexpr bool supports_accumulate()
   {
-    return true;
+    return false;
   }
 
   static constexpr bool supports_bias()
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8q_mopa_1VLx4VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 1, 4, 4, true> transforms = {};
+  StdTransformsSME<operand_type, result_type, 1, 4, 4, true> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8q_mopa_1VLx4VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/generic.cpp
index 56d1a13a72..c8d56dc5e5 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL/generic.cpp
@@ -49,7 +49,7 @@ void sme2_interleaved_nomerge_s8q_mopa_1VLx4VL(const int8_t *const A, const int8
         B(B), kstride_bytes(roundup(K, 4) * sizeof(int8_t)),
         C(C), ldcb(ldc * sizeof(int8_t)),
         M(M), N(N), K(K),
-        min(0), max(0),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias), n_0(n_0),
         accumulator_buffer(accumulator_buffer),
@@ -74,14 +74,13 @@ void sme2_interleaved_nomerge_s8q_mopa_1VLx4VL(const int8_t *const A, const int8
     const long kstride_bytes;
     int8_t *const C;
     const long ldcb;
-    const long M, N, K;
-    int32_t min;
-    int32_t max;
+    const long M, N, K, n_loops, n_tail_iters;
+    int32_t min = std::numeric_limits<int8_t>::min();
+    int32_t max = std::numeric_limits<int8_t>::max();
 
     const int32_t *const bias;
     const int n_0;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -90,131 +89,131 @@ void sme2_interleaved_nomerge_s8q_mopa_1VLx4VL(const int8_t *const A, const int8
   KernelArgs args(A, B, C, ldc, M, N, K, bias, rq, n_0, accumulate, accumulator_buffer);
 
   __asm__ __volatile__(
-      "ldr x15, [%x[args], %[offsetof_flags]]\n"
+      "ldr x14, [%x[args], %[offsetof_flags]]\n"
       ".inst 0xd503477f  // SMSTART ZA\n"
       "ptrue p1.b\n"
       ".inst 0x25207811  // ptrue pn9.b\n"
-      "ldr x14, [%x[args], %[offsetof_accumulator_buffer]]\n"
       "ldr x13, [%x[args], %[offsetof_accumulator_buffer]]\n"
-      "tbz x15, #0, 2f\n"
+      "ldr x11, [%x[args], %[offsetof_accumulator_buffer]]\n"
+      "tbz x14, #0, 2f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5d8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x14]\n"
-      ".inst 0xa041c5c0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c5dc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa040c5a4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x13]\n"
+      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa041c5a8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n"
+      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa042c5a8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n"
+      ".inst 0xc0840502  // mova za2h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa043c5b0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x13, x13, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
-      "ldr w11, [%x[args], %[offsetof_M]]\n"
-      "mov x10, #0x0\n"
+      "ldr w10, [%x[args], %[offsetof_M]]\n"
       "mov x9, #0x0\n"
-      "ldr w28, [%x[args], %[offsetof_N]]\n"
-      "ldr x27, [%x[args], %[offsetof_A]]\n"
+      "mov x28, #0x0\n"
+      "ldr w27, [%x[args], %[offsetof_N]]\n"
+      "ldr x26, [%x[args], %[offsetof_A]]\n"
       "3:"  // M and N loop
-      "mov x26, x27\n"
-      ".inst 0x25bc6530  // whilelt pn8.s, x9, x28, VLx4\n"
-      "tbnz x15, #0, 4f\n"
+      "mov x25, x26\n"
+      ".inst 0x25bb6790  // whilelt pn8.s, x28, x27, VLx4\n"
+      "tbnz x14, #0, 4f\n"
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      ".inst 0xa009c290  // ld1w { z16.s-z19.s }, p8/Z, [x20, x9, LSL #2]\n"
-      ".inst 0xc0902600  // addha za0.s, p1/M, p1/M, z16.s\n"
-      ".inst 0xc0902621  // addha za1.s, p1/M, p1/M, z17.s\n"
-      ".inst 0xc0902642  // addha za2.s, p1/M, p1/M, z18.s\n"
-      ".inst 0xc0902663  // addha za3.s, p1/M, p1/M, z19.s\n"
+      ".inst 0xa11cc289  // ldnt1w { z1.s, z5.s, z9.s, z13.s }, p8/Z, [x20, x28, LSL #2]\n"
+      ".inst 0xc0902420  // addha za0.s, p1/M, p1/M, z1.s\n"
+      ".inst 0xc09024a1  // addha za1.s, p1/M, p1/M, z5.s\n"
+      ".inst 0xc0902522  // addha za2.s, p1/M, p1/M, z9.s\n"
+      ".inst 0xc09025a3  // addha za3.s, p1/M, p1/M, z13.s\n"
       "4:"  // Prepare accumulators: Test for last block
-      "mov x20, x9\n"
-      "mov x21, x10\n"
+      "mov x20, x28\n"
+      "mov x21, x9\n"
       "incw x20, ALL, MUL #4\n"
       "incw x21\n"
-      "cmp x20, x28\n"
-      "mov x20, x15\n"
-      "csel x21, x10, x21, LT\n"
-      "bfm x15, XZR, #0x0, #0x0  // bfc x15, #0x0, #0x1\n"
-      "cmp x21, x11\n"
-      "csel x15, x20, x15, LT\n"
+      "cmp x20, x27\n"
+      "csel x21, x9, x21, LT\n"
+      "mov x20, x14\n"
+      "bfm x14, XZR, #0x0, #0x0  // bfc x14, #0x0, #0x1\n"
+      "cmp x21, x10\n"
+      "csel x14, x20, x14, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x9, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      "ld1b { z5.b }, p1/Z, [x26]\n"
-      ".inst 0xa14086e0  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn9.b/Z, [x23]\n"
-      "ld1b { z31.b }, p1/Z, [x26, #1, MUL VL]\n"
-      ".inst 0xa14186f2  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      "ld1b { z1.b }, p1/Z, [x26, #2, MUL VL]\n"
-      ".inst 0xa14286f0  // ld1b { z16.b, z20.b, z24.b, z28.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      "ld1b { z6.b }, p1/Z, [x26, #3, MUL VL]\n"
-      "addvl x26, x26, #4\n"
-      ".inst 0xa14386e3  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x28, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      "ld1b { z20.b }, p1/Z, [x25]\n"
+      ".inst 0xa04086e5  // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23]\n"
+      "ld1b { z11.b }, p1/Z, [x25, #1, MUL VL]\n"
+      ".inst 0xa04186f9  // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      "ld1b { z2.b }, p1/Z, [x25, #2, MUL VL]\n"
+      ".inst 0xa04286fd  // ldnt1b { z28.b-z31.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      "ld1b { z14.b }, p1/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0xa04386f1  // ldnt1b { z16.b-z19.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0xa08024a0  // smopa za0.s, p1/M, p1/M, z5.b, z0.b\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa08424a1  // smopa za1.s, p1/M, p1/M, z5.b, z4.b\n"
-      ".inst 0xa08824a2  // smopa za2.s, p1/M, p1/M, z5.b, z8.b\n"
-      ".inst 0xa08c24a3  // smopa za3.s, p1/M, p1/M, z5.b, z12.b\n"
-      "ld1b { z5.b }, p1/Z, [x26]\n"
-      ".inst 0xa09227e0  // smopa za0.s, p1/M, p1/M, z31.b, z18.b\n"
-      ".inst 0xa14086e0  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa09627e1  // smopa za1.s, p1/M, p1/M, z31.b, z22.b\n"
-      ".inst 0xa09a27e2  // smopa za2.s, p1/M, p1/M, z31.b, z26.b\n"
-      ".inst 0xa09e27e3  // smopa za3.s, p1/M, p1/M, z31.b, z30.b\n"
-      "ld1b { z31.b }, p1/Z, [x26, #1, MUL VL]\n"
-      ".inst 0xa0902420  // smopa za0.s, p1/M, p1/M, z1.b, z16.b\n"
-      ".inst 0xa14186f2  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa0942421  // smopa za1.s, p1/M, p1/M, z1.b, z20.b\n"
-      ".inst 0xa0982422  // smopa za2.s, p1/M, p1/M, z1.b, z24.b\n"
-      ".inst 0xa09c2423  // smopa za3.s, p1/M, p1/M, z1.b, z28.b\n"
-      "ld1b { z1.b }, p1/Z, [x26, #2, MUL VL]\n"
-      ".inst 0xa14286f0  // ld1b { z16.b, z20.b, z24.b, z28.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      ".inst 0xa08324c0  // smopa za0.s, p1/M, p1/M, z6.b, z3.b\n"
-      ".inst 0xa08724c1  // smopa za1.s, p1/M, p1/M, z6.b, z7.b\n"
-      ".inst 0xa08b24c2  // smopa za2.s, p1/M, p1/M, z6.b, z11.b\n"
-      ".inst 0xa08f24c3  // smopa za3.s, p1/M, p1/M, z6.b, z15.b\n"
-      "ld1b { z6.b }, p1/Z, [x26, #3, MUL VL]\n"
-      "addvl x26, x26, #4\n"
-      ".inst 0xa14386e3  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      ".inst 0xa0842680  // smopa za0.s, p1/M, p1/M, z20.b, z4.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa0852681  // smopa za1.s, p1/M, p1/M, z20.b, z5.b\n"
+      ".inst 0xa0862682  // smopa za2.s, p1/M, p1/M, z20.b, z6.b\n"
+      ".inst 0xa0872683  // smopa za3.s, p1/M, p1/M, z20.b, z7.b\n"
+      "ld1b { z20.b }, p1/Z, [x25]\n"
+      ".inst 0xa0982560  // smopa za0.s, p1/M, p1/M, z11.b, z24.b\n"
+      ".inst 0xa04086e5  // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa0992561  // smopa za1.s, p1/M, p1/M, z11.b, z25.b\n"
+      ".inst 0xa09a2562  // smopa za2.s, p1/M, p1/M, z11.b, z26.b\n"
+      ".inst 0xa09b2563  // smopa za3.s, p1/M, p1/M, z11.b, z27.b\n"
+      "ld1b { z11.b }, p1/Z, [x25, #1, MUL VL]\n"
+      ".inst 0xa09c2440  // smopa za0.s, p1/M, p1/M, z2.b, z28.b\n"
+      ".inst 0xa04186f9  // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa09d2441  // smopa za1.s, p1/M, p1/M, z2.b, z29.b\n"
+      ".inst 0xa09e2442  // smopa za2.s, p1/M, p1/M, z2.b, z30.b\n"
+      ".inst 0xa09f2443  // smopa za3.s, p1/M, p1/M, z2.b, z31.b\n"
+      "ld1b { z2.b }, p1/Z, [x25, #2, MUL VL]\n"
+      ".inst 0xa04286fd  // ldnt1b { z28.b-z31.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      ".inst 0xa09025c0  // smopa za0.s, p1/M, p1/M, z14.b, z16.b\n"
+      ".inst 0xa09125c1  // smopa za1.s, p1/M, p1/M, z14.b, z17.b\n"
+      ".inst 0xa09225c2  // smopa za2.s, p1/M, p1/M, z14.b, z18.b\n"
+      ".inst 0xa09325c3  // smopa za3.s, p1/M, p1/M, z14.b, z19.b\n"
+      "ld1b { z14.b }, p1/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0xa04386f1  // ldnt1b { z16.b-z19.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0xa08024a0  // smopa za0.s, p1/M, p1/M, z5.b, z0.b\n"
-      ".inst 0xa08424a1  // smopa za1.s, p1/M, p1/M, z5.b, z4.b\n"
-      ".inst 0xa08824a2  // smopa za2.s, p1/M, p1/M, z5.b, z8.b\n"
-      ".inst 0xa08c24a3  // smopa za3.s, p1/M, p1/M, z5.b, z12.b\n"
-      ".inst 0xa09227e0  // smopa za0.s, p1/M, p1/M, z31.b, z18.b\n"
-      ".inst 0xa09627e1  // smopa za1.s, p1/M, p1/M, z31.b, z22.b\n"
-      ".inst 0xa09a27e2  // smopa za2.s, p1/M, p1/M, z31.b, z26.b\n"
-      ".inst 0xa09e27e3  // smopa za3.s, p1/M, p1/M, z31.b, z30.b\n"
-      ".inst 0xa0902420  // smopa za0.s, p1/M, p1/M, z1.b, z16.b\n"
-      ".inst 0xa0942421  // smopa za1.s, p1/M, p1/M, z1.b, z20.b\n"
-      ".inst 0xa0982422  // smopa za2.s, p1/M, p1/M, z1.b, z24.b\n"
-      ".inst 0xa09c2423  // smopa za3.s, p1/M, p1/M, z1.b, z28.b\n"
-      ".inst 0xa08324c0  // smopa za0.s, p1/M, p1/M, z6.b, z3.b\n"
-      ".inst 0xa08724c1  // smopa za1.s, p1/M, p1/M, z6.b, z7.b\n"
-      ".inst 0xa08b24c2  // smopa za2.s, p1/M, p1/M, z6.b, z11.b\n"
-      ".inst 0xa08f24c3  // smopa za3.s, p1/M, p1/M, z6.b, z15.b\n"
+      ".inst 0xa0842680  // smopa za0.s, p1/M, p1/M, z20.b, z4.b\n"
+      ".inst 0xa0852681  // smopa za1.s, p1/M, p1/M, z20.b, z5.b\n"
+      ".inst 0xa0862682  // smopa za2.s, p1/M, p1/M, z20.b, z6.b\n"
+      ".inst 0xa0872683  // smopa za3.s, p1/M, p1/M, z20.b, z7.b\n"
+      ".inst 0xa0982560  // smopa za0.s, p1/M, p1/M, z11.b, z24.b\n"
+      ".inst 0xa0992561  // smopa za1.s, p1/M, p1/M, z11.b, z25.b\n"
+      ".inst 0xa09a2562  // smopa za2.s, p1/M, p1/M, z11.b, z26.b\n"
+      ".inst 0xa09b2563  // smopa za3.s, p1/M, p1/M, z11.b, z27.b\n"
+      ".inst 0xa09c2440  // smopa za0.s, p1/M, p1/M, z2.b, z28.b\n"
+      ".inst 0xa09d2441  // smopa za1.s, p1/M, p1/M, z2.b, z29.b\n"
+      ".inst 0xa09e2442  // smopa za2.s, p1/M, p1/M, z2.b, z30.b\n"
+      ".inst 0xa09f2443  // smopa za3.s, p1/M, p1/M, z2.b, z31.b\n"
+      ".inst 0xa09025c0  // smopa za0.s, p1/M, p1/M, z14.b, z16.b\n"
+      ".inst 0xa09125c1  // smopa za1.s, p1/M, p1/M, z14.b, z17.b\n"
+      ".inst 0xa09225c2  // smopa za2.s, p1/M, p1/M, z14.b, z18.b\n"
+      ".inst 0xa09325c3  // smopa za3.s, p1/M, p1/M, z14.b, z19.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      "ld1b { z16.b }, p1/Z, [x26]\n"
-      "subs x20, x20, #0x1\n"
-      "addvl x26, x26, #1\n"
+      "ld1b { z16.b }, p1/Z, [x25]\n"
+      "subs x21, x21, #0x1\n"
+      "addvl x25, x25, #1\n"
       ".inst 0xa04086e4  // ld1b { z4.b-z7.b }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #4\n"
       ".inst 0xa0842600  // smopa za0.s, p1/M, p1/M, z16.b, z4.b\n"
@@ -223,182 +222,182 @@ void sme2_interleaved_nomerge_s8q_mopa_1VLx4VL(const int8_t *const A, const int8
       ".inst 0xa0872603  // smopa za3.s, p1/M, p1/M, z16.b, z7.b\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
-      "ld1w { z15.s }, p1/Z, [x26]\n"
-      "addvl x26, x26, #1\n"
+      "ld1w { z15.s }, p1/Z, [x25]\n"
+      "addvl x25, x25, #1\n"
       ".inst 0xc09125e0  // addva za0.s, p1/M, p1/M, z15.s\n"
       ".inst 0xc09125e1  // addva za1.s, p1/M, p1/M, z15.s\n"
       ".inst 0xc09125e2  // addva za2.s, p1/M, p1/M, z15.s\n"
       ".inst 0xc09125e3  // addva za3.s, p1/M, p1/M, z15.s\n"
-      "tbz x15, #1, 14f\n"
-      "tbz x15, #0, 12f\n"
+      "tbz x14, #1, 14f\n"
+      "tbz x14, #0, 12f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14]\n"
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
-      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa040c5a0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x13]\n"
+      ".inst 0xc0860418  // mova { z24.s-z27.s }, za0h.s[x12]\n"
+      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5a0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n"
+      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
       ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
-      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
-      ".inst 0xa042c5c0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
-      ".inst 0xa060c5bc  // st1w { z28.s-z31.s }, pn9.b, [x13]\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xa061c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13, #0x4, MUL VL]\n"
-      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
+      ".inst 0xa042c5b0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c5b0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5b4  // st1w { z20.s-z23.s }, pn9.b, [x13, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c5b8  // st1w { z24.s-z27.s }, pn9.b, [x13, #0xc, MUL VL]\n"
+      ".inst 0xa060c578  // st1w { z24.s-z27.s }, pn9.b, [x11]\n"
       "addvl x13, x13, #16\n"
+      ".inst 0xa061c564  // st1w { z4.s-z7.s }, pn9.b, [x11, #0x4, MUL VL]\n"
+      ".inst 0xa062c574  // st1w { z20.s-z23.s }, pn9.b, [x11, #0x8, MUL VL]\n"
+      ".inst 0xa063c560  // st1w { z0.s-z3.s }, pn9.b, [x11, #0xc, MUL VL]\n"
+      "addvl x11, x11, #16\n"
       "blt 11b\n"
       "b 21f\n"
       "12:"  // Store to partial result buffer: Store only
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
+      ".inst 0xa060c564  // st1w { z4.s-z7.s }, pn9.b, [x11]\n"
       ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13]\n"
+      ".inst 0xc0860468  // mova { z8.s-z11.s }, za3h.s[x12]\n"
+      ".inst 0xa061c574  // st1w { z20.s-z23.s }, pn9.b, [x11, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5a4  // st1w { z4.s-z7.s }, pn9.b, [x13, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c5ac  // st1w { z12.s-z15.s }, pn9.b, [x13, #0x8, MUL VL]\n"
-      ".inst 0xa063c5b0  // st1w { z16.s-z19.s }, pn9.b, [x13, #0xc, MUL VL]\n"
-      "addvl x13, x13, #16\n"
+      ".inst 0xa062c56c  // st1w { z12.s-z15.s }, pn9.b, [x11, #0x8, MUL VL]\n"
+      ".inst 0xa063c568  // st1w { z8.s-z11.s }, pn9.b, [x11, #0xc, MUL VL]\n"
+      "addvl x11, x11, #16\n"
       "blt 13b\n"
       "b 21f\n"
       "14:"  // Store to output array
-      "ldr x25, [%x[args], %[offsetof_C]]\n"
-      "sub x24, x11, x10\n"
+      "ldr x24, [%x[args], %[offsetof_C]]\n"
+      "add x24, x24, x28\n"  // C += n
+      "sub x23, x10, x9\n"
       "ld1rw { z4.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
-      "ldr x23, [%x[args], %[offsetof_ldcb]]\n"
+      "ldr x22, [%x[args], %[offsetof_ldcb]]\n"
+      "madd x24, x9, x22, x24\n"  // C += m * ldc
       "ld1rw { z5.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
       "ld1rw { z6.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
-      "add x25, x25, x9\n"  // C += n
       "ld1rw { z7.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
-      "madd x25, x10, x23, x25\n"  // C += m * ldc
-      "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "ld1rw { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "ld1rw { z3.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "ld1rw { z8.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
+      "ld1rw { z12.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z13.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z14.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z15.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
       "ld1rw { z21.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
       "ld1rw { z20.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
-      "tbz x15, #2, 15f\n"
-      "ldr w22, [%x[args], %[offsetof_n_0]]\n"
-      "ldr x21, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "tbz x14, #2, 15f\n"
+      "ldr w21, [%x[args], %[offsetof_n_0]]\n"
+      "add x21, x21, x28\n"
+      "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "add x20, x20, x21, LSL #2\n"
+      ".inst 0xa040c284  // ld1w { z4.s-z7.s }, p8/Z, [x20]\n"
       "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n"
-      "add x22, x22, x9\n"
-      "add x21, x21, x22, LSL #2\n"
-      "add x20, x20, x22, LSL #2\n"
-      ".inst 0xa040c2a4  // ld1w { z4.s-z7.s }, p8/Z, [x21]\n"
-      ".inst 0xa040c280  // ld1w { z0.s-z3.s }, p8/Z, [x20]\n"
+      "add x20, x20, x21, LSL #2\n"
+      ".inst 0xa040c28c  // ld1w { z12.s-z15.s }, p8/Z, [x20]\n"
       "15:"  // Store to output array: Load per-channel parameters: End
       "cntw x20\n"
-      "whilelt p0.b, x9, x28\n"
-      "cmp x24, x20\n"
-      "mov x12, #0x0\n"
-      "csel x20, x24, x20, LT\n"
+      "whilelt p0.b, x28, x27\n"
+      "cmp x23, x20\n"
+      "csel x20, x23, x20, LT\n"
       "lsr x21, x20, #0x1\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x1\n"
       "cbz x21, 17f\n"
       "16:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc0860010  // mova { z16.s-z17.s }, za0h.s[x12, 0:1]\n"
-      ".inst 0xc086005e  // mova { z30.s-z31.s }, za1h.s[x12, 0:1]\n"
-      ".inst 0xc086009a  // mova { z26.s-z27.s }, za2h.s[x12, 0:1]\n"
-      ".inst 0xc08600cc  // mova { z12.s-z13.s }, za3h.s[x12, 0:1]\n"
-      ".inst 0xc1a4a410  // sqdmulh { z16.s-z17.s }, { z16.s-z17.s }, z4.s\n"
-      ".inst 0xc1a5a41e  // sqdmulh { z30.s-z31.s }, { z30.s-z31.s }, z5.s\n"
+      ".inst 0xc086001a  // mova { z26.s-z27.s }, za0h.s[x12, 0:1]\n"
+      ".inst 0xc086005c  // mova { z28.s-z29.s }, za1h.s[x12, 0:1]\n"
+      ".inst 0xc1a4a41a  // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z4.s\n"
+      ".inst 0xc0860096  // mova { z22.s-z23.s }, za2h.s[x12, 0:1]\n"
+      ".inst 0xc08600d0  // mova { z16.s-z17.s }, za3h.s[x12, 0:1]\n"
+      ".inst 0xc1a5a41c  // sqdmulh { z28.s-z29.s }, { z28.s-z29.s }, z5.s\n"
+      ".inst 0xc1a6a416  // sqdmulh { z22.s-z23.s }, { z22.s-z23.s }, z6.s\n"
       "add x12, x12, #0x2\n"
-      ".inst 0xc1a6a41a  // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z6.s\n"
       "cmp x12, x21, LSL #1\n"
-      ".inst 0xc1a7a40c  // sqdmulh { z12.s-z13.s }, { z12.s-z13.s }, z7.s\n"
-      ".inst 0xc1a0a230  // srshl { z16.s-z17.s }, { z16.s-z17.s }, z0.s\n"
-      ".inst 0xc1a1a23e  // srshl { z30.s-z31.s }, { z30.s-z31.s }, z1.s\n"
-      ".inst 0xc1a2a23a  // srshl { z26.s-z27.s }, { z26.s-z27.s }, z2.s\n"
-      ".inst 0xc1a3a22c  // srshl { z12.s-z13.s }, { z12.s-z13.s }, z3.s\n"
-      ".inst 0xc1a8a310  // add { z16.s-z17.s }, { z16.s-z17.s }, z8.s\n"
-      ".inst 0xc1a8a31e  // add { z30.s-z31.s }, { z30.s-z31.s }, z8.s\n"
-      ".inst 0xc1a8a31a  // add { z26.s-z27.s }, { z26.s-z27.s }, z8.s\n"
-      ".inst 0xc1a8a30c  // add { z12.s-z13.s }, { z12.s-z13.s }, z8.s\n"
-      ".inst 0xc1b4c6b0  // sclamp { z16.s-z17.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4c6be  // sclamp { z30.s-z31.s }, z21.s, z20.s\n"
+      ".inst 0xc1a7a410  // sqdmulh { z16.s-z17.s }, { z16.s-z17.s }, z7.s\n"
+      ".inst 0xc1aca23a  // srshl { z26.s-z27.s }, { z26.s-z27.s }, z12.s\n"
+      ".inst 0xc1ada23c  // srshl { z28.s-z29.s }, { z28.s-z29.s }, z13.s\n"
+      ".inst 0xc1aea236  // srshl { z22.s-z23.s }, { z22.s-z23.s }, z14.s\n"
+      ".inst 0xc1afa230  // srshl { z16.s-z17.s }, { z16.s-z17.s }, z15.s\n"
+      ".inst 0xc1a0a31a  // add { z26.s-z27.s }, { z26.s-z27.s }, z0.s\n"
+      ".inst 0xc1a0a31c  // add { z28.s-z29.s }, { z28.s-z29.s }, z0.s\n"
+      ".inst 0xc1a0a316  // add { z22.s-z23.s }, { z22.s-z23.s }, z0.s\n"
+      ".inst 0xc1a0a310  // add { z16.s-z17.s }, { z16.s-z17.s }, z0.s\n"
       ".inst 0xc1b4c6ba  // sclamp { z26.s-z27.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4c6ac  // sclamp { z12.s-z13.s }, z21.s, z20.s\n"
-      "uzp1 z19.b, z16.b, z30.b\n"
-      "uzp1 z18.b, z17.b, z31.b\n"
-      "uzp1 z17.b, z26.b, z12.b\n"
-      "uzp1 z16.b, z27.b, z13.b\n"
-      "uzp1 z17.b, z19.b, z17.b\n"
-      "uzp1 z16.b, z18.b, z16.b\n"
-      "st1b { z17.b }, p0, [x25]\n"
-      "add x25, x25, x23\n"
-      "st1b { z16.b }, p0, [x25]\n"
-      "add x25, x25, x23\n"
+      ".inst 0xc1b4c6bc  // sclamp { z28.s-z29.s }, z21.s, z20.s\n"
+      "uzp1 z19.b, z26.b, z28.b\n"
+      ".inst 0xc1b4c6b6  // sclamp { z22.s-z23.s }, z21.s, z20.s\n"
+      ".inst 0xc1b4c6b0  // sclamp { z16.s-z17.s }, z21.s, z20.s\n"
+      "uzp1 z16.b, z22.b, z16.b\n"
+      "uzp1 z18.b, z27.b, z29.b\n"
+      "uzp1 z17.b, z23.b, z17.b\n"
+      "uzp1 z16.b, z19.b, z16.b\n"
+      "st1b { z16.b }, p0, [x24]\n"
+      "add x24, x24, x22\n"
+      "uzp1 z16.b, z18.b, z17.b\n"
+      "st1b { z16.b }, p0, [x24]\n"
+      "add x24, x24, x22\n"
       "blt 16b\n"
       "17:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 18f\n"
       ".inst 0xc086000a  // mova { z10.s-z11.s }, za0h.s[x12, 0:1]\n"
-      ".inst 0xc086005a  // mova { z26.s-z27.s }, za1h.s[x12, 0:1]\n"
-      ".inst 0xc086008e  // mova { z14.s-z15.s }, za2h.s[x12, 0:1]\n"
-      ".inst 0xc08600d6  // mova { z22.s-z23.s }, za3h.s[x12, 0:1]\n"
+      ".inst 0xc0860058  // mova { z24.s-z25.s }, za1h.s[x12, 0:1]\n"
       ".inst 0xc1a4a40a  // sqdmulh { z10.s-z11.s }, { z10.s-z11.s }, z4.s\n"
-      ".inst 0xc1a5a41a  // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z5.s\n"
-      ".inst 0xc1a6a40e  // sqdmulh { z14.s-z15.s }, { z14.s-z15.s }, z6.s\n"
-      ".inst 0xc1a7a416  // sqdmulh { z22.s-z23.s }, { z22.s-z23.s }, z7.s\n"
-      ".inst 0xc1a0a22a  // srshl { z10.s-z11.s }, { z10.s-z11.s }, z0.s\n"
-      ".inst 0xc1a1a23a  // srshl { z26.s-z27.s }, { z26.s-z27.s }, z1.s\n"
-      ".inst 0xc1a2a22e  // srshl { z14.s-z15.s }, { z14.s-z15.s }, z2.s\n"
-      ".inst 0xc1a3a236  // srshl { z22.s-z23.s }, { z22.s-z23.s }, z3.s\n"
-      ".inst 0xc1a8a30a  // add { z10.s-z11.s }, { z10.s-z11.s }, z8.s\n"
-      ".inst 0xc1a8a31a  // add { z26.s-z27.s }, { z26.s-z27.s }, z8.s\n"
-      ".inst 0xc1a8a30e  // add { z14.s-z15.s }, { z14.s-z15.s }, z8.s\n"
-      ".inst 0xc1a8a316  // add { z22.s-z23.s }, { z22.s-z23.s }, z8.s\n"
+      ".inst 0xc086009a  // mova { z26.s-z27.s }, za2h.s[x12, 0:1]\n"
+      ".inst 0xc08600de  // mova { z30.s-z31.s }, za3h.s[x12, 0:1]\n"
+      ".inst 0xc1a5a418  // sqdmulh { z24.s-z25.s }, { z24.s-z25.s }, z5.s\n"
+      ".inst 0xc1a6a41a  // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z6.s\n"
+      ".inst 0xc1a7a41e  // sqdmulh { z30.s-z31.s }, { z30.s-z31.s }, z7.s\n"
+      ".inst 0xc1aca22a  // srshl { z10.s-z11.s }, { z10.s-z11.s }, z12.s\n"
+      ".inst 0xc1ada238  // srshl { z24.s-z25.s }, { z24.s-z25.s }, z13.s\n"
+      ".inst 0xc1aea23a  // srshl { z26.s-z27.s }, { z26.s-z27.s }, z14.s\n"
+      ".inst 0xc1afa23e  // srshl { z30.s-z31.s }, { z30.s-z31.s }, z15.s\n"
+      ".inst 0xc1a0a30a  // add { z10.s-z11.s }, { z10.s-z11.s }, z0.s\n"
+      ".inst 0xc1a0a318  // add { z24.s-z25.s }, { z24.s-z25.s }, z0.s\n"
+      ".inst 0xc1a0a31a  // add { z26.s-z27.s }, { z26.s-z27.s }, z0.s\n"
+      ".inst 0xc1a0a31e  // add { z30.s-z31.s }, { z30.s-z31.s }, z0.s\n"
       ".inst 0xc1b4c6aa  // sclamp { z10.s-z11.s }, z21.s, z20.s\n"
+      ".inst 0xc1b4c6b8  // sclamp { z24.s-z25.s }, z21.s, z20.s\n"
+      "uzp1 z17.b, z10.b, z24.b\n"
       ".inst 0xc1b4c6ba  // sclamp { z26.s-z27.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4c6ae  // sclamp { z14.s-z15.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4c6b6  // sclamp { z22.s-z23.s }, z21.s, z20.s\n"
-      "uzp1 z17.b, z10.b, z26.b\n"
-      "uzp1 z16.b, z14.b, z22.b\n"
+      ".inst 0xc1b4c6be  // sclamp { z30.s-z31.s }, z21.s, z20.s\n"
+      "uzp1 z16.b, z26.b, z30.b\n"
       "uzp1 z16.b, z17.b, z16.b\n"
-      "st1b { z16.b }, p0, [x25]\n"
+      "st1b { z16.b }, p0, [x24]\n"
       "18:"  // Store to output array: Accumulator row 0 oddments: End
       "19:"  // Store to output array: End
-      "tbz x15, #0, 21f\n"
+      "tbz x14, #0, 21f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "20:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5d8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x14]\n"
-      ".inst 0xa041c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5c0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa040c5bc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x13]\n"
+      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa041c5b0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c5ac  // ld1w { z12.s-z15.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n"
+      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa043c5a0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n"
       ".inst 0xc0840403  // mova za3h.s[x12], { z0.s-z3.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x13, x13, #16\n"
       "blt 20b\n"
       "21:"  // End block
-      "incw x9, ALL, MUL #4\n"
-      "cmp x9, x28\n"
+      "incw x28, ALL, MUL #4\n"
+      "cmp x28, x27\n"
       "blt 3b\n"
-      "incw x10\n"
-      "mov x9, #0x0\n"
-      "cmp x10, x11\n"
-      "mov x27, x26\n"
+      "incw x9\n"
+      "cmp x9, x10\n"
+      "mov x28, #0x0\n"
+      "mov x26, x25\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
       :
       : [args] "r" (&args), [offsetof_A] "I" (offsetof(KernelArgs, A)), [offsetof_B] "I" (offsetof(KernelArgs, B)), [offsetof_C] "I" (offsetof(KernelArgs, C)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_M] "I" (offsetof(KernelArgs, M)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_Requantize32_c_offset] "I" (offsetof(Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(Requantize32, minval)), [offsetof_Requantize32_per_channel_muls] "I" (offsetof(Requantize32, per_channel_muls)), [offsetof_Requantize32_per_channel_right_shifts] "I" (offsetof(Requantize32, per_channel_right_shifts)), [offsetof_Requantize32_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_Requantize32_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [offsetof_accumulator_buffer] "I" (offsetof(KernelArgs, accumulator_buffer)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_flags] "I" (offsetof(KernelArgs, flags)), [offsetof_kstride_bytes] "I" (offsetof(KernelArgs, kstride_bytes)), [offsetof_ldcb] "I" (offsetof(KernelArgs, ldcb)), [offsetof_n_0] "I" (offsetof(KernelArgs, n_0)), [rq] "r" (&rq)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL.hpp
index 68b43328a2..b897efe0dc 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
 class cls_sme2_interleaved_nomerge_s8q_mopa_2VLx2VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef int8_t result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, int8_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Requantize32 &rq, const int n_0, bool accumulate, int32_t *const accumulator_buffer);
@@ -61,7 +60,7 @@ public:
 
   static constexpr bool supports_accumulate()
   {
-    return true;
+    return false;
   }
 
   static constexpr bool supports_bias()
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8q_mopa_2VLx2VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 2, 2, 4, true> transforms = {};
+  StdTransformsSME<operand_type, result_type, 2, 2, 4, true> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp
index 8831a224ad..b60573898a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL/generic.cpp
@@ -49,7 +49,7 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
         B(B), kstride_bytes(roundup(K, 4) * sizeof(int8_t)),
         C(C), ldcb(ldc * sizeof(int8_t)),
         M(M), N(N), K(K),
-        min(0), max(0),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias), n_0(n_0),
         accumulator_buffer(accumulator_buffer),
@@ -74,14 +74,13 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
     const long kstride_bytes;
     int8_t *const C;
     const long ldcb;
-    const long M, N, K;
-    int32_t min;
-    int32_t max;
+    const long M, N, K, n_loops, n_tail_iters;
+    int32_t min = std::numeric_limits<int8_t>::min();
+    int32_t max = std::numeric_limits<int8_t>::max();
 
     const int32_t *const bias;
     const int n_0;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -100,17 +99,17 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5f4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xa040c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa041c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -125,108 +124,108 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      ".inst 0xa10a4294  // ld1w { z20.s, z28.s }, p8/Z, [x20, x10, LSL #2]\n"
-      ".inst 0xc0902680  // addha za0.s, p1/M, p1/M, z20.s\n"
-      ".inst 0xc0902781  // addha za1.s, p1/M, p1/M, z28.s\n"
-      ".inst 0xc0902682  // addha za2.s, p1/M, p1/M, z20.s\n"
-      ".inst 0xc0902783  // addha za3.s, p1/M, p1/M, z28.s\n"
+      ".inst 0xa00a4299  // ldnt1w { z24.s-z25.s }, p8/Z, [x20, x10, LSL #2]\n"
+      ".inst 0xc0902700  // addha za0.s, p1/M, p1/M, z24.s\n"
+      ".inst 0xc0902721  // addha za1.s, p1/M, p1/M, z25.s\n"
+      ".inst 0xc0902702  // addha za2.s, p1/M, p1/M, z24.s\n"
+      ".inst 0xc0902723  // addha za3.s, p1/M, p1/M, z25.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20, ALL, MUL #2\n"
       "incw x21, ALL, MUL #2\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa0400778  // ld1b { z24.b-z25.b }, pn9.b/Z, [x27]\n"
-      ".inst 0xa14006f7  // ld1b { z23.b, z31.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa1410776  // ld1b { z22.b, z30.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0xa04106fa  // ld1b { z26.b-z27.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa0420766  // ld1b { z6.b-z7.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14206e0  // ld1b { z0.b, z8.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa043077c  // ld1b { z28.b-z29.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x10, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa1400763  // ld1b { z3.b, z11.b }, pn9.b/Z, [x27]\n"
+      ".inst 0xa14006f9  // ldnt1b { z17.b, z25.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa1410774  // ld1b { z20.b, z28.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa04106f7  // ldnt1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0xa1420775  // ld1b { z21.b, z29.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa14206f8  // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa1430765  // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa04306ec  // ld1b { z12.b-z13.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa14306ef  // ldnt1b { z7.b, z15.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0xa0972700  // smopa za0.s, p1/M, p1/M, z24.b, z23.b\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa09f2701  // smopa za1.s, p1/M, p1/M, z24.b, z31.b\n"
-      ".inst 0xa0972722  // smopa za2.s, p1/M, p1/M, z25.b, z23.b\n"
-      ".inst 0xa09f2723  // smopa za3.s, p1/M, p1/M, z25.b, z31.b\n"
-      ".inst 0xa0400778  // ld1b { z24.b-z25.b }, pn9.b/Z, [x27]\n"
-      ".inst 0xa09a26c0  // smopa za0.s, p1/M, p1/M, z22.b, z26.b\n"
-      ".inst 0xa14006f7  // ld1b { z23.b, z31.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa09b26c1  // smopa za1.s, p1/M, p1/M, z22.b, z27.b\n"
-      ".inst 0xa09a27c2  // smopa za2.s, p1/M, p1/M, z30.b, z26.b\n"
-      ".inst 0xa09b27c3  // smopa za3.s, p1/M, p1/M, z30.b, z27.b\n"
-      ".inst 0xa1410776  // ld1b { z22.b, z30.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0xa08024c0  // smopa za0.s, p1/M, p1/M, z6.b, z0.b\n"
-      ".inst 0xa04106fa  // ld1b { z26.b-z27.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa08824c1  // smopa za1.s, p1/M, p1/M, z6.b, z8.b\n"
-      ".inst 0xa08024e2  // smopa za2.s, p1/M, p1/M, z7.b, z0.b\n"
-      ".inst 0xa08824e3  // smopa za3.s, p1/M, p1/M, z7.b, z8.b\n"
-      ".inst 0xa0420766  // ld1b { z6.b-z7.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14206e0  // ld1b { z0.b, z8.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa08c2780  // smopa za0.s, p1/M, p1/M, z28.b, z12.b\n"
-      ".inst 0xa08d2781  // smopa za1.s, p1/M, p1/M, z28.b, z13.b\n"
-      ".inst 0xa08c27a2  // smopa za2.s, p1/M, p1/M, z29.b, z12.b\n"
-      ".inst 0xa08d27a3  // smopa za3.s, p1/M, p1/M, z29.b, z13.b\n"
-      ".inst 0xa043077c  // ld1b { z28.b-z29.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      ".inst 0xa0912460  // smopa za0.s, p1/M, p1/M, z3.b, z17.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa0992461  // smopa za1.s, p1/M, p1/M, z3.b, z25.b\n"
+      ".inst 0xa0912562  // smopa za2.s, p1/M, p1/M, z11.b, z17.b\n"
+      ".inst 0xa0992563  // smopa za3.s, p1/M, p1/M, z11.b, z25.b\n"
+      ".inst 0xa1400763  // ld1b { z3.b, z11.b }, pn9.b/Z, [x27]\n"
+      ".inst 0xa0962680  // smopa za0.s, p1/M, p1/M, z20.b, z22.b\n"
+      ".inst 0xa14006f9  // ldnt1b { z17.b, z25.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa0972681  // smopa za1.s, p1/M, p1/M, z20.b, z23.b\n"
+      ".inst 0xa0962782  // smopa za2.s, p1/M, p1/M, z28.b, z22.b\n"
+      ".inst 0xa0972783  // smopa za3.s, p1/M, p1/M, z28.b, z23.b\n"
+      ".inst 0xa1410774  // ld1b { z20.b, z28.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa09026a0  // smopa za0.s, p1/M, p1/M, z21.b, z16.b\n"
+      ".inst 0xa04106f7  // ldnt1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0xa09826a1  // smopa za1.s, p1/M, p1/M, z21.b, z24.b\n"
+      ".inst 0xa09027a2  // smopa za2.s, p1/M, p1/M, z29.b, z16.b\n"
+      ".inst 0xa09827a3  // smopa za3.s, p1/M, p1/M, z29.b, z24.b\n"
+      ".inst 0xa1420775  // ld1b { z21.b, z29.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa14206f8  // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa08724a0  // smopa za0.s, p1/M, p1/M, z5.b, z7.b\n"
+      ".inst 0xa08f24a1  // smopa za1.s, p1/M, p1/M, z5.b, z15.b\n"
+      ".inst 0xa08725a2  // smopa za2.s, p1/M, p1/M, z13.b, z7.b\n"
+      ".inst 0xa08f25a3  // smopa za3.s, p1/M, p1/M, z13.b, z15.b\n"
+      ".inst 0xa1430765  // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa04306ec  // ld1b { z12.b-z13.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa14306ef  // ldnt1b { z7.b, z15.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0xa0972700  // smopa za0.s, p1/M, p1/M, z24.b, z23.b\n"
-      ".inst 0xa09f2701  // smopa za1.s, p1/M, p1/M, z24.b, z31.b\n"
-      ".inst 0xa0972722  // smopa za2.s, p1/M, p1/M, z25.b, z23.b\n"
-      ".inst 0xa09f2723  // smopa za3.s, p1/M, p1/M, z25.b, z31.b\n"
-      ".inst 0xa09a26c0  // smopa za0.s, p1/M, p1/M, z22.b, z26.b\n"
-      ".inst 0xa09b26c1  // smopa za1.s, p1/M, p1/M, z22.b, z27.b\n"
-      ".inst 0xa09a27c2  // smopa za2.s, p1/M, p1/M, z30.b, z26.b\n"
-      ".inst 0xa09b27c3  // smopa za3.s, p1/M, p1/M, z30.b, z27.b\n"
-      ".inst 0xa08024c0  // smopa za0.s, p1/M, p1/M, z6.b, z0.b\n"
-      ".inst 0xa08824c1  // smopa za1.s, p1/M, p1/M, z6.b, z8.b\n"
-      ".inst 0xa08024e2  // smopa za2.s, p1/M, p1/M, z7.b, z0.b\n"
-      ".inst 0xa08824e3  // smopa za3.s, p1/M, p1/M, z7.b, z8.b\n"
-      ".inst 0xa08c2780  // smopa za0.s, p1/M, p1/M, z28.b, z12.b\n"
-      ".inst 0xa08d2781  // smopa za1.s, p1/M, p1/M, z28.b, z13.b\n"
-      ".inst 0xa08c27a2  // smopa za2.s, p1/M, p1/M, z29.b, z12.b\n"
-      ".inst 0xa08d27a3  // smopa za3.s, p1/M, p1/M, z29.b, z13.b\n"
+      ".inst 0xa0912460  // smopa za0.s, p1/M, p1/M, z3.b, z17.b\n"
+      ".inst 0xa0992461  // smopa za1.s, p1/M, p1/M, z3.b, z25.b\n"
+      ".inst 0xa0912562  // smopa za2.s, p1/M, p1/M, z11.b, z17.b\n"
+      ".inst 0xa0992563  // smopa za3.s, p1/M, p1/M, z11.b, z25.b\n"
+      ".inst 0xa0962680  // smopa za0.s, p1/M, p1/M, z20.b, z22.b\n"
+      ".inst 0xa0972681  // smopa za1.s, p1/M, p1/M, z20.b, z23.b\n"
+      ".inst 0xa0962782  // smopa za2.s, p1/M, p1/M, z28.b, z22.b\n"
+      ".inst 0xa0972783  // smopa za3.s, p1/M, p1/M, z28.b, z23.b\n"
+      ".inst 0xa09026a0  // smopa za0.s, p1/M, p1/M, z21.b, z16.b\n"
+      ".inst 0xa09826a1  // smopa za1.s, p1/M, p1/M, z21.b, z24.b\n"
+      ".inst 0xa09027a2  // smopa za2.s, p1/M, p1/M, z29.b, z16.b\n"
+      ".inst 0xa09827a3  // smopa za3.s, p1/M, p1/M, z29.b, z24.b\n"
+      ".inst 0xa08724a0  // smopa za0.s, p1/M, p1/M, z5.b, z7.b\n"
+      ".inst 0xa08f24a1  // smopa za1.s, p1/M, p1/M, z5.b, z15.b\n"
+      ".inst 0xa08725a2  // smopa za2.s, p1/M, p1/M, z13.b, z7.b\n"
+      ".inst 0xa08f25a3  // smopa za3.s, p1/M, p1/M, z13.b, z15.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa0400762  // ld1b { z2.b-z3.b }, pn9.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xa1400773  // ld1b { z19.b, z27.b }, pn9.b/Z, [x27]\n"
+      "subs x21, x21, #0x1\n"
       "addvl x27, x27, #2\n"
       ".inst 0xa04006f0  // ld1b { z16.b-z17.b }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #2\n"
-      ".inst 0xa0902440  // smopa za0.s, p1/M, p1/M, z2.b, z16.b\n"
-      ".inst 0xa0912441  // smopa za1.s, p1/M, p1/M, z2.b, z17.b\n"
-      ".inst 0xa0902462  // smopa za2.s, p1/M, p1/M, z3.b, z16.b\n"
-      ".inst 0xa0912463  // smopa za3.s, p1/M, p1/M, z3.b, z17.b\n"
+      ".inst 0xa0902660  // smopa za0.s, p1/M, p1/M, z19.b, z16.b\n"
+      ".inst 0xa0912661  // smopa za1.s, p1/M, p1/M, z19.b, z17.b\n"
+      ".inst 0xa0902762  // smopa za2.s, p1/M, p1/M, z27.b, z16.b\n"
+      ".inst 0xa0912763  // smopa za3.s, p1/M, p1/M, z27.b, z17.b\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
-      ".inst 0xa1404767  // ld1w { z7.s, z15.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040476e  // ld1w { z14.s-z15.s }, pn9.b/Z, [x27]\n"
       "addvl x27, x27, #2\n"
-      ".inst 0xc09124e0  // addva za0.s, p1/M, p1/M, z7.s\n"
-      ".inst 0xc09124e1  // addva za1.s, p1/M, p1/M, z7.s\n"
+      ".inst 0xc09125c0  // addva za0.s, p1/M, p1/M, z14.s\n"
+      ".inst 0xc09125c1  // addva za1.s, p1/M, p1/M, z14.s\n"
       ".inst 0xc09125e2  // addva za2.s, p1/M, p1/M, z15.s\n"
       ".inst 0xc09125e3  // addva za3.s, p1/M, p1/M, z15.s\n"
       "tbz x16, #1, 14f\n"
@@ -234,25 +233,25 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
-      ".inst 0xa042c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
-      ".inst 0xa060c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14]\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xa061c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840703  // mova za3h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa040c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5fc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840781  // mova za1h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
+      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c5fc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840783  // mova za3h.s[x12], { z28.s-z31.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa060c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 24f\n"
@@ -260,71 +259,71 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14]\n"
+      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      ".inst 0xa060c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14]\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
+      ".inst 0xa061c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 24f\n"
       "14:"  // Store to output array
       "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "add x26, x26, x10\n"  // C += n
       "sub x25, x13, x11\n"
-      "ld1rw { z8.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
+      "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
       "ldr x24, [%x[args], %[offsetof_ldcb]]\n"
-      "ld1rw { z9.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
-      "ld1rw { z10.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "add x26, x26, x10\n"  // C += n
-      "ld1rw { z11.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
       "madd x26, x11, x24, x26\n"  // C += m * ldc
-      "ld1rw { z15.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
-      "ld1rw { z21.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
-      "ld1rw { z20.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
+      "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
+      "ld1rw { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z3.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z14.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
+      "ld1rw { z25.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
+      "ld1rw { z24.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
       "tbz x16, #2, 15f\n"
-      "ldr w22, [%x[args], %[offsetof_n_0]]\n"
-      "ldr x21, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "ldr w21, [%x[args], %[offsetof_n_0]]\n"
+      "add x21, x21, x10\n"
+      "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "add x20, x20, x21, LSL #2\n"
+      ".inst 0xa0404280  // ld1w { z0.s-z1.s }, p8/Z, [x20]\n"
       "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n"
-      "add x22, x22, x10\n"
-      "add x21, x21, x22, LSL #2\n"
-      "add x20, x20, x22, LSL #2\n"
-      ".inst 0xa04042a8  // ld1w { z8.s-z9.s }, p8/Z, [x21]\n"
-      ".inst 0xa040428a  // ld1w { z10.s-z11.s }, p8/Z, [x20]\n"
+      "add x20, x20, x21, LSL #2\n"
+      ".inst 0xa0404282  // ld1w { z2.s-z3.s }, p8/Z, [x20]\n"
       "15:"  // Store to output array: Load per-channel parameters: End
       "cntw x23\n"
       "whilelt p0.h, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 17f\n"
       "16:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xc1a8ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z8.s\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      ".inst 0xc1a0ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
+      ".inst 0xc1a1ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a9ac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z9.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1aaaa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z10.s\n"
-      ".inst 0xc1abaa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z11.s\n"
-      ".inst 0xc1afab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z15.s\n"
-      ".inst 0xc1afab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z15.s\n"
-      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cea0  // sclamp { z0.s-z3.s }, z21.s, z20.s\n"
-      "uzp1 z19.h, z28.h, z0.h\n"
-      "uzp1 z18.h, z29.h, z1.h\n"
-      "uzp1 z17.h, z30.h, z2.h\n"
-      "uzp1 z16.h, z31.h, z3.h\n"
-      "st1b { z19.h }, p0, [x26]\n"
+      ".inst 0xc1a2aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      ".inst 0xc1a3aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z3.s\n"
+      ".inst 0xc1aeab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
+      ".inst 0xc1aeab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z14.s\n"
+      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      ".inst 0xc1b8cf28  // sclamp { z8.s-z11.s }, z25.s, z24.s\n"
+      "uzp1 z16.h, z4.h, z8.h\n"
+      "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z18.h }, p0, [x26]\n"
+      "uzp1 z16.h, z5.h, z9.h\n"
+      "uzp1 z17.h, z6.h, z10.h\n"
+      "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "uzp1 z16.h, z7.h, z11.h\n"
       "st1b { z17.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1b { z16.h }, p0, [x26]\n"
@@ -332,59 +331,60 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
       "blt 16b\n"
       "17:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 18f\n"
-      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
-      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
-      ".inst 0xc1a8ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n"
+      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc1a0ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z0.s\n"
+      ".inst 0xc1a1ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a9ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z9.s\n"
-      ".inst 0xc1aaaa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n"
-      ".inst 0xc1abaa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z11.s\n"
-      ".inst 0xc1afab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z15.s\n"
-      ".inst 0xc1afab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z15.s\n"
-      ".inst 0xc1b4cea4  // sclamp { z4.s-z7.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      "uzp1 z16.h, z4.h, z28.h\n"
+      ".inst 0xc1a2aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n"
+      ".inst 0xc1a3aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n"
+      ".inst 0xc1aeab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z14.s\n"
+      ".inst 0xc1aeab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
+      ".inst 0xc1b8cf28  // sclamp { z8.s-z11.s }, z25.s, z24.s\n"
+      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      "uzp1 z16.h, z8.h, z4.h\n"
       "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 18f\n"
       "subs x20, x20, #0x1\n"
-      "uzp1 z16.h, z5.h, z29.h\n"
+      "uzp1 z16.h, z9.h, z5.h\n"
       "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 18f\n"
-      "uzp1 z16.h, z6.h, z30.h\n"
+      "uzp1 z16.h, z10.h, z6.h\n"
       "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "18:"  // Store to output array: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 22f\n"
+      "whilelt p0.h, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x23, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 20f\n"
       "19:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
-      ".inst 0xc1a8ac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z8.s\n"
+      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
+      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
+      ".inst 0xc1a0ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
+      ".inst 0xc1a1ac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z1.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a9ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z9.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1aaaa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z10.s\n"
-      ".inst 0xc1abaa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z11.s\n"
-      ".inst 0xc1afab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z15.s\n"
-      ".inst 0xc1afab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z15.s\n"
-      ".inst 0xc1b4cea0  // sclamp { z0.s-z3.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      "uzp1 z19.h, z0.h, z28.h\n"
-      "uzp1 z18.h, z1.h, z29.h\n"
-      "uzp1 z17.h, z2.h, z30.h\n"
-      "uzp1 z16.h, z3.h, z31.h\n"
-      "st1b { z19.h }, p0, [x26]\n"
+      ".inst 0xc1a2aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      ".inst 0xc1a3aa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z3.s\n"
+      ".inst 0xc1aeab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
+      ".inst 0xc1aeab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z14.s\n"
+      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      ".inst 0xc1b8cf34  // sclamp { z20.s-z23.s }, z25.s, z24.s\n"
+      "uzp1 z16.h, z4.h, z20.h\n"
+      "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z18.h }, p0, [x26]\n"
+      "uzp1 z16.h, z5.h, z21.h\n"
+      "uzp1 z17.h, z6.h, z22.h\n"
+      "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "uzp1 z16.h, z7.h, z23.h\n"
       "st1b { z17.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1b { z16.h }, p0, [x26]\n"
@@ -394,15 +394,15 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
       "cbz x20, 21f\n"
       ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
       ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xc1a8ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n"
+      ".inst 0xc1a0ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
+      ".inst 0xc1a1ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a9ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z9.s\n"
-      ".inst 0xc1aaaa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n"
-      ".inst 0xc1abaa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z11.s\n"
-      ".inst 0xc1afab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z15.s\n"
-      ".inst 0xc1afab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z15.s\n"
-      ".inst 0xc1b4cea4  // sclamp { z4.s-z7.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4ceb0  // sclamp { z16.s-z19.s }, z21.s, z20.s\n"
+      ".inst 0xc1a2aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      ".inst 0xc1a3aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z3.s\n"
+      ".inst 0xc1aeab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
+      ".inst 0xc1aeab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z14.s\n"
+      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
       "uzp1 z16.h, z4.h, z16.h\n"
       "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
@@ -420,25 +420,25 @@ void sme2_interleaved_nomerge_s8q_mopa_2VLx2VL(const int8_t *const A, const int8
       "mov x12, #0x0\n"
       "cntw x20\n"
       "23:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5fc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa041c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
       ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 23b\n"
       "24:"  // End block
       "incw x10, ALL, MUL #2\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #2\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL.hpp
index 51fc52f7b7..5e22847853 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
 class cls_sme2_interleaved_nomerge_s8q_mopa_4VLx1VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef int8_t result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, int8_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Requantize32 &rq, const int n_0, bool accumulate, int32_t *const accumulator_buffer);
@@ -61,7 +60,7 @@ public:
 
   static constexpr bool supports_accumulate()
   {
-    return true;
+    return false;
   }
 
   static constexpr bool supports_bias()
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8q_mopa_4VLx1VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 4, 1, 4, true> transforms = {};
+  StdTransformsSME<operand_type, result_type, 4, 1, 4, true> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp
index df9a866b6d..7b8d34d350 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL/generic.cpp
@@ -49,7 +49,7 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
         B(B), kstride_bytes(roundup(K, 4) * sizeof(int8_t)),
         C(C), ldcb(ldc * sizeof(int8_t)),
         M(M), N(N), K(K),
-        min(0), max(0),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias), n_0(n_0),
         accumulator_buffer(accumulator_buffer),
@@ -74,14 +74,13 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
     const long kstride_bytes;
     int8_t *const C;
     const long ldcb;
-    const long M, N, K;
-    int32_t min;
-    int32_t max;
+    const long M, N, K, n_loops, n_tail_iters;
+    int32_t min = std::numeric_limits<int8_t>::min();
+    int32_t max = std::numeric_limits<int8_t>::max();
 
     const int32_t *const bias;
     const int n_0;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -100,17 +99,17 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa040c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa041c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa042c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa043c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -125,95 +124,95 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "ld1w { z6.s }, p0/Z, [x20, x10, LSL #2]\n"
-      ".inst 0xc09024c0  // addha za0.s, p1/M, p1/M, z6.s\n"
-      ".inst 0xc09024c1  // addha za1.s, p1/M, p1/M, z6.s\n"
-      ".inst 0xc09024c2  // addha za2.s, p1/M, p1/M, z6.s\n"
-      ".inst 0xc09024c3  // addha za3.s, p1/M, p1/M, z6.s\n"
+      "ldnt1w { z8.s }, p0/Z, [x20, x10, LSL #2]\n"
+      ".inst 0xc0902500  // addha za0.s, p1/M, p1/M, z8.s\n"
+      ".inst 0xc0902501  // addha za1.s, p1/M, p1/M, z8.s\n"
+      ".inst 0xc0902502  // addha za2.s, p1/M, p1/M, z8.s\n"
+      ".inst 0xc0902503  // addha za3.s, p1/M, p1/M, z8.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20\n"
       "incw x21, ALL, MUL #4\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa1408360  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn8.b/Z, [x27]\n"
-      "ld1b { z29.b }, p1/Z, [x23]\n"
-      ".inst 0xa1418361  // ld1b { z1.b, z5.b, z9.b, z13.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "ld1b { z19.b }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa1428363  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1b { z20.b }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa0438378  // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x10, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa0408364  // ld1b { z4.b-z7.b }, pn8.b/Z, [x27]\n"
+      "ldnt1b { z14.b }, p1/Z, [x23]\n"
+      ".inst 0xa0418374  // ld1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      "ldnt1b { z31.b }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0xa0428378  // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1b { z13.b }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0xa0438368  // ld1b { z8.b-z11.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1b { z31.b }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1b { z29.b }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0xa09d2400  // smopa za0.s, p1/M, p1/M, z0.b, z29.b\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa09d2481  // smopa za1.s, p1/M, p1/M, z4.b, z29.b\n"
-      ".inst 0xa09d2502  // smopa za2.s, p1/M, p1/M, z8.b, z29.b\n"
-      ".inst 0xa09d2583  // smopa za3.s, p1/M, p1/M, z12.b, z29.b\n"
-      ".inst 0xa1408360  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa0932420  // smopa za0.s, p1/M, p1/M, z1.b, z19.b\n"
-      "ld1b { z29.b }, p1/Z, [x23]\n"
-      ".inst 0xa09324a1  // smopa za1.s, p1/M, p1/M, z5.b, z19.b\n"
-      ".inst 0xa0932522  // smopa za2.s, p1/M, p1/M, z9.b, z19.b\n"
-      ".inst 0xa09325a3  // smopa za3.s, p1/M, p1/M, z13.b, z19.b\n"
-      ".inst 0xa1418361  // ld1b { z1.b, z5.b, z9.b, z13.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0942460  // smopa za0.s, p1/M, p1/M, z3.b, z20.b\n"
-      "ld1b { z19.b }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa09424e1  // smopa za1.s, p1/M, p1/M, z7.b, z20.b\n"
-      ".inst 0xa0942562  // smopa za2.s, p1/M, p1/M, z11.b, z20.b\n"
-      ".inst 0xa09425e3  // smopa za3.s, p1/M, p1/M, z15.b, z20.b\n"
-      ".inst 0xa1428363  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1b { z20.b }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa09f2700  // smopa za0.s, p1/M, p1/M, z24.b, z31.b\n"
-      ".inst 0xa09f2721  // smopa za1.s, p1/M, p1/M, z25.b, z31.b\n"
-      ".inst 0xa09f2742  // smopa za2.s, p1/M, p1/M, z26.b, z31.b\n"
-      ".inst 0xa09f2763  // smopa za3.s, p1/M, p1/M, z27.b, z31.b\n"
-      ".inst 0xa0438378  // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      ".inst 0xa08e2480  // smopa za0.s, p1/M, p1/M, z4.b, z14.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa08e24a1  // smopa za1.s, p1/M, p1/M, z5.b, z14.b\n"
+      ".inst 0xa08e24c2  // smopa za2.s, p1/M, p1/M, z6.b, z14.b\n"
+      ".inst 0xa08e24e3  // smopa za3.s, p1/M, p1/M, z7.b, z14.b\n"
+      ".inst 0xa0408364  // ld1b { z4.b-z7.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa09f2680  // smopa za0.s, p1/M, p1/M, z20.b, z31.b\n"
+      "ldnt1b { z14.b }, p1/Z, [x23]\n"
+      ".inst 0xa09f26a1  // smopa za1.s, p1/M, p1/M, z21.b, z31.b\n"
+      ".inst 0xa09f26c2  // smopa za2.s, p1/M, p1/M, z22.b, z31.b\n"
+      ".inst 0xa09f26e3  // smopa za3.s, p1/M, p1/M, z23.b, z31.b\n"
+      ".inst 0xa0418374  // ld1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa08d2700  // smopa za0.s, p1/M, p1/M, z24.b, z13.b\n"
+      "ldnt1b { z31.b }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0xa08d2721  // smopa za1.s, p1/M, p1/M, z25.b, z13.b\n"
+      ".inst 0xa08d2742  // smopa za2.s, p1/M, p1/M, z26.b, z13.b\n"
+      ".inst 0xa08d2763  // smopa za3.s, p1/M, p1/M, z27.b, z13.b\n"
+      ".inst 0xa0428378  // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1b { z13.b }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0xa09d2500  // smopa za0.s, p1/M, p1/M, z8.b, z29.b\n"
+      ".inst 0xa09d2521  // smopa za1.s, p1/M, p1/M, z9.b, z29.b\n"
+      ".inst 0xa09d2542  // smopa za2.s, p1/M, p1/M, z10.b, z29.b\n"
+      ".inst 0xa09d2563  // smopa za3.s, p1/M, p1/M, z11.b, z29.b\n"
+      ".inst 0xa0438368  // ld1b { z8.b-z11.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1b { z31.b }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1b { z29.b }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0xa09d2400  // smopa za0.s, p1/M, p1/M, z0.b, z29.b\n"
-      ".inst 0xa09d2481  // smopa za1.s, p1/M, p1/M, z4.b, z29.b\n"
-      ".inst 0xa09d2502  // smopa za2.s, p1/M, p1/M, z8.b, z29.b\n"
-      ".inst 0xa09d2583  // smopa za3.s, p1/M, p1/M, z12.b, z29.b\n"
-      ".inst 0xa0932420  // smopa za0.s, p1/M, p1/M, z1.b, z19.b\n"
-      ".inst 0xa09324a1  // smopa za1.s, p1/M, p1/M, z5.b, z19.b\n"
-      ".inst 0xa0932522  // smopa za2.s, p1/M, p1/M, z9.b, z19.b\n"
-      ".inst 0xa09325a3  // smopa za3.s, p1/M, p1/M, z13.b, z19.b\n"
-      ".inst 0xa0942460  // smopa za0.s, p1/M, p1/M, z3.b, z20.b\n"
-      ".inst 0xa09424e1  // smopa za1.s, p1/M, p1/M, z7.b, z20.b\n"
-      ".inst 0xa0942562  // smopa za2.s, p1/M, p1/M, z11.b, z20.b\n"
-      ".inst 0xa09425e3  // smopa za3.s, p1/M, p1/M, z15.b, z20.b\n"
-      ".inst 0xa09f2700  // smopa za0.s, p1/M, p1/M, z24.b, z31.b\n"
-      ".inst 0xa09f2721  // smopa za1.s, p1/M, p1/M, z25.b, z31.b\n"
-      ".inst 0xa09f2742  // smopa za2.s, p1/M, p1/M, z26.b, z31.b\n"
-      ".inst 0xa09f2763  // smopa za3.s, p1/M, p1/M, z27.b, z31.b\n"
+      ".inst 0xa08e2480  // smopa za0.s, p1/M, p1/M, z4.b, z14.b\n"
+      ".inst 0xa08e24a1  // smopa za1.s, p1/M, p1/M, z5.b, z14.b\n"
+      ".inst 0xa08e24c2  // smopa za2.s, p1/M, p1/M, z6.b, z14.b\n"
+      ".inst 0xa08e24e3  // smopa za3.s, p1/M, p1/M, z7.b, z14.b\n"
+      ".inst 0xa09f2680  // smopa za0.s, p1/M, p1/M, z20.b, z31.b\n"
+      ".inst 0xa09f26a1  // smopa za1.s, p1/M, p1/M, z21.b, z31.b\n"
+      ".inst 0xa09f26c2  // smopa za2.s, p1/M, p1/M, z22.b, z31.b\n"
+      ".inst 0xa09f26e3  // smopa za3.s, p1/M, p1/M, z23.b, z31.b\n"
+      ".inst 0xa08d2700  // smopa za0.s, p1/M, p1/M, z24.b, z13.b\n"
+      ".inst 0xa08d2721  // smopa za1.s, p1/M, p1/M, z25.b, z13.b\n"
+      ".inst 0xa08d2742  // smopa za2.s, p1/M, p1/M, z26.b, z13.b\n"
+      ".inst 0xa08d2763  // smopa za3.s, p1/M, p1/M, z27.b, z13.b\n"
+      ".inst 0xa09d2500  // smopa za0.s, p1/M, p1/M, z8.b, z29.b\n"
+      ".inst 0xa09d2521  // smopa za1.s, p1/M, p1/M, z9.b, z29.b\n"
+      ".inst 0xa09d2542  // smopa za2.s, p1/M, p1/M, z10.b, z29.b\n"
+      ".inst 0xa09d2563  // smopa za3.s, p1/M, p1/M, z11.b, z29.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
       ".inst 0xa1408372  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      "subs x21, x21, #0x1\n"
       "addvl x27, x27, #4\n"
       "ld1b { z15.b }, p1/Z, [x23]\n"
       "addvl x23, x23, #1\n"
@@ -234,25 +233,25 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xa041c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
-      ".inst 0xa042c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xa060c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14]\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xa061c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa040c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xa041c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
+      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      ".inst 0xa042c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa043c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c1d8  // st1w { z24.s-z27.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa060c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 30f\n"
@@ -260,56 +259,56 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
+      ".inst 0xa060c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14]\n"
+      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
+      ".inst 0xa061c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c1d0  // st1w { z16.s-z19.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 30f\n"
       "14:"  // Store to output array
       "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "add x26, x26, x10\n"  // C += n
       "sub x25, x13, x11\n"
       "ld1rw { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
       "ldr x24, [%x[args], %[offsetof_ldcb]]\n"
+      "madd x26, x11, x24, x26\n"  // C += m * ldc
       "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
       "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
-      "add x26, x26, x10\n"  // C += n
-      "ld1rw { z25.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
-      "madd x26, x11, x24, x26\n"  // C += m * ldc
-      "ld1rw { z24.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
+      "ld1rw { z21.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
+      "ld1rw { z20.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
       "tbz x16, #2, 15f\n"
-      "ldr w22, [%x[args], %[offsetof_n_0]]\n"
-      "ldr x21, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "ldr w21, [%x[args], %[offsetof_n_0]]\n"
+      "add x21, x21, x10\n"
+      "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "add x20, x20, x21, LSL #2\n"
+      "ld1w { z2.s }, p0/Z, [x20]\n"
       "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n"
-      "add x22, x22, x10\n"
-      "add x21, x21, x22, LSL #2\n"
-      "add x20, x20, x22, LSL #2\n"
-      "ld1w { z2.s }, p0/Z, [x21]\n"
+      "add x20, x20, x21, LSL #2\n"
       "ld1w { z1.s }, p0/Z, [x20]\n"
       "15:"  // Store to output array: Load per-channel parameters: End
       "cntw x23\n"
       "whilelt p0.s, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 17f\n"
       "16:"  // Store to output array: Accumulator row 0 loop
       ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
+      ".inst 0xc1b4ceb0  // sclamp { z16.s-z19.s }, z21.s, z20.s\n"
       "st1b { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1b { z17.s }, p0, [x26]\n"
@@ -321,55 +320,56 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
       "blt 16b\n"
       "17:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 18f\n"
-      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a2ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
-      ".inst 0xc1a1aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
-      ".inst 0xc1a0ab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
-      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
-      "st1b { z4.s }, p0, [x26]\n"
+      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
+      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
+      ".inst 0xc1b4ceb0  // sclamp { z16.s-z19.s }, z21.s, z20.s\n"
+      "st1b { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 18f\n"
       "subs x20, x20, #0x1\n"
-      "st1b { z5.s }, p0, [x26]\n"
+      "st1b { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 18f\n"
-      "st1b { z6.s }, p0, [x26]\n"
+      "st1b { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "18:"  // Store to output array: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 28f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 20f\n"
       "19:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc1a2ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
+      ".inst 0xc1a1aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
-      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
-      "st1b { z16.s }, p0, [x26]\n"
+      ".inst 0xc1a0ab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
+      ".inst 0xc1b4cea4  // sclamp { z4.s-z7.s }, z21.s, z20.s\n"
+      "st1b { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z17.s }, p0, [x26]\n"
+      "st1b { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z18.s }, p0, [x26]\n"
+      "st1b { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z19.s }, p0, [x26]\n"
+      "st1b { z7.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 19b\n"
       "20:"  // Store to output array: Accumulator row 1 oddments
       "cbz x20, 21f\n"
       ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
-      "subs x20, x20, #0x1\n"
       ".inst 0xc1a2ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      "subs x20, x20, #0x1\n"
       ".inst 0xc1a1aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
       ".inst 0xc1a0ab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
-      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      ".inst 0xc1b4cea4  // sclamp { z4.s-z7.s }, z21.s, z20.s\n"
       "st1b { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 21f\n"
@@ -382,113 +382,115 @@ void sme2_interleaved_nomerge_s8q_mopa_4VLx1VL(const int8_t *const A, const int8
       "21:"  // Store to output array: Accumulator row 1 oddments: End
       "subs x25, x25, x22\n"
       "beq 28f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 23f\n"
       "22:"  // Store to output array: Accumulator row 2 loop
-      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
+      ".inst 0xc1a2ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a2ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
+      ".inst 0xc1a1aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1a1aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n"
-      ".inst 0xc1a0ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n"
-      ".inst 0xc1b8cf2c  // sclamp { z12.s-z15.s }, z25.s, z24.s\n"
-      "st1b { z12.s }, p0, [x26]\n"
+      ".inst 0xc1a0ab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z0.s\n"
+      ".inst 0xc1b4cea8  // sclamp { z8.s-z11.s }, z21.s, z20.s\n"
+      "st1b { z8.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z13.s }, p0, [x26]\n"
+      "st1b { z9.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z14.s }, p0, [x26]\n"
+      "st1b { z10.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z15.s }, p0, [x26]\n"
+      "st1b { z11.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 22b\n"
       "23:"  // Store to output array: Accumulator row 2 oddments
       "cbz x20, 24f\n"
-      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc1a2ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
-      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
-      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
-      "st1b { z16.s }, p0, [x26]\n"
+      ".inst 0xc1a1aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n"
+      ".inst 0xc1a0ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n"
+      ".inst 0xc1b4ceac  // sclamp { z12.s-z15.s }, z21.s, z20.s\n"
+      "st1b { z12.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 24f\n"
       "subs x20, x20, #0x1\n"
-      "st1b { z17.s }, p0, [x26]\n"
+      "st1b { z13.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 24f\n"
-      "st1b { z18.s }, p0, [x26]\n"
+      "st1b { z14.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "24:"  // Store to output array: Accumulator row 2 oddments: End
       "subs x25, x25, x22\n"
       "beq 28f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x23, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 26f\n"
       "25:"  // Store to output array: Accumulator row 3 loop
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xc1a2ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
+      ".inst 0xc1a1aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
-      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
-      "st1b { z16.s }, p0, [x26]\n"
+      ".inst 0xc1a0ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n"
+      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
+      "st1b { z28.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z17.s }, p0, [x26]\n"
+      "st1b { z29.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z18.s }, p0, [x26]\n"
+      "st1b { z30.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z19.s }, p0, [x26]\n"
+      "st1b { z31.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 25b\n"
       "26:"  // Store to output array: Accumulator row 3 oddments
       "cbz x20, 27f\n"
-      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xc1a2ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a2ac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z2.s\n"
-      ".inst 0xc1a1aa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z1.s\n"
-      ".inst 0xc1a0ab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z0.s\n"
-      ".inst 0xc1b8cf34  // sclamp { z20.s-z23.s }, z25.s, z24.s\n"
-      "st1b { z20.s }, p0, [x26]\n"
+      ".inst 0xc1a1aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n"
+      ".inst 0xc1a0ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n"
+      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
+      "st1b { z28.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 27f\n"
       "subs x20, x20, #0x1\n"
-      "st1b { z21.s }, p0, [x26]\n"
+      "st1b { z29.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 27f\n"
-      "st1b { z22.s }, p0, [x26]\n"
+      "st1b { z30.s }, p0, [x26]\n"
       "27:"  // Store to output array: Accumulator row 3 oddments: End
       "28:"  // Store to output array: End
       "tbz x16, #0, 30f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "29:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa041c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
       ".inst 0xc0840403  // mova za3h.s[x12], { z0.s-z3.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 29b\n"
       "30:"  // End block
       "incw x10\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #4\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL.hpp
index fe8f5383bd..7792192856 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL(const int8_t *const A, const
 class cls_sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, float *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const DequantizeFloat &dq, const float *const late_bias, const Activation act, bool accumulate, int32_t *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 1, 4, 4> transforms = {};
+  StdTransformsSME<operand_type, result_type, 1, 4, 4> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL/generic.cpp
index bf60c61fc0..4b26a6578c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL/generic.cpp
@@ -40,8 +40,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL(const int8_t *const A, const
       const int8_t *const B,
       float *const C, const int ldc,
       const int M, const int N, const int K,
-      const int32_t *const bias,
-      const DequantizeFloat &, const float *const late_bias, const Activation act,
+      const int32_t *const bias, const float *const late_bias, const Activation act,
       bool accumulate,
       int32_t *const accumulator_buffer
     ) : A(A),
@@ -95,7 +94,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_1VLx4VL(const int8_t *const A, const
   };
 
   // Construct arguments for this kernel
-  KernelArgs args(A, B, C, ldc, M, N, K, bias, dq, late_bias, act, accumulate, accumulator_buffer);
+  KernelArgs args(A, B, C, ldc, M, N, K, bias, late_bias, act, accumulate, accumulator_buffer);
 
   __asm__ __volatile__(
       "ldr x13, [%x[args], %[offsetof_flags]]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL.hpp
index edb6737dc7..df2c9c0ca3 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL(const int8_t *const A, const
 class cls_sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, float *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const DequantizeFloat &dq, const float *const late_bias, const Activation act, bool accumulate, int32_t *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 2, 2, 4> transforms = {};
+  StdTransformsSME<operand_type, result_type, 2, 2, 4> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL/generic.cpp
index ff649f11f6..1631fae8e9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL/generic.cpp
@@ -40,8 +40,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL(const int8_t *const A, const
       const int8_t *const B,
       float *const C, const int ldc,
       const int M, const int N, const int K,
-      const int32_t *const bias,
-      const DequantizeFloat &, const float *const late_bias, const Activation act,
+      const int32_t *const bias, const float *const late_bias, const Activation act,
       bool accumulate,
       int32_t *const accumulator_buffer
     ) : A(A),
@@ -95,7 +94,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_2VLx2VL(const int8_t *const A, const
   };
 
   // Construct arguments for this kernel
-  KernelArgs args(A, B, C, ldc, M, N, K, bias, dq, late_bias, act, accumulate, accumulator_buffer);
+  KernelArgs args(A, B, C, ldc, M, N, K, bias, late_bias, act, accumulate, accumulator_buffer);
 
   __asm__ __volatile__(
       "ldr x16, [%x[args], %[offsetof_flags]]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL.hpp
index 112f5ef0e8..70952f4f03 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL(const int8_t *const A, const
 class cls_sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef float result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, float *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const DequantizeFloat &dq, const float *const late_bias, const Activation act, bool accumulate, int32_t *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 4, 1, 4> transforms = {};
+  StdTransformsSME<operand_type, result_type, 4, 1, 4> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL/generic.cpp
index a08ea8311a..bafb16bca8 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL/generic.cpp
@@ -40,8 +40,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL(const int8_t *const A, const
       const int8_t *const B,
       float *const C, const int ldc,
       const int M, const int N, const int K,
-      const int32_t *const bias,
-      const DequantizeFloat &, const float *const late_bias, const Activation act,
+      const int32_t *const bias, const float *const late_bias, const Activation act,
       bool accumulate,
       int32_t *const accumulator_buffer
     ) : A(A),
@@ -95,7 +94,7 @@ void sme2_interleaved_nomerge_s8qfp32_mopa_4VLx1VL(const int8_t *const A, const
   };
 
   // Construct arguments for this kernel
-  KernelArgs args(A, B, C, ldc, M, N, K, bias, dq, late_bias, act, accumulate, accumulator_buffer);
+  KernelArgs args(A, B, C, ldc, M, N, K, bias, late_bias, act, accumulate, accumulator_buffer);
 
   __asm__ __volatile__(
       "ldr x16, [%x[args], %[offsetof_flags]]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL.hpp
index 0c8de041cb..84386009a0 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
 class cls_sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef int32_t result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, int32_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Activation, bool accumulate, int32_t *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 1, 4, 4> transforms = {};
+  StdTransformsSME<operand_type, result_type, 1, 4, 4> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp
index a643fb265b..67c759410a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
         B(B), kstride_bytes(roundup(K, 4) * sizeof(int8_t)),
         C(C), ldcb(ldc * sizeof(int32_t)),
         M(M), N(N), K(K),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias),
         accumulator_buffer(accumulator_buffer),
@@ -68,11 +69,10 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
     const long kstride_bytes;
     int32_t *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
 
     const int32_t *const bias;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -92,16 +92,16 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
       ".inst 0xa040c564  // ld1w { z4.s-z7.s }, pn9.b/Z, [x11]\n"
-      ".inst 0xa041c57c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
-      ".inst 0xa042c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
-      ".inst 0xa043c574  // ld1w { z20.s-z23.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
       ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
-      "addvl x11, x11, #16\n"
-      ".inst 0xc0840781  // mova za1h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa041c56c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
+      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa042c57c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
+      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa043c574  // ld1w { z20.s-z23.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
       ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x11, x11, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w9, [%x[args], %[offsetof_M]]\n"
@@ -116,102 +116,102 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      ".inst 0xa01bc288  // ld1w { z8.s-z11.s }, p8/Z, [x20, x27, LSL #2]\n"
-      ".inst 0xc0900100  // addha za0.s, p0/M, p0/M, z8.s\n"
-      ".inst 0xc0900121  // addha za1.s, p0/M, p0/M, z9.s\n"
-      ".inst 0xc0900142  // addha za2.s, p0/M, p0/M, z10.s\n"
-      ".inst 0xc0900163  // addha za3.s, p0/M, p0/M, z11.s\n"
+      ".inst 0xa11bc29b  // ldnt1w { z19.s, z23.s, z27.s, z31.s }, p8/Z, [x20, x27, LSL #2]\n"
+      ".inst 0xc0900260  // addha za0.s, p0/M, p0/M, z19.s\n"
+      ".inst 0xc09002e1  // addha za1.s, p0/M, p0/M, z23.s\n"
+      ".inst 0xc0900362  // addha za2.s, p0/M, p0/M, z27.s\n"
+      ".inst 0xc09003e3  // addha za3.s, p0/M, p0/M, z31.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x27\n"
       "mov x21, x28\n"
       "incw x20, ALL, MUL #4\n"
       "incw x21\n"
       "cmp x20, x26\n"
-      "mov x20, x13\n"
       "csel x21, x28, x21, LT\n"
+      "mov x20, x13\n"
       "bfm x13, XZR, #0x0, #0x0  // bfc x13, #0x0, #0x1\n"
       "cmp x21, x9\n"
       "csel x13, x20, x13, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x27, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      "ld1b { z21.b }, p0/Z, [x24]\n"
-      ".inst 0xa04086f8  // ld1b { z24.b-z27.b }, pn9.b/Z, [x23]\n"
-      "ld1b { z6.b }, p0/Z, [x24, #1, MUL VL]\n"
-      ".inst 0xa14186e1  // ld1b { z1.b, z5.b, z9.b, z13.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      "ld1b { z31.b }, p0/Z, [x24, #2, MUL VL]\n"
-      ".inst 0xa14286e3  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      "ld1b { z23.b }, p0/Z, [x24, #3, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x27, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      "ld1b { z30.b }, p0/Z, [x24]\n"
+      ".inst 0xa04086e1  // ldnt1b { z0.b-z3.b }, pn9.b/Z, [x23]\n"
+      "ld1b { z21.b }, p0/Z, [x24, #1, MUL VL]\n"
+      ".inst 0xa04186f9  // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      "ld1b { z28.b }, p0/Z, [x24, #2, MUL VL]\n"
+      ".inst 0xa04286e5  // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      "ld1b { z11.b }, p0/Z, [x24, #3, MUL VL]\n"
       "addvl x24, x24, #4\n"
-      ".inst 0xa14386e0  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      ".inst 0xa04386f1  // ldnt1b { z16.b-z19.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "ble 7f\n"
       "6:"  // K loop
+      ".inst 0xa08003c0  // smopa za0.s, p0/M, p0/M, z30.b, z0.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa08103c1  // smopa za1.s, p0/M, p0/M, z30.b, z1.b\n"
+      ".inst 0xa08203c2  // smopa za2.s, p0/M, p0/M, z30.b, z2.b\n"
+      ".inst 0xa08303c3  // smopa za3.s, p0/M, p0/M, z30.b, z3.b\n"
+      "ld1b { z30.b }, p0/Z, [x24]\n"
       ".inst 0xa09802a0  // smopa za0.s, p0/M, p0/M, z21.b, z24.b\n"
-      "subs x21, x21, #0x1\n"
+      ".inst 0xa04086e1  // ldnt1b { z0.b-z3.b }, pn9.b/Z, [x23]\n"
       ".inst 0xa09902a1  // smopa za1.s, p0/M, p0/M, z21.b, z25.b\n"
       ".inst 0xa09a02a2  // smopa za2.s, p0/M, p0/M, z21.b, z26.b\n"
       ".inst 0xa09b02a3  // smopa za3.s, p0/M, p0/M, z21.b, z27.b\n"
-      "ld1b { z21.b }, p0/Z, [x24]\n"
-      ".inst 0xa08100c0  // smopa za0.s, p0/M, p0/M, z6.b, z1.b\n"
-      ".inst 0xa04086f8  // ld1b { z24.b-z27.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa08500c1  // smopa za1.s, p0/M, p0/M, z6.b, z5.b\n"
-      ".inst 0xa08900c2  // smopa za2.s, p0/M, p0/M, z6.b, z9.b\n"
-      ".inst 0xa08d00c3  // smopa za3.s, p0/M, p0/M, z6.b, z13.b\n"
-      "ld1b { z6.b }, p0/Z, [x24, #1, MUL VL]\n"
-      ".inst 0xa08303e0  // smopa za0.s, p0/M, p0/M, z31.b, z3.b\n"
-      ".inst 0xa14186e1  // ld1b { z1.b, z5.b, z9.b, z13.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa08703e1  // smopa za1.s, p0/M, p0/M, z31.b, z7.b\n"
-      ".inst 0xa08b03e2  // smopa za2.s, p0/M, p0/M, z31.b, z11.b\n"
-      ".inst 0xa08f03e3  // smopa za3.s, p0/M, p0/M, z31.b, z15.b\n"
-      "ld1b { z31.b }, p0/Z, [x24, #2, MUL VL]\n"
-      ".inst 0xa14286e3  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      ".inst 0xa08002e0  // smopa za0.s, p0/M, p0/M, z23.b, z0.b\n"
-      ".inst 0xa08402e1  // smopa za1.s, p0/M, p0/M, z23.b, z4.b\n"
-      ".inst 0xa08802e2  // smopa za2.s, p0/M, p0/M, z23.b, z8.b\n"
-      ".inst 0xa08c02e3  // smopa za3.s, p0/M, p0/M, z23.b, z12.b\n"
-      "ld1b { z23.b }, p0/Z, [x24, #3, MUL VL]\n"
+      "ld1b { z21.b }, p0/Z, [x24, #1, MUL VL]\n"
+      ".inst 0xa0840380  // smopa za0.s, p0/M, p0/M, z28.b, z4.b\n"
+      ".inst 0xa04186f9  // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa0850381  // smopa za1.s, p0/M, p0/M, z28.b, z5.b\n"
+      ".inst 0xa0860382  // smopa za2.s, p0/M, p0/M, z28.b, z6.b\n"
+      ".inst 0xa0870383  // smopa za3.s, p0/M, p0/M, z28.b, z7.b\n"
+      "ld1b { z28.b }, p0/Z, [x24, #2, MUL VL]\n"
+      ".inst 0xa04286e5  // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      ".inst 0xa0900160  // smopa za0.s, p0/M, p0/M, z11.b, z16.b\n"
+      ".inst 0xa0910161  // smopa za1.s, p0/M, p0/M, z11.b, z17.b\n"
+      ".inst 0xa0920162  // smopa za2.s, p0/M, p0/M, z11.b, z18.b\n"
+      ".inst 0xa0930163  // smopa za3.s, p0/M, p0/M, z11.b, z19.b\n"
+      "ld1b { z11.b }, p0/Z, [x24, #3, MUL VL]\n"
       "addvl x24, x24, #4\n"
-      ".inst 0xa14386e0  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      ".inst 0xa04386f1  // ldnt1b { z16.b-z19.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "bgt 6b\n"
       "7:"  // K loop tail
+      ".inst 0xa08003c0  // smopa za0.s, p0/M, p0/M, z30.b, z0.b\n"
+      ".inst 0xa08103c1  // smopa za1.s, p0/M, p0/M, z30.b, z1.b\n"
+      ".inst 0xa08203c2  // smopa za2.s, p0/M, p0/M, z30.b, z2.b\n"
+      ".inst 0xa08303c3  // smopa za3.s, p0/M, p0/M, z30.b, z3.b\n"
       ".inst 0xa09802a0  // smopa za0.s, p0/M, p0/M, z21.b, z24.b\n"
       ".inst 0xa09902a1  // smopa za1.s, p0/M, p0/M, z21.b, z25.b\n"
       ".inst 0xa09a02a2  // smopa za2.s, p0/M, p0/M, z21.b, z26.b\n"
       ".inst 0xa09b02a3  // smopa za3.s, p0/M, p0/M, z21.b, z27.b\n"
-      ".inst 0xa08100c0  // smopa za0.s, p0/M, p0/M, z6.b, z1.b\n"
-      ".inst 0xa08500c1  // smopa za1.s, p0/M, p0/M, z6.b, z5.b\n"
-      ".inst 0xa08900c2  // smopa za2.s, p0/M, p0/M, z6.b, z9.b\n"
-      ".inst 0xa08d00c3  // smopa za3.s, p0/M, p0/M, z6.b, z13.b\n"
-      ".inst 0xa08303e0  // smopa za0.s, p0/M, p0/M, z31.b, z3.b\n"
-      ".inst 0xa08703e1  // smopa za1.s, p0/M, p0/M, z31.b, z7.b\n"
-      ".inst 0xa08b03e2  // smopa za2.s, p0/M, p0/M, z31.b, z11.b\n"
-      ".inst 0xa08f03e3  // smopa za3.s, p0/M, p0/M, z31.b, z15.b\n"
-      ".inst 0xa08002e0  // smopa za0.s, p0/M, p0/M, z23.b, z0.b\n"
-      ".inst 0xa08402e1  // smopa za1.s, p0/M, p0/M, z23.b, z4.b\n"
-      ".inst 0xa08802e2  // smopa za2.s, p0/M, p0/M, z23.b, z8.b\n"
-      ".inst 0xa08c02e3  // smopa za3.s, p0/M, p0/M, z23.b, z12.b\n"
+      ".inst 0xa0840380  // smopa za0.s, p0/M, p0/M, z28.b, z4.b\n"
+      ".inst 0xa0850381  // smopa za1.s, p0/M, p0/M, z28.b, z5.b\n"
+      ".inst 0xa0860382  // smopa za2.s, p0/M, p0/M, z28.b, z6.b\n"
+      ".inst 0xa0870383  // smopa za3.s, p0/M, p0/M, z28.b, z7.b\n"
+      ".inst 0xa0900160  // smopa za0.s, p0/M, p0/M, z11.b, z16.b\n"
+      ".inst 0xa0910161  // smopa za1.s, p0/M, p0/M, z11.b, z17.b\n"
+      ".inst 0xa0920162  // smopa za2.s, p0/M, p0/M, z11.b, z18.b\n"
+      ".inst 0xa0930163  // smopa za3.s, p0/M, p0/M, z11.b, z19.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      "ld1b { z14.b }, p0/Z, [x24]\n"
-      "subs x20, x20, #0x1\n"
+      "ld1b { z22.b }, p0/Z, [x24]\n"
+      "subs x21, x21, #0x1\n"
       "addvl x24, x24, #1\n"
-      ".inst 0xa14086e1  // ld1b { z1.b, z5.b, z9.b, z13.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa14086f1  // ld1b { z17.b, z21.b, z25.b, z29.b }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #4\n"
-      ".inst 0xa08101c0  // smopa za0.s, p0/M, p0/M, z14.b, z1.b\n"
-      ".inst 0xa08501c1  // smopa za1.s, p0/M, p0/M, z14.b, z5.b\n"
-      ".inst 0xa08901c2  // smopa za2.s, p0/M, p0/M, z14.b, z9.b\n"
-      ".inst 0xa08d01c3  // smopa za3.s, p0/M, p0/M, z14.b, z13.b\n"
+      ".inst 0xa09102c0  // smopa za0.s, p0/M, p0/M, z22.b, z17.b\n"
+      ".inst 0xa09502c1  // smopa za1.s, p0/M, p0/M, z22.b, z21.b\n"
+      ".inst 0xa09902c2  // smopa za2.s, p0/M, p0/M, z22.b, z25.b\n"
+      ".inst 0xa09d02c3  // smopa za3.s, p0/M, p0/M, z22.b, z29.b\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x13, #1, 14f\n"
@@ -219,25 +219,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c578  // ld1w { z24.s-z27.s }, pn9.b/Z, [x11]\n"
-      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
-      ".inst 0xa041c56c  // ld1w { z12.s-z15.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
+      ".inst 0xa040c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
+      ".inst 0xa041c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xc086045c  // mova { z28.s-z31.s }, za2h.s[x12]\n"
-      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
-      ".inst 0xa042c574  // ld1w { z20.s-z23.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
-      ".inst 0xa043c564  // ld1w { z4.s-z7.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x11, x11, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xa060c548  // st1w { z8.s-z11.s }, pn9.b, [x10]\n"
-      ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
-      ".inst 0xa061c550  // st1w { z16.s-z19.s }, pn9.b, [x10, #0x4, MUL VL]\n"
-      ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
+      ".inst 0xa042c560  // ld1w { z0.s-z3.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
+      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa043c574  // ld1w { z20.s-z23.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c55c  // st1w { z28.s-z31.s }, pn9.b, [x10, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c540  // st1w { z0.s-z3.s }, pn9.b, [x10, #0xc, MUL VL]\n"
+      ".inst 0xa060c544  // st1w { z4.s-z7.s }, pn9.b, [x10]\n"
+      "addvl x11, x11, #16\n"
+      ".inst 0xa061c54c  // st1w { z12.s-z15.s }, pn9.b, [x10, #0x4, MUL VL]\n"
+      ".inst 0xa062c55c  // st1w { z28.s-z31.s }, pn9.b, [x10, #0x8, MUL VL]\n"
+      ".inst 0xa063c550  // st1w { z16.s-z19.s }, pn9.b, [x10, #0xc, MUL VL]\n"
       "addvl x10, x10, #16\n"
       "blt 11b\n"
       "b 20f\n"
@@ -245,16 +245,16 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
-      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
-      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c544  // st1w { z4.s-z7.s }, pn9.b, [x10]\n"
+      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa060c54c  // st1w { z12.s-z15.s }, pn9.b, [x10]\n"
+      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
+      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
+      ".inst 0xa061c544  // st1w { z4.s-z7.s }, pn9.b, [x10, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c548  // st1w { z8.s-z11.s }, pn9.b, [x10, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c558  // st1w { z24.s-z27.s }, pn9.b, [x10, #0x8, MUL VL]\n"
-      ".inst 0xa063c550  // st1w { z16.s-z19.s }, pn9.b, [x10, #0xc, MUL VL]\n"
+      ".inst 0xa062c540  // st1w { z0.s-z3.s }, pn9.b, [x10, #0x8, MUL VL]\n"
+      ".inst 0xa063c558  // st1w { z24.s-z27.s }, pn9.b, [x10, #0xc, MUL VL]\n"
       "addvl x10, x10, #16\n"
       "blt 13b\n"
       "b 20f\n"
@@ -264,11 +264,11 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
       "cntw x20\n"
       "ldr x22, [%x[args], %[offsetof_ldcb]]\n"
       "cmp x21, x20\n"
-      "mov x12, #0x0\n"
       "csel x20, x21, x20, LT\n"
       "add x23, x23, x27, LSL #2\n"  // C += n
       "lsr x21, x20, #0x2\n"
       "madd x23, x28, x22, x23\n"  // C += m * ldc
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Accumulator row 0 loop
@@ -278,55 +278,55 @@ void sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL(const int8_t *const A, const in
       ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
       ".inst 0xa160c2e0  // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x23]\n"
       "add x23, x23, x22\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xa160c2e1  // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x23]\n"
       "add x23, x23, x22\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa160c2e2  // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x23]\n"
       "add x23, x23, x22\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa160c2e3  // st1w { z3.s, z7.s, z11.s, z15.s }, p8, [x23]\n"
       "add x23, x23, x22\n"
       "blt 15b\n"
       "16:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
-      ".inst 0xa160c2e0  // st1w { z0.s, z4.s, z8.s, z12.s }, p8, [x23]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xa160c2f0  // st1w { z16.s, z20.s, z24.s, z28.s }, p8, [x23]\n"
       "add x23, x23, x22\n"
       "beq 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xa160c2e1  // st1w { z1.s, z5.s, z9.s, z13.s }, p8, [x23]\n"
+      ".inst 0xa160c2f1  // st1w { z17.s, z21.s, z25.s, z29.s }, p8, [x23]\n"
       "add x23, x23, x22\n"
       "beq 17f\n"
-      ".inst 0xa160c2e2  // st1w { z2.s, z6.s, z10.s, z14.s }, p8, [x23]\n"
+      ".inst 0xa160c2f2  // st1w { z18.s, z22.s, z26.s, z30.s }, p8, [x23]\n"
       "17:"  // Store to output array: Accumulator row 0 oddments: End
       "18:"  // Store to output array: End
       "tbz x13, #0, 20f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "19:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c560  // ld1w { z0.s-z3.s }, pn9.b/Z, [x11]\n"
-      ".inst 0xa041c57c  // ld1w { z28.s-z31.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
-      ".inst 0xa042c564  // ld1w { z4.s-z7.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
+      ".inst 0xa040c568  // ld1w { z8.s-z11.s }, pn9.b/Z, [x11]\n"
+      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa041c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c570  // ld1w { z16.s-z19.s }, pn9.b/Z, [x11, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c568  // ld1w { z8.s-z11.s }, pn9.b/Z, [x11, #0xc, MUL VL]\n"
-      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
-      "addvl x11, x11, #16\n"
-      ".inst 0xc0840781  // mova za1h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
       ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x11, x11, #16\n"
       "blt 19b\n"
       "20:"  // End block
       "incw x27, ALL, MUL #4\n"
       "cmp x27, x26\n"
       "blt 3b\n"
       "incw x28\n"
-      "mov x27, #0x0\n"
       "cmp x28, x9\n"
+      "mov x27, #0x0\n"
       "mov x25, x24\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL.hpp
index 074a8819f9..2899d7553c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
 class cls_sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef int32_t result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, int32_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Activation, bool accumulate, int32_t *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 2, 2, 4> transforms = {};
+  StdTransformsSME<operand_type, result_type, 2, 2, 4> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp
index ae14cd7d50..7f44e5ffe5 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
         B(B), kstride_bytes(roundup(K, 4) * sizeof(int8_t)),
         C(C), ldcb(ldc * sizeof(int32_t)),
         M(M), N(N), K(K),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias),
         accumulator_buffer(accumulator_buffer),
@@ -68,11 +69,10 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
     const long kstride_bytes;
     int32_t *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
 
     const int32_t *const bias;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -91,17 +91,17 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5fc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840781  // mova za1h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
-      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa040c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa041c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa042c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa043c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840403  // mova za3h.s[x12], { z0.s-z3.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -116,102 +116,102 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      ".inst 0xa10a4296  // ld1w { z22.s, z30.s }, p8/Z, [x20, x10, LSL #2]\n"
-      ".inst 0xc09002c0  // addha za0.s, p0/M, p0/M, z22.s\n"
-      ".inst 0xc09003c1  // addha za1.s, p0/M, p0/M, z30.s\n"
-      ".inst 0xc09002c2  // addha za2.s, p0/M, p0/M, z22.s\n"
-      ".inst 0xc09003c3  // addha za3.s, p0/M, p0/M, z30.s\n"
+      ".inst 0xa00a4295  // ldnt1w { z20.s-z21.s }, p8/Z, [x20, x10, LSL #2]\n"
+      ".inst 0xc0900280  // addha za0.s, p0/M, p0/M, z20.s\n"
+      ".inst 0xc09002a1  // addha za1.s, p0/M, p0/M, z21.s\n"
+      ".inst 0xc0900282  // addha za2.s, p0/M, p0/M, z20.s\n"
+      ".inst 0xc09002a3  // addha za3.s, p0/M, p0/M, z21.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20, ALL, MUL #2\n"
       "incw x21, ALL, MUL #2\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa1400766  // ld1b { z6.b, z14.b }, pn9.b/Z, [x27]\n"
-      ".inst 0xa04006e2  // ld1b { z2.b-z3.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa041077a  // ld1b { z26.b-z27.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0xa04106f6  // ld1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa1420765  // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14206f5  // ld1b { z21.b, z29.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa0430760  // ld1b { z0.b-z1.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x10, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa040077c  // ld1b { z28.b-z29.b }, pn9.b/Z, [x27]\n"
+      ".inst 0xa14006e8  // ldnt1b { z0.b, z8.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa0410762  // ld1b { z2.b-z3.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa14106ff  // ldnt1b { z23.b, z31.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0xa042076e  // ld1b { z14.b-z15.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa14206f8  // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa0430764  // ld1b { z4.b-z5.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa14306f1  // ld1b { z17.b, z25.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa04306f5  // ldnt1b { z20.b-z21.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0xa08200c0  // smopa za0.s, p0/M, p0/M, z6.b, z2.b\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa08300c1  // smopa za1.s, p0/M, p0/M, z6.b, z3.b\n"
-      ".inst 0xa08201c2  // smopa za2.s, p0/M, p0/M, z14.b, z2.b\n"
-      ".inst 0xa08301c3  // smopa za3.s, p0/M, p0/M, z14.b, z3.b\n"
-      ".inst 0xa1400766  // ld1b { z6.b, z14.b }, pn9.b/Z, [x27]\n"
-      ".inst 0xa0960340  // smopa za0.s, p0/M, p0/M, z26.b, z22.b\n"
-      ".inst 0xa04006e2  // ld1b { z2.b-z3.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa0970341  // smopa za1.s, p0/M, p0/M, z26.b, z23.b\n"
-      ".inst 0xa0960362  // smopa za2.s, p0/M, p0/M, z27.b, z22.b\n"
-      ".inst 0xa0970363  // smopa za3.s, p0/M, p0/M, z27.b, z23.b\n"
-      ".inst 0xa041077a  // ld1b { z26.b-z27.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0xa09500a0  // smopa za0.s, p0/M, p0/M, z5.b, z21.b\n"
-      ".inst 0xa04106f6  // ld1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa09d00a1  // smopa za1.s, p0/M, p0/M, z5.b, z29.b\n"
-      ".inst 0xa09501a2  // smopa za2.s, p0/M, p0/M, z13.b, z21.b\n"
-      ".inst 0xa09d01a3  // smopa za3.s, p0/M, p0/M, z13.b, z29.b\n"
-      ".inst 0xa1420765  // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14206f5  // ld1b { z21.b, z29.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa0910000  // smopa za0.s, p0/M, p0/M, z0.b, z17.b\n"
-      ".inst 0xa0990001  // smopa za1.s, p0/M, p0/M, z0.b, z25.b\n"
-      ".inst 0xa0910022  // smopa za2.s, p0/M, p0/M, z1.b, z17.b\n"
-      ".inst 0xa0990023  // smopa za3.s, p0/M, p0/M, z1.b, z25.b\n"
-      ".inst 0xa0430760  // ld1b { z0.b-z1.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      ".inst 0xa0800380  // smopa za0.s, p0/M, p0/M, z28.b, z0.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa0880381  // smopa za1.s, p0/M, p0/M, z28.b, z8.b\n"
+      ".inst 0xa08003a2  // smopa za2.s, p0/M, p0/M, z29.b, z0.b\n"
+      ".inst 0xa08803a3  // smopa za3.s, p0/M, p0/M, z29.b, z8.b\n"
+      ".inst 0xa040077c  // ld1b { z28.b-z29.b }, pn9.b/Z, [x27]\n"
+      ".inst 0xa0970040  // smopa za0.s, p0/M, p0/M, z2.b, z23.b\n"
+      ".inst 0xa14006e8  // ldnt1b { z0.b, z8.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa09f0041  // smopa za1.s, p0/M, p0/M, z2.b, z31.b\n"
+      ".inst 0xa0970062  // smopa za2.s, p0/M, p0/M, z3.b, z23.b\n"
+      ".inst 0xa09f0063  // smopa za3.s, p0/M, p0/M, z3.b, z31.b\n"
+      ".inst 0xa0410762  // ld1b { z2.b-z3.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa09001c0  // smopa za0.s, p0/M, p0/M, z14.b, z16.b\n"
+      ".inst 0xa14106ff  // ldnt1b { z23.b, z31.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0xa09801c1  // smopa za1.s, p0/M, p0/M, z14.b, z24.b\n"
+      ".inst 0xa09001e2  // smopa za2.s, p0/M, p0/M, z15.b, z16.b\n"
+      ".inst 0xa09801e3  // smopa za3.s, p0/M, p0/M, z15.b, z24.b\n"
+      ".inst 0xa042076e  // ld1b { z14.b-z15.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa14206f8  // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa0940080  // smopa za0.s, p0/M, p0/M, z4.b, z20.b\n"
+      ".inst 0xa0950081  // smopa za1.s, p0/M, p0/M, z4.b, z21.b\n"
+      ".inst 0xa09400a2  // smopa za2.s, p0/M, p0/M, z5.b, z20.b\n"
+      ".inst 0xa09500a3  // smopa za3.s, p0/M, p0/M, z5.b, z21.b\n"
+      ".inst 0xa0430764  // ld1b { z4.b-z5.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa14306f1  // ld1b { z17.b, z25.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa04306f5  // ldnt1b { z20.b-z21.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0xa08200c0  // smopa za0.s, p0/M, p0/M, z6.b, z2.b\n"
-      ".inst 0xa08300c1  // smopa za1.s, p0/M, p0/M, z6.b, z3.b\n"
-      ".inst 0xa08201c2  // smopa za2.s, p0/M, p0/M, z14.b, z2.b\n"
-      ".inst 0xa08301c3  // smopa za3.s, p0/M, p0/M, z14.b, z3.b\n"
-      ".inst 0xa0960340  // smopa za0.s, p0/M, p0/M, z26.b, z22.b\n"
-      ".inst 0xa0970341  // smopa za1.s, p0/M, p0/M, z26.b, z23.b\n"
-      ".inst 0xa0960362  // smopa za2.s, p0/M, p0/M, z27.b, z22.b\n"
-      ".inst 0xa0970363  // smopa za3.s, p0/M, p0/M, z27.b, z23.b\n"
-      ".inst 0xa09500a0  // smopa za0.s, p0/M, p0/M, z5.b, z21.b\n"
-      ".inst 0xa09d00a1  // smopa za1.s, p0/M, p0/M, z5.b, z29.b\n"
-      ".inst 0xa09501a2  // smopa za2.s, p0/M, p0/M, z13.b, z21.b\n"
-      ".inst 0xa09d01a3  // smopa za3.s, p0/M, p0/M, z13.b, z29.b\n"
-      ".inst 0xa0910000  // smopa za0.s, p0/M, p0/M, z0.b, z17.b\n"
-      ".inst 0xa0990001  // smopa za1.s, p0/M, p0/M, z0.b, z25.b\n"
-      ".inst 0xa0910022  // smopa za2.s, p0/M, p0/M, z1.b, z17.b\n"
-      ".inst 0xa0990023  // smopa za3.s, p0/M, p0/M, z1.b, z25.b\n"
+      ".inst 0xa0800380  // smopa za0.s, p0/M, p0/M, z28.b, z0.b\n"
+      ".inst 0xa0880381  // smopa za1.s, p0/M, p0/M, z28.b, z8.b\n"
+      ".inst 0xa08003a2  // smopa za2.s, p0/M, p0/M, z29.b, z0.b\n"
+      ".inst 0xa08803a3  // smopa za3.s, p0/M, p0/M, z29.b, z8.b\n"
+      ".inst 0xa0970040  // smopa za0.s, p0/M, p0/M, z2.b, z23.b\n"
+      ".inst 0xa09f0041  // smopa za1.s, p0/M, p0/M, z2.b, z31.b\n"
+      ".inst 0xa0970062  // smopa za2.s, p0/M, p0/M, z3.b, z23.b\n"
+      ".inst 0xa09f0063  // smopa za3.s, p0/M, p0/M, z3.b, z31.b\n"
+      ".inst 0xa09001c0  // smopa za0.s, p0/M, p0/M, z14.b, z16.b\n"
+      ".inst 0xa09801c1  // smopa za1.s, p0/M, p0/M, z14.b, z24.b\n"
+      ".inst 0xa09001e2  // smopa za2.s, p0/M, p0/M, z15.b, z16.b\n"
+      ".inst 0xa09801e3  // smopa za3.s, p0/M, p0/M, z15.b, z24.b\n"
+      ".inst 0xa0940080  // smopa za0.s, p0/M, p0/M, z4.b, z20.b\n"
+      ".inst 0xa0950081  // smopa za1.s, p0/M, p0/M, z4.b, z21.b\n"
+      ".inst 0xa09400a2  // smopa za2.s, p0/M, p0/M, z5.b, z20.b\n"
+      ".inst 0xa09500a3  // smopa za3.s, p0/M, p0/M, z5.b, z21.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa1400777  // ld1b { z23.b, z31.b }, pn9.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xa1400774  // ld1b { z20.b, z28.b }, pn9.b/Z, [x27]\n"
+      "subs x21, x21, #0x1\n"
       "addvl x27, x27, #2\n"
       ".inst 0xa14006e7  // ld1b { z7.b, z15.b }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #2\n"
-      ".inst 0xa08702e0  // smopa za0.s, p0/M, p0/M, z23.b, z7.b\n"
-      ".inst 0xa08f02e1  // smopa za1.s, p0/M, p0/M, z23.b, z15.b\n"
-      ".inst 0xa08703e2  // smopa za2.s, p0/M, p0/M, z31.b, z7.b\n"
-      ".inst 0xa08f03e3  // smopa za3.s, p0/M, p0/M, z31.b, z15.b\n"
+      ".inst 0xa0870280  // smopa za0.s, p0/M, p0/M, z20.b, z7.b\n"
+      ".inst 0xa08f0281  // smopa za1.s, p0/M, p0/M, z20.b, z15.b\n"
+      ".inst 0xa0870382  // smopa za2.s, p0/M, p0/M, z28.b, z7.b\n"
+      ".inst 0xa08f0383  // smopa za3.s, p0/M, p0/M, z28.b, z15.b\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x16, #1, 14f\n"
@@ -219,25 +219,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
-      ".inst 0xa042c5f4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
+      ".inst 0xa042c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840702  // mova za2h.s[x12], { z24.s-z27.s }\n"
       ".inst 0xa043c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xa060c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14]\n"
-      ".inst 0xc0840682  // mova za2h.s[x12], { z20.s-z23.s }\n"
-      ".inst 0xa061c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa060c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 23f\n"
@@ -245,16 +245,16 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
       ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
-      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
-      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14]\n"
-      "add x12, x12, #0x4\n"
+      ".inst 0xa060c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14]\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
       ".inst 0xa061c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n"
+      "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c5d8  // st1w { z24.s-z27.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 23f\n"
@@ -264,25 +264,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
       "cntw x24\n"
       "ldr x23, [%x[args], %[offsetof_ldcb]]\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x24, LT\n"
       "add x26, x26, x10, LSL #2\n"  // C += n
       "lsr x21, x22, #0x2\n"
       "madd x26, x11, x23, x26\n"  // C += m * ldc
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      ".inst 0xc0860438  // mova { z24.s-z27.s }, za1h.s[x12]\n"
-      ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
+      ".inst 0xa1604344  // st1w { z4.s, z12.s }, p8, [x26]\n"
+      "add x26, x26, x23\n"
+      ".inst 0xa1604345  // st1w { z5.s, z13.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
+      ".inst 0xa1604346  // st1w { z6.s, z14.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
-      "add x26, x26, x23\n"
-      ".inst 0xa1604353  // st1w { z19.s, z27.s }, p8, [x26]\n"
+      ".inst 0xa1604347  // st1w { z7.s, z15.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
       "blt 15b\n"
       "16:"  // Store to output array: Accumulator row 0 oddments
@@ -303,9 +303,9 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
       "subs x25, x25, x22\n"
       "beq 21f\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x24, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 19f\n"
       "18:"  // Store to output array: Accumulator row 1 loop
@@ -313,53 +313,53 @@ void sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL(const int8_t *const A, const in
       ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
       ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xa1604353  // st1w { z19.s, z27.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
       "blt 18b\n"
       "19:"  // Store to output array: Accumulator row 1 oddments
       "cbz x20, 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
-      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
-      ".inst 0xa1604350  // st1w { z16.s, z24.s }, p8, [x26]\n"
+      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
+      ".inst 0xc0860468  // mova { z8.s-z11.s }, za3h.s[x12]\n"
+      ".inst 0xa1604340  // st1w { z0.s, z8.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
       "beq 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xa1604351  // st1w { z17.s, z25.s }, p8, [x26]\n"
+      ".inst 0xa1604341  // st1w { z1.s, z9.s }, p8, [x26]\n"
       "add x26, x26, x23\n"
       "beq 20f\n"
-      ".inst 0xa1604352  // st1w { z18.s, z26.s }, p8, [x26]\n"
+      ".inst 0xa1604342  // st1w { z2.s, z10.s }, p8, [x26]\n"
       "20:"  // Store to output array: Accumulator row 1 oddments: End
       "21:"  // Store to output array: End
       "tbz x16, #0, 23f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "22:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xa040c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa041c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
       ".inst 0xa042c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
       ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa043c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
       ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 22b\n"
       "23:"  // End block
       "incw x10, ALL, MUL #2\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #2\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL.hpp
index 6b1dca0e2a..0c29ab9991 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
 class cls_sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL
 {
 public:
-  typedef int8_t lhs_operand_type;
-  typedef int8_t rhs_operand_type;
+  typedef int8_t operand_type;
   typedef int32_t result_type;
 
   typedef void (*kern_type)(const int8_t *const A, const int8_t *const B, int32_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Activation, bool accumulate, int32_t *const accumulator_buffer);
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 4, 1, 4> transforms = {};
+  StdTransformsSME<operand_type, result_type, 4, 1, 4> transforms = {};
 
   cls_sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp
index 03c19c46f5..932bd6b595 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL/generic.cpp
@@ -48,6 +48,7 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
         B(B), kstride_bytes(roundup(K, 4) * sizeof(int8_t)),
         C(C), ldcb(ldc * sizeof(int32_t)),
         M(M), N(N), K(K),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias),
         accumulator_buffer(accumulator_buffer),
@@ -68,11 +69,10 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
     const long kstride_bytes;
     int32_t *const C;
     const long ldcb;
-    const long M, N, K;
+    const long M, N, K, n_loops, n_tail_iters;
 
     const int32_t *const bias;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -91,17 +91,17 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840481  // mova za1h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xa040c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa041c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840781  // mova za1h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa042c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840703  // mova za3h.s[x12], { z24.s-z27.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -116,102 +116,102 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "ld1w { z1.s }, p0/Z, [x20, x10, LSL #2]\n"
-      ".inst 0xc0902420  // addha za0.s, p1/M, p1/M, z1.s\n"
-      ".inst 0xc0902421  // addha za1.s, p1/M, p1/M, z1.s\n"
-      ".inst 0xc0902422  // addha za2.s, p1/M, p1/M, z1.s\n"
-      ".inst 0xc0902423  // addha za3.s, p1/M, p1/M, z1.s\n"
+      "ldnt1w { z17.s }, p0/Z, [x20, x10, LSL #2]\n"
+      ".inst 0xc0902620  // addha za0.s, p1/M, p1/M, z17.s\n"
+      ".inst 0xc0902621  // addha za1.s, p1/M, p1/M, z17.s\n"
+      ".inst 0xc0902622  // addha za2.s, p1/M, p1/M, z17.s\n"
+      ".inst 0xc0902623  // addha za3.s, p1/M, p1/M, z17.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20\n"
       "incw x21, ALL, MUL #4\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa1408370  // ld1b { z16.b, z20.b, z24.b, z28.b }, pn8.b/Z, [x27]\n"
-      "ld1b { z0.b }, p1/Z, [x23]\n"
-      ".inst 0xa041836c  // ld1b { z12.b-z15.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "ld1b { z10.b }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa1428371  // ld1b { z17.b, z21.b, z25.b, z29.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1b { z18.b }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa1438373  // ld1b { z19.b, z23.b, z27.b, z31.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x10, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa1408372  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n"
+      "ldnt1b { z12.b }, p1/Z, [x23]\n"
+      ".inst 0xa1418370  // ld1b { z16.b, z20.b, z24.b, z28.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      "ldnt1b { z5.b }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0xa1428363  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1b { z4.b }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0xa1438362  // ld1b { z2.b, z6.b, z10.b, z14.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1b { z7.b }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1b { z19.b }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0xa0802600  // smopa za0.s, p1/M, p1/M, z16.b, z0.b\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa0802681  // smopa za1.s, p1/M, p1/M, z20.b, z0.b\n"
-      ".inst 0xa0802702  // smopa za2.s, p1/M, p1/M, z24.b, z0.b\n"
-      ".inst 0xa0802783  // smopa za3.s, p1/M, p1/M, z28.b, z0.b\n"
-      ".inst 0xa1408370  // ld1b { z16.b, z20.b, z24.b, z28.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa08a2580  // smopa za0.s, p1/M, p1/M, z12.b, z10.b\n"
-      "ld1b { z0.b }, p1/Z, [x23]\n"
-      ".inst 0xa08a25a1  // smopa za1.s, p1/M, p1/M, z13.b, z10.b\n"
-      ".inst 0xa08a25c2  // smopa za2.s, p1/M, p1/M, z14.b, z10.b\n"
-      ".inst 0xa08a25e3  // smopa za3.s, p1/M, p1/M, z15.b, z10.b\n"
-      ".inst 0xa041836c  // ld1b { z12.b-z15.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa0922620  // smopa za0.s, p1/M, p1/M, z17.b, z18.b\n"
-      "ld1b { z10.b }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa09226a1  // smopa za1.s, p1/M, p1/M, z21.b, z18.b\n"
-      ".inst 0xa0922722  // smopa za2.s, p1/M, p1/M, z25.b, z18.b\n"
-      ".inst 0xa09227a3  // smopa za3.s, p1/M, p1/M, z29.b, z18.b\n"
-      ".inst 0xa1428371  // ld1b { z17.b, z21.b, z25.b, z29.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1b { z18.b }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa0872660  // smopa za0.s, p1/M, p1/M, z19.b, z7.b\n"
-      ".inst 0xa08726e1  // smopa za1.s, p1/M, p1/M, z23.b, z7.b\n"
-      ".inst 0xa0872762  // smopa za2.s, p1/M, p1/M, z27.b, z7.b\n"
-      ".inst 0xa08727e3  // smopa za3.s, p1/M, p1/M, z31.b, z7.b\n"
-      ".inst 0xa1438373  // ld1b { z19.b, z23.b, z27.b, z31.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      ".inst 0xa08c2640  // smopa za0.s, p1/M, p1/M, z18.b, z12.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa08c26c1  // smopa za1.s, p1/M, p1/M, z22.b, z12.b\n"
+      ".inst 0xa08c2742  // smopa za2.s, p1/M, p1/M, z26.b, z12.b\n"
+      ".inst 0xa08c27c3  // smopa za3.s, p1/M, p1/M, z30.b, z12.b\n"
+      ".inst 0xa1408372  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa0852600  // smopa za0.s, p1/M, p1/M, z16.b, z5.b\n"
+      "ldnt1b { z12.b }, p1/Z, [x23]\n"
+      ".inst 0xa0852681  // smopa za1.s, p1/M, p1/M, z20.b, z5.b\n"
+      ".inst 0xa0852702  // smopa za2.s, p1/M, p1/M, z24.b, z5.b\n"
+      ".inst 0xa0852783  // smopa za3.s, p1/M, p1/M, z28.b, z5.b\n"
+      ".inst 0xa1418370  // ld1b { z16.b, z20.b, z24.b, z28.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa0842460  // smopa za0.s, p1/M, p1/M, z3.b, z4.b\n"
+      "ldnt1b { z5.b }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0xa08424e1  // smopa za1.s, p1/M, p1/M, z7.b, z4.b\n"
+      ".inst 0xa0842562  // smopa za2.s, p1/M, p1/M, z11.b, z4.b\n"
+      ".inst 0xa08425e3  // smopa za3.s, p1/M, p1/M, z15.b, z4.b\n"
+      ".inst 0xa1428363  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1b { z4.b }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0xa0932440  // smopa za0.s, p1/M, p1/M, z2.b, z19.b\n"
+      ".inst 0xa09324c1  // smopa za1.s, p1/M, p1/M, z6.b, z19.b\n"
+      ".inst 0xa0932542  // smopa za2.s, p1/M, p1/M, z10.b, z19.b\n"
+      ".inst 0xa09325c3  // smopa za3.s, p1/M, p1/M, z14.b, z19.b\n"
+      ".inst 0xa1438362  // ld1b { z2.b, z6.b, z10.b, z14.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1b { z7.b }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1b { z19.b }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0xa0802600  // smopa za0.s, p1/M, p1/M, z16.b, z0.b\n"
-      ".inst 0xa0802681  // smopa za1.s, p1/M, p1/M, z20.b, z0.b\n"
-      ".inst 0xa0802702  // smopa za2.s, p1/M, p1/M, z24.b, z0.b\n"
-      ".inst 0xa0802783  // smopa za3.s, p1/M, p1/M, z28.b, z0.b\n"
-      ".inst 0xa08a2580  // smopa za0.s, p1/M, p1/M, z12.b, z10.b\n"
-      ".inst 0xa08a25a1  // smopa za1.s, p1/M, p1/M, z13.b, z10.b\n"
-      ".inst 0xa08a25c2  // smopa za2.s, p1/M, p1/M, z14.b, z10.b\n"
-      ".inst 0xa08a25e3  // smopa za3.s, p1/M, p1/M, z15.b, z10.b\n"
-      ".inst 0xa0922620  // smopa za0.s, p1/M, p1/M, z17.b, z18.b\n"
-      ".inst 0xa09226a1  // smopa za1.s, p1/M, p1/M, z21.b, z18.b\n"
-      ".inst 0xa0922722  // smopa za2.s, p1/M, p1/M, z25.b, z18.b\n"
-      ".inst 0xa09227a3  // smopa za3.s, p1/M, p1/M, z29.b, z18.b\n"
-      ".inst 0xa0872660  // smopa za0.s, p1/M, p1/M, z19.b, z7.b\n"
-      ".inst 0xa08726e1  // smopa za1.s, p1/M, p1/M, z23.b, z7.b\n"
-      ".inst 0xa0872762  // smopa za2.s, p1/M, p1/M, z27.b, z7.b\n"
-      ".inst 0xa08727e3  // smopa za3.s, p1/M, p1/M, z31.b, z7.b\n"
+      ".inst 0xa08c2640  // smopa za0.s, p1/M, p1/M, z18.b, z12.b\n"
+      ".inst 0xa08c26c1  // smopa za1.s, p1/M, p1/M, z22.b, z12.b\n"
+      ".inst 0xa08c2742  // smopa za2.s, p1/M, p1/M, z26.b, z12.b\n"
+      ".inst 0xa08c27c3  // smopa za3.s, p1/M, p1/M, z30.b, z12.b\n"
+      ".inst 0xa0852600  // smopa za0.s, p1/M, p1/M, z16.b, z5.b\n"
+      ".inst 0xa0852681  // smopa za1.s, p1/M, p1/M, z20.b, z5.b\n"
+      ".inst 0xa0852702  // smopa za2.s, p1/M, p1/M, z24.b, z5.b\n"
+      ".inst 0xa0852783  // smopa za3.s, p1/M, p1/M, z28.b, z5.b\n"
+      ".inst 0xa0842460  // smopa za0.s, p1/M, p1/M, z3.b, z4.b\n"
+      ".inst 0xa08424e1  // smopa za1.s, p1/M, p1/M, z7.b, z4.b\n"
+      ".inst 0xa0842562  // smopa za2.s, p1/M, p1/M, z11.b, z4.b\n"
+      ".inst 0xa08425e3  // smopa za3.s, p1/M, p1/M, z15.b, z4.b\n"
+      ".inst 0xa0932440  // smopa za0.s, p1/M, p1/M, z2.b, z19.b\n"
+      ".inst 0xa09324c1  // smopa za1.s, p1/M, p1/M, z6.b, z19.b\n"
+      ".inst 0xa0932542  // smopa za2.s, p1/M, p1/M, z10.b, z19.b\n"
+      ".inst 0xa09325c3  // smopa za3.s, p1/M, p1/M, z14.b, z19.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa1408372  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xa0408368  // ld1b { z8.b-z11.b }, pn8.b/Z, [x27]\n"
+      "subs x21, x21, #0x1\n"
       "addvl x27, x27, #4\n"
       "ld1b { z15.b }, p1/Z, [x23]\n"
       "addvl x23, x23, #1\n"
-      ".inst 0xa08f2640  // smopa za0.s, p1/M, p1/M, z18.b, z15.b\n"
-      ".inst 0xa08f26c1  // smopa za1.s, p1/M, p1/M, z22.b, z15.b\n"
-      ".inst 0xa08f2742  // smopa za2.s, p1/M, p1/M, z26.b, z15.b\n"
-      ".inst 0xa08f27c3  // smopa za3.s, p1/M, p1/M, z30.b, z15.b\n"
+      ".inst 0xa08f2500  // smopa za0.s, p1/M, p1/M, z8.b, z15.b\n"
+      ".inst 0xa08f2521  // smopa za1.s, p1/M, p1/M, z9.b, z15.b\n"
+      ".inst 0xa08f2542  // smopa za2.s, p1/M, p1/M, z10.b, z15.b\n"
+      ".inst 0xa08f2563  // smopa za3.s, p1/M, p1/M, z11.b, z15.b\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
       "tbz x16, #1, 14f\n"
@@ -219,25 +219,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xa041c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
-      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
+      ".inst 0xa040c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xa041c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
+      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
       ".inst 0xa042c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840680  // mova za0h.s[x12], { z20.s-z23.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
-      ".inst 0xa060c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14]\n"
       ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
-      ".inst 0xa061c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840783  // mova za3h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa043c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa060c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 29f\n"
@@ -247,14 +247,14 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
       "13:"  // Store to partial result buffer: Store only: Loop
       ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
       ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
-      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
-      ".inst 0xc0860464  // mova { z4.s-z7.s }, za3h.s[x12]\n"
       ".inst 0xa060c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14]\n"
-      "add x12, x12, #0x4\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
       ".inst 0xa061c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14, #0x4, MUL VL]\n"
+      "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
       ".inst 0xa062c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa063c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 29f\n"
@@ -264,94 +264,94 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
       "cntw x24\n"
       "ldr x23, [%x[args], %[offsetof_ldcb]]\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x24, LT\n"
       "add x26, x26, x10, LSL #2\n"  // C += n
       "lsr x21, x22, #0x2\n"
       "madd x26, x11, x23, x26\n"  // C += m * ldc
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 16f\n"
       "15:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z12.s }, p0, [x26]\n"
+      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
+      "st1w { z8.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z13.s }, p0, [x26]\n"
+      "st1w { z9.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1w { z14.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z10.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1w { z15.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z11.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "blt 15b\n"
       "16:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 17f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      "st1w { z16.s }, p0, [x26]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      "st1w { z4.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 17f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z17.s }, p0, [x26]\n"
+      "st1w { z5.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 17f\n"
-      "st1w { z18.s }, p0, [x26]\n"
+      "st1w { z6.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "17:"  // Store to output array: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x24, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 19f\n"
       "18:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z0.s }, p0, [x26]\n"
+      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
+      "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z1.s }, p0, [x26]\n"
+      "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1w { z2.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1w { z3.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z19.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "blt 18b\n"
       "19:"  // Store to output array: Accumulator row 1 oddments
       "cbz x20, 20f\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
-      "st1w { z12.s }, p0, [x26]\n"
+      ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
+      "st1w { z20.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 20f\n"
       "subs x20, x20, #0x1\n"
-      "st1w { z13.s }, p0, [x26]\n"
+      "st1w { z21.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "beq 20f\n"
-      "st1w { z14.s }, p0, [x26]\n"
+      "st1w { z22.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "20:"  // Store to output array: Accumulator row 1 oddments: End
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x24, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 22f\n"
       "21:"  // Store to output array: Accumulator row 2 loop
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      "add x12, x12, #0x4\n"
-      "st1w { z8.s }, p0, [x26]\n"
+      ".inst 0xc0860458  // mova { z24.s-z27.s }, za2h.s[x12]\n"
+      "st1w { z24.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "cmp x12, x21, LSL #2\n"
-      "st1w { z9.s }, p0, [x26]\n"
+      "st1w { z25.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1w { z10.s }, p0, [x26]\n"
+      "add x12, x12, #0x4\n"
+      "st1w { z26.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "st1w { z11.s }, p0, [x26]\n"
+      "cmp x12, x21, LSL #2\n"
+      "st1w { z27.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "blt 21b\n"
       "22:"  // Store to output array: Accumulator row 2 oddments
@@ -371,21 +371,21 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
       "subs x25, x25, x22\n"
       "beq 27f\n"
       "cmp x25, x24\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x24, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 25f\n"
       "24:"  // Store to output array: Accumulator row 3 loop
       ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       "st1w { z16.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
-      "cmp x12, x21, LSL #2\n"
       "st1w { z17.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
+      "add x12, x12, #0x4\n"
       "st1w { z18.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
+      "cmp x12, x21, LSL #2\n"
       "st1w { z19.s }, p0, [x26]\n"
       "add x26, x26, x23\n"
       "blt 24b\n"
@@ -407,25 +407,25 @@ void sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL(const int8_t *const A, const in
       "mov x12, #0x0\n"
       "cntw x20\n"
       "28:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xa040c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
       ".inst 0xa041c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
       ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa042c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
       ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 28b\n"
       "29:"  // End block
       "incw x10\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #4\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL.hpp
index f384e0f491..f540d3fa24 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin
 class cls_sme2_interleaved_nomerge_u8q_mopa_1VLx4VL
 {
 public:
-  typedef uint8_t lhs_operand_type;
-  typedef uint8_t rhs_operand_type;
+  typedef uint8_t operand_type;
   typedef uint8_t result_type;
 
   typedef void (*kern_type)(const uint8_t *const A, const uint8_t *const B, uint8_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Requantize32 &rq, const int n_0, bool accumulate, int32_t *const accumulator_buffer);
@@ -61,7 +60,7 @@ public:
 
   static constexpr bool supports_accumulate()
   {
-    return true;
+    return false;
   }
 
   static constexpr bool supports_bias()
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_u8q_mopa_1VLx4VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 1, 4, 4, true> transforms = {};
+  StdTransformsSME<operand_type, result_type, 1, 4, 4, true> transforms = {};
 
   cls_sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp
index 0482a5ea0f..0a468e0ff7 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_1VLx4VL/generic.cpp
@@ -49,7 +49,7 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin
         B(B), kstride_bytes(roundup(K, 4) * sizeof(uint8_t)),
         C(C), ldcb(ldc * sizeof(uint8_t)),
         M(M), N(N), K(K),
-        min(0), max(0),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias), n_0(n_0),
         accumulator_buffer(accumulator_buffer),
@@ -74,14 +74,13 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin
     const long kstride_bytes;
     uint8_t *const C;
     const long ldcb;
-    const long M, N, K;
-    int32_t min;
-    int32_t max;
+    const long M, N, K, n_loops, n_tail_iters;
+    int32_t min = std::numeric_limits<uint8_t>::min();
+    int32_t max = std::numeric_limits<uint8_t>::max();
 
     const int32_t *const bias;
     const int n_0;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -90,131 +89,131 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin
   KernelArgs args(A, B, C, ldc, M, N, K, bias, rq, n_0, accumulate, accumulator_buffer);
 
   __asm__ __volatile__(
-      "ldr x15, [%x[args], %[offsetof_flags]]\n"
+      "ldr x14, [%x[args], %[offsetof_flags]]\n"
       ".inst 0xd503477f  // SMSTART ZA\n"
       "ptrue p1.b\n"
       ".inst 0x25207811  // ptrue pn9.b\n"
-      "ldr x14, [%x[args], %[offsetof_accumulator_buffer]]\n"
       "ldr x13, [%x[args], %[offsetof_accumulator_buffer]]\n"
-      "tbz x15, #0, 2f\n"
+      "ldr x11, [%x[args], %[offsetof_accumulator_buffer]]\n"
+      "tbz x14, #0, 2f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5d8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x14]\n"
-      ".inst 0xa041c5c0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c5dc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa040c5a4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x13]\n"
+      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa041c5a8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n"
+      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa042c5a8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n"
+      ".inst 0xc0840502  // mova za2h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa043c5b0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x13, x13, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
-      "ldr w11, [%x[args], %[offsetof_M]]\n"
-      "mov x10, #0x0\n"
+      "ldr w10, [%x[args], %[offsetof_M]]\n"
       "mov x9, #0x0\n"
-      "ldr w28, [%x[args], %[offsetof_N]]\n"
-      "ldr x27, [%x[args], %[offsetof_A]]\n"
+      "mov x28, #0x0\n"
+      "ldr w27, [%x[args], %[offsetof_N]]\n"
+      "ldr x26, [%x[args], %[offsetof_A]]\n"
       "3:"  // M and N loop
-      "mov x26, x27\n"
-      ".inst 0x25bc6530  // whilelt pn8.s, x9, x28, VLx4\n"
-      "tbnz x15, #0, 4f\n"
+      "mov x25, x26\n"
+      ".inst 0x25bb6790  // whilelt pn8.s, x28, x27, VLx4\n"
+      "tbnz x14, #0, 4f\n"
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      ".inst 0xa009c290  // ld1w { z16.s-z19.s }, p8/Z, [x20, x9, LSL #2]\n"
-      ".inst 0xc0902600  // addha za0.s, p1/M, p1/M, z16.s\n"
-      ".inst 0xc0902621  // addha za1.s, p1/M, p1/M, z17.s\n"
-      ".inst 0xc0902642  // addha za2.s, p1/M, p1/M, z18.s\n"
-      ".inst 0xc0902663  // addha za3.s, p1/M, p1/M, z19.s\n"
+      ".inst 0xa11cc289  // ldnt1w { z1.s, z5.s, z9.s, z13.s }, p8/Z, [x20, x28, LSL #2]\n"
+      ".inst 0xc0902420  // addha za0.s, p1/M, p1/M, z1.s\n"
+      ".inst 0xc09024a1  // addha za1.s, p1/M, p1/M, z5.s\n"
+      ".inst 0xc0902522  // addha za2.s, p1/M, p1/M, z9.s\n"
+      ".inst 0xc09025a3  // addha za3.s, p1/M, p1/M, z13.s\n"
       "4:"  // Prepare accumulators: Test for last block
-      "mov x20, x9\n"
-      "mov x21, x10\n"
+      "mov x20, x28\n"
+      "mov x21, x9\n"
       "incw x20, ALL, MUL #4\n"
       "incw x21\n"
-      "cmp x20, x28\n"
-      "mov x20, x15\n"
-      "csel x21, x10, x21, LT\n"
-      "bfm x15, XZR, #0x0, #0x0  // bfc x15, #0x0, #0x1\n"
-      "cmp x21, x11\n"
-      "csel x15, x20, x15, LT\n"
+      "cmp x20, x27\n"
+      "csel x21, x9, x21, LT\n"
+      "mov x20, x14\n"
+      "bfm x14, XZR, #0x0, #0x0  // bfc x14, #0x0, #0x1\n"
+      "cmp x21, x10\n"
+      "csel x14, x20, x14, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x9, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      "ld1b { z5.b }, p1/Z, [x26]\n"
-      ".inst 0xa14086e0  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn9.b/Z, [x23]\n"
-      "ld1b { z31.b }, p1/Z, [x26, #1, MUL VL]\n"
-      ".inst 0xa14186f2  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      "ld1b { z1.b }, p1/Z, [x26, #2, MUL VL]\n"
-      ".inst 0xa14286f0  // ld1b { z16.b, z20.b, z24.b, z28.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      "ld1b { z6.b }, p1/Z, [x26, #3, MUL VL]\n"
-      "addvl x26, x26, #4\n"
-      ".inst 0xa14386e3  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x28, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      "ld1b { z20.b }, p1/Z, [x25]\n"
+      ".inst 0xa04086e5  // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23]\n"
+      "ld1b { z11.b }, p1/Z, [x25, #1, MUL VL]\n"
+      ".inst 0xa04186f9  // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      "ld1b { z2.b }, p1/Z, [x25, #2, MUL VL]\n"
+      ".inst 0xa04286fd  // ldnt1b { z28.b-z31.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      "ld1b { z14.b }, p1/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0xa04386f1  // ldnt1b { z16.b-z19.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0xa1a024a0  // umopa za0.s, p1/M, p1/M, z5.b, z0.b\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa1a424a1  // umopa za1.s, p1/M, p1/M, z5.b, z4.b\n"
-      ".inst 0xa1a824a2  // umopa za2.s, p1/M, p1/M, z5.b, z8.b\n"
-      ".inst 0xa1ac24a3  // umopa za3.s, p1/M, p1/M, z5.b, z12.b\n"
-      "ld1b { z5.b }, p1/Z, [x26]\n"
-      ".inst 0xa1b227e0  // umopa za0.s, p1/M, p1/M, z31.b, z18.b\n"
-      ".inst 0xa14086e0  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa1b627e1  // umopa za1.s, p1/M, p1/M, z31.b, z22.b\n"
-      ".inst 0xa1ba27e2  // umopa za2.s, p1/M, p1/M, z31.b, z26.b\n"
-      ".inst 0xa1be27e3  // umopa za3.s, p1/M, p1/M, z31.b, z30.b\n"
-      "ld1b { z31.b }, p1/Z, [x26, #1, MUL VL]\n"
-      ".inst 0xa1b02420  // umopa za0.s, p1/M, p1/M, z1.b, z16.b\n"
-      ".inst 0xa14186f2  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa1b42421  // umopa za1.s, p1/M, p1/M, z1.b, z20.b\n"
-      ".inst 0xa1b82422  // umopa za2.s, p1/M, p1/M, z1.b, z24.b\n"
-      ".inst 0xa1bc2423  // umopa za3.s, p1/M, p1/M, z1.b, z28.b\n"
-      "ld1b { z1.b }, p1/Z, [x26, #2, MUL VL]\n"
-      ".inst 0xa14286f0  // ld1b { z16.b, z20.b, z24.b, z28.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
-      ".inst 0xa1a324c0  // umopa za0.s, p1/M, p1/M, z6.b, z3.b\n"
-      ".inst 0xa1a724c1  // umopa za1.s, p1/M, p1/M, z6.b, z7.b\n"
-      ".inst 0xa1ab24c2  // umopa za2.s, p1/M, p1/M, z6.b, z11.b\n"
-      ".inst 0xa1af24c3  // umopa za3.s, p1/M, p1/M, z6.b, z15.b\n"
-      "ld1b { z6.b }, p1/Z, [x26, #3, MUL VL]\n"
-      "addvl x26, x26, #4\n"
-      ".inst 0xa14386e3  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
+      ".inst 0xa1a42680  // umopa za0.s, p1/M, p1/M, z20.b, z4.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa1a52681  // umopa za1.s, p1/M, p1/M, z20.b, z5.b\n"
+      ".inst 0xa1a62682  // umopa za2.s, p1/M, p1/M, z20.b, z6.b\n"
+      ".inst 0xa1a72683  // umopa za3.s, p1/M, p1/M, z20.b, z7.b\n"
+      "ld1b { z20.b }, p1/Z, [x25]\n"
+      ".inst 0xa1b82560  // umopa za0.s, p1/M, p1/M, z11.b, z24.b\n"
+      ".inst 0xa04086e5  // ldnt1b { z4.b-z7.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa1b92561  // umopa za1.s, p1/M, p1/M, z11.b, z25.b\n"
+      ".inst 0xa1ba2562  // umopa za2.s, p1/M, p1/M, z11.b, z26.b\n"
+      ".inst 0xa1bb2563  // umopa za3.s, p1/M, p1/M, z11.b, z27.b\n"
+      "ld1b { z11.b }, p1/Z, [x25, #1, MUL VL]\n"
+      ".inst 0xa1bc2440  // umopa za0.s, p1/M, p1/M, z2.b, z28.b\n"
+      ".inst 0xa04186f9  // ldnt1b { z24.b-z27.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa1bd2441  // umopa za1.s, p1/M, p1/M, z2.b, z29.b\n"
+      ".inst 0xa1be2442  // umopa za2.s, p1/M, p1/M, z2.b, z30.b\n"
+      ".inst 0xa1bf2443  // umopa za3.s, p1/M, p1/M, z2.b, z31.b\n"
+      "ld1b { z2.b }, p1/Z, [x25, #2, MUL VL]\n"
+      ".inst 0xa04286fd  // ldnt1b { z28.b-z31.b }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
+      ".inst 0xa1b025c0  // umopa za0.s, p1/M, p1/M, z14.b, z16.b\n"
+      ".inst 0xa1b125c1  // umopa za1.s, p1/M, p1/M, z14.b, z17.b\n"
+      ".inst 0xa1b225c2  // umopa za2.s, p1/M, p1/M, z14.b, z18.b\n"
+      ".inst 0xa1b325c3  // umopa za3.s, p1/M, p1/M, z14.b, z19.b\n"
+      "ld1b { z14.b }, p1/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0xa04386f1  // ldnt1b { z16.b-z19.b }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
       "addvl x23, x23, #16\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0xa1a024a0  // umopa za0.s, p1/M, p1/M, z5.b, z0.b\n"
-      ".inst 0xa1a424a1  // umopa za1.s, p1/M, p1/M, z5.b, z4.b\n"
-      ".inst 0xa1a824a2  // umopa za2.s, p1/M, p1/M, z5.b, z8.b\n"
-      ".inst 0xa1ac24a3  // umopa za3.s, p1/M, p1/M, z5.b, z12.b\n"
-      ".inst 0xa1b227e0  // umopa za0.s, p1/M, p1/M, z31.b, z18.b\n"
-      ".inst 0xa1b627e1  // umopa za1.s, p1/M, p1/M, z31.b, z22.b\n"
-      ".inst 0xa1ba27e2  // umopa za2.s, p1/M, p1/M, z31.b, z26.b\n"
-      ".inst 0xa1be27e3  // umopa za3.s, p1/M, p1/M, z31.b, z30.b\n"
-      ".inst 0xa1b02420  // umopa za0.s, p1/M, p1/M, z1.b, z16.b\n"
-      ".inst 0xa1b42421  // umopa za1.s, p1/M, p1/M, z1.b, z20.b\n"
-      ".inst 0xa1b82422  // umopa za2.s, p1/M, p1/M, z1.b, z24.b\n"
-      ".inst 0xa1bc2423  // umopa za3.s, p1/M, p1/M, z1.b, z28.b\n"
-      ".inst 0xa1a324c0  // umopa za0.s, p1/M, p1/M, z6.b, z3.b\n"
-      ".inst 0xa1a724c1  // umopa za1.s, p1/M, p1/M, z6.b, z7.b\n"
-      ".inst 0xa1ab24c2  // umopa za2.s, p1/M, p1/M, z6.b, z11.b\n"
-      ".inst 0xa1af24c3  // umopa za3.s, p1/M, p1/M, z6.b, z15.b\n"
+      ".inst 0xa1a42680  // umopa za0.s, p1/M, p1/M, z20.b, z4.b\n"
+      ".inst 0xa1a52681  // umopa za1.s, p1/M, p1/M, z20.b, z5.b\n"
+      ".inst 0xa1a62682  // umopa za2.s, p1/M, p1/M, z20.b, z6.b\n"
+      ".inst 0xa1a72683  // umopa za3.s, p1/M, p1/M, z20.b, z7.b\n"
+      ".inst 0xa1b82560  // umopa za0.s, p1/M, p1/M, z11.b, z24.b\n"
+      ".inst 0xa1b92561  // umopa za1.s, p1/M, p1/M, z11.b, z25.b\n"
+      ".inst 0xa1ba2562  // umopa za2.s, p1/M, p1/M, z11.b, z26.b\n"
+      ".inst 0xa1bb2563  // umopa za3.s, p1/M, p1/M, z11.b, z27.b\n"
+      ".inst 0xa1bc2440  // umopa za0.s, p1/M, p1/M, z2.b, z28.b\n"
+      ".inst 0xa1bd2441  // umopa za1.s, p1/M, p1/M, z2.b, z29.b\n"
+      ".inst 0xa1be2442  // umopa za2.s, p1/M, p1/M, z2.b, z30.b\n"
+      ".inst 0xa1bf2443  // umopa za3.s, p1/M, p1/M, z2.b, z31.b\n"
+      ".inst 0xa1b025c0  // umopa za0.s, p1/M, p1/M, z14.b, z16.b\n"
+      ".inst 0xa1b125c1  // umopa za1.s, p1/M, p1/M, z14.b, z17.b\n"
+      ".inst 0xa1b225c2  // umopa za2.s, p1/M, p1/M, z14.b, z18.b\n"
+      ".inst 0xa1b325c3  // umopa za3.s, p1/M, p1/M, z14.b, z19.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      "ld1b { z16.b }, p1/Z, [x26]\n"
-      "subs x20, x20, #0x1\n"
-      "addvl x26, x26, #1\n"
+      "ld1b { z16.b }, p1/Z, [x25]\n"
+      "subs x21, x21, #0x1\n"
+      "addvl x25, x25, #1\n"
       ".inst 0xa04086e4  // ld1b { z4.b-z7.b }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #4\n"
       ".inst 0xa1a42600  // umopa za0.s, p1/M, p1/M, z16.b, z4.b\n"
@@ -223,182 +222,182 @@ void sme2_interleaved_nomerge_u8q_mopa_1VLx4VL(const uint8_t *const A, const uin
       ".inst 0xa1a72603  // umopa za3.s, p1/M, p1/M, z16.b, z7.b\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
-      "ld1w { z15.s }, p1/Z, [x26]\n"
-      "addvl x26, x26, #1\n"
+      "ld1w { z15.s }, p1/Z, [x25]\n"
+      "addvl x25, x25, #1\n"
       ".inst 0xc09125e0  // addva za0.s, p1/M, p1/M, z15.s\n"
       ".inst 0xc09125e1  // addva za1.s, p1/M, p1/M, z15.s\n"
       ".inst 0xc09125e2  // addva za2.s, p1/M, p1/M, z15.s\n"
       ".inst 0xc09125e3  // addva za3.s, p1/M, p1/M, z15.s\n"
-      "tbz x15, #1, 14f\n"
-      "tbz x15, #0, 12f\n"
+      "tbz x14, #1, 14f\n"
+      "tbz x14, #0, 12f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14]\n"
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
-      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5d0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa040c5a0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x13]\n"
+      ".inst 0xc0860418  // mova { z24.s-z27.s }, za0h.s[x12]\n"
+      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5a0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n"
+      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
       ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
-      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
-      ".inst 0xa042c5c0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
-      ".inst 0xa060c5bc  // st1w { z28.s-z31.s }, pn9.b, [x13]\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xa061c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13, #0x4, MUL VL]\n"
-      ".inst 0xc0840583  // mova za3h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
+      ".inst 0xa042c5b0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c5b0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5b4  // st1w { z20.s-z23.s }, pn9.b, [x13, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c5b8  // st1w { z24.s-z27.s }, pn9.b, [x13, #0xc, MUL VL]\n"
+      ".inst 0xa060c578  // st1w { z24.s-z27.s }, pn9.b, [x11]\n"
       "addvl x13, x13, #16\n"
+      ".inst 0xa061c564  // st1w { z4.s-z7.s }, pn9.b, [x11, #0x4, MUL VL]\n"
+      ".inst 0xa062c574  // st1w { z20.s-z23.s }, pn9.b, [x11, #0x8, MUL VL]\n"
+      ".inst 0xa063c560  // st1w { z0.s-z3.s }, pn9.b, [x11, #0xc, MUL VL]\n"
+      "addvl x11, x11, #16\n"
       "blt 11b\n"
       "b 21f\n"
       "12:"  // Store to partial result buffer: Store only
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0860434  // mova { z20.s-z23.s }, za1h.s[x12]\n"
+      ".inst 0xa060c564  // st1w { z4.s-z7.s }, pn9.b, [x11]\n"
       ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5a8  // st1w { z8.s-z11.s }, pn9.b, [x13]\n"
+      ".inst 0xc0860468  // mova { z8.s-z11.s }, za3h.s[x12]\n"
+      ".inst 0xa061c574  // st1w { z20.s-z23.s }, pn9.b, [x11, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5a4  // st1w { z4.s-z7.s }, pn9.b, [x13, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c5ac  // st1w { z12.s-z15.s }, pn9.b, [x13, #0x8, MUL VL]\n"
-      ".inst 0xa063c5b0  // st1w { z16.s-z19.s }, pn9.b, [x13, #0xc, MUL VL]\n"
-      "addvl x13, x13, #16\n"
+      ".inst 0xa062c56c  // st1w { z12.s-z15.s }, pn9.b, [x11, #0x8, MUL VL]\n"
+      ".inst 0xa063c568  // st1w { z8.s-z11.s }, pn9.b, [x11, #0xc, MUL VL]\n"
+      "addvl x11, x11, #16\n"
       "blt 13b\n"
       "b 21f\n"
       "14:"  // Store to output array
-      "ldr x25, [%x[args], %[offsetof_C]]\n"
-      "sub x24, x11, x10\n"
+      "ldr x24, [%x[args], %[offsetof_C]]\n"
+      "add x24, x24, x28\n"  // C += n
+      "sub x23, x10, x9\n"
       "ld1rw { z4.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
-      "ldr x23, [%x[args], %[offsetof_ldcb]]\n"
+      "ldr x22, [%x[args], %[offsetof_ldcb]]\n"
+      "madd x24, x9, x22, x24\n"  // C += m * ldc
       "ld1rw { z5.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
       "ld1rw { z6.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
-      "add x25, x25, x9\n"  // C += n
       "ld1rw { z7.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
-      "madd x25, x10, x23, x25\n"  // C += m * ldc
-      "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "ld1rw { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "ld1rw { z3.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "ld1rw { z8.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
+      "ld1rw { z12.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z13.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z14.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z15.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
       "ld1rw { z21.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
       "ld1rw { z20.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
-      "tbz x15, #2, 15f\n"
-      "ldr w22, [%x[args], %[offsetof_n_0]]\n"
-      "ldr x21, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "tbz x14, #2, 15f\n"
+      "ldr w21, [%x[args], %[offsetof_n_0]]\n"
+      "add x21, x21, x28\n"
+      "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "add x20, x20, x21, LSL #2\n"
+      ".inst 0xa040c284  // ld1w { z4.s-z7.s }, p8/Z, [x20]\n"
       "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n"
-      "add x22, x22, x9\n"
-      "add x21, x21, x22, LSL #2\n"
-      "add x20, x20, x22, LSL #2\n"
-      ".inst 0xa040c2a4  // ld1w { z4.s-z7.s }, p8/Z, [x21]\n"
-      ".inst 0xa040c280  // ld1w { z0.s-z3.s }, p8/Z, [x20]\n"
+      "add x20, x20, x21, LSL #2\n"
+      ".inst 0xa040c28c  // ld1w { z12.s-z15.s }, p8/Z, [x20]\n"
       "15:"  // Store to output array: Load per-channel parameters: End
       "cntw x20\n"
-      "whilelt p0.b, x9, x28\n"
-      "cmp x24, x20\n"
-      "mov x12, #0x0\n"
-      "csel x20, x24, x20, LT\n"
+      "whilelt p0.b, x28, x27\n"
+      "cmp x23, x20\n"
+      "csel x20, x23, x20, LT\n"
       "lsr x21, x20, #0x1\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x1\n"
       "cbz x21, 17f\n"
       "16:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc0860010  // mova { z16.s-z17.s }, za0h.s[x12, 0:1]\n"
-      ".inst 0xc086005e  // mova { z30.s-z31.s }, za1h.s[x12, 0:1]\n"
-      ".inst 0xc086009a  // mova { z26.s-z27.s }, za2h.s[x12, 0:1]\n"
-      ".inst 0xc08600cc  // mova { z12.s-z13.s }, za3h.s[x12, 0:1]\n"
-      ".inst 0xc1a4a410  // sqdmulh { z16.s-z17.s }, { z16.s-z17.s }, z4.s\n"
-      ".inst 0xc1a5a41e  // sqdmulh { z30.s-z31.s }, { z30.s-z31.s }, z5.s\n"
+      ".inst 0xc086001a  // mova { z26.s-z27.s }, za0h.s[x12, 0:1]\n"
+      ".inst 0xc086005c  // mova { z28.s-z29.s }, za1h.s[x12, 0:1]\n"
+      ".inst 0xc1a4a41a  // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z4.s\n"
+      ".inst 0xc0860096  // mova { z22.s-z23.s }, za2h.s[x12, 0:1]\n"
+      ".inst 0xc08600d0  // mova { z16.s-z17.s }, za3h.s[x12, 0:1]\n"
+      ".inst 0xc1a5a41c  // sqdmulh { z28.s-z29.s }, { z28.s-z29.s }, z5.s\n"
+      ".inst 0xc1a6a416  // sqdmulh { z22.s-z23.s }, { z22.s-z23.s }, z6.s\n"
       "add x12, x12, #0x2\n"
-      ".inst 0xc1a6a41a  // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z6.s\n"
       "cmp x12, x21, LSL #1\n"
-      ".inst 0xc1a7a40c  // sqdmulh { z12.s-z13.s }, { z12.s-z13.s }, z7.s\n"
-      ".inst 0xc1a0a230  // srshl { z16.s-z17.s }, { z16.s-z17.s }, z0.s\n"
-      ".inst 0xc1a1a23e  // srshl { z30.s-z31.s }, { z30.s-z31.s }, z1.s\n"
-      ".inst 0xc1a2a23a  // srshl { z26.s-z27.s }, { z26.s-z27.s }, z2.s\n"
-      ".inst 0xc1a3a22c  // srshl { z12.s-z13.s }, { z12.s-z13.s }, z3.s\n"
-      ".inst 0xc1a8a310  // add { z16.s-z17.s }, { z16.s-z17.s }, z8.s\n"
-      ".inst 0xc1a8a31e  // add { z30.s-z31.s }, { z30.s-z31.s }, z8.s\n"
-      ".inst 0xc1a8a31a  // add { z26.s-z27.s }, { z26.s-z27.s }, z8.s\n"
-      ".inst 0xc1a8a30c  // add { z12.s-z13.s }, { z12.s-z13.s }, z8.s\n"
-      ".inst 0xc1b4c6b0  // sclamp { z16.s-z17.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4c6be  // sclamp { z30.s-z31.s }, z21.s, z20.s\n"
+      ".inst 0xc1a7a410  // sqdmulh { z16.s-z17.s }, { z16.s-z17.s }, z7.s\n"
+      ".inst 0xc1aca23a  // srshl { z26.s-z27.s }, { z26.s-z27.s }, z12.s\n"
+      ".inst 0xc1ada23c  // srshl { z28.s-z29.s }, { z28.s-z29.s }, z13.s\n"
+      ".inst 0xc1aea236  // srshl { z22.s-z23.s }, { z22.s-z23.s }, z14.s\n"
+      ".inst 0xc1afa230  // srshl { z16.s-z17.s }, { z16.s-z17.s }, z15.s\n"
+      ".inst 0xc1a0a31a  // add { z26.s-z27.s }, { z26.s-z27.s }, z0.s\n"
+      ".inst 0xc1a0a31c  // add { z28.s-z29.s }, { z28.s-z29.s }, z0.s\n"
+      ".inst 0xc1a0a316  // add { z22.s-z23.s }, { z22.s-z23.s }, z0.s\n"
+      ".inst 0xc1a0a310  // add { z16.s-z17.s }, { z16.s-z17.s }, z0.s\n"
       ".inst 0xc1b4c6ba  // sclamp { z26.s-z27.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4c6ac  // sclamp { z12.s-z13.s }, z21.s, z20.s\n"
-      "uzp1 z19.b, z16.b, z30.b\n"
-      "uzp1 z18.b, z17.b, z31.b\n"
-      "uzp1 z17.b, z26.b, z12.b\n"
-      "uzp1 z16.b, z27.b, z13.b\n"
-      "uzp1 z17.b, z19.b, z17.b\n"
-      "uzp1 z16.b, z18.b, z16.b\n"
-      "st1b { z17.b }, p0, [x25]\n"
-      "add x25, x25, x23\n"
-      "st1b { z16.b }, p0, [x25]\n"
-      "add x25, x25, x23\n"
+      ".inst 0xc1b4c6bc  // sclamp { z28.s-z29.s }, z21.s, z20.s\n"
+      "uzp1 z19.b, z26.b, z28.b\n"
+      ".inst 0xc1b4c6b6  // sclamp { z22.s-z23.s }, z21.s, z20.s\n"
+      ".inst 0xc1b4c6b0  // sclamp { z16.s-z17.s }, z21.s, z20.s\n"
+      "uzp1 z16.b, z22.b, z16.b\n"
+      "uzp1 z18.b, z27.b, z29.b\n"
+      "uzp1 z17.b, z23.b, z17.b\n"
+      "uzp1 z16.b, z19.b, z16.b\n"
+      "st1b { z16.b }, p0, [x24]\n"
+      "add x24, x24, x22\n"
+      "uzp1 z16.b, z18.b, z17.b\n"
+      "st1b { z16.b }, p0, [x24]\n"
+      "add x24, x24, x22\n"
       "blt 16b\n"
       "17:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 18f\n"
       ".inst 0xc086000a  // mova { z10.s-z11.s }, za0h.s[x12, 0:1]\n"
-      ".inst 0xc086005a  // mova { z26.s-z27.s }, za1h.s[x12, 0:1]\n"
-      ".inst 0xc086008e  // mova { z14.s-z15.s }, za2h.s[x12, 0:1]\n"
-      ".inst 0xc08600d6  // mova { z22.s-z23.s }, za3h.s[x12, 0:1]\n"
+      ".inst 0xc0860058  // mova { z24.s-z25.s }, za1h.s[x12, 0:1]\n"
       ".inst 0xc1a4a40a  // sqdmulh { z10.s-z11.s }, { z10.s-z11.s }, z4.s\n"
-      ".inst 0xc1a5a41a  // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z5.s\n"
-      ".inst 0xc1a6a40e  // sqdmulh { z14.s-z15.s }, { z14.s-z15.s }, z6.s\n"
-      ".inst 0xc1a7a416  // sqdmulh { z22.s-z23.s }, { z22.s-z23.s }, z7.s\n"
-      ".inst 0xc1a0a22a  // srshl { z10.s-z11.s }, { z10.s-z11.s }, z0.s\n"
-      ".inst 0xc1a1a23a  // srshl { z26.s-z27.s }, { z26.s-z27.s }, z1.s\n"
-      ".inst 0xc1a2a22e  // srshl { z14.s-z15.s }, { z14.s-z15.s }, z2.s\n"
-      ".inst 0xc1a3a236  // srshl { z22.s-z23.s }, { z22.s-z23.s }, z3.s\n"
-      ".inst 0xc1a8a30a  // add { z10.s-z11.s }, { z10.s-z11.s }, z8.s\n"
-      ".inst 0xc1a8a31a  // add { z26.s-z27.s }, { z26.s-z27.s }, z8.s\n"
-      ".inst 0xc1a8a30e  // add { z14.s-z15.s }, { z14.s-z15.s }, z8.s\n"
-      ".inst 0xc1a8a316  // add { z22.s-z23.s }, { z22.s-z23.s }, z8.s\n"
+      ".inst 0xc086009a  // mova { z26.s-z27.s }, za2h.s[x12, 0:1]\n"
+      ".inst 0xc08600de  // mova { z30.s-z31.s }, za3h.s[x12, 0:1]\n"
+      ".inst 0xc1a5a418  // sqdmulh { z24.s-z25.s }, { z24.s-z25.s }, z5.s\n"
+      ".inst 0xc1a6a41a  // sqdmulh { z26.s-z27.s }, { z26.s-z27.s }, z6.s\n"
+      ".inst 0xc1a7a41e  // sqdmulh { z30.s-z31.s }, { z30.s-z31.s }, z7.s\n"
+      ".inst 0xc1aca22a  // srshl { z10.s-z11.s }, { z10.s-z11.s }, z12.s\n"
+      ".inst 0xc1ada238  // srshl { z24.s-z25.s }, { z24.s-z25.s }, z13.s\n"
+      ".inst 0xc1aea23a  // srshl { z26.s-z27.s }, { z26.s-z27.s }, z14.s\n"
+      ".inst 0xc1afa23e  // srshl { z30.s-z31.s }, { z30.s-z31.s }, z15.s\n"
+      ".inst 0xc1a0a30a  // add { z10.s-z11.s }, { z10.s-z11.s }, z0.s\n"
+      ".inst 0xc1a0a318  // add { z24.s-z25.s }, { z24.s-z25.s }, z0.s\n"
+      ".inst 0xc1a0a31a  // add { z26.s-z27.s }, { z26.s-z27.s }, z0.s\n"
+      ".inst 0xc1a0a31e  // add { z30.s-z31.s }, { z30.s-z31.s }, z0.s\n"
       ".inst 0xc1b4c6aa  // sclamp { z10.s-z11.s }, z21.s, z20.s\n"
+      ".inst 0xc1b4c6b8  // sclamp { z24.s-z25.s }, z21.s, z20.s\n"
+      "uzp1 z17.b, z10.b, z24.b\n"
       ".inst 0xc1b4c6ba  // sclamp { z26.s-z27.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4c6ae  // sclamp { z14.s-z15.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4c6b6  // sclamp { z22.s-z23.s }, z21.s, z20.s\n"
-      "uzp1 z17.b, z10.b, z26.b\n"
-      "uzp1 z16.b, z14.b, z22.b\n"
+      ".inst 0xc1b4c6be  // sclamp { z30.s-z31.s }, z21.s, z20.s\n"
+      "uzp1 z16.b, z26.b, z30.b\n"
       "uzp1 z16.b, z17.b, z16.b\n"
-      "st1b { z16.b }, p0, [x25]\n"
+      "st1b { z16.b }, p0, [x24]\n"
       "18:"  // Store to output array: Accumulator row 0 oddments: End
       "19:"  // Store to output array: End
-      "tbz x15, #0, 21f\n"
+      "tbz x14, #0, 21f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "20:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5d8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x14]\n"
-      ".inst 0xa041c5cc  // ld1w { z12.s-z15.s }, pn9.b/Z, [x14, #0x4, MUL VL]\n"
-      ".inst 0xa042c5c4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa043c5c0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x14, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x14, x14, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa040c5bc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x13]\n"
+      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa041c5b0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x13, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c5ac  // ld1w { z12.s-z15.s }, pn9.b/Z, [x13, #0x8, MUL VL]\n"
+      ".inst 0xc0840582  // mova za2h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xa043c5a0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x13, #0xc, MUL VL]\n"
       ".inst 0xc0840403  // mova za3h.s[x12], { z0.s-z3.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x13, x13, #16\n"
       "blt 20b\n"
       "21:"  // End block
-      "incw x9, ALL, MUL #4\n"
-      "cmp x9, x28\n"
+      "incw x28, ALL, MUL #4\n"
+      "cmp x28, x27\n"
       "blt 3b\n"
-      "incw x10\n"
-      "mov x9, #0x0\n"
-      "cmp x10, x11\n"
-      "mov x27, x26\n"
+      "incw x9\n"
+      "cmp x9, x10\n"
+      "mov x28, #0x0\n"
+      "mov x26, x25\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
       :
       : [args] "r" (&args), [offsetof_A] "I" (offsetof(KernelArgs, A)), [offsetof_B] "I" (offsetof(KernelArgs, B)), [offsetof_C] "I" (offsetof(KernelArgs, C)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_M] "I" (offsetof(KernelArgs, M)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_Requantize32_c_offset] "I" (offsetof(Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(Requantize32, minval)), [offsetof_Requantize32_per_channel_muls] "I" (offsetof(Requantize32, per_channel_muls)), [offsetof_Requantize32_per_channel_right_shifts] "I" (offsetof(Requantize32, per_channel_right_shifts)), [offsetof_Requantize32_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_Requantize32_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [offsetof_accumulator_buffer] "I" (offsetof(KernelArgs, accumulator_buffer)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_flags] "I" (offsetof(KernelArgs, flags)), [offsetof_kstride_bytes] "I" (offsetof(KernelArgs, kstride_bytes)), [offsetof_ldcb] "I" (offsetof(KernelArgs, ldcb)), [offsetof_n_0] "I" (offsetof(KernelArgs, n_0)), [rq] "r" (&rq)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL.hpp
index a2621e85f4..8f5880bcea 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
 class cls_sme2_interleaved_nomerge_u8q_mopa_2VLx2VL
 {
 public:
-  typedef uint8_t lhs_operand_type;
-  typedef uint8_t rhs_operand_type;
+  typedef uint8_t operand_type;
   typedef uint8_t result_type;
 
   typedef void (*kern_type)(const uint8_t *const A, const uint8_t *const B, uint8_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Requantize32 &rq, const int n_0, bool accumulate, int32_t *const accumulator_buffer);
@@ -61,7 +60,7 @@ public:
 
   static constexpr bool supports_accumulate()
   {
-    return true;
+    return false;
   }
 
   static constexpr bool supports_bias()
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_u8q_mopa_2VLx2VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 2, 2, 4, true> transforms = {};
+  StdTransformsSME<operand_type, result_type, 2, 2, 4, true> transforms = {};
 
   cls_sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/generic.cpp
index 51b93d3636..8e8524a780 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_2VLx2VL/generic.cpp
@@ -49,7 +49,7 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
         B(B), kstride_bytes(roundup(K, 4) * sizeof(uint8_t)),
         C(C), ldcb(ldc * sizeof(uint8_t)),
         M(M), N(N), K(K),
-        min(0), max(0),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias), n_0(n_0),
         accumulator_buffer(accumulator_buffer),
@@ -74,14 +74,13 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
     const long kstride_bytes;
     uint8_t *const C;
     const long ldcb;
-    const long M, N, K;
-    int32_t min;
-    int32_t max;
+    const long M, N, K, n_loops, n_tail_iters;
+    int32_t min = std::numeric_limits<uint8_t>::min();
+    int32_t max = std::numeric_limits<uint8_t>::max();
 
     const int32_t *const bias;
     const int n_0;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -100,17 +99,17 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5f4  // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840501  // mova za1h.s[x12], { z8.s-z11.s }\n"
-      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
-      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
+      ".inst 0xa040c5e8  // ld1w { z8.s-z11.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
+      ".inst 0xa041c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840401  // mova za1h.s[x12], { z0.s-z3.s }\n"
+      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -125,108 +124,108 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      ".inst 0xa10a4294  // ld1w { z20.s, z28.s }, p8/Z, [x20, x10, LSL #2]\n"
-      ".inst 0xc0902680  // addha za0.s, p1/M, p1/M, z20.s\n"
-      ".inst 0xc0902781  // addha za1.s, p1/M, p1/M, z28.s\n"
-      ".inst 0xc0902682  // addha za2.s, p1/M, p1/M, z20.s\n"
-      ".inst 0xc0902783  // addha za3.s, p1/M, p1/M, z28.s\n"
+      ".inst 0xa00a4299  // ldnt1w { z24.s-z25.s }, p8/Z, [x20, x10, LSL #2]\n"
+      ".inst 0xc0902700  // addha za0.s, p1/M, p1/M, z24.s\n"
+      ".inst 0xc0902721  // addha za1.s, p1/M, p1/M, z25.s\n"
+      ".inst 0xc0902702  // addha za2.s, p1/M, p1/M, z24.s\n"
+      ".inst 0xc0902723  // addha za3.s, p1/M, p1/M, z25.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20, ALL, MUL #2\n"
       "incw x21, ALL, MUL #2\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa0400778  // ld1b { z24.b-z25.b }, pn9.b/Z, [x27]\n"
-      ".inst 0xa14006f7  // ld1b { z23.b, z31.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa1410776  // ld1b { z22.b, z30.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0xa04106fa  // ld1b { z26.b-z27.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa0420766  // ld1b { z6.b-z7.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14206e0  // ld1b { z0.b, z8.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa043077c  // ld1b { z28.b-z29.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x10, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa1400763  // ld1b { z3.b, z11.b }, pn9.b/Z, [x27]\n"
+      ".inst 0xa14006f9  // ldnt1b { z17.b, z25.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa1410774  // ld1b { z20.b, z28.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa04106f7  // ldnt1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0xa1420775  // ld1b { z21.b, z29.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa14206f8  // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa1430765  // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa04306ec  // ld1b { z12.b-z13.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa14306ef  // ldnt1b { z7.b, z15.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0xa1b72700  // umopa za0.s, p1/M, p1/M, z24.b, z23.b\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa1bf2701  // umopa za1.s, p1/M, p1/M, z24.b, z31.b\n"
-      ".inst 0xa1b72722  // umopa za2.s, p1/M, p1/M, z25.b, z23.b\n"
-      ".inst 0xa1bf2723  // umopa za3.s, p1/M, p1/M, z25.b, z31.b\n"
-      ".inst 0xa0400778  // ld1b { z24.b-z25.b }, pn9.b/Z, [x27]\n"
-      ".inst 0xa1ba26c0  // umopa za0.s, p1/M, p1/M, z22.b, z26.b\n"
-      ".inst 0xa14006f7  // ld1b { z23.b, z31.b }, pn9.b/Z, [x23]\n"
-      ".inst 0xa1bb26c1  // umopa za1.s, p1/M, p1/M, z22.b, z27.b\n"
-      ".inst 0xa1ba27c2  // umopa za2.s, p1/M, p1/M, z30.b, z26.b\n"
-      ".inst 0xa1bb27c3  // umopa za3.s, p1/M, p1/M, z30.b, z27.b\n"
-      ".inst 0xa1410776  // ld1b { z22.b, z30.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
-      ".inst 0xa1a024c0  // umopa za0.s, p1/M, p1/M, z6.b, z0.b\n"
-      ".inst 0xa04106fa  // ld1b { z26.b-z27.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
-      ".inst 0xa1a824c1  // umopa za1.s, p1/M, p1/M, z6.b, z8.b\n"
-      ".inst 0xa1a024e2  // umopa za2.s, p1/M, p1/M, z7.b, z0.b\n"
-      ".inst 0xa1a824e3  // umopa za3.s, p1/M, p1/M, z7.b, z8.b\n"
-      ".inst 0xa0420766  // ld1b { z6.b-z7.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa14206e0  // ld1b { z0.b, z8.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
-      ".inst 0xa1ac2780  // umopa za0.s, p1/M, p1/M, z28.b, z12.b\n"
-      ".inst 0xa1ad2781  // umopa za1.s, p1/M, p1/M, z28.b, z13.b\n"
-      ".inst 0xa1ac27a2  // umopa za2.s, p1/M, p1/M, z29.b, z12.b\n"
-      ".inst 0xa1ad27a3  // umopa za3.s, p1/M, p1/M, z29.b, z13.b\n"
-      ".inst 0xa043077c  // ld1b { z28.b-z29.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
+      ".inst 0xa1b12460  // umopa za0.s, p1/M, p1/M, z3.b, z17.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa1b92461  // umopa za1.s, p1/M, p1/M, z3.b, z25.b\n"
+      ".inst 0xa1b12562  // umopa za2.s, p1/M, p1/M, z11.b, z17.b\n"
+      ".inst 0xa1b92563  // umopa za3.s, p1/M, p1/M, z11.b, z25.b\n"
+      ".inst 0xa1400763  // ld1b { z3.b, z11.b }, pn9.b/Z, [x27]\n"
+      ".inst 0xa1b62680  // umopa za0.s, p1/M, p1/M, z20.b, z22.b\n"
+      ".inst 0xa14006f9  // ldnt1b { z17.b, z25.b }, pn9.b/Z, [x23]\n"
+      ".inst 0xa1b72681  // umopa za1.s, p1/M, p1/M, z20.b, z23.b\n"
+      ".inst 0xa1b62782  // umopa za2.s, p1/M, p1/M, z28.b, z22.b\n"
+      ".inst 0xa1b72783  // umopa za3.s, p1/M, p1/M, z28.b, z23.b\n"
+      ".inst 0xa1410774  // ld1b { z20.b, z28.b }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
+      ".inst 0xa1b026a0  // umopa za0.s, p1/M, p1/M, z21.b, z16.b\n"
+      ".inst 0xa04106f7  // ldnt1b { z22.b-z23.b }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
+      ".inst 0xa1b826a1  // umopa za1.s, p1/M, p1/M, z21.b, z24.b\n"
+      ".inst 0xa1b027a2  // umopa za2.s, p1/M, p1/M, z29.b, z16.b\n"
+      ".inst 0xa1b827a3  // umopa za3.s, p1/M, p1/M, z29.b, z24.b\n"
+      ".inst 0xa1420775  // ld1b { z21.b, z29.b }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa14206f8  // ldnt1b { z16.b, z24.b }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
+      ".inst 0xa1a724a0  // umopa za0.s, p1/M, p1/M, z5.b, z7.b\n"
+      ".inst 0xa1af24a1  // umopa za1.s, p1/M, p1/M, z5.b, z15.b\n"
+      ".inst 0xa1a725a2  // umopa za2.s, p1/M, p1/M, z13.b, z7.b\n"
+      ".inst 0xa1af25a3  // umopa za3.s, p1/M, p1/M, z13.b, z15.b\n"
+      ".inst 0xa1430765  // ld1b { z5.b, z13.b }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
       "addvl x27, x27, #8\n"
-      ".inst 0xa04306ec  // ld1b { z12.b-z13.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
+      ".inst 0xa14306ef  // ldnt1b { z7.b, z15.b }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
       "addvl x23, x23, #8\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0xa1b72700  // umopa za0.s, p1/M, p1/M, z24.b, z23.b\n"
-      ".inst 0xa1bf2701  // umopa za1.s, p1/M, p1/M, z24.b, z31.b\n"
-      ".inst 0xa1b72722  // umopa za2.s, p1/M, p1/M, z25.b, z23.b\n"
-      ".inst 0xa1bf2723  // umopa za3.s, p1/M, p1/M, z25.b, z31.b\n"
-      ".inst 0xa1ba26c0  // umopa za0.s, p1/M, p1/M, z22.b, z26.b\n"
-      ".inst 0xa1bb26c1  // umopa za1.s, p1/M, p1/M, z22.b, z27.b\n"
-      ".inst 0xa1ba27c2  // umopa za2.s, p1/M, p1/M, z30.b, z26.b\n"
-      ".inst 0xa1bb27c3  // umopa za3.s, p1/M, p1/M, z30.b, z27.b\n"
-      ".inst 0xa1a024c0  // umopa za0.s, p1/M, p1/M, z6.b, z0.b\n"
-      ".inst 0xa1a824c1  // umopa za1.s, p1/M, p1/M, z6.b, z8.b\n"
-      ".inst 0xa1a024e2  // umopa za2.s, p1/M, p1/M, z7.b, z0.b\n"
-      ".inst 0xa1a824e3  // umopa za3.s, p1/M, p1/M, z7.b, z8.b\n"
-      ".inst 0xa1ac2780  // umopa za0.s, p1/M, p1/M, z28.b, z12.b\n"
-      ".inst 0xa1ad2781  // umopa za1.s, p1/M, p1/M, z28.b, z13.b\n"
-      ".inst 0xa1ac27a2  // umopa za2.s, p1/M, p1/M, z29.b, z12.b\n"
-      ".inst 0xa1ad27a3  // umopa za3.s, p1/M, p1/M, z29.b, z13.b\n"
+      ".inst 0xa1b12460  // umopa za0.s, p1/M, p1/M, z3.b, z17.b\n"
+      ".inst 0xa1b92461  // umopa za1.s, p1/M, p1/M, z3.b, z25.b\n"
+      ".inst 0xa1b12562  // umopa za2.s, p1/M, p1/M, z11.b, z17.b\n"
+      ".inst 0xa1b92563  // umopa za3.s, p1/M, p1/M, z11.b, z25.b\n"
+      ".inst 0xa1b62680  // umopa za0.s, p1/M, p1/M, z20.b, z22.b\n"
+      ".inst 0xa1b72681  // umopa za1.s, p1/M, p1/M, z20.b, z23.b\n"
+      ".inst 0xa1b62782  // umopa za2.s, p1/M, p1/M, z28.b, z22.b\n"
+      ".inst 0xa1b72783  // umopa za3.s, p1/M, p1/M, z28.b, z23.b\n"
+      ".inst 0xa1b026a0  // umopa za0.s, p1/M, p1/M, z21.b, z16.b\n"
+      ".inst 0xa1b826a1  // umopa za1.s, p1/M, p1/M, z21.b, z24.b\n"
+      ".inst 0xa1b027a2  // umopa za2.s, p1/M, p1/M, z29.b, z16.b\n"
+      ".inst 0xa1b827a3  // umopa za3.s, p1/M, p1/M, z29.b, z24.b\n"
+      ".inst 0xa1a724a0  // umopa za0.s, p1/M, p1/M, z5.b, z7.b\n"
+      ".inst 0xa1af24a1  // umopa za1.s, p1/M, p1/M, z5.b, z15.b\n"
+      ".inst 0xa1a725a2  // umopa za2.s, p1/M, p1/M, z13.b, z7.b\n"
+      ".inst 0xa1af25a3  // umopa za3.s, p1/M, p1/M, z13.b, z15.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
-      ".inst 0xa0400762  // ld1b { z2.b-z3.b }, pn9.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      ".inst 0xa1400773  // ld1b { z19.b, z27.b }, pn9.b/Z, [x27]\n"
+      "subs x21, x21, #0x1\n"
       "addvl x27, x27, #2\n"
       ".inst 0xa04006f0  // ld1b { z16.b-z17.b }, pn9.b/Z, [x23]\n"
       "addvl x23, x23, #2\n"
-      ".inst 0xa1b02440  // umopa za0.s, p1/M, p1/M, z2.b, z16.b\n"
-      ".inst 0xa1b12441  // umopa za1.s, p1/M, p1/M, z2.b, z17.b\n"
-      ".inst 0xa1b02462  // umopa za2.s, p1/M, p1/M, z3.b, z16.b\n"
-      ".inst 0xa1b12463  // umopa za3.s, p1/M, p1/M, z3.b, z17.b\n"
+      ".inst 0xa1b02660  // umopa za0.s, p1/M, p1/M, z19.b, z16.b\n"
+      ".inst 0xa1b12661  // umopa za1.s, p1/M, p1/M, z19.b, z17.b\n"
+      ".inst 0xa1b02762  // umopa za2.s, p1/M, p1/M, z27.b, z16.b\n"
+      ".inst 0xa1b12763  // umopa za3.s, p1/M, p1/M, z27.b, z17.b\n"
       "bgt 9b\n"
       "10:"  // K oddments: End
-      ".inst 0xa1404767  // ld1w { z7.s, z15.s }, pn9.b/Z, [x27]\n"
+      ".inst 0xa040476e  // ld1w { z14.s-z15.s }, pn9.b/Z, [x27]\n"
       "addvl x27, x27, #2\n"
-      ".inst 0xc09124e0  // addva za0.s, p1/M, p1/M, z7.s\n"
-      ".inst 0xc09124e1  // addva za1.s, p1/M, p1/M, z7.s\n"
+      ".inst 0xc09125c0  // addva za0.s, p1/M, p1/M, z14.s\n"
+      ".inst 0xc09125c1  // addva za1.s, p1/M, p1/M, z14.s\n"
       ".inst 0xc09125e2  // addva za2.s, p1/M, p1/M, z15.s\n"
       ".inst 0xc09125e3  // addva za3.s, p1/M, p1/M, z15.s\n"
       "tbz x16, #1, 14f\n"
@@ -234,25 +233,25 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
-      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
-      ".inst 0xa041c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
-      ".inst 0xa042c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c5f8  // ld1w { z24.s-z27.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
-      ".inst 0xa060c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14]\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xa061c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840703  // mova za3h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa040c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      ".inst 0xa041c5fc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840781  // mova za1h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
+      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa043c5fc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840783  // mova za3h.s[x12], { z28.s-z31.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa060c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c5c4  // st1w { z4.s-z7.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 24f\n"
@@ -260,71 +259,71 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c5dc  // st1w { z28.s-z31.s }, pn9.b, [x14]\n"
+      ".inst 0xc086040c  // mova { z12.s-z15.s }, za0h.s[x12]\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      ".inst 0xa060c5cc  // st1w { z12.s-z15.s }, pn9.b, [x14]\n"
+      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
+      ".inst 0xa061c5c8  // st1w { z8.s-z11.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c5c0  // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c5d0  // st1w { z16.s-z19.s }, pn9.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c5d4  // st1w { z20.s-z23.s }, pn9.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 24f\n"
       "14:"  // Store to output array
       "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "add x26, x26, x10\n"  // C += n
       "sub x25, x13, x11\n"
-      "ld1rw { z8.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
+      "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
       "ldr x24, [%x[args], %[offsetof_ldcb]]\n"
-      "ld1rw { z9.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
-      "ld1rw { z10.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
-      "add x26, x26, x10\n"  // C += n
-      "ld1rw { z11.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
       "madd x26, x11, x24, x26\n"  // C += m * ldc
-      "ld1rw { z15.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
-      "ld1rw { z21.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
-      "ld1rw { z20.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
+      "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
+      "ld1rw { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z3.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
+      "ld1rw { z14.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
+      "ld1rw { z25.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
+      "ld1rw { z24.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
       "tbz x16, #2, 15f\n"
-      "ldr w22, [%x[args], %[offsetof_n_0]]\n"
-      "ldr x21, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "ldr w21, [%x[args], %[offsetof_n_0]]\n"
+      "add x21, x21, x10\n"
+      "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "add x20, x20, x21, LSL #2\n"
+      ".inst 0xa0404280  // ld1w { z0.s-z1.s }, p8/Z, [x20]\n"
       "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n"
-      "add x22, x22, x10\n"
-      "add x21, x21, x22, LSL #2\n"
-      "add x20, x20, x22, LSL #2\n"
-      ".inst 0xa04042a8  // ld1w { z8.s-z9.s }, p8/Z, [x21]\n"
-      ".inst 0xa040428a  // ld1w { z10.s-z11.s }, p8/Z, [x20]\n"
+      "add x20, x20, x21, LSL #2\n"
+      ".inst 0xa0404282  // ld1w { z2.s-z3.s }, p8/Z, [x20]\n"
       "15:"  // Store to output array: Load per-channel parameters: End
       "cntw x23\n"
       "whilelt p0.h, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 17f\n"
       "16:"  // Store to output array: Accumulator row 0 loop
-      ".inst 0xc086041c  // mova { z28.s-z31.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xc1a8ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z8.s\n"
+      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0860428  // mova { z8.s-z11.s }, za1h.s[x12]\n"
+      ".inst 0xc1a0ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
+      ".inst 0xc1a1ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a9ac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z9.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1aaaa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z10.s\n"
-      ".inst 0xc1abaa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z11.s\n"
-      ".inst 0xc1afab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z15.s\n"
-      ".inst 0xc1afab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z15.s\n"
-      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cea0  // sclamp { z0.s-z3.s }, z21.s, z20.s\n"
-      "uzp1 z19.h, z28.h, z0.h\n"
-      "uzp1 z18.h, z29.h, z1.h\n"
-      "uzp1 z17.h, z30.h, z2.h\n"
-      "uzp1 z16.h, z31.h, z3.h\n"
-      "st1b { z19.h }, p0, [x26]\n"
+      ".inst 0xc1a2aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      ".inst 0xc1a3aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z3.s\n"
+      ".inst 0xc1aeab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
+      ".inst 0xc1aeab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z14.s\n"
+      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      ".inst 0xc1b8cf28  // sclamp { z8.s-z11.s }, z25.s, z24.s\n"
+      "uzp1 z16.h, z4.h, z8.h\n"
+      "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z18.h }, p0, [x26]\n"
+      "uzp1 z16.h, z5.h, z9.h\n"
+      "uzp1 z17.h, z6.h, z10.h\n"
+      "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "uzp1 z16.h, z7.h, z11.h\n"
       "st1b { z17.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1b { z16.h }, p0, [x26]\n"
@@ -332,59 +331,60 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
       "blt 16b\n"
       "17:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 18f\n"
-      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
-      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
-      ".inst 0xc1a8ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n"
+      ".inst 0xc0860408  // mova { z8.s-z11.s }, za0h.s[x12]\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc1a0ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z0.s\n"
+      ".inst 0xc1a1ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a9ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z9.s\n"
-      ".inst 0xc1aaaa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n"
-      ".inst 0xc1abaa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z11.s\n"
-      ".inst 0xc1afab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z15.s\n"
-      ".inst 0xc1afab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z15.s\n"
-      ".inst 0xc1b4cea4  // sclamp { z4.s-z7.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      "uzp1 z16.h, z4.h, z28.h\n"
+      ".inst 0xc1a2aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n"
+      ".inst 0xc1a3aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n"
+      ".inst 0xc1aeab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z14.s\n"
+      ".inst 0xc1aeab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
+      ".inst 0xc1b8cf28  // sclamp { z8.s-z11.s }, z25.s, z24.s\n"
+      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      "uzp1 z16.h, z8.h, z4.h\n"
       "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 18f\n"
       "subs x20, x20, #0x1\n"
-      "uzp1 z16.h, z5.h, z29.h\n"
+      "uzp1 z16.h, z9.h, z5.h\n"
       "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 18f\n"
-      "uzp1 z16.h, z6.h, z30.h\n"
+      "uzp1 z16.h, z10.h, z6.h\n"
       "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "18:"  // Store to output array: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 22f\n"
+      "whilelt p0.h, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x23, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 20f\n"
       "19:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860440  // mova { z0.s-z3.s }, za2h.s[x12]\n"
-      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
-      ".inst 0xc1a8ac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z8.s\n"
+      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
+      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
+      ".inst 0xc1a0ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
+      ".inst 0xc1a1ac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z1.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a9ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z9.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1aaaa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z10.s\n"
-      ".inst 0xc1abaa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z11.s\n"
-      ".inst 0xc1afab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z15.s\n"
-      ".inst 0xc1afab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z15.s\n"
-      ".inst 0xc1b4cea0  // sclamp { z0.s-z3.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
-      "uzp1 z19.h, z0.h, z28.h\n"
-      "uzp1 z18.h, z1.h, z29.h\n"
-      "uzp1 z17.h, z2.h, z30.h\n"
-      "uzp1 z16.h, z3.h, z31.h\n"
-      "st1b { z19.h }, p0, [x26]\n"
+      ".inst 0xc1a2aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      ".inst 0xc1a3aa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z3.s\n"
+      ".inst 0xc1aeab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
+      ".inst 0xc1aeab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z14.s\n"
+      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      ".inst 0xc1b8cf34  // sclamp { z20.s-z23.s }, z25.s, z24.s\n"
+      "uzp1 z16.h, z4.h, z20.h\n"
+      "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z18.h }, p0, [x26]\n"
+      "uzp1 z16.h, z5.h, z21.h\n"
+      "uzp1 z17.h, z6.h, z22.h\n"
+      "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
+      "uzp1 z16.h, z7.h, z23.h\n"
       "st1b { z17.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1b { z16.h }, p0, [x26]\n"
@@ -394,15 +394,15 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
       "cbz x20, 21f\n"
       ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
       ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xc1a8ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n"
+      ".inst 0xc1a0ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
+      ".inst 0xc1a1ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a9ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z9.s\n"
-      ".inst 0xc1aaaa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z10.s\n"
-      ".inst 0xc1abaa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z11.s\n"
-      ".inst 0xc1afab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z15.s\n"
-      ".inst 0xc1afab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z15.s\n"
-      ".inst 0xc1b4cea4  // sclamp { z4.s-z7.s }, z21.s, z20.s\n"
-      ".inst 0xc1b4ceb0  // sclamp { z16.s-z19.s }, z21.s, z20.s\n"
+      ".inst 0xc1a2aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      ".inst 0xc1a3aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z3.s\n"
+      ".inst 0xc1aeab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
+      ".inst 0xc1aeab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z14.s\n"
+      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
       "uzp1 z16.h, z4.h, z16.h\n"
       "st1b { z16.h }, p0, [x26]\n"
       "add x26, x26, x24\n"
@@ -420,25 +420,25 @@ void sme2_interleaved_nomerge_u8q_mopa_2VLx2VL(const uint8_t *const A, const uin
       "mov x12, #0x0\n"
       "cntw x20\n"
       "23:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c5e0  // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n"
-      ".inst 0xa041c5ec  // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c5fc  // ld1w { z28.s-z31.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa041c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c5f0  // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c5e4  // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840400  // mova za0h.s[x12], { z0.s-z3.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
       ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 23b\n"
       "24:"  // End block
       "incw x10, ALL, MUL #2\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #2\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL.hpp
index dbf62cbb8a..0665468517 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL.hpp
@@ -37,8 +37,7 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
 class cls_sme2_interleaved_nomerge_u8q_mopa_4VLx1VL
 {
 public:
-  typedef uint8_t lhs_operand_type;
-  typedef uint8_t rhs_operand_type;
+  typedef uint8_t operand_type;
   typedef uint8_t result_type;
 
   typedef void (*kern_type)(const uint8_t *const A, const uint8_t *const B, uint8_t *const C, int ldc, const int M, const int N, const int K, const int32_t *const bias, const Requantize32 &rq, const int n_0, bool accumulate, int32_t *const accumulator_buffer);
@@ -61,7 +60,7 @@ public:
 
   static constexpr bool supports_accumulate()
   {
-    return true;
+    return false;
   }
 
   static constexpr bool supports_bias()
@@ -82,7 +81,7 @@ public:
   // Default to the generic kernel
   kern_type kernel = sme2_interleaved_nomerge_u8q_mopa_4VLx1VL;
 
-  StdTransformsSME<lhs_operand_type, result_type, 4, 1, 4, true> transforms = {};
+  StdTransformsSME<operand_type, result_type, 4, 1, 4, true> transforms = {};
 
   cls_sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const CPUInfo *)
   {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/generic.cpp
index 775a3bf3d2..2239b3f1be 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_interleaved_nomerge_u8q_mopa_4VLx1VL/generic.cpp
@@ -49,7 +49,7 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
         B(B), kstride_bytes(roundup(K, 4) * sizeof(uint8_t)),
         C(C), ldcb(ldc * sizeof(uint8_t)),
         M(M), N(N), K(K),
-        min(0), max(0),
+        n_loops(((K / 4) - 1) / 2), n_tail_iters(((K / 4) - 1) % 2),
 
         bias(bias), n_0(n_0),
         accumulator_buffer(accumulator_buffer),
@@ -74,14 +74,13 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
     const long kstride_bytes;
     uint8_t *const C;
     const long ldcb;
-    const long M, N, K;
-    int32_t min;
-    int32_t max;
+    const long M, N, K, n_loops, n_tail_iters;
+    int32_t min = std::numeric_limits<uint8_t>::min();
+    int32_t max = std::numeric_limits<uint8_t>::max();
 
     const int32_t *const bias;
     const int n_0;
 
-
     int32_t *const accumulator_buffer;
     uint64_t flags;
   };
@@ -100,17 +99,17 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
       "mov x12, #0x0\n"
       "cntw x20\n"
       "1:"  // Initial accumulator load from buffer: Loop
-      ".inst 0xa040c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840500  // mova za0h.s[x12], { z8.s-z11.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840402  // mova za2h.s[x12], { z0.s-z3.s }\n"
-      ".inst 0xc0840483  // mova za3h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa040c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840600  // mova za0h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa041c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840701  // mova za1h.s[x12], { z24.s-z27.s }\n"
+      ".inst 0xa042c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa043c1f4  // ld1w { z20.s-z23.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840683  // mova za3h.s[x12], { z20.s-z23.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 1b\n"
       "2:"  // Initial accumulator load from buffer: End
       "ldr w13, [%x[args], %[offsetof_M]]\n"
@@ -125,95 +124,95 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
       "ldr x20, [%x[args], %[offsetof_bias]]\n"
       ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
       "cbz x20, 5f\n"
-      "ld1w { z6.s }, p0/Z, [x20, x10, LSL #2]\n"
-      ".inst 0xc09024c0  // addha za0.s, p1/M, p1/M, z6.s\n"
-      ".inst 0xc09024c1  // addha za1.s, p1/M, p1/M, z6.s\n"
-      ".inst 0xc09024c2  // addha za2.s, p1/M, p1/M, z6.s\n"
-      ".inst 0xc09024c3  // addha za3.s, p1/M, p1/M, z6.s\n"
+      "ldnt1w { z8.s }, p0/Z, [x20, x10, LSL #2]\n"
+      ".inst 0xc0902500  // addha za0.s, p1/M, p1/M, z8.s\n"
+      ".inst 0xc0902501  // addha za1.s, p1/M, p1/M, z8.s\n"
+      ".inst 0xc0902502  // addha za2.s, p1/M, p1/M, z8.s\n"
+      ".inst 0xc0902503  // addha za3.s, p1/M, p1/M, z8.s\n"
       "4:"  // Prepare accumulators: Test for last block
       "mov x20, x10\n"
       "mov x21, x11\n"
       "incw x20\n"
       "incw x21, ALL, MUL #4\n"
       "cmp x20, x9\n"
-      "mov x20, x16\n"
       "csel x21, x11, x21, LT\n"
+      "mov x20, x16\n"
       "bfm x16, XZR, #0x0, #0x0  // bfc x16, #0x0, #0x1\n"
       "cmp x21, x13\n"
       "csel x16, x20, x16, LT\n"
       "5:"  // Prepare accumulators: End
       "ldr x20, [%x[args], %[offsetof_K]]\n"
-      "ldr x23, [%x[args], %[offsetof_B]]\n"
-      "ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
       "add x20, x20, #0x3\n"
       "lsr x20, x20, #0x2\n"
-      "lsr x21, x20, #0x2\n"
-      "madd x23, x10, x22, x23\n"  // bptr = B + n * kstride_bytes
-      "and x20, x20, #0x3\n"
-      "cbz x21, 8f\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa1408360  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn8.b/Z, [x27]\n"
-      "ld1b { z29.b }, p1/Z, [x23]\n"
-      ".inst 0xa1418361  // ld1b { z1.b, z5.b, z9.b, z13.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      "ld1b { z19.b }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa1428363  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1b { z20.b }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa0438378  // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      "ldr x23, [%x[args], %[offsetof_B]]\n"
+      "lsr x22, x20, #0x2\n"
+      "and x21, x20, #0x3\n"
+      "ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
+      "madd x23, x10, x20, x23\n"  // bptr = B + n * kstride_bytes
+      "cbz x22, 8f\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa0408364  // ld1b { z4.b-z7.b }, pn8.b/Z, [x27]\n"
+      "ldnt1b { z14.b }, p1/Z, [x23]\n"
+      ".inst 0xa0418374  // ld1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      "ldnt1b { z31.b }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0xa0428378  // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1b { z13.b }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0xa0438368  // ld1b { z8.b-z11.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1b { z31.b }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1b { z29.b }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "ble 7f\n"
       "6:"  // K loop
-      ".inst 0xa1bd2400  // umopa za0.s, p1/M, p1/M, z0.b, z29.b\n"
-      "subs x21, x21, #0x1\n"
-      ".inst 0xa1bd2481  // umopa za1.s, p1/M, p1/M, z4.b, z29.b\n"
-      ".inst 0xa1bd2502  // umopa za2.s, p1/M, p1/M, z8.b, z29.b\n"
-      ".inst 0xa1bd2583  // umopa za3.s, p1/M, p1/M, z12.b, z29.b\n"
-      ".inst 0xa1408360  // ld1b { z0.b, z4.b, z8.b, z12.b }, pn8.b/Z, [x27]\n"
-      ".inst 0xa1b32420  // umopa za0.s, p1/M, p1/M, z1.b, z19.b\n"
-      "ld1b { z29.b }, p1/Z, [x23]\n"
-      ".inst 0xa1b324a1  // umopa za1.s, p1/M, p1/M, z5.b, z19.b\n"
-      ".inst 0xa1b32522  // umopa za2.s, p1/M, p1/M, z9.b, z19.b\n"
-      ".inst 0xa1b325a3  // umopa za3.s, p1/M, p1/M, z13.b, z19.b\n"
-      ".inst 0xa1418361  // ld1b { z1.b, z5.b, z9.b, z13.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
-      ".inst 0xa1b42460  // umopa za0.s, p1/M, p1/M, z3.b, z20.b\n"
-      "ld1b { z19.b }, p1/Z, [x23, #1, MUL VL]\n"
-      ".inst 0xa1b424e1  // umopa za1.s, p1/M, p1/M, z7.b, z20.b\n"
-      ".inst 0xa1b42562  // umopa za2.s, p1/M, p1/M, z11.b, z20.b\n"
-      ".inst 0xa1b425e3  // umopa za3.s, p1/M, p1/M, z15.b, z20.b\n"
-      ".inst 0xa1428363  // ld1b { z3.b, z7.b, z11.b, z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
-      "ld1b { z20.b }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0xa1bf2700  // umopa za0.s, p1/M, p1/M, z24.b, z31.b\n"
-      ".inst 0xa1bf2721  // umopa za1.s, p1/M, p1/M, z25.b, z31.b\n"
-      ".inst 0xa1bf2742  // umopa za2.s, p1/M, p1/M, z26.b, z31.b\n"
-      ".inst 0xa1bf2763  // umopa za3.s, p1/M, p1/M, z27.b, z31.b\n"
-      ".inst 0xa0438378  // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
+      ".inst 0xa1ae2480  // umopa za0.s, p1/M, p1/M, z4.b, z14.b\n"
+      "subs x22, x22, #0x1\n"
+      ".inst 0xa1ae24a1  // umopa za1.s, p1/M, p1/M, z5.b, z14.b\n"
+      ".inst 0xa1ae24c2  // umopa za2.s, p1/M, p1/M, z6.b, z14.b\n"
+      ".inst 0xa1ae24e3  // umopa za3.s, p1/M, p1/M, z7.b, z14.b\n"
+      ".inst 0xa0408364  // ld1b { z4.b-z7.b }, pn8.b/Z, [x27]\n"
+      ".inst 0xa1bf2680  // umopa za0.s, p1/M, p1/M, z20.b, z31.b\n"
+      "ldnt1b { z14.b }, p1/Z, [x23]\n"
+      ".inst 0xa1bf26a1  // umopa za1.s, p1/M, p1/M, z21.b, z31.b\n"
+      ".inst 0xa1bf26c2  // umopa za2.s, p1/M, p1/M, z22.b, z31.b\n"
+      ".inst 0xa1bf26e3  // umopa za3.s, p1/M, p1/M, z23.b, z31.b\n"
+      ".inst 0xa0418374  // ld1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n"
+      ".inst 0xa1ad2700  // umopa za0.s, p1/M, p1/M, z24.b, z13.b\n"
+      "ldnt1b { z31.b }, p1/Z, [x23, #1, MUL VL]\n"
+      ".inst 0xa1ad2721  // umopa za1.s, p1/M, p1/M, z25.b, z13.b\n"
+      ".inst 0xa1ad2742  // umopa za2.s, p1/M, p1/M, z26.b, z13.b\n"
+      ".inst 0xa1ad2763  // umopa za3.s, p1/M, p1/M, z27.b, z13.b\n"
+      ".inst 0xa0428378  // ld1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n"
+      "ldnt1b { z13.b }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0xa1bd2500  // umopa za0.s, p1/M, p1/M, z8.b, z29.b\n"
+      ".inst 0xa1bd2521  // umopa za1.s, p1/M, p1/M, z9.b, z29.b\n"
+      ".inst 0xa1bd2542  // umopa za2.s, p1/M, p1/M, z10.b, z29.b\n"
+      ".inst 0xa1bd2563  // umopa za3.s, p1/M, p1/M, z11.b, z29.b\n"
+      ".inst 0xa0438368  // ld1b { z8.b-z11.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n"
       "addvl x27, x27, #16\n"
-      "ld1b { z31.b }, p1/Z, [x23, #3, MUL VL]\n"
+      "ldnt1b { z29.b }, p1/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
       "bgt 6b\n"
       "7:"  // K loop tail
-      ".inst 0xa1bd2400  // umopa za0.s, p1/M, p1/M, z0.b, z29.b\n"
-      ".inst 0xa1bd2481  // umopa za1.s, p1/M, p1/M, z4.b, z29.b\n"
-      ".inst 0xa1bd2502  // umopa za2.s, p1/M, p1/M, z8.b, z29.b\n"
-      ".inst 0xa1bd2583  // umopa za3.s, p1/M, p1/M, z12.b, z29.b\n"
-      ".inst 0xa1b32420  // umopa za0.s, p1/M, p1/M, z1.b, z19.b\n"
-      ".inst 0xa1b324a1  // umopa za1.s, p1/M, p1/M, z5.b, z19.b\n"
-      ".inst 0xa1b32522  // umopa za2.s, p1/M, p1/M, z9.b, z19.b\n"
-      ".inst 0xa1b325a3  // umopa za3.s, p1/M, p1/M, z13.b, z19.b\n"
-      ".inst 0xa1b42460  // umopa za0.s, p1/M, p1/M, z3.b, z20.b\n"
-      ".inst 0xa1b424e1  // umopa za1.s, p1/M, p1/M, z7.b, z20.b\n"
-      ".inst 0xa1b42562  // umopa za2.s, p1/M, p1/M, z11.b, z20.b\n"
-      ".inst 0xa1b425e3  // umopa za3.s, p1/M, p1/M, z15.b, z20.b\n"
-      ".inst 0xa1bf2700  // umopa za0.s, p1/M, p1/M, z24.b, z31.b\n"
-      ".inst 0xa1bf2721  // umopa za1.s, p1/M, p1/M, z25.b, z31.b\n"
-      ".inst 0xa1bf2742  // umopa za2.s, p1/M, p1/M, z26.b, z31.b\n"
-      ".inst 0xa1bf2763  // umopa za3.s, p1/M, p1/M, z27.b, z31.b\n"
+      ".inst 0xa1ae2480  // umopa za0.s, p1/M, p1/M, z4.b, z14.b\n"
+      ".inst 0xa1ae24a1  // umopa za1.s, p1/M, p1/M, z5.b, z14.b\n"
+      ".inst 0xa1ae24c2  // umopa za2.s, p1/M, p1/M, z6.b, z14.b\n"
+      ".inst 0xa1ae24e3  // umopa za3.s, p1/M, p1/M, z7.b, z14.b\n"
+      ".inst 0xa1bf2680  // umopa za0.s, p1/M, p1/M, z20.b, z31.b\n"
+      ".inst 0xa1bf26a1  // umopa za1.s, p1/M, p1/M, z21.b, z31.b\n"
+      ".inst 0xa1bf26c2  // umopa za2.s, p1/M, p1/M, z22.b, z31.b\n"
+      ".inst 0xa1bf26e3  // umopa za3.s, p1/M, p1/M, z23.b, z31.b\n"
+      ".inst 0xa1ad2700  // umopa za0.s, p1/M, p1/M, z24.b, z13.b\n"
+      ".inst 0xa1ad2721  // umopa za1.s, p1/M, p1/M, z25.b, z13.b\n"
+      ".inst 0xa1ad2742  // umopa za2.s, p1/M, p1/M, z26.b, z13.b\n"
+      ".inst 0xa1ad2763  // umopa za3.s, p1/M, p1/M, z27.b, z13.b\n"
+      ".inst 0xa1bd2500  // umopa za0.s, p1/M, p1/M, z8.b, z29.b\n"
+      ".inst 0xa1bd2521  // umopa za1.s, p1/M, p1/M, z9.b, z29.b\n"
+      ".inst 0xa1bd2542  // umopa za2.s, p1/M, p1/M, z10.b, z29.b\n"
+      ".inst 0xa1bd2563  // umopa za3.s, p1/M, p1/M, z11.b, z29.b\n"
       "8:"  // K oddments
-      "cbz x20, 10f\n"
+      "cbz x21, 10f\n"
       "9:"  // K oddments: Loop
       ".inst 0xa1408372  // ld1b { z18.b, z22.b, z26.b, z30.b }, pn8.b/Z, [x27]\n"
-      "subs x20, x20, #0x1\n"
+      "subs x21, x21, #0x1\n"
       "addvl x27, x27, #4\n"
       "ld1b { z15.b }, p1/Z, [x23]\n"
       "addvl x23, x23, #1\n"
@@ -234,25 +233,25 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
       "mov x12, #0x0\n"
       "cntw x20\n"
       "11:"  // Store to partial result buffer: Store and refill: Loop
-      ".inst 0xa040c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xc0860414  // mova { z20.s-z23.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xa041c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
-      ".inst 0xc0860478  // mova { z24.s-z27.s }, za3h.s[x12]\n"
-      ".inst 0xa042c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
-      ".inst 0xa043c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840480  // mova za0h.s[x12], { z4.s-z7.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xa060c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14]\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
-      ".inst 0xa061c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0x4, MUL VL]\n"
-      ".inst 0xc0840603  // mova za3h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa040c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc0840580  // mova za0h.s[x12], { z12.s-z15.s }\n"
+      ".inst 0xc086043c  // mova { z28.s-z31.s }, za1h.s[x12]\n"
+      ".inst 0xa041c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
+      ".inst 0xc086046c  // mova { z12.s-z15.s }, za3h.s[x12]\n"
+      ".inst 0xa042c1e4  // ld1w { z4.s-z7.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840482  // mova za2h.s[x12], { z4.s-z7.s }\n"
+      ".inst 0xa043c1e8  // ld1w { z8.s-z11.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
+      ".inst 0xc0840503  // mova za3h.s[x12], { z8.s-z11.s }\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa062c1c8  // st1w { z8.s-z11.s }, pn8.b, [x14, #0x8, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa063c1d8  // st1w { z24.s-z27.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa060c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14]\n"
+      "addvl x15, x15, #16\n"
+      ".inst 0xa061c1dc  // st1w { z28.s-z31.s }, pn8.b, [x14, #0x4, MUL VL]\n"
+      ".inst 0xa062c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 11b\n"
       "b 30f\n"
@@ -260,56 +259,56 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
       "mov x12, #0x0\n"
       "cntw x20\n"
       "13:"  // Store to partial result buffer: Store only: Loop
-      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
-      ".inst 0xc0860420  // mova { z0.s-z3.s }, za1h.s[x12]\n"
-      ".inst 0xc0860454  // mova { z20.s-z23.s }, za2h.s[x12]\n"
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
-      ".inst 0xa060c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14]\n"
+      ".inst 0xc0860400  // mova { z0.s-z3.s }, za0h.s[x12]\n"
+      ".inst 0xc086042c  // mova { z12.s-z15.s }, za1h.s[x12]\n"
+      ".inst 0xa060c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14]\n"
+      ".inst 0xc0860444  // mova { z4.s-z7.s }, za2h.s[x12]\n"
+      ".inst 0xc0860460  // mova { z0.s-z3.s }, za3h.s[x12]\n"
+      ".inst 0xa061c1cc  // st1w { z12.s-z15.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xa061c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0x4, MUL VL]\n"
       "cmp x12, x20\n"
-      ".inst 0xa062c1d4  // st1w { z20.s-z23.s }, pn8.b, [x14, #0x8, MUL VL]\n"
-      ".inst 0xa063c1d0  // st1w { z16.s-z19.s }, pn8.b, [x14, #0xc, MUL VL]\n"
+      ".inst 0xa062c1c4  // st1w { z4.s-z7.s }, pn8.b, [x14, #0x8, MUL VL]\n"
+      ".inst 0xa063c1c0  // st1w { z0.s-z3.s }, pn8.b, [x14, #0xc, MUL VL]\n"
       "addvl x14, x14, #16\n"
       "blt 13b\n"
       "b 30f\n"
       "14:"  // Store to output array
       "ldr x26, [%x[args], %[offsetof_C]]\n"
+      "add x26, x26, x10\n"  // C += n
       "sub x25, x13, x11\n"
       "ld1rw { z2.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_mul]]\n"
       "ldr x24, [%x[args], %[offsetof_ldcb]]\n"
+      "madd x26, x11, x24, x26\n"  // C += m * ldc
       "ld1rw { z1.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_per_layer_right_shift]]\n"
       "ld1rw { z0.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_c_offset]]\n"
-      "add x26, x26, x10\n"  // C += n
-      "ld1rw { z25.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
-      "madd x26, x11, x24, x26\n"  // C += m * ldc
-      "ld1rw { z24.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
+      "ld1rw { z21.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_minval]]\n"
+      "ld1rw { z20.s }, p1/Z, [%x[rq], %[offsetof_Requantize32_maxval]]\n"
       "tbz x16, #2, 15f\n"
-      "ldr w22, [%x[args], %[offsetof_n_0]]\n"
-      "ldr x21, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "ldr w21, [%x[args], %[offsetof_n_0]]\n"
+      "add x21, x21, x10\n"
+      "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_muls]]\n"
+      "add x20, x20, x21, LSL #2\n"
+      "ld1w { z2.s }, p0/Z, [x20]\n"
       "ldr x20, [%x[rq], %[offsetof_Requantize32_per_channel_right_shifts]]\n"
-      "add x22, x22, x10\n"
-      "add x21, x21, x22, LSL #2\n"
-      "add x20, x20, x22, LSL #2\n"
-      "ld1w { z2.s }, p0/Z, [x21]\n"
+      "add x20, x20, x21, LSL #2\n"
       "ld1w { z1.s }, p0/Z, [x20]\n"
       "15:"  // Store to output array: Load per-channel parameters: End
       "cntw x23\n"
       "whilelt p0.s, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 17f\n"
       "16:"  // Store to output array: Accumulator row 0 loop
       ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
-      "add x12, x12, #0x4\n"
       ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
-      "cmp x12, x21, LSL #2\n"
+      "add x12, x12, #0x4\n"
       ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
+      "cmp x12, x21, LSL #2\n"
       ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
+      ".inst 0xc1b4ceb0  // sclamp { z16.s-z19.s }, z21.s, z20.s\n"
       "st1b { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "st1b { z17.s }, p0, [x26]\n"
@@ -321,55 +320,56 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
       "blt 16b\n"
       "17:"  // Store to output array: Accumulator row 0 oddments
       "cbz x20, 18f\n"
-      ".inst 0xc0860404  // mova { z4.s-z7.s }, za0h.s[x12]\n"
+      ".inst 0xc0860410  // mova { z16.s-z19.s }, za0h.s[x12]\n"
+      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a2ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
-      ".inst 0xc1a1aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
-      ".inst 0xc1a0ab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
-      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
-      "st1b { z4.s }, p0, [x26]\n"
+      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
+      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
+      ".inst 0xc1b4ceb0  // sclamp { z16.s-z19.s }, z21.s, z20.s\n"
+      "st1b { z16.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 18f\n"
       "subs x20, x20, #0x1\n"
-      "st1b { z5.s }, p0, [x26]\n"
+      "st1b { z17.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 18f\n"
-      "st1b { z6.s }, p0, [x26]\n"
+      "st1b { z18.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "18:"  // Store to output array: Accumulator row 0 oddments: End
       "subs x25, x25, x22\n"
       "beq 28f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 20f\n"
       "19:"  // Store to output array: Accumulator row 1 loop
-      ".inst 0xc0860430  // mova { z16.s-z19.s }, za1h.s[x12]\n"
+      ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
+      ".inst 0xc1a2ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
+      ".inst 0xc1a1aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
-      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
-      "st1b { z16.s }, p0, [x26]\n"
+      ".inst 0xc1a0ab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
+      ".inst 0xc1b4cea4  // sclamp { z4.s-z7.s }, z21.s, z20.s\n"
+      "st1b { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z17.s }, p0, [x26]\n"
+      "st1b { z5.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z18.s }, p0, [x26]\n"
+      "st1b { z6.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z19.s }, p0, [x26]\n"
+      "st1b { z7.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 19b\n"
       "20:"  // Store to output array: Accumulator row 1 oddments
       "cbz x20, 21f\n"
       ".inst 0xc0860424  // mova { z4.s-z7.s }, za1h.s[x12]\n"
-      "subs x20, x20, #0x1\n"
       ".inst 0xc1a2ac04  // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n"
+      "subs x20, x20, #0x1\n"
       ".inst 0xc1a1aa24  // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n"
       ".inst 0xc1a0ab04  // add { z4.s-z7.s }, { z4.s-z7.s }, z0.s\n"
-      ".inst 0xc1b8cf24  // sclamp { z4.s-z7.s }, z25.s, z24.s\n"
+      ".inst 0xc1b4cea4  // sclamp { z4.s-z7.s }, z21.s, z20.s\n"
       "st1b { z4.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 21f\n"
@@ -382,113 +382,115 @@ void sme2_interleaved_nomerge_u8q_mopa_4VLx1VL(const uint8_t *const A, const uin
       "21:"  // Store to output array: Accumulator row 1 oddments: End
       "subs x25, x25, x22\n"
       "beq 28f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x22, x25, x23, LT\n"
       "lsr x21, x22, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x22, #0x3\n"
       "cbz x21, 23f\n"
       "22:"  // Store to output array: Accumulator row 2 loop
-      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc0860448  // mova { z8.s-z11.s }, za2h.s[x12]\n"
+      ".inst 0xc1a2ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a2ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
+      ".inst 0xc1a1aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1a1aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n"
-      ".inst 0xc1a0ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n"
-      ".inst 0xc1b8cf2c  // sclamp { z12.s-z15.s }, z25.s, z24.s\n"
-      "st1b { z12.s }, p0, [x26]\n"
+      ".inst 0xc1a0ab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z0.s\n"
+      ".inst 0xc1b4cea8  // sclamp { z8.s-z11.s }, z21.s, z20.s\n"
+      "st1b { z8.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z13.s }, p0, [x26]\n"
+      "st1b { z9.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z14.s }, p0, [x26]\n"
+      "st1b { z10.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z15.s }, p0, [x26]\n"
+      "st1b { z11.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 22b\n"
       "23:"  // Store to output array: Accumulator row 2 oddments
       "cbz x20, 24f\n"
-      ".inst 0xc0860450  // mova { z16.s-z19.s }, za2h.s[x12]\n"
+      ".inst 0xc086044c  // mova { z12.s-z15.s }, za2h.s[x12]\n"
+      ".inst 0xc1a2ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
-      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
-      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
-      "st1b { z16.s }, p0, [x26]\n"
+      ".inst 0xc1a1aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n"
+      ".inst 0xc1a0ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n"
+      ".inst 0xc1b4ceac  // sclamp { z12.s-z15.s }, z21.s, z20.s\n"
+      "st1b { z12.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 24f\n"
       "subs x20, x20, #0x1\n"
-      "st1b { z17.s }, p0, [x26]\n"
+      "st1b { z13.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 24f\n"
-      "st1b { z18.s }, p0, [x26]\n"
+      "st1b { z14.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "24:"  // Store to output array: Accumulator row 2 oddments: End
       "subs x25, x25, x22\n"
       "beq 28f\n"
+      "whilelt p0.s, x10, x9\n"
       "cmp x25, x23\n"
-      "mov x12, #0x0\n"
       "csel x20, x25, x23, LT\n"
       "lsr x21, x20, #0x2\n"
+      "mov x12, #0x0\n"
       "and x20, x20, #0x3\n"
       "cbz x21, 26f\n"
       "25:"  // Store to output array: Accumulator row 3 loop
-      ".inst 0xc0860470  // mova { z16.s-z19.s }, za3h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xc1a2ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n"
       "add x12, x12, #0x4\n"
-      ".inst 0xc1a2ac10  // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n"
+      ".inst 0xc1a1aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n"
       "cmp x12, x21, LSL #2\n"
-      ".inst 0xc1a1aa30  // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n"
-      ".inst 0xc1a0ab10  // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n"
-      ".inst 0xc1b8cf30  // sclamp { z16.s-z19.s }, z25.s, z24.s\n"
-      "st1b { z16.s }, p0, [x26]\n"
+      ".inst 0xc1a0ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n"
+      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
+      "st1b { z28.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z17.s }, p0, [x26]\n"
+      "st1b { z29.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z18.s }, p0, [x26]\n"
+      "st1b { z30.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
-      "st1b { z19.s }, p0, [x26]\n"
+      "st1b { z31.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "blt 25b\n"
       "26:"  // Store to output array: Accumulator row 3 oddments
       "cbz x20, 27f\n"
-      ".inst 0xc0860474  // mova { z20.s-z23.s }, za3h.s[x12]\n"
+      ".inst 0xc086047c  // mova { z28.s-z31.s }, za3h.s[x12]\n"
+      ".inst 0xc1a2ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n"
       "subs x20, x20, #0x1\n"
-      ".inst 0xc1a2ac14  // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z2.s\n"
-      ".inst 0xc1a1aa34  // srshl { z20.s-z23.s }, { z20.s-z23.s }, z1.s\n"
-      ".inst 0xc1a0ab14  // add { z20.s-z23.s }, { z20.s-z23.s }, z0.s\n"
-      ".inst 0xc1b8cf34  // sclamp { z20.s-z23.s }, z25.s, z24.s\n"
-      "st1b { z20.s }, p0, [x26]\n"
+      ".inst 0xc1a1aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n"
+      ".inst 0xc1a0ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n"
+      ".inst 0xc1b4cebc  // sclamp { z28.s-z31.s }, z21.s, z20.s\n"
+      "st1b { z28.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 27f\n"
       "subs x20, x20, #0x1\n"
-      "st1b { z21.s }, p0, [x26]\n"
+      "st1b { z29.s }, p0, [x26]\n"
       "add x26, x26, x24\n"
       "beq 27f\n"
-      "st1b { z22.s }, p0, [x26]\n"
+      "st1b { z30.s }, p0, [x26]\n"
       "27:"  // Store to output array: Accumulator row 3 oddments: End
       "28:"  // Store to output array: End
       "tbz x16, #0, 30f\n"
       "mov x12, #0x0\n"
       "cntw x20\n"
       "29:"  // Store to output array: Refill accumulators: Loop
-      ".inst 0xa040c1f8  // ld1w { z24.s-z27.s }, pn8.b/Z, [x15]\n"
-      ".inst 0xa041c1ec  // ld1w { z12.s-z15.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
-      ".inst 0xa042c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xa040c1fc  // ld1w { z28.s-z31.s }, pn8.b/Z, [x15]\n"
+      ".inst 0xc0840780  // mova za0h.s[x12], { z28.s-z31.s }\n"
+      ".inst 0xa041c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x4, MUL VL]\n"
+      ".inst 0xc0840601  // mova za1h.s[x12], { z16.s-z19.s }\n"
+      ".inst 0xa042c1f0  // ld1w { z16.s-z19.s }, pn8.b/Z, [x15, #0x8, MUL VL]\n"
+      ".inst 0xc0840602  // mova za2h.s[x12], { z16.s-z19.s }\n"
       ".inst 0xa043c1e0  // ld1w { z0.s-z3.s }, pn8.b/Z, [x15, #0xc, MUL VL]\n"
-      ".inst 0xc0840700  // mova za0h.s[x12], { z24.s-z27.s }\n"
-      "addvl x15, x15, #16\n"
-      ".inst 0xc0840581  // mova za1h.s[x12], { z12.s-z15.s }\n"
-      ".inst 0xc0840782  // mova za2h.s[x12], { z28.s-z31.s }\n"
       ".inst 0xc0840403  // mova za3h.s[x12], { z0.s-z3.s }\n"
       "add x12, x12, #0x4\n"
       "cmp x12, x20\n"
+      "addvl x15, x15, #16\n"
       "blt 29b\n"
       "30:"  // End block
       "incw x10\n"
       "cmp x10, x9\n"
       "blt 3b\n"
       "incw x11, ALL, MUL #4\n"
-      "mov x10, #0x0\n"
       "cmp x11, x13\n"
+      "mov x10, #0x0\n"
       "mov x28, x27\n"
       "blt 3b\n"
       ".inst 0xd503467f  // SMSTOP\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp
index ac21a980d3..4cf20bef91 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp
@@ -82,7 +82,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 8, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 8, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp
index 61c38db3cb..7fc723ecad 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL/generic.cpp
@@ -50,19 +50,18 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -83,7 +82,6 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -109,15 +107,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "bgt 29f\n"
       "beq 15f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -125,12 +123,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 3f\n"
       "mov x11, x12\n"
       "3:"  // Height 1: B setup done
@@ -145,26 +143,26 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cbz x15, 4f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x15, x15, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "b 6f\n"
       "4:"  // Height 1: no bias
       "tbz %x[flags], #0, 5f\n"
-      "ld1w { z19.s }, p4/Z, [x13]\n"
+      "ld1w { z16.s }, p4/Z, [x13]\n"
       "ld1w { z18.s }, p3/Z, [x13, #1, MUL VL]\n"
+      "zip1 z8.d, z16.d, z12.d\n"
+      "zip2 z12.d, z16.d, z12.d\n"
       "ld1w { z17.s }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
-      "zip1 z8.d, z19.d, z12.d\n"
-      "zip2 z12.d, z19.d, z12.d\n"
       "zip1 z9.d, z18.d, z13.d\n"
       "zip2 z13.d, z18.d, z13.d\n"
       "zip1 z10.d, z17.d, z14.d\n"
@@ -185,8 +183,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "7:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 8f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -202,96 +200,96 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ble 11f\n"
       "10:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
+      "ld1rqh { z20.h }, p0/Z, [x26]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z20.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "trn1 z19.d, z20.d, z18.d\n"
-      "trn2 z20.d, z20.d, z18.d\n"
-      ".inst 0x6471e668  // bfmmla z8.s, z19.h, z17.h\n"
-      "ld1h { z1.h }, p5/Z, [x11]\n"
-      ".inst 0x6470e66c  // bfmmla z12.s, z19.h, z16.h\n"
+      ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
+      ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x11]\n"
       "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n"
-      ".inst 0x6461e669  // bfmmla z9.s, z19.h, z1.h\n"
-      "ld1h { z18.h }, p5/Z, [x10]\n"
-      ".inst 0x6470e66d  // bfmmla z13.s, z19.h, z16.h\n"
+      ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
+      ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6472e66a  // bfmmla z10.s, z19.h, z18.h\n"
+      ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
+      ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x9]\n"
-      ".inst 0x6470e66e  // bfmmla z14.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
-      ".inst 0x6471e66b  // bfmmla z11.s, z19.h, z17.h\n"
+      "trn2 z20.d, z20.d, z19.d\n"
+      ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
+      ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x12, #2, MUL VL]\n"
-      ".inst 0x6470e66f  // bfmmla z15.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       ".inst 0x6471e688  // bfmmla z8.s, z20.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x11, #2, MUL VL]\n"
       ".inst 0x6470e68c  // bfmmla z12.s, z20.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x11, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x11, #3, MUL VL]\n"
-      "addvl x11, x11, #4\n"
       ".inst 0x6471e689  // bfmmla z9.s, z20.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6470e68d  // bfmmla z13.s, z20.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
-      ".inst 0x6471e68a  // bfmmla z10.s, z20.h, z17.h\n"
+      "ld1h { z16.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #3, MUL VL]\n"
+      ".inst 0x6470e68a  // bfmmla z10.s, z20.h, z16.h\n"
+      ".inst 0x6471e68e  // bfmmla z14.s, z20.h, z17.h\n"
       "ld1h { z17.h }, p5/Z, [x9, #2, MUL VL]\n"
-      ".inst 0x6470e68e  // bfmmla z14.s, z20.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6471e68b  // bfmmla z11.s, z20.h, z17.h\n"
       ".inst 0x6470e68f  // bfmmla z15.s, z20.h, z16.h\n"
+      "add x26, x26, #0x10\n"
+      "addvl x12, x12, #4\n"
+      "addvl x11, x11, #4\n"
+      "addvl x10, x10, #4\n"
+      "addvl x9, x9, #4\n"
       "bgt 10b\n"
       "11:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x12]\n"
-      "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
-      "addvl x12, x12, #2\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1h { z17.h }, p5/Z, [x12]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
       ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x11]\n"
       ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x11]\n"
       "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
-      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
       ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
       ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x9]\n"
       ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x9]\n"
       "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
+      "subs x27, x27, #0x4\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
       ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
+      "addvl x12, x12, #2\n"
+      "addvl x11, x11, #2\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
       "ble 12f\n"
-      "ld1h { z16.h }, p5/Z, [x12]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z17.h }, p5/Z, [x12]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
+      ".inst 0x6471e428  // bfmmla z8.s, z1.h, z17.h\n"
+      ".inst 0x6470e42c  // bfmmla z12.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x11]\n"
+      "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n"
+      ".inst 0x6471e429  // bfmmla z9.s, z1.h, z17.h\n"
+      ".inst 0x6470e42d  // bfmmla z13.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x6471e42a  // bfmmla z10.s, z1.h, z17.h\n"
+      ".inst 0x6470e42e  // bfmmla z14.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x9]\n"
+      "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
+      ".inst 0x6471e42b  // bfmmla z11.s, z1.h, z17.h\n"
+      ".inst 0x6470e42f  // bfmmla z15.s, z1.h, z16.h\n"
       "addvl x12, x12, #2\n"
-      ".inst 0x6470e428  // bfmmla z8.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x11]\n"
-      ".inst 0x6471e42c  // bfmmla z12.s, z1.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x11, #1, MUL VL]\n"
       "addvl x11, x11, #2\n"
-      ".inst 0x6470e429  // bfmmla z9.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10]\n"
-      ".inst 0x6471e42d  // bfmmla z13.s, z1.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #1, MUL VL]\n"
       "addvl x10, x10, #2\n"
-      ".inst 0x6470e42a  // bfmmla z10.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x9]\n"
-      ".inst 0x6471e42e  // bfmmla z14.s, z1.h, z17.h\n"
-      "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n"
       "addvl x9, x9, #2\n"
-      ".inst 0x6470e42b  // bfmmla z11.s, z1.h, z16.h\n"
-      ".inst 0x6460e42f  // bfmmla z15.s, z1.h, z0.h\n"
       "12:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -302,14 +300,14 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "uzp1 z10.d, z10.d, z14.d\n"
       "uzp1 z11.d, z11.d, z15.d\n"
       "tbz %x[flags], #1, 13f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
-      "fmin z8.s, p5/M, z8.s, z17.s\n"
-      "fmin z9.s, p5/M, z9.s, z17.s\n"
-      "fmin z10.s, p5/M, z10.s, z17.s\n"
-      "fmin z11.s, p5/M, z11.s, z17.s\n"
+      "fmin z8.s, p5/M, z8.s, z21.s\n"
+      "fmin z9.s, p5/M, z9.s, z21.s\n"
+      "fmin z10.s, p5/M, z10.s, z21.s\n"
+      "fmin z11.s, p5/M, z11.s, z21.s\n"
       "fmax z8.s, p5/M, z8.s, z16.s\n"
       "fmax z9.s, p5/M, z9.s, z16.s\n"
       "fmax z10.s, p5/M, z10.s, z16.s\n"
@@ -327,15 +325,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "b 86f\n"
       "15:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "16:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -343,12 +341,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 17f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 17f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 17f\n"
       "mov x11, x12\n"
       "17:"  // Height 2: B setup done
@@ -363,38 +361,38 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cbz x15, 18f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x15, x15, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "b 20f\n"
       "18:"  // Height 2: no bias
       "tbz %x[flags], #0, 19f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z18.s }, p4/Z, [x13]\n"
-      "ld1w { z16.s }, p3/Z, [x13, #1, MUL VL]\n"
-      "ld1w { z5.s }, p2/Z, [x13, #2, MUL VL]\n"
-      "ld1w { z17.s }, p1/Z, [x13, #3, MUL VL]\n"
       "add x20, x13, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x13]\n"
+      "ld1w { z18.s }, p3/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z17.s }, p2/Z, [x13, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
+      "zip1 z9.d, z18.d, z13.d\n"
+      "zip2 z13.d, z18.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z8.d, z18.d, z12.d\n"
-      "zip2 z12.d, z18.d, z12.d\n"
-      "zip1 z9.d, z16.d, z13.d\n"
-      "zip2 z13.d, z16.d, z13.d\n"
-      "zip1 z10.d, z5.d, z14.d\n"
-      "zip2 z14.d, z5.d, z14.d\n"
-      "zip1 z11.d, z17.d, z15.d\n"
-      "zip2 z15.d, z17.d, z15.d\n"
+      "zip1 z10.d, z17.d, z14.d\n"
+      "zip2 z14.d, z17.d, z14.d\n"
+      "zip1 z11.d, z16.d, z15.d\n"
+      "zip2 z15.d, z16.d, z15.d\n"
       "b 20f\n"
       "19:"  // Height 2: no accumulate
       "mov z8.b, #0x0\n"
@@ -409,8 +407,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "21:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 22f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -429,99 +427,99 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ble 25f\n"
       "24:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
+      "ld1rqh { z20.h }, p0/Z, [x26]\n"
+      "ld1rqh { z19.h }, p0/Z, [x25]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
-      "ld1h { z20.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z19.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqh { z16.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "trn1 z18.d, z19.d, z16.d\n"
-      "trn2 z19.d, z19.d, z16.d\n"
+      "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
       ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
+      ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x11]\n"
-      ".inst 0x6474e64c  // bfmmla z12.s, z18.h, z20.h\n"
       "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n"
       ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
       ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x9]\n"
       ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x9]\n"
       "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
+      "trn2 z20.d, z20.d, z19.d\n"
       ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x12, #2, MUL VL]\n"
       ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x12, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
-      ".inst 0x6471e668  // bfmmla z8.s, z19.h, z17.h\n"
+      ".inst 0x6471e688  // bfmmla z8.s, z20.h, z17.h\n"
+      ".inst 0x6470e68c  // bfmmla z12.s, z20.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x11, #2, MUL VL]\n"
-      ".inst 0x6470e66c  // bfmmla z12.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x11, #3, MUL VL]\n"
-      "addvl x11, x11, #4\n"
-      ".inst 0x6471e669  // bfmmla z9.s, z19.h, z17.h\n"
+      ".inst 0x6471e689  // bfmmla z9.s, z20.h, z17.h\n"
+      ".inst 0x6470e68d  // bfmmla z13.s, z20.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6470e66d  // bfmmla z13.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
-      ".inst 0x6471e66a  // bfmmla z10.s, z19.h, z17.h\n"
+      ".inst 0x6471e68a  // bfmmla z10.s, z20.h, z17.h\n"
+      ".inst 0x6470e68e  // bfmmla z14.s, z20.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x9, #2, MUL VL]\n"
-      ".inst 0x6470e66e  // bfmmla z14.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x9, #3, MUL VL]\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x6471e68b  // bfmmla z11.s, z20.h, z17.h\n"
+      ".inst 0x6470e68f  // bfmmla z15.s, z20.h, z16.h\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "addvl x12, x12, #4\n"
+      "addvl x11, x11, #4\n"
+      "addvl x10, x10, #4\n"
       "addvl x9, x9, #4\n"
-      ".inst 0x6471e66b  // bfmmla z11.s, z19.h, z17.h\n"
-      ".inst 0x6470e66f  // bfmmla z15.s, z19.h, z16.h\n"
       "bgt 24b\n"
       "25:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x12]\n"
-      "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
-      "addvl x12, x12, #2\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
       "ld1rqh { z19.h }, p0/Z, [x25]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1h { z17.h }, p5/Z, [x12]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
       ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x11]\n"
       ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x11]\n"
       "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
-      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
       ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
       ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x9]\n"
       ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x9]\n"
       "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
+      "subs x27, x27, #0x4\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
       ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
-      "ble 26f\n"
-      "ld1h { z16.h }, p5/Z, [x12]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #1, MUL VL]\n"
       "addvl x12, x12, #2\n"
-      ".inst 0x6470e428  // bfmmla z8.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x11]\n"
-      ".inst 0x6471e42c  // bfmmla z12.s, z1.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x11, #1, MUL VL]\n"
       "addvl x11, x11, #2\n"
-      ".inst 0x6470e429  // bfmmla z9.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10]\n"
-      ".inst 0x6471e42d  // bfmmla z13.s, z1.h, z17.h\n"
-      "ld1h { z26.h }, p5/Z, [x10, #1, MUL VL]\n"
       "addvl x10, x10, #2\n"
-      ".inst 0x6470e42a  // bfmmla z10.s, z1.h, z16.h\n"
-      "ld1h { z17.h }, p5/Z, [x9]\n"
-      ".inst 0x647ae42e  // bfmmla z14.s, z1.h, z26.h\n"
-      "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
       "addvl x9, x9, #2\n"
-      ".inst 0x6471e42b  // bfmmla z11.s, z1.h, z17.h\n"
+      "ble 26f\n"
+      "ld1h { z17.h }, p5/Z, [x12]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
+      ".inst 0x6471e428  // bfmmla z8.s, z1.h, z17.h\n"
+      ".inst 0x6470e42c  // bfmmla z12.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x11]\n"
+      "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n"
+      ".inst 0x6471e429  // bfmmla z9.s, z1.h, z17.h\n"
+      ".inst 0x6470e42d  // bfmmla z13.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x6471e42a  // bfmmla z10.s, z1.h, z17.h\n"
+      ".inst 0x6470e42e  // bfmmla z14.s, z1.h, z16.h\n"
+      "ld1h { z22.h }, p5/Z, [x9]\n"
+      "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
+      ".inst 0x6476e42b  // bfmmla z11.s, z1.h, z22.h\n"
       ".inst 0x6470e42f  // bfmmla z15.s, z1.h, z16.h\n"
+      "addvl x12, x12, #2\n"
+      "addvl x11, x11, #2\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
       "26:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -530,17 +528,17 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z7.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
+      "add x25, x13, x20, LSL #2\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x26, x13, x20, LSL #2\n"
       "tbz %x[flags], #1, 27f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z17.s\n"
       "fmin z12.s, p5/M, z12.s, z17.s\n"
@@ -564,10 +562,10 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
       "28:"  // Height 2: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -575,15 +573,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "b 86f\n"
       "29:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "30:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -591,12 +589,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 31f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 31f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 31f\n"
       "mov x11, x12\n"
       "31:"  // Height 3: B setup done
@@ -611,15 +609,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cbz x15, 32f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x15, x15, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -634,36 +632,36 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "32:"  // Height 3: no bias
       "tbz %x[flags], #0, 33f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z26.s }, p4/Z, [x13]\n"
-      "ld1w { z25.s }, p3/Z, [x13, #1, MUL VL]\n"
-      "ld1w { z24.s }, p2/Z, [x13, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
       "add x21, x13, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x13]\n"
+      "ld1w { z17.s }, p3/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x13, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x20]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "zip1 z8.d, z26.d, z12.d\n"
-      "zip2 z12.d, z26.d, z12.d\n"
-      "ld1w { z2.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z9.d, z25.d, z13.d\n"
-      "zip2 z13.d, z25.d, z13.d\n"
-      "zip1 z10.d, z24.d, z14.d\n"
-      "zip2 z14.d, z24.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
-      "zip1 z19.d, z2.d, z23.d\n"
-      "zip2 z23.d, z2.d, z23.d\n"
+      "zip1 z19.d, z24.d, z23.d\n"
+      "zip2 z23.d, z24.d, z23.d\n"
       "b 34f\n"
       "33:"  // Height 3: no accumulate
       "mov z8.b, #0x0\n"
@@ -686,8 +684,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "35:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 36f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -709,130 +707,130 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ble 39f\n"
       "38:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x12]\n"
-      "ld1h { z30.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z29.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
+      "ld1rqh { z30.h }, p0/Z, [x26]\n"
       "ld1rqh { z24.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
       "ld1rqh { z28.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "trn1 z27.d, z29.d, z24.d\n"
-      "trn2 z29.d, z29.d, z24.d\n"
-      "trn1 z26.d, z28.d, z31.d\n"
-      "trn2 z28.d, z28.d, z31.d\n"
+      "trn1 z27.d, z30.d, z24.d\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "ld1h { z25.h }, p5/Z, [x12]\n"
+      "trn1 z26.d, z28.d, z29.d\n"
+      "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n"
       ".inst 0x6479e768  // bfmmla z8.s, z27.h, z25.h\n"
-      ".inst 0x647ee76c  // bfmmla z12.s, z27.h, z30.h\n"
       ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e76c  // bfmmla z12.s, z27.h, z24.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x11]\n"
-      ".inst 0x647ee754  // bfmmla z20.s, z26.h, z30.h\n"
       "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n"
       ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
+      "trn2 z28.d, z28.d, z29.d\n"
       ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x9]\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x12, #2, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
       ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
-      ".inst 0x6479e7a8  // bfmmla z8.s, z29.h, z25.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6479e7c8  // bfmmla z8.s, z30.h, z25.h\n"
       ".inst 0x6479e790  // bfmmla z16.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x11, #2, MUL VL]\n"
-      ".inst 0x6478e7ac  // bfmmla z12.s, z29.h, z24.h\n"
+      "addvl x12, x12, #4\n"
+      ".inst 0x6478e7cc  // bfmmla z12.s, z30.h, z24.h\n"
       ".inst 0x6478e794  // bfmmla z20.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x11, #3, MUL VL]\n"
       "addvl x11, x11, #4\n"
-      ".inst 0x6479e7a9  // bfmmla z9.s, z29.h, z25.h\n"
+      ".inst 0x6479e7c9  // bfmmla z9.s, z30.h, z25.h\n"
       ".inst 0x6479e791  // bfmmla z17.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6478e7ad  // bfmmla z13.s, z29.h, z24.h\n"
+      ".inst 0x6478e7cd  // bfmmla z13.s, z30.h, z24.h\n"
       ".inst 0x6478e795  // bfmmla z21.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
-      ".inst 0x6479e7aa  // bfmmla z10.s, z29.h, z25.h\n"
+      ".inst 0x6479e7ca  // bfmmla z10.s, z30.h, z25.h\n"
       ".inst 0x6479e792  // bfmmla z18.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x9, #2, MUL VL]\n"
-      ".inst 0x6478e7ae  // bfmmla z14.s, z29.h, z24.h\n"
+      ".inst 0x6478e7ce  // bfmmla z14.s, z30.h, z24.h\n"
       ".inst 0x6478e796  // bfmmla z22.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      ".inst 0x6479e7ab  // bfmmla z11.s, z29.h, z25.h\n"
+      ".inst 0x6479e7cb  // bfmmla z11.s, z30.h, z25.h\n"
       ".inst 0x6479e793  // bfmmla z19.s, z28.h, z25.h\n"
-      ".inst 0x6478e7af  // bfmmla z15.s, z29.h, z24.h\n"
+      ".inst 0x6478e7cf  // bfmmla z15.s, z30.h, z24.h\n"
       ".inst 0x6478e797  // bfmmla z23.s, z28.h, z24.h\n"
       "bgt 38b\n"
       "39:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x12]\n"
-      "ld1h { z28.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
-      "addvl x12, x12, #2\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
       "ld1rqh { z24.h }, p0/Z, [x25]\n"
       "ld1rqh { z3.h }, p0/Z, [x24]\n"
       "trn1 z27.d, z1.d, z24.d\n"
       "trn2 z1.d, z1.d, z24.d\n"
-      "trn1 z26.d, z3.d, z29.d\n"
+      "ld1h { z25.h }, p5/Z, [x12]\n"
+      "trn1 z26.d, z3.d, z28.d\n"
+      "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n"
       ".inst 0x6479e768  // bfmmla z8.s, z27.h, z25.h\n"
-      ".inst 0x647ce76c  // bfmmla z12.s, z27.h, z28.h\n"
-      "trn2 z3.d, z3.d, z29.d\n"
       ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e76c  // bfmmla z12.s, z27.h, z24.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x11]\n"
-      ".inst 0x647ce754  // bfmmla z20.s, z26.h, z28.h\n"
       "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
       ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x4\n"
       ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
+      "trn2 z3.d, z3.d, z28.d\n"
       ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x9]\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
+      "addvl x11, x11, #2\n"
       ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
       ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
       ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "ble 40f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
       ".inst 0x6479e428  // bfmmla z8.s, z1.h, z25.h\n"
       ".inst 0x6479e470  // bfmmla z16.s, z3.h, z25.h\n"
-      "ld1h { z25.h }, p5/Z, [x11]\n"
       ".inst 0x6478e42c  // bfmmla z12.s, z1.h, z24.h\n"
       ".inst 0x6478e474  // bfmmla z20.s, z3.h, z24.h\n"
+      "ld1h { z25.h }, p5/Z, [x11]\n"
       "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
       ".inst 0x6479e429  // bfmmla z9.s, z1.h, z25.h\n"
       ".inst 0x6479e471  // bfmmla z17.s, z3.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6478e42d  // bfmmla z13.s, z1.h, z24.h\n"
       ".inst 0x6478e475  // bfmmla z21.s, z3.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
+      "addvl x11, x11, #2\n"
       ".inst 0x6479e42a  // bfmmla z10.s, z1.h, z25.h\n"
       ".inst 0x6479e472  // bfmmla z18.s, z3.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x9]\n"
+      "addvl x10, x10, #2\n"
       ".inst 0x6478e42e  // bfmmla z14.s, z1.h, z24.h\n"
       ".inst 0x6478e476  // bfmmla z22.s, z3.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n"
@@ -847,24 +845,24 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 35b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
       "uzp1 z7.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x13, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 z16.d, z16.d, z20.d\n"
       "uzp1 z17.d, z17.d, z21.d\n"
       "uzp1 z18.d, z18.d, z22.d\n"
       "uzp1 z19.d, z19.d, z23.d\n"
       "tbz %x[flags], #1, 41f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z25.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z25.s }, p5/Z, [x21]\n"
       "ld1rw { z24.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z25.s\n"
       "fmin z12.s, p5/M, z12.s, z25.s\n"
@@ -896,14 +894,14 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
       "42:"  // Height 3: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -911,15 +909,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "b 86f\n"
       "43:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "44:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -927,12 +925,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 45f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 45f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 45f\n"
       "mov x11, x12\n"
       "45:"  // Height 4: B setup done
@@ -947,15 +945,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cbz x15, 46f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x15, x15, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -970,37 +968,37 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "46:"  // Height 4: no bias
       "tbz %x[flags], #0, 47f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x13]\n"
-      "ld1w { z22.s }, p3/Z, [x13, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x13, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
       "add x22, x13, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x13]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x13, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x21]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x20]\n"
-      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x20]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
@@ -1027,8 +1025,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "49:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 50f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1053,133 +1051,133 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ble 53f\n"
       "52:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z31.h }, p5/Z, [x12]\n"
-      "ld1h { z30.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z29.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqh { z25.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqh { z30.h }, p0/Z, [x26]\n"
+      "ld1rqh { z24.h }, p0/Z, [x25]\n"
+      "trn1 z29.d, z30.d, z24.d\n"
       "ld1rqh { z28.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqh { z24.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "trn1 z27.d, z29.d, z25.d\n"
-      "trn2 z29.d, z29.d, z25.d\n"
-      "trn1 z26.d, z28.d, z24.d\n"
-      "trn2 z28.d, z28.d, z24.d\n"
-      ".inst 0x647fe768  // bfmmla z8.s, z27.h, z31.h\n"
-      ".inst 0x647ee76c  // bfmmla z12.s, z27.h, z30.h\n"
-      ".inst 0x647fe750  // bfmmla z16.s, z26.h, z31.h\n"
+      "ld1rqh { z27.h }, p0/Z, [x23]\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "trn1 z26.d, z28.d, z27.d\n"
+      "ld1h { z25.h }, p5/Z, [x12]\n"
+      "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n"
+      ".inst 0x6479e7a8  // bfmmla z8.s, z29.h, z25.h\n"
+      ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e7ac  // bfmmla z12.s, z29.h, z24.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x11]\n"
-      ".inst 0x647ee754  // bfmmla z20.s, z26.h, z30.h\n"
       "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n"
-      ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
+      ".inst 0x6479e7a9  // bfmmla z9.s, z29.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
+      "trn2 z28.d, z28.d, z27.d\n"
+      ".inst 0x6478e7ad  // bfmmla z13.s, z29.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
+      "sub x27, x27, #0x8\n"
+      ".inst 0x6479e7aa  // bfmmla z10.s, z29.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x9]\n"
-      ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x6478e7ae  // bfmmla z14.s, z29.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n"
-      ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
+      "add x26, x26, #0x10\n"
+      ".inst 0x6479e7ab  // bfmmla z11.s, z29.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x12, #2, MUL VL]\n"
-      ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x6478e7af  // bfmmla z15.s, z29.h, z24.h\n"
       ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
-      ".inst 0x6479e7a8  // bfmmla z8.s, z29.h, z25.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6479e7c8  // bfmmla z8.s, z30.h, z25.h\n"
       ".inst 0x6479e790  // bfmmla z16.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x11, #2, MUL VL]\n"
-      ".inst 0x6478e7ac  // bfmmla z12.s, z29.h, z24.h\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6478e7cc  // bfmmla z12.s, z30.h, z24.h\n"
       ".inst 0x6478e794  // bfmmla z20.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x11, #3, MUL VL]\n"
-      "addvl x11, x11, #4\n"
-      ".inst 0x6479e7a9  // bfmmla z9.s, z29.h, z25.h\n"
+      "addvl x12, x12, #4\n"
+      ".inst 0x6479e7c9  // bfmmla z9.s, z30.h, z25.h\n"
       ".inst 0x6479e791  // bfmmla z17.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6478e7ad  // bfmmla z13.s, z29.h, z24.h\n"
+      "addvl x11, x11, #4\n"
+      ".inst 0x6478e7cd  // bfmmla z13.s, z30.h, z24.h\n"
       ".inst 0x6478e795  // bfmmla z21.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
-      ".inst 0x6479e7aa  // bfmmla z10.s, z29.h, z25.h\n"
+      ".inst 0x6479e7ca  // bfmmla z10.s, z30.h, z25.h\n"
       ".inst 0x6479e792  // bfmmla z18.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x9, #2, MUL VL]\n"
-      ".inst 0x6478e7ae  // bfmmla z14.s, z29.h, z24.h\n"
+      ".inst 0x6478e7ce  // bfmmla z14.s, z30.h, z24.h\n"
       ".inst 0x6478e796  // bfmmla z22.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      ".inst 0x6479e7ab  // bfmmla z11.s, z29.h, z25.h\n"
+      ".inst 0x6479e7cb  // bfmmla z11.s, z30.h, z25.h\n"
       ".inst 0x6479e793  // bfmmla z19.s, z28.h, z25.h\n"
-      ".inst 0x6478e7af  // bfmmla z15.s, z29.h, z24.h\n"
+      ".inst 0x6478e7cf  // bfmmla z15.s, z30.h, z24.h\n"
       ".inst 0x6478e797  // bfmmla z23.s, z28.h, z24.h\n"
       "bgt 52b\n"
       "53:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x12]\n"
-      "ld1h { z28.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
-      "addvl x12, x12, #2\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "ld1rqh { z25.h }, p0/Z, [x25]\n"
+      "ld1rqh { z24.h }, p0/Z, [x25]\n"
+      "trn1 z28.d, z1.d, z24.d\n"
       "ld1rqh { z3.h }, p0/Z, [x24]\n"
-      "ld1rqh { z24.h }, p0/Z, [x23]\n"
-      "trn1 z27.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      "trn1 z26.d, z3.d, z24.d\n"
-      ".inst 0x647de768  // bfmmla z8.s, z27.h, z29.h\n"
-      ".inst 0x647ce76c  // bfmmla z12.s, z27.h, z28.h\n"
-      "trn2 z3.d, z3.d, z24.d\n"
-      ".inst 0x647de750  // bfmmla z16.s, z26.h, z29.h\n"
+      "ld1rqh { z27.h }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z24.d\n"
+      "trn1 z26.d, z3.d, z27.d\n"
+      "ld1h { z25.h }, p5/Z, [x12]\n"
+      "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n"
+      ".inst 0x6479e788  // bfmmla z8.s, z28.h, z25.h\n"
+      ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e78c  // bfmmla z12.s, z28.h, z24.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x11]\n"
-      ".inst 0x647ce754  // bfmmla z20.s, z26.h, z28.h\n"
       "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
-      ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
+      ".inst 0x6479e789  // bfmmla z9.s, z28.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
+      "subs x27, x27, #0x4\n"
+      ".inst 0x6478e78d  // bfmmla z13.s, z28.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
-      ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
+      "trn2 z3.d, z3.d, z27.d\n"
+      ".inst 0x6479e78a  // bfmmla z10.s, z28.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x9]\n"
-      ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
+      "addvl x12, x12, #2\n"
+      ".inst 0x6478e78e  // bfmmla z14.s, z28.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
-      ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
+      "addvl x11, x11, #2\n"
+      ".inst 0x6479e78b  // bfmmla z11.s, z28.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
-      ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      ".inst 0x6478e78f  // bfmmla z15.s, z28.h, z24.h\n"
       ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "ble 54f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
       ".inst 0x6479e428  // bfmmla z8.s, z1.h, z25.h\n"
       ".inst 0x6479e470  // bfmmla z16.s, z3.h, z25.h\n"
-      "ld1h { z25.h }, p5/Z, [x11]\n"
       ".inst 0x6478e42c  // bfmmla z12.s, z1.h, z24.h\n"
       ".inst 0x6478e474  // bfmmla z20.s, z3.h, z24.h\n"
+      "ld1h { z25.h }, p5/Z, [x11]\n"
       "ld1h { z24.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
       ".inst 0x6479e429  // bfmmla z9.s, z1.h, z25.h\n"
       ".inst 0x6479e471  // bfmmla z17.s, z3.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6478e42d  // bfmmla z13.s, z1.h, z24.h\n"
       ".inst 0x6478e475  // bfmmla z21.s, z3.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
+      "addvl x11, x11, #2\n"
       ".inst 0x6479e42a  // bfmmla z10.s, z1.h, z25.h\n"
       ".inst 0x6479e472  // bfmmla z18.s, z3.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x9]\n"
+      "addvl x10, x10, #2\n"
       ".inst 0x6478e42e  // bfmmla z14.s, z1.h, z24.h\n"
       ".inst 0x6478e476  // bfmmla z22.s, z3.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x9, #1, MUL VL]\n"
@@ -1194,17 +1192,17 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 49b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z7.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
       "uzp1 z20.d, z17.d, z21.d\n"
@@ -1214,9 +1212,9 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
       "tbz %x[flags], #1, 55f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z24.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z24.s }, p5/Z, [x21]\n"
       "ld1rw { z23.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z24.s\n"
       "fmin z12.s, p5/M, z12.s, z24.s\n"
@@ -1256,18 +1254,18 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
       "56:"  // Height 4: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1275,15 +1273,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "b 86f\n"
       "57:"  // Height 5
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "58:"  // Height 5: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -1291,12 +1289,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 59f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 59f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 59f\n"
       "mov x11, x12\n"
       "59:"  // Height 5: B setup done
@@ -1311,15 +1309,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cbz x15, 60f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x15, x15, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -1342,46 +1340,46 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "60:"  // Height 5: no bias
       "tbz %x[flags], #0, 61f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x13]\n"
-      "ld1w { z22.s }, p3/Z, [x13, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x13, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
       "add x23, x13, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x13]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x13, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x22]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x21]\n"
-      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x20]\n"
-      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
+      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x21]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x20]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z19.d, z24.d, z23.d\n"
       "zip2 z23.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z24.d, z25.d, z28.d\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
@@ -1420,8 +1418,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "63:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 64f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1449,120 +1447,120 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ble 67f\n"
       "66:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z1.h }, p5/Z, [x12]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z6.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqh { z3.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqh { z1.h }, p0/Z, [x25]\n"
       "ld1rqh { z7.h }, p0/Z, [x24]\n"
       "ld1rqh { z2.h }, p0/Z, [x23]\n"
-      "ld1rqh { z5.h }, p0/Z, [x22]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
+      "trn1 z5.d, z6.d, z1.d\n"
+      "trn2 z6.d, z6.d, z1.d\n"
+      "ld1rqh { z4.h }, p0/Z, [x22]\n"
       "trn1 z3.d, z7.d, z2.d\n"
       "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
+      "ld1h { z1.h }, p5/Z, [x12]\n"
+      "trn1 z2.d, z4.d, z0.d\n"
+      "trn2 z4.d, z4.d, z0.d\n"
       "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n"
-      ".inst 0x6461e488  // bfmmla z8.s, z4.h, z1.h\n"
+      ".inst 0x6461e4a8  // bfmmla z8.s, z5.h, z1.h\n"
       ".inst 0x6461e470  // bfmmla z16.s, z3.h, z1.h\n"
       ".inst 0x6461e458  // bfmmla z24.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x11]\n"
-      ".inst 0x6460e48c  // bfmmla z12.s, z4.h, z0.h\n"
+      "sub x27, x27, #0x8\n"
+      ".inst 0x6460e4ac  // bfmmla z12.s, z5.h, z0.h\n"
       ".inst 0x6460e474  // bfmmla z20.s, z3.h, z0.h\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6460e45c  // bfmmla z28.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n"
-      ".inst 0x6461e489  // bfmmla z9.s, z4.h, z1.h\n"
+      ".inst 0x6461e4a9  // bfmmla z9.s, z5.h, z1.h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6461e471  // bfmmla z17.s, z3.h, z1.h\n"
       ".inst 0x6461e459  // bfmmla z25.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10]\n"
-      ".inst 0x6460e48d  // bfmmla z13.s, z4.h, z0.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6460e4ad  // bfmmla z13.s, z5.h, z0.h\n"
       ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6460e45d  // bfmmla z29.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6461e48a  // bfmmla z10.s, z4.h, z1.h\n"
+      ".inst 0x6461e4aa  // bfmmla z10.s, z5.h, z1.h\n"
       ".inst 0x6461e472  // bfmmla z18.s, z3.h, z1.h\n"
       ".inst 0x6461e45a  // bfmmla z26.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x9]\n"
-      ".inst 0x6460e48e  // bfmmla z14.s, z4.h, z0.h\n"
+      ".inst 0x6460e4ae  // bfmmla z14.s, z5.h, z0.h\n"
       ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
       ".inst 0x6460e45e  // bfmmla z30.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n"
-      ".inst 0x6461e48b  // bfmmla z11.s, z4.h, z1.h\n"
+      ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
       ".inst 0x6461e473  // bfmmla z19.s, z3.h, z1.h\n"
       ".inst 0x6461e45b  // bfmmla z27.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x12, #2, MUL VL]\n"
-      ".inst 0x6460e48f  // bfmmla z15.s, z4.h, z0.h\n"
+      ".inst 0x6460e4af  // bfmmla z15.s, z5.h, z0.h\n"
       ".inst 0x6460e477  // bfmmla z23.s, z3.h, z0.h\n"
       ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       ".inst 0x6461e4c8  // bfmmla z8.s, z6.h, z1.h\n"
+      "addvl x12, x12, #4\n"
       ".inst 0x6461e4f0  // bfmmla z16.s, z7.h, z1.h\n"
-      ".inst 0x6461e4b8  // bfmmla z24.s, z5.h, z1.h\n"
+      ".inst 0x6461e498  // bfmmla z24.s, z4.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x11, #2, MUL VL]\n"
       ".inst 0x6460e4cc  // bfmmla z12.s, z6.h, z0.h\n"
       ".inst 0x6460e4f4  // bfmmla z20.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bc  // bfmmla z28.s, z5.h, z0.h\n"
+      ".inst 0x6460e49c  // bfmmla z28.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x11, #3, MUL VL]\n"
-      "addvl x11, x11, #4\n"
       ".inst 0x6461e4c9  // bfmmla z9.s, z6.h, z1.h\n"
+      "addvl x11, x11, #4\n"
       ".inst 0x6461e4f1  // bfmmla z17.s, z7.h, z1.h\n"
-      ".inst 0x6461e4b9  // bfmmla z25.s, z5.h, z1.h\n"
+      ".inst 0x6461e499  // bfmmla z25.s, z4.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6460e4cd  // bfmmla z13.s, z6.h, z0.h\n"
       ".inst 0x6460e4f5  // bfmmla z21.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
+      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       ".inst 0x6461e4ca  // bfmmla z10.s, z6.h, z1.h\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x6461e4f2  // bfmmla z18.s, z7.h, z1.h\n"
-      ".inst 0x6461e4ba  // bfmmla z26.s, z5.h, z1.h\n"
+      ".inst 0x6461e49a  // bfmmla z26.s, z4.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x9, #2, MUL VL]\n"
       ".inst 0x6460e4ce  // bfmmla z14.s, z6.h, z0.h\n"
       ".inst 0x6460e4f6  // bfmmla z22.s, z7.h, z0.h\n"
-      ".inst 0x6460e4be  // bfmmla z30.s, z5.h, z0.h\n"
+      ".inst 0x6460e49e  // bfmmla z30.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
       ".inst 0x6461e4cb  // bfmmla z11.s, z6.h, z1.h\n"
+      "addvl x9, x9, #4\n"
       ".inst 0x6461e4f3  // bfmmla z19.s, z7.h, z1.h\n"
-      ".inst 0x6461e4bb  // bfmmla z27.s, z5.h, z1.h\n"
+      ".inst 0x6461e49b  // bfmmla z27.s, z4.h, z1.h\n"
       ".inst 0x6460e4cf  // bfmmla z15.s, z6.h, z0.h\n"
       ".inst 0x6460e4f7  // bfmmla z23.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bf  // bfmmla z31.s, z5.h, z0.h\n"
+      ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
       "bgt 66b\n"
       "67:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z2.h }, p5/Z, [x12]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "ld1rqh { z6.h }, p0/Z, [x25]\n"
+      "ld1rqh { z4.h }, p0/Z, [x25]\n"
       "ld1rqh { z3.h }, p0/Z, [x24]\n"
-      "ld1rqh { z4.h }, p0/Z, [x23]\n"
+      "ld1rqh { z2.h }, p0/Z, [x23]\n"
+      "trn1 z7.d, z1.d, z4.d\n"
+      "trn2 z1.d, z1.d, z4.d\n"
       "ld1rqh { z5.h }, p0/Z, [x22]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
+      "ld1h { z2.h }, p5/Z, [x12]\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
       "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
       ".inst 0x6462e4e8  // bfmmla z8.s, z7.h, z2.h\n"
       ".inst 0x6462e4d0  // bfmmla z16.s, z6.h, z2.h\n"
       ".inst 0x6462e498  // bfmmla z24.s, z4.h, z2.h\n"
       "ld1h { z2.h }, p5/Z, [x11]\n"
+      "subs x27, x27, #0x4\n"
       ".inst 0x6460e4ec  // bfmmla z12.s, z7.h, z0.h\n"
       ".inst 0x6460e4d4  // bfmmla z20.s, z6.h, z0.h\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6460e49c  // bfmmla z28.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
       ".inst 0x6462e4e9  // bfmmla z9.s, z7.h, z2.h\n"
+      "addvl x11, x11, #2\n"
       ".inst 0x6462e4d1  // bfmmla z17.s, z6.h, z2.h\n"
       ".inst 0x6462e499  // bfmmla z25.s, z4.h, z2.h\n"
       "ld1h { z2.h }, p5/Z, [x10]\n"
@@ -1570,8 +1568,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6460e4d5  // bfmmla z21.s, z6.h, z0.h\n"
       ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
       ".inst 0x6462e4ea  // bfmmla z10.s, z7.h, z2.h\n"
+      "addvl x10, x10, #2\n"
       ".inst 0x6462e4d2  // bfmmla z18.s, z6.h, z2.h\n"
       ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
       "ld1h { z2.h }, p5/Z, [x9]\n"
@@ -1579,8 +1577,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6460e4d6  // bfmmla z22.s, z6.h, z0.h\n"
       ".inst 0x6460e49e  // bfmmla z30.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
       ".inst 0x6462e4eb  // bfmmla z11.s, z7.h, z2.h\n"
+      "addvl x9, x9, #2\n"
       ".inst 0x6462e4d3  // bfmmla z19.s, z6.h, z2.h\n"
       ".inst 0x6462e49b  // bfmmla z27.s, z4.h, z2.h\n"
       ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
@@ -1589,12 +1587,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ble 68f\n"
       "ld1h { z2.h }, p5/Z, [x12]\n"
       "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
       ".inst 0x6462e428  // bfmmla z8.s, z1.h, z2.h\n"
       ".inst 0x6462e470  // bfmmla z16.s, z3.h, z2.h\n"
       ".inst 0x6462e4b8  // bfmmla z24.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x11]\n"
       ".inst 0x6460e42c  // bfmmla z12.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x11]\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6460e474  // bfmmla z20.s, z3.h, z0.h\n"
       ".inst 0x6460e4bc  // bfmmla z28.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n"
@@ -1602,8 +1600,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6462e429  // bfmmla z9.s, z1.h, z2.h\n"
       ".inst 0x6462e471  // bfmmla z17.s, z3.h, z2.h\n"
       ".inst 0x6462e4b9  // bfmmla z25.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x10]\n"
       ".inst 0x6460e42d  // bfmmla z13.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x10]\n"
       ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
       ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
@@ -1611,8 +1609,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6462e42a  // bfmmla z10.s, z1.h, z2.h\n"
       ".inst 0x6462e472  // bfmmla z18.s, z3.h, z2.h\n"
       ".inst 0x6462e4ba  // bfmmla z26.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x9]\n"
       ".inst 0x6460e42e  // bfmmla z14.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x9]\n"
       ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
       ".inst 0x6460e4be  // bfmmla z30.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n"
@@ -1629,20 +1627,20 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 63b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z7.d, z8.d, z12.d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
@@ -1654,9 +1652,9 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "uzp1 z26.d, z26.d, z30.d\n"
       "uzp1 z27.d, z27.d, z31.d\n"
       "tbz %x[flags], #1, 69f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z0.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z0.s }, p5/Z, [x21]\n"
       "ld1rw { z23.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z0.s\n"
       "fmin z12.s, p5/M, z12.s, z0.s\n"
@@ -1704,22 +1702,22 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
       "70:"  // Height 5: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1727,19 +1725,18 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "b 86f\n"
       "71:"  // Height 6
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0x18\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0x18\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "72:"  // Height 6: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -1747,12 +1744,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 73f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 73f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 73f\n"
       "mov x11, x12\n"
       "73:"  // Height 6: B setup done
@@ -1767,15 +1764,15 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cbz x15, 74f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x15, x15, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -1798,54 +1795,54 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "74:"  // Height 6: no bias
       "tbz %x[flags], #0, 75f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x13]\n"
-      "ld1w { z22.s }, p3/Z, [x13, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x13, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
       "add x24, x13, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "ld1w { z17.s }, p4/Z, [x13]\n"
       "add x22, x23, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x24]\n"
-      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z18.s }, p3/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x13, #2, MUL VL]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z16.s }, p1/Z, [x13, #3, MUL VL]\n"
+      "ld1w { z12.s }, p4/Z, [x24]\n"
+      "zip1 z8.d, z17.d, z12.d\n"
+      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
+      "zip2 z12.d, z17.d, z12.d\n"
+      "zip1 z9.d, z18.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x23]\n"
+      "zip2 z13.d, z18.d, z13.d\n"
+      "zip1 z10.d, z20.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip2 z14.d, z20.d, z14.d\n"
+      "zip1 z11.d, z16.d, z15.d\n"
       "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
       "ld1w { z20.s }, p4/Z, [x22]\n"
-      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x21]\n"
-      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
       "zip2 z20.d, z17.d, z20.d\n"
-      "ld1w { z28.s }, p4/Z, [x20]\n"
-      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x21]\n"
       "zip2 z21.d, z18.d, z21.d\n"
-      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z28.s }, p4/Z, [x20]\n"
       "zip2 z23.d, z24.d, z23.d\n"
       "zip1 z24.d, z25.d, z28.d\n"
+      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
+      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip2 z29.d, z26.d, z29.d\n"
       "zip1 z26.d, z27.d, z30.d\n"
       "zip2 z30.d, z27.d, z30.d\n"
@@ -1881,8 +1878,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "77:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 78f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1913,123 +1910,123 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ble 81f\n"
       "80:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z1.h }, p5/Z, [x12]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z6.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqh { z3.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqh { z7.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqh { z2.h }, p0/Z, [x23]\n"
-      "ld1rqh { z5.h }, p0/Z, [x22]\n"
+      "ld1rqh { z7.h }, p0/Z, [x26]\n"
+      "ld1rqh { z0.h }, p0/Z, [x25]\n"
+      "trn1 z6.d, z7.d, z0.d\n"
+      "ld1rqh { z5.h }, p0/Z, [x24]\n"
+      "ld1rqh { z1.h }, p0/Z, [x23]\n"
+      "trn2 z7.d, z7.d, z0.d\n"
+      "trn1 z4.d, z5.d, z1.d\n"
+      "ld1rqh { z3.h }, p0/Z, [x22]\n"
       "ld1rqh { z0.h }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
-      "add x21, x21, #0x10\n"
-      "trn1 z3.d, z7.d, z2.d\n"
-      "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
+      "trn2 z5.d, z5.d, z1.d\n"
+      "trn1 z2.d, z3.d, z0.d\n"
+      "trn2 z3.d, z3.d, z0.d\n"
+      "ld1h { z1.h }, p5/Z, [x12]\n"
       "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n"
-      ".inst 0x6461e488  // bfmmla z8.s, z4.h, z1.h\n"
-      ".inst 0x6461e470  // bfmmla z16.s, z3.h, z1.h\n"
+      ".inst 0x6461e4c8  // bfmmla z8.s, z6.h, z1.h\n"
+      ".inst 0x6461e490  // bfmmla z16.s, z4.h, z1.h\n"
       ".inst 0x6461e458  // bfmmla z24.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x11]\n"
-      ".inst 0x6460e48c  // bfmmla z12.s, z4.h, z0.h\n"
-      ".inst 0x6460e474  // bfmmla z20.s, z3.h, z0.h\n"
+      "sub x27, x27, #0x8\n"
+      ".inst 0x6460e4cc  // bfmmla z12.s, z6.h, z0.h\n"
+      ".inst 0x6460e494  // bfmmla z20.s, z4.h, z0.h\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6460e45c  // bfmmla z28.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n"
-      ".inst 0x6461e489  // bfmmla z9.s, z4.h, z1.h\n"
-      ".inst 0x6461e471  // bfmmla z17.s, z3.h, z1.h\n"
+      ".inst 0x6461e4c9  // bfmmla z9.s, z6.h, z1.h\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x6461e491  // bfmmla z17.s, z4.h, z1.h\n"
       ".inst 0x6461e459  // bfmmla z25.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10]\n"
-      ".inst 0x6460e48d  // bfmmla z13.s, z4.h, z0.h\n"
-      ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6460e4cd  // bfmmla z13.s, z6.h, z0.h\n"
+      ".inst 0x6460e495  // bfmmla z21.s, z4.h, z0.h\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6460e45d  // bfmmla z29.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6461e48a  // bfmmla z10.s, z4.h, z1.h\n"
-      ".inst 0x6461e472  // bfmmla z18.s, z3.h, z1.h\n"
+      ".inst 0x6461e4ca  // bfmmla z10.s, z6.h, z1.h\n"
+      "add x21, x21, #0x10\n"
+      ".inst 0x6461e492  // bfmmla z18.s, z4.h, z1.h\n"
       ".inst 0x6461e45a  // bfmmla z26.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x9]\n"
-      ".inst 0x6460e48e  // bfmmla z14.s, z4.h, z0.h\n"
-      ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
+      ".inst 0x6460e4ce  // bfmmla z14.s, z6.h, z0.h\n"
+      ".inst 0x6460e496  // bfmmla z22.s, z4.h, z0.h\n"
       ".inst 0x6460e45e  // bfmmla z30.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n"
-      ".inst 0x6461e48b  // bfmmla z11.s, z4.h, z1.h\n"
-      ".inst 0x6461e473  // bfmmla z19.s, z3.h, z1.h\n"
+      ".inst 0x6461e4cb  // bfmmla z11.s, z6.h, z1.h\n"
+      ".inst 0x6461e493  // bfmmla z19.s, z4.h, z1.h\n"
       ".inst 0x6461e45b  // bfmmla z27.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x12, #2, MUL VL]\n"
-      ".inst 0x6460e48f  // bfmmla z15.s, z4.h, z0.h\n"
-      ".inst 0x6460e477  // bfmmla z23.s, z3.h, z0.h\n"
+      ".inst 0x6460e4cf  // bfmmla z15.s, z6.h, z0.h\n"
+      ".inst 0x6460e497  // bfmmla z23.s, z4.h, z0.h\n"
       ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x12, #3, MUL VL]\n"
+      ".inst 0x6461e4e8  // bfmmla z8.s, z7.h, z1.h\n"
       "addvl x12, x12, #4\n"
-      ".inst 0x6461e4c8  // bfmmla z8.s, z6.h, z1.h\n"
-      ".inst 0x6461e4f0  // bfmmla z16.s, z7.h, z1.h\n"
-      ".inst 0x6461e4b8  // bfmmla z24.s, z5.h, z1.h\n"
+      ".inst 0x6461e4b0  // bfmmla z16.s, z5.h, z1.h\n"
+      ".inst 0x6461e478  // bfmmla z24.s, z3.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x11, #2, MUL VL]\n"
-      ".inst 0x6460e4cc  // bfmmla z12.s, z6.h, z0.h\n"
-      ".inst 0x6460e4f4  // bfmmla z20.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bc  // bfmmla z28.s, z5.h, z0.h\n"
+      ".inst 0x6460e4ec  // bfmmla z12.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b4  // bfmmla z20.s, z5.h, z0.h\n"
+      ".inst 0x6460e47c  // bfmmla z28.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x11, #3, MUL VL]\n"
+      ".inst 0x6461e4e9  // bfmmla z9.s, z7.h, z1.h\n"
       "addvl x11, x11, #4\n"
-      ".inst 0x6461e4c9  // bfmmla z9.s, z6.h, z1.h\n"
-      ".inst 0x6461e4f1  // bfmmla z17.s, z7.h, z1.h\n"
-      ".inst 0x6461e4b9  // bfmmla z25.s, z5.h, z1.h\n"
+      ".inst 0x6461e4b1  // bfmmla z17.s, z5.h, z1.h\n"
+      ".inst 0x6461e479  // bfmmla z25.s, z3.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6460e4cd  // bfmmla z13.s, z6.h, z0.h\n"
-      ".inst 0x6460e4f5  // bfmmla z21.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
+      ".inst 0x6460e4ed  // bfmmla z13.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b5  // bfmmla z21.s, z5.h, z0.h\n"
+      ".inst 0x6460e47d  // bfmmla z29.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
+      ".inst 0x6461e4ea  // bfmmla z10.s, z7.h, z1.h\n"
       "addvl x10, x10, #4\n"
-      ".inst 0x6461e4ca  // bfmmla z10.s, z6.h, z1.h\n"
-      ".inst 0x6461e4f2  // bfmmla z18.s, z7.h, z1.h\n"
-      ".inst 0x6461e4ba  // bfmmla z26.s, z5.h, z1.h\n"
+      ".inst 0x6461e4b2  // bfmmla z18.s, z5.h, z1.h\n"
+      ".inst 0x6461e47a  // bfmmla z26.s, z3.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x9, #2, MUL VL]\n"
-      ".inst 0x6460e4ce  // bfmmla z14.s, z6.h, z0.h\n"
-      ".inst 0x6460e4f6  // bfmmla z22.s, z7.h, z0.h\n"
-      ".inst 0x6460e4be  // bfmmla z30.s, z5.h, z0.h\n"
+      ".inst 0x6460e4ee  // bfmmla z14.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b6  // bfmmla z22.s, z5.h, z0.h\n"
+      ".inst 0x6460e47e  // bfmmla z30.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x9, #3, MUL VL]\n"
+      ".inst 0x6461e4eb  // bfmmla z11.s, z7.h, z1.h\n"
       "addvl x9, x9, #4\n"
-      ".inst 0x6461e4cb  // bfmmla z11.s, z6.h, z1.h\n"
-      ".inst 0x6461e4f3  // bfmmla z19.s, z7.h, z1.h\n"
-      ".inst 0x6461e4bb  // bfmmla z27.s, z5.h, z1.h\n"
-      ".inst 0x6460e4cf  // bfmmla z15.s, z6.h, z0.h\n"
-      ".inst 0x6460e4f7  // bfmmla z23.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bf  // bfmmla z31.s, z5.h, z0.h\n"
+      ".inst 0x6461e4b3  // bfmmla z19.s, z5.h, z1.h\n"
+      ".inst 0x6461e47b  // bfmmla z27.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b7  // bfmmla z23.s, z5.h, z0.h\n"
+      ".inst 0x6460e47f  // bfmmla z31.s, z3.h, z0.h\n"
       "bgt 80b\n"
       "81:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z2.h }, p5/Z, [x12]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "ld1rqh { z6.h }, p0/Z, [x25]\n"
+      "ld1rqh { z0.h }, p0/Z, [x25]\n"
+      "trn1 z7.d, z1.d, z0.d\n"
       "ld1rqh { z3.h }, p0/Z, [x24]\n"
-      "ld1rqh { z4.h }, p0/Z, [x23]\n"
+      "ld1rqh { z2.h }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z0.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
       "ld1rqh { z5.h }, p0/Z, [x22]\n"
       "ld1rqh { z0.h }, p0/Z, [x21]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
+      "ld1h { z2.h }, p5/Z, [x12]\n"
       "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
       ".inst 0x6462e4e8  // bfmmla z8.s, z7.h, z2.h\n"
       ".inst 0x6462e4d0  // bfmmla z16.s, z6.h, z2.h\n"
       ".inst 0x6462e498  // bfmmla z24.s, z4.h, z2.h\n"
       "ld1h { z2.h }, p5/Z, [x11]\n"
+      "subs x27, x27, #0x4\n"
       ".inst 0x6460e4ec  // bfmmla z12.s, z7.h, z0.h\n"
       ".inst 0x6460e4d4  // bfmmla z20.s, z6.h, z0.h\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6460e49c  // bfmmla z28.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
       ".inst 0x6462e4e9  // bfmmla z9.s, z7.h, z2.h\n"
+      "addvl x11, x11, #2\n"
       ".inst 0x6462e4d1  // bfmmla z17.s, z6.h, z2.h\n"
       ".inst 0x6462e499  // bfmmla z25.s, z4.h, z2.h\n"
       "ld1h { z2.h }, p5/Z, [x10]\n"
@@ -2037,8 +2034,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6460e4d5  // bfmmla z21.s, z6.h, z0.h\n"
       ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
       ".inst 0x6462e4ea  // bfmmla z10.s, z7.h, z2.h\n"
+      "addvl x10, x10, #2\n"
       ".inst 0x6462e4d2  // bfmmla z18.s, z6.h, z2.h\n"
       ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
       "ld1h { z2.h }, p5/Z, [x9]\n"
@@ -2046,8 +2043,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6460e4d6  // bfmmla z22.s, z6.h, z0.h\n"
       ".inst 0x6460e49e  // bfmmla z30.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
       ".inst 0x6462e4eb  // bfmmla z11.s, z7.h, z2.h\n"
+      "addvl x9, x9, #2\n"
       ".inst 0x6462e4d3  // bfmmla z19.s, z6.h, z2.h\n"
       ".inst 0x6462e49b  // bfmmla z27.s, z4.h, z2.h\n"
       ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
@@ -2056,12 +2053,12 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "ble 82f\n"
       "ld1h { z2.h }, p5/Z, [x12]\n"
       "ld1h { z0.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
       ".inst 0x6462e428  // bfmmla z8.s, z1.h, z2.h\n"
       ".inst 0x6462e470  // bfmmla z16.s, z3.h, z2.h\n"
       ".inst 0x6462e4b8  // bfmmla z24.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x11]\n"
       ".inst 0x6460e42c  // bfmmla z12.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x11]\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6460e474  // bfmmla z20.s, z3.h, z0.h\n"
       ".inst 0x6460e4bc  // bfmmla z28.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x11, #1, MUL VL]\n"
@@ -2069,8 +2066,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6462e429  // bfmmla z9.s, z1.h, z2.h\n"
       ".inst 0x6462e471  // bfmmla z17.s, z3.h, z2.h\n"
       ".inst 0x6462e4b9  // bfmmla z25.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x10]\n"
       ".inst 0x6460e42d  // bfmmla z13.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x10]\n"
       ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
       ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
@@ -2078,8 +2075,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6462e42a  // bfmmla z10.s, z1.h, z2.h\n"
       ".inst 0x6462e472  // bfmmla z18.s, z3.h, z2.h\n"
       ".inst 0x6462e4ba  // bfmmla z26.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x9]\n"
       ".inst 0x6460e42e  // bfmmla z14.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x9]\n"
       ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
       ".inst 0x6460e4be  // bfmmla z30.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x9, #1, MUL VL]\n"
@@ -2096,21 +2093,21 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 77b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z7.d, z8.d, z12.d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
@@ -2126,9 +2123,9 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "uzp1 z30.d, z27.d, z31.d\n"
       "uzp2 z27.d, z27.d, z31.d\n"
       "tbz %x[flags], #1, 83f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p5/Z, [x21]\n"
       "ld1rw { z0.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z1.s\n"
       "fmin z12.s, p5/M, z12.s, z1.s\n"
@@ -2184,26 +2181,26 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z23.s }, p4, [x23]\n"
-      "st1w { z28.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z29.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z30.s }, p1, [x23, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x22]\n"
-      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z23.s }, p4, [x22]\n"
+      "st1w { z28.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z29.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z30.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x21]\n"
+      "st1w { z25.s }, p3, [x21, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x21, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x21, #3, MUL VL]\n"
       "84:"  // Height 6: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -2220,8 +2217,8 @@ void sve_ffhybrid_bf16fp32_mmla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "86:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL.hpp
index 44b766c2d5..8e83f1cb2c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL.hpp
@@ -82,7 +82,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 4, 1> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/a64fx.cpp
index 406f20bad7..ffad168b44 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/a64fx.cpp
@@ -49,19 +49,18 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const __fp16 *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -82,7 +81,6 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -108,15 +106,15 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "bgt 27f\n"
       "beq 14f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -124,12 +122,12 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 3f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 3f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 3f\n"
       "mov x11, x12\n"
       "3:"  // Height 1: B setup done
@@ -164,8 +162,8 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "7:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 8f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -183,41 +181,41 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "ld1h { z7.h }, p4/Z, [x11]\n"
       "ble 11f\n"
       "10:"  // Height 1: Multiply loop: Main loop
-      "addvl x12, x12, #1\n"
       "fmla z8.h, p4/M, z6.h, z0.h\n"
-      "ld1h { z17.h }, p4/Z, [x10]\n"
-      "addvl x11, x11, #1\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z17.h }, p4/Z, [x10]\n"
       "ld1h { z16.h }, p4/Z, [x9]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "fmla z10.h, p4/M, z17.h, z0.h\n"
+      "fmla z11.h, p4/M, z16.h, z0.h\n"
       "add x26, x26, #0x2\n"
       "subs x27, x27, #0x1\n"
+      "ld1rh { z0.h }, p4/Z, [x26]\n"
+      "ld1h { z6.h }, p4/Z, [x12]\n"
       "addvl x10, x10, #1\n"
       "addvl x9, x9, #1\n"
-      "fmla z10.h, p4/M, z17.h, z0.h\n"
-      "ld1h { z6.h }, p4/Z, [x12]\n"
-      "fmla z11.h, p4/M, z16.h, z0.h\n"
-      "ld1rh { z0.h }, p4/Z, [x26]\n"
       "ld1h { z7.h }, p4/Z, [x11]\n"
       "bgt 10b\n"
       "11:"  // Height 1: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "fmla z8.h, p4/M, z6.h, z0.h\n"
-      "ld1h { z17.h }, p4/Z, [x10]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z17.h }, p4/Z, [x10]\n"
       "ld1h { z16.h }, p4/Z, [x9]\n"
       "add x28, x28, #0x1\n"
+      "cmp x28, x20\n"
+      "fmla z10.h, p4/M, z17.h, z0.h\n"
+      "fmla z11.h, p4/M, z16.h, z0.h\n"
       "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
       "addvl x10, x10, #1\n"
-      "cmp x28, x20\n"
       "addvl x9, x9, #1\n"
-      "fmla z10.h, p4/M, z17.h, z0.h\n"
-      "fmla z11.h, p4/M, z16.h, z0.h\n"
       "bne 7b\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z17.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z17.h }, p4/Z, [x21]\n"
       "ld1rh { z16.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z17.h\n"
       "fmin z9.h, p4/M, z9.h, z17.h\n"
@@ -240,15 +238,15 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "b 80f\n"
       "14:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -256,12 +254,12 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 16f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 16f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 16f\n"
       "mov x11, x12\n"
       "16:"  // Height 2: B setup done
@@ -276,22 +274,22 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "cbz x15, 17f\n"
       "ld1h { z8.h }, p4/Z, [x15]\n"
       "ld1h { z9.h }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x15, x15, #4\n"
       "b 19f\n"
       "17:"  // Height 2: no bias
       "tbz %x[flags], #0, 18f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x13, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x13]\n"
       "ld1h { z9.h }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x13, #2, MUL VL]\n"
       "ld1h { z11.h }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x20, x13, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x20]\n"
       "ld1h { z13.h }, p2/Z, [x20, #1, MUL VL]\n"
       "ld1h { z14.h }, p1/Z, [x20, #2, MUL VL]\n"
@@ -310,8 +308,8 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "20:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 21f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -333,26 +331,26 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "ld1h { z7.h }, p4/Z, [x11]\n"
       "ble 24f\n"
       "23:"  // Height 2: Multiply loop: Main loop
-      "addvl x12, x12, #1\n"
       "fmla z8.h, p4/M, z6.h, z0.h\n"
       "fmla z12.h, p4/M, z6.h, z1.h\n"
       "ld1h { z17.h }, p4/Z, [x10]\n"
-      "addvl x11, x11, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "ld1h { z16.h }, p4/Z, [x9]\n"
+      "addvl x11, x11, #1\n"
       "add x26, x26, #0x2\n"
       "subs x27, x27, #0x1\n"
-      "add x25, x25, #0x2\n"
-      "addvl x10, x10, #1\n"
       "fmla z10.h, p4/M, z17.h, z0.h\n"
       "fmla z14.h, p4/M, z17.h, z1.h\n"
-      "addvl x9, x9, #1\n"
-      "ld1h { z6.h }, p4/Z, [x12]\n"
+      "add x25, x25, #0x2\n"
       "fmla z11.h, p4/M, z16.h, z0.h\n"
-      "ld1rh { z0.h }, p4/Z, [x26]\n"
       "fmla z15.h, p4/M, z16.h, z1.h\n"
+      "addvl x10, x10, #1\n"
+      "ld1rh { z0.h }, p4/Z, [x26]\n"
       "ld1rh { z1.h }, p4/Z, [x25]\n"
+      "addvl x9, x9, #1\n"
+      "ld1h { z6.h }, p4/Z, [x12]\n"
       "ld1h { z7.h }, p4/Z, [x11]\n"
       "bgt 23b\n"
       "24:"  // Height 2: Multiply loop: Main loop skip
@@ -364,22 +362,22 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "ld1h { z16.h }, p4/Z, [x9]\n"
       "add x28, x28, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "cmp x28, x20\n"
-      "addvl x10, x10, #1\n"
       "fmla z10.h, p4/M, z17.h, z0.h\n"
       "fmla z14.h, p4/M, z17.h, z1.h\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z11.h, p4/M, z16.h, z0.h\n"
       "fmla z15.h, p4/M, z16.h, z1.h\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "bne 20b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z17.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z17.h }, p4/Z, [x21]\n"
       "ld1rh { z16.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z17.h\n"
       "fmin z9.h, p4/M, z9.h, z17.h\n"
@@ -403,10 +401,10 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
       "26:"  // Height 2: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -414,15 +412,15 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "b 80f\n"
       "27:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -430,12 +428,12 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 29f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 29f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 29f\n"
       "mov x11, x12\n"
       "29:"  // Height 3: B setup done
@@ -450,27 +448,27 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "cbz x15, 30f\n"
       "ld1h { z8.h }, p4/Z, [x15]\n"
       "ld1h { z9.h }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 32f\n"
       "30:"  // Height 3: no bias
       "tbz %x[flags], #0, 31f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x13, x20, LSL #1\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x13]\n"
       "ld1h { z9.h }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x13, #2, MUL VL]\n"
       "ld1h { z11.h }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x21, x13, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x21]\n"
       "ld1h { z13.h }, p2/Z, [x21, #1, MUL VL]\n"
       "ld1h { z14.h }, p1/Z, [x21, #2, MUL VL]\n"
@@ -497,8 +495,8 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "33:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 34f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -524,13 +522,13 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "ld1h { z7.h }, p4/Z, [x11]\n"
       "ble 37f\n"
       "36:"  // Height 3: Multiply loop: Main loop
-      "addvl x12, x12, #1\n"
       "fmla z8.h, p4/M, z6.h, z0.h\n"
       "fmla z12.h, p4/M, z6.h, z1.h\n"
+      "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
-      "ld1h { z21.h }, p4/Z, [x10]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z21.h }, p4/Z, [x10]\n"
       "add x26, x26, #0x2\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
@@ -538,18 +536,18 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "subs x27, x27, #0x1\n"
       "add x25, x25, #0x2\n"
       "add x24, x24, #0x2\n"
-      "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.h, p4/M, z21.h, z0.h\n"
       "fmla z14.h, p4/M, z21.h, z1.h\n"
       "fmla z18.h, p4/M, z21.h, z2.h\n"
-      "ld1h { z6.h }, p4/Z, [x12]\n"
       "fmla z11.h, p4/M, z20.h, z0.h\n"
-      "ld1rh { z0.h }, p4/Z, [x26]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, p4/M, z20.h, z1.h\n"
       "fmla z19.h, p4/M, z20.h, z2.h\n"
+      "ld1rh { z0.h }, p4/Z, [x26]\n"
       "ld1rh { z1.h }, p4/Z, [x25]\n"
       "ld1rh { z2.h }, p4/Z, [x24]\n"
+      "ld1h { z6.h }, p4/Z, [x12]\n"
       "ld1h { z7.h }, p4/Z, [x11]\n"
       "bgt 36b\n"
       "37:"  // Height 3: Multiply loop: Main loop skip
@@ -558,30 +556,30 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z12.h, p4/M, z6.h, z1.h\n"
       "add x28, x28, #0x1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
-      "ld1h { z21.h }, p4/Z, [x10]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
-      "addvl x12, x12, #1\n"
+      "ld1h { z21.h }, p4/Z, [x10]\n"
+      "cmp x28, x20\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
       "ld1h { z20.h }, p4/Z, [x9]\n"
-      "addvl x11, x11, #1\n"
-      "cmp x28, x20\n"
-      "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.h, p4/M, z21.h, z0.h\n"
       "fmla z14.h, p4/M, z21.h, z1.h\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.h, p4/M, z21.h, z2.h\n"
       "fmla z11.h, p4/M, z20.h, z0.h\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, p4/M, z20.h, z1.h\n"
       "fmla z19.h, p4/M, z20.h, z2.h\n"
       "bne 33b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z21.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z21.h }, p4/Z, [x21]\n"
       "ld1rh { z20.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z21.h\n"
       "fmin z9.h, p4/M, z9.h, z21.h\n"
@@ -613,14 +611,14 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p3, [x25]\n"
-      "st1h { z17.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p1, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p3, [x24]\n"
+      "st1h { z17.h }, p2, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p0, [x24, #3, MUL VL]\n"
       "39:"  // Height 3: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -628,15 +626,15 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "b 80f\n"
       "40:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -644,12 +642,12 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 42f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 42f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 42f\n"
       "mov x11, x12\n"
       "42:"  // Height 4: B setup done
@@ -664,18 +662,18 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "cbz x15, 43f\n"
       "ld1h { z8.h }, p4/Z, [x15]\n"
       "ld1h { z9.h }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -683,13 +681,13 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "43:"  // Height 4: no bias
       "tbz %x[flags], #0, 44f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x13, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x13]\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z9.h }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x13, #2, MUL VL]\n"
       "ld1h { z11.h }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x22, x13, x20, LSL #1\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x22]\n"
       "ld1h { z13.h }, p2/Z, [x22, #1, MUL VL]\n"
       "ld1h { z14.h }, p1/Z, [x22, #2, MUL VL]\n"
@@ -724,8 +722,8 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "46:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 47f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -755,9 +753,9 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "ld1h { z7.h }, p4/Z, [x11]\n"
       "ble 50f\n"
       "49:"  // Height 4: Multiply loop: Main loop
-      "addvl x12, x12, #1\n"
       "fmla z8.h, p4/M, z6.h, z0.h\n"
       "fmla z12.h, p4/M, z6.h, z1.h\n"
+      "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
       "fmla z20.h, p4/M, z6.h, z3.h\n"
@@ -775,9 +773,9 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z10.h, p4/M, z25.h, z0.h\n"
       "fmla z14.h, p4/M, z25.h, z1.h\n"
       "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
       "fmla z18.h, p4/M, z25.h, z2.h\n"
       "fmla z22.h, p4/M, z25.h, z3.h\n"
+      "addvl x9, x9, #1\n"
       "ld1h { z6.h }, p4/Z, [x12]\n"
       "fmla z11.h, p4/M, z24.h, z0.h\n"
       "fmla z15.h, p4/M, z24.h, z1.h\n"
@@ -797,18 +795,18 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z16.h, p4/M, z6.h, z2.h\n"
       "fmla z20.h, p4/M, z6.h, z3.h\n"
       "ld1h { z25.h }, p4/Z, [x10]\n"
-      "addvl x12, x12, #1\n"
+      "cmp x28, x20\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
+      "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
-      "addvl x10, x10, #1\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
       "fmla z21.h, p4/M, z7.h, z3.h\n"
       "ld1h { z24.h }, p4/Z, [x9]\n"
-      "cmp x28, x20\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.h, p4/M, z25.h, z0.h\n"
       "fmla z14.h, p4/M, z25.h, z1.h\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, p4/M, z25.h, z2.h\n"
       "fmla z22.h, p4/M, z25.h, z3.h\n"
       "fmla z11.h, p4/M, z24.h, z0.h\n"
@@ -817,13 +815,13 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z23.h, p4/M, z24.h, z3.h\n"
       "bne 46b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z25.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z25.h }, p4/Z, [x21]\n"
       "ld1rh { z24.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z25.h\n"
       "fmin z9.h, p4/M, z9.h, z25.h\n"
@@ -863,18 +861,18 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p3, [x25]\n"
-      "st1h { z17.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p1, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p0, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p3, [x24]\n"
-      "st1h { z21.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p1, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p0, [x24, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p3, [x24]\n"
+      "st1h { z17.h }, p2, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p0, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p3, [x23]\n"
+      "st1h { z21.h }, p2, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p1, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p0, [x23, #3, MUL VL]\n"
       "52:"  // Height 4: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -882,15 +880,15 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "b 80f\n"
       "53:"  // Height 5
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "54:"  // Height 5: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -898,12 +896,12 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 55f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 55f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 55f\n"
       "mov x11, x12\n"
       "55:"  // Height 5: B setup done
@@ -918,18 +916,18 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "cbz x15, 56f\n"
       "ld1h { z8.h }, p4/Z, [x15]\n"
       "ld1h { z9.h }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -941,16 +939,16 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "56:"  // Height 5: no bias
       "tbz %x[flags], #0, 57f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x13]\n"
+      "add x21, x22, x20, LSL #1\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z9.h }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x13, #2, MUL VL]\n"
       "ld1h { z11.h }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x23, x13, x20, LSL #1\n"
-      "add x22, x23, x20, LSL #1\n"
-      "add x21, x22, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x23]\n"
       "ld1h { z13.h }, p2/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z14.h }, p1/Z, [x23, #2, MUL VL]\n"
       "ld1h { z15.h }, p0/Z, [x23, #3, MUL VL]\n"
       "ld1h { z16.h }, p3/Z, [x22]\n"
@@ -991,8 +989,8 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "59:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 60f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1035,8 +1033,8 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "add x26, x26, #0x2\n"
       "subs x27, x27, #0x1\n"
       "fmla z24.h, p4/M, z6.h, z4.h\n"
-      "ld1h { z29.h }, p4/Z, [x10]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z29.h }, p4/Z, [x10]\n"
       "add x25, x25, #0x2\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
@@ -1046,23 +1044,23 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z25.h, p4/M, z7.h, z4.h\n"
       "ld1h { z28.h }, p4/Z, [x9]\n"
       "add x22, x22, #0x2\n"
-      "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.h, p4/M, z29.h, z0.h\n"
       "fmla z14.h, p4/M, z29.h, z1.h\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, p4/M, z29.h, z2.h\n"
       "fmla z22.h, p4/M, z29.h, z3.h\n"
       "fmla z26.h, p4/M, z29.h, z4.h\n"
-      "ld1h { z6.h }, p4/Z, [x12]\n"
       "fmla z11.h, p4/M, z28.h, z0.h\n"
       "ld1rh { z0.h }, p4/Z, [x26]\n"
+      "ld1h { z6.h }, p4/Z, [x12]\n"
       "fmla z15.h, p4/M, z28.h, z1.h\n"
-      "ld1rh { z1.h }, p4/Z, [x25]\n"
       "fmla z19.h, p4/M, z28.h, z2.h\n"
+      "ld1rh { z1.h }, p4/Z, [x25]\n"
       "ld1rh { z2.h }, p4/Z, [x24]\n"
       "fmla z23.h, p4/M, z28.h, z3.h\n"
-      "ld1rh { z3.h }, p4/Z, [x23]\n"
       "fmla z27.h, p4/M, z28.h, z4.h\n"
+      "ld1rh { z3.h }, p4/Z, [x23]\n"
       "ld1rh { z4.h }, p4/Z, [x22]\n"
       "ld1h { z7.h }, p4/Z, [x11]\n"
       "bgt 62b\n"
@@ -1073,15 +1071,15 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
       "fmla z20.h, p4/M, z6.h, z3.h\n"
+      "cmp x28, x20\n"
       "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z24.h, p4/M, z6.h, z4.h\n"
-      "ld1h { z29.h }, p4/Z, [x10]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
-      "addvl x10, x10, #1\n"
+      "ld1h { z29.h }, p4/Z, [x10]\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
-      "cmp x28, x20\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.h, p4/M, z7.h, z3.h\n"
       "fmla z25.h, p4/M, z7.h, z4.h\n"
       "ld1h { z28.h }, p4/Z, [x9]\n"
@@ -1098,14 +1096,14 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z27.h, p4/M, z28.h, z4.h\n"
       "bne 59b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z29.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z29.h }, p4/Z, [x21]\n"
       "ld1rh { z28.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z29.h\n"
       "fmin z9.h, p4/M, z9.h, z29.h\n"
@@ -1153,22 +1151,22 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p3, [x25]\n"
-      "st1h { z17.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p1, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p0, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p3, [x24]\n"
-      "st1h { z21.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p1, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p0, [x24, #3, MUL VL]\n"
-      "st1h { z24.h }, p3, [x23]\n"
-      "st1h { z25.h }, p2, [x23, #1, MUL VL]\n"
-      "st1h { z26.h }, p1, [x23, #2, MUL VL]\n"
-      "st1h { z27.h }, p0, [x23, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p3, [x24]\n"
+      "st1h { z17.h }, p2, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p0, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p3, [x23]\n"
+      "st1h { z21.h }, p2, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p1, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p0, [x23, #3, MUL VL]\n"
+      "st1h { z24.h }, p3, [x22]\n"
+      "st1h { z25.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z26.h }, p1, [x22, #2, MUL VL]\n"
+      "st1h { z27.h }, p0, [x22, #3, MUL VL]\n"
       "65:"  // Height 5: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1176,19 +1174,18 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "b 80f\n"
       "66:"  // Height 6
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0xc\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0xc\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "67:"  // Height 6: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -1196,12 +1193,12 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 68f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 68f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 68f\n"
       "mov x11, x12\n"
       "68:"  // Height 6: B setup done
@@ -1216,18 +1213,18 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "cbz x15, 69f\n"
       "ld1h { z8.h }, p4/Z, [x15]\n"
       "ld1h { z9.h }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1243,17 +1240,17 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "69:"  // Height 6: no bias
       "tbz %x[flags], #0, 70f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x13, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x13]\n"
+      "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "ld1h { z9.h }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x13, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z11.h }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x24, x13, x20, LSL #1\n"
-      "add x23, x24, x20, LSL #1\n"
-      "add x22, x23, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x24]\n"
       "ld1h { z13.h }, p2/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z14.h }, p1/Z, [x24, #2, MUL VL]\n"
       "ld1h { z15.h }, p0/Z, [x24, #3, MUL VL]\n"
       "ld1h { z16.h }, p3/Z, [x23]\n"
@@ -1302,8 +1299,8 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "72:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 73f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1365,9 +1362,9 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z29.h, p4/M, z7.h, z5.h\n"
       "ld1h { z7.h }, p4/Z, [x9]\n"
       "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.h, p4/M, z6.h, z0.h\n"
       "fmla z14.h, p4/M, z6.h, z1.h\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, p4/M, z6.h, z2.h\n"
       "fmla z22.h, p4/M, z6.h, z3.h\n"
       "fmla z26.h, p4/M, z6.h, z4.h\n"
@@ -1394,15 +1391,15 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
       "fmla z20.h, p4/M, z6.h, z3.h\n"
+      "cmp x28, x20\n"
       "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z24.h, p4/M, z6.h, z4.h\n"
       "fmla z28.h, p4/M, z6.h, z5.h\n"
       "ld1h { z6.h }, p4/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
-      "cmp x28, x20\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
       "fmla z21.h, p4/M, z7.h, z3.h\n"
       "fmla z25.h, p4/M, z7.h, z4.h\n"
@@ -1423,15 +1420,15 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "fmla z31.h, p4/M, z7.h, z5.h\n"
       "bne 72b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "tbz %x[flags], #1, 77f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z1.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z1.h }, p4/Z, [x21]\n"
       "ld1rh { z0.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z1.h\n"
       "fmin z9.h, p4/M, z9.h, z1.h\n"
@@ -1487,26 +1484,26 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p3, [x25]\n"
-      "st1h { z17.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p1, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p0, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p3, [x24]\n"
-      "st1h { z21.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p1, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p0, [x24, #3, MUL VL]\n"
-      "st1h { z24.h }, p3, [x23]\n"
-      "st1h { z25.h }, p2, [x23, #1, MUL VL]\n"
-      "st1h { z26.h }, p1, [x23, #2, MUL VL]\n"
-      "st1h { z27.h }, p0, [x23, #3, MUL VL]\n"
-      "st1h { z28.h }, p3, [x22]\n"
-      "st1h { z29.h }, p2, [x22, #1, MUL VL]\n"
-      "st1h { z30.h }, p1, [x22, #2, MUL VL]\n"
-      "st1h { z31.h }, p0, [x22, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p3, [x24]\n"
+      "st1h { z17.h }, p2, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p0, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p3, [x23]\n"
+      "st1h { z21.h }, p2, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p1, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p0, [x23, #3, MUL VL]\n"
+      "st1h { z24.h }, p3, [x22]\n"
+      "st1h { z25.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z26.h }, p1, [x22, #2, MUL VL]\n"
+      "st1h { z27.h }, p0, [x22, #3, MUL VL]\n"
+      "st1h { z28.h }, p3, [x21]\n"
+      "st1h { z29.h }, p2, [x21, #1, MUL VL]\n"
+      "st1h { z30.h }, p1, [x21, #2, MUL VL]\n"
+      "st1h { z31.h }, p0, [x21, #3, MUL VL]\n"
       "78:"  // Height 6: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1523,8 +1520,8 @@ void sve_ffhybrid_fp16_mla_6x4VL_a64fx (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp
index 78a84fd89b..d5ccf3476c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp16_mla_6x4VL/generic.cpp
@@ -49,19 +49,18 @@ void sve_ffhybrid_fp16_mla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const __fp16 *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -82,7 +81,6 @@ void sve_ffhybrid_fp16_mla_6x4VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -108,15 +106,15 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "bgt 29f\n"
       "beq 15f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -124,12 +122,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 3f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 3f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 3f\n"
       "mov x11, x12\n"
       "3:"  // Height 1: B setup done
@@ -164,8 +162,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "7:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 8f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -181,201 +179,201 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 11f\n"
       "10:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x12]\n"
-      "ld1h { z16.h }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "fmla z8.h, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x12]\n"
+      "fmla z8.h, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x11]\n"
       "fmla z9.h, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10]\n"
+      "fmla z10.h, z16.h, z0.h[0]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "fmla z10.h, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #1, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #1, MUL VL]\n"
+      "fmla z8.h, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n"
-      "fmla z8.h, z17.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[1]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "fmla z10.h, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
-      "fmla z10.h, z17.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[1]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #2, MUL VL]\n"
+      "fmla z8.h, z16.h, z0.h[2]\n"
       "ld1h { z16.h }, p5/Z, [x11, #2, MUL VL]\n"
-      "fmla z8.h, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[2]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z10.h, z16.h, z0.h[2]\n"
       "ld1h { z16.h }, p5/Z, [x9, #2, MUL VL]\n"
-      "fmla z10.h, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #3, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[2]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #3, MUL VL]\n"
+      "fmla z8.h, z16.h, z0.h[3]\n"
       "ld1h { z16.h }, p5/Z, [x11, #3, MUL VL]\n"
-      "fmla z8.h, z17.h, z0.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[3]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
+      "fmla z10.h, z16.h, z0.h[3]\n"
       "ld1h { z16.h }, p5/Z, [x9, #3, MUL VL]\n"
-      "fmla z10.h, z17.h, z0.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #4, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[3]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #4, MUL VL]\n"
+      "fmla z8.h, z16.h, z0.h[4]\n"
       "ld1h { z16.h }, p5/Z, [x11, #4, MUL VL]\n"
-      "fmla z8.h, z17.h, z0.h[4]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[4]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "fmla z10.h, z16.h, z0.h[4]\n"
       "ld1h { z16.h }, p5/Z, [x9, #4, MUL VL]\n"
-      "fmla z10.h, z17.h, z0.h[4]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #5, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[4]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #5, MUL VL]\n"
+      "fmla z8.h, z16.h, z0.h[5]\n"
       "ld1h { z16.h }, p5/Z, [x11, #5, MUL VL]\n"
-      "fmla z8.h, z17.h, z0.h[5]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[5]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
+      "fmla z10.h, z16.h, z0.h[5]\n"
       "ld1h { z16.h }, p5/Z, [x9, #5, MUL VL]\n"
-      "fmla z10.h, z17.h, z0.h[5]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #6, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[5]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #6, MUL VL]\n"
+      "fmla z8.h, z16.h, z0.h[6]\n"
       "ld1h { z16.h }, p5/Z, [x11, #6, MUL VL]\n"
-      "fmla z8.h, z17.h, z0.h[6]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[6]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #6, MUL VL]\n"
+      "fmla z10.h, z16.h, z0.h[6]\n"
       "ld1h { z16.h }, p5/Z, [x9, #6, MUL VL]\n"
-      "fmla z10.h, z17.h, z0.h[6]\n"
-      "ld1h { z17.h }, p5/Z, [x12, #7, MUL VL]\n"
-      "addvl x12, x12, #8\n"
       "fmla z11.h, z16.h, z0.h[6]\n"
+      "ld1h { z16.h }, p5/Z, [x12, #7, MUL VL]\n"
+      "fmla z8.h, z16.h, z0.h[7]\n"
       "ld1h { z16.h }, p5/Z, [x11, #7, MUL VL]\n"
-      "addvl x11, x11, #8\n"
-      "fmla z8.h, z17.h, z0.h[7]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       "fmla z9.h, z16.h, z0.h[7]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #7, MUL VL]\n"
+      "sub x27, x27, #0x8\n"
       "ld1h { z16.h }, p5/Z, [x9, #7, MUL VL]\n"
-      "addvl x9, x9, #8\n"
+      "cmp x27, #0x8\n"
       "fmla z10.h, z17.h, z0.h[7]\n"
       "fmla z11.h, z16.h, z0.h[7]\n"
+      "add x26, x26, #0x10\n"
+      "addvl x12, x12, #8\n"
+      "addvl x11, x11, #8\n"
+      "addvl x10, x10, #8\n"
+      "addvl x9, x9, #8\n"
       "bgt 10b\n"
       "11:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x12]\n"
+      "ld1rqh { z0.h }, p0/Z, [x26]\n"
+      "ld1h { z16.h }, p5/Z, [x12]\n"
+      "fmla z8.h, z16.h, z0.h[0]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
+      "fmla z9.h, z16.h, z0.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "subs x27, x27, #0x1\n"
+      "ld1h { z16.h }, p5/Z, [x9]\n"
+      "fmla z10.h, z17.h, z0.h[0]\n"
+      "fmla z11.h, z16.h, z0.h[0]\n"
       "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
-      "ld1rqh { z0.h }, p0/Z, [x26]\n"
-      "fmla z8.h, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
       "addvl x10, x10, #1\n"
-      "fmla z9.h, z16.h, z0.h[0]\n"
-      "ld1h { z16.h }, p5/Z, [x9]\n"
       "addvl x9, x9, #1\n"
-      "fmla z10.h, z17.h, z0.h[0]\n"
-      "fmla z11.h, z16.h, z0.h[0]\n"
       "ble 12f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[1]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[1]\n"
       "fmla z11.h, z16.h, z0.h[1]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 12f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[2]\n"
       "fmla z11.h, z16.h, z0.h[2]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 12f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[3]\n"
       "fmla z11.h, z16.h, z0.h[3]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 12f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[4]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[4]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[4]\n"
       "fmla z11.h, z16.h, z0.h[4]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 12f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[5]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[5]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[5]\n"
       "fmla z11.h, z16.h, z0.h[5]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 12f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[6]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[6]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[6]\n"
       "fmla z11.h, z16.h, z0.h[6]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 12f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[7]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[7]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.h, z17.h, z0.h[7]\n"
       "fmla z11.h, z16.h, z0.h[7]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "12:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 7b\n"
       "tbz %x[flags], #1, 13f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z17.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z17.h }, p5/Z, [x21]\n"
       "ld1rh { z16.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z17.h\n"
       "fmin z9.h, p5/M, z9.h, z17.h\n"
@@ -398,15 +396,15 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "b 86f\n"
       "15:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "16:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -414,12 +412,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 17f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 17f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 17f\n"
       "mov x11, x12\n"
       "17:"  // Height 2: B setup done
@@ -434,22 +432,22 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cbz x15, 18f\n"
       "ld1h { z8.h }, p5/Z, [x15]\n"
       "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x15, x15, #4\n"
       "b 20f\n"
       "18:"  // Height 2: no bias
       "tbz %x[flags], #0, 19f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x13, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x13]\n"
       "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x20, x13, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x20]\n"
       "ld1h { z13.h }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1h { z14.h }, p2/Z, [x20, #2, MUL VL]\n"
@@ -468,8 +466,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "21:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 22f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -488,29 +486,29 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 25f\n"
       "24:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x12]\n"
-      "ld1h { z16.h }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
+      "ld1h { z17.h }, p5/Z, [x12]\n"
+      "ld1h { z16.h }, p5/Z, [x11]\n"
       "fmla z8.h, z17.h, z1.h[0]\n"
       "fmla z12.h, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
       "fmla z9.h, z16.h, z1.h[0]\n"
       "fmla z13.h, z16.h, z0.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
       "fmla z10.h, z17.h, z1.h[0]\n"
       "fmla z14.h, z17.h, z0.h[0]\n"
       "ld1h { z17.h }, p5/Z, [x12, #1, MUL VL]\n"
+      "cmp x27, #0x8\n"
       "fmla z11.h, z16.h, z1.h[0]\n"
       "fmla z15.h, z16.h, z0.h[0]\n"
       "ld1h { z16.h }, p5/Z, [x11, #1, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       "fmla z8.h, z17.h, z1.h[1]\n"
       "fmla z12.h, z17.h, z0.h[1]\n"
       "ld1h { z17.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       "fmla z9.h, z16.h, z1.h[1]\n"
       "fmla z13.h, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x9, #1, MUL VL]\n"
@@ -597,161 +595,161 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "bgt 24b\n"
       "25:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x12]\n"
-      "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
+      "ld1h { z17.h }, p5/Z, [x12]\n"
+      "ld1h { z16.h }, p5/Z, [x11]\n"
       "fmla z8.h, z17.h, z0.h[0]\n"
       "fmla z12.h, z17.h, z1.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[0]\n"
       "fmla z13.h, z16.h, z1.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.h, z17.h, z0.h[0]\n"
       "fmla z14.h, z17.h, z1.h[0]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z11.h, z16.h, z0.h[0]\n"
       "fmla z15.h, z16.h, z1.h[0]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[1]\n"
       "fmla z12.h, z17.h, z1.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[1]\n"
       "fmla z13.h, z16.h, z1.h[1]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[1]\n"
       "fmla z14.h, z17.h, z1.h[1]\n"
+      "addvl x12, x12, #1\n"
       "fmla z11.h, z16.h, z0.h[1]\n"
       "fmla z15.h, z16.h, z1.h[1]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[2]\n"
       "fmla z12.h, z17.h, z1.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[2]\n"
       "fmla z13.h, z16.h, z1.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[2]\n"
       "fmla z14.h, z17.h, z1.h[2]\n"
+      "addvl x12, x12, #1\n"
       "fmla z11.h, z16.h, z0.h[2]\n"
       "fmla z15.h, z16.h, z1.h[2]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[3]\n"
       "fmla z12.h, z17.h, z1.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[3]\n"
       "fmla z13.h, z16.h, z1.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[3]\n"
       "fmla z14.h, z17.h, z1.h[3]\n"
+      "addvl x12, x12, #1\n"
       "fmla z11.h, z16.h, z0.h[3]\n"
       "fmla z15.h, z16.h, z1.h[3]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[4]\n"
       "fmla z12.h, z17.h, z1.h[4]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[4]\n"
       "fmla z13.h, z16.h, z1.h[4]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[4]\n"
       "fmla z14.h, z17.h, z1.h[4]\n"
+      "addvl x12, x12, #1\n"
       "fmla z11.h, z16.h, z0.h[4]\n"
       "fmla z15.h, z16.h, z1.h[4]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[5]\n"
       "fmla z12.h, z17.h, z1.h[5]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[5]\n"
       "fmla z13.h, z16.h, z1.h[5]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[5]\n"
       "fmla z14.h, z17.h, z1.h[5]\n"
+      "addvl x12, x12, #1\n"
       "fmla z11.h, z16.h, z0.h[5]\n"
       "fmla z15.h, z16.h, z1.h[5]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[6]\n"
       "fmla z12.h, z17.h, z1.h[6]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[6]\n"
       "fmla z13.h, z16.h, z1.h[6]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[6]\n"
       "fmla z14.h, z17.h, z1.h[6]\n"
+      "addvl x12, x12, #1\n"
       "fmla z11.h, z16.h, z0.h[6]\n"
       "fmla z15.h, z16.h, z1.h[6]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1h { z17.h }, p5/Z, [x12]\n"
       "ld1h { z16.h }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z17.h, z0.h[7]\n"
       "fmla z12.h, z17.h, z1.h[7]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z16.h, z0.h[7]\n"
       "fmla z13.h, z16.h, z1.h[7]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.h, z17.h, z0.h[7]\n"
       "fmla z14.h, z17.h, z1.h[7]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z11.h, z16.h, z0.h[7]\n"
       "fmla z15.h, z16.h, z1.h[7]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "26:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 21b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "tbz %x[flags], #1, 27f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z17.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z17.h }, p5/Z, [x21]\n"
       "ld1rh { z16.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z17.h\n"
       "fmin z9.h, p5/M, z9.h, z17.h\n"
@@ -775,10 +773,10 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
       "28:"  // Height 2: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -786,15 +784,15 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "b 86f\n"
       "29:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "30:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -802,12 +800,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 31f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 31f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 31f\n"
       "mov x11, x12\n"
       "31:"  // Height 3: B setup done
@@ -822,27 +820,27 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cbz x15, 32f\n"
       "ld1h { z8.h }, p5/Z, [x15]\n"
       "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 34f\n"
       "32:"  // Height 3: no bias
       "tbz %x[flags], #0, 33f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x13, x20, LSL #1\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x13]\n"
       "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x21, x13, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x21]\n"
       "ld1h { z13.h }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1h { z14.h }, p2/Z, [x21, #2, MUL VL]\n"
@@ -869,8 +867,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "35:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 36f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -892,126 +890,126 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 39f\n"
       "38:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z21.h }, p5/Z, [x12]\n"
-      "ld1h { z20.h }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z2.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z0.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1h { z21.h }, p5/Z, [x12]\n"
       "fmla z8.h, z21.h, z2.h[0]\n"
       "fmla z12.h, z21.h, z1.h[0]\n"
-      "fmla z9.h, z20.h, z2.h[0]\n"
-      "fmla z13.h, z20.h, z1.h[0]\n"
+      "ld1h { z20.h }, p5/Z, [x11]\n"
       "fmla z16.h, z21.h, z0.h[0]\n"
+      "fmla z9.h, z20.h, z2.h[0]\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
+      "fmla z13.h, z20.h, z1.h[0]\n"
       "fmla z17.h, z20.h, z0.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
+      "cmp x27, #0x8\n"
       "fmla z10.h, z21.h, z2.h[0]\n"
       "fmla z14.h, z21.h, z1.h[0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "fmla z18.h, z21.h, z0.h[0]\n"
-      "ld1h { z21.h }, p5/Z, [x12, #1, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[0]\n"
+      "ld1h { z21.h }, p5/Z, [x12, #1, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       "fmla z15.h, z20.h, z1.h[0]\n"
       "fmla z19.h, z20.h, z0.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x11, #1, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[1]\n"
       "fmla z12.h, z21.h, z1.h[1]\n"
       "fmla z16.h, z21.h, z0.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[1]\n"
       "fmla z17.h, z20.h, z0.h[1]\n"
       "ld1h { z20.h }, p5/Z, [x9, #1, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[1]\n"
       "fmla z14.h, z21.h, z1.h[1]\n"
       "fmla z18.h, z21.h, z0.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[1]\n"
       "fmla z19.h, z20.h, z0.h[1]\n"
       "ld1h { z20.h }, p5/Z, [x11, #2, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[2]\n"
       "fmla z12.h, z21.h, z1.h[2]\n"
       "fmla z16.h, z21.h, z0.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[2]\n"
       "fmla z17.h, z20.h, z0.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x9, #2, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[2]\n"
       "fmla z14.h, z21.h, z1.h[2]\n"
       "fmla z18.h, z21.h, z0.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x12, #3, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x12, #3, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[2]\n"
       "fmla z19.h, z20.h, z0.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x11, #3, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[3]\n"
       "fmla z12.h, z21.h, z1.h[3]\n"
       "fmla z16.h, z21.h, z0.h[3]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[3]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[3]\n"
       "fmla z17.h, z20.h, z0.h[3]\n"
       "ld1h { z20.h }, p5/Z, [x9, #3, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[3]\n"
       "fmla z14.h, z21.h, z1.h[3]\n"
       "fmla z18.h, z21.h, z0.h[3]\n"
-      "ld1h { z21.h }, p5/Z, [x12, #4, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[3]\n"
+      "ld1h { z21.h }, p5/Z, [x12, #4, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[3]\n"
       "fmla z19.h, z20.h, z0.h[3]\n"
       "ld1h { z20.h }, p5/Z, [x11, #4, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[4]\n"
       "fmla z12.h, z21.h, z1.h[4]\n"
       "fmla z16.h, z21.h, z0.h[4]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[4]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[4]\n"
       "fmla z17.h, z20.h, z0.h[4]\n"
       "ld1h { z20.h }, p5/Z, [x9, #4, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[4]\n"
       "fmla z14.h, z21.h, z1.h[4]\n"
       "fmla z18.h, z21.h, z0.h[4]\n"
-      "ld1h { z21.h }, p5/Z, [x12, #5, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[4]\n"
+      "ld1h { z21.h }, p5/Z, [x12, #5, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[4]\n"
       "fmla z19.h, z20.h, z0.h[4]\n"
       "ld1h { z20.h }, p5/Z, [x11, #5, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[5]\n"
       "fmla z12.h, z21.h, z1.h[5]\n"
       "fmla z16.h, z21.h, z0.h[5]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[5]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[5]\n"
       "fmla z17.h, z20.h, z0.h[5]\n"
       "ld1h { z20.h }, p5/Z, [x9, #5, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[5]\n"
       "fmla z14.h, z21.h, z1.h[5]\n"
       "fmla z18.h, z21.h, z0.h[5]\n"
-      "ld1h { z21.h }, p5/Z, [x12, #6, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[5]\n"
+      "ld1h { z21.h }, p5/Z, [x12, #6, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[5]\n"
       "fmla z19.h, z20.h, z0.h[5]\n"
       "ld1h { z20.h }, p5/Z, [x11, #6, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[6]\n"
       "fmla z12.h, z21.h, z1.h[6]\n"
       "fmla z16.h, z21.h, z0.h[6]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[6]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[6]\n"
       "fmla z17.h, z20.h, z0.h[6]\n"
       "ld1h { z20.h }, p5/Z, [x9, #6, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[6]\n"
       "fmla z14.h, z21.h, z1.h[6]\n"
       "fmla z18.h, z21.h, z0.h[6]\n"
+      "fmla z11.h, z20.h, z2.h[6]\n"
       "ld1h { z21.h }, p5/Z, [x12, #7, MUL VL]\n"
       "addvl x12, x12, #8\n"
-      "fmla z11.h, z20.h, z2.h[6]\n"
       "fmla z15.h, z20.h, z1.h[6]\n"
       "fmla z19.h, z20.h, z0.h[6]\n"
       "ld1h { z20.h }, p5/Z, [x11, #7, MUL VL]\n"
@@ -1019,9 +1017,9 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z8.h, z21.h, z2.h[7]\n"
       "fmla z12.h, z21.h, z1.h[7]\n"
       "fmla z16.h, z21.h, z0.h[7]\n"
+      "fmla z9.h, z20.h, z2.h[7]\n"
       "ld1h { z21.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #8\n"
-      "fmla z9.h, z20.h, z2.h[7]\n"
       "fmla z13.h, z20.h, z1.h[7]\n"
       "fmla z17.h, z20.h, z0.h[7]\n"
       "ld1h { z20.h }, p5/Z, [x9, #7, MUL VL]\n"
@@ -1035,179 +1033,179 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "bgt 38b\n"
       "39:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z21.h }, p5/Z, [x12]\n"
-      "ld1h { z20.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
+      "ld1h { z21.h }, p5/Z, [x12]\n"
       "fmla z8.h, z21.h, z0.h[0]\n"
       "fmla z12.h, z21.h, z1.h[0]\n"
-      "fmla z9.h, z20.h, z0.h[0]\n"
-      "fmla z13.h, z20.h, z1.h[0]\n"
+      "ld1h { z20.h }, p5/Z, [x11]\n"
       "fmla z16.h, z21.h, z2.h[0]\n"
+      "fmla z9.h, z20.h, z0.h[0]\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "fmla z13.h, z20.h, z1.h[0]\n"
       "fmla z17.h, z20.h, z2.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.h, z21.h, z0.h[0]\n"
       "fmla z14.h, z21.h, z1.h[0]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.h, z21.h, z2.h[0]\n"
       "fmla z11.h, z20.h, z0.h[0]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, z20.h, z1.h[0]\n"
       "fmla z19.h, z20.h, z2.h[0]\n"
       "ble 40f\n"
       "ld1h { z21.h }, p5/Z, [x12]\n"
       "ld1h { z20.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z21.h, z0.h[1]\n"
       "fmla z12.h, z21.h, z1.h[1]\n"
       "fmla z16.h, z21.h, z2.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z20.h, z0.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[1]\n"
       "fmla z17.h, z20.h, z2.h[1]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.h, z21.h, z0.h[1]\n"
       "fmla z14.h, z21.h, z1.h[1]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.h, z21.h, z2.h[1]\n"
       "fmla z11.h, z20.h, z0.h[1]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, z20.h, z1.h[1]\n"
       "fmla z19.h, z20.h, z2.h[1]\n"
       "ble 40f\n"
       "ld1h { z21.h }, p5/Z, [x12]\n"
       "ld1h { z20.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z21.h, z0.h[2]\n"
       "fmla z12.h, z21.h, z1.h[2]\n"
       "fmla z16.h, z21.h, z2.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z20.h, z0.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[2]\n"
       "fmla z17.h, z20.h, z2.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.h, z21.h, z0.h[2]\n"
       "fmla z14.h, z21.h, z1.h[2]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.h, z21.h, z2.h[2]\n"
       "fmla z11.h, z20.h, z0.h[2]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, z20.h, z1.h[2]\n"
       "fmla z19.h, z20.h, z2.h[2]\n"
       "ble 40f\n"
       "ld1h { z21.h }, p5/Z, [x12]\n"
       "ld1h { z20.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z21.h, z0.h[3]\n"
       "fmla z12.h, z21.h, z1.h[3]\n"
       "fmla z16.h, z21.h, z2.h[3]\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z20.h, z0.h[3]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[3]\n"
       "fmla z17.h, z20.h, z2.h[3]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.h, z21.h, z0.h[3]\n"
       "fmla z14.h, z21.h, z1.h[3]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.h, z21.h, z2.h[3]\n"
       "fmla z11.h, z20.h, z0.h[3]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, z20.h, z1.h[3]\n"
       "fmla z19.h, z20.h, z2.h[3]\n"
       "ble 40f\n"
       "ld1h { z21.h }, p5/Z, [x12]\n"
       "ld1h { z20.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z21.h, z0.h[4]\n"
       "fmla z12.h, z21.h, z1.h[4]\n"
       "fmla z16.h, z21.h, z2.h[4]\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z20.h, z0.h[4]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[4]\n"
       "fmla z17.h, z20.h, z2.h[4]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.h, z21.h, z0.h[4]\n"
       "fmla z14.h, z21.h, z1.h[4]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.h, z21.h, z2.h[4]\n"
       "fmla z11.h, z20.h, z0.h[4]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, z20.h, z1.h[4]\n"
       "fmla z19.h, z20.h, z2.h[4]\n"
       "ble 40f\n"
       "ld1h { z21.h }, p5/Z, [x12]\n"
       "ld1h { z20.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z21.h, z0.h[5]\n"
       "fmla z12.h, z21.h, z1.h[5]\n"
       "fmla z16.h, z21.h, z2.h[5]\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z20.h, z0.h[5]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[5]\n"
       "fmla z17.h, z20.h, z2.h[5]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.h, z21.h, z0.h[5]\n"
       "fmla z14.h, z21.h, z1.h[5]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.h, z21.h, z2.h[5]\n"
       "fmla z11.h, z20.h, z0.h[5]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, z20.h, z1.h[5]\n"
       "fmla z19.h, z20.h, z2.h[5]\n"
       "ble 40f\n"
       "ld1h { z21.h }, p5/Z, [x12]\n"
       "ld1h { z20.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z21.h, z0.h[6]\n"
       "fmla z12.h, z21.h, z1.h[6]\n"
       "fmla z16.h, z21.h, z2.h[6]\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z20.h, z0.h[6]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[6]\n"
       "fmla z17.h, z20.h, z2.h[6]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.h, z21.h, z0.h[6]\n"
       "fmla z14.h, z21.h, z1.h[6]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.h, z21.h, z2.h[6]\n"
       "fmla z11.h, z20.h, z0.h[6]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.h, z20.h, z1.h[6]\n"
       "fmla z19.h, z20.h, z2.h[6]\n"
       "ble 40f\n"
       "ld1h { z21.h }, p5/Z, [x12]\n"
       "ld1h { z20.h }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z21.h, z0.h[7]\n"
       "fmla z12.h, z21.h, z1.h[7]\n"
       "fmla z16.h, z21.h, z2.h[7]\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.h, z20.h, z0.h[7]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
+      "addvl x12, x12, #1\n"
       "fmla z13.h, z20.h, z1.h[7]\n"
       "fmla z17.h, z20.h, z2.h[7]\n"
       "ld1h { z20.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z10.h, z21.h, z0.h[7]\n"
       "fmla z14.h, z21.h, z1.h[7]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, z21.h, z2.h[7]\n"
       "fmla z11.h, z20.h, z0.h[7]\n"
       "fmla z15.h, z20.h, z1.h[7]\n"
@@ -1218,12 +1216,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 35b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "tbz %x[flags], #1, 41f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z21.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z21.h }, p5/Z, [x21]\n"
       "ld1rh { z20.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z21.h\n"
       "fmin z9.h, p5/M, z9.h, z21.h\n"
@@ -1255,14 +1253,14 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p4, [x25]\n"
-      "st1h { z17.h }, p3, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p4, [x24]\n"
+      "st1h { z17.h }, p3, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p1, [x24, #3, MUL VL]\n"
       "42:"  // Height 3: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1270,15 +1268,15 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "b 86f\n"
       "43:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "44:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -1286,12 +1284,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 45f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 45f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 45f\n"
       "mov x11, x12\n"
       "45:"  // Height 4: B setup done
@@ -1306,18 +1304,18 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cbz x15, 46f\n"
       "ld1h { z8.h }, p5/Z, [x15]\n"
       "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1325,13 +1323,13 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "46:"  // Height 4: no bias
       "tbz %x[flags], #0, 47f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x13, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x13]\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x22, x13, x20, LSL #1\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x22]\n"
       "ld1h { z13.h }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1h { z14.h }, p2/Z, [x22, #2, MUL VL]\n"
@@ -1366,8 +1364,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "49:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 50f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1392,25 +1390,25 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 53f\n"
       "52:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x12]\n"
-      "ld1h { z24.h }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z3.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z2.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z1.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
+      "ld1h { z25.h }, p5/Z, [x12]\n"
+      "ld1h { z24.h }, p5/Z, [x11]\n"
       "fmla z8.h, z25.h, z3.h[0]\n"
       "fmla z12.h, z25.h, z2.h[0]\n"
-      "fmla z9.h, z24.h, z3.h[0]\n"
-      "fmla z13.h, z24.h, z2.h[0]\n"
       "fmla z16.h, z25.h, z1.h[0]\n"
       "fmla z20.h, z25.h, z0.h[0]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
+      "add x25, x25, #0x10\n"
+      "fmla z9.h, z24.h, z3.h[0]\n"
+      "fmla z13.h, z24.h, z2.h[0]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "fmla z17.h, z24.h, z1.h[0]\n"
       "fmla z21.h, z24.h, z0.h[0]\n"
       "ld1h { z24.h }, p5/Z, [x9]\n"
@@ -1569,22 +1567,22 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "bgt 52b\n"
       "53:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x12]\n"
-      "ld1h { z24.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      "ld1h { z25.h }, p5/Z, [x12]\n"
+      "ld1h { z24.h }, p5/Z, [x11]\n"
       "fmla z8.h, z25.h, z0.h[0]\n"
       "fmla z12.h, z25.h, z1.h[0]\n"
-      "fmla z9.h, z24.h, z0.h[0]\n"
-      "fmla z13.h, z24.h, z1.h[0]\n"
       "fmla z16.h, z25.h, z2.h[0]\n"
       "fmla z20.h, z25.h, z3.h[0]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
+      "addvl x12, x12, #1\n"
+      "fmla z9.h, z24.h, z0.h[0]\n"
+      "fmla z13.h, z24.h, z1.h[0]\n"
+      "addvl x11, x11, #1\n"
       "addvl x10, x10, #1\n"
       "fmla z17.h, z24.h, z2.h[0]\n"
       "fmla z21.h, z24.h, z3.h[0]\n"
@@ -1601,23 +1599,23 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 54f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z25.h, z0.h[1]\n"
       "fmla z12.h, z25.h, z1.h[1]\n"
       "fmla z16.h, z25.h, z2.h[1]\n"
       "fmla z20.h, z25.h, z3.h[1]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[1]\n"
       "fmla z13.h, z24.h, z1.h[1]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z17.h, z24.h, z2.h[1]\n"
       "fmla z21.h, z24.h, z3.h[1]\n"
       "ld1h { z24.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.h, z25.h, z0.h[1]\n"
       "fmla z14.h, z25.h, z1.h[1]\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, z25.h, z2.h[1]\n"
       "fmla z22.h, z25.h, z3.h[1]\n"
       "fmla z11.h, z24.h, z0.h[1]\n"
@@ -1627,23 +1625,23 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 54f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z25.h, z0.h[2]\n"
       "fmla z12.h, z25.h, z1.h[2]\n"
       "fmla z16.h, z25.h, z2.h[2]\n"
       "fmla z20.h, z25.h, z3.h[2]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[2]\n"
       "fmla z13.h, z24.h, z1.h[2]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z17.h, z24.h, z2.h[2]\n"
       "fmla z21.h, z24.h, z3.h[2]\n"
       "ld1h { z24.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.h, z25.h, z0.h[2]\n"
       "fmla z14.h, z25.h, z1.h[2]\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, z25.h, z2.h[2]\n"
       "fmla z22.h, z25.h, z3.h[2]\n"
       "fmla z11.h, z24.h, z0.h[2]\n"
@@ -1653,23 +1651,23 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 54f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z25.h, z0.h[3]\n"
       "fmla z12.h, z25.h, z1.h[3]\n"
       "fmla z16.h, z25.h, z2.h[3]\n"
       "fmla z20.h, z25.h, z3.h[3]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[3]\n"
       "fmla z13.h, z24.h, z1.h[3]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z17.h, z24.h, z2.h[3]\n"
       "fmla z21.h, z24.h, z3.h[3]\n"
       "ld1h { z24.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.h, z25.h, z0.h[3]\n"
       "fmla z14.h, z25.h, z1.h[3]\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, z25.h, z2.h[3]\n"
       "fmla z22.h, z25.h, z3.h[3]\n"
       "fmla z11.h, z24.h, z0.h[3]\n"
@@ -1679,23 +1677,23 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 54f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z25.h, z0.h[4]\n"
       "fmla z12.h, z25.h, z1.h[4]\n"
       "fmla z16.h, z25.h, z2.h[4]\n"
       "fmla z20.h, z25.h, z3.h[4]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[4]\n"
       "fmla z13.h, z24.h, z1.h[4]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z17.h, z24.h, z2.h[4]\n"
       "fmla z21.h, z24.h, z3.h[4]\n"
       "ld1h { z24.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.h, z25.h, z0.h[4]\n"
       "fmla z14.h, z25.h, z1.h[4]\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, z25.h, z2.h[4]\n"
       "fmla z22.h, z25.h, z3.h[4]\n"
       "fmla z11.h, z24.h, z0.h[4]\n"
@@ -1705,23 +1703,23 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 54f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z25.h, z0.h[5]\n"
       "fmla z12.h, z25.h, z1.h[5]\n"
       "fmla z16.h, z25.h, z2.h[5]\n"
       "fmla z20.h, z25.h, z3.h[5]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[5]\n"
       "fmla z13.h, z24.h, z1.h[5]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z17.h, z24.h, z2.h[5]\n"
       "fmla z21.h, z24.h, z3.h[5]\n"
       "ld1h { z24.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.h, z25.h, z0.h[5]\n"
       "fmla z14.h, z25.h, z1.h[5]\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, z25.h, z2.h[5]\n"
       "fmla z22.h, z25.h, z3.h[5]\n"
       "fmla z11.h, z24.h, z0.h[5]\n"
@@ -1731,23 +1729,23 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 54f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z25.h, z0.h[6]\n"
       "fmla z12.h, z25.h, z1.h[6]\n"
       "fmla z16.h, z25.h, z2.h[6]\n"
       "fmla z20.h, z25.h, z3.h[6]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[6]\n"
       "fmla z13.h, z24.h, z1.h[6]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z17.h, z24.h, z2.h[6]\n"
       "fmla z21.h, z24.h, z3.h[6]\n"
       "ld1h { z24.h }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.h, z25.h, z0.h[6]\n"
       "fmla z14.h, z25.h, z1.h[6]\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.h, z25.h, z2.h[6]\n"
       "fmla z22.h, z25.h, z3.h[6]\n"
       "fmla z11.h, z24.h, z0.h[6]\n"
@@ -1757,16 +1755,16 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 54f\n"
       "ld1h { z25.h }, p5/Z, [x12]\n"
       "ld1h { z24.h }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z25.h, z0.h[7]\n"
       "fmla z12.h, z25.h, z1.h[7]\n"
       "fmla z16.h, z25.h, z2.h[7]\n"
       "fmla z20.h, z25.h, z3.h[7]\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z9.h, z24.h, z0.h[7]\n"
       "fmla z13.h, z24.h, z1.h[7]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.h, z24.h, z2.h[7]\n"
       "fmla z21.h, z24.h, z3.h[7]\n"
       "ld1h { z24.h }, p5/Z, [x9]\n"
@@ -1785,13 +1783,13 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 49b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "tbz %x[flags], #1, 55f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z25.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z25.h }, p5/Z, [x21]\n"
       "ld1rh { z24.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z25.h\n"
       "fmin z9.h, p5/M, z9.h, z25.h\n"
@@ -1831,18 +1829,18 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p4, [x25]\n"
-      "st1h { z17.h }, p3, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p1, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p4, [x24]\n"
-      "st1h { z21.h }, p3, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p1, [x24, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p4, [x24]\n"
+      "st1h { z17.h }, p3, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p1, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p4, [x23]\n"
+      "st1h { z21.h }, p3, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p1, [x23, #3, MUL VL]\n"
       "56:"  // Height 4: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1850,15 +1848,15 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "b 86f\n"
       "57:"  // Height 5
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "58:"  // Height 5: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -1866,12 +1864,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 59f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 59f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 59f\n"
       "mov x11, x12\n"
       "59:"  // Height 5: B setup done
@@ -1886,18 +1884,18 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cbz x15, 60f\n"
       "ld1h { z8.h }, p5/Z, [x15]\n"
       "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1909,16 +1907,16 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "60:"  // Height 5: no bias
       "tbz %x[flags], #0, 61f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x13]\n"
+      "add x21, x22, x20, LSL #1\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x23, x13, x20, LSL #1\n"
-      "add x22, x23, x20, LSL #1\n"
-      "add x21, x22, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x23]\n"
       "ld1h { z13.h }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z14.h }, p2/Z, [x23, #2, MUL VL]\n"
       "ld1h { z15.h }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1h { z16.h }, p4/Z, [x22]\n"
@@ -1959,8 +1957,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "63:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 64f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1988,29 +1986,29 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 67f\n"
       "66:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x12]\n"
-      "ld1h { z28.h }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z4.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z3.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z1.h }, p0/Z, [x23]\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1h { z29.h }, p5/Z, [x12]\n"
       "fmla z8.h, z29.h, z4.h[0]\n"
       "fmla z12.h, z29.h, z3.h[0]\n"
-      "fmla z9.h, z28.h, z4.h[0]\n"
+      "ld1h { z28.h }, p5/Z, [x11]\n"
       "fmla z16.h, z29.h, z2.h[0]\n"
       "fmla z20.h, z29.h, z1.h[0]\n"
+      "add x25, x25, #0x10\n"
       "fmla z24.h, z29.h, z0.h[0]\n"
-      "fmla z13.h, z28.h, z3.h[0]\n"
+      "fmla z9.h, z28.h, z4.h[0]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
+      "add x24, x24, #0x10\n"
+      "fmla z13.h, z28.h, z3.h[0]\n"
       "fmla z17.h, z28.h, z2.h[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z21.h, z28.h, z1.h[0]\n"
       "fmla z25.h, z28.h, z0.h[0]\n"
       "ld1h { z28.h }, p5/Z, [x9]\n"
@@ -2019,8 +2017,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[0]\n"
       "fmla z22.h, z29.h, z1.h[0]\n"
       "fmla z26.h, z29.h, z0.h[0]\n"
-      "ld1h { z29.h }, p5/Z, [x12, #1, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[0]\n"
+      "ld1h { z29.h }, p5/Z, [x12, #1, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[0]\n"
       "fmla z19.h, z28.h, z2.h[0]\n"
       "fmla z23.h, z28.h, z1.h[0]\n"
@@ -2031,8 +2029,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[1]\n"
       "fmla z20.h, z29.h, z1.h[1]\n"
       "fmla z24.h, z29.h, z0.h[1]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[1]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[1]\n"
       "fmla z17.h, z28.h, z2.h[1]\n"
       "fmla z21.h, z28.h, z1.h[1]\n"
@@ -2043,8 +2041,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[1]\n"
       "fmla z22.h, z29.h, z1.h[1]\n"
       "fmla z26.h, z29.h, z0.h[1]\n"
-      "ld1h { z29.h }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[1]\n"
+      "ld1h { z29.h }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[1]\n"
       "fmla z19.h, z28.h, z2.h[1]\n"
       "fmla z23.h, z28.h, z1.h[1]\n"
@@ -2055,8 +2053,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[2]\n"
       "fmla z20.h, z29.h, z1.h[2]\n"
       "fmla z24.h, z29.h, z0.h[2]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[2]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[2]\n"
       "fmla z17.h, z28.h, z2.h[2]\n"
       "fmla z21.h, z28.h, z1.h[2]\n"
@@ -2067,8 +2065,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[2]\n"
       "fmla z22.h, z29.h, z1.h[2]\n"
       "fmla z26.h, z29.h, z0.h[2]\n"
-      "ld1h { z29.h }, p5/Z, [x12, #3, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[2]\n"
+      "ld1h { z29.h }, p5/Z, [x12, #3, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[2]\n"
       "fmla z19.h, z28.h, z2.h[2]\n"
       "fmla z23.h, z28.h, z1.h[2]\n"
@@ -2079,8 +2077,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[3]\n"
       "fmla z20.h, z29.h, z1.h[3]\n"
       "fmla z24.h, z29.h, z0.h[3]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[3]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[3]\n"
       "fmla z17.h, z28.h, z2.h[3]\n"
       "fmla z21.h, z28.h, z1.h[3]\n"
@@ -2091,8 +2089,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[3]\n"
       "fmla z22.h, z29.h, z1.h[3]\n"
       "fmla z26.h, z29.h, z0.h[3]\n"
-      "ld1h { z29.h }, p5/Z, [x12, #4, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[3]\n"
+      "ld1h { z29.h }, p5/Z, [x12, #4, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[3]\n"
       "fmla z19.h, z28.h, z2.h[3]\n"
       "fmla z23.h, z28.h, z1.h[3]\n"
@@ -2103,8 +2101,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[4]\n"
       "fmla z20.h, z29.h, z1.h[4]\n"
       "fmla z24.h, z29.h, z0.h[4]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[4]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[4]\n"
       "fmla z17.h, z28.h, z2.h[4]\n"
       "fmla z21.h, z28.h, z1.h[4]\n"
@@ -2115,8 +2113,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[4]\n"
       "fmla z22.h, z29.h, z1.h[4]\n"
       "fmla z26.h, z29.h, z0.h[4]\n"
-      "ld1h { z29.h }, p5/Z, [x12, #5, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[4]\n"
+      "ld1h { z29.h }, p5/Z, [x12, #5, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[4]\n"
       "fmla z19.h, z28.h, z2.h[4]\n"
       "fmla z23.h, z28.h, z1.h[4]\n"
@@ -2127,8 +2125,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[5]\n"
       "fmla z20.h, z29.h, z1.h[5]\n"
       "fmla z24.h, z29.h, z0.h[5]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[5]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[5]\n"
       "fmla z17.h, z28.h, z2.h[5]\n"
       "fmla z21.h, z28.h, z1.h[5]\n"
@@ -2139,8 +2137,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[5]\n"
       "fmla z22.h, z29.h, z1.h[5]\n"
       "fmla z26.h, z29.h, z0.h[5]\n"
-      "ld1h { z29.h }, p5/Z, [x12, #6, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[5]\n"
+      "ld1h { z29.h }, p5/Z, [x12, #6, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[5]\n"
       "fmla z19.h, z28.h, z2.h[5]\n"
       "fmla z23.h, z28.h, z1.h[5]\n"
@@ -2151,8 +2149,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[6]\n"
       "fmla z20.h, z29.h, z1.h[6]\n"
       "fmla z24.h, z29.h, z0.h[6]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[6]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[6]\n"
       "fmla z17.h, z28.h, z2.h[6]\n"
       "fmla z21.h, z28.h, z1.h[6]\n"
@@ -2163,30 +2161,30 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[6]\n"
       "fmla z22.h, z29.h, z1.h[6]\n"
       "fmla z26.h, z29.h, z0.h[6]\n"
+      "fmla z11.h, z28.h, z4.h[6]\n"
       "ld1h { z29.h }, p5/Z, [x12, #7, MUL VL]\n"
       "addvl x12, x12, #8\n"
-      "fmla z11.h, z28.h, z4.h[6]\n"
       "fmla z15.h, z28.h, z3.h[6]\n"
       "fmla z19.h, z28.h, z2.h[6]\n"
       "fmla z23.h, z28.h, z1.h[6]\n"
       "fmla z27.h, z28.h, z0.h[6]\n"
       "ld1h { z28.h }, p5/Z, [x11, #7, MUL VL]\n"
-      "fmla z8.h, z29.h, z4.h[7]\n"
       "addvl x11, x11, #8\n"
+      "fmla z8.h, z29.h, z4.h[7]\n"
       "fmla z12.h, z29.h, z3.h[7]\n"
       "fmla z16.h, z29.h, z2.h[7]\n"
       "fmla z20.h, z29.h, z1.h[7]\n"
       "fmla z24.h, z29.h, z0.h[7]\n"
+      "fmla z9.h, z28.h, z4.h[7]\n"
       "ld1h { z29.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #8\n"
-      "fmla z9.h, z28.h, z4.h[7]\n"
       "fmla z13.h, z28.h, z3.h[7]\n"
       "fmla z17.h, z28.h, z2.h[7]\n"
       "fmla z21.h, z28.h, z1.h[7]\n"
       "fmla z25.h, z28.h, z0.h[7]\n"
       "ld1h { z28.h }, p5/Z, [x9, #7, MUL VL]\n"
-      "fmla z10.h, z29.h, z4.h[7]\n"
       "addvl x9, x9, #8\n"
+      "fmla z10.h, z29.h, z4.h[7]\n"
       "fmla z14.h, z29.h, z3.h[7]\n"
       "fmla z18.h, z29.h, z2.h[7]\n"
       "fmla z22.h, z29.h, z1.h[7]\n"
@@ -2199,25 +2197,25 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "bgt 66b\n"
       "67:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x12]\n"
-      "ld1h { z28.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
       "ld1rqh { z4.h }, p0/Z, [x22]\n"
+      "ld1h { z29.h }, p5/Z, [x12]\n"
       "fmla z8.h, z29.h, z0.h[0]\n"
       "fmla z12.h, z29.h, z1.h[0]\n"
-      "fmla z9.h, z28.h, z0.h[0]\n"
-      "fmla z13.h, z28.h, z1.h[0]\n"
+      "ld1h { z28.h }, p5/Z, [x11]\n"
       "fmla z16.h, z29.h, z2.h[0]\n"
       "fmla z20.h, z29.h, z3.h[0]\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z29.h, z4.h[0]\n"
-      "fmla z17.h, z28.h, z2.h[0]\n"
+      "fmla z9.h, z28.h, z0.h[0]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
+      "addvl x11, x11, #1\n"
+      "fmla z13.h, z28.h, z1.h[0]\n"
+      "fmla z17.h, z28.h, z2.h[0]\n"
       "addvl x10, x10, #1\n"
       "fmla z21.h, z28.h, z3.h[0]\n"
       "fmla z25.h, z28.h, z4.h[0]\n"
@@ -2236,19 +2234,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 68f\n"
       "ld1h { z29.h }, p5/Z, [x12]\n"
       "ld1h { z28.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z29.h, z0.h[1]\n"
       "fmla z12.h, z29.h, z1.h[1]\n"
       "fmla z16.h, z29.h, z2.h[1]\n"
       "fmla z20.h, z29.h, z3.h[1]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z29.h, z4.h[1]\n"
       "fmla z9.h, z28.h, z0.h[1]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.h, z28.h, z1.h[1]\n"
       "fmla z17.h, z28.h, z2.h[1]\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.h, z28.h, z3.h[1]\n"
       "fmla z25.h, z28.h, z4.h[1]\n"
       "ld1h { z28.h }, p5/Z, [x9]\n"
@@ -2266,19 +2264,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 68f\n"
       "ld1h { z29.h }, p5/Z, [x12]\n"
       "ld1h { z28.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z29.h, z0.h[2]\n"
       "fmla z12.h, z29.h, z1.h[2]\n"
       "fmla z16.h, z29.h, z2.h[2]\n"
       "fmla z20.h, z29.h, z3.h[2]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z29.h, z4.h[2]\n"
       "fmla z9.h, z28.h, z0.h[2]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.h, z28.h, z1.h[2]\n"
       "fmla z17.h, z28.h, z2.h[2]\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.h, z28.h, z3.h[2]\n"
       "fmla z25.h, z28.h, z4.h[2]\n"
       "ld1h { z28.h }, p5/Z, [x9]\n"
@@ -2296,19 +2294,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 68f\n"
       "ld1h { z29.h }, p5/Z, [x12]\n"
       "ld1h { z28.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z29.h, z0.h[3]\n"
       "fmla z12.h, z29.h, z1.h[3]\n"
       "fmla z16.h, z29.h, z2.h[3]\n"
       "fmla z20.h, z29.h, z3.h[3]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z29.h, z4.h[3]\n"
       "fmla z9.h, z28.h, z0.h[3]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.h, z28.h, z1.h[3]\n"
       "fmla z17.h, z28.h, z2.h[3]\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.h, z28.h, z3.h[3]\n"
       "fmla z25.h, z28.h, z4.h[3]\n"
       "ld1h { z28.h }, p5/Z, [x9]\n"
@@ -2326,19 +2324,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 68f\n"
       "ld1h { z29.h }, p5/Z, [x12]\n"
       "ld1h { z28.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z29.h, z0.h[4]\n"
       "fmla z12.h, z29.h, z1.h[4]\n"
       "fmla z16.h, z29.h, z2.h[4]\n"
       "fmla z20.h, z29.h, z3.h[4]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z29.h, z4.h[4]\n"
       "fmla z9.h, z28.h, z0.h[4]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.h, z28.h, z1.h[4]\n"
       "fmla z17.h, z28.h, z2.h[4]\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.h, z28.h, z3.h[4]\n"
       "fmla z25.h, z28.h, z4.h[4]\n"
       "ld1h { z28.h }, p5/Z, [x9]\n"
@@ -2356,19 +2354,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 68f\n"
       "ld1h { z29.h }, p5/Z, [x12]\n"
       "ld1h { z28.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z29.h, z0.h[5]\n"
       "fmla z12.h, z29.h, z1.h[5]\n"
       "fmla z16.h, z29.h, z2.h[5]\n"
       "fmla z20.h, z29.h, z3.h[5]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z29.h, z4.h[5]\n"
       "fmla z9.h, z28.h, z0.h[5]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.h, z28.h, z1.h[5]\n"
       "fmla z17.h, z28.h, z2.h[5]\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.h, z28.h, z3.h[5]\n"
       "fmla z25.h, z28.h, z4.h[5]\n"
       "ld1h { z28.h }, p5/Z, [x9]\n"
@@ -2386,19 +2384,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 68f\n"
       "ld1h { z29.h }, p5/Z, [x12]\n"
       "ld1h { z28.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z29.h, z0.h[6]\n"
       "fmla z12.h, z29.h, z1.h[6]\n"
       "fmla z16.h, z29.h, z2.h[6]\n"
       "fmla z20.h, z29.h, z3.h[6]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z29.h, z4.h[6]\n"
       "fmla z9.h, z28.h, z0.h[6]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.h, z28.h, z1.h[6]\n"
       "fmla z17.h, z28.h, z2.h[6]\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.h, z28.h, z3.h[6]\n"
       "fmla z25.h, z28.h, z4.h[6]\n"
       "ld1h { z28.h }, p5/Z, [x9]\n"
@@ -2416,12 +2414,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 68f\n"
       "ld1h { z29.h }, p5/Z, [x12]\n"
       "ld1h { z28.h }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z29.h, z0.h[7]\n"
       "fmla z12.h, z29.h, z1.h[7]\n"
       "fmla z16.h, z29.h, z2.h[7]\n"
       "fmla z20.h, z29.h, z3.h[7]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z24.h, z29.h, z4.h[7]\n"
       "fmla z9.h, z28.h, z0.h[7]\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
@@ -2448,14 +2446,14 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 63b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "tbz %x[flags], #1, 69f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z29.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z29.h }, p5/Z, [x21]\n"
       "ld1rh { z28.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z29.h\n"
       "fmin z9.h, p5/M, z9.h, z29.h\n"
@@ -2503,22 +2501,22 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p4, [x25]\n"
-      "st1h { z17.h }, p3, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p1, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p4, [x24]\n"
-      "st1h { z21.h }, p3, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p1, [x24, #3, MUL VL]\n"
-      "st1h { z24.h }, p4, [x23]\n"
-      "st1h { z25.h }, p3, [x23, #1, MUL VL]\n"
-      "st1h { z26.h }, p2, [x23, #2, MUL VL]\n"
-      "st1h { z27.h }, p1, [x23, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p4, [x24]\n"
+      "st1h { z17.h }, p3, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p1, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p4, [x23]\n"
+      "st1h { z21.h }, p3, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p1, [x23, #3, MUL VL]\n"
+      "st1h { z24.h }, p4, [x22]\n"
+      "st1h { z25.h }, p3, [x22, #1, MUL VL]\n"
+      "st1h { z26.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z27.h }, p1, [x22, #3, MUL VL]\n"
       "70:"  // Height 5: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -2526,19 +2524,18 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "b 86f\n"
       "71:"  // Height 6
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0xc\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0xc\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "72:"  // Height 6: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cnth x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #1\n"
+      "cnth x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
       "add x20, x9, x20, LSL #1\n"
@@ -2546,12 +2543,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 73f\n"
       "dech x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 73f\n"
       "dech x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 73f\n"
       "mov x11, x12\n"
       "73:"  // Height 6: B setup done
@@ -2566,18 +2563,18 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cbz x15, 74f\n"
       "ld1h { z8.h }, p5/Z, [x15]\n"
       "ld1h { z9.h }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -2593,17 +2590,17 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "74:"  // Height 6: no bias
       "tbz %x[flags], #0, 75f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x13, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x13]\n"
+      "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "ld1h { z9.h }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x13, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z11.h }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x24, x13, x20, LSL #1\n"
-      "add x23, x24, x20, LSL #1\n"
-      "add x22, x23, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x24]\n"
       "ld1h { z13.h }, p3/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z14.h }, p2/Z, [x24, #2, MUL VL]\n"
       "ld1h { z15.h }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1h { z16.h }, p4/Z, [x23]\n"
@@ -2652,8 +2649,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "77:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 78f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2684,29 +2681,29 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 81f\n"
       "80:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z1.h }, p5/Z, [x12]\n"
-      "ld1h { z0.h }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z7.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z6.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z5.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z4.h }, p0/Z, [x23]\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "ld1rqh { z3.h }, p0/Z, [x22]\n"
       "ld1rqh { z2.h }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1h { z1.h }, p5/Z, [x12]\n"
+      "ld1h { z0.h }, p5/Z, [x11]\n"
       "fmla z8.h, z1.h, z7.h[0]\n"
       "fmla z12.h, z1.h, z6.h[0]\n"
-      "add x21, x21, #0x10\n"
       "fmla z16.h, z1.h, z5.h[0]\n"
       "fmla z20.h, z1.h, z4.h[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z24.h, z1.h, z3.h[0]\n"
       "fmla z28.h, z1.h, z2.h[0]\n"
       "ld1h { z1.h }, p5/Z, [x10]\n"
+      "add x21, x21, #0x10\n"
       "fmla z9.h, z0.h, z7.h[0]\n"
       "fmla z13.h, z0.h, z6.h[0]\n"
       "fmla z17.h, z0.h, z5.h[0]\n"
@@ -2929,27 +2926,27 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "bgt 80b\n"
       "81:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z7.h }, p5/Z, [x12]\n"
-      "ld1h { z6.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
       "ld1rqh { z4.h }, p0/Z, [x22]\n"
       "ld1rqh { z5.h }, p0/Z, [x21]\n"
+      "ld1h { z7.h }, p5/Z, [x12]\n"
+      "ld1h { z6.h }, p5/Z, [x11]\n"
       "fmla z8.h, z7.h, z0.h[0]\n"
       "fmla z12.h, z7.h, z1.h[0]\n"
-      "fmla z9.h, z6.h, z0.h[0]\n"
-      "fmla z13.h, z6.h, z1.h[0]\n"
       "fmla z16.h, z7.h, z2.h[0]\n"
       "fmla z20.h, z7.h, z3.h[0]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z24.h, z7.h, z4.h[0]\n"
       "fmla z28.h, z7.h, z5.h[0]\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "addvl x10, x10, #1\n"
+      "fmla z9.h, z6.h, z0.h[0]\n"
+      "fmla z13.h, z6.h, z1.h[0]\n"
       "fmla z17.h, z6.h, z2.h[0]\n"
       "fmla z21.h, z6.h, z3.h[0]\n"
       "fmla z25.h, z6.h, z4.h[0]\n"
@@ -2971,19 +2968,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 82f\n"
       "ld1h { z7.h }, p5/Z, [x12]\n"
       "ld1h { z6.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z7.h, z0.h[1]\n"
       "fmla z12.h, z7.h, z1.h[1]\n"
       "fmla z16.h, z7.h, z2.h[1]\n"
       "fmla z20.h, z7.h, z3.h[1]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z7.h, z4.h[1]\n"
       "fmla z28.h, z7.h, z5.h[1]\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.h, z6.h, z0.h[1]\n"
       "fmla z13.h, z6.h, z1.h[1]\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.h, z6.h, z2.h[1]\n"
       "fmla z21.h, z6.h, z3.h[1]\n"
       "fmla z25.h, z6.h, z4.h[1]\n"
@@ -3005,19 +3002,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 82f\n"
       "ld1h { z7.h }, p5/Z, [x12]\n"
       "ld1h { z6.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z7.h, z0.h[2]\n"
       "fmla z12.h, z7.h, z1.h[2]\n"
       "fmla z16.h, z7.h, z2.h[2]\n"
       "fmla z20.h, z7.h, z3.h[2]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z7.h, z4.h[2]\n"
       "fmla z28.h, z7.h, z5.h[2]\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.h, z6.h, z0.h[2]\n"
       "fmla z13.h, z6.h, z1.h[2]\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.h, z6.h, z2.h[2]\n"
       "fmla z21.h, z6.h, z3.h[2]\n"
       "fmla z25.h, z6.h, z4.h[2]\n"
@@ -3039,19 +3036,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 82f\n"
       "ld1h { z7.h }, p5/Z, [x12]\n"
       "ld1h { z6.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z7.h, z0.h[3]\n"
       "fmla z12.h, z7.h, z1.h[3]\n"
       "fmla z16.h, z7.h, z2.h[3]\n"
       "fmla z20.h, z7.h, z3.h[3]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z7.h, z4.h[3]\n"
       "fmla z28.h, z7.h, z5.h[3]\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.h, z6.h, z0.h[3]\n"
       "fmla z13.h, z6.h, z1.h[3]\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.h, z6.h, z2.h[3]\n"
       "fmla z21.h, z6.h, z3.h[3]\n"
       "fmla z25.h, z6.h, z4.h[3]\n"
@@ -3073,19 +3070,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 82f\n"
       "ld1h { z7.h }, p5/Z, [x12]\n"
       "ld1h { z6.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z7.h, z0.h[4]\n"
       "fmla z12.h, z7.h, z1.h[4]\n"
       "fmla z16.h, z7.h, z2.h[4]\n"
       "fmla z20.h, z7.h, z3.h[4]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z7.h, z4.h[4]\n"
       "fmla z28.h, z7.h, z5.h[4]\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.h, z6.h, z0.h[4]\n"
       "fmla z13.h, z6.h, z1.h[4]\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.h, z6.h, z2.h[4]\n"
       "fmla z21.h, z6.h, z3.h[4]\n"
       "fmla z25.h, z6.h, z4.h[4]\n"
@@ -3107,19 +3104,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 82f\n"
       "ld1h { z7.h }, p5/Z, [x12]\n"
       "ld1h { z6.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z7.h, z0.h[5]\n"
       "fmla z12.h, z7.h, z1.h[5]\n"
       "fmla z16.h, z7.h, z2.h[5]\n"
       "fmla z20.h, z7.h, z3.h[5]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z7.h, z4.h[5]\n"
       "fmla z28.h, z7.h, z5.h[5]\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.h, z6.h, z0.h[5]\n"
       "fmla z13.h, z6.h, z1.h[5]\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.h, z6.h, z2.h[5]\n"
       "fmla z21.h, z6.h, z3.h[5]\n"
       "fmla z25.h, z6.h, z4.h[5]\n"
@@ -3141,19 +3138,19 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 82f\n"
       "ld1h { z7.h }, p5/Z, [x12]\n"
       "ld1h { z6.h }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z7.h, z0.h[6]\n"
       "fmla z12.h, z7.h, z1.h[6]\n"
       "fmla z16.h, z7.h, z2.h[6]\n"
       "fmla z20.h, z7.h, z3.h[6]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.h, z7.h, z4.h[6]\n"
       "fmla z28.h, z7.h, z5.h[6]\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.h, z6.h, z0.h[6]\n"
       "fmla z13.h, z6.h, z1.h[6]\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.h, z6.h, z2.h[6]\n"
       "fmla z21.h, z6.h, z3.h[6]\n"
       "fmla z25.h, z6.h, z4.h[6]\n"
@@ -3175,12 +3172,12 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "ble 82f\n"
       "ld1h { z7.h }, p5/Z, [x12]\n"
       "ld1h { z6.h }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.h, z7.h, z0.h[7]\n"
       "fmla z12.h, z7.h, z1.h[7]\n"
       "fmla z16.h, z7.h, z2.h[7]\n"
       "fmla z20.h, z7.h, z3.h[7]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z24.h, z7.h, z4.h[7]\n"
       "fmla z28.h, z7.h, z5.h[7]\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
@@ -3211,15 +3208,15 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 77b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x13, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "tbz %x[flags], #1, 83f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z1.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z1.h }, p5/Z, [x21]\n"
       "ld1rh { z0.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z1.h\n"
       "fmin z9.h, p5/M, z9.h, z1.h\n"
@@ -3275,26 +3272,26 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x13, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p4, [x25]\n"
-      "st1h { z17.h }, p3, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p1, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p4, [x24]\n"
-      "st1h { z21.h }, p3, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p1, [x24, #3, MUL VL]\n"
-      "st1h { z24.h }, p4, [x23]\n"
-      "st1h { z25.h }, p3, [x23, #1, MUL VL]\n"
-      "st1h { z26.h }, p2, [x23, #2, MUL VL]\n"
-      "st1h { z27.h }, p1, [x23, #3, MUL VL]\n"
-      "st1h { z28.h }, p4, [x22]\n"
-      "st1h { z29.h }, p3, [x22, #1, MUL VL]\n"
-      "st1h { z30.h }, p2, [x22, #2, MUL VL]\n"
-      "st1h { z31.h }, p1, [x22, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p4, [x24]\n"
+      "st1h { z17.h }, p3, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p1, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p4, [x23]\n"
+      "st1h { z21.h }, p3, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p1, [x23, #3, MUL VL]\n"
+      "st1h { z24.h }, p4, [x22]\n"
+      "st1h { z25.h }, p3, [x22, #1, MUL VL]\n"
+      "st1h { z26.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z27.h }, p1, [x22, #3, MUL VL]\n"
+      "st1h { z28.h }, p4, [x21]\n"
+      "st1h { z29.h }, p3, [x21, #1, MUL VL]\n"
+      "st1h { z30.h }, p2, [x21, #2, MUL VL]\n"
+      "st1h { z31.h }, p1, [x21, #3, MUL VL]\n"
       "84:"  // Height 6: Writeback done
       "dech x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -3311,8 +3308,8 @@ void sve_ffhybrid_fp16_mla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "86:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL.hpp
index 847103646c..ffa365b8a0 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL.hpp
@@ -82,7 +82,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 4, 1> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp
index 51e2b3722a..53cd52fe56 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/a64fx.cpp
@@ -49,19 +49,18 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -82,7 +81,6 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -108,15 +106,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "bgt 27f\n"
       "beq 14f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -124,12 +122,12 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 3f\n"
       "mov x11, x12\n"
       "3:"  // Height 1: B setup done
@@ -164,8 +162,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "7:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 8f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -183,41 +181,41 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "ld1w { z7.s }, p4/Z, [x11]\n"
       "ble 11f\n"
       "10:"  // Height 1: Multiply loop: Main loop
-      "addvl x12, x12, #1\n"
       "fmla z8.s, p4/M, z6.s, z0.s\n"
-      "ld1w { z17.s }, p4/Z, [x10]\n"
-      "addvl x11, x11, #1\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z17.s }, p4/Z, [x10]\n"
       "ld1w { z16.s }, p4/Z, [x9]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "fmla z10.s, p4/M, z17.s, z0.s\n"
+      "fmla z11.s, p4/M, z16.s, z0.s\n"
       "add x26, x26, #0x4\n"
       "subs x27, x27, #0x1\n"
+      "ld1rw { z0.s }, p4/Z, [x26]\n"
+      "ld1w { z6.s }, p4/Z, [x12]\n"
       "addvl x10, x10, #1\n"
       "addvl x9, x9, #1\n"
-      "fmla z10.s, p4/M, z17.s, z0.s\n"
-      "ld1w { z6.s }, p4/Z, [x12]\n"
-      "fmla z11.s, p4/M, z16.s, z0.s\n"
-      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1w { z7.s }, p4/Z, [x11]\n"
       "bgt 10b\n"
       "11:"  // Height 1: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "fmla z8.s, p4/M, z6.s, z0.s\n"
-      "ld1w { z17.s }, p4/Z, [x10]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z17.s }, p4/Z, [x10]\n"
       "ld1w { z16.s }, p4/Z, [x9]\n"
       "add x28, x28, #0x1\n"
+      "cmp x28, x20\n"
+      "fmla z10.s, p4/M, z17.s, z0.s\n"
+      "fmla z11.s, p4/M, z16.s, z0.s\n"
       "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
       "addvl x10, x10, #1\n"
-      "cmp x28, x20\n"
       "addvl x9, x9, #1\n"
-      "fmla z10.s, p4/M, z17.s, z0.s\n"
-      "fmla z11.s, p4/M, z16.s, z0.s\n"
       "bne 7b\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p4/Z, [x21]\n"
       "ld1rw { z16.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z17.s\n"
       "fmin z9.s, p4/M, z9.s, z17.s\n"
@@ -240,15 +238,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "b 80f\n"
       "14:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -256,12 +254,12 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 16f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 16f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 16f\n"
       "mov x11, x12\n"
       "16:"  // Height 2: B setup done
@@ -276,22 +274,22 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "cbz x15, 17f\n"
       "ld1w { z8.s }, p4/Z, [x15]\n"
       "ld1w { z9.s }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x15, x15, #4\n"
       "b 19f\n"
       "17:"  // Height 2: no bias
       "tbz %x[flags], #0, 18f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x13, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x13]\n"
       "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x20, x13, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x20]\n"
       "ld1w { z13.s }, p2/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x20, #2, MUL VL]\n"
@@ -310,8 +308,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "20:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 21f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -333,26 +331,26 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "ld1w { z7.s }, p4/Z, [x11]\n"
       "ble 24f\n"
       "23:"  // Height 2: Multiply loop: Main loop
-      "addvl x12, x12, #1\n"
       "fmla z8.s, p4/M, z6.s, z0.s\n"
       "fmla z12.s, p4/M, z6.s, z1.s\n"
       "ld1w { z17.s }, p4/Z, [x10]\n"
-      "addvl x11, x11, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "ld1w { z16.s }, p4/Z, [x9]\n"
+      "addvl x11, x11, #1\n"
       "add x26, x26, #0x4\n"
       "subs x27, x27, #0x1\n"
-      "add x25, x25, #0x4\n"
-      "addvl x10, x10, #1\n"
       "fmla z10.s, p4/M, z17.s, z0.s\n"
       "fmla z14.s, p4/M, z17.s, z1.s\n"
-      "addvl x9, x9, #1\n"
-      "ld1w { z6.s }, p4/Z, [x12]\n"
+      "add x25, x25, #0x4\n"
       "fmla z11.s, p4/M, z16.s, z0.s\n"
-      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "fmla z15.s, p4/M, z16.s, z1.s\n"
+      "addvl x10, x10, #1\n"
+      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1rw { z1.s }, p4/Z, [x25]\n"
+      "addvl x9, x9, #1\n"
+      "ld1w { z6.s }, p4/Z, [x12]\n"
       "ld1w { z7.s }, p4/Z, [x11]\n"
       "bgt 23b\n"
       "24:"  // Height 2: Multiply loop: Main loop skip
@@ -364,22 +362,22 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "ld1w { z16.s }, p4/Z, [x9]\n"
       "add x28, x28, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "cmp x28, x20\n"
-      "addvl x10, x10, #1\n"
       "fmla z10.s, p4/M, z17.s, z0.s\n"
       "fmla z14.s, p4/M, z17.s, z1.s\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z11.s, p4/M, z16.s, z0.s\n"
       "fmla z15.s, p4/M, z16.s, z1.s\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "bne 20b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p4/Z, [x21]\n"
       "ld1rw { z16.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z17.s\n"
       "fmin z9.s, p4/M, z9.s, z17.s\n"
@@ -403,10 +401,10 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
       "26:"  // Height 2: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -414,15 +412,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "b 80f\n"
       "27:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -430,12 +428,12 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 29f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 29f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 29f\n"
       "mov x11, x12\n"
       "29:"  // Height 3: B setup done
@@ -450,27 +448,27 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "cbz x15, 30f\n"
       "ld1w { z8.s }, p4/Z, [x15]\n"
       "ld1w { z9.s }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 32f\n"
       "30:"  // Height 3: no bias
       "tbz %x[flags], #0, 31f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x13, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x13]\n"
       "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x21, x13, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x21]\n"
       "ld1w { z13.s }, p2/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x21, #2, MUL VL]\n"
@@ -497,8 +495,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "33:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 34f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -524,13 +522,13 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "ld1w { z7.s }, p4/Z, [x11]\n"
       "ble 37f\n"
       "36:"  // Height 3: Multiply loop: Main loop
-      "addvl x12, x12, #1\n"
       "fmla z8.s, p4/M, z6.s, z0.s\n"
       "fmla z12.s, p4/M, z6.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
-      "ld1w { z21.s }, p4/Z, [x10]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z21.s }, p4/Z, [x10]\n"
       "add x26, x26, #0x4\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
@@ -538,18 +536,18 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "subs x27, x27, #0x1\n"
       "add x25, x25, #0x4\n"
       "add x24, x24, #0x4\n"
-      "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.s, p4/M, z21.s, z0.s\n"
       "fmla z14.s, p4/M, z21.s, z1.s\n"
       "fmla z18.s, p4/M, z21.s, z2.s\n"
-      "ld1w { z6.s }, p4/Z, [x12]\n"
       "fmla z11.s, p4/M, z20.s, z0.s\n"
-      "ld1rw { z0.s }, p4/Z, [x26]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.s, p4/M, z20.s, z1.s\n"
       "fmla z19.s, p4/M, z20.s, z2.s\n"
+      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
+      "ld1w { z6.s }, p4/Z, [x12]\n"
       "ld1w { z7.s }, p4/Z, [x11]\n"
       "bgt 36b\n"
       "37:"  // Height 3: Multiply loop: Main loop skip
@@ -558,30 +556,30 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z12.s, p4/M, z6.s, z1.s\n"
       "add x28, x28, #0x1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
-      "ld1w { z21.s }, p4/Z, [x10]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
-      "addvl x12, x12, #1\n"
+      "ld1w { z21.s }, p4/Z, [x10]\n"
+      "cmp x28, x20\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
       "ld1w { z20.s }, p4/Z, [x9]\n"
-      "addvl x11, x11, #1\n"
-      "cmp x28, x20\n"
-      "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.s, p4/M, z21.s, z0.s\n"
       "fmla z14.s, p4/M, z21.s, z1.s\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.s, p4/M, z21.s, z2.s\n"
       "fmla z11.s, p4/M, z20.s, z0.s\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.s, p4/M, z20.s, z1.s\n"
       "fmla z19.s, p4/M, z20.s, z2.s\n"
       "bne 33b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z21.s }, p4/Z, [x21]\n"
       "ld1rw { z20.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z21.s\n"
       "fmin z9.s, p4/M, z9.s, z21.s\n"
@@ -613,14 +611,14 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p3, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p3, [x24]\n"
+      "st1w { z17.s }, p2, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p0, [x24, #3, MUL VL]\n"
       "39:"  // Height 3: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -628,15 +626,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "b 80f\n"
       "40:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -644,12 +642,12 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 42f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 42f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 42f\n"
       "mov x11, x12\n"
       "42:"  // Height 4: B setup done
@@ -664,18 +662,18 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "cbz x15, 43f\n"
       "ld1w { z8.s }, p4/Z, [x15]\n"
       "ld1w { z9.s }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -683,13 +681,13 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "43:"  // Height 4: no bias
       "tbz %x[flags], #0, 44f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x13, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x13]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x22, x13, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x22]\n"
       "ld1w { z13.s }, p2/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x22, #2, MUL VL]\n"
@@ -724,8 +722,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "46:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 47f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -755,9 +753,9 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "ld1w { z7.s }, p4/Z, [x11]\n"
       "ble 50f\n"
       "49:"  // Height 4: Multiply loop: Main loop
-      "addvl x12, x12, #1\n"
       "fmla z8.s, p4/M, z6.s, z0.s\n"
       "fmla z12.s, p4/M, z6.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
       "fmla z20.s, p4/M, z6.s, z3.s\n"
@@ -775,9 +773,9 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z10.s, p4/M, z25.s, z0.s\n"
       "fmla z14.s, p4/M, z25.s, z1.s\n"
       "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
       "fmla z18.s, p4/M, z25.s, z2.s\n"
       "fmla z22.s, p4/M, z25.s, z3.s\n"
+      "addvl x9, x9, #1\n"
       "ld1w { z6.s }, p4/Z, [x12]\n"
       "fmla z11.s, p4/M, z24.s, z0.s\n"
       "fmla z15.s, p4/M, z24.s, z1.s\n"
@@ -797,18 +795,18 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z16.s, p4/M, z6.s, z2.s\n"
       "fmla z20.s, p4/M, z6.s, z3.s\n"
       "ld1w { z25.s }, p4/Z, [x10]\n"
-      "addvl x12, x12, #1\n"
+      "cmp x28, x20\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
-      "addvl x10, x10, #1\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
       "fmla z21.s, p4/M, z7.s, z3.s\n"
       "ld1w { z24.s }, p4/Z, [x9]\n"
-      "cmp x28, x20\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.s, p4/M, z25.s, z0.s\n"
       "fmla z14.s, p4/M, z25.s, z1.s\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.s, p4/M, z25.s, z2.s\n"
       "fmla z22.s, p4/M, z25.s, z3.s\n"
       "fmla z11.s, p4/M, z24.s, z0.s\n"
@@ -817,13 +815,13 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z23.s, p4/M, z24.s, z3.s\n"
       "bne 46b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z25.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z25.s }, p4/Z, [x21]\n"
       "ld1rw { z24.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z25.s\n"
       "fmin z9.s, p4/M, z9.s, z25.s\n"
@@ -863,18 +861,18 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p3, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p0, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p3, [x24]\n"
-      "st1w { z21.s }, p2, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p1, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p0, [x24, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p3, [x24]\n"
+      "st1w { z17.s }, p2, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p0, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p3, [x23]\n"
+      "st1w { z21.s }, p2, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p1, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p0, [x23, #3, MUL VL]\n"
       "52:"  // Height 4: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -882,15 +880,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "b 80f\n"
       "53:"  // Height 5
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "54:"  // Height 5: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -898,12 +896,12 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 55f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 55f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 55f\n"
       "mov x11, x12\n"
       "55:"  // Height 5: B setup done
@@ -918,18 +916,18 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "cbz x15, 56f\n"
       "ld1w { z8.s }, p4/Z, [x15]\n"
       "ld1w { z9.s }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -941,16 +939,16 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "56:"  // Height 5: no bias
       "tbz %x[flags], #0, 57f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x13]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x23, x13, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x23]\n"
       "ld1w { z13.s }, p2/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p1/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p0/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p3/Z, [x22]\n"
@@ -991,8 +989,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "59:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 60f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1035,8 +1033,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "add x26, x26, #0x4\n"
       "subs x27, x27, #0x1\n"
       "fmla z24.s, p4/M, z6.s, z4.s\n"
-      "ld1w { z29.s }, p4/Z, [x10]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z29.s }, p4/Z, [x10]\n"
       "add x25, x25, #0x4\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
@@ -1046,23 +1044,23 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z25.s, p4/M, z7.s, z4.s\n"
       "ld1w { z28.s }, p4/Z, [x9]\n"
       "add x22, x22, #0x4\n"
-      "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.s, p4/M, z29.s, z0.s\n"
       "fmla z14.s, p4/M, z29.s, z1.s\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.s, p4/M, z29.s, z2.s\n"
       "fmla z22.s, p4/M, z29.s, z3.s\n"
       "fmla z26.s, p4/M, z29.s, z4.s\n"
-      "ld1w { z6.s }, p4/Z, [x12]\n"
       "fmla z11.s, p4/M, z28.s, z0.s\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
+      "ld1w { z6.s }, p4/Z, [x12]\n"
       "fmla z15.s, p4/M, z28.s, z1.s\n"
-      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "fmla z19.s, p4/M, z28.s, z2.s\n"
+      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
       "fmla z23.s, p4/M, z28.s, z3.s\n"
-      "ld1rw { z3.s }, p4/Z, [x23]\n"
       "fmla z27.s, p4/M, z28.s, z4.s\n"
+      "ld1rw { z3.s }, p4/Z, [x23]\n"
       "ld1rw { z4.s }, p4/Z, [x22]\n"
       "ld1w { z7.s }, p4/Z, [x11]\n"
       "bgt 62b\n"
@@ -1073,15 +1071,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
       "fmla z20.s, p4/M, z6.s, z3.s\n"
+      "cmp x28, x20\n"
       "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z24.s, p4/M, z6.s, z4.s\n"
-      "ld1w { z29.s }, p4/Z, [x10]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
-      "addvl x10, x10, #1\n"
+      "ld1w { z29.s }, p4/Z, [x10]\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
-      "cmp x28, x20\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.s, p4/M, z7.s, z3.s\n"
       "fmla z25.s, p4/M, z7.s, z4.s\n"
       "ld1w { z28.s }, p4/Z, [x9]\n"
@@ -1098,14 +1096,14 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z27.s, p4/M, z28.s, z4.s\n"
       "bne 59b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z29.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z29.s }, p4/Z, [x21]\n"
       "ld1rw { z28.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z29.s\n"
       "fmin z9.s, p4/M, z9.s, z29.s\n"
@@ -1153,22 +1151,22 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p3, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p0, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p3, [x24]\n"
-      "st1w { z21.s }, p2, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p1, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p0, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p3, [x23]\n"
-      "st1w { z25.s }, p2, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p1, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p0, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p3, [x24]\n"
+      "st1w { z17.s }, p2, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p0, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p3, [x23]\n"
+      "st1w { z21.s }, p2, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p1, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p0, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p3, [x22]\n"
+      "st1w { z25.s }, p2, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p0, [x22, #3, MUL VL]\n"
       "65:"  // Height 5: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1176,19 +1174,18 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "b 80f\n"
       "66:"  // Height 6
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0x18\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0x18\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "67:"  // Height 6: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -1196,12 +1193,12 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 68f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 68f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 68f\n"
       "mov x11, x12\n"
       "68:"  // Height 6: B setup done
@@ -1216,18 +1213,18 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "cbz x15, 69f\n"
       "ld1w { z8.s }, p4/Z, [x15]\n"
       "ld1w { z9.s }, p4/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p4/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1243,17 +1240,17 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "69:"  // Height 6: no bias
       "tbz %x[flags], #0, 70f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x13, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x13]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x13, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p0/Z, [x13, #3, MUL VL]\n"
-      "add x24, x13, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x24]\n"
       "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p3/Z, [x23]\n"
@@ -1302,8 +1299,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "72:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 73f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1365,9 +1362,9 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z29.s, p4/M, z7.s, z5.s\n"
       "ld1w { z7.s }, p4/Z, [x9]\n"
       "addvl x10, x10, #1\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.s, p4/M, z6.s, z0.s\n"
       "fmla z14.s, p4/M, z6.s, z1.s\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.s, p4/M, z6.s, z2.s\n"
       "fmla z22.s, p4/M, z6.s, z3.s\n"
       "fmla z26.s, p4/M, z6.s, z4.s\n"
@@ -1394,15 +1391,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
       "fmla z20.s, p4/M, z6.s, z3.s\n"
+      "cmp x28, x20\n"
       "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z24.s, p4/M, z6.s, z4.s\n"
       "fmla z28.s, p4/M, z6.s, z5.s\n"
       "ld1w { z6.s }, p4/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
-      "cmp x28, x20\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
       "fmla z21.s, p4/M, z7.s, z3.s\n"
       "fmla z25.s, p4/M, z7.s, z4.s\n"
@@ -1423,15 +1420,15 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "fmla z31.s, p4/M, z7.s, z5.s\n"
       "bne 72b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "tbz %x[flags], #1, 77f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p4/Z, [x21]\n"
       "ld1rw { z0.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z1.s\n"
       "fmin z9.s, p4/M, z9.s, z1.s\n"
@@ -1487,26 +1484,26 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p3, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p0, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p3, [x24]\n"
-      "st1w { z21.s }, p2, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p1, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p0, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p3, [x23]\n"
-      "st1w { z25.s }, p2, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p1, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p0, [x23, #3, MUL VL]\n"
-      "st1w { z28.s }, p3, [x22]\n"
-      "st1w { z29.s }, p2, [x22, #1, MUL VL]\n"
-      "st1w { z30.s }, p1, [x22, #2, MUL VL]\n"
-      "st1w { z31.s }, p0, [x22, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p3, [x24]\n"
+      "st1w { z17.s }, p2, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p0, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p3, [x23]\n"
+      "st1w { z21.s }, p2, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p1, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p0, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p3, [x22]\n"
+      "st1w { z25.s }, p2, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p0, [x22, #3, MUL VL]\n"
+      "st1w { z28.s }, p3, [x21]\n"
+      "st1w { z29.s }, p2, [x21, #1, MUL VL]\n"
+      "st1w { z30.s }, p1, [x21, #2, MUL VL]\n"
+      "st1w { z31.s }, p0, [x21, #3, MUL VL]\n"
       "78:"  // Height 6: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1523,8 +1520,8 @@ void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp
index 299dec5b3c..8f12e9ee62 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp
@@ -49,19 +49,18 @@ void sve_ffhybrid_fp32_mla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -82,7 +81,6 @@ void sve_ffhybrid_fp32_mla_6x4VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -108,15 +106,15 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "bgt 29f\n"
       "beq 15f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -124,12 +122,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 3f\n"
       "mov x11, x12\n"
       "3:"  // Height 1: B setup done
@@ -164,8 +162,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "7:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 8f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -181,113 +179,113 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 11f\n"
       "10:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z17.s }, p5/Z, [x12]\n"
-      "ld1w { z16.s }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "fmla z8.s, z17.s, z0.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
+      "ld1w { z16.s }, p5/Z, [x12]\n"
+      "fmla z8.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p5/Z, [x11]\n"
       "fmla z9.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p5/Z, [x10]\n"
+      "fmla z10.s, z16.s, z0.s[0]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
-      "fmla z10.s, z17.s, z0.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x12, #1, MUL VL]\n"
       "fmla z11.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p5/Z, [x12, #1, MUL VL]\n"
+      "fmla z8.s, z16.s, z0.s[1]\n"
       "ld1w { z16.s }, p5/Z, [x11, #1, MUL VL]\n"
-      "fmla z8.s, z17.s, z0.s[1]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[1]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
+      "fmla z10.s, z16.s, z0.s[1]\n"
       "ld1w { z16.s }, p5/Z, [x9, #1, MUL VL]\n"
-      "fmla z10.s, z17.s, z0.s[1]\n"
-      "ld1w { z17.s }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z11.s, z16.s, z0.s[1]\n"
+      "ld1w { z16.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "fmla z8.s, z16.s, z0.s[2]\n"
       "ld1w { z16.s }, p5/Z, [x11, #2, MUL VL]\n"
-      "fmla z8.s, z17.s, z0.s[2]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[2]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z10.s, z16.s, z0.s[2]\n"
       "ld1w { z16.s }, p5/Z, [x9, #2, MUL VL]\n"
-      "fmla z10.s, z17.s, z0.s[2]\n"
-      "ld1w { z17.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "fmla z11.s, z16.s, z0.s[2]\n"
+      "ld1w { z16.s }, p5/Z, [x12, #3, MUL VL]\n"
+      "fmla z8.s, z16.s, z0.s[3]\n"
       "ld1w { z16.s }, p5/Z, [x11, #3, MUL VL]\n"
-      "addvl x11, x11, #4\n"
-      "fmla z8.s, z17.s, z0.s[3]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z9.s, z16.s, z0.s[3]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #3, MUL VL]\n"
+      "sub x27, x27, #0x4\n"
       "ld1w { z16.s }, p5/Z, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
+      "cmp x27, #0x4\n"
       "fmla z10.s, z17.s, z0.s[3]\n"
       "fmla z11.s, z16.s, z0.s[3]\n"
+      "add x26, x26, #0x10\n"
+      "addvl x12, x12, #4\n"
+      "addvl x11, x11, #4\n"
+      "addvl x10, x10, #4\n"
+      "addvl x9, x9, #4\n"
       "bgt 10b\n"
       "11:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z17.s }, p5/Z, [x12]\n"
+      "ld1rqw { z0.s }, p0/Z, [x26]\n"
+      "ld1w { z16.s }, p5/Z, [x12]\n"
+      "fmla z8.s, z16.s, z0.s[0]\n"
       "ld1w { z16.s }, p5/Z, [x11]\n"
+      "fmla z9.s, z16.s, z0.s[0]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "subs x27, x27, #0x1\n"
+      "ld1w { z16.s }, p5/Z, [x9]\n"
+      "fmla z10.s, z17.s, z0.s[0]\n"
+      "fmla z11.s, z16.s, z0.s[0]\n"
       "addvl x12, x12, #1\n"
       "addvl x11, x11, #1\n"
-      "ld1rqw { z0.s }, p0/Z, [x26]\n"
-      "fmla z8.s, z17.s, z0.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
       "addvl x10, x10, #1\n"
-      "fmla z9.s, z16.s, z0.s[0]\n"
-      "ld1w { z16.s }, p5/Z, [x9]\n"
       "addvl x9, x9, #1\n"
-      "fmla z10.s, z17.s, z0.s[0]\n"
-      "fmla z11.s, z16.s, z0.s[0]\n"
       "ble 12f\n"
       "ld1w { z17.s }, p5/Z, [x12]\n"
       "ld1w { z16.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z17.s, z0.s[1]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z16.s, z0.s[1]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.s, z17.s, z0.s[1]\n"
       "fmla z11.s, z16.s, z0.s[1]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 12f\n"
       "ld1w { z17.s }, p5/Z, [x12]\n"
       "ld1w { z16.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z17.s, z0.s[2]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z16.s, z0.s[2]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.s, z17.s, z0.s[2]\n"
       "fmla z11.s, z16.s, z0.s[2]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 12f\n"
       "ld1w { z17.s }, p5/Z, [x12]\n"
       "ld1w { z16.s }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z17.s, z0.s[3]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z16.s, z0.s[3]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.s, z17.s, z0.s[3]\n"
       "fmla z11.s, z16.s, z0.s[3]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "12:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 7b\n"
       "tbz %x[flags], #1, 13f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z17.s\n"
       "fmin z9.s, p5/M, z9.s, z17.s\n"
@@ -310,15 +308,15 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "b 86f\n"
       "15:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "16:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -326,12 +324,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 17f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 17f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 17f\n"
       "mov x11, x12\n"
       "17:"  // Height 2: B setup done
@@ -346,22 +344,22 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cbz x15, 18f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x15, x15, #4\n"
       "b 20f\n"
       "18:"  // Height 2: no bias
       "tbz %x[flags], #0, 19f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x13, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x13]\n"
       "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x20, x13, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
@@ -380,8 +378,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "21:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 22f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -400,29 +398,29 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 25f\n"
       "24:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z17.s }, p5/Z, [x12]\n"
-      "ld1w { z16.s }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z1.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
+      "ld1w { z17.s }, p5/Z, [x12]\n"
+      "ld1w { z16.s }, p5/Z, [x11]\n"
       "fmla z8.s, z17.s, z1.s[0]\n"
       "fmla z12.s, z17.s, z0.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
       "fmla z9.s, z16.s, z1.s[0]\n"
       "fmla z13.s, z16.s, z0.s[0]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
       "fmla z10.s, z17.s, z1.s[0]\n"
       "fmla z14.s, z17.s, z0.s[0]\n"
       "ld1w { z17.s }, p5/Z, [x12, #1, MUL VL]\n"
+      "cmp x27, #0x4\n"
       "fmla z11.s, z16.s, z1.s[0]\n"
       "fmla z15.s, z16.s, z0.s[0]\n"
       "ld1w { z16.s }, p5/Z, [x11, #1, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       "fmla z8.s, z17.s, z1.s[1]\n"
       "fmla z12.s, z17.s, z0.s[1]\n"
       "ld1w { z17.s }, p5/Z, [x10, #1, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       "fmla z9.s, z16.s, z1.s[1]\n"
       "fmla z13.s, z16.s, z0.s[1]\n"
       "ld1w { z16.s }, p5/Z, [x9, #1, MUL VL]\n"
@@ -461,89 +459,89 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "bgt 24b\n"
       "25:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z17.s }, p5/Z, [x12]\n"
-      "ld1w { z16.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
+      "ld1w { z17.s }, p5/Z, [x12]\n"
+      "ld1w { z16.s }, p5/Z, [x11]\n"
       "fmla z8.s, z17.s, z0.s[0]\n"
       "fmla z12.s, z17.s, z1.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z16.s, z0.s[0]\n"
       "fmla z13.s, z16.s, z1.s[0]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.s, z17.s, z0.s[0]\n"
       "fmla z14.s, z17.s, z1.s[0]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z11.s, z16.s, z0.s[0]\n"
       "fmla z15.s, z16.s, z1.s[0]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1w { z17.s }, p5/Z, [x12]\n"
       "ld1w { z16.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z17.s, z0.s[1]\n"
       "fmla z12.s, z17.s, z1.s[1]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z16.s, z0.s[1]\n"
       "fmla z13.s, z16.s, z1.s[1]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.s, z17.s, z0.s[1]\n"
       "fmla z14.s, z17.s, z1.s[1]\n"
+      "addvl x12, x12, #1\n"
       "fmla z11.s, z16.s, z0.s[1]\n"
       "fmla z15.s, z16.s, z1.s[1]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1w { z17.s }, p5/Z, [x12]\n"
       "ld1w { z16.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z17.s, z0.s[2]\n"
       "fmla z12.s, z17.s, z1.s[2]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z16.s, z0.s[2]\n"
       "fmla z13.s, z16.s, z1.s[2]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.s, z17.s, z0.s[2]\n"
       "fmla z14.s, z17.s, z1.s[2]\n"
+      "addvl x12, x12, #1\n"
       "fmla z11.s, z16.s, z0.s[2]\n"
       "fmla z15.s, z16.s, z1.s[2]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "ble 26f\n"
       "ld1w { z17.s }, p5/Z, [x12]\n"
       "ld1w { z16.s }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z17.s, z0.s[3]\n"
       "fmla z12.s, z17.s, z1.s[3]\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z16.s, z0.s[3]\n"
       "fmla z13.s, z16.s, z1.s[3]\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
       "fmla z10.s, z17.s, z0.s[3]\n"
       "fmla z14.s, z17.s, z1.s[3]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z11.s, z16.s, z0.s[3]\n"
       "fmla z15.s, z16.s, z1.s[3]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "26:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 21b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "tbz %x[flags], #1, 27f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z17.s\n"
       "fmin z9.s, p5/M, z9.s, z17.s\n"
@@ -567,10 +565,10 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
       "28:"  // Height 2: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -578,15 +576,15 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "b 86f\n"
       "29:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "30:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -594,12 +592,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 31f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 31f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 31f\n"
       "mov x11, x12\n"
       "31:"  // Height 3: B setup done
@@ -614,27 +612,27 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cbz x15, 32f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 34f\n"
       "32:"  // Height 3: no bias
       "tbz %x[flags], #0, 33f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x13, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x13]\n"
       "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x21, x13, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
@@ -661,8 +659,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "35:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 36f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -684,62 +682,62 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 39f\n"
       "38:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z21.s }, p5/Z, [x12]\n"
-      "ld1w { z20.s }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "ld1rqw { z0.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1w { z21.s }, p5/Z, [x12]\n"
       "fmla z8.s, z21.s, z2.s[0]\n"
       "fmla z12.s, z21.s, z1.s[0]\n"
-      "fmla z9.s, z20.s, z2.s[0]\n"
-      "fmla z13.s, z20.s, z1.s[0]\n"
+      "ld1w { z20.s }, p5/Z, [x11]\n"
       "fmla z16.s, z21.s, z0.s[0]\n"
+      "fmla z9.s, z20.s, z2.s[0]\n"
       "ld1w { z21.s }, p5/Z, [x10]\n"
+      "fmla z13.s, z20.s, z1.s[0]\n"
       "fmla z17.s, z20.s, z0.s[0]\n"
       "ld1w { z20.s }, p5/Z, [x9]\n"
+      "cmp x27, #0x4\n"
       "fmla z10.s, z21.s, z2.s[0]\n"
       "fmla z14.s, z21.s, z1.s[0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "fmla z18.s, z21.s, z0.s[0]\n"
-      "ld1w { z21.s }, p5/Z, [x12, #1, MUL VL]\n"
       "fmla z11.s, z20.s, z2.s[0]\n"
+      "ld1w { z21.s }, p5/Z, [x12, #1, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       "fmla z15.s, z20.s, z1.s[0]\n"
       "fmla z19.s, z20.s, z0.s[0]\n"
       "ld1w { z20.s }, p5/Z, [x11, #1, MUL VL]\n"
       "fmla z8.s, z21.s, z2.s[1]\n"
       "fmla z12.s, z21.s, z1.s[1]\n"
       "fmla z16.s, z21.s, z0.s[1]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.s, z20.s, z2.s[1]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z13.s, z20.s, z1.s[1]\n"
       "fmla z17.s, z20.s, z0.s[1]\n"
       "ld1w { z20.s }, p5/Z, [x9, #1, MUL VL]\n"
       "fmla z10.s, z21.s, z2.s[1]\n"
       "fmla z14.s, z21.s, z1.s[1]\n"
       "fmla z18.s, z21.s, z0.s[1]\n"
-      "ld1w { z21.s }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z11.s, z20.s, z2.s[1]\n"
+      "ld1w { z21.s }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z15.s, z20.s, z1.s[1]\n"
       "fmla z19.s, z20.s, z0.s[1]\n"
       "ld1w { z20.s }, p5/Z, [x11, #2, MUL VL]\n"
       "fmla z8.s, z21.s, z2.s[2]\n"
       "fmla z12.s, z21.s, z1.s[2]\n"
       "fmla z16.s, z21.s, z0.s[2]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z20.s, z2.s[2]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.s, z20.s, z1.s[2]\n"
       "fmla z17.s, z20.s, z0.s[2]\n"
       "ld1w { z20.s }, p5/Z, [x9, #2, MUL VL]\n"
       "fmla z10.s, z21.s, z2.s[2]\n"
       "fmla z14.s, z21.s, z1.s[2]\n"
       "fmla z18.s, z21.s, z0.s[2]\n"
+      "fmla z11.s, z20.s, z2.s[2]\n"
       "ld1w { z21.s }, p5/Z, [x12, #3, MUL VL]\n"
       "addvl x12, x12, #4\n"
-      "fmla z11.s, z20.s, z2.s[2]\n"
       "fmla z15.s, z20.s, z1.s[2]\n"
       "fmla z19.s, z20.s, z0.s[2]\n"
       "ld1w { z20.s }, p5/Z, [x11, #3, MUL VL]\n"
@@ -747,9 +745,9 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "fmla z8.s, z21.s, z2.s[3]\n"
       "fmla z12.s, z21.s, z1.s[3]\n"
       "fmla z16.s, z21.s, z0.s[3]\n"
+      "fmla z9.s, z20.s, z2.s[3]\n"
       "ld1w { z21.s }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
-      "fmla z9.s, z20.s, z2.s[3]\n"
       "fmla z13.s, z20.s, z1.s[3]\n"
       "fmla z17.s, z20.s, z0.s[3]\n"
       "ld1w { z20.s }, p5/Z, [x9, #3, MUL VL]\n"
@@ -763,91 +761,91 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "bgt 38b\n"
       "39:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z21.s }, p5/Z, [x12]\n"
-      "ld1w { z20.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
+      "ld1w { z21.s }, p5/Z, [x12]\n"
       "fmla z8.s, z21.s, z0.s[0]\n"
       "fmla z12.s, z21.s, z1.s[0]\n"
-      "fmla z9.s, z20.s, z0.s[0]\n"
-      "fmla z13.s, z20.s, z1.s[0]\n"
+      "ld1w { z20.s }, p5/Z, [x11]\n"
       "fmla z16.s, z21.s, z2.s[0]\n"
+      "fmla z9.s, z20.s, z0.s[0]\n"
       "ld1w { z21.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "fmla z13.s, z20.s, z1.s[0]\n"
       "fmla z17.s, z20.s, z2.s[0]\n"
       "ld1w { z20.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.s, z21.s, z0.s[0]\n"
       "fmla z14.s, z21.s, z1.s[0]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.s, z21.s, z2.s[0]\n"
       "fmla z11.s, z20.s, z0.s[0]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.s, z20.s, z1.s[0]\n"
       "fmla z19.s, z20.s, z2.s[0]\n"
       "ble 40f\n"
       "ld1w { z21.s }, p5/Z, [x12]\n"
       "ld1w { z20.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z21.s, z0.s[1]\n"
       "fmla z12.s, z21.s, z1.s[1]\n"
       "fmla z16.s, z21.s, z2.s[1]\n"
-      "ld1w { z21.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z20.s, z0.s[1]\n"
+      "ld1w { z21.s }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.s, z20.s, z1.s[1]\n"
       "fmla z17.s, z20.s, z2.s[1]\n"
       "ld1w { z20.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.s, z21.s, z0.s[1]\n"
       "fmla z14.s, z21.s, z1.s[1]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.s, z21.s, z2.s[1]\n"
       "fmla z11.s, z20.s, z0.s[1]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.s, z20.s, z1.s[1]\n"
       "fmla z19.s, z20.s, z2.s[1]\n"
       "ble 40f\n"
       "ld1w { z21.s }, p5/Z, [x12]\n"
       "ld1w { z20.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z21.s, z0.s[2]\n"
       "fmla z12.s, z21.s, z1.s[2]\n"
       "fmla z16.s, z21.s, z2.s[2]\n"
-      "ld1w { z21.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z20.s, z0.s[2]\n"
+      "ld1w { z21.s }, p5/Z, [x10]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.s, z20.s, z1.s[2]\n"
       "fmla z17.s, z20.s, z2.s[2]\n"
       "ld1w { z20.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z10.s, z21.s, z0.s[2]\n"
       "fmla z14.s, z21.s, z1.s[2]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z18.s, z21.s, z2.s[2]\n"
       "fmla z11.s, z20.s, z0.s[2]\n"
+      "addvl x9, x9, #1\n"
       "fmla z15.s, z20.s, z1.s[2]\n"
       "fmla z19.s, z20.s, z2.s[2]\n"
       "ble 40f\n"
       "ld1w { z21.s }, p5/Z, [x12]\n"
       "ld1w { z20.s }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z21.s, z0.s[3]\n"
       "fmla z12.s, z21.s, z1.s[3]\n"
       "fmla z16.s, z21.s, z2.s[3]\n"
-      "ld1w { z21.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
       "fmla z9.s, z20.s, z0.s[3]\n"
+      "ld1w { z21.s }, p5/Z, [x10]\n"
+      "addvl x12, x12, #1\n"
       "fmla z13.s, z20.s, z1.s[3]\n"
       "fmla z17.s, z20.s, z2.s[3]\n"
       "ld1w { z20.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z10.s, z21.s, z0.s[3]\n"
       "fmla z14.s, z21.s, z1.s[3]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.s, z21.s, z2.s[3]\n"
       "fmla z11.s, z20.s, z0.s[3]\n"
       "fmla z15.s, z20.s, z1.s[3]\n"
@@ -858,12 +856,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 35b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "tbz %x[flags], #1, 41f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z21.s }, p5/Z, [x21]\n"
       "ld1rw { z20.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z21.s\n"
       "fmin z9.s, p5/M, z9.s, z21.s\n"
@@ -895,14 +893,14 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
       "42:"  // Height 3: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -910,15 +908,15 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "b 86f\n"
       "43:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "44:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -926,12 +924,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 45f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 45f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 45f\n"
       "mov x11, x12\n"
       "45:"  // Height 4: B setup done
@@ -946,18 +944,18 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cbz x15, 46f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -965,13 +963,13 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "46:"  // Height 4: no bias
       "tbz %x[flags], #0, 47f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x13, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x13]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x22, x13, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
@@ -1006,8 +1004,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "49:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 50f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1032,25 +1030,25 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 53f\n"
       "52:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z25.s }, p5/Z, [x12]\n"
-      "ld1w { z24.s }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z3.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z2.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "ld1rqw { z1.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "cmp x27, #0x4\n"
+      "add x26, x26, #0x10\n"
+      "ld1w { z25.s }, p5/Z, [x12]\n"
+      "ld1w { z24.s }, p5/Z, [x11]\n"
       "fmla z8.s, z25.s, z3.s[0]\n"
       "fmla z12.s, z25.s, z2.s[0]\n"
-      "fmla z9.s, z24.s, z3.s[0]\n"
-      "fmla z13.s, z24.s, z2.s[0]\n"
       "fmla z16.s, z25.s, z1.s[0]\n"
       "fmla z20.s, z25.s, z0.s[0]\n"
       "ld1w { z25.s }, p5/Z, [x10]\n"
+      "add x25, x25, #0x10\n"
+      "fmla z9.s, z24.s, z3.s[0]\n"
+      "fmla z13.s, z24.s, z2.s[0]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "fmla z17.s, z24.s, z1.s[0]\n"
       "fmla z21.s, z24.s, z0.s[0]\n"
       "ld1w { z24.s }, p5/Z, [x9]\n"
@@ -1129,22 +1127,22 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "bgt 52b\n"
       "53:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z25.s }, p5/Z, [x12]\n"
-      "ld1w { z24.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
       "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      "ld1w { z25.s }, p5/Z, [x12]\n"
+      "ld1w { z24.s }, p5/Z, [x11]\n"
       "fmla z8.s, z25.s, z0.s[0]\n"
       "fmla z12.s, z25.s, z1.s[0]\n"
-      "fmla z9.s, z24.s, z0.s[0]\n"
-      "fmla z13.s, z24.s, z1.s[0]\n"
       "fmla z16.s, z25.s, z2.s[0]\n"
       "fmla z20.s, z25.s, z3.s[0]\n"
       "ld1w { z25.s }, p5/Z, [x10]\n"
+      "addvl x12, x12, #1\n"
+      "fmla z9.s, z24.s, z0.s[0]\n"
+      "fmla z13.s, z24.s, z1.s[0]\n"
+      "addvl x11, x11, #1\n"
       "addvl x10, x10, #1\n"
       "fmla z17.s, z24.s, z2.s[0]\n"
       "fmla z21.s, z24.s, z3.s[0]\n"
@@ -1161,23 +1159,23 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 54f\n"
       "ld1w { z25.s }, p5/Z, [x12]\n"
       "ld1w { z24.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z25.s, z0.s[1]\n"
       "fmla z12.s, z25.s, z1.s[1]\n"
       "fmla z16.s, z25.s, z2.s[1]\n"
       "fmla z20.s, z25.s, z3.s[1]\n"
       "ld1w { z25.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.s, z24.s, z0.s[1]\n"
       "fmla z13.s, z24.s, z1.s[1]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z17.s, z24.s, z2.s[1]\n"
       "fmla z21.s, z24.s, z3.s[1]\n"
       "ld1w { z24.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.s, z25.s, z0.s[1]\n"
       "fmla z14.s, z25.s, z1.s[1]\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.s, z25.s, z2.s[1]\n"
       "fmla z22.s, z25.s, z3.s[1]\n"
       "fmla z11.s, z24.s, z0.s[1]\n"
@@ -1187,23 +1185,23 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 54f\n"
       "ld1w { z25.s }, p5/Z, [x12]\n"
       "ld1w { z24.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z25.s, z0.s[2]\n"
       "fmla z12.s, z25.s, z1.s[2]\n"
       "fmla z16.s, z25.s, z2.s[2]\n"
       "fmla z20.s, z25.s, z3.s[2]\n"
       "ld1w { z25.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.s, z24.s, z0.s[2]\n"
       "fmla z13.s, z24.s, z1.s[2]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z17.s, z24.s, z2.s[2]\n"
       "fmla z21.s, z24.s, z3.s[2]\n"
       "ld1w { z24.s }, p5/Z, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z10.s, z25.s, z0.s[2]\n"
       "fmla z14.s, z25.s, z1.s[2]\n"
+      "addvl x9, x9, #1\n"
       "fmla z18.s, z25.s, z2.s[2]\n"
       "fmla z22.s, z25.s, z3.s[2]\n"
       "fmla z11.s, z24.s, z0.s[2]\n"
@@ -1213,16 +1211,16 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 54f\n"
       "ld1w { z25.s }, p5/Z, [x12]\n"
       "ld1w { z24.s }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z25.s, z0.s[3]\n"
       "fmla z12.s, z25.s, z1.s[3]\n"
       "fmla z16.s, z25.s, z2.s[3]\n"
       "fmla z20.s, z25.s, z3.s[3]\n"
       "ld1w { z25.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x12, x12, #1\n"
       "fmla z9.s, z24.s, z0.s[3]\n"
       "fmla z13.s, z24.s, z1.s[3]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.s, z24.s, z2.s[3]\n"
       "fmla z21.s, z24.s, z3.s[3]\n"
       "ld1w { z24.s }, p5/Z, [x9]\n"
@@ -1241,13 +1239,13 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 49b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "tbz %x[flags], #1, 55f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z25.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z25.s }, p5/Z, [x21]\n"
       "ld1rw { z24.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z25.s\n"
       "fmin z9.s, p5/M, z9.s, z25.s\n"
@@ -1287,18 +1285,18 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
       "56:"  // Height 4: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1306,15 +1304,15 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "b 86f\n"
       "57:"  // Height 5
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "58:"  // Height 5: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -1322,12 +1320,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 59f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 59f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 59f\n"
       "mov x11, x12\n"
       "59:"  // Height 5: B setup done
@@ -1342,18 +1340,18 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cbz x15, 60f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1365,16 +1363,16 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "60:"  // Height 5: no bias
       "tbz %x[flags], #0, 61f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x13]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x23, x13, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x22]\n"
@@ -1415,8 +1413,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "63:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 64f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1444,29 +1442,29 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 67f\n"
       "66:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z29.s }, p5/Z, [x12]\n"
-      "ld1w { z28.s }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z4.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z3.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z1.s }, p0/Z, [x23]\n"
+      "cmp x27, #0x4\n"
+      "add x26, x26, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1w { z29.s }, p5/Z, [x12]\n"
       "fmla z8.s, z29.s, z4.s[0]\n"
       "fmla z12.s, z29.s, z3.s[0]\n"
-      "fmla z9.s, z28.s, z4.s[0]\n"
+      "ld1w { z28.s }, p5/Z, [x11]\n"
       "fmla z16.s, z29.s, z2.s[0]\n"
       "fmla z20.s, z29.s, z1.s[0]\n"
+      "add x25, x25, #0x10\n"
       "fmla z24.s, z29.s, z0.s[0]\n"
-      "fmla z13.s, z28.s, z3.s[0]\n"
+      "fmla z9.s, z28.s, z4.s[0]\n"
       "ld1w { z29.s }, p5/Z, [x10]\n"
+      "add x24, x24, #0x10\n"
+      "fmla z13.s, z28.s, z3.s[0]\n"
       "fmla z17.s, z28.s, z2.s[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z21.s, z28.s, z1.s[0]\n"
       "fmla z25.s, z28.s, z0.s[0]\n"
       "ld1w { z28.s }, p5/Z, [x9]\n"
@@ -1475,8 +1473,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "fmla z18.s, z29.s, z2.s[0]\n"
       "fmla z22.s, z29.s, z1.s[0]\n"
       "fmla z26.s, z29.s, z0.s[0]\n"
-      "ld1w { z29.s }, p5/Z, [x12, #1, MUL VL]\n"
       "fmla z11.s, z28.s, z4.s[0]\n"
+      "ld1w { z29.s }, p5/Z, [x12, #1, MUL VL]\n"
       "fmla z15.s, z28.s, z3.s[0]\n"
       "fmla z19.s, z28.s, z2.s[0]\n"
       "fmla z23.s, z28.s, z1.s[0]\n"
@@ -1487,8 +1485,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "fmla z16.s, z29.s, z2.s[1]\n"
       "fmla z20.s, z29.s, z1.s[1]\n"
       "fmla z24.s, z29.s, z0.s[1]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.s, z28.s, z4.s[1]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z13.s, z28.s, z3.s[1]\n"
       "fmla z17.s, z28.s, z2.s[1]\n"
       "fmla z21.s, z28.s, z1.s[1]\n"
@@ -1499,8 +1497,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "fmla z18.s, z29.s, z2.s[1]\n"
       "fmla z22.s, z29.s, z1.s[1]\n"
       "fmla z26.s, z29.s, z0.s[1]\n"
-      "ld1w { z29.s }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z11.s, z28.s, z4.s[1]\n"
+      "ld1w { z29.s }, p5/Z, [x12, #2, MUL VL]\n"
       "fmla z15.s, z28.s, z3.s[1]\n"
       "fmla z19.s, z28.s, z2.s[1]\n"
       "fmla z23.s, z28.s, z1.s[1]\n"
@@ -1511,8 +1509,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "fmla z16.s, z29.s, z2.s[2]\n"
       "fmla z20.s, z29.s, z1.s[2]\n"
       "fmla z24.s, z29.s, z0.s[2]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z28.s, z4.s[2]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.s, z28.s, z3.s[2]\n"
       "fmla z17.s, z28.s, z2.s[2]\n"
       "fmla z21.s, z28.s, z1.s[2]\n"
@@ -1523,30 +1521,30 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "fmla z18.s, z29.s, z2.s[2]\n"
       "fmla z22.s, z29.s, z1.s[2]\n"
       "fmla z26.s, z29.s, z0.s[2]\n"
+      "fmla z11.s, z28.s, z4.s[2]\n"
       "ld1w { z29.s }, p5/Z, [x12, #3, MUL VL]\n"
       "addvl x12, x12, #4\n"
-      "fmla z11.s, z28.s, z4.s[2]\n"
       "fmla z15.s, z28.s, z3.s[2]\n"
       "fmla z19.s, z28.s, z2.s[2]\n"
       "fmla z23.s, z28.s, z1.s[2]\n"
       "fmla z27.s, z28.s, z0.s[2]\n"
       "ld1w { z28.s }, p5/Z, [x11, #3, MUL VL]\n"
-      "fmla z8.s, z29.s, z4.s[3]\n"
       "addvl x11, x11, #4\n"
+      "fmla z8.s, z29.s, z4.s[3]\n"
       "fmla z12.s, z29.s, z3.s[3]\n"
       "fmla z16.s, z29.s, z2.s[3]\n"
       "fmla z20.s, z29.s, z1.s[3]\n"
       "fmla z24.s, z29.s, z0.s[3]\n"
+      "fmla z9.s, z28.s, z4.s[3]\n"
       "ld1w { z29.s }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
-      "fmla z9.s, z28.s, z4.s[3]\n"
       "fmla z13.s, z28.s, z3.s[3]\n"
       "fmla z17.s, z28.s, z2.s[3]\n"
       "fmla z21.s, z28.s, z1.s[3]\n"
       "fmla z25.s, z28.s, z0.s[3]\n"
       "ld1w { z28.s }, p5/Z, [x9, #3, MUL VL]\n"
-      "fmla z10.s, z29.s, z4.s[3]\n"
       "addvl x9, x9, #4\n"
+      "fmla z10.s, z29.s, z4.s[3]\n"
       "fmla z14.s, z29.s, z3.s[3]\n"
       "fmla z18.s, z29.s, z2.s[3]\n"
       "fmla z22.s, z29.s, z1.s[3]\n"
@@ -1559,25 +1557,25 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "bgt 66b\n"
       "67:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z29.s }, p5/Z, [x12]\n"
-      "ld1w { z28.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
       "ld1rqw { z3.s }, p0/Z, [x23]\n"
       "ld1rqw { z4.s }, p0/Z, [x22]\n"
+      "ld1w { z29.s }, p5/Z, [x12]\n"
       "fmla z8.s, z29.s, z0.s[0]\n"
       "fmla z12.s, z29.s, z1.s[0]\n"
-      "fmla z9.s, z28.s, z0.s[0]\n"
-      "fmla z13.s, z28.s, z1.s[0]\n"
+      "ld1w { z28.s }, p5/Z, [x11]\n"
       "fmla z16.s, z29.s, z2.s[0]\n"
       "fmla z20.s, z29.s, z3.s[0]\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.s, z29.s, z4.s[0]\n"
-      "fmla z17.s, z28.s, z2.s[0]\n"
+      "fmla z9.s, z28.s, z0.s[0]\n"
       "ld1w { z29.s }, p5/Z, [x10]\n"
+      "addvl x11, x11, #1\n"
+      "fmla z13.s, z28.s, z1.s[0]\n"
+      "fmla z17.s, z28.s, z2.s[0]\n"
       "addvl x10, x10, #1\n"
       "fmla z21.s, z28.s, z3.s[0]\n"
       "fmla z25.s, z28.s, z4.s[0]\n"
@@ -1596,19 +1594,19 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 68f\n"
       "ld1w { z29.s }, p5/Z, [x12]\n"
       "ld1w { z28.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z29.s, z0.s[1]\n"
       "fmla z12.s, z29.s, z1.s[1]\n"
       "fmla z16.s, z29.s, z2.s[1]\n"
       "fmla z20.s, z29.s, z3.s[1]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.s, z29.s, z4.s[1]\n"
       "fmla z9.s, z28.s, z0.s[1]\n"
       "ld1w { z29.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.s, z28.s, z1.s[1]\n"
       "fmla z17.s, z28.s, z2.s[1]\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.s, z28.s, z3.s[1]\n"
       "fmla z25.s, z28.s, z4.s[1]\n"
       "ld1w { z28.s }, p5/Z, [x9]\n"
@@ -1626,19 +1624,19 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 68f\n"
       "ld1w { z29.s }, p5/Z, [x12]\n"
       "ld1w { z28.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z29.s, z0.s[2]\n"
       "fmla z12.s, z29.s, z1.s[2]\n"
       "fmla z16.s, z29.s, z2.s[2]\n"
       "fmla z20.s, z29.s, z3.s[2]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.s, z29.s, z4.s[2]\n"
       "fmla z9.s, z28.s, z0.s[2]\n"
       "ld1w { z29.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z13.s, z28.s, z1.s[2]\n"
       "fmla z17.s, z28.s, z2.s[2]\n"
+      "addvl x10, x10, #1\n"
       "fmla z21.s, z28.s, z3.s[2]\n"
       "fmla z25.s, z28.s, z4.s[2]\n"
       "ld1w { z28.s }, p5/Z, [x9]\n"
@@ -1656,12 +1654,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 68f\n"
       "ld1w { z29.s }, p5/Z, [x12]\n"
       "ld1w { z28.s }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z29.s, z0.s[3]\n"
       "fmla z12.s, z29.s, z1.s[3]\n"
       "fmla z16.s, z29.s, z2.s[3]\n"
       "fmla z20.s, z29.s, z3.s[3]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z24.s, z29.s, z4.s[3]\n"
       "fmla z9.s, z28.s, z0.s[3]\n"
       "ld1w { z29.s }, p5/Z, [x10]\n"
@@ -1688,14 +1686,14 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 63b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "tbz %x[flags], #1, 69f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z29.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z29.s }, p5/Z, [x21]\n"
       "ld1rw { z28.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z29.s\n"
       "fmin z9.s, p5/M, z9.s, z29.s\n"
@@ -1743,22 +1741,22 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
       "70:"  // Height 5: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -1766,19 +1764,18 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "b 86f\n"
       "71:"  // Height 6
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0x18\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0x18\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "72:"  // Height 6: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #3\n"
       "add x11, x12, x20, LSL #2\n"
+      "cntw x21, ALL, MUL #3\n"
       "add x10, x11, x20, LSL #2\n"
       "add x9, x10, x20, LSL #2\n"
       "add x20, x9, x20, LSL #2\n"
@@ -1786,12 +1783,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 73f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 73f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 73f\n"
       "mov x11, x12\n"
       "73:"  // Height 6: B setup done
@@ -1806,18 +1803,18 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cbz x15, 74f\n"
       "ld1w { z8.s }, p5/Z, [x15]\n"
       "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
-      "addvl x15, x15, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x15, x15, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1833,17 +1830,17 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "74:"  // Height 6: no bias
       "tbz %x[flags], #0, 75f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x13, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x13]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n"
-      "add x24, x13, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x24]\n"
       "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x23]\n"
@@ -1892,8 +1889,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "77:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 78f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1924,29 +1921,29 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 81f\n"
       "80:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z1.s }, p5/Z, [x12]\n"
-      "ld1w { z0.s }, p5/Z, [x11]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z7.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z6.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "ld1rqw { z5.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z4.s }, p0/Z, [x23]\n"
+      "cmp x27, #0x4\n"
+      "add x26, x26, #0x10\n"
       "ld1rqw { z3.s }, p0/Z, [x22]\n"
       "ld1rqw { z2.s }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1w { z1.s }, p5/Z, [x12]\n"
+      "ld1w { z0.s }, p5/Z, [x11]\n"
       "fmla z8.s, z1.s, z7.s[0]\n"
       "fmla z12.s, z1.s, z6.s[0]\n"
-      "add x21, x21, #0x10\n"
       "fmla z16.s, z1.s, z5.s[0]\n"
       "fmla z20.s, z1.s, z4.s[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z24.s, z1.s, z3.s[0]\n"
       "fmla z28.s, z1.s, z2.s[0]\n"
       "ld1w { z1.s }, p5/Z, [x10]\n"
+      "add x21, x21, #0x10\n"
       "fmla z9.s, z0.s, z7.s[0]\n"
       "fmla z13.s, z0.s, z6.s[0]\n"
       "fmla z17.s, z0.s, z5.s[0]\n"
@@ -2057,27 +2054,27 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "bgt 80b\n"
       "81:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z7.s }, p5/Z, [x12]\n"
-      "ld1w { z6.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
       "ld1rqw { z3.s }, p0/Z, [x23]\n"
       "ld1rqw { z4.s }, p0/Z, [x22]\n"
       "ld1rqw { z5.s }, p0/Z, [x21]\n"
+      "ld1w { z7.s }, p5/Z, [x12]\n"
+      "ld1w { z6.s }, p5/Z, [x11]\n"
       "fmla z8.s, z7.s, z0.s[0]\n"
       "fmla z12.s, z7.s, z1.s[0]\n"
-      "fmla z9.s, z6.s, z0.s[0]\n"
-      "fmla z13.s, z6.s, z1.s[0]\n"
       "fmla z16.s, z7.s, z2.s[0]\n"
       "fmla z20.s, z7.s, z3.s[0]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z24.s, z7.s, z4.s[0]\n"
       "fmla z28.s, z7.s, z5.s[0]\n"
       "ld1w { z7.s }, p5/Z, [x10]\n"
       "addvl x10, x10, #1\n"
+      "fmla z9.s, z6.s, z0.s[0]\n"
+      "fmla z13.s, z6.s, z1.s[0]\n"
       "fmla z17.s, z6.s, z2.s[0]\n"
       "fmla z21.s, z6.s, z3.s[0]\n"
       "fmla z25.s, z6.s, z4.s[0]\n"
@@ -2099,19 +2096,19 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 82f\n"
       "ld1w { z7.s }, p5/Z, [x12]\n"
       "ld1w { z6.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z7.s, z0.s[1]\n"
       "fmla z12.s, z7.s, z1.s[1]\n"
       "fmla z16.s, z7.s, z2.s[1]\n"
       "fmla z20.s, z7.s, z3.s[1]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.s, z7.s, z4.s[1]\n"
       "fmla z28.s, z7.s, z5.s[1]\n"
       "ld1w { z7.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.s, z6.s, z0.s[1]\n"
       "fmla z13.s, z6.s, z1.s[1]\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.s, z6.s, z2.s[1]\n"
       "fmla z21.s, z6.s, z3.s[1]\n"
       "fmla z25.s, z6.s, z4.s[1]\n"
@@ -2133,19 +2130,19 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 82f\n"
       "ld1w { z7.s }, p5/Z, [x12]\n"
       "ld1w { z6.s }, p5/Z, [x11]\n"
-      "subs x27, x27, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z7.s, z0.s[2]\n"
       "fmla z12.s, z7.s, z1.s[2]\n"
       "fmla z16.s, z7.s, z2.s[2]\n"
       "fmla z20.s, z7.s, z3.s[2]\n"
+      "subs x27, x27, #0x1\n"
+      "addvl x12, x12, #1\n"
       "fmla z24.s, z7.s, z4.s[2]\n"
       "fmla z28.s, z7.s, z5.s[2]\n"
       "ld1w { z7.s }, p5/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z9.s, z6.s, z0.s[2]\n"
       "fmla z13.s, z6.s, z1.s[2]\n"
+      "addvl x10, x10, #1\n"
       "fmla z17.s, z6.s, z2.s[2]\n"
       "fmla z21.s, z6.s, z3.s[2]\n"
       "fmla z25.s, z6.s, z4.s[2]\n"
@@ -2167,12 +2164,12 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "ble 82f\n"
       "ld1w { z7.s }, p5/Z, [x12]\n"
       "ld1w { z6.s }, p5/Z, [x11]\n"
-      "addvl x12, x12, #1\n"
-      "addvl x11, x11, #1\n"
       "fmla z8.s, z7.s, z0.s[3]\n"
       "fmla z12.s, z7.s, z1.s[3]\n"
       "fmla z16.s, z7.s, z2.s[3]\n"
       "fmla z20.s, z7.s, z3.s[3]\n"
+      "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
       "fmla z24.s, z7.s, z4.s[3]\n"
       "fmla z28.s, z7.s, z5.s[3]\n"
       "ld1w { z7.s }, p5/Z, [x10]\n"
@@ -2203,15 +2200,15 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 77b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x13, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x13, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "tbz %x[flags], #1, 83f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p5/Z, [x21]\n"
       "ld1rw { z0.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z1.s\n"
       "fmin z9.s, p5/M, z9.s, z1.s\n"
@@ -2267,26 +2264,26 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x13, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x13, #3, MUL VL]\n"
       "addvl x13, x13, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
-      "st1w { z28.s }, p4, [x22]\n"
-      "st1w { z29.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z30.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z31.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z28.s }, p4, [x21]\n"
+      "st1w { z29.s }, p3, [x21, #1, MUL VL]\n"
+      "st1w { z30.s }, p2, [x21, #2, MUL VL]\n"
+      "st1w { z31.s }, p1, [x21, #3, MUL VL]\n"
       "84:"  // Height 6: Writeback done
       "decw x14, ALL, MUL #4\n"
       "cmp x14, XZR\n"
@@ -2303,8 +2300,8 @@ void sve_ffhybrid_fp32_mla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "86:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp
index 90112a823b..23f686a902 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp
@@ -82,14 +82,16 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 4, 12, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 4, 12, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
         if (std::is_same<T, float>::value) {
             switch (ci->get_cpu_model()) {
+                case CPUModel::V1:
+                    return { 28.74 };
                 default:
-                    return { 32.35 };
+                    return { 15.27 };
             }
         }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp
index 0e98cc6def..c2b6dd1030 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp
@@ -50,19 +50,18 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -83,7 +82,6 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     ka.B_stride = B_stride;
     switch(act.type) {
         default:
@@ -106,17 +104,17 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "bgt 29f\n"
       "beq 15f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #5\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #5\n"
       "add x28, x9, x20, LSL #1\n"
       "add x27, x28, x20, LSL #1\n"
       "add x20, x27, x20, LSL #1\n"
@@ -124,20 +122,20 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x27, x12\n"
       "cmp x14, x21\n"
+      "mov x27, x12\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x28, x12\n"
       "cmp x14, x21\n"
+      "mov x28, x12\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 3f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 3f\n"
       "mov x11, x12\n"
       "3:"  // Height 1: B setup done
@@ -156,19 +154,19 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "cbz x15, 4f\n"
       "ld1w { z8.s }, p7/Z, [x15]\n"
       "ld1w { z9.s }, p7/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p7/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p7/Z, [x15, #3, MUL VL]\n"
-      "ld1w { z12.s }, p7/Z, [x15, #4, MUL VL]\n"
-      "ld1w { z13.s }, p7/Z, [x15, #5, MUL VL]\n"
-      "addvl x15, x15, #6\n"
       "zip2 z14.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p7/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p7/Z, [x15, #3, MUL VL]\n"
       "zip2 z15.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
+      "ld1w { z12.s }, p7/Z, [x15, #4, MUL VL]\n"
+      "ld1w { z13.s }, p7/Z, [x15, #5, MUL VL]\n"
       "zip2 z16.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
       "zip2 z17.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
+      "addvl x15, x15, #6\n"
       "zip2 z18.d, z12.d, z12.d\n"
       "zip1 z12.d, z12.d, z12.d\n"
       "zip2 z19.d, z13.d, z13.d\n"
@@ -176,16 +174,16 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "b 6f\n"
       "4:"  // Height 1: no bias
       "tbz %x[flags], #0, 5f\n"
-      "ld1w { z25.s }, p6/Z, [x13]\n"
-      "ld1w { z24.s }, p5/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z21.s }, p6/Z, [x13]\n"
+      "ld1w { z20.s }, p5/Z, [x13, #1, MUL VL]\n"
+      "zip1 z8.d, z21.d, z14.d\n"
+      "zip2 z14.d, z21.d, z14.d\n"
       "ld1w { z23.s }, p4/Z, [x13, #2, MUL VL]\n"
       "ld1w { z22.s }, p3/Z, [x13, #3, MUL VL]\n"
+      "zip1 z9.d, z20.d, z15.d\n"
+      "zip2 z15.d, z20.d, z15.d\n"
       "ld1w { z21.s }, p2/Z, [x13, #4, MUL VL]\n"
       "ld1w { z20.s }, p1/Z, [x13, #5, MUL VL]\n"
-      "zip1 z8.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "zip1 z9.d, z24.d, z15.d\n"
-      "zip2 z15.d, z24.d, z15.d\n"
       "zip1 z10.d, z23.d, z16.d\n"
       "zip2 z16.d, z23.d, z16.d\n"
       "zip1 z11.d, z22.d, z17.d\n"
@@ -212,8 +210,8 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "mov x26, #0x0\n"
       "7:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 8f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -229,78 +227,78 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "ble 11f\n"
       "10:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z23.h }, p7/Z, [x12]\n"
-      "ld1h { z22.h }, p7/Z, [x12, #1, MUL VL]\n"
-      "sub x25, x25, #0x4\n"
-      "ld1h { z21.h }, p7/Z, [x11]\n"
-      "ld1h { z20.h }, p7/Z, [x11, #1, MUL VL]\n"
-      "cmp x25, #0x4\n"
-      "addvl x12, x12, #2\n"
-      "addvl x11, x11, #2\n"
       "ld1rqw { z24.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       ".inst 0x658abf18  // bfcvt z24.h, p7/M, z24.s\n"
       "uzp1 z24.h, z24.h, z24.h\n"
-      ".inst 0x6477e708  // bfmmla z8.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x10]\n"
-      ".inst 0x6476e70e  // bfmmla z14.s, z24.h, z22.h\n"
-      "ld1h { z22.h }, p7/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
+      "ld1h { z21.h }, p7/Z, [x12]\n"
+      "ld1h { z20.h }, p7/Z, [x12, #1, MUL VL]\n"
+      ".inst 0x6475e708  // bfmmla z8.s, z24.h, z21.h\n"
+      ".inst 0x6474e70e  // bfmmla z14.s, z24.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x11]\n"
+      "ld1h { z20.h }, p7/Z, [x11, #1, MUL VL]\n"
       ".inst 0x6475e709  // bfmmla z9.s, z24.h, z21.h\n"
-      "ld1h { z21.h }, p7/Z, [x9]\n"
       ".inst 0x6474e70f  // bfmmla z15.s, z24.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x10]\n"
+      "ld1h { z20.h }, p7/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x6475e70a  // bfmmla z10.s, z24.h, z21.h\n"
+      ".inst 0x6474e710  // bfmmla z16.s, z24.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x9]\n"
       "ld1h { z20.h }, p7/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
-      ".inst 0x6477e70a  // bfmmla z10.s, z24.h, z23.h\n"
+      ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
+      ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
       "ld1h { z23.h }, p7/Z, [x28]\n"
-      ".inst 0x6476e710  // bfmmla z16.s, z24.h, z22.h\n"
       "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #2\n"
-      ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
+      "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       "ld1h { z21.h }, p7/Z, [x27]\n"
-      ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
       "ld1h { z20.h }, p7/Z, [x27, #1, MUL VL]\n"
-      "addvl x27, x27, #2\n"
       ".inst 0x6477e70c  // bfmmla z12.s, z24.h, z23.h\n"
       ".inst 0x6476e712  // bfmmla z18.s, z24.h, z22.h\n"
       ".inst 0x6475e70d  // bfmmla z13.s, z24.h, z21.h\n"
       ".inst 0x6474e713  // bfmmla z19.s, z24.h, z20.h\n"
+      "add x24, x24, #0x10\n"
+      "addvl x12, x12, #2\n"
+      "addvl x11, x11, #2\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      "addvl x28, x28, #2\n"
+      "addvl x27, x27, #2\n"
       "bgt 10b\n"
       "11:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z23.h }, p7/Z, [x12]\n"
-      "ld1h { z22.h }, p7/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
+      "ld1rqw { z22.s }, p0/Z, [x24]\n"
+      ".inst 0x658abed6  // bfcvt z22.h, p7/M, z22.s\n"
+      "uzp1 z22.h, z22.h, z22.h\n"
+      "ld1h { z21.h }, p7/Z, [x12]\n"
+      "ld1h { z20.h }, p7/Z, [x12, #1, MUL VL]\n"
+      ".inst 0x6475e6c8  // bfmmla z8.s, z22.h, z21.h\n"
+      ".inst 0x6474e6ce  // bfmmla z14.s, z22.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x11]\n"
       "ld1h { z20.h }, p7/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
-      "ld1rqw { z24.s }, p0/Z, [x24]\n"
-      ".inst 0x658abf18  // bfcvt z24.h, p7/M, z24.s\n"
-      "uzp1 z24.h, z24.h, z24.h\n"
-      ".inst 0x6477e708  // bfmmla z8.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x10]\n"
-      ".inst 0x6476e70e  // bfmmla z14.s, z24.h, z22.h\n"
-      "ld1h { z22.h }, p7/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
-      ".inst 0x6475e709  // bfmmla z9.s, z24.h, z21.h\n"
+      ".inst 0x6475e6c9  // bfmmla z9.s, z22.h, z21.h\n"
+      ".inst 0x6474e6cf  // bfmmla z15.s, z22.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x10]\n"
+      "ld1h { z20.h }, p7/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x6475e6ca  // bfmmla z10.s, z22.h, z21.h\n"
+      ".inst 0x6474e6d0  // bfmmla z16.s, z22.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x9]\n"
-      ".inst 0x6474e70f  // bfmmla z15.s, z24.h, z20.h\n"
       "ld1h { z20.h }, p7/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
-      ".inst 0x6477e70a  // bfmmla z10.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x28]\n"
-      ".inst 0x6476e710  // bfmmla z16.s, z24.h, z22.h\n"
-      "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #2\n"
-      ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
+      ".inst 0x6475e6cb  // bfmmla z11.s, z22.h, z21.h\n"
+      ".inst 0x6474e6d1  // bfmmla z17.s, z22.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x28]\n"
+      "ld1h { z20.h }, p7/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x6475e6cc  // bfmmla z12.s, z22.h, z21.h\n"
+      ".inst 0x6474e6d2  // bfmmla z18.s, z22.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x27]\n"
-      ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
       "ld1h { z20.h }, p7/Z, [x27, #1, MUL VL]\n"
+      ".inst 0x6475e6cd  // bfmmla z13.s, z22.h, z21.h\n"
+      ".inst 0x6474e6d3  // bfmmla z19.s, z22.h, z20.h\n"
+      "addvl x12, x12, #2\n"
+      "addvl x11, x11, #2\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      "addvl x28, x28, #2\n"
       "addvl x27, x27, #2\n"
-      ".inst 0x6477e70c  // bfmmla z12.s, z24.h, z23.h\n"
-      ".inst 0x6476e712  // bfmmla z18.s, z24.h, z22.h\n"
-      ".inst 0x6475e70d  // bfmmla z13.s, z24.h, z21.h\n"
-      ".inst 0x6474e713  // bfmmla z19.s, z24.h, z20.h\n"
       "12:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -313,9 +311,9 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "uzp1 z12.d, z12.d, z18.d\n"
       "uzp1 z13.d, z13.d, z19.d\n"
       "tbz %x[flags], #1, 13f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p7/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z21.s }, p7/Z, [x21]\n"
       "ld1rw { z20.s }, p7/Z, [x20]\n"
       "fmin z8.s, p7/M, z8.s, z21.s\n"
       "fmin z9.s, p7/M, z9.s, z21.s\n"
@@ -344,17 +342,17 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "b 58f\n"
       "15:"  // Height 2
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "16:"  // Height 2: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #5\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #5\n"
       "add x28, x9, x20, LSL #1\n"
       "add x27, x28, x20, LSL #1\n"
       "add x20, x27, x20, LSL #1\n"
@@ -362,20 +360,20 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 17f\n"
       "decw x21\n"
-      "mov x27, x12\n"
       "cmp x14, x21\n"
+      "mov x27, x12\n"
       "bgt 17f\n"
       "decw x21\n"
-      "mov x28, x12\n"
       "cmp x14, x21\n"
+      "mov x28, x12\n"
       "bgt 17f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 17f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 17f\n"
       "mov x11, x12\n"
       "17:"  // Height 2: B setup done
@@ -394,19 +392,19 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "cbz x15, 18f\n"
       "ld1w { z8.s }, p7/Z, [x15]\n"
       "ld1w { z9.s }, p7/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p7/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p7/Z, [x15, #3, MUL VL]\n"
-      "ld1w { z12.s }, p7/Z, [x15, #4, MUL VL]\n"
-      "ld1w { z13.s }, p7/Z, [x15, #5, MUL VL]\n"
-      "addvl x15, x15, #6\n"
       "zip2 z14.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p7/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p7/Z, [x15, #3, MUL VL]\n"
       "zip2 z15.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
+      "ld1w { z12.s }, p7/Z, [x15, #4, MUL VL]\n"
+      "ld1w { z13.s }, p7/Z, [x15, #5, MUL VL]\n"
       "zip2 z16.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
       "zip2 z17.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
+      "addvl x15, x15, #6\n"
       "zip2 z18.d, z12.d, z12.d\n"
       "zip1 z12.d, z12.d, z12.d\n"
       "zip2 z19.d, z13.d, z13.d\n"
@@ -415,25 +413,25 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "18:"  // Height 2: no bias
       "tbz %x[flags], #0, 19f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z25.s }, p6/Z, [x13]\n"
-      "ld1w { z24.s }, p5/Z, [x13, #1, MUL VL]\n"
-      "ld1w { z23.s }, p4/Z, [x13, #2, MUL VL]\n"
+      "add x20, x13, x20, LSL #2\n"
+      "ld1w { z16.s }, p6/Z, [x13]\n"
+      "ld1w { z17.s }, p5/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x13, #2, MUL VL]\n"
       "ld1w { z22.s }, p3/Z, [x13, #3, MUL VL]\n"
       "ld1w { z21.s }, p2/Z, [x13, #4, MUL VL]\n"
-      "add x20, x13, x20, LSL #2\n"
       "ld1w { z20.s }, p1/Z, [x13, #5, MUL VL]\n"
       "ld1w { z14.s }, p6/Z, [x20]\n"
+      "zip1 z8.d, z16.d, z14.d\n"
+      "zip2 z14.d, z16.d, z14.d\n"
       "ld1w { z15.s }, p5/Z, [x20, #1, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x20, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z15.d\n"
+      "zip2 z15.d, z17.d, z15.d\n"
       "ld1w { z17.s }, p3/Z, [x20, #3, MUL VL]\n"
       "ld1w { z18.s }, p2/Z, [x20, #4, MUL VL]\n"
+      "zip1 z10.d, z19.d, z16.d\n"
+      "zip2 z16.d, z19.d, z16.d\n"
       "ld1w { z19.s }, p1/Z, [x20, #5, MUL VL]\n"
-      "zip1 z8.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "zip1 z9.d, z24.d, z15.d\n"
-      "zip2 z15.d, z24.d, z15.d\n"
-      "zip1 z10.d, z23.d, z16.d\n"
-      "zip2 z16.d, z23.d, z16.d\n"
       "zip1 z11.d, z22.d, z17.d\n"
       "zip2 z17.d, z22.d, z17.d\n"
       "zip1 z12.d, z21.d, z18.d\n"
@@ -458,8 +456,8 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "mov x26, #0x0\n"
       "21:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 22f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -478,87 +476,87 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "ble 25f\n"
       "24:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z23.h }, p7/Z, [x12]\n"
-      "ld1h { z22.h }, p7/Z, [x12, #1, MUL VL]\n"
-      "sub x25, x25, #0x4\n"
-      "ld1h { z21.h }, p7/Z, [x11]\n"
-      "ld1h { z25.h }, p7/Z, [x11, #1, MUL VL]\n"
-      "cmp x25, #0x4\n"
-      "addvl x12, x12, #2\n"
-      "addvl x11, x11, #2\n"
       "ld1rqw { z24.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z20.s }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
       ".inst 0x658abf18  // bfcvt z24.h, p7/M, z24.s\n"
       ".inst 0x658abe94  // bfcvt z20.h, p7/M, z20.s\n"
       "uzp1 z24.h, z24.h, z24.h\n"
+      "ld1h { z23.h }, p7/Z, [x12]\n"
+      "ld1h { z22.h }, p7/Z, [x12, #1, MUL VL]\n"
       "uzp1 z20.h, z20.h, z20.h\n"
       "trn1 z24.d, z24.d, z20.d\n"
+      "ld1h { z21.h }, p7/Z, [x11]\n"
+      "ld1h { z20.h }, p7/Z, [x11, #1, MUL VL]\n"
       ".inst 0x6477e708  // bfmmla z8.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x10]\n"
       ".inst 0x6476e70e  // bfmmla z14.s, z24.h, z22.h\n"
+      "ld1h { z23.h }, p7/Z, [x10]\n"
       "ld1h { z22.h }, p7/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
       ".inst 0x6475e709  // bfmmla z9.s, z24.h, z21.h\n"
+      ".inst 0x6474e70f  // bfmmla z15.s, z24.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x9]\n"
-      ".inst 0x6479e70f  // bfmmla z15.s, z24.h, z25.h\n"
       "ld1h { z20.h }, p7/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
       ".inst 0x6477e70a  // bfmmla z10.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x28]\n"
       ".inst 0x6476e710  // bfmmla z16.s, z24.h, z22.h\n"
+      "ld1h { z23.h }, p7/Z, [x28]\n"
       "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #2\n"
       ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
-      "ld1h { z21.h }, p7/Z, [x27]\n"
       ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x27]\n"
       "ld1h { z20.h }, p7/Z, [x27, #1, MUL VL]\n"
-      "addvl x27, x27, #2\n"
+      "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       ".inst 0x6477e70c  // bfmmla z12.s, z24.h, z23.h\n"
       ".inst 0x6476e712  // bfmmla z18.s, z24.h, z22.h\n"
       ".inst 0x6475e70d  // bfmmla z13.s, z24.h, z21.h\n"
       ".inst 0x6474e713  // bfmmla z19.s, z24.h, z20.h\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
+      "addvl x12, x12, #2\n"
+      "addvl x11, x11, #2\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      "addvl x28, x28, #2\n"
+      "addvl x27, x27, #2\n"
       "bgt 24b\n"
       "25:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z23.h }, p7/Z, [x12]\n"
-      "ld1h { z22.h }, p7/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
-      "ld1h { z21.h }, p7/Z, [x11]\n"
-      "ld1h { z25.h }, p7/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
       "ld1rqw { z24.s }, p0/Z, [x24]\n"
       "ld1rqw { z20.s }, p0/Z, [x23]\n"
       ".inst 0x658abf18  // bfcvt z24.h, p7/M, z24.s\n"
       ".inst 0x658abe94  // bfcvt z20.h, p7/M, z20.s\n"
       "uzp1 z24.h, z24.h, z24.h\n"
+      "ld1h { z23.h }, p7/Z, [x12]\n"
+      "ld1h { z22.h }, p7/Z, [x12, #1, MUL VL]\n"
       "uzp1 z20.h, z20.h, z20.h\n"
       "trn1 z24.d, z24.d, z20.d\n"
+      "ld1h { z21.h }, p7/Z, [x11]\n"
+      "ld1h { z20.h }, p7/Z, [x11, #1, MUL VL]\n"
       ".inst 0x6477e708  // bfmmla z8.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x10]\n"
       ".inst 0x6476e70e  // bfmmla z14.s, z24.h, z22.h\n"
+      "ld1h { z23.h }, p7/Z, [x10]\n"
       "ld1h { z22.h }, p7/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
       ".inst 0x6475e709  // bfmmla z9.s, z24.h, z21.h\n"
+      ".inst 0x6474e70f  // bfmmla z15.s, z24.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x9]\n"
-      ".inst 0x6479e70f  // bfmmla z15.s, z24.h, z25.h\n"
       "ld1h { z20.h }, p7/Z, [x9, #1, MUL VL]\n"
-      "addvl x9, x9, #2\n"
       ".inst 0x6477e70a  // bfmmla z10.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x28]\n"
       ".inst 0x6476e710  // bfmmla z16.s, z24.h, z22.h\n"
+      "ld1h { z23.h }, p7/Z, [x28]\n"
       "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #2\n"
       ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
-      "ld1h { z21.h }, p7/Z, [x27]\n"
       ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x27]\n"
       "ld1h { z20.h }, p7/Z, [x27, #1, MUL VL]\n"
-      "addvl x27, x27, #2\n"
       ".inst 0x6477e70c  // bfmmla z12.s, z24.h, z23.h\n"
       ".inst 0x6476e712  // bfmmla z18.s, z24.h, z22.h\n"
+      "addvl x12, x12, #2\n"
+      "addvl x11, x11, #2\n"
       ".inst 0x6475e70d  // bfmmla z13.s, z24.h, z21.h\n"
       ".inst 0x6474e713  // bfmmla z19.s, z24.h, z20.h\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      "addvl x28, x28, #2\n"
+      "addvl x27, x27, #2\n"
       "26:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -567,21 +565,21 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z4.d, z8.d, z14.d\n"
       "uzp2 z8.d, z8.d, z14.d\n"
+      "add x23, x13, x20, LSL #2\n"
       "uzp1 z14.d, z9.d, z15.d\n"
       "uzp2 z9.d, z9.d, z15.d\n"
       "uzp1 z15.d, z10.d, z16.d\n"
       "uzp2 z10.d, z10.d, z16.d\n"
       "uzp1 z16.d, z11.d, z17.d\n"
       "uzp2 z11.d, z11.d, z17.d\n"
-      "add x24, x13, x20, LSL #2\n"
       "uzp1 z17.d, z12.d, z18.d\n"
       "uzp2 z12.d, z12.d, z18.d\n"
       "uzp1 z18.d, z13.d, z19.d\n"
       "uzp2 z13.d, z13.d, z19.d\n"
       "tbz %x[flags], #1, 27f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z20.s }, p7/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z20.s }, p7/Z, [x21]\n"
       "ld1rw { z19.s }, p7/Z, [x20]\n"
       "fmin z4.s, p7/M, z4.s, z20.s\n"
       "fmin z14.s, p7/M, z14.s, z20.s\n"
@@ -615,12 +613,12 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "st1w { z17.s }, p2, [x13, #4, MUL VL]\n"
       "st1w { z18.s }, p1, [x13, #5, MUL VL]\n"
       "addvl x13, x13, #6\n"
-      "st1w { z8.s }, p6, [x24]\n"
-      "st1w { z9.s }, p5, [x24, #1, MUL VL]\n"
-      "st1w { z10.s }, p4, [x24, #2, MUL VL]\n"
-      "st1w { z11.s }, p3, [x24, #3, MUL VL]\n"
-      "st1w { z12.s }, p2, [x24, #4, MUL VL]\n"
-      "st1w { z13.s }, p1, [x24, #5, MUL VL]\n"
+      "st1w { z8.s }, p6, [x23]\n"
+      "st1w { z9.s }, p5, [x23, #1, MUL VL]\n"
+      "st1w { z10.s }, p4, [x23, #2, MUL VL]\n"
+      "st1w { z11.s }, p3, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p2, [x23, #4, MUL VL]\n"
+      "st1w { z13.s }, p1, [x23, #5, MUL VL]\n"
       "28:"  // Height 2: Writeback done
       "decw x14, ALL, MUL #6\n"
       "cmp x14, XZR\n"
@@ -628,17 +626,17 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "b 58f\n"
       "29:"  // Height 3
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x15, %x[bias]\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x13, %x[output_ptr]\n"
       "30:"  // Height 3: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #5\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #5\n"
       "add x28, x9, x20, LSL #1\n"
       "add x27, x28, x20, LSL #1\n"
       "add x20, x27, x20, LSL #1\n"
@@ -646,20 +644,20 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 31f\n"
       "decw x21\n"
-      "mov x27, x12\n"
       "cmp x14, x21\n"
+      "mov x27, x12\n"
       "bgt 31f\n"
       "decw x21\n"
-      "mov x28, x12\n"
       "cmp x14, x21\n"
+      "mov x28, x12\n"
       "bgt 31f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 31f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 31f\n"
       "mov x11, x12\n"
       "31:"  // Height 3: B setup done
@@ -678,19 +676,19 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "cbz x15, 32f\n"
       "ld1w { z8.s }, p7/Z, [x15]\n"
       "ld1w { z9.s }, p7/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p7/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p7/Z, [x15, #3, MUL VL]\n"
-      "ld1w { z12.s }, p7/Z, [x15, #4, MUL VL]\n"
-      "ld1w { z13.s }, p7/Z, [x15, #5, MUL VL]\n"
-      "addvl x15, x15, #6\n"
       "zip2 z14.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p7/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p7/Z, [x15, #3, MUL VL]\n"
       "zip2 z15.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
+      "ld1w { z12.s }, p7/Z, [x15, #4, MUL VL]\n"
+      "ld1w { z13.s }, p7/Z, [x15, #5, MUL VL]\n"
       "zip2 z16.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
       "zip2 z17.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
+      "addvl x15, x15, #6\n"
       "zip2 z18.d, z12.d, z12.d\n"
       "zip1 z12.d, z12.d, z12.d\n"
       "zip2 z19.d, z13.d, z13.d\n"
@@ -711,38 +709,38 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "32:"  // Height 3: no bias
       "tbz %x[flags], #0, 33f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z22.s }, p6/Z, [x13]\n"
-      "ld1w { z24.s }, p5/Z, [x13, #1, MUL VL]\n"
-      "ld1w { z0.s }, p4/Z, [x13, #2, MUL VL]\n"
-      "ld1w { z2.s }, p3/Z, [x13, #3, MUL VL]\n"
-      "ld1w { z1.s }, p2/Z, [x13, #4, MUL VL]\n"
       "add x21, x13, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z16.s }, p6/Z, [x13]\n"
+      "ld1w { z17.s }, p5/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x13, #2, MUL VL]\n"
+      "ld1w { z22.s }, p3/Z, [x13, #3, MUL VL]\n"
+      "ld1w { z24.s }, p2/Z, [x13, #4, MUL VL]\n"
       "ld1w { z20.s }, p1/Z, [x13, #5, MUL VL]\n"
       "ld1w { z14.s }, p6/Z, [x21]\n"
+      "zip1 z8.d, z16.d, z14.d\n"
+      "zip2 z14.d, z16.d, z14.d\n"
       "ld1w { z15.s }, p5/Z, [x21, #1, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x21, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z15.d\n"
+      "zip2 z15.d, z17.d, z15.d\n"
       "ld1w { z17.s }, p3/Z, [x21, #3, MUL VL]\n"
       "ld1w { z18.s }, p2/Z, [x21, #4, MUL VL]\n"
+      "zip1 z10.d, z19.d, z16.d\n"
+      "zip2 z16.d, z19.d, z16.d\n"
       "ld1w { z19.s }, p1/Z, [x21, #5, MUL VL]\n"
       "ld1w { z21.s }, p6/Z, [x20]\n"
-      "zip1 z8.d, z22.d, z14.d\n"
-      "zip2 z14.d, z22.d, z14.d\n"
+      "zip1 z11.d, z22.d, z17.d\n"
+      "zip2 z17.d, z22.d, z17.d\n"
       "ld1w { z22.s }, p5/Z, [x20, #1, MUL VL]\n"
       "ld1w { z23.s }, p4/Z, [x20, #2, MUL VL]\n"
-      "zip1 z9.d, z24.d, z15.d\n"
-      "zip2 z15.d, z24.d, z15.d\n"
+      "zip1 z12.d, z24.d, z18.d\n"
+      "zip2 z18.d, z24.d, z18.d\n"
       "ld1w { z24.s }, p3/Z, [x20, #3, MUL VL]\n"
       "ld1w { z25.s }, p2/Z, [x20, #4, MUL VL]\n"
-      "zip1 z10.d, z0.d, z16.d\n"
-      "zip2 z16.d, z0.d, z16.d\n"
-      "ld1w { z0.s }, p1/Z, [x20, #5, MUL VL]\n"
-      "zip1 z11.d, z2.d, z17.d\n"
-      "zip2 z17.d, z2.d, z17.d\n"
-      "zip1 z12.d, z1.d, z18.d\n"
-      "zip2 z18.d, z1.d, z18.d\n"
       "zip1 z13.d, z20.d, z19.d\n"
       "zip2 z19.d, z20.d, z19.d\n"
+      "ld1w { z0.s }, p1/Z, [x20, #5, MUL VL]\n"
       "zip1 z20.d, z21.d, z26.d\n"
       "zip2 z26.d, z21.d, z26.d\n"
       "zip1 z21.d, z22.d, z27.d\n"
@@ -785,8 +783,8 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "mov x26, #0x0\n"
       "35:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 36f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -808,111 +806,111 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "ble 39f\n"
       "38:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z3.h }, p7/Z, [x12]\n"
-      "ld1h { z2.h }, p7/Z, [x12, #1, MUL VL]\n"
-      "sub x25, x25, #0x4\n"
-      "ld1h { z1.h }, p7/Z, [x11]\n"
-      "ld1h { z6.h }, p7/Z, [x11, #1, MUL VL]\n"
-      "cmp x25, #0x4\n"
-      "addvl x12, x12, #2\n"
-      "addvl x11, x11, #2\n"
       "ld1rqw { z5.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "ld1rqw { z4.s }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
       ".inst 0x658abca5  // bfcvt z5.h, p7/M, z5.s\n"
+      "ld1rqw { z4.s }, p0/Z, [x22]\n"
       ".inst 0x658abc00  // bfcvt z0.h, p7/M, z0.s\n"
       "uzp1 z5.h, z5.h, z5.h\n"
-      ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
+      "ld1h { z3.h }, p7/Z, [x12]\n"
       "uzp1 z0.h, z0.h, z0.h\n"
+      ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
+      "ld1h { z2.h }, p7/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z1.h }, p7/Z, [x11]\n"
       "trn1 z5.d, z5.d, z0.d\n"
       "uzp1 z4.h, z4.h, z4.h\n"
+      "ld1h { z0.h }, p7/Z, [x11, #1, MUL VL]\n"
       ".inst 0x6463e4a8  // bfmmla z8.s, z5.h, z3.h\n"
       ".inst 0x6463e494  // bfmmla z20.s, z4.h, z3.h\n"
-      "ld1h { z3.h }, p7/Z, [x10]\n"
       ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
+      "ld1h { z3.h }, p7/Z, [x10]\n"
+      "sub x25, x25, #0x4\n"
       ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
-      "ld1h { z2.h }, p7/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
       ".inst 0x6461e4a9  // bfmmla z9.s, z5.h, z1.h\n"
+      "ld1h { z2.h }, p7/Z, [x10, #1, MUL VL]\n"
+      "cmp x25, #0x4\n"
       ".inst 0x6461e495  // bfmmla z21.s, z4.h, z1.h\n"
-      ".inst 0x6466e4af  // bfmmla z15.s, z5.h, z6.h\n"
+      ".inst 0x6460e4af  // bfmmla z15.s, z5.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x9]\n"
-      ".inst 0x6466e49b  // bfmmla z27.s, z4.h, z6.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6460e49b  // bfmmla z27.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x9, #1, MUL VL]\n"
       ".inst 0x6463e4aa  // bfmmla z10.s, z5.h, z3.h\n"
-      "addvl x9, x9, #2\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6463e496  // bfmmla z22.s, z4.h, z3.h\n"
       ".inst 0x6462e4b0  // bfmmla z16.s, z5.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
       ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #2\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
       ".inst 0x6460e4b1  // bfmmla z17.s, z5.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x27]\n"
+      "addvl x11, x11, #2\n"
       ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x27, #1, MUL VL]\n"
       ".inst 0x6463e4ac  // bfmmla z12.s, z5.h, z3.h\n"
-      "addvl x27, x27, #2\n"
+      "addvl x10, x10, #2\n"
       ".inst 0x6463e498  // bfmmla z24.s, z4.h, z3.h\n"
       ".inst 0x6462e4b2  // bfmmla z18.s, z5.h, z2.h\n"
+      "addvl x9, x9, #2\n"
+      "addvl x28, x28, #2\n"
       ".inst 0x6462e49e  // bfmmla z30.s, z4.h, z2.h\n"
       ".inst 0x6461e4ad  // bfmmla z13.s, z5.h, z1.h\n"
+      "addvl x27, x27, #2\n"
       ".inst 0x6461e499  // bfmmla z25.s, z4.h, z1.h\n"
       ".inst 0x6460e4b3  // bfmmla z19.s, z5.h, z0.h\n"
       ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
       "bgt 38b\n"
       "39:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z3.h }, p7/Z, [x12]\n"
-      "ld1h { z2.h }, p7/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
-      "ld1h { z6.h }, p7/Z, [x11]\n"
-      "ld1h { z1.h }, p7/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
       "ld1rqw { z5.s }, p0/Z, [x24]\n"
       "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "ld1rqw { z4.s }, p0/Z, [x22]\n"
       ".inst 0x658abca5  // bfcvt z5.h, p7/M, z5.s\n"
+      "ld1rqw { z4.s }, p0/Z, [x22]\n"
       ".inst 0x658abc00  // bfcvt z0.h, p7/M, z0.s\n"
       "uzp1 z5.h, z5.h, z5.h\n"
+      "ld1h { z3.h }, p7/Z, [x12]\n"
       "uzp1 z0.h, z0.h, z0.h\n"
       ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
+      "ld1h { z2.h }, p7/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z1.h }, p7/Z, [x11]\n"
       "trn1 z5.d, z5.d, z0.d\n"
       "uzp1 z4.h, z4.h, z4.h\n"
+      "ld1h { z0.h }, p7/Z, [x11, #1, MUL VL]\n"
       ".inst 0x6463e4a8  // bfmmla z8.s, z5.h, z3.h\n"
-      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
-      ".inst 0x6466e4a9  // bfmmla z9.s, z5.h, z6.h\n"
-      ".inst 0x6461e4af  // bfmmla z15.s, z5.h, z1.h\n"
       ".inst 0x6463e494  // bfmmla z20.s, z4.h, z3.h\n"
+      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x10]\n"
+      "addvl x12, x12, #2\n"
       ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
+      ".inst 0x6461e4a9  // bfmmla z9.s, z5.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6466e495  // bfmmla z21.s, z4.h, z6.h\n"
-      ".inst 0x6461e49b  // bfmmla z27.s, z4.h, z1.h\n"
+      "addvl x11, x11, #2\n"
+      ".inst 0x6461e495  // bfmmla z21.s, z4.h, z1.h\n"
+      ".inst 0x6460e4af  // bfmmla z15.s, z5.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x9]\n"
-      "ld1h { z0.h }, p7/Z, [x9, #1, MUL VL]\n"
       "addvl x10, x10, #2\n"
-      "addvl x9, x9, #2\n"
+      ".inst 0x6460e49b  // bfmmla z27.s, z4.h, z0.h\n"
+      "ld1h { z0.h }, p7/Z, [x9, #1, MUL VL]\n"
       ".inst 0x6463e4aa  // bfmmla z10.s, z5.h, z3.h\n"
+      "addvl x9, x9, #2\n"
       ".inst 0x6463e496  // bfmmla z22.s, z4.h, z3.h\n"
       ".inst 0x6462e4b0  // bfmmla z16.s, z5.h, z2.h\n"
-      ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28]\n"
-      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
       ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
-      ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
-      "ld1h { z1.h }, p7/Z, [x27]\n"
+      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
       "addvl x28, x28, #2\n"
+      ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
       ".inst 0x6460e4b1  // bfmmla z17.s, z5.h, z0.h\n"
+      "ld1h { z1.h }, p7/Z, [x27]\n"
       ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x27, #1, MUL VL]\n"
-      "addvl x27, x27, #2\n"
       ".inst 0x6463e4ac  // bfmmla z12.s, z5.h, z3.h\n"
+      "addvl x27, x27, #2\n"
       ".inst 0x6463e498  // bfmmla z24.s, z4.h, z3.h\n"
       ".inst 0x6462e4b2  // bfmmla z18.s, z5.h, z2.h\n"
       ".inst 0x6462e49e  // bfmmla z30.s, z4.h, z2.h\n"
@@ -926,16 +924,16 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "cmp x26, x20\n"
       "bne 35b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #2\n"
       "uzp1 z4.d, z8.d, z14.d\n"
       "uzp2 z8.d, z8.d, z14.d\n"
       "uzp1 z14.d, z9.d, z15.d\n"
       "uzp2 z9.d, z9.d, z15.d\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp1 z15.d, z10.d, z16.d\n"
       "uzp2 z10.d, z10.d, z16.d\n"
-      "add x24, x13, x20, LSL #2\n"
       "uzp1 z16.d, z11.d, z17.d\n"
       "uzp2 z11.d, z11.d, z17.d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z17.d, z12.d, z18.d\n"
       "uzp2 z12.d, z12.d, z18.d\n"
       "uzp1 z18.d, z13.d, z19.d\n"
@@ -947,9 +945,9 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "uzp1 z24.d, z24.d, z30.d\n"
       "uzp1 z25.d, z25.d, z31.d\n"
       "tbz %x[flags], #1, 41f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z0.s }, p7/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z0.s }, p7/Z, [x21]\n"
       "ld1rw { z19.s }, p7/Z, [x20]\n"
       "fmin z4.s, p7/M, z4.s, z0.s\n"
       "fmin z14.s, p7/M, z14.s, z0.s\n"
@@ -995,18 +993,18 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "st1w { z17.s }, p2, [x13, #4, MUL VL]\n"
       "st1w { z18.s }, p1, [x13, #5, MUL VL]\n"
       "addvl x13, x13, #6\n"
-      "st1w { z8.s }, p6, [x24]\n"
-      "st1w { z9.s }, p5, [x24, #1, MUL VL]\n"
-      "st1w { z10.s }, p4, [x24, #2, MUL VL]\n"
-      "st1w { z11.s }, p3, [x24, #3, MUL VL]\n"
-      "st1w { z12.s }, p2, [x24, #4, MUL VL]\n"
-      "st1w { z13.s }, p1, [x24, #5, MUL VL]\n"
-      "st1w { z20.s }, p6, [x23]\n"
-      "st1w { z21.s }, p5, [x23, #1, MUL VL]\n"
-      "st1w { z22.s }, p4, [x23, #2, MUL VL]\n"
-      "st1w { z23.s }, p3, [x23, #3, MUL VL]\n"
-      "st1w { z24.s }, p2, [x23, #4, MUL VL]\n"
-      "st1w { z25.s }, p1, [x23, #5, MUL VL]\n"
+      "st1w { z8.s }, p6, [x23]\n"
+      "st1w { z9.s }, p5, [x23, #1, MUL VL]\n"
+      "st1w { z10.s }, p4, [x23, #2, MUL VL]\n"
+      "st1w { z11.s }, p3, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p2, [x23, #4, MUL VL]\n"
+      "st1w { z13.s }, p1, [x23, #5, MUL VL]\n"
+      "st1w { z20.s }, p6, [x22]\n"
+      "st1w { z21.s }, p5, [x22, #1, MUL VL]\n"
+      "st1w { z22.s }, p4, [x22, #2, MUL VL]\n"
+      "st1w { z23.s }, p3, [x22, #3, MUL VL]\n"
+      "st1w { z24.s }, p2, [x22, #4, MUL VL]\n"
+      "st1w { z25.s }, p1, [x22, #5, MUL VL]\n"
       "42:"  // Height 3: Writeback done
       "decw x14, ALL, MUL #6\n"
       "cmp x14, XZR\n"
@@ -1014,21 +1012,20 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "b 58f\n"
       "43:"  // Height 4
       "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n"
-      "mov x21, #0x10\n"
       "ldr x14, [%x[args_ptr], %[offsetof_N]]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x21, #0x10\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "madd x21, x20, x21, x13\n"
-      "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x15, %x[bias]\n"
+      "mov x13, %x[output_ptr]\n"
+      "madd %x[output_ptr], x20, x21, %x[output_ptr]\n"
       "44:"  // Height 4: Column loop
       "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x21, ALL, MUL #5\n"
       "add x11, x12, x20, LSL #1\n"
       "add x10, x11, x20, LSL #1\n"
       "add x9, x10, x20, LSL #1\n"
+      "cntw x21, ALL, MUL #5\n"
       "add x28, x9, x20, LSL #1\n"
       "add x27, x28, x20, LSL #1\n"
       "add x20, x27, x20, LSL #1\n"
@@ -1036,20 +1033,20 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "bgt 45f\n"
       "decw x21\n"
-      "mov x27, x12\n"
       "cmp x14, x21\n"
+      "mov x27, x12\n"
       "bgt 45f\n"
       "decw x21\n"
-      "mov x28, x12\n"
       "cmp x14, x21\n"
+      "mov x28, x12\n"
       "bgt 45f\n"
       "decw x21\n"
-      "mov x9, x12\n"
       "cmp x14, x21\n"
+      "mov x9, x12\n"
       "bgt 45f\n"
       "decw x21\n"
-      "mov x10, x12\n"
       "cmp x14, x21\n"
+      "mov x10, x12\n"
       "bgt 45f\n"
       "mov x11, x12\n"
       "45:"  // Height 4: B setup done
@@ -1068,19 +1065,19 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "cbz x15, 46f\n"
       "ld1w { z8.s }, p7/Z, [x15]\n"
       "ld1w { z9.s }, p7/Z, [x15, #1, MUL VL]\n"
-      "ld1w { z10.s }, p7/Z, [x15, #2, MUL VL]\n"
-      "ld1w { z11.s }, p7/Z, [x15, #3, MUL VL]\n"
-      "ld1w { z12.s }, p7/Z, [x15, #4, MUL VL]\n"
-      "ld1w { z13.s }, p7/Z, [x15, #5, MUL VL]\n"
-      "addvl x15, x15, #6\n"
       "zip2 z14.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p7/Z, [x15, #2, MUL VL]\n"
+      "ld1w { z11.s }, p7/Z, [x15, #3, MUL VL]\n"
       "zip2 z15.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
+      "ld1w { z12.s }, p7/Z, [x15, #4, MUL VL]\n"
+      "ld1w { z13.s }, p7/Z, [x15, #5, MUL VL]\n"
       "zip2 z16.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
       "zip2 z17.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
+      "addvl x15, x15, #6\n"
       "zip2 z18.d, z12.d, z12.d\n"
       "zip1 z12.d, z12.d, z12.d\n"
       "zip2 z19.d, z13.d, z13.d\n"
@@ -1101,51 +1098,51 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "46:"  // Height 4: no bias
       "tbz %x[flags], #0, 47f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z22.s }, p6/Z, [x13]\n"
-      "ld1w { z24.s }, p5/Z, [x13, #1, MUL VL]\n"
-      "ld1w { z26.s }, p4/Z, [x13, #2, MUL VL]\n"
-      "ld1w { z27.s }, p3/Z, [x13, #3, MUL VL]\n"
-      "ld1w { z29.s }, p2/Z, [x13, #4, MUL VL]\n"
       "add x22, x13, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z16.s }, p6/Z, [x13]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p5/Z, [x13, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x13, #2, MUL VL]\n"
+      "ld1w { z22.s }, p3/Z, [x13, #3, MUL VL]\n"
+      "ld1w { z24.s }, p2/Z, [x13, #4, MUL VL]\n"
       "ld1w { z20.s }, p1/Z, [x13, #5, MUL VL]\n"
       "ld1w { z14.s }, p6/Z, [x22]\n"
+      "zip1 z8.d, z16.d, z14.d\n"
+      "zip2 z14.d, z16.d, z14.d\n"
       "ld1w { z15.s }, p5/Z, [x22, #1, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x22, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z15.d\n"
+      "zip2 z15.d, z17.d, z15.d\n"
       "ld1w { z17.s }, p3/Z, [x22, #3, MUL VL]\n"
       "ld1w { z18.s }, p2/Z, [x22, #4, MUL VL]\n"
+      "zip1 z10.d, z19.d, z16.d\n"
+      "zip2 z16.d, z19.d, z16.d\n"
       "ld1w { z19.s }, p1/Z, [x22, #5, MUL VL]\n"
       "ld1w { z21.s }, p6/Z, [x21]\n"
-      "zip1 z8.d, z22.d, z14.d\n"
-      "zip2 z14.d, z22.d, z14.d\n"
+      "zip1 z11.d, z22.d, z17.d\n"
+      "zip2 z17.d, z22.d, z17.d\n"
       "ld1w { z22.s }, p5/Z, [x21, #1, MUL VL]\n"
       "ld1w { z23.s }, p4/Z, [x21, #2, MUL VL]\n"
-      "zip1 z9.d, z24.d, z15.d\n"
-      "zip2 z15.d, z24.d, z15.d\n"
+      "zip1 z12.d, z24.d, z18.d\n"
+      "zip2 z18.d, z24.d, z18.d\n"
       "ld1w { z24.s }, p3/Z, [x21, #3, MUL VL]\n"
       "ld1w { z25.s }, p2/Z, [x21, #4, MUL VL]\n"
-      "zip1 z10.d, z26.d, z16.d\n"
-      "zip2 z16.d, z26.d, z16.d\n"
-      "ld1w { z0.s }, p1/Z, [x21, #5, MUL VL]\n"
-      "ld1w { z26.s }, p6/Z, [x20]\n"
-      "zip1 z11.d, z27.d, z17.d\n"
-      "zip2 z17.d, z27.d, z17.d\n"
-      "ld1w { z27.s }, p5/Z, [x20, #1, MUL VL]\n"
-      "ld1w { z28.s }, p4/Z, [x20, #2, MUL VL]\n"
-      "zip1 z12.d, z29.d, z18.d\n"
-      "zip2 z18.d, z29.d, z18.d\n"
-      "ld1w { z29.s }, p3/Z, [x20, #3, MUL VL]\n"
-      "ld1w { z30.s }, p2/Z, [x20, #4, MUL VL]\n"
       "zip1 z13.d, z20.d, z19.d\n"
       "zip2 z19.d, z20.d, z19.d\n"
-      "ld1w { z31.s }, p1/Z, [x20, #5, MUL VL]\n"
+      "ld1w { z0.s }, p1/Z, [x21, #5, MUL VL]\n"
+      "ld1w { z26.s }, p6/Z, [x20]\n"
       "zip1 z20.d, z21.d, z26.d\n"
       "zip2 z26.d, z21.d, z26.d\n"
+      "ld1w { z27.s }, p5/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z28.s }, p4/Z, [x20, #2, MUL VL]\n"
       "zip1 z21.d, z22.d, z27.d\n"
       "zip2 z27.d, z22.d, z27.d\n"
+      "ld1w { z29.s }, p3/Z, [x20, #3, MUL VL]\n"
+      "ld1w { z30.s }, p2/Z, [x20, #4, MUL VL]\n"
       "zip1 z22.d, z23.d, z28.d\n"
       "zip2 z28.d, z23.d, z28.d\n"
+      "ld1w { z31.s }, p1/Z, [x20, #5, MUL VL]\n"
       "zip1 z23.d, z24.d, z29.d\n"
       "zip2 z29.d, z24.d, z29.d\n"
       "zip1 z24.d, z25.d, z30.d\n"
@@ -1182,8 +1179,8 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "mov x26, #0x0\n"
       "49:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 50f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1208,144 +1205,144 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "ble 53f\n"
       "52:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z3.h }, p7/Z, [x12]\n"
-      "ld1h { z2.h }, p7/Z, [x12, #1, MUL VL]\n"
-      "sub x25, x25, #0x4\n"
-      "ld1h { z7.h }, p7/Z, [x11]\n"
-      "ld1h { z6.h }, p7/Z, [x11, #1, MUL VL]\n"
-      "cmp x25, #0x4\n"
-      "addvl x12, x12, #2\n"
-      "addvl x11, x11, #2\n"
-      "ld1rqw { z5.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqw { z1.s }, p0/Z, [x23]\n"
-      "ld1rqw { z4.s }, p0/Z, [x22]\n"
-      "ld1rqw { z0.s }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
+      "ld1rqw { z7.s }, p0/Z, [x24]\n"
+      "ld1rqw { z6.s }, p0/Z, [x23]\n"
+      ".inst 0x658abce7  // bfcvt z7.h, p7/M, z7.s\n"
+      "ld1rqw { z5.s }, p0/Z, [x22]\n"
+      "ld1rqw { z4.s }, p0/Z, [x21]\n"
+      ".inst 0x658abcc6  // bfcvt z6.h, p7/M, z6.s\n"
       ".inst 0x658abca5  // bfcvt z5.h, p7/M, z5.s\n"
-      ".inst 0x658abc21  // bfcvt z1.h, p7/M, z1.s\n"
       ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
-      ".inst 0x658abc00  // bfcvt z0.h, p7/M, z0.s\n"
+      "uzp1 z7.h, z7.h, z7.h\n"
+      "ld1h { z3.h }, p7/Z, [x12]\n"
+      "ld1h { z2.h }, p7/Z, [x12, #1, MUL VL]\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
       "uzp1 z5.h, z5.h, z5.h\n"
-      "uzp1 z1.h, z1.h, z1.h\n"
+      "ld1h { z1.h }, p7/Z, [x11]\n"
+      "ld1h { z0.h }, p7/Z, [x11, #1, MUL VL]\n"
       "uzp1 z4.h, z4.h, z4.h\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
-      "trn1 z5.d, z5.d, z1.d\n"
-      "trn1 z4.d, z4.d, z0.d\n"
-      ".inst 0x6463e4a8  // bfmmla z8.s, z5.h, z3.h\n"
-      ".inst 0x6463e494  // bfmmla z20.s, z4.h, z3.h\n"
+      "trn1 z7.d, z7.d, z6.d\n"
+      ".inst 0x6463e4e8  // bfmmla z8.s, z7.h, z3.h\n"
+      "sub x25, x25, #0x4\n"
+      "trn1 z5.d, z5.d, z4.d\n"
+      ".inst 0x6463e4b4  // bfmmla z20.s, z5.h, z3.h\n"
+      ".inst 0x6462e4ee  // bfmmla z14.s, z7.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x10]\n"
-      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
-      ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
+      ".inst 0x6462e4ba  // bfmmla z26.s, z5.h, z2.h\n"
+      ".inst 0x6461e4e9  // bfmmla z9.s, z7.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
-      ".inst 0x6467e4a9  // bfmmla z9.s, z5.h, z7.h\n"
-      ".inst 0x6467e495  // bfmmla z21.s, z4.h, z7.h\n"
-      ".inst 0x6466e4af  // bfmmla z15.s, z5.h, z6.h\n"
+      "cmp x25, #0x4\n"
+      ".inst 0x6461e4b5  // bfmmla z21.s, z5.h, z1.h\n"
+      ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x9]\n"
-      ".inst 0x6466e49b  // bfmmla z27.s, z4.h, z6.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6460e4bb  // bfmmla z27.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x9, #1, MUL VL]\n"
-      ".inst 0x6463e4aa  // bfmmla z10.s, z5.h, z3.h\n"
-      "addvl x9, x9, #2\n"
-      ".inst 0x6463e496  // bfmmla z22.s, z4.h, z3.h\n"
-      ".inst 0x6462e4b0  // bfmmla z16.s, z5.h, z2.h\n"
+      ".inst 0x6463e4ea  // bfmmla z10.s, z7.h, z3.h\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6463e4b6  // bfmmla z22.s, z5.h, z3.h\n"
+      ".inst 0x6462e4f0  // bfmmla z16.s, z7.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28]\n"
-      ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
-      ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
+      "add x22, x22, #0x10\n"
+      ".inst 0x6462e4bc  // bfmmla z28.s, z5.h, z2.h\n"
+      ".inst 0x6461e4eb  // bfmmla z11.s, z7.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #2\n"
-      ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
-      ".inst 0x6460e4b1  // bfmmla z17.s, z5.h, z0.h\n"
+      "add x21, x21, #0x10\n"
+      ".inst 0x6461e4b7  // bfmmla z23.s, z5.h, z1.h\n"
+      ".inst 0x6460e4f1  // bfmmla z17.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x27]\n"
-      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
+      "addvl x12, x12, #2\n"
+      ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x27, #1, MUL VL]\n"
-      ".inst 0x6463e4ac  // bfmmla z12.s, z5.h, z3.h\n"
+      ".inst 0x6463e4ec  // bfmmla z12.s, z7.h, z3.h\n"
+      "addvl x11, x11, #2\n"
+      ".inst 0x6463e4b8  // bfmmla z24.s, z5.h, z3.h\n"
+      ".inst 0x6462e4f2  // bfmmla z18.s, z7.h, z2.h\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      ".inst 0x6462e4be  // bfmmla z30.s, z5.h, z2.h\n"
+      ".inst 0x6461e4ed  // bfmmla z13.s, z7.h, z1.h\n"
+      "addvl x28, x28, #2\n"
       "addvl x27, x27, #2\n"
-      ".inst 0x6463e498  // bfmmla z24.s, z4.h, z3.h\n"
-      ".inst 0x6462e4b2  // bfmmla z18.s, z5.h, z2.h\n"
-      ".inst 0x6462e49e  // bfmmla z30.s, z4.h, z2.h\n"
-      ".inst 0x6461e4ad  // bfmmla z13.s, z5.h, z1.h\n"
-      ".inst 0x6461e499  // bfmmla z25.s, z4.h, z1.h\n"
-      ".inst 0x6460e4b3  // bfmmla z19.s, z5.h, z0.h\n"
-      ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
+      ".inst 0x6461e4b9  // bfmmla z25.s, z5.h, z1.h\n"
+      ".inst 0x6460e4f3  // bfmmla z19.s, z7.h, z0.h\n"
+      ".inst 0x6460e4bf  // bfmmla z31.s, z5.h, z0.h\n"
       "bgt 52b\n"
       "53:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z3.h }, p7/Z, [x12]\n"
-      "ld1h { z2.h }, p7/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
-      "ld1h { z7.h }, p7/Z, [x11]\n"
-      "ld1h { z6.h }, p7/Z, [x11, #1, MUL VL]\n"
-      "addvl x11, x11, #2\n"
-      "ld1rqw { z5.s }, p0/Z, [x24]\n"
-      "ld1rqw { z1.s }, p0/Z, [x23]\n"
-      "ld1rqw { z4.s }, p0/Z, [x22]\n"
-      "ld1rqw { z0.s }, p0/Z, [x21]\n"
+      "ld1rqw { z7.s }, p0/Z, [x24]\n"
+      "ld1rqw { z6.s }, p0/Z, [x23]\n"
+      ".inst 0x658abce7  // bfcvt z7.h, p7/M, z7.s\n"
+      "ld1rqw { z5.s }, p0/Z, [x22]\n"
+      "ld1rqw { z4.s }, p0/Z, [x21]\n"
+      ".inst 0x658abcc6  // bfcvt z6.h, p7/M, z6.s\n"
       ".inst 0x658abca5  // bfcvt z5.h, p7/M, z5.s\n"
-      ".inst 0x658abc21  // bfcvt z1.h, p7/M, z1.s\n"
       ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
-      ".inst 0x658abc00  // bfcvt z0.h, p7/M, z0.s\n"
+      "uzp1 z7.h, z7.h, z7.h\n"
+      "ld1h { z3.h }, p7/Z, [x12]\n"
+      "ld1h { z2.h }, p7/Z, [x12, #1, MUL VL]\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
       "uzp1 z5.h, z5.h, z5.h\n"
-      "uzp1 z1.h, z1.h, z1.h\n"
+      "ld1h { z1.h }, p7/Z, [x11]\n"
+      "ld1h { z0.h }, p7/Z, [x11, #1, MUL VL]\n"
       "uzp1 z4.h, z4.h, z4.h\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
-      "trn1 z5.d, z5.d, z1.d\n"
-      "trn1 z4.d, z4.d, z0.d\n"
-      ".inst 0x6463e4a8  // bfmmla z8.s, z5.h, z3.h\n"
-      ".inst 0x6463e494  // bfmmla z20.s, z4.h, z3.h\n"
+      "trn1 z7.d, z7.d, z6.d\n"
+      ".inst 0x6463e4e8  // bfmmla z8.s, z7.h, z3.h\n"
+      "addvl x12, x12, #2\n"
+      "trn1 z5.d, z5.d, z4.d\n"
+      ".inst 0x6463e4b4  // bfmmla z20.s, z5.h, z3.h\n"
+      ".inst 0x6462e4ee  // bfmmla z14.s, z7.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x10]\n"
-      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
-      ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
+      ".inst 0x6462e4ba  // bfmmla z26.s, z5.h, z2.h\n"
+      ".inst 0x6461e4e9  // bfmmla z9.s, z7.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x10, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
-      ".inst 0x6467e4a9  // bfmmla z9.s, z5.h, z7.h\n"
-      ".inst 0x6467e495  // bfmmla z21.s, z4.h, z7.h\n"
-      ".inst 0x6466e4af  // bfmmla z15.s, z5.h, z6.h\n"
+      "addvl x11, x11, #2\n"
+      ".inst 0x6461e4b5  // bfmmla z21.s, z5.h, z1.h\n"
+      ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x9]\n"
-      ".inst 0x6466e49b  // bfmmla z27.s, z4.h, z6.h\n"
+      "addvl x10, x10, #2\n"
+      ".inst 0x6460e4bb  // bfmmla z27.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x9, #1, MUL VL]\n"
-      ".inst 0x6463e4aa  // bfmmla z10.s, z5.h, z3.h\n"
+      ".inst 0x6463e4ea  // bfmmla z10.s, z7.h, z3.h\n"
       "addvl x9, x9, #2\n"
-      ".inst 0x6463e496  // bfmmla z22.s, z4.h, z3.h\n"
-      ".inst 0x6462e4b0  // bfmmla z16.s, z5.h, z2.h\n"
+      ".inst 0x6463e4b6  // bfmmla z22.s, z5.h, z3.h\n"
+      ".inst 0x6462e4f0  // bfmmla z16.s, z7.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28]\n"
-      ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
-      ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
+      ".inst 0x6462e4bc  // bfmmla z28.s, z5.h, z2.h\n"
+      ".inst 0x6461e4eb  // bfmmla z11.s, z7.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
       "addvl x28, x28, #2\n"
-      ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
-      ".inst 0x6460e4b1  // bfmmla z17.s, z5.h, z0.h\n"
+      ".inst 0x6461e4b7  // bfmmla z23.s, z5.h, z1.h\n"
+      ".inst 0x6460e4f1  // bfmmla z17.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x27]\n"
-      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
+      ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x27, #1, MUL VL]\n"
-      ".inst 0x6463e4ac  // bfmmla z12.s, z5.h, z3.h\n"
+      ".inst 0x6463e4ec  // bfmmla z12.s, z7.h, z3.h\n"
       "addvl x27, x27, #2\n"
-      ".inst 0x6463e498  // bfmmla z24.s, z4.h, z3.h\n"
-      ".inst 0x6462e4b2  // bfmmla z18.s, z5.h, z2.h\n"
-      ".inst 0x6462e49e  // bfmmla z30.s, z4.h, z2.h\n"
-      ".inst 0x6461e4ad  // bfmmla z13.s, z5.h, z1.h\n"
-      ".inst 0x6461e499  // bfmmla z25.s, z4.h, z1.h\n"
-      ".inst 0x6460e4b3  // bfmmla z19.s, z5.h, z0.h\n"
-      ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
+      ".inst 0x6463e4b8  // bfmmla z24.s, z5.h, z3.h\n"
+      ".inst 0x6462e4f2  // bfmmla z18.s, z7.h, z2.h\n"
+      ".inst 0x6462e4be  // bfmmla z30.s, z5.h, z2.h\n"
+      ".inst 0x6461e4ed  // bfmmla z13.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b9  // bfmmla z25.s, z5.h, z1.h\n"
+      ".inst 0x6460e4f3  // bfmmla z19.s, z7.h, z0.h\n"
+      ".inst 0x6460e4bf  // bfmmla z31.s, z5.h, z0.h\n"
       "54:"  // Height 4: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
       "cmp x26, x20\n"
       "bne 49b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x13, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp1 z4.d, z8.d, z14.d\n"
       "uzp2 z8.d, z8.d, z14.d\n"
       "uzp1 z14.d, z9.d, z15.d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z15.d\n"
       "uzp1 z15.d, z10.d, z16.d\n"
       "uzp2 z10.d, z10.d, z16.d\n"
-      "add x24, x13, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z16.d, z11.d, z17.d\n"
       "uzp2 z11.d, z11.d, z17.d\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 z17.d, z12.d, z18.d\n"
       "uzp2 z12.d, z12.d, z18.d\n"
       "uzp1 z18.d, z13.d, z19.d\n"
@@ -1363,9 +1360,9 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "uzp1 z30.d, z25.d, z31.d\n"
       "uzp2 z25.d, z25.d, z31.d\n"
       "tbz %x[flags], #1, 55f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p7/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p7/Z, [x21]\n"
       "ld1rw { z0.s }, p7/Z, [x20]\n"
       "fmin z4.s, p7/M, z4.s, z1.s\n"
       "fmin z14.s, p7/M, z14.s, z1.s\n"
@@ -1423,24 +1420,24 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "st1w { z17.s }, p2, [x13, #4, MUL VL]\n"
       "st1w { z18.s }, p1, [x13, #5, MUL VL]\n"
       "addvl x13, x13, #6\n"
-      "st1w { z8.s }, p6, [x24]\n"
-      "st1w { z9.s }, p5, [x24, #1, MUL VL]\n"
-      "st1w { z10.s }, p4, [x24, #2, MUL VL]\n"
-      "st1w { z11.s }, p3, [x24, #3, MUL VL]\n"
-      "st1w { z12.s }, p2, [x24, #4, MUL VL]\n"
-      "st1w { z13.s }, p1, [x24, #5, MUL VL]\n"
-      "st1w { z19.s }, p6, [x23]\n"
-      "st1w { z26.s }, p5, [x23, #1, MUL VL]\n"
-      "st1w { z27.s }, p4, [x23, #2, MUL VL]\n"
-      "st1w { z28.s }, p3, [x23, #3, MUL VL]\n"
-      "st1w { z29.s }, p2, [x23, #4, MUL VL]\n"
-      "st1w { z30.s }, p1, [x23, #5, MUL VL]\n"
-      "st1w { z20.s }, p6, [x22]\n"
-      "st1w { z21.s }, p5, [x22, #1, MUL VL]\n"
-      "st1w { z22.s }, p4, [x22, #2, MUL VL]\n"
-      "st1w { z23.s }, p3, [x22, #3, MUL VL]\n"
-      "st1w { z24.s }, p2, [x22, #4, MUL VL]\n"
-      "st1w { z25.s }, p1, [x22, #5, MUL VL]\n"
+      "st1w { z8.s }, p6, [x23]\n"
+      "st1w { z9.s }, p5, [x23, #1, MUL VL]\n"
+      "st1w { z10.s }, p4, [x23, #2, MUL VL]\n"
+      "st1w { z11.s }, p3, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p2, [x23, #4, MUL VL]\n"
+      "st1w { z13.s }, p1, [x23, #5, MUL VL]\n"
+      "st1w { z19.s }, p6, [x22]\n"
+      "st1w { z26.s }, p5, [x22, #1, MUL VL]\n"
+      "st1w { z27.s }, p4, [x22, #2, MUL VL]\n"
+      "st1w { z28.s }, p3, [x22, #3, MUL VL]\n"
+      "st1w { z29.s }, p2, [x22, #4, MUL VL]\n"
+      "st1w { z30.s }, p1, [x22, #5, MUL VL]\n"
+      "st1w { z20.s }, p6, [x21]\n"
+      "st1w { z21.s }, p5, [x21, #1, MUL VL]\n"
+      "st1w { z22.s }, p4, [x21, #2, MUL VL]\n"
+      "st1w { z23.s }, p3, [x21, #3, MUL VL]\n"
+      "st1w { z24.s }, p2, [x21, #4, MUL VL]\n"
+      "st1w { z25.s }, p1, [x21, #5, MUL VL]\n"
       "56:"  // Height 4: Writeback done
       "decw x14, ALL, MUL #6\n"
       "cmp x14, XZR\n"
@@ -1457,8 +1454,8 @@ void sve_ffhybrid_fp32bf16fp32_mmla_4x6VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "58:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_dot_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_dot_8x3VL.hpp
deleted file mode 100644
index 70ed2ee4c7..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_dot_8x3VL.hpp
+++ /dev/null
@@ -1,110 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include "../std_transforms_sve.hpp"
-#include "../bfloat.hpp"
-#include "../kernel_weight_format.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    const bfloat16 *, const bfloat16 *, size_t, \
-    float *, int, size_t, int
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void sve_ffinterleaved_bf16fp32_dot_8x3VL( ARGLIST );
-
-class cls_sve_ffinterleaved_bf16fp32_dot_8x3VL
-{
-public:
-    typedef bfloat16 lhs_operand_type;
-    typedef bfloat16 rhs_operand_type;
-    typedef float result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 8;
-    }
-
-    static unsigned int out_width()
-    {
-        return get_vector_length<float>() * 3;
-    }
-    static unsigned int stripe_width()
-    {
-        return get_vector_length<float>();
-    }
-
-    static KernelWeightFormat kernel_weight_format()
-    {
-        return KernelWeightFormat::VL1VL_BL32;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 2;
-    }
-
-
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 2, 1> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 2, 1, true> transforms_quantized = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-
-        if (std::is_same<T, bfloat16>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 20.92, 7.74, 4.14 };
-            }
-        }
-
-
-        if (std::is_same<T, float>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 20.92, 5.18, 4.37 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=sve_ffinterleaved_bf16fp32_dot_8x3VL;
-    cls_sve_ffinterleaved_bf16fp32_dot_8x3VL(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_dot_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_dot_8x3VL/generic.cpp
deleted file mode 100644
index 26192718b5..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_dot_8x3VL/generic.cpp
+++ /dev/null
@@ -1,274 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include <cstddef>
-#include "../../bfloat.hpp"
-
-namespace arm_gemm {
-
-void sve_ffinterleaved_bf16fp32_dot_8x3VL(
-    const bfloat16 *Apanel,
-    const bfloat16 *Bpanel,
-    size_t B_stride,
-    float *Cpanel,
-    int ablocks,
-    size_t N,
-    int K) {
-
-    struct KernelArgs {
-        size_t K = {};
-        const bfloat16 *Bpanel = {};
-        size_t N = {};
-        size_t B_stride = {};
-        const bfloat16 *cur_B_ptr = {};
-    } ka;
-
-    ka.K = (K/2) - 1;
-    ka.Bpanel = Bpanel;
-    ka.N = N;
-    ka.B_stride = B_stride;
-
-    __asm__ __volatile__(
-      "ptrue p0.b\n"
-      "1:"  // Height loop
-      "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
-      "ldr x26, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x25, %x[Apanel]\n"
-      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "2:"  // Width loop
-      "ldr x24, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
-      "cntw x23, ALL, MUL #2\n"
-      "mov %x[Apanel], x25\n"
-      "add x22, x24, x20, LSL #1\n"
-      "cmp x26, x23\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
-      "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
-      "bgt 3f\n"
-      "decw x23\n"
-      "mov x21, x24\n"
-      "cmp x26, x23\n"
-      "bgt 3f\n"
-      "mov x22, x24\n"
-      "3:"  // B setup done
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
-      "mov z8.b, #0x0\n"
-      "mov z9.b, #0x0\n"
-      "mov z10.b, #0x0\n"
-      "mov z11.b, #0x0\n"
-      "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
-      "mov z12.b, #0x0\n"
-      "mov z13.b, #0x0\n"
-      "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
-      "cmp x20, #0x2\n"
-      "mov z14.b, #0x0\n"
-      "mov z15.b, #0x0\n"
-      "mov z16.b, #0x0\n"
-      "mov z17.b, #0x0\n"
-      "ld1h { z4.h }, p0/Z, [x24]\n"
-      "mov z18.b, #0x0\n"
-      "mov z19.b, #0x0\n"
-      "ld1h { z5.h }, p0/Z, [x22]\n"
-      "mov z20.b, #0x0\n"
-      "mov z21.b, #0x0\n"
-      "ld1h { z6.h }, p0/Z, [x21]\n"
-      "mov z22.b, #0x0\n"
-      "mov z23.b, #0x0\n"
-      "mov z24.b, #0x0\n"
-      "mov z25.b, #0x0\n"
-      "mov z26.b, #0x0\n"
-      "mov z27.b, #0x0\n"
-      "mov z28.b, #0x0\n"
-      "mov z29.b, #0x0\n"
-      "mov z30.b, #0x0\n"
-      "mov z31.b, #0x0\n"
-      "blt 5f\n"
-      "4:"  // main loop head
-      ".inst 0x64604088  // bfdot z8.s, z4.h, z0.h[0]\n"
-      ".inst 0x6468408b  // bfdot z11.s, z4.h, z0.h[1]\n"
-      "ld1rqh { z3.h }, p0/Z, [%x[Apanel], #32]\n"
-      ".inst 0x6470408e  // bfdot z14.s, z4.h, z0.h[2]\n"
-      ".inst 0x64784091  // bfdot z17.s, z4.h, z0.h[3]\n"
-      "ld1rqh { z7.h }, p0/Z, [%x[Apanel], #48]\n"
-      ".inst 0x64614094  // bfdot z20.s, z4.h, z1.h[0]\n"
-      ".inst 0x64694097  // bfdot z23.s, z4.h, z1.h[1]\n"
-      "sub x20, x20, #0x2\n"
-      ".inst 0x6471409a  // bfdot z26.s, z4.h, z1.h[2]\n"
-      ".inst 0x6479409d  // bfdot z29.s, z4.h, z1.h[3]\n"
-      "ld1h { z4.h }, p0/Z, [x24, #1, MUL VL]\n"
-      ".inst 0x646040a9  // bfdot z9.s, z5.h, z0.h[0]\n"
-      ".inst 0x646840ac  // bfdot z12.s, z5.h, z0.h[1]\n"
-      "addvl x24, x24, #2\n"
-      ".inst 0x647040af  // bfdot z15.s, z5.h, z0.h[2]\n"
-      ".inst 0x647840b2  // bfdot z18.s, z5.h, z0.h[3]\n"
-      "cmp x20, #0x2\n"
-      ".inst 0x646140b5  // bfdot z21.s, z5.h, z1.h[0]\n"
-      ".inst 0x646940b8  // bfdot z24.s, z5.h, z1.h[1]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
-      ".inst 0x647140bb  // bfdot z27.s, z5.h, z1.h[2]\n"
-      ".inst 0x647940be  // bfdot z30.s, z5.h, z1.h[3]\n"
-      "ld1h { z5.h }, p0/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x646040ca  // bfdot z10.s, z6.h, z0.h[0]\n"
-      ".inst 0x646840cd  // bfdot z13.s, z6.h, z0.h[1]\n"
-      "addvl x22, x22, #2\n"
-      ".inst 0x647040d0  // bfdot z16.s, z6.h, z0.h[2]\n"
-      ".inst 0x647840d3  // bfdot z19.s, z6.h, z0.h[3]\n"
-      "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
-      ".inst 0x646140d6  // bfdot z22.s, z6.h, z1.h[0]\n"
-      ".inst 0x646940d9  // bfdot z25.s, z6.h, z1.h[1]\n"
-      ".inst 0x647140dc  // bfdot z28.s, z6.h, z1.h[2]\n"
-      ".inst 0x647940df  // bfdot z31.s, z6.h, z1.h[3]\n"
-      "ld1h { z2.h }, p0/Z, [x21, #1, MUL VL]\n"
-      "addvl x21, x21, #2\n"
-      ".inst 0x64634088  // bfdot z8.s, z4.h, z3.h[0]\n"
-      ".inst 0x646b408b  // bfdot z11.s, z4.h, z3.h[1]\n"
-      ".inst 0x6473408e  // bfdot z14.s, z4.h, z3.h[2]\n"
-      ".inst 0x647b4091  // bfdot z17.s, z4.h, z3.h[3]\n"
-      "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
-      ".inst 0x64674094  // bfdot z20.s, z4.h, z7.h[0]\n"
-      ".inst 0x646f4097  // bfdot z23.s, z4.h, z7.h[1]\n"
-      ".inst 0x6477409a  // bfdot z26.s, z4.h, z7.h[2]\n"
-      ".inst 0x647f409d  // bfdot z29.s, z4.h, z7.h[3]\n"
-      "ld1h { z4.h }, p0/Z, [x24]\n"
-      ".inst 0x646340a9  // bfdot z9.s, z5.h, z3.h[0]\n"
-      ".inst 0x646b40ac  // bfdot z12.s, z5.h, z3.h[1]\n"
-      ".inst 0x647340af  // bfdot z15.s, z5.h, z3.h[2]\n"
-      ".inst 0x647b40b2  // bfdot z18.s, z5.h, z3.h[3]\n"
-      ".inst 0x646740b5  // bfdot z21.s, z5.h, z7.h[0]\n"
-      ".inst 0x646f40b8  // bfdot z24.s, z5.h, z7.h[1]\n"
-      ".inst 0x647740bb  // bfdot z27.s, z5.h, z7.h[2]\n"
-      ".inst 0x647f40be  // bfdot z30.s, z5.h, z7.h[3]\n"
-      "ld1h { z5.h }, p0/Z, [x22]\n"
-      ".inst 0x6463404a  // bfdot z10.s, z2.h, z3.h[0]\n"
-      ".inst 0x646b404d  // bfdot z13.s, z2.h, z3.h[1]\n"
-      ".inst 0x64734050  // bfdot z16.s, z2.h, z3.h[2]\n"
-      ".inst 0x647b4053  // bfdot z19.s, z2.h, z3.h[3]\n"
-      ".inst 0x64674056  // bfdot z22.s, z2.h, z7.h[0]\n"
-      ".inst 0x646f4059  // bfdot z25.s, z2.h, z7.h[1]\n"
-      ".inst 0x6477405c  // bfdot z28.s, z2.h, z7.h[2]\n"
-      ".inst 0x647f405f  // bfdot z31.s, z2.h, z7.h[3]\n"
-      "ld1h { z6.h }, p0/Z, [x21]\n"
-      "bge 4b\n"
-      "5:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "addvl x24, x24, #1\n"
-      ".inst 0x64604088  // bfdot z8.s, z4.h, z0.h[0]\n"
-      "addvl x22, x22, #1\n"
-      "addvl x21, x21, #1\n"
-      ".inst 0x6468408b  // bfdot z11.s, z4.h, z0.h[1]\n"
-      ".inst 0x6470408e  // bfdot z14.s, z4.h, z0.h[2]\n"
-      ".inst 0x64784091  // bfdot z17.s, z4.h, z0.h[3]\n"
-      ".inst 0x64614094  // bfdot z20.s, z4.h, z1.h[0]\n"
-      ".inst 0x64694097  // bfdot z23.s, z4.h, z1.h[1]\n"
-      ".inst 0x6471409a  // bfdot z26.s, z4.h, z1.h[2]\n"
-      ".inst 0x6479409d  // bfdot z29.s, z4.h, z1.h[3]\n"
-      ".inst 0x646040a9  // bfdot z9.s, z5.h, z0.h[0]\n"
-      ".inst 0x646840ac  // bfdot z12.s, z5.h, z0.h[1]\n"
-      ".inst 0x647040af  // bfdot z15.s, z5.h, z0.h[2]\n"
-      ".inst 0x647840b2  // bfdot z18.s, z5.h, z0.h[3]\n"
-      ".inst 0x646140b5  // bfdot z21.s, z5.h, z1.h[0]\n"
-      ".inst 0x646940b8  // bfdot z24.s, z5.h, z1.h[1]\n"
-      ".inst 0x647140bb  // bfdot z27.s, z5.h, z1.h[2]\n"
-      ".inst 0x647940be  // bfdot z30.s, z5.h, z1.h[3]\n"
-      ".inst 0x646040ca  // bfdot z10.s, z6.h, z0.h[0]\n"
-      ".inst 0x646840cd  // bfdot z13.s, z6.h, z0.h[1]\n"
-      ".inst 0x647040d0  // bfdot z16.s, z6.h, z0.h[2]\n"
-      ".inst 0x647840d3  // bfdot z19.s, z6.h, z0.h[3]\n"
-      ".inst 0x646140d6  // bfdot z22.s, z6.h, z1.h[0]\n"
-      ".inst 0x646940d9  // bfdot z25.s, z6.h, z1.h[1]\n"
-      ".inst 0x647140dc  // bfdot z28.s, z6.h, z1.h[2]\n"
-      ".inst 0x647940df  // bfdot z31.s, z6.h, z1.h[3]\n"
-      "cbz x20, 6f\n"
-      "ld1rqh { z4.h }, p0/Z, [%x[Apanel]]\n"
-      "ld1rqh { z3.h }, p0/Z, [%x[Apanel], #16]\n"
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "ld1h { z2.h }, p0/Z, [x24]\n"
-      "ld1h { z1.h }, p0/Z, [x22]\n"
-      "ld1h { z0.h }, p0/Z, [x21]\n"
-      ".inst 0x64644048  // bfdot z8.s, z2.h, z4.h[0]\n"
-      ".inst 0x646c404b  // bfdot z11.s, z2.h, z4.h[1]\n"
-      ".inst 0x6474404e  // bfdot z14.s, z2.h, z4.h[2]\n"
-      ".inst 0x647c4051  // bfdot z17.s, z2.h, z4.h[3]\n"
-      ".inst 0x64634054  // bfdot z20.s, z2.h, z3.h[0]\n"
-      ".inst 0x646b4057  // bfdot z23.s, z2.h, z3.h[1]\n"
-      ".inst 0x6473405a  // bfdot z26.s, z2.h, z3.h[2]\n"
-      ".inst 0x647b405d  // bfdot z29.s, z2.h, z3.h[3]\n"
-      ".inst 0x64644029  // bfdot z9.s, z1.h, z4.h[0]\n"
-      ".inst 0x646c402c  // bfdot z12.s, z1.h, z4.h[1]\n"
-      ".inst 0x6474402f  // bfdot z15.s, z1.h, z4.h[2]\n"
-      ".inst 0x647c4032  // bfdot z18.s, z1.h, z4.h[3]\n"
-      ".inst 0x64634035  // bfdot z21.s, z1.h, z3.h[0]\n"
-      ".inst 0x646b4038  // bfdot z24.s, z1.h, z3.h[1]\n"
-      ".inst 0x6473403b  // bfdot z27.s, z1.h, z3.h[2]\n"
-      ".inst 0x647b403e  // bfdot z30.s, z1.h, z3.h[3]\n"
-      ".inst 0x6464400a  // bfdot z10.s, z0.h, z4.h[0]\n"
-      ".inst 0x646c400d  // bfdot z13.s, z0.h, z4.h[1]\n"
-      ".inst 0x64744010  // bfdot z16.s, z0.h, z4.h[2]\n"
-      ".inst 0x647c4013  // bfdot z19.s, z0.h, z4.h[3]\n"
-      ".inst 0x64634016  // bfdot z22.s, z0.h, z3.h[0]\n"
-      ".inst 0x646b4019  // bfdot z25.s, z0.h, z3.h[1]\n"
-      ".inst 0x6473401c  // bfdot z28.s, z0.h, z3.h[2]\n"
-      ".inst 0x647b401f  // bfdot z31.s, z0.h, z3.h[3]\n"
-      "6:"  // multiply loop done
-      "decw x26, ALL, MUL #3\n"
-      "st1w { z8.s }, p0, [%x[Cpanel]]\n"
-      "st1w { z9.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "st1w { z10.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "st1w { z11.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "cmp x26, XZR\n"
-      "st1w { z12.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
-      "st1w { z13.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
-      "st1w { z14.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
-      "st1w { z15.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
-      "addvl %x[Cpanel], %x[Cpanel], #16\n"
-      "st1w { z16.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
-      "st1w { z18.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
-      "st1w { z19.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
-      "st1w { z20.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
-      "st1w { z21.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
-      "st1w { z22.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
-      "st1w { z23.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
-      "st1w { z24.s }, p0, [%x[Cpanel]]\n"
-      "st1w { z25.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "st1w { z26.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "st1w { z27.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "st1w { z28.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
-      "st1w { z29.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
-      "st1w { z30.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
-      "st1w { z31.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
-      "addvl %x[Cpanel], %x[Cpanel], #8\n"
-      "bgt 2b\n"
-      "subs %x[ablocks], %x[ablocks], #0x1\n"
-      "bne 1b\n"
-      : [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks)
-      : [args_ptr] "r" (&ka), [offsetof_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr))
-      : "cc", "memory", "p0", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
-    );
-}
-
-} // namespace arm_gemm
-#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp
index 8695a9b53c..1fe5f48da6 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp
@@ -41,8 +41,7 @@ void sve_ffinterleaved_bf16fp32_mmla_8x3VL( ARGLIST );
 class cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL
 {
 public:
-    typedef bfloat16 lhs_operand_type;
-    typedef bfloat16 rhs_operand_type;
+    typedef bfloat16 operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -73,8 +72,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 4, 2> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 4, 2, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
@@ -89,8 +88,10 @@ public:
 
         if (std::is_same<T, float>::value) {
             switch (ci->get_cpu_model()) {
-                default:
-                    return { 39.66, 5.18, 4.37 };
+                case CPUModel::V1:
+                    return { 53.48, 4.23, 6.53 };
+                default:                
+                    return { 29.07, 2.76, 5.39 };
             }
         }
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL/generic.cpp
index 69adb67a6a..09bc24051a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL/generic.cpp
@@ -55,52 +55,52 @@ void sve_ffinterleaved_bf16fp32_mmla_8x3VL(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x26, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x25, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x25, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x24, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
       "cntw x23, ALL, MUL #2\n"
-      "mov %x[Apanel], x25\n"
       "add x22, x24, x20, LSL #1\n"
-      "cmp x26, x23\n"
       "add x21, x22, x20, LSL #1\n"
       "add x20, x21, x20, LSL #1\n"
+      "cmp x26, x23\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov %x[Apanel], x25\n"
       "bgt 3f\n"
       "decw x23\n"
-      "mov x21, x24\n"
       "cmp x26, x23\n"
+      "mov x21, x24\n"
       "bgt 3f\n"
       "mov x22, x24\n"
       "3:"  // B setup done
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
       "mov z10.b, #0x0\n"
-      "mov z11.b, #0x0\n"
       "ld1h { z4.h }, p0/Z, [x24]\n"
+      "mov z11.b, #0x0\n"
       "mov z12.b, #0x0\n"
-      "mov z13.b, #0x0\n"
       "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
-      "cmp x20, #0x2\n"
+      "mov z13.b, #0x0\n"
       "mov z14.b, #0x0\n"
+      "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
       "mov z15.b, #0x0\n"
       "mov z16.b, #0x0\n"
+      "ld1h { z5.h }, p0/Z, [x24, #1, MUL VL]\n"
       "mov z17.b, #0x0\n"
-      "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
       "mov z18.b, #0x0\n"
+      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #32]\n"
       "mov z19.b, #0x0\n"
-      "ld1h { z5.h }, p0/Z, [x24, #1, MUL VL]\n"
       "mov z20.b, #0x0\n"
+      "addvl x24, x24, #2\n"
       "mov z21.b, #0x0\n"
-      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #32]\n"
       "mov z22.b, #0x0\n"
+      "add %x[Apanel], %x[Apanel], #0x30\n"
       "mov z23.b, #0x0\n"
-      "addvl x24, x24, #2\n"
       "mov z24.b, #0x0\n"
       "mov z25.b, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
       "mov z26.b, #0x0\n"
       "mov z27.b, #0x0\n"
       "mov z28.b, #0x0\n"
@@ -114,78 +114,78 @@ void sve_ffinterleaved_bf16fp32_mmla_8x3VL(
       ".inst 0x6465e40b  // bfmmla z11.s, z0.h, z5.h\n"
       ".inst 0x6464e42e  // bfmmla z14.s, z1.h, z4.h\n"
       ".inst 0x6465e431  // bfmmla z17.s, z1.h, z5.h\n"
-      "ld1h { z3.h }, p0/Z, [x22]\n"
+      "ld1h { z7.h }, p0/Z, [x22]\n"
       ".inst 0x6464e454  // bfmmla z20.s, z2.h, z4.h\n"
       ".inst 0x6465e457  // bfmmla z23.s, z2.h, z5.h\n"
-      "ld1h { z7.h }, p0/Z, [x22, #1, MUL VL]\n"
+      "ld1h { z3.h }, p0/Z, [x22, #1, MUL VL]\n"
       ".inst 0x6464e4da  // bfmmla z26.s, z6.h, z4.h\n"
       ".inst 0x6465e4dd  // bfmmla z29.s, z6.h, z5.h\n"
-      "ld1h { z4.h }, p0/Z, [x21]\n"
-      "ld1h { z5.h }, p0/Z, [x21, #1, MUL VL]\n"
-      ".inst 0x6463e409  // bfmmla z9.s, z0.h, z3.h\n"
+      "ld1h { z5.h }, p0/Z, [x21]\n"
+      "ld1h { z4.h }, p0/Z, [x21, #1, MUL VL]\n"
+      ".inst 0x6467e409  // bfmmla z9.s, z0.h, z7.h\n"
+      ".inst 0x6463e40c  // bfmmla z12.s, z0.h, z3.h\n"
+      ".inst 0x6467e42f  // bfmmla z15.s, z1.h, z7.h\n"
+      ".inst 0x6463e432  // bfmmla z18.s, z1.h, z3.h\n"
       "sub x20, x20, #0x2\n"
-      ".inst 0x6467e40c  // bfmmla z12.s, z0.h, z7.h\n"
-      ".inst 0x6463e42f  // bfmmla z15.s, z1.h, z3.h\n"
+      ".inst 0x6467e455  // bfmmla z21.s, z2.h, z7.h\n"
+      ".inst 0x6463e458  // bfmmla z24.s, z2.h, z3.h\n"
       "cmp x20, #0x2\n"
-      ".inst 0x6467e432  // bfmmla z18.s, z1.h, z7.h\n"
-      ".inst 0x6463e455  // bfmmla z21.s, z2.h, z3.h\n"
-      ".inst 0x6467e458  // bfmmla z24.s, z2.h, z7.h\n"
-      ".inst 0x6463e4db  // bfmmla z27.s, z6.h, z3.h\n"
+      ".inst 0x6467e4db  // bfmmla z27.s, z6.h, z7.h\n"
+      ".inst 0x6463e4de  // bfmmla z30.s, z6.h, z3.h\n"
       "ld1h { z3.h }, p0/Z, [x24]\n"
-      ".inst 0x6467e4de  // bfmmla z30.s, z6.h, z7.h\n"
-      ".inst 0x6464e40a  // bfmmla z10.s, z0.h, z4.h\n"
-      "ld1h { z7.h }, p0/Z, [x24, #1, MUL VL]\n"
-      ".inst 0x6465e40d  // bfmmla z13.s, z0.h, z5.h\n"
-      ".inst 0x6464e430  // bfmmla z16.s, z1.h, z4.h\n"
+      ".inst 0x6465e40a  // bfmmla z10.s, z0.h, z5.h\n"
+      ".inst 0x6464e40d  // bfmmla z13.s, z0.h, z4.h\n"
       "ld1rqh { z0.h }, p0/Z, [%x[Apanel], #16]\n"
-      ".inst 0x6465e433  // bfmmla z19.s, z1.h, z5.h\n"
-      ".inst 0x6464e456  // bfmmla z22.s, z2.h, z4.h\n"
+      ".inst 0x6465e430  // bfmmla z16.s, z1.h, z5.h\n"
+      ".inst 0x6464e433  // bfmmla z19.s, z1.h, z4.h\n"
       "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #32]\n"
-      ".inst 0x6465e459  // bfmmla z25.s, z2.h, z5.h\n"
-      ".inst 0x6464e4dc  // bfmmla z28.s, z6.h, z4.h\n"
-      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #48]\n"
-      ".inst 0x6465e4df  // bfmmla z31.s, z6.h, z5.h\n"
+      ".inst 0x6465e456  // bfmmla z22.s, z2.h, z5.h\n"
+      ".inst 0x6464e459  // bfmmla z25.s, z2.h, z4.h\n"
+      "ld1h { z7.h }, p0/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x6465e4dc  // bfmmla z28.s, z6.h, z5.h\n"
+      ".inst 0x6464e4df  // bfmmla z31.s, z6.h, z4.h\n"
+      "ld1rqh { z5.h }, p0/Z, [%x[Apanel], #48]\n"
       "ld1rqh { z6.h }, p0/Z, [%x[Apanel], #64]\n"
       ".inst 0x6463e408  // bfmmla z8.s, z0.h, z3.h\n"
       ".inst 0x6467e40b  // bfmmla z11.s, z0.h, z7.h\n"
       ".inst 0x6463e42e  // bfmmla z14.s, z1.h, z3.h\n"
-      "ld1h { z4.h }, p0/Z, [x22, #2, MUL VL]\n"
       ".inst 0x6467e431  // bfmmla z17.s, z1.h, z7.h\n"
-      ".inst 0x6463e454  // bfmmla z20.s, z2.h, z3.h\n"
-      "ld1h { z5.h }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x6467e457  // bfmmla z23.s, z2.h, z7.h\n"
+      "ld1h { z2.h }, p0/Z, [x22, #2, MUL VL]\n"
+      ".inst 0x6463e4b4  // bfmmla z20.s, z5.h, z3.h\n"
+      ".inst 0x6467e4b7  // bfmmla z23.s, z5.h, z7.h\n"
+      "ld1h { z4.h }, p0/Z, [x22, #3, MUL VL]\n"
       ".inst 0x6463e4da  // bfmmla z26.s, z6.h, z3.h\n"
-      "ld1h { z3.h }, p0/Z, [x21, #2, MUL VL]\n"
       ".inst 0x6467e4dd  // bfmmla z29.s, z6.h, z7.h\n"
+      "ld1h { z3.h }, p0/Z, [x21, #2, MUL VL]\n"
       "ld1h { z7.h }, p0/Z, [x21, #3, MUL VL]\n"
-      ".inst 0x6464e409  // bfmmla z9.s, z0.h, z4.h\n"
-      ".inst 0x6465e40c  // bfmmla z12.s, z0.h, z5.h\n"
-      ".inst 0x6464e42f  // bfmmla z15.s, z1.h, z4.h\n"
+      ".inst 0x6462e409  // bfmmla z9.s, z0.h, z2.h\n"
+      ".inst 0x6464e40c  // bfmmla z12.s, z0.h, z4.h\n"
+      ".inst 0x6462e42f  // bfmmla z15.s, z1.h, z2.h\n"
+      ".inst 0x6464e432  // bfmmla z18.s, z1.h, z4.h\n"
       "addvl x22, x22, #4\n"
-      ".inst 0x6465e432  // bfmmla z18.s, z1.h, z5.h\n"
-      ".inst 0x6464e455  // bfmmla z21.s, z2.h, z4.h\n"
+      ".inst 0x6462e4b5  // bfmmla z21.s, z5.h, z2.h\n"
+      ".inst 0x6464e4b8  // bfmmla z24.s, z5.h, z4.h\n"
       "addvl x21, x21, #4\n"
-      ".inst 0x6465e458  // bfmmla z24.s, z2.h, z5.h\n"
-      ".inst 0x6464e4db  // bfmmla z27.s, z6.h, z4.h\n"
+      ".inst 0x6462e4db  // bfmmla z27.s, z6.h, z2.h\n"
+      ".inst 0x6464e4de  // bfmmla z30.s, z6.h, z4.h\n"
       "ld1h { z4.h }, p0/Z, [x24, #2, MUL VL]\n"
-      ".inst 0x6465e4de  // bfmmla z30.s, z6.h, z5.h\n"
       ".inst 0x6463e40a  // bfmmla z10.s, z0.h, z3.h\n"
-      "ld1h { z5.h }, p0/Z, [x24, #3, MUL VL]\n"
       ".inst 0x6467e40d  // bfmmla z13.s, z0.h, z7.h\n"
-      ".inst 0x6463e430  // bfmmla z16.s, z1.h, z3.h\n"
       "ld1rqh { z0.h }, p0/Z, [%x[Apanel], #80]\n"
+      ".inst 0x6463e430  // bfmmla z16.s, z1.h, z3.h\n"
       ".inst 0x6467e433  // bfmmla z19.s, z1.h, z7.h\n"
-      ".inst 0x6463e456  // bfmmla z22.s, z2.h, z3.h\n"
       "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #96]\n"
-      ".inst 0x6467e459  // bfmmla z25.s, z2.h, z7.h\n"
+      ".inst 0x6463e4b6  // bfmmla z22.s, z5.h, z3.h\n"
+      ".inst 0x6467e4b9  // bfmmla z25.s, z5.h, z7.h\n"
+      "ld1h { z5.h }, p0/Z, [x24, #3, MUL VL]\n"
       ".inst 0x6463e4dc  // bfmmla z28.s, z6.h, z3.h\n"
-      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #112]\n"
       ".inst 0x6467e4df  // bfmmla z31.s, z6.h, z7.h\n"
+      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #112]\n"
       "add %x[Apanel], %x[Apanel], #0x80\n"
       "addvl x24, x24, #4\n"
       "bge 4b\n"
       "5:"  // main loop skip
-      "ld1rqh { z3.h }, p0/Z, [%x[Apanel]]\n"
+      "ld1rqh { z7.h }, p0/Z, [%x[Apanel]]\n"
       ".inst 0x6464e408  // bfmmla z8.s, z0.h, z4.h\n"
       ".inst 0x6465e40b  // bfmmla z11.s, z0.h, z5.h\n"
       ".inst 0x6464e42e  // bfmmla z14.s, z1.h, z4.h\n"
@@ -193,52 +193,52 @@ void sve_ffinterleaved_bf16fp32_mmla_8x3VL(
       "ld1h { z6.h }, p0/Z, [x22]\n"
       ".inst 0x6464e454  // bfmmla z20.s, z2.h, z4.h\n"
       ".inst 0x6465e457  // bfmmla z23.s, z2.h, z5.h\n"
-      "ld1h { z7.h }, p0/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x6464e47a  // bfmmla z26.s, z3.h, z4.h\n"
-      ".inst 0x6465e47d  // bfmmla z29.s, z3.h, z5.h\n"
+      "ld1h { z3.h }, p0/Z, [x22, #1, MUL VL]\n"
+      ".inst 0x6464e4fa  // bfmmla z26.s, z7.h, z4.h\n"
+      ".inst 0x6465e4fd  // bfmmla z29.s, z7.h, z5.h\n"
       "ld1h { z5.h }, p0/Z, [x21]\n"
       "ld1h { z4.h }, p0/Z, [x21, #1, MUL VL]\n"
       ".inst 0x6466e409  // bfmmla z9.s, z0.h, z6.h\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      ".inst 0x6467e40c  // bfmmla z12.s, z0.h, z7.h\n"
+      ".inst 0x6463e40c  // bfmmla z12.s, z0.h, z3.h\n"
       ".inst 0x6466e42f  // bfmmla z15.s, z1.h, z6.h\n"
-      "addvl x22, x22, #2\n"
-      ".inst 0x6467e432  // bfmmla z18.s, z1.h, z7.h\n"
+      ".inst 0x6463e432  // bfmmla z18.s, z1.h, z3.h\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       ".inst 0x6466e455  // bfmmla z21.s, z2.h, z6.h\n"
+      ".inst 0x6463e458  // bfmmla z24.s, z2.h, z3.h\n"
+      "addvl x22, x22, #2\n"
+      ".inst 0x6466e4fb  // bfmmla z27.s, z7.h, z6.h\n"
+      ".inst 0x6463e4fe  // bfmmla z30.s, z7.h, z3.h\n"
       "addvl x21, x21, #2\n"
-      ".inst 0x6467e458  // bfmmla z24.s, z2.h, z7.h\n"
-      ".inst 0x6466e47b  // bfmmla z27.s, z3.h, z6.h\n"
-      ".inst 0x6467e47e  // bfmmla z30.s, z3.h, z7.h\n"
       ".inst 0x6465e40a  // bfmmla z10.s, z0.h, z5.h\n"
       ".inst 0x6464e40d  // bfmmla z13.s, z0.h, z4.h\n"
       ".inst 0x6465e430  // bfmmla z16.s, z1.h, z5.h\n"
       ".inst 0x6464e433  // bfmmla z19.s, z1.h, z4.h\n"
       ".inst 0x6465e456  // bfmmla z22.s, z2.h, z5.h\n"
       ".inst 0x6464e459  // bfmmla z25.s, z2.h, z4.h\n"
-      ".inst 0x6465e47c  // bfmmla z28.s, z3.h, z5.h\n"
-      ".inst 0x6464e47f  // bfmmla z31.s, z3.h, z4.h\n"
+      ".inst 0x6465e4fc  // bfmmla z28.s, z7.h, z5.h\n"
+      ".inst 0x6464e4ff  // bfmmla z31.s, z7.h, z4.h\n"
       "cbz x20, 6f\n"
       "ld1h { z1.h }, p0/Z, [x24]\n"
       "ld1rqh { z7.h }, p0/Z, [%x[Apanel]]\n"
+      ".inst 0x6461e4e8  // bfmmla z8.s, z7.h, z1.h\n"
       "ld1rqh { z6.h }, p0/Z, [%x[Apanel], #16]\n"
       "ld1h { z0.h }, p0/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x6460e4eb  // bfmmla z11.s, z7.h, z0.h\n"
       "ld1rqh { z5.h }, p0/Z, [%x[Apanel], #32]\n"
       "ld1rqh { z4.h }, p0/Z, [%x[Apanel], #48]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
-      ".inst 0x6461e4e8  // bfmmla z8.s, z7.h, z1.h\n"
-      "ld1h { z3.h }, p0/Z, [x22]\n"
-      "ld1h { z2.h }, p0/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x6460e4eb  // bfmmla z11.s, z7.h, z0.h\n"
       ".inst 0x6461e4ce  // bfmmla z14.s, z6.h, z1.h\n"
       ".inst 0x6460e4d1  // bfmmla z17.s, z6.h, z0.h\n"
       ".inst 0x6461e4b4  // bfmmla z20.s, z5.h, z1.h\n"
+      "ld1h { z3.h }, p0/Z, [x22]\n"
       ".inst 0x6460e4b7  // bfmmla z23.s, z5.h, z0.h\n"
       ".inst 0x6461e49a  // bfmmla z26.s, z4.h, z1.h\n"
-      "ld1h { z1.h }, p0/Z, [x21]\n"
+      "ld1h { z2.h }, p0/Z, [x22, #1, MUL VL]\n"
       ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
+      "ld1h { z1.h }, p0/Z, [x21]\n"
       "ld1h { z0.h }, p0/Z, [x21, #1, MUL VL]\n"
       ".inst 0x6463e4e9  // bfmmla z9.s, z7.h, z3.h\n"
       ".inst 0x6462e4ec  // bfmmla z12.s, z7.h, z2.h\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
       ".inst 0x6463e4cf  // bfmmla z15.s, z6.h, z3.h\n"
       ".inst 0x6462e4d2  // bfmmla z18.s, z6.h, z2.h\n"
       ".inst 0x6463e4b5  // bfmmla z21.s, z5.h, z3.h\n"
@@ -255,53 +255,53 @@ void sve_ffinterleaved_bf16fp32_mmla_8x3VL(
       ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
       "6:"  // multiply loop done
       "decw x26, ALL, MUL #3\n"
-      "uzp1 z2.d, z8.d, z11.d\n"
+      "uzp1 z0.d, z8.d, z11.d\n"
       "uzp2 z8.d, z8.d, z11.d\n"
       "uzp1 z1.d, z9.d, z12.d\n"
       "uzp2 z9.d, z9.d, z12.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel]]\n"
       "uzp1 z0.d, z10.d, z13.d\n"
       "uzp2 z10.d, z10.d, z13.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel]]\n"
-      "uzp1 z3.d, z14.d, z17.d\n"
-      "uzp2 z14.d, z14.d, z17.d\n"
       "st1w { z1.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "uzp1 z17.d, z15.d, z18.d\n"
-      "cmp x26, XZR\n"
       "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "uzp2 z15.d, z15.d, z18.d\n"
-      "uzp1 z2.d, z16.d, z19.d\n"
+      "uzp1 z2.d, z14.d, z17.d\n"
+      "uzp2 z14.d, z14.d, z17.d\n"
       "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "uzp2 z16.d, z16.d, z19.d\n"
-      "uzp1 z1.d, z20.d, z23.d\n"
+      "uzp1 z1.d, z15.d, z18.d\n"
+      "cmp x26, XZR\n"
       "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
-      "uzp2 z20.d, z20.d, z23.d\n"
-      "uzp1 z0.d, z21.d, z24.d\n"
+      "uzp2 z15.d, z15.d, z18.d\n"
+      "uzp1 z17.d, z16.d, z19.d\n"
       "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
+      "uzp2 z16.d, z16.d, z19.d\n"
+      "uzp1 z0.d, z20.d, z23.d\n"
+      "st1w { z2.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
+      "uzp2 z20.d, z20.d, z23.d\n"
+      "uzp1 z23.d, z21.d, z24.d\n"
+      "st1w { z1.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
+      "addvl %x[Cpanel], %x[Cpanel], #16\n"
       "uzp2 z21.d, z21.d, z24.d\n"
-      "uzp1 z23.d, z22.d, z25.d\n"
-      "st1w { z3.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
+      "st1w { z17.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
+      "uzp1 z19.d, z22.d, z25.d\n"
       "uzp2 z22.d, z22.d, z25.d\n"
-      "uzp1 z19.d, z26.d, z29.d\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
-      "addvl %x[Cpanel], %x[Cpanel], #16\n"
-      "uzp2 z26.d, z26.d, z29.d\n"
-      "uzp1 z18.d, z27.d, z30.d\n"
-      "uzp2 z27.d, z27.d, z30.d\n"
-      "uzp1 z17.d, z28.d, z31.d\n"
-      "uzp2 z28.d, z28.d, z31.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
       "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
+      "uzp1 z18.d, z26.d, z29.d\n"
+      "uzp2 z26.d, z26.d, z29.d\n"
       "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
+      "uzp1 z17.d, z27.d, z30.d\n"
+      "uzp2 z27.d, z27.d, z30.d\n"
       "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
-      "st1w { z23.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
+      "uzp1 z16.d, z28.d, z31.d\n"
+      "uzp2 z28.d, z28.d, z31.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
+      "st1w { z23.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
+      "st1w { z19.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
       "st1w { z20.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
       "st1w { z21.s }, p0, [%x[Cpanel]]\n"
       "st1w { z22.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "st1w { z19.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "st1w { z18.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+      "st1w { z18.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+      "st1w { z17.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+      "st1w { z16.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
       "st1w { z26.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "st1w { z27.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
       "st1w { z28.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL.hpp
index 22cd8be2b0..13ad2404e3 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL.hpp
@@ -41,8 +41,7 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx( ARGLIST );
 class cls_sve_ffinterleaved_fp16_mla_8x3VL
 {
 public:
-    typedef __fp16 lhs_operand_type;
-    typedef __fp16 rhs_operand_type;
+    typedef __fp16 operand_type;
     typedef __fp16 result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -73,8 +72,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 1, 1> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp
index 6749fc720a..0389fb043a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/a64fx.cpp
@@ -54,52 +54,52 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x26, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x25, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x25, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x24, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
       "cnth x23, ALL, MUL #2\n"
-      "mov %x[Apanel], x25\n"
       "add x22, x24, x20, LSL #1\n"
-      "cmp x26, x23\n"
       "add x21, x22, x20, LSL #1\n"
       "add x20, x21, x20, LSL #1\n"
+      "cmp x26, x23\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov %x[Apanel], x25\n"
       "bgt 3f\n"
       "dech x23\n"
-      "mov x21, x24\n"
       "cmp x26, x23\n"
+      "mov x21, x24\n"
       "bgt 3f\n"
       "mov x22, x24\n"
       "3:"  // B setup done
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
       "mov z10.b, #0x0\n"
-      "mov z11.b, #0x0\n"
       "ld1h { z0.h }, p0/Z, [x24]\n"
+      "mov z11.b, #0x0\n"
       "mov z12.b, #0x0\n"
-      "mov z13.b, #0x0\n"
       "ld1h { z1.h }, p0/Z, [x22]\n"
-      "cmp x20, #0x2\n"
+      "mov z13.b, #0x0\n"
       "mov z14.b, #0x0\n"
+      "ld1h { z2.h }, p0/Z, [x21]\n"
       "mov z15.b, #0x0\n"
       "mov z16.b, #0x0\n"
+      "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n"
       "mov z17.b, #0x0\n"
-      "ld1h { z2.h }, p0/Z, [x21]\n"
       "mov z18.b, #0x0\n"
+      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n"
       "mov z19.b, #0x0\n"
-      "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n"
       "mov z20.b, #0x0\n"
+      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n"
       "mov z21.b, #0x0\n"
-      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n"
       "mov z22.b, #0x0\n"
+      "ld1rh { z6.h }, p0/Z, [%x[Apanel], #6]\n"
       "mov z23.b, #0x0\n"
-      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n"
       "mov z24.b, #0x0\n"
       "mov z25.b, #0x0\n"
-      "ld1rh { z6.h }, p0/Z, [%x[Apanel], #6]\n"
       "mov z26.b, #0x0\n"
       "mov z27.b, #0x0\n"
       "mov z28.b, #0x0\n"
@@ -116,12 +116,12 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx(
       "fmla z11.h, p0/M, z0.h, z4.h\n"
       "fmla z12.h, p0/M, z1.h, z4.h\n"
       "fmla z13.h, p0/M, z2.h, z4.h\n"
-      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #10]\n"
+      "ld1rh { z7.h }, p0/Z, [%x[Apanel], #10]\n"
       "fmla z14.h, p0/M, z0.h, z5.h\n"
       "fmla z15.h, p0/M, z1.h, z5.h\n"
       "cmp x20, #0x2\n"
       "fmla z16.h, p0/M, z2.h, z5.h\n"
-      "ld1rh { z7.h }, p0/Z, [%x[Apanel], #12]\n"
+      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #12]\n"
       "fmla z17.h, p0/M, z0.h, z6.h\n"
       "fmla z18.h, p0/M, z1.h, z6.h\n"
       "fmla z19.h, p0/M, z2.h, z6.h\n"
@@ -129,60 +129,60 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx(
       "fmla z20.h, p0/M, z0.h, z3.h\n"
       "fmla z21.h, p0/M, z1.h, z3.h\n"
       "fmla z22.h, p0/M, z2.h, z3.h\n"
-      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #16]\n"
-      "fmla z23.h, p0/M, z0.h, z4.h\n"
-      "fmla z24.h, p0/M, z1.h, z4.h\n"
-      "fmla z25.h, p0/M, z2.h, z4.h\n"
-      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #18]\n"
-      "fmla z26.h, p0/M, z0.h, z7.h\n"
-      "fmla z27.h, p0/M, z1.h, z7.h\n"
-      "fmla z28.h, p0/M, z2.h, z7.h\n"
-      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #20]\n"
+      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #16]\n"
+      "fmla z23.h, p0/M, z0.h, z7.h\n"
+      "fmla z24.h, p0/M, z1.h, z7.h\n"
+      "fmla z25.h, p0/M, z2.h, z7.h\n"
+      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #18]\n"
+      "fmla z26.h, p0/M, z0.h, z4.h\n"
+      "fmla z27.h, p0/M, z1.h, z4.h\n"
+      "fmla z28.h, p0/M, z2.h, z4.h\n"
+      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #20]\n"
       "fmla z29.h, p0/M, z0.h, z6.h\n"
       "ld1h { z7.h }, p0/Z, [x24, #1, MUL VL]\n"
       "fmla z30.h, p0/M, z1.h, z6.h\n"
       "fmla z31.h, p0/M, z2.h, z6.h\n"
       "ld1h { z6.h }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1h { z2.h }, p0/Z, [x21, #1, MUL VL]\n"
-      "addvl x24, x24, #2\n"
+      "fmla z8.h, p0/M, z7.h, z3.h\n"
       "ld1rh { z1.h }, p0/Z, [%x[Apanel], #22]\n"
-      "addvl x22, x22, #2\n"
-      "addvl x21, x21, #2\n"
-      "fmla z8.h, p0/M, z7.h, z5.h\n"
-      "fmla z11.h, p0/M, z7.h, z4.h\n"
-      "fmla z9.h, p0/M, z6.h, z5.h\n"
-      "fmla z12.h, p0/M, z6.h, z4.h\n"
-      "fmla z10.h, p0/M, z2.h, z5.h\n"
-      "fmla z13.h, p0/M, z2.h, z4.h\n"
-      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #24]\n"
-      "fmla z14.h, p0/M, z7.h, z3.h\n"
-      "fmla z15.h, p0/M, z6.h, z3.h\n"
-      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #26]\n"
-      "fmla z16.h, p0/M, z2.h, z3.h\n"
-      "fmla z17.h, p0/M, z7.h, z1.h\n"
+      "fmla z9.h, p0/M, z6.h, z3.h\n"
+      "fmla z10.h, p0/M, z2.h, z3.h\n"
+      "fmla z11.h, p0/M, z7.h, z5.h\n"
+      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #24]\n"
+      "fmla z12.h, p0/M, z6.h, z5.h\n"
+      "fmla z13.h, p0/M, z2.h, z5.h\n"
+      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #26]\n"
+      "fmla z14.h, p0/M, z7.h, z4.h\n"
+      "fmla z15.h, p0/M, z6.h, z4.h\n"
+      "addvl x24, x24, #2\n"
+      "fmla z16.h, p0/M, z2.h, z4.h\n"
       "ld1rh { z0.h }, p0/Z, [%x[Apanel], #28]\n"
+      "fmla z17.h, p0/M, z7.h, z1.h\n"
       "fmla z18.h, p0/M, z6.h, z1.h\n"
       "fmla z19.h, p0/M, z2.h, z1.h\n"
       "ld1rh { z1.h }, p0/Z, [%x[Apanel], #30]\n"
+      "addvl x22, x22, #2\n"
+      "addvl x21, x21, #2\n"
       "add %x[Apanel], %x[Apanel], #0x20\n"
-      "fmla z20.h, p0/M, z7.h, z5.h\n"
-      "fmla z21.h, p0/M, z6.h, z5.h\n"
-      "fmla z22.h, p0/M, z2.h, z5.h\n"
-      "fmla z23.h, p0/M, z7.h, z4.h\n"
+      "fmla z20.h, p0/M, z7.h, z3.h\n"
+      "fmla z21.h, p0/M, z6.h, z3.h\n"
+      "fmla z22.h, p0/M, z2.h, z3.h\n"
       "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n"
-      "fmla z24.h, p0/M, z6.h, z4.h\n"
-      "fmla z25.h, p0/M, z2.h, z4.h\n"
-      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n"
+      "fmla z23.h, p0/M, z7.h, z5.h\n"
+      "fmla z24.h, p0/M, z6.h, z5.h\n"
+      "fmla z25.h, p0/M, z2.h, z5.h\n"
       "fmla z26.h, p0/M, z7.h, z0.h\n"
+      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n"
       "fmla z27.h, p0/M, z6.h, z0.h\n"
       "fmla z28.h, p0/M, z2.h, z0.h\n"
+      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n"
       "fmla z29.h, p0/M, z7.h, z1.h\n"
       "ld1h { z0.h }, p0/Z, [x24]\n"
       "fmla z30.h, p0/M, z6.h, z1.h\n"
       "fmla z31.h, p0/M, z2.h, z1.h\n"
       "ld1h { z1.h }, p0/Z, [x22]\n"
       "ld1h { z2.h }, p0/Z, [x21]\n"
-      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n"
       "ld1rh { z6.h }, p0/Z, [%x[Apanel], #6]\n"
       "bge 4b\n"
       "5:"  // main loop skip
@@ -204,12 +204,12 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx(
       "fmla z18.h, p0/M, z1.h, z6.h\n"
       "fmla z19.h, p0/M, z2.h, z6.h\n"
       "ld1rh { z3.h }, p0/Z, [%x[Apanel], #14]\n"
-      "addvl x21, x21, #1\n"
       "fmla z20.h, p0/M, z0.h, z7.h\n"
       "fmla z21.h, p0/M, z1.h, z7.h\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
+      "addvl x21, x21, #1\n"
       "fmla z22.h, p0/M, z2.h, z7.h\n"
       "fmla z23.h, p0/M, z0.h, z4.h\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       "fmla z24.h, p0/M, z1.h, z4.h\n"
       "fmla z25.h, p0/M, z2.h, z4.h\n"
       "fmla z26.h, p0/M, z0.h, z5.h\n"
@@ -223,19 +223,19 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx(
       "ld1h { z5.h }, p0/Z, [x22]\n"
       "ld1h { z4.h }, p0/Z, [x21]\n"
       "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n"
+      "fmla z8.h, p0/M, z6.h, z3.h\n"
       "ld1rh { z2.h }, p0/Z, [%x[Apanel], #2]\n"
       "ld1rh { z1.h }, p0/Z, [%x[Apanel], #4]\n"
-      "ld1rh { z0.h }, p0/Z, [%x[Apanel], #6]\n"
-      "fmla z8.h, p0/M, z6.h, z3.h\n"
       "fmla z9.h, p0/M, z5.h, z3.h\n"
+      "ld1rh { z0.h }, p0/Z, [%x[Apanel], #6]\n"
       "fmla z10.h, p0/M, z4.h, z3.h\n"
       "fmla z11.h, p0/M, z6.h, z2.h\n"
-      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #8]\n"
       "fmla z12.h, p0/M, z5.h, z2.h\n"
       "fmla z13.h, p0/M, z4.h, z2.h\n"
-      "ld1rh { z2.h }, p0/Z, [%x[Apanel], #10]\n"
+      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #8]\n"
       "fmla z14.h, p0/M, z6.h, z1.h\n"
       "fmla z15.h, p0/M, z5.h, z1.h\n"
+      "ld1rh { z2.h }, p0/Z, [%x[Apanel], #10]\n"
       "fmla z16.h, p0/M, z4.h, z1.h\n"
       "fmla z17.h, p0/M, z6.h, z0.h\n"
       "ld1rh { z1.h }, p0/Z, [%x[Apanel], #12]\n"
@@ -258,10 +258,10 @@ void sve_ffinterleaved_fp16_mla_8x3VL_a64fx(
       "6:"  // multiply loop done
       "dech x26, ALL, MUL #3\n"
       "st1h { z8.h }, p0, [%x[Cpanel]]\n"
+      "cmp x26, XZR\n"
       "st1h { z9.h }, p0, [%x[Cpanel], #1, MUL VL]\n"
       "st1h { z10.h }, p0, [%x[Cpanel], #2, MUL VL]\n"
       "st1h { z11.h }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "cmp x26, XZR\n"
       "st1h { z12.h }, p0, [%x[Cpanel], #4, MUL VL]\n"
       "st1h { z13.h }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "st1h { z14.h }, p0, [%x[Cpanel], #6, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/generic.cpp
index 204bfdd658..bc23dc28b0 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp16_mla_8x3VL/generic.cpp
@@ -54,43 +54,43 @@ void sve_ffinterleaved_fp16_mla_8x3VL(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x26, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x25, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x25, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x24, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
       "cnth x23, ALL, MUL #2\n"
-      "mov %x[Apanel], x25\n"
       "add x22, x24, x20, LSL #1\n"
-      "cmp x26, x23\n"
       "add x21, x22, x20, LSL #1\n"
       "add x20, x21, x20, LSL #1\n"
+      "cmp x26, x23\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov %x[Apanel], x25\n"
       "bgt 3f\n"
       "dech x23\n"
-      "mov x21, x24\n"
       "cmp x26, x23\n"
+      "mov x21, x24\n"
       "bgt 3f\n"
       "mov x22, x24\n"
       "3:"  // B setup done
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
       "mov z10.b, #0x0\n"
-      "mov z11.b, #0x0\n"
       "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
+      "mov z11.b, #0x0\n"
       "mov z12.b, #0x0\n"
-      "mov z13.b, #0x0\n"
       "ld1h { z2.h }, p0/Z, [x24]\n"
-      "cmp x20, #0x2\n"
+      "mov z13.b, #0x0\n"
       "mov z14.b, #0x0\n"
+      "ld1h { z3.h }, p0/Z, [x22]\n"
       "mov z15.b, #0x0\n"
       "mov z16.b, #0x0\n"
+      "ld1h { z4.h }, p0/Z, [x21]\n"
       "mov z17.b, #0x0\n"
-      "ld1h { z3.h }, p0/Z, [x22]\n"
       "mov z18.b, #0x0\n"
       "mov z19.b, #0x0\n"
-      "ld1h { z4.h }, p0/Z, [x21]\n"
       "mov z20.b, #0x0\n"
       "mov z21.b, #0x0\n"
       "mov z22.b, #0x0\n"
@@ -169,18 +169,18 @@ void sve_ffinterleaved_fp16_mla_8x3VL(
       "fmla z31.h, z1.h, z7.h[7]\n"
       "bge 4b\n"
       "5:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      "addvl x24, x24, #1\n"
       "fmla z8.h, z2.h, z0.h[0]\n"
-      "addvl x22, x22, #1\n"
-      "addvl x21, x21, #1\n"
       "fmla z11.h, z2.h, z0.h[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       "fmla z14.h, z2.h, z0.h[2]\n"
       "fmla z17.h, z2.h, z0.h[3]\n"
+      "addvl x24, x24, #1\n"
       "fmla z20.h, z2.h, z0.h[4]\n"
       "fmla z23.h, z2.h, z0.h[5]\n"
+      "addvl x22, x22, #1\n"
       "fmla z26.h, z2.h, z0.h[6]\n"
       "fmla z29.h, z2.h, z0.h[7]\n"
+      "addvl x21, x21, #1\n"
       "fmla z9.h, z3.h, z0.h[0]\n"
       "fmla z12.h, z3.h, z0.h[1]\n"
       "fmla z15.h, z3.h, z0.h[2]\n"
@@ -200,13 +200,13 @@ void sve_ffinterleaved_fp16_mla_8x3VL(
       "cbz x20, 6f\n"
       "ld1rqh { z3.h }, p0/Z, [%x[Apanel]]\n"
       "ld1h { z2.h }, p0/Z, [x24]\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
+      "fmla z8.h, z2.h, z3.h[0]\n"
       "ld1h { z1.h }, p0/Z, [x22]\n"
       "ld1h { z0.h }, p0/Z, [x21]\n"
-      "fmla z8.h, z2.h, z3.h[0]\n"
       "fmla z11.h, z2.h, z3.h[1]\n"
       "fmla z14.h, z2.h, z3.h[2]\n"
       "fmla z17.h, z2.h, z3.h[3]\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       "fmla z20.h, z2.h, z3.h[4]\n"
       "fmla z23.h, z2.h, z3.h[5]\n"
       "fmla z26.h, z2.h, z3.h[6]\n"
@@ -230,10 +230,10 @@ void sve_ffinterleaved_fp16_mla_8x3VL(
       "6:"  // multiply loop done
       "dech x26, ALL, MUL #3\n"
       "st1h { z8.h }, p0, [%x[Cpanel]]\n"
+      "cmp x26, XZR\n"
       "st1h { z9.h }, p0, [%x[Cpanel], #1, MUL VL]\n"
       "st1h { z10.h }, p0, [%x[Cpanel], #2, MUL VL]\n"
       "st1h { z11.h }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "cmp x26, XZR\n"
       "st1h { z12.h }, p0, [%x[Cpanel], #4, MUL VL]\n"
       "st1h { z13.h }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "st1h { z14.h }, p0, [%x[Cpanel], #6, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL.hpp
index ad52e2a9b3..21e811497a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL.hpp
@@ -41,8 +41,7 @@ void sve_ffinterleaved_fp32_mla_8x3VL_a64fx( ARGLIST );
 class cls_sve_ffinterleaved_fp32_mla_8x3VL
 {
 public:
-    typedef float lhs_operand_type;
-    typedef float rhs_operand_type;
+    typedef float operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -73,8 +72,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 1, 1> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/a64fx.cpp
index 6135cd4bae..d67c01a574 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/a64fx.cpp
@@ -54,52 +54,52 @@ void sve_ffinterleaved_fp32_mla_8x3VL_a64fx(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x26, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x25, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x25, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x24, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
       "cntw x23, ALL, MUL #2\n"
-      "mov %x[Apanel], x25\n"
       "add x22, x24, x20, LSL #2\n"
-      "cmp x26, x23\n"
       "add x21, x22, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "cmp x26, x23\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov %x[Apanel], x25\n"
       "bgt 3f\n"
       "decw x23\n"
-      "mov x21, x24\n"
       "cmp x26, x23\n"
+      "mov x21, x24\n"
       "bgt 3f\n"
       "mov x22, x24\n"
       "3:"  // B setup done
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
       "mov z10.b, #0x0\n"
-      "mov z11.b, #0x0\n"
       "ld1w { z0.s }, p0/Z, [x24]\n"
+      "mov z11.b, #0x0\n"
       "mov z12.b, #0x0\n"
-      "mov z13.b, #0x0\n"
       "ld1w { z1.s }, p0/Z, [x22]\n"
-      "cmp x20, #0x2\n"
+      "mov z13.b, #0x0\n"
       "mov z14.b, #0x0\n"
+      "ld1w { z2.s }, p0/Z, [x21]\n"
       "mov z15.b, #0x0\n"
       "mov z16.b, #0x0\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
       "mov z17.b, #0x0\n"
-      "ld1w { z2.s }, p0/Z, [x21]\n"
       "mov z18.b, #0x0\n"
+      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "mov z19.b, #0x0\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
       "mov z20.b, #0x0\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "mov z21.b, #0x0\n"
-      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "mov z22.b, #0x0\n"
+      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "mov z23.b, #0x0\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "mov z24.b, #0x0\n"
       "mov z25.b, #0x0\n"
-      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "mov z26.b, #0x0\n"
       "mov z27.b, #0x0\n"
       "mov z28.b, #0x0\n"
@@ -116,12 +116,12 @@ void sve_ffinterleaved_fp32_mla_8x3VL_a64fx(
       "fmla z11.s, p0/M, z0.s, z4.s\n"
       "fmla z12.s, p0/M, z1.s, z4.s\n"
       "fmla z13.s, p0/M, z2.s, z4.s\n"
-      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #20]\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #20]\n"
       "fmla z14.s, p0/M, z0.s, z5.s\n"
       "fmla z15.s, p0/M, z1.s, z5.s\n"
       "cmp x20, #0x2\n"
       "fmla z16.s, p0/M, z2.s, z5.s\n"
-      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #24]\n"
+      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #24]\n"
       "fmla z17.s, p0/M, z0.s, z6.s\n"
       "fmla z18.s, p0/M, z1.s, z6.s\n"
       "fmla z19.s, p0/M, z2.s, z6.s\n"
@@ -129,60 +129,60 @@ void sve_ffinterleaved_fp32_mla_8x3VL_a64fx(
       "fmla z20.s, p0/M, z0.s, z3.s\n"
       "fmla z21.s, p0/M, z1.s, z3.s\n"
       "fmla z22.s, p0/M, z2.s, z3.s\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #32]\n"
-      "fmla z23.s, p0/M, z0.s, z4.s\n"
-      "fmla z24.s, p0/M, z1.s, z4.s\n"
-      "fmla z25.s, p0/M, z2.s, z4.s\n"
-      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #36]\n"
-      "fmla z26.s, p0/M, z0.s, z7.s\n"
-      "fmla z27.s, p0/M, z1.s, z7.s\n"
-      "fmla z28.s, p0/M, z2.s, z7.s\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #40]\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #32]\n"
+      "fmla z23.s, p0/M, z0.s, z7.s\n"
+      "fmla z24.s, p0/M, z1.s, z7.s\n"
+      "fmla z25.s, p0/M, z2.s, z7.s\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #36]\n"
+      "fmla z26.s, p0/M, z0.s, z4.s\n"
+      "fmla z27.s, p0/M, z1.s, z4.s\n"
+      "fmla z28.s, p0/M, z2.s, z4.s\n"
+      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #40]\n"
       "fmla z29.s, p0/M, z0.s, z6.s\n"
       "ld1w { z7.s }, p0/Z, [x24, #1, MUL VL]\n"
       "fmla z30.s, p0/M, z1.s, z6.s\n"
       "fmla z31.s, p0/M, z2.s, z6.s\n"
       "ld1w { z6.s }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1w { z2.s }, p0/Z, [x21, #1, MUL VL]\n"
-      "addvl x24, x24, #2\n"
+      "fmla z8.s, p0/M, z7.s, z3.s\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #44]\n"
-      "addvl x22, x22, #2\n"
-      "addvl x21, x21, #2\n"
-      "fmla z8.s, p0/M, z7.s, z5.s\n"
-      "fmla z11.s, p0/M, z7.s, z4.s\n"
-      "fmla z9.s, p0/M, z6.s, z5.s\n"
-      "fmla z12.s, p0/M, z6.s, z4.s\n"
-      "fmla z10.s, p0/M, z2.s, z5.s\n"
-      "fmla z13.s, p0/M, z2.s, z4.s\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #48]\n"
-      "fmla z14.s, p0/M, z7.s, z3.s\n"
-      "fmla z15.s, p0/M, z6.s, z3.s\n"
-      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #52]\n"
-      "fmla z16.s, p0/M, z2.s, z3.s\n"
-      "fmla z17.s, p0/M, z7.s, z1.s\n"
+      "fmla z9.s, p0/M, z6.s, z3.s\n"
+      "fmla z10.s, p0/M, z2.s, z3.s\n"
+      "fmla z11.s, p0/M, z7.s, z5.s\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #48]\n"
+      "fmla z12.s, p0/M, z6.s, z5.s\n"
+      "fmla z13.s, p0/M, z2.s, z5.s\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #52]\n"
+      "fmla z14.s, p0/M, z7.s, z4.s\n"
+      "fmla z15.s, p0/M, z6.s, z4.s\n"
+      "addvl x24, x24, #2\n"
+      "fmla z16.s, p0/M, z2.s, z4.s\n"
       "ld1rw { z0.s }, p0/Z, [%x[Apanel], #56]\n"
+      "fmla z17.s, p0/M, z7.s, z1.s\n"
       "fmla z18.s, p0/M, z6.s, z1.s\n"
       "fmla z19.s, p0/M, z2.s, z1.s\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #60]\n"
+      "addvl x22, x22, #2\n"
+      "addvl x21, x21, #2\n"
       "add %x[Apanel], %x[Apanel], #0x40\n"
-      "fmla z20.s, p0/M, z7.s, z5.s\n"
-      "fmla z21.s, p0/M, z6.s, z5.s\n"
-      "fmla z22.s, p0/M, z2.s, z5.s\n"
-      "fmla z23.s, p0/M, z7.s, z4.s\n"
+      "fmla z20.s, p0/M, z7.s, z3.s\n"
+      "fmla z21.s, p0/M, z6.s, z3.s\n"
+      "fmla z22.s, p0/M, z2.s, z3.s\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
-      "fmla z24.s, p0/M, z6.s, z4.s\n"
-      "fmla z25.s, p0/M, z2.s, z4.s\n"
-      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
+      "fmla z23.s, p0/M, z7.s, z5.s\n"
+      "fmla z24.s, p0/M, z6.s, z5.s\n"
+      "fmla z25.s, p0/M, z2.s, z5.s\n"
       "fmla z26.s, p0/M, z7.s, z0.s\n"
+      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "fmla z27.s, p0/M, z6.s, z0.s\n"
       "fmla z28.s, p0/M, z2.s, z0.s\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "fmla z29.s, p0/M, z7.s, z1.s\n"
       "ld1w { z0.s }, p0/Z, [x24]\n"
       "fmla z30.s, p0/M, z6.s, z1.s\n"
       "fmla z31.s, p0/M, z2.s, z1.s\n"
       "ld1w { z1.s }, p0/Z, [x22]\n"
       "ld1w { z2.s }, p0/Z, [x21]\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "bge 4b\n"
       "5:"  // main loop skip
@@ -204,12 +204,12 @@ void sve_ffinterleaved_fp32_mla_8x3VL_a64fx(
       "fmla z18.s, p0/M, z1.s, z6.s\n"
       "fmla z19.s, p0/M, z2.s, z6.s\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel], #28]\n"
-      "addvl x21, x21, #1\n"
       "fmla z20.s, p0/M, z0.s, z7.s\n"
       "fmla z21.s, p0/M, z1.s, z7.s\n"
-      "add %x[Apanel], %x[Apanel], #0x20\n"
+      "addvl x21, x21, #1\n"
       "fmla z22.s, p0/M, z2.s, z7.s\n"
       "fmla z23.s, p0/M, z0.s, z4.s\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "fmla z24.s, p0/M, z1.s, z4.s\n"
       "fmla z25.s, p0/M, z2.s, z4.s\n"
       "fmla z26.s, p0/M, z0.s, z5.s\n"
@@ -223,19 +223,19 @@ void sve_ffinterleaved_fp32_mla_8x3VL_a64fx(
       "ld1w { z5.s }, p0/Z, [x22]\n"
       "ld1w { z4.s }, p0/Z, [x21]\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
+      "fmla z8.s, p0/M, z6.s, z3.s\n"
       "ld1rw { z2.s }, p0/Z, [%x[Apanel], #4]\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #8]\n"
-      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
-      "fmla z8.s, p0/M, z6.s, z3.s\n"
       "fmla z9.s, p0/M, z5.s, z3.s\n"
+      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
       "fmla z10.s, p0/M, z4.s, z3.s\n"
       "fmla z11.s, p0/M, z6.s, z2.s\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
       "fmla z12.s, p0/M, z5.s, z2.s\n"
       "fmla z13.s, p0/M, z4.s, z2.s\n"
-      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
       "fmla z14.s, p0/M, z6.s, z1.s\n"
       "fmla z15.s, p0/M, z5.s, z1.s\n"
+      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
       "fmla z16.s, p0/M, z4.s, z1.s\n"
       "fmla z17.s, p0/M, z6.s, z0.s\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #24]\n"
@@ -258,10 +258,10 @@ void sve_ffinterleaved_fp32_mla_8x3VL_a64fx(
       "6:"  // multiply loop done
       "decw x26, ALL, MUL #3\n"
       "st1w { z8.s }, p0, [%x[Cpanel]]\n"
+      "cmp x26, XZR\n"
       "st1w { z9.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
       "st1w { z10.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
       "st1w { z11.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "cmp x26, XZR\n"
       "st1w { z12.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
       "st1w { z13.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "st1w { z14.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/generic.cpp
index 05262b50cb..5f29a6ce3c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffinterleaved_fp32_mla_8x3VL/generic.cpp
@@ -54,46 +54,46 @@ void sve_ffinterleaved_fp32_mla_8x3VL(
       "1:"  // Height loop
       "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n"
       "ldr x26, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov x25, %x[Apanel]\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov x25, %x[Apanel]\n"
       "2:"  // Width loop
       "ldr x24, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
       "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n"
       "cntw x23, ALL, MUL #2\n"
-      "mov %x[Apanel], x25\n"
       "add x22, x24, x20, LSL #2\n"
-      "cmp x26, x23\n"
       "add x21, x22, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "cmp x26, x23\n"
       "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
+      "mov %x[Apanel], x25\n"
       "bgt 3f\n"
       "decw x23\n"
-      "mov x21, x24\n"
       "cmp x26, x23\n"
+      "mov x21, x24\n"
       "bgt 3f\n"
       "mov x22, x24\n"
       "3:"  // B setup done
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
       "mov z10.b, #0x0\n"
-      "mov z11.b, #0x0\n"
       "ld1rqw { z0.s }, p0/Z, [%x[Apanel]]\n"
+      "mov z11.b, #0x0\n"
       "mov z12.b, #0x0\n"
-      "mov z13.b, #0x0\n"
       "ld1rqw { z1.s }, p0/Z, [%x[Apanel], #16]\n"
-      "cmp x20, #0x2\n"
+      "mov z13.b, #0x0\n"
       "mov z14.b, #0x0\n"
+      "ld1w { z4.s }, p0/Z, [x24]\n"
       "mov z15.b, #0x0\n"
       "mov z16.b, #0x0\n"
+      "ld1w { z5.s }, p0/Z, [x22]\n"
       "mov z17.b, #0x0\n"
-      "ld1w { z4.s }, p0/Z, [x24]\n"
       "mov z18.b, #0x0\n"
+      "ld1w { z6.s }, p0/Z, [x21]\n"
       "mov z19.b, #0x0\n"
-      "ld1w { z5.s }, p0/Z, [x22]\n"
       "mov z20.b, #0x0\n"
       "mov z21.b, #0x0\n"
-      "ld1w { z6.s }, p0/Z, [x21]\n"
       "mov z22.b, #0x0\n"
       "mov z23.b, #0x0\n"
       "mov z24.b, #0x0\n"
@@ -172,18 +172,18 @@ void sve_ffinterleaved_fp32_mla_8x3VL(
       "ld1w { z6.s }, p0/Z, [x21]\n"
       "bge 4b\n"
       "5:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "addvl x24, x24, #1\n"
       "fmla z8.s, z4.s, z0.s[0]\n"
-      "addvl x22, x22, #1\n"
-      "addvl x21, x21, #1\n"
       "fmla z11.s, z4.s, z0.s[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "fmla z14.s, z4.s, z0.s[2]\n"
       "fmla z17.s, z4.s, z0.s[3]\n"
+      "addvl x24, x24, #1\n"
       "fmla z20.s, z4.s, z1.s[0]\n"
       "fmla z23.s, z4.s, z1.s[1]\n"
+      "addvl x22, x22, #1\n"
       "fmla z26.s, z4.s, z1.s[2]\n"
       "fmla z29.s, z4.s, z1.s[3]\n"
+      "addvl x21, x21, #1\n"
       "fmla z9.s, z5.s, z0.s[0]\n"
       "fmla z12.s, z5.s, z0.s[1]\n"
       "fmla z15.s, z5.s, z0.s[2]\n"
@@ -206,8 +206,8 @@ void sve_ffinterleaved_fp32_mla_8x3VL(
       "add %x[Apanel], %x[Apanel], #0x20\n"
       "ld1w { z2.s }, p0/Z, [x24]\n"
       "ld1w { z1.s }, p0/Z, [x22]\n"
-      "ld1w { z0.s }, p0/Z, [x21]\n"
       "fmla z8.s, z2.s, z4.s[0]\n"
+      "ld1w { z0.s }, p0/Z, [x21]\n"
       "fmla z11.s, z2.s, z4.s[1]\n"
       "fmla z14.s, z2.s, z4.s[2]\n"
       "fmla z17.s, z2.s, z4.s[3]\n"
@@ -234,10 +234,10 @@ void sve_ffinterleaved_fp32_mla_8x3VL(
       "6:"  // multiply loop done
       "decw x26, ALL, MUL #3\n"
       "st1w { z8.s }, p0, [%x[Cpanel]]\n"
+      "cmp x26, XZR\n"
       "st1w { z9.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
       "st1w { z10.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
       "st1w { z11.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "cmp x26, XZR\n"
       "st1w { z12.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
       "st1w { z13.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "st1w { z14.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL.hpp
index d1f3c31de3..4de4101148 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 4, 2> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 2> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL/generic.cpp
index 739ee24050..688e7377b9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_dot_6x4VL/generic.cpp
@@ -48,19 +48,18 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -81,7 +80,6 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -105,10 +103,10 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cmp %x[M], #0x2\n"
       "bgt 27f\n"
       "beq 14f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -141,8 +139,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "mov x28, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -158,98 +156,98 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      ".inst 0x64604228  // bfdot z8.s, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10]\n"
+      ".inst 0x64604208  // bfdot z8.s, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64604209  // bfdot z9.s, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #2, MUL VL]\n"
+      ".inst 0x6460420a  // bfdot z10.s, z16.h, z0.h[0]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6460422a  // bfdot z10.s, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6460420b  // bfdot z11.s, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #4, MUL VL]\n"
+      ".inst 0x64684208  // bfdot z8.s, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x64684228  // bfdot z8.s, z17.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x64684209  // bfdot z9.s, z16.h, z0.h[1]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #6, MUL VL]\n"
+      ".inst 0x6468420a  // bfdot z10.s, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x6468422a  // bfdot z10.s, z17.h, z0.h[1]\n"
       ".inst 0x6468420b  // bfdot z11.s, z16.h, z0.h[1]\n"
       "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-7, MUL VL]\n"
       ".inst 0x64704228  // bfdot z8.s, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x64704209  // bfdot z9.s, z16.h, z0.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-5, MUL VL]\n"
       ".inst 0x6470422a  // bfdot z10.s, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x6470420b  // bfdot z11.s, z16.h, z0.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-3, MUL VL]\n"
       ".inst 0x64784228  // bfdot z8.s, z17.h, z0.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x64784209  // bfdot z9.s, z16.h, z0.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-1, MUL VL]\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6478422a  // bfdot z10.s, z17.h, z0.h[3]\n"
       ".inst 0x6478420b  // bfdot z11.s, z16.h, z0.h[3]\n"
+      "add x26, x26, #0x10\n"
       "bgt 9b\n"
       "10:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
-      ".inst 0x64604228  // bfdot z8.s, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10]\n"
+      ".inst 0x64604208  // bfdot z8.s, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64604209  // bfdot z9.s, z16.h, z0.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x2\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       ".inst 0x6460422a  // bfdot z10.s, z17.h, z0.h[0]\n"
       ".inst 0x6460420b  // bfdot z11.s, z16.h, z0.h[0]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x64684228  // bfdot z8.s, z17.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64684209  // bfdot z9.s, z16.h, z0.h[1]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x6468422a  // bfdot z10.s, z17.h, z0.h[1]\n"
       ".inst 0x6468420b  // bfdot z11.s, z16.h, z0.h[1]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x64704228  // bfdot z8.s, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64704209  // bfdot z9.s, z16.h, z0.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x6470422a  // bfdot z10.s, z17.h, z0.h[2]\n"
       ".inst 0x6470420b  // bfdot z11.s, z16.h, z0.h[2]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64784228  // bfdot z8.s, z17.h, z0.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64784209  // bfdot z9.s, z16.h, z0.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       ".inst 0x6478422a  // bfdot z10.s, z17.h, z0.h[3]\n"
       ".inst 0x6478420b  // bfdot z11.s, z16.h, z0.h[3]\n"
+      "addvl x10, x10, #4\n"
       "11:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 6b\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z17.s\n"
       "fmin z9.s, p5/M, z9.s, z17.s\n"
@@ -271,10 +269,10 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "bgt 2b\n"
       "b 80f\n"
       "14:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -287,22 +285,22 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cbz x12, 16f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x12, x12, #4\n"
       "b 18f\n"
       "16:"  // Height 2: no bias
       "tbz %x[flags], #0, 17f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
@@ -321,8 +319,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "mov x28, #0x0\n"
       "19:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -341,38 +339,38 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 23f\n"
       "22:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64614228  // bfdot z8.s, z17.h, z1.h[0]\n"
       ".inst 0x6460422c  // bfdot z12.s, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64614209  // bfdot z9.s, z16.h, z1.h[0]\n"
       ".inst 0x6460420d  // bfdot z13.s, z16.h, z0.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6461422a  // bfdot z10.s, z17.h, z1.h[0]\n"
       ".inst 0x6460422e  // bfdot z14.s, z17.h, z0.h[0]\n"
       "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6461420b  // bfdot z11.s, z16.h, z1.h[0]\n"
       ".inst 0x6460420f  // bfdot z15.s, z16.h, z0.h[0]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x64694228  // bfdot z8.s, z17.h, z1.h[1]\n"
       ".inst 0x6468422c  // bfdot z12.s, z17.h, z0.h[1]\n"
       "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x64694209  // bfdot z9.s, z16.h, z1.h[1]\n"
       ".inst 0x6468420d  // bfdot z13.s, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
       ".inst 0x6469422a  // bfdot z10.s, z17.h, z1.h[1]\n"
       ".inst 0x6468422e  // bfdot z14.s, z17.h, z0.h[1]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       ".inst 0x6469420b  // bfdot z11.s, z16.h, z1.h[1]\n"
       ".inst 0x6468420f  // bfdot z15.s, z16.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-7, MUL VL]\n"
       ".inst 0x64714228  // bfdot z8.s, z17.h, z1.h[2]\n"
       ".inst 0x6470422c  // bfdot z12.s, z17.h, z0.h[2]\n"
@@ -399,50 +397,50 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "bgt 22b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x2\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64604228  // bfdot z8.s, z17.h, z0.h[0]\n"
       ".inst 0x6461422c  // bfdot z12.s, z17.h, z1.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64604209  // bfdot z9.s, z16.h, z0.h[0]\n"
       ".inst 0x6461420d  // bfdot z13.s, z16.h, z1.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       ".inst 0x6460422a  // bfdot z10.s, z17.h, z0.h[0]\n"
       ".inst 0x6461422e  // bfdot z14.s, z17.h, z1.h[0]\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x6460420b  // bfdot z11.s, z16.h, z0.h[0]\n"
       ".inst 0x6461420f  // bfdot z15.s, z16.h, z1.h[0]\n"
       "ble 24f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x64684228  // bfdot z8.s, z17.h, z0.h[1]\n"
       ".inst 0x6469422c  // bfdot z12.s, z17.h, z1.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64684209  // bfdot z9.s, z16.h, z0.h[1]\n"
       ".inst 0x6469420d  // bfdot z13.s, z16.h, z1.h[1]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x6468422a  // bfdot z10.s, z17.h, z0.h[1]\n"
       ".inst 0x6469422e  // bfdot z14.s, z17.h, z1.h[1]\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x6468420b  // bfdot z11.s, z16.h, z0.h[1]\n"
       ".inst 0x6469420f  // bfdot z15.s, z16.h, z1.h[1]\n"
       "ble 24f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x64704228  // bfdot z8.s, z17.h, z0.h[2]\n"
       ".inst 0x6471422c  // bfdot z12.s, z17.h, z1.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64704209  // bfdot z9.s, z16.h, z0.h[2]\n"
       ".inst 0x6471420d  // bfdot z13.s, z16.h, z1.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x6470422a  // bfdot z10.s, z17.h, z0.h[2]\n"
       ".inst 0x6471422e  // bfdot z14.s, z17.h, z1.h[2]\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x6470420b  // bfdot z11.s, z16.h, z0.h[2]\n"
       ".inst 0x6471420f  // bfdot z15.s, z16.h, z1.h[2]\n"
       "ble 24f\n"
@@ -450,13 +448,13 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64784228  // bfdot z8.s, z17.h, z0.h[3]\n"
       ".inst 0x6479422c  // bfdot z12.s, z17.h, z1.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64784209  // bfdot z9.s, z16.h, z0.h[3]\n"
       ".inst 0x6479420d  // bfdot z13.s, z16.h, z1.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       ".inst 0x6478422a  // bfdot z10.s, z17.h, z0.h[3]\n"
       ".inst 0x6479422e  // bfdot z14.s, z17.h, z1.h[3]\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x6478420b  // bfdot z11.s, z16.h, z0.h[3]\n"
       ".inst 0x6479420f  // bfdot z15.s, z16.h, z1.h[3]\n"
       "24:"  // Height 2: Multiply loop: multiply skip
@@ -465,11 +463,11 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 19b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z17.s\n"
       "fmin z9.s, p5/M, z9.s, z17.s\n"
@@ -493,20 +491,20 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
       "26:"  // Height 2: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 15b\n"
       "b 80f\n"
       "27:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -519,27 +517,27 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cbz x12, 29f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 31f\n"
       "29:"  // Height 3: no bias
       "tbz %x[flags], #0, 30f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
@@ -566,8 +564,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "mov x28, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -589,37 +587,37 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 36f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z2.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z0.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
       ".inst 0x646242a8  // bfdot z8.s, z21.h, z2.h[0]\n"
       ".inst 0x646142ac  // bfdot z12.s, z21.h, z1.h[0]\n"
-      ".inst 0x64624289  // bfdot z9.s, z20.h, z2.h[0]\n"
-      ".inst 0x6461428d  // bfdot z13.s, z20.h, z1.h[0]\n"
+      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x646042b0  // bfdot z16.s, z21.h, z0.h[0]\n"
+      ".inst 0x64624289  // bfdot z9.s, z20.h, z2.h[0]\n"
       "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      ".inst 0x6461428d  // bfdot z13.s, z20.h, z1.h[0]\n"
       ".inst 0x64604291  // bfdot z17.s, z20.h, z0.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
+      "cmp x27, #0x8\n"
       ".inst 0x646242aa  // bfdot z10.s, z21.h, z2.h[0]\n"
       ".inst 0x646142ae  // bfdot z14.s, z21.h, z1.h[0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x646042b2  // bfdot z18.s, z21.h, z0.h[0]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6462428b  // bfdot z11.s, z20.h, z2.h[0]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6461428f  // bfdot z15.s, z20.h, z1.h[0]\n"
       ".inst 0x64604293  // bfdot z19.s, z20.h, z0.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x646a42a8  // bfdot z8.s, z21.h, z2.h[1]\n"
       ".inst 0x646942ac  // bfdot z12.s, z21.h, z1.h[1]\n"
       ".inst 0x646842b0  // bfdot z16.s, z21.h, z0.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x646a4289  // bfdot z9.s, z20.h, z2.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6469428d  // bfdot z13.s, z20.h, z1.h[1]\n"
       ".inst 0x64684291  // bfdot z17.s, z20.h, z0.h[1]\n"
       "ld1h { z20.h }, p5/Z, [x10, #7, MUL VL]\n"
@@ -628,31 +626,31 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x646942ae  // bfdot z14.s, z21.h, z1.h[1]\n"
       ".inst 0x646842b2  // bfdot z18.s, z21.h, z0.h[1]\n"
       ".inst 0x646a428b  // bfdot z11.s, z20.h, z2.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-8, MUL VL]\n"
       ".inst 0x6469428f  // bfdot z15.s, z20.h, z1.h[1]\n"
       ".inst 0x64684293  // bfdot z19.s, z20.h, z0.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-7, MUL VL]\n"
       ".inst 0x647242a8  // bfdot z8.s, z21.h, z2.h[2]\n"
       ".inst 0x647142ac  // bfdot z12.s, z21.h, z1.h[2]\n"
       ".inst 0x647042b0  // bfdot z16.s, z21.h, z0.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x64724289  // bfdot z9.s, z20.h, z2.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x6471428d  // bfdot z13.s, z20.h, z1.h[2]\n"
       ".inst 0x64704291  // bfdot z17.s, z20.h, z0.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-5, MUL VL]\n"
       ".inst 0x647242aa  // bfdot z10.s, z21.h, z2.h[2]\n"
       ".inst 0x647142ae  // bfdot z14.s, z21.h, z1.h[2]\n"
       ".inst 0x647042b2  // bfdot z18.s, z21.h, z0.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x6472428b  // bfdot z11.s, z20.h, z2.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x6471428f  // bfdot z15.s, z20.h, z1.h[2]\n"
       ".inst 0x64704293  // bfdot z19.s, z20.h, z0.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-3, MUL VL]\n"
       ".inst 0x647a42a8  // bfdot z8.s, z21.h, z2.h[3]\n"
       ".inst 0x647942ac  // bfdot z12.s, z21.h, z1.h[3]\n"
       ".inst 0x647842b0  // bfdot z16.s, z21.h, z0.h[3]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x647a4289  // bfdot z9.s, z20.h, z2.h[3]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x6479428d  // bfdot z13.s, z20.h, z1.h[3]\n"
       ".inst 0x64784291  // bfdot z17.s, z20.h, z0.h[3]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-1, MUL VL]\n"
@@ -665,18 +663,18 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "bgt 35b\n"
       "36:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x2\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
       ".inst 0x646042a8  // bfdot z8.s, z21.h, z0.h[0]\n"
       ".inst 0x646142ac  // bfdot z12.s, z21.h, z1.h[0]\n"
-      ".inst 0x64604289  // bfdot z9.s, z20.h, z0.h[0]\n"
-      ".inst 0x6461428d  // bfdot z13.s, z20.h, z1.h[0]\n"
+      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x646242b0  // bfdot z16.s, z21.h, z2.h[0]\n"
+      ".inst 0x64604289  // bfdot z9.s, z20.h, z0.h[0]\n"
       "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      ".inst 0x6461428d  // bfdot z13.s, z20.h, z1.h[0]\n"
       ".inst 0x64624291  // bfdot z17.s, z20.h, z2.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
@@ -689,12 +687,12 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 37f\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x646842a8  // bfdot z8.s, z21.h, z0.h[1]\n"
       ".inst 0x646942ac  // bfdot z12.s, z21.h, z1.h[1]\n"
       ".inst 0x646a42b0  // bfdot z16.s, z21.h, z2.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64684289  // bfdot z9.s, z20.h, z0.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x6469428d  // bfdot z13.s, z20.h, z1.h[1]\n"
       ".inst 0x646a4291  // bfdot z17.s, z20.h, z2.h[1]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -708,12 +706,12 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 37f\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x647042a8  // bfdot z8.s, z21.h, z0.h[2]\n"
       ".inst 0x647142ac  // bfdot z12.s, z21.h, z1.h[2]\n"
       ".inst 0x647242b0  // bfdot z16.s, z21.h, z2.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64704289  // bfdot z9.s, z20.h, z0.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x6471428d  // bfdot z13.s, z20.h, z1.h[2]\n"
       ".inst 0x64724291  // bfdot z17.s, z20.h, z2.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -730,8 +728,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x647842a8  // bfdot z8.s, z21.h, z0.h[3]\n"
       ".inst 0x647942ac  // bfdot z12.s, z21.h, z1.h[3]\n"
       ".inst 0x647a42b0  // bfdot z16.s, z21.h, z2.h[3]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64784289  // bfdot z9.s, z20.h, z0.h[3]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6479428d  // bfdot z13.s, z20.h, z1.h[3]\n"
       ".inst 0x647a4291  // bfdot z17.s, z20.h, z2.h[3]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -748,12 +746,12 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z21.s }, p5/Z, [x21]\n"
       "ld1rw { z20.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z21.s\n"
       "fmin z9.s, p5/M, z9.s, z21.s\n"
@@ -785,24 +783,24 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
       "39:"  // Height 3: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 28b\n"
       "b 80f\n"
       "40:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -815,18 +813,18 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cbz x12, 42f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -834,13 +832,13 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "42:"  // Height 4: no bias
       "tbz %x[flags], #0, 43f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
@@ -875,8 +873,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "mov x28, #0x0\n"
       "45:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -901,25 +899,25 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 49f\n"
       "48:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x10]\n"
-      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z3.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z2.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z1.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64634328  // bfdot z8.s, z25.h, z3.h[0]\n"
       ".inst 0x6462432c  // bfdot z12.s, z25.h, z2.h[0]\n"
-      ".inst 0x64634309  // bfdot z9.s, z24.h, z3.h[0]\n"
-      ".inst 0x6462430d  // bfdot z13.s, z24.h, z2.h[0]\n"
       ".inst 0x64614330  // bfdot z16.s, z25.h, z1.h[0]\n"
       ".inst 0x64604334  // bfdot z20.s, z25.h, z0.h[0]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x64634309  // bfdot z9.s, z24.h, z3.h[0]\n"
+      ".inst 0x6462430d  // bfdot z13.s, z24.h, z2.h[0]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x64614311  // bfdot z17.s, z24.h, z1.h[0]\n"
       ".inst 0x64604315  // bfdot z21.s, z24.h, z0.h[0]\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -948,9 +946,9 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x646a432e  // bfdot z14.s, z25.h, z2.h[1]\n"
       ".inst 0x64694332  // bfdot z18.s, z25.h, z1.h[1]\n"
       ".inst 0x64684336  // bfdot z22.s, z25.h, z0.h[1]\n"
+      "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
       ".inst 0x646b430b  // bfdot z11.s, z24.h, z3.h[1]\n"
       ".inst 0x646a430f  // bfdot z15.s, z24.h, z2.h[1]\n"
-      "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
       ".inst 0x64694313  // bfdot z19.s, z24.h, z1.h[1]\n"
       ".inst 0x64684317  // bfdot z23.s, z24.h, z0.h[1]\n"
       "ld1h { z24.h }, p5/Z, [x10, #-7, MUL VL]\n"
@@ -995,20 +993,20 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "bgt 48b\n"
       "49:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x10]\n"
-      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x2\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64604328  // bfdot z8.s, z25.h, z0.h[0]\n"
       ".inst 0x6461432c  // bfdot z12.s, z25.h, z1.h[0]\n"
-      ".inst 0x64604309  // bfdot z9.s, z24.h, z0.h[0]\n"
-      ".inst 0x6461430d  // bfdot z13.s, z24.h, z1.h[0]\n"
       ".inst 0x64624330  // bfdot z16.s, z25.h, z2.h[0]\n"
       ".inst 0x64634334  // bfdot z20.s, z25.h, z3.h[0]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      ".inst 0x64604309  // bfdot z9.s, z24.h, z0.h[0]\n"
+      ".inst 0x6461430d  // bfdot z13.s, z24.h, z1.h[0]\n"
       ".inst 0x64624311  // bfdot z17.s, z24.h, z2.h[0]\n"
       ".inst 0x64634315  // bfdot z21.s, z24.h, z3.h[0]\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1024,12 +1022,12 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x64684328  // bfdot z8.s, z25.h, z0.h[1]\n"
       ".inst 0x6469432c  // bfdot z12.s, z25.h, z1.h[1]\n"
       ".inst 0x646a4330  // bfdot z16.s, z25.h, z2.h[1]\n"
       ".inst 0x646b4334  // bfdot z20.s, z25.h, z3.h[1]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x64684309  // bfdot z9.s, z24.h, z0.h[1]\n"
       ".inst 0x6469430d  // bfdot z13.s, z24.h, z1.h[1]\n"
       ".inst 0x646a4311  // bfdot z17.s, z24.h, z2.h[1]\n"
@@ -1047,12 +1045,12 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x64704328  // bfdot z8.s, z25.h, z0.h[2]\n"
       ".inst 0x6471432c  // bfdot z12.s, z25.h, z1.h[2]\n"
       ".inst 0x64724330  // bfdot z16.s, z25.h, z2.h[2]\n"
       ".inst 0x64734334  // bfdot z20.s, z25.h, z3.h[2]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x64704309  // bfdot z9.s, z24.h, z0.h[2]\n"
       ".inst 0x6471430d  // bfdot z13.s, z24.h, z1.h[2]\n"
       ".inst 0x64724311  // bfdot z17.s, z24.h, z2.h[2]\n"
@@ -1095,13 +1093,13 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z25.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z25.s }, p5/Z, [x21]\n"
       "ld1rw { z24.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z25.s\n"
       "fmin z9.s, p5/M, z9.s, z25.s\n"
@@ -1141,28 +1139,28 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
       "52:"  // Height 4: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 41b\n"
       "b 80f\n"
       "53:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "54:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1175,18 +1173,18 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cbz x12, 55f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1198,16 +1196,16 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "55:"  // Height 5: no bias
       "tbz %x[flags], #0, 56f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x22]\n"
@@ -1248,8 +1246,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "mov x28, #0x0\n"
       "58:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 59f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1277,29 +1275,29 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 62f\n"
       "61:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z4.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z3.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z1.h }, p0/Z, [x23]\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1h { z29.h }, p5/Z, [x10]\n"
       ".inst 0x646443a8  // bfdot z8.s, z29.h, z4.h[0]\n"
       ".inst 0x646343ac  // bfdot z12.s, z29.h, z3.h[0]\n"
-      ".inst 0x64644389  // bfdot z9.s, z28.h, z4.h[0]\n"
+      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x646243b0  // bfdot z16.s, z29.h, z2.h[0]\n"
       ".inst 0x646143b4  // bfdot z20.s, z29.h, z1.h[0]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x646043b8  // bfdot z24.s, z29.h, z0.h[0]\n"
-      ".inst 0x6463438d  // bfdot z13.s, z28.h, z3.h[0]\n"
+      ".inst 0x64644389  // bfdot z9.s, z28.h, z4.h[0]\n"
       "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6463438d  // bfdot z13.s, z28.h, z3.h[0]\n"
       ".inst 0x64624391  // bfdot z17.s, z28.h, z2.h[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x64614395  // bfdot z21.s, z28.h, z1.h[0]\n"
       ".inst 0x64604399  // bfdot z25.s, z28.h, z0.h[0]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1308,8 +1306,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x646243b2  // bfdot z18.s, z29.h, z2.h[0]\n"
       ".inst 0x646143b6  // bfdot z22.s, z29.h, z1.h[0]\n"
       ".inst 0x646043ba  // bfdot z26.s, z29.h, z0.h[0]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6464438b  // bfdot z11.s, z28.h, z4.h[0]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6463438f  // bfdot z15.s, z28.h, z3.h[0]\n"
       ".inst 0x64624393  // bfdot z19.s, z28.h, z2.h[0]\n"
       ".inst 0x64614397  // bfdot z23.s, z28.h, z1.h[0]\n"
@@ -1320,8 +1318,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x646a43b0  // bfdot z16.s, z29.h, z2.h[1]\n"
       ".inst 0x646943b4  // bfdot z20.s, z29.h, z1.h[1]\n"
       ".inst 0x646843b8  // bfdot z24.s, z29.h, z0.h[1]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x646c4389  // bfdot z9.s, z28.h, z4.h[1]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x646b438d  // bfdot z13.s, z28.h, z3.h[1]\n"
       ".inst 0x646a4391  // bfdot z17.s, z28.h, z2.h[1]\n"
       ".inst 0x64694395  // bfdot z21.s, z28.h, z1.h[1]\n"
@@ -1334,8 +1332,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x646943b6  // bfdot z22.s, z29.h, z1.h[1]\n"
       ".inst 0x646843ba  // bfdot z26.s, z29.h, z0.h[1]\n"
       ".inst 0x646c438b  // bfdot z11.s, z28.h, z4.h[1]\n"
-      ".inst 0x646b438f  // bfdot z15.s, z28.h, z3.h[1]\n"
       "ld1h { z29.h }, p5/Z, [x10, #-8, MUL VL]\n"
+      ".inst 0x646b438f  // bfdot z15.s, z28.h, z3.h[1]\n"
       ".inst 0x646a4393  // bfdot z19.s, z28.h, z2.h[1]\n"
       ".inst 0x64694397  // bfdot z23.s, z28.h, z1.h[1]\n"
       ".inst 0x6468439b  // bfdot z27.s, z28.h, z0.h[1]\n"
@@ -1345,8 +1343,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x647243b0  // bfdot z16.s, z29.h, z2.h[2]\n"
       ".inst 0x647143b4  // bfdot z20.s, z29.h, z1.h[2]\n"
       ".inst 0x647043b8  // bfdot z24.s, z29.h, z0.h[2]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x64744389  // bfdot z9.s, z28.h, z4.h[2]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x6473438d  // bfdot z13.s, z28.h, z3.h[2]\n"
       ".inst 0x64724391  // bfdot z17.s, z28.h, z2.h[2]\n"
       ".inst 0x64714395  // bfdot z21.s, z28.h, z1.h[2]\n"
@@ -1357,8 +1355,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x647243b2  // bfdot z18.s, z29.h, z2.h[2]\n"
       ".inst 0x647143b6  // bfdot z22.s, z29.h, z1.h[2]\n"
       ".inst 0x647043ba  // bfdot z26.s, z29.h, z0.h[2]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x6474438b  // bfdot z11.s, z28.h, z4.h[2]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x6473438f  // bfdot z15.s, z28.h, z3.h[2]\n"
       ".inst 0x64724393  // bfdot z19.s, z28.h, z2.h[2]\n"
       ".inst 0x64714397  // bfdot z23.s, z28.h, z1.h[2]\n"
@@ -1369,8 +1367,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x647a43b0  // bfdot z16.s, z29.h, z2.h[3]\n"
       ".inst 0x647943b4  // bfdot z20.s, z29.h, z1.h[3]\n"
       ".inst 0x647843b8  // bfdot z24.s, z29.h, z0.h[3]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x647c4389  // bfdot z9.s, z28.h, z4.h[3]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x647b438d  // bfdot z13.s, z28.h, z3.h[3]\n"
       ".inst 0x647a4391  // bfdot z17.s, z28.h, z2.h[3]\n"
       ".inst 0x64794395  // bfdot z21.s, z28.h, z1.h[3]\n"
@@ -1389,23 +1387,23 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "bgt 61b\n"
       "62:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x2\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
       "ld1rqh { z4.h }, p0/Z, [x22]\n"
+      "ld1h { z29.h }, p5/Z, [x10]\n"
       ".inst 0x646043a8  // bfdot z8.s, z29.h, z0.h[0]\n"
       ".inst 0x646143ac  // bfdot z12.s, z29.h, z1.h[0]\n"
-      ".inst 0x64604389  // bfdot z9.s, z28.h, z0.h[0]\n"
-      ".inst 0x6461438d  // bfdot z13.s, z28.h, z1.h[0]\n"
+      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x646243b0  // bfdot z16.s, z29.h, z2.h[0]\n"
       ".inst 0x646343b4  // bfdot z20.s, z29.h, z3.h[0]\n"
       ".inst 0x646443b8  // bfdot z24.s, z29.h, z4.h[0]\n"
-      ".inst 0x64624391  // bfdot z17.s, z28.h, z2.h[0]\n"
+      ".inst 0x64604389  // bfdot z9.s, z28.h, z0.h[0]\n"
       "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
+      ".inst 0x6461438d  // bfdot z13.s, z28.h, z1.h[0]\n"
+      ".inst 0x64624391  // bfdot z17.s, z28.h, z2.h[0]\n"
       ".inst 0x64634395  // bfdot z21.s, z28.h, z3.h[0]\n"
       ".inst 0x64644399  // bfdot z25.s, z28.h, z4.h[0]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1423,21 +1421,21 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 63f\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
       "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x646843a8  // bfdot z8.s, z29.h, z0.h[1]\n"
       ".inst 0x646943ac  // bfdot z12.s, z29.h, z1.h[1]\n"
       ".inst 0x646a43b0  // bfdot z16.s, z29.h, z2.h[1]\n"
       ".inst 0x646b43b4  // bfdot z20.s, z29.h, z3.h[1]\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x646c43b8  // bfdot z24.s, z29.h, z4.h[1]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64684389  // bfdot z9.s, z28.h, z0.h[1]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6469438d  // bfdot z13.s, z28.h, z1.h[1]\n"
       ".inst 0x646a4391  // bfdot z17.s, z28.h, z2.h[1]\n"
       ".inst 0x646b4395  // bfdot z21.s, z28.h, z3.h[1]\n"
       ".inst 0x646c4399  // bfdot z25.s, z28.h, z4.h[1]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x646843aa  // bfdot z10.s, z29.h, z0.h[1]\n"
       "addvl x10, x10, #4\n"
+      ".inst 0x646843aa  // bfdot z10.s, z29.h, z0.h[1]\n"
       ".inst 0x646943ae  // bfdot z14.s, z29.h, z1.h[1]\n"
       ".inst 0x646a43b2  // bfdot z18.s, z29.h, z2.h[1]\n"
       ".inst 0x646b43b6  // bfdot z22.s, z29.h, z3.h[1]\n"
@@ -1450,21 +1448,21 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 63f\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
       "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x647043a8  // bfdot z8.s, z29.h, z0.h[2]\n"
       ".inst 0x647143ac  // bfdot z12.s, z29.h, z1.h[2]\n"
       ".inst 0x647243b0  // bfdot z16.s, z29.h, z2.h[2]\n"
       ".inst 0x647343b4  // bfdot z20.s, z29.h, z3.h[2]\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x647443b8  // bfdot z24.s, z29.h, z4.h[2]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64704389  // bfdot z9.s, z28.h, z0.h[2]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6471438d  // bfdot z13.s, z28.h, z1.h[2]\n"
       ".inst 0x64724391  // bfdot z17.s, z28.h, z2.h[2]\n"
       ".inst 0x64734395  // bfdot z21.s, z28.h, z3.h[2]\n"
       ".inst 0x64744399  // bfdot z25.s, z28.h, z4.h[2]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x647043aa  // bfdot z10.s, z29.h, z0.h[2]\n"
       "addvl x10, x10, #4\n"
+      ".inst 0x647043aa  // bfdot z10.s, z29.h, z0.h[2]\n"
       ".inst 0x647143ae  // bfdot z14.s, z29.h, z1.h[2]\n"
       ".inst 0x647243b2  // bfdot z18.s, z29.h, z2.h[2]\n"
       ".inst 0x647343b6  // bfdot z22.s, z29.h, z3.h[2]\n"
@@ -1482,8 +1480,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       ".inst 0x647a43b0  // bfdot z16.s, z29.h, z2.h[3]\n"
       ".inst 0x647b43b4  // bfdot z20.s, z29.h, z3.h[3]\n"
       ".inst 0x647c43b8  // bfdot z24.s, z29.h, z4.h[3]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x64784389  // bfdot z9.s, z28.h, z0.h[3]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6479438d  // bfdot z13.s, z28.h, z1.h[3]\n"
       ".inst 0x647a4391  // bfdot z17.s, z28.h, z2.h[3]\n"
       ".inst 0x647b4395  // bfdot z21.s, z28.h, z3.h[3]\n"
@@ -1506,14 +1504,14 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 58b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z29.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z29.s }, p5/Z, [x21]\n"
       "ld1rw { z28.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z29.s\n"
       "fmin z9.s, p5/M, z9.s, z29.s\n"
@@ -1561,22 +1559,22 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
       "65:"  // Height 5: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1584,13 +1582,12 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "b 80f\n"
       "66:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "67:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1603,18 +1600,18 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cbz x12, 68f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1630,17 +1627,17 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "68:"  // Height 6: no bias
       "tbz %x[flags], #0, 69f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x24]\n"
       "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x23]\n"
@@ -1689,8 +1686,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "mov x28, #0x0\n"
       "71:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 72f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1721,29 +1718,29 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 75f\n"
       "74:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z1.h }, p5/Z, [x10]\n"
-      "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z7.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z6.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z5.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z4.h }, p0/Z, [x23]\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "ld1rqh { z3.h }, p0/Z, [x22]\n"
       "ld1rqh { z2.h }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1h { z1.h }, p5/Z, [x10]\n"
+      "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x64674028  // bfdot z8.s, z1.h, z7.h[0]\n"
       ".inst 0x6466402c  // bfdot z12.s, z1.h, z6.h[0]\n"
-      "add x21, x21, #0x10\n"
       ".inst 0x64654030  // bfdot z16.s, z1.h, z5.h[0]\n"
       ".inst 0x64644034  // bfdot z20.s, z1.h, z4.h[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x64634038  // bfdot z24.s, z1.h, z3.h[0]\n"
       ".inst 0x6462403c  // bfdot z28.s, z1.h, z2.h[0]\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x64674009  // bfdot z9.s, z0.h, z7.h[0]\n"
       ".inst 0x6466400d  // bfdot z13.s, z0.h, z6.h[0]\n"
       ".inst 0x64654011  // bfdot z17.s, z0.h, z5.h[0]\n"
@@ -1851,24 +1848,24 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "bgt 74b\n"
       "75:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z7.h }, p5/Z, [x10]\n"
-      "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x2\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
       "ld1rqh { z4.h }, p0/Z, [x22]\n"
       "ld1rqh { z5.h }, p0/Z, [x21]\n"
+      "ld1h { z7.h }, p5/Z, [x10]\n"
+      "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x646040e8  // bfdot z8.s, z7.h, z0.h[0]\n"
       ".inst 0x646140ec  // bfdot z12.s, z7.h, z1.h[0]\n"
-      ".inst 0x646040c9  // bfdot z9.s, z6.h, z0.h[0]\n"
-      ".inst 0x646140cd  // bfdot z13.s, z6.h, z1.h[0]\n"
       ".inst 0x646240f0  // bfdot z16.s, z7.h, z2.h[0]\n"
       ".inst 0x646340f4  // bfdot z20.s, z7.h, z3.h[0]\n"
       ".inst 0x646440f8  // bfdot z24.s, z7.h, z4.h[0]\n"
       ".inst 0x646540fc  // bfdot z28.s, z7.h, z5.h[0]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      ".inst 0x646040c9  // bfdot z9.s, z6.h, z0.h[0]\n"
+      ".inst 0x646140cd  // bfdot z13.s, z6.h, z1.h[0]\n"
       ".inst 0x646240d1  // bfdot z17.s, z6.h, z2.h[0]\n"
       ".inst 0x646340d5  // bfdot z21.s, z6.h, z3.h[0]\n"
       ".inst 0x646440d9  // bfdot z25.s, z6.h, z4.h[0]\n"
@@ -1890,23 +1887,23 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 76f\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x646840e8  // bfdot z8.s, z7.h, z0.h[1]\n"
       ".inst 0x646940ec  // bfdot z12.s, z7.h, z1.h[1]\n"
       ".inst 0x646a40f0  // bfdot z16.s, z7.h, z2.h[1]\n"
       ".inst 0x646b40f4  // bfdot z20.s, z7.h, z3.h[1]\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x646c40f8  // bfdot z24.s, z7.h, z4.h[1]\n"
       ".inst 0x646d40fc  // bfdot z28.s, z7.h, z5.h[1]\n"
-      ".inst 0x646840c9  // bfdot z9.s, z6.h, z0.h[1]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      ".inst 0x646840c9  // bfdot z9.s, z6.h, z0.h[1]\n"
       ".inst 0x646940cd  // bfdot z13.s, z6.h, z1.h[1]\n"
       ".inst 0x646a40d1  // bfdot z17.s, z6.h, z2.h[1]\n"
       ".inst 0x646b40d5  // bfdot z21.s, z6.h, z3.h[1]\n"
       ".inst 0x646c40d9  // bfdot z25.s, z6.h, z4.h[1]\n"
       ".inst 0x646d40dd  // bfdot z29.s, z6.h, z5.h[1]\n"
       "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x646840ea  // bfdot z10.s, z7.h, z0.h[1]\n"
       "addvl x10, x10, #4\n"
+      ".inst 0x646840ea  // bfdot z10.s, z7.h, z0.h[1]\n"
       ".inst 0x646940ee  // bfdot z14.s, z7.h, z1.h[1]\n"
       ".inst 0x646a40f2  // bfdot z18.s, z7.h, z2.h[1]\n"
       ".inst 0x646b40f6  // bfdot z22.s, z7.h, z3.h[1]\n"
@@ -1921,23 +1918,23 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "ble 76f\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x2\n"
       ".inst 0x647040e8  // bfdot z8.s, z7.h, z0.h[2]\n"
       ".inst 0x647140ec  // bfdot z12.s, z7.h, z1.h[2]\n"
       ".inst 0x647240f0  // bfdot z16.s, z7.h, z2.h[2]\n"
       ".inst 0x647340f4  // bfdot z20.s, z7.h, z3.h[2]\n"
+      "subs x27, x27, #0x2\n"
       ".inst 0x647440f8  // bfdot z24.s, z7.h, z4.h[2]\n"
       ".inst 0x647540fc  // bfdot z28.s, z7.h, z5.h[2]\n"
-      ".inst 0x647040c9  // bfdot z9.s, z6.h, z0.h[2]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      ".inst 0x647040c9  // bfdot z9.s, z6.h, z0.h[2]\n"
       ".inst 0x647140cd  // bfdot z13.s, z6.h, z1.h[2]\n"
       ".inst 0x647240d1  // bfdot z17.s, z6.h, z2.h[2]\n"
       ".inst 0x647340d5  // bfdot z21.s, z6.h, z3.h[2]\n"
       ".inst 0x647440d9  // bfdot z25.s, z6.h, z4.h[2]\n"
       ".inst 0x647540dd  // bfdot z29.s, z6.h, z5.h[2]\n"
       "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x647040ea  // bfdot z10.s, z7.h, z0.h[2]\n"
       "addvl x10, x10, #4\n"
+      ".inst 0x647040ea  // bfdot z10.s, z7.h, z0.h[2]\n"
       ".inst 0x647140ee  // bfdot z14.s, z7.h, z1.h[2]\n"
       ".inst 0x647240f2  // bfdot z18.s, z7.h, z2.h[2]\n"
       ".inst 0x647340f6  // bfdot z22.s, z7.h, z3.h[2]\n"
@@ -1985,15 +1982,15 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 71b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "tbz %x[flags], #1, 77f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p5/Z, [x21]\n"
       "ld1rw { z0.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z1.s\n"
       "fmin z9.s, p5/M, z9.s, z1.s\n"
@@ -2049,26 +2046,26 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
-      "st1w { z28.s }, p4, [x22]\n"
-      "st1w { z29.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z30.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z31.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z28.s }, p4, [x21]\n"
+      "st1w { z29.s }, p3, [x21, #1, MUL VL]\n"
+      "st1w { z30.s }, p2, [x21, #2, MUL VL]\n"
+      "st1w { z31.s }, p1, [x21, #3, MUL VL]\n"
       "78:"  // Height 6: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -2085,8 +2082,8 @@ void sve_hybrid_bf16fp32_dot_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL.hpp
index 325499b7a3..d7ca55c295 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 8, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 8, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp
index 64788ab092..ad997b0034 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_bf16fp32_mmla_6x4VL/generic.cpp
@@ -48,19 +48,18 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -81,7 +80,6 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -105,10 +103,10 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cmp %x[M], #0x2\n"
       "bgt 27f\n"
       "beq 14f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -121,26 +119,26 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cbz x12, 3f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "b 5f\n"
       "3:"  // Height 1: no bias
       "tbz %x[flags], #0, 4f\n"
-      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z16.s }, p4/Z, [x9]\n"
       "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "zip1 z8.d, z16.d, z12.d\n"
+      "zip2 z12.d, z16.d, z12.d\n"
       "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "zip1 z8.d, z19.d, z12.d\n"
-      "zip2 z12.d, z19.d, z12.d\n"
       "zip1 z9.d, z18.d, z13.d\n"
       "zip2 z13.d, z18.d, z13.d\n"
       "zip1 z10.d, z17.d, z14.d\n"
@@ -161,8 +159,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -178,87 +176,87 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
+      "ld1rqh { z20.h }, p0/Z, [x26]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z20.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "trn1 z19.d, z20.d, z18.d\n"
-      "trn2 z20.d, z20.d, z18.d\n"
-      ".inst 0x6471e668  // bfmmla z8.s, z19.h, z17.h\n"
-      "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6470e66c  // bfmmla z12.s, z19.h, z16.h\n"
+      ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
+      ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6461e669  // bfmmla z9.s, z19.h, z1.h\n"
-      "ld1h { z18.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6470e66d  // bfmmla z13.s, z19.h, z16.h\n"
+      ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
+      ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6472e66a  // bfmmla z10.s, z19.h, z18.h\n"
+      ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
+      ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6470e66e  // bfmmla z14.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x6471e66b  // bfmmla z11.s, z19.h, z17.h\n"
-      ".inst 0x6470e66f  // bfmmla z15.s, z19.h, z16.h\n"
+      "trn2 z20.d, z20.d, z19.d\n"
+      ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
+      ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-7, MUL VL]\n"
       ".inst 0x6471e688  // bfmmla z8.s, z20.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x6470e68c  // bfmmla z12.s, z20.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-5, MUL VL]\n"
       ".inst 0x6471e689  // bfmmla z9.s, z20.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x6470e68d  // bfmmla z13.s, z20.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x6471e68a  // bfmmla z10.s, z20.h, z17.h\n"
+      "ld1h { z16.h }, p5/Z, [x10, #-4, MUL VL]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-3, MUL VL]\n"
+      ".inst 0x6470e68a  // bfmmla z10.s, z20.h, z16.h\n"
+      ".inst 0x6471e68e  // bfmmla z14.s, z20.h, z17.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x6470e68e  // bfmmla z14.s, z20.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #-1, MUL VL]\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6471e68b  // bfmmla z11.s, z20.h, z17.h\n"
       ".inst 0x6470e68f  // bfmmla z15.s, z20.h, z16.h\n"
+      "add x26, x26, #0x10\n"
       "bgt 9b\n"
       "10:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "subs x27, x27, #0x4\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
       ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
+      "addvl x10, x10, #8\n"
       "ble 11f\n"
-      "ld1h { z16.h }, p5/Z, [x10]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6470e428  // bfmmla z8.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6471e42c  // bfmmla z12.s, z1.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6470e429  // bfmmla z9.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6471e42d  // bfmmla z13.s, z1.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6470e42a  // bfmmla z10.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6471e42e  // bfmmla z14.s, z1.h, z17.h\n"
-      "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x6471e428  // bfmmla z8.s, z1.h, z17.h\n"
+      ".inst 0x6470e42c  // bfmmla z12.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
+      ".inst 0x6471e429  // bfmmla z9.s, z1.h, z17.h\n"
+      ".inst 0x6470e42d  // bfmmla z13.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
+      ".inst 0x6471e42a  // bfmmla z10.s, z1.h, z17.h\n"
+      ".inst 0x6470e42e  // bfmmla z14.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
+      ".inst 0x6471e42b  // bfmmla z11.s, z1.h, z17.h\n"
+      ".inst 0x6470e42f  // bfmmla z15.s, z1.h, z16.h\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6470e42b  // bfmmla z11.s, z1.h, z16.h\n"
-      ".inst 0x6460e42f  // bfmmla z15.s, z1.h, z0.h\n"
       "11:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -269,14 +267,14 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "uzp1 z10.d, z10.d, z14.d\n"
       "uzp1 z11.d, z11.d, z15.d\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
-      "fmin z8.s, p5/M, z8.s, z17.s\n"
-      "fmin z9.s, p5/M, z9.s, z17.s\n"
-      "fmin z10.s, p5/M, z10.s, z17.s\n"
-      "fmin z11.s, p5/M, z11.s, z17.s\n"
+      "fmin z8.s, p5/M, z8.s, z21.s\n"
+      "fmin z9.s, p5/M, z9.s, z21.s\n"
+      "fmin z10.s, p5/M, z10.s, z21.s\n"
+      "fmin z11.s, p5/M, z11.s, z21.s\n"
       "fmax z8.s, p5/M, z8.s, z16.s\n"
       "fmax z9.s, p5/M, z9.s, z16.s\n"
       "fmax z10.s, p5/M, z10.s, z16.s\n"
@@ -293,10 +291,10 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "bgt 2b\n"
       "b 80f\n"
       "14:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -309,38 +307,38 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cbz x12, 16f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "b 18f\n"
       "16:"  // Height 2: no bias
       "tbz %x[flags], #0, 17f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z18.s }, p4/Z, [x9]\n"
-      "ld1w { z16.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z5.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z17.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x20, x9, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
+      "zip1 z9.d, z18.d, z13.d\n"
+      "zip2 z13.d, z18.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z8.d, z18.d, z12.d\n"
-      "zip2 z12.d, z18.d, z12.d\n"
-      "zip1 z9.d, z16.d, z13.d\n"
-      "zip2 z13.d, z16.d, z13.d\n"
-      "zip1 z10.d, z5.d, z14.d\n"
-      "zip2 z14.d, z5.d, z14.d\n"
-      "zip1 z11.d, z17.d, z15.d\n"
-      "zip2 z15.d, z17.d, z15.d\n"
+      "zip1 z10.d, z17.d, z14.d\n"
+      "zip2 z14.d, z17.d, z14.d\n"
+      "zip1 z11.d, z16.d, z15.d\n"
+      "zip2 z15.d, z16.d, z15.d\n"
       "b 18f\n"
       "17:"  // Height 2: no accumulate
       "mov z8.b, #0x0\n"
@@ -355,8 +353,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "19:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -375,90 +373,90 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "ble 23f\n"
       "22:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
+      "ld1rqh { z20.h }, p0/Z, [x26]\n"
+      "ld1rqh { z19.h }, p0/Z, [x25]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z19.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqh { z16.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "trn1 z18.d, z19.d, z16.d\n"
-      "trn2 z19.d, z19.d, z16.d\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
+      ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6474e64c  // bfmmla z12.s, z18.h, z20.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
+      "trn2 z20.d, z20.d, z19.d\n"
       ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
       ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x6471e668  // bfmmla z8.s, z19.h, z17.h\n"
+      ".inst 0x6471e688  // bfmmla z8.s, z20.h, z17.h\n"
+      ".inst 0x6470e68c  // bfmmla z12.s, z20.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x6470e66c  // bfmmla z12.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x6471e669  // bfmmla z9.s, z19.h, z17.h\n"
+      ".inst 0x6471e689  // bfmmla z9.s, z20.h, z17.h\n"
+      ".inst 0x6470e68d  // bfmmla z13.s, z20.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x6470e66d  // bfmmla z13.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x6471e66a  // bfmmla z10.s, z19.h, z17.h\n"
+      ".inst 0x6471e68a  // bfmmla z10.s, z20.h, z17.h\n"
+      ".inst 0x6470e68e  // bfmmla z14.s, z20.h, z16.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x6470e66e  // bfmmla z14.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x6471e66b  // bfmmla z11.s, z19.h, z17.h\n"
-      ".inst 0x6470e66f  // bfmmla z15.s, z19.h, z16.h\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x6471e68b  // bfmmla z11.s, z20.h, z17.h\n"
+      ".inst 0x6470e68f  // bfmmla z15.s, z20.h, z16.h\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "bgt 22b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
       "ld1rqh { z19.h }, p0/Z, [x25]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "subs x27, x27, #0x4\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
       ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
+      "addvl x10, x10, #8\n"
       "ble 24f\n"
-      "ld1h { z16.h }, p5/Z, [x10]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6470e428  // bfmmla z8.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6471e42c  // bfmmla z12.s, z1.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6470e429  // bfmmla z9.s, z1.h, z16.h\n"
-      "ld1h { z16.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6471e42d  // bfmmla z13.s, z1.h, z17.h\n"
-      "ld1h { z26.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6470e42a  // bfmmla z10.s, z1.h, z16.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x647ae42e  // bfmmla z14.s, z1.h, z26.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x6471e428  // bfmmla z8.s, z1.h, z17.h\n"
+      ".inst 0x6470e42c  // bfmmla z12.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
+      ".inst 0x6471e429  // bfmmla z9.s, z1.h, z17.h\n"
+      ".inst 0x6470e42d  // bfmmla z13.s, z1.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
+      ".inst 0x6471e42a  // bfmmla z10.s, z1.h, z17.h\n"
+      ".inst 0x6470e42e  // bfmmla z14.s, z1.h, z16.h\n"
+      "ld1h { z22.h }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x6471e42b  // bfmmla z11.s, z1.h, z17.h\n"
+      ".inst 0x6476e42b  // bfmmla z11.s, z1.h, z22.h\n"
       ".inst 0x6470e42f  // bfmmla z15.s, z1.h, z16.h\n"
+      "addvl x10, x10, #8\n"
       "24:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -467,17 +465,17 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z7.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
+      "add x25, x9, x20, LSL #2\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x26, x9, x20, LSL #2\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z17.s\n"
       "fmin z12.s, p5/M, z12.s, z17.s\n"
@@ -501,20 +499,20 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
       "26:"  // Height 2: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 15b\n"
       "b 80f\n"
       "27:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -527,15 +525,15 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cbz x12, 29f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -550,36 +548,36 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "29:"  // Height 3: no bias
       "tbz %x[flags], #0, 30f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z26.s }, p4/Z, [x9]\n"
-      "ld1w { z25.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z24.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x21, x9, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x20]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "zip1 z8.d, z26.d, z12.d\n"
-      "zip2 z12.d, z26.d, z12.d\n"
-      "ld1w { z2.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z9.d, z25.d, z13.d\n"
-      "zip2 z13.d, z25.d, z13.d\n"
-      "zip1 z10.d, z24.d, z14.d\n"
-      "zip2 z14.d, z24.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
-      "zip1 z19.d, z2.d, z23.d\n"
-      "zip2 z23.d, z2.d, z23.d\n"
+      "zip1 z19.d, z24.d, z23.d\n"
+      "zip2 z23.d, z24.d, z23.d\n"
       "b 31f\n"
       "30:"  // Height 3: no accumulate
       "mov z8.b, #0x0\n"
@@ -602,8 +600,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -625,92 +623,92 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "ble 36f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x10]\n"
-      "ld1h { z30.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z29.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
+      "ld1rqh { z30.h }, p0/Z, [x26]\n"
       "ld1rqh { z24.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
       "ld1rqh { z28.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "trn1 z27.d, z29.d, z24.d\n"
-      "trn2 z29.d, z29.d, z24.d\n"
-      "trn1 z26.d, z28.d, z31.d\n"
-      "trn2 z28.d, z28.d, z31.d\n"
+      "trn1 z27.d, z30.d, z24.d\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "trn1 z26.d, z28.d, z29.d\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6479e768  // bfmmla z8.s, z27.h, z25.h\n"
-      ".inst 0x647ee76c  // bfmmla z12.s, z27.h, z30.h\n"
       ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e76c  // bfmmla z12.s, z27.h, z24.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x647ee754  // bfmmla z20.s, z26.h, z30.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "trn2 z28.d, z28.d, z29.d\n"
       ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n"
+      "sub x27, x27, #0x8\n"
       ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n"
+      "cmp x27, #0x8\n"
       ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
       ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
+      "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
       ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
-      "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z24.h }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x6479e7a8  // bfmmla z8.s, z29.h, z25.h\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x6479e7c8  // bfmmla z8.s, z30.h, z25.h\n"
       ".inst 0x6479e790  // bfmmla z16.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x6478e7ac  // bfmmla z12.s, z29.h, z24.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6478e7cc  // bfmmla z12.s, z30.h, z24.h\n"
       ".inst 0x6478e794  // bfmmla z20.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x6479e7a9  // bfmmla z9.s, z29.h, z25.h\n"
+      ".inst 0x6479e7c9  // bfmmla z9.s, z30.h, z25.h\n"
       ".inst 0x6479e791  // bfmmla z17.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x6478e7ad  // bfmmla z13.s, z29.h, z24.h\n"
+      ".inst 0x6478e7cd  // bfmmla z13.s, z30.h, z24.h\n"
       ".inst 0x6478e795  // bfmmla z21.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x6479e7aa  // bfmmla z10.s, z29.h, z25.h\n"
+      ".inst 0x6479e7ca  // bfmmla z10.s, z30.h, z25.h\n"
       ".inst 0x6479e792  // bfmmla z18.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x6478e7ae  // bfmmla z14.s, z29.h, z24.h\n"
+      ".inst 0x6478e7ce  // bfmmla z14.s, z30.h, z24.h\n"
       ".inst 0x6478e796  // bfmmla z22.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x6479e7ab  // bfmmla z11.s, z29.h, z25.h\n"
+      ".inst 0x6479e7cb  // bfmmla z11.s, z30.h, z25.h\n"
       ".inst 0x6479e793  // bfmmla z19.s, z28.h, z25.h\n"
-      ".inst 0x6478e7af  // bfmmla z15.s, z29.h, z24.h\n"
+      ".inst 0x6478e7cf  // bfmmla z15.s, z30.h, z24.h\n"
       ".inst 0x6478e797  // bfmmla z23.s, z28.h, z24.h\n"
       "bgt 35b\n"
       "36:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
       "ld1rqh { z24.h }, p0/Z, [x25]\n"
       "ld1rqh { z3.h }, p0/Z, [x24]\n"
       "trn1 z27.d, z1.d, z24.d\n"
       "trn2 z1.d, z1.d, z24.d\n"
-      "trn1 z26.d, z3.d, z29.d\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "trn1 z26.d, z3.d, z28.d\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6479e768  // bfmmla z8.s, z27.h, z25.h\n"
-      ".inst 0x647ce76c  // bfmmla z12.s, z27.h, z28.h\n"
-      "trn2 z3.d, z3.d, z29.d\n"
       ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e76c  // bfmmla z12.s, z27.h, z24.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x647ce754  // bfmmla z20.s, z26.h, z28.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n"
+      "trn2 z3.d, z3.d, z28.d\n"
       ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n"
@@ -727,9 +725,9 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6479e428  // bfmmla z8.s, z1.h, z25.h\n"
       ".inst 0x6479e470  // bfmmla z16.s, z3.h, z25.h\n"
-      "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6478e42c  // bfmmla z12.s, z1.h, z24.h\n"
       ".inst 0x6478e474  // bfmmla z20.s, z3.h, z24.h\n"
+      "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6479e429  // bfmmla z9.s, z1.h, z25.h\n"
       ".inst 0x6479e471  // bfmmla z17.s, z3.h, z25.h\n"
@@ -754,24 +752,24 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "uzp1 z7.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x9, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 z16.d, z16.d, z20.d\n"
       "uzp1 z17.d, z17.d, z21.d\n"
       "uzp1 z18.d, z18.d, z22.d\n"
       "uzp1 z19.d, z19.d, z23.d\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z25.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z25.s }, p5/Z, [x21]\n"
       "ld1rw { z24.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z25.s\n"
       "fmin z12.s, p5/M, z12.s, z25.s\n"
@@ -803,24 +801,24 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
       "39:"  // Height 3: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 28b\n"
       "b 80f\n"
       "40:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -833,15 +831,15 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cbz x12, 42f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -856,37 +854,37 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "42:"  // Height 4: no bias
       "tbz %x[flags], #0, 43f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x22, x9, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x21]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x20]\n"
-      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x20]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
@@ -913,8 +911,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "45:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -939,114 +937,114 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "ble 49f\n"
       "48:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z31.h }, p5/Z, [x10]\n"
-      "ld1h { z30.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z29.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqh { z25.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqh { z30.h }, p0/Z, [x26]\n"
+      "ld1rqh { z24.h }, p0/Z, [x25]\n"
+      "trn1 z29.d, z30.d, z24.d\n"
       "ld1rqh { z28.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqh { z24.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "trn1 z27.d, z29.d, z25.d\n"
-      "trn2 z29.d, z29.d, z25.d\n"
-      "trn1 z26.d, z28.d, z24.d\n"
-      "trn2 z28.d, z28.d, z24.d\n"
-      ".inst 0x647fe768  // bfmmla z8.s, z27.h, z31.h\n"
-      ".inst 0x647ee76c  // bfmmla z12.s, z27.h, z30.h\n"
-      ".inst 0x647fe750  // bfmmla z16.s, z26.h, z31.h\n"
+      "ld1rqh { z27.h }, p0/Z, [x23]\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "trn1 z26.d, z28.d, z27.d\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x6479e7a8  // bfmmla z8.s, z29.h, z25.h\n"
+      ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e7ac  // bfmmla z12.s, z29.h, z24.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x647ee754  // bfmmla z20.s, z26.h, z30.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
+      ".inst 0x6479e7a9  // bfmmla z9.s, z29.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
+      "trn2 z28.d, z28.d, z27.d\n"
+      ".inst 0x6478e7ad  // bfmmla z13.s, z29.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
+      "sub x27, x27, #0x8\n"
+      ".inst 0x6479e7aa  // bfmmla z10.s, z29.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
+      "cmp x27, #0x8\n"
+      ".inst 0x6478e7ae  // bfmmla z14.s, z29.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
+      ".inst 0x6479e7ab  // bfmmla z11.s, z29.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
-      ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
-      ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
+      "add x26, x26, #0x10\n"
+      ".inst 0x6478e7af  // bfmmla z15.s, z29.h, z24.h\n"
+      ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x6479e7a8  // bfmmla z8.s, z29.h, z25.h\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x6479e7c8  // bfmmla z8.s, z30.h, z25.h\n"
       ".inst 0x6479e790  // bfmmla z16.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x6478e7ac  // bfmmla z12.s, z29.h, z24.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6478e7cc  // bfmmla z12.s, z30.h, z24.h\n"
       ".inst 0x6478e794  // bfmmla z20.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x6479e7a9  // bfmmla z9.s, z29.h, z25.h\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6479e7c9  // bfmmla z9.s, z30.h, z25.h\n"
       ".inst 0x6479e791  // bfmmla z17.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x6478e7ad  // bfmmla z13.s, z29.h, z24.h\n"
+      ".inst 0x6478e7cd  // bfmmla z13.s, z30.h, z24.h\n"
       ".inst 0x6478e795  // bfmmla z21.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x6479e7aa  // bfmmla z10.s, z29.h, z25.h\n"
+      ".inst 0x6479e7ca  // bfmmla z10.s, z30.h, z25.h\n"
       ".inst 0x6479e792  // bfmmla z18.s, z28.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x6478e7ae  // bfmmla z14.s, z29.h, z24.h\n"
+      ".inst 0x6478e7ce  // bfmmla z14.s, z30.h, z24.h\n"
       ".inst 0x6478e796  // bfmmla z22.s, z28.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x6479e7ab  // bfmmla z11.s, z29.h, z25.h\n"
+      ".inst 0x6479e7cb  // bfmmla z11.s, z30.h, z25.h\n"
       ".inst 0x6479e793  // bfmmla z19.s, z28.h, z25.h\n"
-      ".inst 0x6478e7af  // bfmmla z15.s, z29.h, z24.h\n"
+      ".inst 0x6478e7cf  // bfmmla z15.s, z30.h, z24.h\n"
       ".inst 0x6478e797  // bfmmla z23.s, z28.h, z24.h\n"
       "bgt 48b\n"
       "49:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "ld1rqh { z25.h }, p0/Z, [x25]\n"
+      "ld1rqh { z24.h }, p0/Z, [x25]\n"
+      "trn1 z28.d, z1.d, z24.d\n"
       "ld1rqh { z3.h }, p0/Z, [x24]\n"
-      "ld1rqh { z24.h }, p0/Z, [x23]\n"
-      "trn1 z27.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      "trn1 z26.d, z3.d, z24.d\n"
-      ".inst 0x647de768  // bfmmla z8.s, z27.h, z29.h\n"
-      ".inst 0x647ce76c  // bfmmla z12.s, z27.h, z28.h\n"
-      "trn2 z3.d, z3.d, z24.d\n"
-      ".inst 0x647de750  // bfmmla z16.s, z26.h, z29.h\n"
+      "ld1rqh { z27.h }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z24.d\n"
+      "trn1 z26.d, z3.d, z27.d\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x6479e788  // bfmmla z8.s, z28.h, z25.h\n"
+      ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e78c  // bfmmla z12.s, z28.h, z24.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x647ce754  // bfmmla z20.s, z26.h, z28.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
+      ".inst 0x6479e789  // bfmmla z9.s, z28.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
+      "subs x27, x27, #0x4\n"
+      ".inst 0x6478e78d  // bfmmla z13.s, z28.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
+      "trn2 z3.d, z3.d, z27.d\n"
+      ".inst 0x6479e78a  // bfmmla z10.s, z28.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
+      ".inst 0x6478e78e  // bfmmla z14.s, z28.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
+      ".inst 0x6479e78b  // bfmmla z11.s, z28.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
-      ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
+      ".inst 0x6478e78f  // bfmmla z15.s, z28.h, z24.h\n"
       ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6479e428  // bfmmla z8.s, z1.h, z25.h\n"
       ".inst 0x6479e470  // bfmmla z16.s, z3.h, z25.h\n"
-      "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6478e42c  // bfmmla z12.s, z1.h, z24.h\n"
       ".inst 0x6478e474  // bfmmla z20.s, z3.h, z24.h\n"
+      "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6479e429  // bfmmla z9.s, z1.h, z25.h\n"
       ".inst 0x6479e471  // bfmmla z17.s, z3.h, z25.h\n"
@@ -1071,17 +1069,17 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z7.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
       "uzp1 z20.d, z17.d, z21.d\n"
@@ -1091,9 +1089,9 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z24.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z24.s }, p5/Z, [x21]\n"
       "ld1rw { z23.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z24.s\n"
       "fmin z12.s, p5/M, z12.s, z24.s\n"
@@ -1133,28 +1131,28 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
       "52:"  // Height 4: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 41b\n"
       "b 80f\n"
       "53:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "54:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1167,15 +1165,15 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cbz x12, 55f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -1198,46 +1196,46 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "55:"  // Height 5: no bias
       "tbz %x[flags], #0, 56f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x22]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x21]\n"
-      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x20]\n"
-      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
+      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x21]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x20]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z19.d, z24.d, z23.d\n"
       "zip2 z23.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z24.d, z25.d, z28.d\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
@@ -1276,8 +1274,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "58:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 59f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1305,103 +1303,102 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "ble 62f\n"
       "61:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z1.h }, p5/Z, [x10]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z6.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqh { z3.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqh { z1.h }, p0/Z, [x25]\n"
       "ld1rqh { z7.h }, p0/Z, [x24]\n"
       "ld1rqh { z2.h }, p0/Z, [x23]\n"
-      "ld1rqh { z5.h }, p0/Z, [x22]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
+      "trn1 z5.d, z6.d, z1.d\n"
+      "trn2 z6.d, z6.d, z1.d\n"
+      "ld1rqh { z4.h }, p0/Z, [x22]\n"
       "trn1 z3.d, z7.d, z2.d\n"
       "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
+      "ld1h { z1.h }, p5/Z, [x10]\n"
+      "trn1 z2.d, z4.d, z0.d\n"
+      "trn2 z4.d, z4.d, z0.d\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6461e488  // bfmmla z8.s, z4.h, z1.h\n"
+      ".inst 0x6461e4a8  // bfmmla z8.s, z5.h, z1.h\n"
       ".inst 0x6461e470  // bfmmla z16.s, z3.h, z1.h\n"
       ".inst 0x6461e458  // bfmmla z24.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6460e48c  // bfmmla z12.s, z4.h, z0.h\n"
+      "sub x27, x27, #0x8\n"
+      ".inst 0x6460e4ac  // bfmmla z12.s, z5.h, z0.h\n"
       ".inst 0x6460e474  // bfmmla z20.s, z3.h, z0.h\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6460e45c  // bfmmla z28.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6461e489  // bfmmla z9.s, z4.h, z1.h\n"
+      ".inst 0x6461e4a9  // bfmmla z9.s, z5.h, z1.h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6461e471  // bfmmla z17.s, z3.h, z1.h\n"
       ".inst 0x6461e459  // bfmmla z25.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6460e48d  // bfmmla z13.s, z4.h, z0.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6460e4ad  // bfmmla z13.s, z5.h, z0.h\n"
       ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6460e45d  // bfmmla z29.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6461e48a  // bfmmla z10.s, z4.h, z1.h\n"
+      ".inst 0x6461e4aa  // bfmmla z10.s, z5.h, z1.h\n"
       ".inst 0x6461e472  // bfmmla z18.s, z3.h, z1.h\n"
       ".inst 0x6461e45a  // bfmmla z26.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6460e48e  // bfmmla z14.s, z4.h, z0.h\n"
+      ".inst 0x6460e4ae  // bfmmla z14.s, z5.h, z0.h\n"
       ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
       ".inst 0x6460e45e  // bfmmla z30.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x6461e48b  // bfmmla z11.s, z4.h, z1.h\n"
+      ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
       ".inst 0x6461e473  // bfmmla z19.s, z3.h, z1.h\n"
       ".inst 0x6461e45b  // bfmmla z27.s, z2.h, z1.h\n"
-      ".inst 0x6460e48f  // bfmmla z15.s, z4.h, z0.h\n"
+      "ld1h { z1.h }, p5/Z, [x10, #-8, MUL VL]\n"
+      ".inst 0x6460e4af  // bfmmla z15.s, z5.h, z0.h\n"
       ".inst 0x6460e477  // bfmmla z23.s, z3.h, z0.h\n"
       ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
-      "ld1h { z1.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z0.h }, p5/Z, [x10, #-7, MUL VL]\n"
       ".inst 0x6461e4c8  // bfmmla z8.s, z6.h, z1.h\n"
       ".inst 0x6461e4f0  // bfmmla z16.s, z7.h, z1.h\n"
-      ".inst 0x6461e4b8  // bfmmla z24.s, z5.h, z1.h\n"
+      ".inst 0x6461e498  // bfmmla z24.s, z4.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x6460e4cc  // bfmmla z12.s, z6.h, z0.h\n"
       ".inst 0x6460e4f4  // bfmmla z20.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bc  // bfmmla z28.s, z5.h, z0.h\n"
+      ".inst 0x6460e49c  // bfmmla z28.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #-5, MUL VL]\n"
       ".inst 0x6461e4c9  // bfmmla z9.s, z6.h, z1.h\n"
       ".inst 0x6461e4f1  // bfmmla z17.s, z7.h, z1.h\n"
-      ".inst 0x6461e4b9  // bfmmla z25.s, z5.h, z1.h\n"
+      ".inst 0x6461e499  // bfmmla z25.s, z4.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x6460e4cd  // bfmmla z13.s, z6.h, z0.h\n"
       ".inst 0x6460e4f5  // bfmmla z21.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
+      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #-3, MUL VL]\n"
       ".inst 0x6461e4ca  // bfmmla z10.s, z6.h, z1.h\n"
       ".inst 0x6461e4f2  // bfmmla z18.s, z7.h, z1.h\n"
-      ".inst 0x6461e4ba  // bfmmla z26.s, z5.h, z1.h\n"
+      ".inst 0x6461e49a  // bfmmla z26.s, z4.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x6460e4ce  // bfmmla z14.s, z6.h, z0.h\n"
       ".inst 0x6460e4f6  // bfmmla z22.s, z7.h, z0.h\n"
-      ".inst 0x6460e4be  // bfmmla z30.s, z5.h, z0.h\n"
+      ".inst 0x6460e49e  // bfmmla z30.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #-1, MUL VL]\n"
       ".inst 0x6461e4cb  // bfmmla z11.s, z6.h, z1.h\n"
       ".inst 0x6461e4f3  // bfmmla z19.s, z7.h, z1.h\n"
-      ".inst 0x6461e4bb  // bfmmla z27.s, z5.h, z1.h\n"
+      ".inst 0x6461e49b  // bfmmla z27.s, z4.h, z1.h\n"
       ".inst 0x6460e4cf  // bfmmla z15.s, z6.h, z0.h\n"
       ".inst 0x6460e4f7  // bfmmla z23.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bf  // bfmmla z31.s, z5.h, z0.h\n"
+      ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
       "bgt 61b\n"
       "62:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z2.h }, p5/Z, [x10]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "ld1rqh { z6.h }, p0/Z, [x25]\n"
+      "ld1rqh { z4.h }, p0/Z, [x25]\n"
       "ld1rqh { z3.h }, p0/Z, [x24]\n"
-      "ld1rqh { z4.h }, p0/Z, [x23]\n"
+      "ld1rqh { z2.h }, p0/Z, [x23]\n"
+      "trn1 z7.d, z1.d, z4.d\n"
+      "trn2 z1.d, z1.d, z4.d\n"
       "ld1rqh { z5.h }, p0/Z, [x22]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
+      "ld1h { z2.h }, p5/Z, [x10]\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
@@ -1409,6 +1406,7 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6462e4d0  // bfmmla z16.s, z6.h, z2.h\n"
       ".inst 0x6462e498  // bfmmla z24.s, z4.h, z2.h\n"
       "ld1h { z2.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       ".inst 0x6460e4ec  // bfmmla z12.s, z7.h, z0.h\n"
       ".inst 0x6460e4d4  // bfmmla z20.s, z6.h, z0.h\n"
       ".inst 0x6460e49c  // bfmmla z28.s, z4.h, z0.h\n"
@@ -1429,8 +1427,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6460e4d6  // bfmmla z22.s, z6.h, z0.h\n"
       ".inst 0x6460e49e  // bfmmla z30.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x6462e4eb  // bfmmla z11.s, z7.h, z2.h\n"
+      "addvl x10, x10, #8\n"
       ".inst 0x6462e4d3  // bfmmla z19.s, z6.h, z2.h\n"
       ".inst 0x6462e49b  // bfmmla z27.s, z4.h, z2.h\n"
       ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
@@ -1442,24 +1440,24 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6462e428  // bfmmla z8.s, z1.h, z2.h\n"
       ".inst 0x6462e470  // bfmmla z16.s, z3.h, z2.h\n"
       ".inst 0x6462e4b8  // bfmmla z24.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6460e42c  // bfmmla z12.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6460e474  // bfmmla z20.s, z3.h, z0.h\n"
       ".inst 0x6460e4bc  // bfmmla z28.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6462e429  // bfmmla z9.s, z1.h, z2.h\n"
       ".inst 0x6462e471  // bfmmla z17.s, z3.h, z2.h\n"
       ".inst 0x6462e4b9  // bfmmla z25.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6460e42d  // bfmmla z13.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
       ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x6462e42a  // bfmmla z10.s, z1.h, z2.h\n"
       ".inst 0x6462e472  // bfmmla z18.s, z3.h, z2.h\n"
       ".inst 0x6462e4ba  // bfmmla z26.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6460e42e  // bfmmla z14.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
       ".inst 0x6460e4be  // bfmmla z30.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
@@ -1476,20 +1474,20 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 58b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z7.d, z8.d, z12.d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
@@ -1501,9 +1499,9 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "uzp1 z26.d, z26.d, z30.d\n"
       "uzp1 z27.d, z27.d, z31.d\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z0.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z0.s }, p5/Z, [x21]\n"
       "ld1rw { z23.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z0.s\n"
       "fmin z12.s, p5/M, z12.s, z0.s\n"
@@ -1551,22 +1549,22 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
       "65:"  // Height 5: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1574,13 +1572,12 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "b 80f\n"
       "66:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "67:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1593,15 +1590,15 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cbz x12, 68f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -1624,54 +1621,54 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "68:"  // Height 6: no bias
       "tbz %x[flags], #0, 69f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "ld1w { z17.s }, p4/Z, [x9]\n"
       "add x22, x23, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x24]\n"
-      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x9, #2, MUL VL]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
+      "ld1w { z12.s }, p4/Z, [x24]\n"
+      "zip1 z8.d, z17.d, z12.d\n"
+      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
+      "zip2 z12.d, z17.d, z12.d\n"
+      "zip1 z9.d, z18.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x23]\n"
+      "zip2 z13.d, z18.d, z13.d\n"
+      "zip1 z10.d, z20.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip2 z14.d, z20.d, z14.d\n"
+      "zip1 z11.d, z16.d, z15.d\n"
       "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
       "ld1w { z20.s }, p4/Z, [x22]\n"
-      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x21]\n"
-      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
       "zip2 z20.d, z17.d, z20.d\n"
-      "ld1w { z28.s }, p4/Z, [x20]\n"
-      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x21]\n"
       "zip2 z21.d, z18.d, z21.d\n"
-      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z28.s }, p4/Z, [x20]\n"
       "zip2 z23.d, z24.d, z23.d\n"
       "zip1 z24.d, z25.d, z28.d\n"
+      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
+      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip2 z29.d, z26.d, z29.d\n"
       "zip1 z26.d, z27.d, z30.d\n"
       "zip2 z30.d, z27.d, z30.d\n"
@@ -1707,8 +1704,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "71:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 72f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1739,113 +1736,113 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "ble 75f\n"
       "74:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z1.h }, p5/Z, [x10]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
-      "ld1rqh { z6.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqh { z3.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqh { z7.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqh { z2.h }, p0/Z, [x23]\n"
-      "ld1rqh { z5.h }, p0/Z, [x22]\n"
+      "ld1rqh { z7.h }, p0/Z, [x26]\n"
+      "ld1rqh { z0.h }, p0/Z, [x25]\n"
+      "trn1 z6.d, z7.d, z0.d\n"
+      "ld1rqh { z5.h }, p0/Z, [x24]\n"
+      "ld1rqh { z1.h }, p0/Z, [x23]\n"
+      "trn2 z7.d, z7.d, z0.d\n"
+      "trn1 z4.d, z5.d, z1.d\n"
+      "ld1rqh { z3.h }, p0/Z, [x22]\n"
       "ld1rqh { z0.h }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
-      "add x21, x21, #0x10\n"
-      "trn1 z3.d, z7.d, z2.d\n"
-      "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
+      "trn2 z5.d, z5.d, z1.d\n"
+      "trn1 z2.d, z3.d, z0.d\n"
+      "trn2 z3.d, z3.d, z0.d\n"
+      "ld1h { z1.h }, p5/Z, [x10]\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x6461e488  // bfmmla z8.s, z4.h, z1.h\n"
-      ".inst 0x6461e470  // bfmmla z16.s, z3.h, z1.h\n"
+      ".inst 0x6461e4c8  // bfmmla z8.s, z6.h, z1.h\n"
+      ".inst 0x6461e490  // bfmmla z16.s, z4.h, z1.h\n"
       ".inst 0x6461e458  // bfmmla z24.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6460e48c  // bfmmla z12.s, z4.h, z0.h\n"
-      ".inst 0x6460e474  // bfmmla z20.s, z3.h, z0.h\n"
+      "sub x27, x27, #0x8\n"
+      ".inst 0x6460e4cc  // bfmmla z12.s, z6.h, z0.h\n"
+      ".inst 0x6460e494  // bfmmla z20.s, z4.h, z0.h\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x6460e45c  // bfmmla z28.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6461e489  // bfmmla z9.s, z4.h, z1.h\n"
-      ".inst 0x6461e471  // bfmmla z17.s, z3.h, z1.h\n"
+      ".inst 0x6461e4c9  // bfmmla z9.s, z6.h, z1.h\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x6461e491  // bfmmla z17.s, z4.h, z1.h\n"
       ".inst 0x6461e459  // bfmmla z25.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6460e48d  // bfmmla z13.s, z4.h, z0.h\n"
-      ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6460e4cd  // bfmmla z13.s, z6.h, z0.h\n"
+      ".inst 0x6460e495  // bfmmla z21.s, z4.h, z0.h\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6460e45d  // bfmmla z29.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6461e48a  // bfmmla z10.s, z4.h, z1.h\n"
-      ".inst 0x6461e472  // bfmmla z18.s, z3.h, z1.h\n"
+      ".inst 0x6461e4ca  // bfmmla z10.s, z6.h, z1.h\n"
+      "add x21, x21, #0x10\n"
+      ".inst 0x6461e492  // bfmmla z18.s, z4.h, z1.h\n"
       ".inst 0x6461e45a  // bfmmla z26.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6460e48e  // bfmmla z14.s, z4.h, z0.h\n"
-      ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
+      ".inst 0x6460e4ce  // bfmmla z14.s, z6.h, z0.h\n"
+      ".inst 0x6460e496  // bfmmla z22.s, z4.h, z0.h\n"
       ".inst 0x6460e45e  // bfmmla z30.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x6461e48b  // bfmmla z11.s, z4.h, z1.h\n"
-      ".inst 0x6461e473  // bfmmla z19.s, z3.h, z1.h\n"
+      ".inst 0x6461e4cb  // bfmmla z11.s, z6.h, z1.h\n"
+      ".inst 0x6461e493  // bfmmla z19.s, z4.h, z1.h\n"
       ".inst 0x6461e45b  // bfmmla z27.s, z2.h, z1.h\n"
-      ".inst 0x6460e48f  // bfmmla z15.s, z4.h, z0.h\n"
-      ".inst 0x6460e477  // bfmmla z23.s, z3.h, z0.h\n"
-      ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #-8, MUL VL]\n"
+      ".inst 0x6460e4cf  // bfmmla z15.s, z6.h, z0.h\n"
+      ".inst 0x6460e497  // bfmmla z23.s, z4.h, z0.h\n"
+      ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x6461e4c8  // bfmmla z8.s, z6.h, z1.h\n"
-      ".inst 0x6461e4f0  // bfmmla z16.s, z7.h, z1.h\n"
-      ".inst 0x6461e4b8  // bfmmla z24.s, z5.h, z1.h\n"
+      ".inst 0x6461e4e8  // bfmmla z8.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b0  // bfmmla z16.s, z5.h, z1.h\n"
+      ".inst 0x6461e478  // bfmmla z24.s, z3.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x6460e4cc  // bfmmla z12.s, z6.h, z0.h\n"
-      ".inst 0x6460e4f4  // bfmmla z20.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bc  // bfmmla z28.s, z5.h, z0.h\n"
+      ".inst 0x6460e4ec  // bfmmla z12.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b4  // bfmmla z20.s, z5.h, z0.h\n"
+      ".inst 0x6460e47c  // bfmmla z28.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x6461e4c9  // bfmmla z9.s, z6.h, z1.h\n"
-      ".inst 0x6461e4f1  // bfmmla z17.s, z7.h, z1.h\n"
-      ".inst 0x6461e4b9  // bfmmla z25.s, z5.h, z1.h\n"
+      ".inst 0x6461e4e9  // bfmmla z9.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b1  // bfmmla z17.s, z5.h, z1.h\n"
+      ".inst 0x6461e479  // bfmmla z25.s, z3.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x6460e4cd  // bfmmla z13.s, z6.h, z0.h\n"
-      ".inst 0x6460e4f5  // bfmmla z21.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
+      ".inst 0x6460e4ed  // bfmmla z13.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b5  // bfmmla z21.s, z5.h, z0.h\n"
+      ".inst 0x6460e47d  // bfmmla z29.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x6461e4ca  // bfmmla z10.s, z6.h, z1.h\n"
-      ".inst 0x6461e4f2  // bfmmla z18.s, z7.h, z1.h\n"
-      ".inst 0x6461e4ba  // bfmmla z26.s, z5.h, z1.h\n"
+      ".inst 0x6461e4ea  // bfmmla z10.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b2  // bfmmla z18.s, z5.h, z1.h\n"
+      ".inst 0x6461e47a  // bfmmla z26.s, z3.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x6460e4ce  // bfmmla z14.s, z6.h, z0.h\n"
-      ".inst 0x6460e4f6  // bfmmla z22.s, z7.h, z0.h\n"
-      ".inst 0x6460e4be  // bfmmla z30.s, z5.h, z0.h\n"
+      ".inst 0x6460e4ee  // bfmmla z14.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b6  // bfmmla z22.s, z5.h, z0.h\n"
+      ".inst 0x6460e47e  // bfmmla z30.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x6461e4cb  // bfmmla z11.s, z6.h, z1.h\n"
-      ".inst 0x6461e4f3  // bfmmla z19.s, z7.h, z1.h\n"
-      ".inst 0x6461e4bb  // bfmmla z27.s, z5.h, z1.h\n"
-      ".inst 0x6460e4cf  // bfmmla z15.s, z6.h, z0.h\n"
-      ".inst 0x6460e4f7  // bfmmla z23.s, z7.h, z0.h\n"
-      ".inst 0x6460e4bf  // bfmmla z31.s, z5.h, z0.h\n"
+      ".inst 0x6461e4eb  // bfmmla z11.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b3  // bfmmla z19.s, z5.h, z1.h\n"
+      ".inst 0x6461e47b  // bfmmla z27.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b7  // bfmmla z23.s, z5.h, z0.h\n"
+      ".inst 0x6460e47f  // bfmmla z31.s, z3.h, z0.h\n"
       "bgt 74b\n"
       "75:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z2.h }, p5/Z, [x10]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "ld1rqh { z6.h }, p0/Z, [x25]\n"
+      "ld1rqh { z0.h }, p0/Z, [x25]\n"
+      "trn1 z7.d, z1.d, z0.d\n"
       "ld1rqh { z3.h }, p0/Z, [x24]\n"
-      "ld1rqh { z4.h }, p0/Z, [x23]\n"
+      "ld1rqh { z2.h }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z0.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
       "ld1rqh { z5.h }, p0/Z, [x22]\n"
       "ld1rqh { z0.h }, p0/Z, [x21]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
+      "ld1h { z2.h }, p5/Z, [x10]\n"
       "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6462e4e8  // bfmmla z8.s, z7.h, z2.h\n"
       ".inst 0x6462e4d0  // bfmmla z16.s, z6.h, z2.h\n"
       ".inst 0x6462e498  // bfmmla z24.s, z4.h, z2.h\n"
       "ld1h { z2.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       ".inst 0x6460e4ec  // bfmmla z12.s, z7.h, z0.h\n"
       ".inst 0x6460e4d4  // bfmmla z20.s, z6.h, z0.h\n"
       ".inst 0x6460e49c  // bfmmla z28.s, z4.h, z0.h\n"
@@ -1866,8 +1863,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6460e4d6  // bfmmla z22.s, z6.h, z0.h\n"
       ".inst 0x6460e49e  // bfmmla z30.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x6462e4eb  // bfmmla z11.s, z7.h, z2.h\n"
+      "addvl x10, x10, #8\n"
       ".inst 0x6462e4d3  // bfmmla z19.s, z6.h, z2.h\n"
       ".inst 0x6462e49b  // bfmmla z27.s, z4.h, z2.h\n"
       ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
@@ -1879,24 +1876,24 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       ".inst 0x6462e428  // bfmmla z8.s, z1.h, z2.h\n"
       ".inst 0x6462e470  // bfmmla z16.s, z3.h, z2.h\n"
       ".inst 0x6462e4b8  // bfmmla z24.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6460e42c  // bfmmla z12.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6460e474  // bfmmla z20.s, z3.h, z0.h\n"
       ".inst 0x6460e4bc  // bfmmla z28.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6462e429  // bfmmla z9.s, z1.h, z2.h\n"
       ".inst 0x6462e471  // bfmmla z17.s, z3.h, z2.h\n"
       ".inst 0x6462e4b9  // bfmmla z25.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6460e42d  // bfmmla z13.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
       ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x6462e42a  // bfmmla z10.s, z1.h, z2.h\n"
       ".inst 0x6462e472  // bfmmla z18.s, z3.h, z2.h\n"
       ".inst 0x6462e4ba  // bfmmla z26.s, z5.h, z2.h\n"
-      "ld1h { z2.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6460e42e  // bfmmla z14.s, z1.h, z0.h\n"
+      "ld1h { z2.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
       ".inst 0x6460e4be  // bfmmla z30.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
@@ -1913,21 +1910,21 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 71b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z7.d, z8.d, z12.d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
@@ -1943,9 +1940,9 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "uzp1 z30.d, z27.d, z31.d\n"
       "uzp2 z27.d, z27.d, z31.d\n"
       "tbz %x[flags], #1, 77f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p5/Z, [x21]\n"
       "ld1rw { z0.s }, p5/Z, [x20]\n"
       "fmin z7.s, p5/M, z7.s, z1.s\n"
       "fmin z12.s, p5/M, z12.s, z1.s\n"
@@ -2001,26 +1998,26 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z23.s }, p4, [x23]\n"
-      "st1w { z28.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z29.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z30.s }, p1, [x23, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x22]\n"
-      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z23.s }, p4, [x22]\n"
+      "st1w { z28.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z29.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z30.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x21]\n"
+      "st1w { z25.s }, p3, [x21, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x21, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x21, #3, MUL VL]\n"
       "78:"  // Height 6: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -2037,8 +2034,8 @@ void sve_hybrid_bf16fp32_mmla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp
index ec754b6435..86d6ecab54 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 4, 1> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/a64fx.cpp
index 93462461ad..8578733628 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/a64fx.cpp
@@ -47,19 +47,18 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const __fp16 *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -104,10 +102,10 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "cmp %x[M], #0x2\n"
       "bgt 25f\n"
       "beq 13f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.h, x20, x11\n"
@@ -140,8 +138,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -160,14 +158,14 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop
       "fmla z8.h, p4/M, z6.h, z0.h\n"
-      "ld1h { z17.h }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z17.h }, p4/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p4/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
       "add x26, x26, #0x2\n"
-      "subs x27, x27, #0x1\n"
       "fmla z10.h, p4/M, z17.h, z0.h\n"
       "fmla z11.h, p4/M, z16.h, z0.h\n"
+      "subs x27, x27, #0x1\n"
       "ld1rh { z0.h }, p4/Z, [x26]\n"
       "ld1h { z6.h }, p4/Z, [x10]\n"
       "ld1h { z7.h }, p4/Z, [x10, #1, MUL VL]\n"
@@ -175,19 +173,19 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "10:"  // Height 1: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "fmla z8.h, p4/M, z6.h, z0.h\n"
-      "ld1h { z17.h }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z17.h }, p4/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p4/Z, [x10, #3, MUL VL]\n"
       "add x28, x28, #0x1\n"
-      "addvl x10, x10, #4\n"
       "cmp x28, x20\n"
       "fmla z10.h, p4/M, z17.h, z0.h\n"
       "fmla z11.h, p4/M, z16.h, z0.h\n"
+      "addvl x10, x10, #4\n"
       "bne 6b\n"
       "tbz %x[flags], #1, 11f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z17.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z17.h }, p4/Z, [x21]\n"
       "ld1rh { z16.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z17.h\n"
       "fmin z9.h, p4/M, z9.h, z17.h\n"
@@ -209,10 +207,10 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "bgt 2b\n"
       "b 74f\n"
       "13:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "14:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.h, x20, x11\n"
@@ -225,22 +223,22 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "cbz x12, 15f\n"
       "ld1h { z8.h }, p4/Z, [x12]\n"
       "ld1h { z9.h }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x12, x12, #4\n"
       "b 17f\n"
       "15:"  // Height 2: no bias
       "tbz %x[flags], #0, 16f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x9]\n"
       "ld1h { z9.h }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1h { z11.h }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x20]\n"
       "ld1h { z13.h }, p2/Z, [x20, #1, MUL VL]\n"
       "ld1h { z14.h }, p1/Z, [x20, #2, MUL VL]\n"
@@ -259,8 +257,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "18:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 19f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -295,8 +293,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z10.h, p4/M, z17.h, z0.h\n"
       "fmla z14.h, p4/M, z17.h, z1.h\n"
       "fmla z11.h, p4/M, z16.h, z0.h\n"
-      "ld1rh { z0.h }, p4/Z, [x26]\n"
       "fmla z15.h, p4/M, z16.h, z1.h\n"
+      "ld1rh { z0.h }, p4/Z, [x26]\n"
       "ld1rh { z1.h }, p4/Z, [x25]\n"
       "ld1h { z6.h }, p4/Z, [x10]\n"
       "ld1h { z7.h }, p4/Z, [x10, #1, MUL VL]\n"
@@ -310,19 +308,19 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "ld1h { z16.h }, p4/Z, [x10, #3, MUL VL]\n"
       "add x28, x28, #0x1\n"
-      "addvl x10, x10, #4\n"
       "cmp x28, x20\n"
       "fmla z10.h, p4/M, z17.h, z0.h\n"
       "fmla z14.h, p4/M, z17.h, z1.h\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, p4/M, z16.h, z0.h\n"
       "fmla z15.h, p4/M, z16.h, z1.h\n"
       "bne 18b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "tbz %x[flags], #1, 23f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z17.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z17.h }, p4/Z, [x21]\n"
       "ld1rh { z16.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z17.h\n"
       "fmin z9.h, p4/M, z9.h, z17.h\n"
@@ -346,20 +344,20 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
       "24:"  // Height 2: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 14b\n"
       "b 74f\n"
       "25:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "26:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.h, x20, x11\n"
@@ -372,27 +370,27 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "cbz x12, 27f\n"
       "ld1h { z8.h }, p4/Z, [x12]\n"
       "ld1h { z9.h }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 29f\n"
       "27:"  // Height 3: no bias
       "tbz %x[flags], #0, 28f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #1\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x9]\n"
       "ld1h { z9.h }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1h { z11.h }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x21]\n"
       "ld1h { z13.h }, p2/Z, [x21, #1, MUL VL]\n"
       "ld1h { z14.h }, p1/Z, [x21, #2, MUL VL]\n"
@@ -419,8 +417,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "30:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 31f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -451,8 +449,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "add x26, x26, #0x2\n"
       "subs x27, x27, #0x1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
-      "ld1h { z21.h }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z21.h }, p4/Z, [x10, #2, MUL VL]\n"
       "add x25, x25, #0x2\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
@@ -464,11 +462,11 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z18.h, p4/M, z21.h, z2.h\n"
       "fmla z11.h, p4/M, z20.h, z0.h\n"
       "ld1rh { z0.h }, p4/Z, [x26]\n"
+      "ld1h { z6.h }, p4/Z, [x10]\n"
       "fmla z15.h, p4/M, z20.h, z1.h\n"
-      "ld1rh { z1.h }, p4/Z, [x25]\n"
       "fmla z19.h, p4/M, z20.h, z2.h\n"
+      "ld1rh { z1.h }, p4/Z, [x25]\n"
       "ld1rh { z2.h }, p4/Z, [x24]\n"
-      "ld1h { z6.h }, p4/Z, [x10]\n"
       "ld1h { z7.h }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 33b\n"
       "34:"  // Height 3: Multiply loop: Main loop skip
@@ -477,13 +475,13 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z12.h, p4/M, z6.h, z1.h\n"
       "add x28, x28, #0x1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
-      "ld1h { z21.h }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z21.h }, p4/Z, [x10, #2, MUL VL]\n"
+      "cmp x28, x20\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
       "ld1h { z20.h }, p4/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
-      "cmp x28, x20\n"
       "fmla z10.h, p4/M, z21.h, z0.h\n"
       "fmla z14.h, p4/M, z21.h, z1.h\n"
       "fmla z18.h, p4/M, z21.h, z2.h\n"
@@ -492,12 +490,12 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z19.h, p4/M, z20.h, z2.h\n"
       "bne 30b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "tbz %x[flags], #1, 35f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z21.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z21.h }, p4/Z, [x21]\n"
       "ld1rh { z20.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z21.h\n"
       "fmin z9.h, p4/M, z9.h, z21.h\n"
@@ -529,24 +527,24 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p3, [x25]\n"
-      "st1h { z17.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p1, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p3, [x24]\n"
+      "st1h { z17.h }, p2, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p0, [x24, #3, MUL VL]\n"
       "36:"  // Height 3: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 26b\n"
       "b 74f\n"
       "37:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "38:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.h, x20, x11\n"
@@ -559,18 +557,18 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "cbz x12, 39f\n"
       "ld1h { z8.h }, p4/Z, [x12]\n"
       "ld1h { z9.h }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -578,13 +576,13 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "39:"  // Height 4: no bias
       "tbz %x[flags], #0, 40f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x9]\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z9.h }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1h { z11.h }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #1\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x22]\n"
       "ld1h { z13.h }, p2/Z, [x22, #1, MUL VL]\n"
       "ld1h { z14.h }, p1/Z, [x22, #2, MUL VL]\n"
@@ -619,8 +617,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "42:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 43f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -670,6 +668,7 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z14.h, p4/M, z25.h, z1.h\n"
       "fmla z18.h, p4/M, z25.h, z2.h\n"
       "fmla z22.h, p4/M, z25.h, z3.h\n"
+      "ld1h { z6.h }, p4/Z, [x10]\n"
       "fmla z11.h, p4/M, z24.h, z0.h\n"
       "fmla z15.h, p4/M, z24.h, z1.h\n"
       "ld1rh { z0.h }, p4/Z, [x26]\n"
@@ -678,7 +677,6 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z23.h, p4/M, z24.h, z3.h\n"
       "ld1rh { z2.h }, p4/Z, [x24]\n"
       "ld1rh { z3.h }, p4/Z, [x23]\n"
-      "ld1h { z6.h }, p4/Z, [x10]\n"
       "ld1h { z7.h }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 45b\n"
       "46:"  // Height 4: Multiply loop: Main loop skip
@@ -689,15 +687,15 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z16.h, p4/M, z6.h, z2.h\n"
       "fmla z20.h, p4/M, z6.h, z3.h\n"
       "ld1h { z25.h }, p4/Z, [x10, #2, MUL VL]\n"
+      "cmp x28, x20\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
       "fmla z21.h, p4/M, z7.h, z3.h\n"
       "ld1h { z24.h }, p4/Z, [x10, #3, MUL VL]\n"
-      "cmp x28, x20\n"
+      "addvl x10, x10, #4\n"
       "fmla z10.h, p4/M, z25.h, z0.h\n"
       "fmla z14.h, p4/M, z25.h, z1.h\n"
-      "addvl x10, x10, #4\n"
       "fmla z18.h, p4/M, z25.h, z2.h\n"
       "fmla z22.h, p4/M, z25.h, z3.h\n"
       "fmla z11.h, p4/M, z24.h, z0.h\n"
@@ -706,13 +704,13 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z23.h, p4/M, z24.h, z3.h\n"
       "bne 42b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "tbz %x[flags], #1, 47f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z25.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z25.h }, p4/Z, [x21]\n"
       "ld1rh { z24.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z25.h\n"
       "fmin z9.h, p4/M, z9.h, z25.h\n"
@@ -752,28 +750,28 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p3, [x25]\n"
-      "st1h { z17.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p1, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p0, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p3, [x24]\n"
-      "st1h { z21.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p1, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p0, [x24, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p3, [x24]\n"
+      "st1h { z17.h }, p2, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p0, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p3, [x23]\n"
+      "st1h { z21.h }, p2, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p1, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p0, [x23, #3, MUL VL]\n"
       "48:"  // Height 4: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 38b\n"
       "b 74f\n"
       "49:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "50:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.h, x20, x11\n"
@@ -786,18 +784,18 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "cbz x12, 51f\n"
       "ld1h { z8.h }, p4/Z, [x12]\n"
       "ld1h { z9.h }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -809,16 +807,16 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "51:"  // Height 5: no bias
       "tbz %x[flags], #0, 52f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x9]\n"
+      "add x21, x22, x20, LSL #1\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z9.h }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1h { z11.h }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #1\n"
-      "add x22, x23, x20, LSL #1\n"
-      "add x21, x22, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x23]\n"
       "ld1h { z13.h }, p2/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z14.h }, p1/Z, [x23, #2, MUL VL]\n"
       "ld1h { z15.h }, p0/Z, [x23, #3, MUL VL]\n"
       "ld1h { z16.h }, p3/Z, [x22]\n"
@@ -859,8 +857,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "54:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 55f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -903,8 +901,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "add x25, x25, #0x2\n"
       "add x24, x24, #0x2\n"
       "fmla z24.h, p4/M, z6.h, z4.h\n"
-      "ld1h { z29.h }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z29.h }, p4/Z, [x10, #2, MUL VL]\n"
       "add x23, x23, #0x2\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
@@ -922,12 +920,12 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "ld1rh { z0.h }, p4/Z, [x26]\n"
       "ld1h { z6.h }, p4/Z, [x10]\n"
       "fmla z15.h, p4/M, z28.h, z1.h\n"
-      "ld1rh { z1.h }, p4/Z, [x25]\n"
       "fmla z19.h, p4/M, z28.h, z2.h\n"
+      "ld1rh { z1.h }, p4/Z, [x25]\n"
       "ld1rh { z2.h }, p4/Z, [x24]\n"
       "fmla z23.h, p4/M, z28.h, z3.h\n"
-      "ld1rh { z3.h }, p4/Z, [x23]\n"
       "fmla z27.h, p4/M, z28.h, z4.h\n"
+      "ld1rh { z3.h }, p4/Z, [x23]\n"
       "ld1rh { z4.h }, p4/Z, [x22]\n"
       "ld1h { z7.h }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 57b\n"
@@ -938,12 +936,12 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
       "fmla z20.h, p4/M, z6.h, z3.h\n"
+      "cmp x28, x20\n"
       "fmla z24.h, p4/M, z6.h, z4.h\n"
-      "ld1h { z29.h }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
+      "ld1h { z29.h }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
-      "cmp x28, x20\n"
       "fmla z21.h, p4/M, z7.h, z3.h\n"
       "fmla z25.h, p4/M, z7.h, z4.h\n"
       "ld1h { z28.h }, p4/Z, [x10, #3, MUL VL]\n"
@@ -960,14 +958,14 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z27.h, p4/M, z28.h, z4.h\n"
       "bne 54b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "tbz %x[flags], #1, 59f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z29.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z29.h }, p4/Z, [x21]\n"
       "ld1rh { z28.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z29.h\n"
       "fmin z9.h, p4/M, z9.h, z29.h\n"
@@ -1015,22 +1013,22 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p3, [x25]\n"
-      "st1h { z17.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p1, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p0, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p3, [x24]\n"
-      "st1h { z21.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p1, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p0, [x24, #3, MUL VL]\n"
-      "st1h { z24.h }, p3, [x23]\n"
-      "st1h { z25.h }, p2, [x23, #1, MUL VL]\n"
-      "st1h { z26.h }, p1, [x23, #2, MUL VL]\n"
-      "st1h { z27.h }, p0, [x23, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p3, [x24]\n"
+      "st1h { z17.h }, p2, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p0, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p3, [x23]\n"
+      "st1h { z21.h }, p2, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p1, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p0, [x23, #3, MUL VL]\n"
+      "st1h { z24.h }, p3, [x22]\n"
+      "st1h { z25.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z26.h }, p1, [x22, #2, MUL VL]\n"
+      "st1h { z27.h }, p0, [x22, #3, MUL VL]\n"
       "60:"  // Height 5: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1038,13 +1036,12 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "b 74f\n"
       "61:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0xc\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "62:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.h, x20, x11\n"
@@ -1057,18 +1054,18 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "cbz x12, 63f\n"
       "ld1h { z8.h }, p4/Z, [x12]\n"
       "ld1h { z9.h }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1084,17 +1081,17 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "63:"  // Height 6: no bias
       "tbz %x[flags], #0, 64f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "ld1h { z8.h }, p3/Z, [x9]\n"
+      "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "ld1h { z9.h }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p1/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z11.h }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #1\n"
-      "add x23, x24, x20, LSL #1\n"
-      "add x22, x23, x20, LSL #1\n"
       "ld1h { z12.h }, p3/Z, [x24]\n"
       "ld1h { z13.h }, p2/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z14.h }, p1/Z, [x24, #2, MUL VL]\n"
       "ld1h { z15.h }, p0/Z, [x24, #3, MUL VL]\n"
       "ld1h { z16.h }, p3/Z, [x23]\n"
@@ -1143,8 +1140,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "66:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 67f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1232,12 +1229,12 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "fmla z16.h, p4/M, z6.h, z2.h\n"
       "fmla z20.h, p4/M, z6.h, z3.h\n"
+      "cmp x28, x20\n"
       "fmla z24.h, p4/M, z6.h, z4.h\n"
       "fmla z28.h, p4/M, z6.h, z5.h\n"
       "ld1h { z6.h }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, p4/M, z7.h, z0.h\n"
       "fmla z13.h, p4/M, z7.h, z1.h\n"
-      "cmp x28, x20\n"
       "fmla z17.h, p4/M, z7.h, z2.h\n"
       "fmla z21.h, p4/M, z7.h, z3.h\n"
       "fmla z25.h, p4/M, z7.h, z4.h\n"
@@ -1258,15 +1255,15 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "fmla z31.h, p4/M, z7.h, z5.h\n"
       "bne 66b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "tbz %x[flags], #1, 71f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z1.h }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z1.h }, p4/Z, [x21]\n"
       "ld1rh { z0.h }, p4/Z, [x20]\n"
       "fmin z8.h, p4/M, z8.h, z1.h\n"
       "fmin z9.h, p4/M, z9.h, z1.h\n"
@@ -1322,26 +1319,26 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "st1h { z10.h }, p1, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p3, [x26]\n"
-      "st1h { z13.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p1, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p0, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p3, [x25]\n"
-      "st1h { z17.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p1, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p0, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p3, [x24]\n"
-      "st1h { z21.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p1, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p0, [x24, #3, MUL VL]\n"
-      "st1h { z24.h }, p3, [x23]\n"
-      "st1h { z25.h }, p2, [x23, #1, MUL VL]\n"
-      "st1h { z26.h }, p1, [x23, #2, MUL VL]\n"
-      "st1h { z27.h }, p0, [x23, #3, MUL VL]\n"
-      "st1h { z28.h }, p3, [x22]\n"
-      "st1h { z29.h }, p2, [x22, #1, MUL VL]\n"
-      "st1h { z30.h }, p1, [x22, #2, MUL VL]\n"
-      "st1h { z31.h }, p0, [x22, #3, MUL VL]\n"
+      "st1h { z12.h }, p3, [x25]\n"
+      "st1h { z13.h }, p2, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p1, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p0, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p3, [x24]\n"
+      "st1h { z17.h }, p2, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p0, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p3, [x23]\n"
+      "st1h { z21.h }, p2, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p1, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p0, [x23, #3, MUL VL]\n"
+      "st1h { z24.h }, p3, [x22]\n"
+      "st1h { z25.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z26.h }, p1, [x22, #2, MUL VL]\n"
+      "st1h { z27.h }, p0, [x22, #3, MUL VL]\n"
+      "st1h { z28.h }, p3, [x21]\n"
+      "st1h { z29.h }, p2, [x21, #1, MUL VL]\n"
+      "st1h { z30.h }, p1, [x21, #2, MUL VL]\n"
+      "st1h { z31.h }, p0, [x21, #3, MUL VL]\n"
       "72:"  // Height 6: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1358,8 +1355,8 @@ void sve_hybrid_fp16_mla_6x4VL_a64fx (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "74:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp
index 231472bcd0..64c4dfcd0e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp
@@ -47,19 +47,18 @@ void sve_hybrid_fp16_mla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const __fp16 *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void sve_hybrid_fp16_mla_6x4VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -104,10 +102,10 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cmp %x[M], #0x2\n"
       "bgt 27f\n"
       "beq 14f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.h, x20, x11\n"
@@ -140,8 +138,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -157,56 +155,53 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "fmla z8.h, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10]\n"
+      "fmla z8.h, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z10.h, z16.h, z0.h[0]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "fmla z8.h, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
-      "fmla z8.h, z17.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[1]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #6, MUL VL]\n"
+      "fmla z10.h, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      "fmla z10.h, z17.h, z0.h[1]\n"
       "fmla z11.h, z16.h, z0.h[1]\n"
       "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-5, MUL VL]\n"
       "fmla z10.h, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-3, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-1, MUL VL]\n"
       "fmla z10.h, z17.h, z0.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
       "fmla z11.h, z16.h, z0.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[4]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[4]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z10.h, z17.h, z0.h[4]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[4]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[5]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[5]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
       "fmla z10.h, z17.h, z0.h[5]\n"
@@ -214,118 +209,121 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[6]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[6]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-5, MUL VL]\n"
       "fmla z10.h, z17.h, z0.h[6]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.h, z16.h, z0.h[6]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-3, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[7]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[7]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-1, MUL VL]\n"
+      "sub x27, x27, #0x8\n"
+      "cmp x27, #0x8\n"
       "fmla z10.h, z17.h, z0.h[7]\n"
       "fmla z11.h, z16.h, z0.h[7]\n"
+      "add x26, x26, #0x10\n"
       "bgt 9b\n"
       "10:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
-      "fmla z8.h, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z16.h }, p5/Z, [x10]\n"
+      "fmla z8.h, z16.h, z0.h[0]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z10.h, z17.h, z0.h[0]\n"
       "fmla z11.h, z16.h, z0.h[0]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[1]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[1]\n"
       "fmla z11.h, z16.h, z0.h[1]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[2]\n"
       "fmla z11.h, z16.h, z0.h[2]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[3]\n"
       "fmla z11.h, z16.h, z0.h[3]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[4]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[4]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[4]\n"
       "fmla z11.h, z16.h, z0.h[4]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[5]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[5]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[5]\n"
       "fmla z11.h, z16.h, z0.h[5]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[6]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[6]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[6]\n"
       "fmla z11.h, z16.h, z0.h[6]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[7]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[7]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z10.h, z17.h, z0.h[7]\n"
       "fmla z11.h, z16.h, z0.h[7]\n"
+      "addvl x10, x10, #4\n"
       "11:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 6b\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z17.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z17.h }, p5/Z, [x21]\n"
       "ld1rh { z16.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z17.h\n"
       "fmin z9.h, p5/M, z9.h, z17.h\n"
@@ -347,10 +345,10 @@ void sve_hybrid_fp16_mla_6x4VL (
       "bgt 2b\n"
       "b 80f\n"
       "14:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.h, x20, x11\n"
@@ -363,22 +361,22 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cbz x12, 16f\n"
       "ld1h { z8.h }, p5/Z, [x12]\n"
       "ld1h { z9.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x12, x12, #4\n"
       "b 18f\n"
       "16:"  // Height 2: no bias
       "tbz %x[flags], #0, 17f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x9]\n"
       "ld1h { z9.h }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1h { z11.h }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x20]\n"
       "ld1h { z13.h }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1h { z14.h }, p2/Z, [x20, #2, MUL VL]\n"
@@ -397,8 +395,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "19:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -417,38 +415,38 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 23f\n"
       "22:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z1.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z17.h, z1.h[0]\n"
       "fmla z12.h, z17.h, z0.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z1.h[0]\n"
       "fmla z13.h, z16.h, z0.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z10.h, z17.h, z1.h[0]\n"
       "fmla z14.h, z17.h, z0.h[0]\n"
       "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "cmp x27, #0x8\n"
       "fmla z11.h, z16.h, z1.h[0]\n"
       "fmla z15.h, z16.h, z0.h[0]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       "fmla z8.h, z17.h, z1.h[1]\n"
       "fmla z12.h, z17.h, z0.h[1]\n"
       "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       "fmla z9.h, z16.h, z1.h[1]\n"
       "fmla z13.h, z16.h, z0.h[1]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
       "fmla z10.h, z17.h, z1.h[1]\n"
       "fmla z14.h, z17.h, z0.h[1]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z11.h, z16.h, z1.h[1]\n"
       "fmla z15.h, z16.h, z0.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.h, z17.h, z1.h[2]\n"
       "fmla z12.h, z17.h, z0.h[2]\n"
@@ -495,9 +493,9 @@ void sve_hybrid_fp16_mla_6x4VL (
       "addvl x10, x10, #16\n"
       "fmla z10.h, z17.h, z1.h[5]\n"
       "fmla z14.h, z17.h, z0.h[5]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z11.h, z16.h, z1.h[5]\n"
       "fmla z15.h, z16.h, z0.h[5]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.h, z17.h, z1.h[6]\n"
       "fmla z12.h, z17.h, z0.h[6]\n"
@@ -524,110 +522,110 @@ void sve_hybrid_fp16_mla_6x4VL (
       "bgt 22b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[0]\n"
       "fmla z12.h, z17.h, z1.h[0]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[0]\n"
       "fmla z13.h, z16.h, z1.h[0]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z10.h, z17.h, z0.h[0]\n"
       "fmla z14.h, z17.h, z1.h[0]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, z16.h, z0.h[0]\n"
       "fmla z15.h, z16.h, z1.h[0]\n"
       "ble 24f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[1]\n"
       "fmla z12.h, z17.h, z1.h[1]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[1]\n"
       "fmla z13.h, z16.h, z1.h[1]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[1]\n"
       "fmla z14.h, z17.h, z1.h[1]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, z16.h, z0.h[1]\n"
       "fmla z15.h, z16.h, z1.h[1]\n"
       "ble 24f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[2]\n"
       "fmla z12.h, z17.h, z1.h[2]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[2]\n"
       "fmla z13.h, z16.h, z1.h[2]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[2]\n"
       "fmla z14.h, z17.h, z1.h[2]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, z16.h, z0.h[2]\n"
       "fmla z15.h, z16.h, z1.h[2]\n"
       "ble 24f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[3]\n"
       "fmla z12.h, z17.h, z1.h[3]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[3]\n"
       "fmla z13.h, z16.h, z1.h[3]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[3]\n"
       "fmla z14.h, z17.h, z1.h[3]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, z16.h, z0.h[3]\n"
       "fmla z15.h, z16.h, z1.h[3]\n"
       "ble 24f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[4]\n"
       "fmla z12.h, z17.h, z1.h[4]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[4]\n"
       "fmla z13.h, z16.h, z1.h[4]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[4]\n"
       "fmla z14.h, z17.h, z1.h[4]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, z16.h, z0.h[4]\n"
       "fmla z15.h, z16.h, z1.h[4]\n"
       "ble 24f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[5]\n"
       "fmla z12.h, z17.h, z1.h[5]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[5]\n"
       "fmla z13.h, z16.h, z1.h[5]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[5]\n"
       "fmla z14.h, z17.h, z1.h[5]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, z16.h, z0.h[5]\n"
       "fmla z15.h, z16.h, z1.h[5]\n"
       "ble 24f\n"
       "ld1h { z17.h }, p5/Z, [x10]\n"
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z17.h, z0.h[6]\n"
       "fmla z12.h, z17.h, z1.h[6]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[6]\n"
       "fmla z13.h, z16.h, z1.h[6]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.h, z17.h, z0.h[6]\n"
       "fmla z14.h, z17.h, z1.h[6]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, z16.h, z0.h[6]\n"
       "fmla z15.h, z16.h, z1.h[6]\n"
       "ble 24f\n"
@@ -635,13 +633,13 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z17.h, z0.h[7]\n"
       "fmla z12.h, z17.h, z1.h[7]\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z16.h, z0.h[7]\n"
       "fmla z13.h, z16.h, z1.h[7]\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z10.h, z17.h, z0.h[7]\n"
       "fmla z14.h, z17.h, z1.h[7]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.h, z16.h, z0.h[7]\n"
       "fmla z15.h, z16.h, z1.h[7]\n"
       "24:"  // Height 2: Multiply loop: multiply skip
@@ -650,11 +648,11 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 19b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z17.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z17.h }, p5/Z, [x21]\n"
       "ld1rh { z16.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z17.h\n"
       "fmin z9.h, p5/M, z9.h, z17.h\n"
@@ -678,20 +676,20 @@ void sve_hybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
       "26:"  // Height 2: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 15b\n"
       "b 80f\n"
       "27:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.h, x20, x11\n"
@@ -704,27 +702,27 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cbz x12, 29f\n"
       "ld1h { z8.h }, p5/Z, [x12]\n"
       "ld1h { z9.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 31f\n"
       "29:"  // Height 3: no bias
       "tbz %x[flags], #0, 30f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #1\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x9]\n"
       "ld1h { z9.h }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1h { z11.h }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x21]\n"
       "ld1h { z13.h }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1h { z14.h }, p2/Z, [x21, #2, MUL VL]\n"
@@ -751,8 +749,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -774,37 +772,37 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 36f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z2.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z0.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
       "fmla z8.h, z21.h, z2.h[0]\n"
       "fmla z12.h, z21.h, z1.h[0]\n"
-      "fmla z9.h, z20.h, z2.h[0]\n"
-      "fmla z13.h, z20.h, z1.h[0]\n"
+      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z16.h, z21.h, z0.h[0]\n"
+      "fmla z9.h, z20.h, z2.h[0]\n"
       "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z13.h, z20.h, z1.h[0]\n"
       "fmla z17.h, z20.h, z0.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
+      "cmp x27, #0x8\n"
       "fmla z10.h, z21.h, z2.h[0]\n"
       "fmla z14.h, z21.h, z1.h[0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "fmla z18.h, z21.h, z0.h[0]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[0]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       "fmla z15.h, z20.h, z1.h[0]\n"
       "fmla z19.h, z20.h, z0.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[1]\n"
       "fmla z12.h, z21.h, z1.h[1]\n"
       "fmla z16.h, z21.h, z0.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[1]\n"
       "fmla z17.h, z20.h, z0.h[1]\n"
       "ld1h { z20.h }, p5/Z, [x10, #7, MUL VL]\n"
@@ -813,63 +811,63 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z14.h, z21.h, z1.h[1]\n"
       "fmla z18.h, z21.h, z0.h[1]\n"
       "fmla z11.h, z20.h, z2.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[1]\n"
       "fmla z19.h, z20.h, z0.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[2]\n"
       "fmla z12.h, z21.h, z1.h[2]\n"
       "fmla z16.h, z21.h, z0.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[2]\n"
       "fmla z17.h, z20.h, z0.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-5, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[2]\n"
       "fmla z14.h, z21.h, z1.h[2]\n"
       "fmla z18.h, z21.h, z0.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[2]\n"
       "fmla z19.h, z20.h, z0.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-3, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[3]\n"
       "fmla z12.h, z21.h, z1.h[3]\n"
       "fmla z16.h, z21.h, z0.h[3]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[3]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[3]\n"
       "fmla z17.h, z20.h, z0.h[3]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-1, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[3]\n"
       "fmla z14.h, z21.h, z1.h[3]\n"
       "fmla z18.h, z21.h, z0.h[3]\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
       "fmla z11.h, z20.h, z2.h[3]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
       "fmla z15.h, z20.h, z1.h[3]\n"
       "fmla z19.h, z20.h, z0.h[3]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[4]\n"
       "fmla z12.h, z21.h, z1.h[4]\n"
       "fmla z16.h, z21.h, z0.h[4]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[4]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[4]\n"
       "fmla z17.h, z20.h, z0.h[4]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[4]\n"
       "fmla z14.h, z21.h, z1.h[4]\n"
       "fmla z18.h, z21.h, z0.h[4]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[4]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[4]\n"
       "fmla z19.h, z20.h, z0.h[4]\n"
       "ld1h { z20.h }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[5]\n"
       "fmla z12.h, z21.h, z1.h[5]\n"
       "fmla z16.h, z21.h, z0.h[5]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[5]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[5]\n"
       "fmla z17.h, z20.h, z0.h[5]\n"
       "ld1h { z20.h }, p5/Z, [x10, #7, MUL VL]\n"
@@ -878,31 +876,31 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z14.h, z21.h, z1.h[5]\n"
       "fmla z18.h, z21.h, z0.h[5]\n"
       "fmla z11.h, z20.h, z2.h[5]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[5]\n"
       "fmla z19.h, z20.h, z0.h[5]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[6]\n"
       "fmla z12.h, z21.h, z1.h[6]\n"
       "fmla z16.h, z21.h, z0.h[6]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[6]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[6]\n"
       "fmla z17.h, z20.h, z0.h[6]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-5, MUL VL]\n"
       "fmla z10.h, z21.h, z2.h[6]\n"
       "fmla z14.h, z21.h, z1.h[6]\n"
       "fmla z18.h, z21.h, z0.h[6]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.h, z20.h, z2.h[6]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z15.h, z20.h, z1.h[6]\n"
       "fmla z19.h, z20.h, z0.h[6]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-3, MUL VL]\n"
       "fmla z8.h, z21.h, z2.h[7]\n"
       "fmla z12.h, z21.h, z1.h[7]\n"
       "fmla z16.h, z21.h, z0.h[7]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.h, z20.h, z2.h[7]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[7]\n"
       "fmla z17.h, z20.h, z0.h[7]\n"
       "ld1h { z20.h }, p5/Z, [x10, #-1, MUL VL]\n"
@@ -915,18 +913,18 @@ void sve_hybrid_fp16_mla_6x4VL (
       "bgt 35b\n"
       "36:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z21.h }, p5/Z, [x10]\n"
-      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
+      "ld1h { z21.h }, p5/Z, [x10]\n"
       "fmla z8.h, z21.h, z0.h[0]\n"
       "fmla z12.h, z21.h, z1.h[0]\n"
-      "fmla z9.h, z20.h, z0.h[0]\n"
-      "fmla z13.h, z20.h, z1.h[0]\n"
+      "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z16.h, z21.h, z2.h[0]\n"
+      "fmla z9.h, z20.h, z0.h[0]\n"
       "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z13.h, z20.h, z1.h[0]\n"
       "fmla z17.h, z20.h, z2.h[0]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
@@ -939,12 +937,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 37f\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z21.h, z0.h[1]\n"
       "fmla z12.h, z21.h, z1.h[1]\n"
       "fmla z16.h, z21.h, z2.h[1]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z0.h[1]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[1]\n"
       "fmla z17.h, z20.h, z2.h[1]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -958,12 +956,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 37f\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z21.h, z0.h[2]\n"
       "fmla z12.h, z21.h, z1.h[2]\n"
       "fmla z16.h, z21.h, z2.h[2]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z0.h[2]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[2]\n"
       "fmla z17.h, z20.h, z2.h[2]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -977,12 +975,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 37f\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z21.h, z0.h[3]\n"
       "fmla z12.h, z21.h, z1.h[3]\n"
       "fmla z16.h, z21.h, z2.h[3]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z0.h[3]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[3]\n"
       "fmla z17.h, z20.h, z2.h[3]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -996,12 +994,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 37f\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z21.h, z0.h[4]\n"
       "fmla z12.h, z21.h, z1.h[4]\n"
       "fmla z16.h, z21.h, z2.h[4]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z0.h[4]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[4]\n"
       "fmla z17.h, z20.h, z2.h[4]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1015,12 +1013,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 37f\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z21.h, z0.h[5]\n"
       "fmla z12.h, z21.h, z1.h[5]\n"
       "fmla z16.h, z21.h, z2.h[5]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z0.h[5]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[5]\n"
       "fmla z17.h, z20.h, z2.h[5]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1034,12 +1032,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 37f\n"
       "ld1h { z21.h }, p5/Z, [x10]\n"
       "ld1h { z20.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z21.h, z0.h[6]\n"
       "fmla z12.h, z21.h, z1.h[6]\n"
       "fmla z16.h, z21.h, z2.h[6]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z0.h[6]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.h, z20.h, z1.h[6]\n"
       "fmla z17.h, z20.h, z2.h[6]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1056,8 +1054,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z8.h, z21.h, z0.h[7]\n"
       "fmla z12.h, z21.h, z1.h[7]\n"
       "fmla z16.h, z21.h, z2.h[7]\n"
-      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z20.h, z0.h[7]\n"
+      "ld1h { z21.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z20.h, z1.h[7]\n"
       "fmla z17.h, z20.h, z2.h[7]\n"
       "ld1h { z20.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1074,12 +1072,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
+      "add x24, x25, x20, LSL #1\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z21.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z21.h }, p5/Z, [x21]\n"
       "ld1rh { z20.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z21.h\n"
       "fmin z9.h, p5/M, z9.h, z21.h\n"
@@ -1111,24 +1109,24 @@ void sve_hybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p4, [x25]\n"
-      "st1h { z17.h }, p3, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p4, [x24]\n"
+      "st1h { z17.h }, p3, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p1, [x24, #3, MUL VL]\n"
       "39:"  // Height 3: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 28b\n"
       "b 80f\n"
       "40:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.h, x20, x11\n"
@@ -1141,18 +1139,18 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cbz x12, 42f\n"
       "ld1h { z8.h }, p5/Z, [x12]\n"
       "ld1h { z9.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1160,13 +1158,13 @@ void sve_hybrid_fp16_mla_6x4VL (
       "42:"  // Height 4: no bias
       "tbz %x[flags], #0, 43f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x9]\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z9.h }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1h { z11.h }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #1\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x22]\n"
       "ld1h { z13.h }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1h { z14.h }, p2/Z, [x22, #2, MUL VL]\n"
@@ -1201,8 +1199,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "45:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1227,25 +1225,25 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 49f\n"
       "48:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x10]\n"
-      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z3.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z2.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z1.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z25.h, z3.h[0]\n"
       "fmla z12.h, z25.h, z2.h[0]\n"
-      "fmla z9.h, z24.h, z3.h[0]\n"
-      "fmla z13.h, z24.h, z2.h[0]\n"
       "fmla z16.h, z25.h, z1.h[0]\n"
       "fmla z20.h, z25.h, z0.h[0]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      "fmla z9.h, z24.h, z3.h[0]\n"
+      "fmla z13.h, z24.h, z2.h[0]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "fmla z17.h, z24.h, z1.h[0]\n"
       "fmla z21.h, z24.h, z0.h[0]\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1274,9 +1272,9 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z14.h, z25.h, z2.h[1]\n"
       "fmla z18.h, z25.h, z1.h[1]\n"
       "fmla z22.h, z25.h, z0.h[1]\n"
+      "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z11.h, z24.h, z3.h[1]\n"
       "fmla z15.h, z24.h, z2.h[1]\n"
-      "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z19.h, z24.h, z1.h[1]\n"
       "fmla z23.h, z24.h, z0.h[1]\n"
       "ld1h { z24.h }, p5/Z, [x10, #-7, MUL VL]\n"
@@ -1355,9 +1353,9 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z14.h, z25.h, z2.h[5]\n"
       "fmla z18.h, z25.h, z1.h[5]\n"
       "fmla z22.h, z25.h, z0.h[5]\n"
+      "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z11.h, z24.h, z3.h[5]\n"
       "fmla z15.h, z24.h, z2.h[5]\n"
-      "ld1h { z25.h }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z19.h, z24.h, z1.h[5]\n"
       "fmla z23.h, z24.h, z0.h[5]\n"
       "ld1h { z24.h }, p5/Z, [x10, #-7, MUL VL]\n"
@@ -1402,20 +1400,20 @@ void sve_hybrid_fp16_mla_6x4VL (
       "bgt 48b\n"
       "49:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x10]\n"
-      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z25.h, z0.h[0]\n"
       "fmla z12.h, z25.h, z1.h[0]\n"
-      "fmla z9.h, z24.h, z0.h[0]\n"
-      "fmla z13.h, z24.h, z1.h[0]\n"
       "fmla z16.h, z25.h, z2.h[0]\n"
       "fmla z20.h, z25.h, z3.h[0]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.h, z24.h, z0.h[0]\n"
+      "fmla z13.h, z24.h, z1.h[0]\n"
       "fmla z17.h, z24.h, z2.h[0]\n"
       "fmla z21.h, z24.h, z3.h[0]\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1431,12 +1429,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z25.h, z0.h[1]\n"
       "fmla z12.h, z25.h, z1.h[1]\n"
       "fmla z16.h, z25.h, z2.h[1]\n"
       "fmla z20.h, z25.h, z3.h[1]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[1]\n"
       "fmla z13.h, z24.h, z1.h[1]\n"
       "fmla z17.h, z24.h, z2.h[1]\n"
@@ -1454,12 +1452,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z25.h, z0.h[2]\n"
       "fmla z12.h, z25.h, z1.h[2]\n"
       "fmla z16.h, z25.h, z2.h[2]\n"
       "fmla z20.h, z25.h, z3.h[2]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[2]\n"
       "fmla z13.h, z24.h, z1.h[2]\n"
       "fmla z17.h, z24.h, z2.h[2]\n"
@@ -1477,12 +1475,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z25.h, z0.h[3]\n"
       "fmla z12.h, z25.h, z1.h[3]\n"
       "fmla z16.h, z25.h, z2.h[3]\n"
       "fmla z20.h, z25.h, z3.h[3]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[3]\n"
       "fmla z13.h, z24.h, z1.h[3]\n"
       "fmla z17.h, z24.h, z2.h[3]\n"
@@ -1500,12 +1498,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z25.h, z0.h[4]\n"
       "fmla z12.h, z25.h, z1.h[4]\n"
       "fmla z16.h, z25.h, z2.h[4]\n"
       "fmla z20.h, z25.h, z3.h[4]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[4]\n"
       "fmla z13.h, z24.h, z1.h[4]\n"
       "fmla z17.h, z24.h, z2.h[4]\n"
@@ -1523,12 +1521,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z25.h, z0.h[5]\n"
       "fmla z12.h, z25.h, z1.h[5]\n"
       "fmla z16.h, z25.h, z2.h[5]\n"
       "fmla z20.h, z25.h, z3.h[5]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[5]\n"
       "fmla z13.h, z24.h, z1.h[5]\n"
       "fmla z17.h, z24.h, z2.h[5]\n"
@@ -1546,12 +1544,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 50f\n"
       "ld1h { z25.h }, p5/Z, [x10]\n"
       "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z25.h, z0.h[6]\n"
       "fmla z12.h, z25.h, z1.h[6]\n"
       "fmla z16.h, z25.h, z2.h[6]\n"
       "fmla z20.h, z25.h, z3.h[6]\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.h, z24.h, z0.h[6]\n"
       "fmla z13.h, z24.h, z1.h[6]\n"
       "fmla z17.h, z24.h, z2.h[6]\n"
@@ -1594,13 +1592,13 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z25.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z25.h }, p5/Z, [x21]\n"
       "ld1rh { z24.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z25.h\n"
       "fmin z9.h, p5/M, z9.h, z25.h\n"
@@ -1640,28 +1638,28 @@ void sve_hybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p4, [x25]\n"
-      "st1h { z17.h }, p3, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p1, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p4, [x24]\n"
-      "st1h { z21.h }, p3, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p1, [x24, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p4, [x24]\n"
+      "st1h { z17.h }, p3, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p1, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p4, [x23]\n"
+      "st1h { z21.h }, p3, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p1, [x23, #3, MUL VL]\n"
       "52:"  // Height 4: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 41b\n"
       "b 80f\n"
       "53:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "54:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.h, x20, x11\n"
@@ -1674,18 +1672,18 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cbz x12, 55f\n"
       "ld1h { z8.h }, p5/Z, [x12]\n"
       "ld1h { z9.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1697,16 +1695,16 @@ void sve_hybrid_fp16_mla_6x4VL (
       "55:"  // Height 5: no bias
       "tbz %x[flags], #0, 56f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x9]\n"
+      "add x21, x22, x20, LSL #1\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z9.h }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1h { z11.h }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #1\n"
-      "add x22, x23, x20, LSL #1\n"
-      "add x21, x22, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x23]\n"
       "ld1h { z13.h }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z14.h }, p2/Z, [x23, #2, MUL VL]\n"
       "ld1h { z15.h }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1h { z16.h }, p4/Z, [x22]\n"
@@ -1747,8 +1745,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "58:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 59f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1776,29 +1774,29 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 62f\n"
       "61:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z4.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z3.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z1.h }, p0/Z, [x23]\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "ld1rqh { z0.h }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1h { z29.h }, p5/Z, [x10]\n"
       "fmla z8.h, z29.h, z4.h[0]\n"
       "fmla z12.h, z29.h, z3.h[0]\n"
-      "fmla z9.h, z28.h, z4.h[0]\n"
+      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z16.h, z29.h, z2.h[0]\n"
       "fmla z20.h, z29.h, z1.h[0]\n"
+      "add x25, x25, #0x10\n"
       "fmla z24.h, z29.h, z0.h[0]\n"
-      "fmla z13.h, z28.h, z3.h[0]\n"
+      "fmla z9.h, z28.h, z4.h[0]\n"
       "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      "fmla z13.h, z28.h, z3.h[0]\n"
       "fmla z17.h, z28.h, z2.h[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z21.h, z28.h, z1.h[0]\n"
       "fmla z25.h, z28.h, z0.h[0]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1807,8 +1805,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[0]\n"
       "fmla z22.h, z29.h, z1.h[0]\n"
       "fmla z26.h, z29.h, z0.h[0]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[0]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[0]\n"
       "fmla z19.h, z28.h, z2.h[0]\n"
       "fmla z23.h, z28.h, z1.h[0]\n"
@@ -1819,8 +1817,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[1]\n"
       "fmla z20.h, z29.h, z1.h[1]\n"
       "fmla z24.h, z29.h, z0.h[1]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[1]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[1]\n"
       "fmla z17.h, z28.h, z2.h[1]\n"
       "fmla z21.h, z28.h, z1.h[1]\n"
@@ -1833,8 +1831,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z22.h, z29.h, z1.h[1]\n"
       "fmla z26.h, z29.h, z0.h[1]\n"
       "fmla z11.h, z28.h, z4.h[1]\n"
-      "fmla z15.h, z28.h, z3.h[1]\n"
       "ld1h { z29.h }, p5/Z, [x10, #-8, MUL VL]\n"
+      "fmla z15.h, z28.h, z3.h[1]\n"
       "fmla z19.h, z28.h, z2.h[1]\n"
       "fmla z23.h, z28.h, z1.h[1]\n"
       "fmla z27.h, z28.h, z0.h[1]\n"
@@ -1844,8 +1842,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[2]\n"
       "fmla z20.h, z29.h, z1.h[2]\n"
       "fmla z24.h, z29.h, z0.h[2]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[2]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[2]\n"
       "fmla z17.h, z28.h, z2.h[2]\n"
       "fmla z21.h, z28.h, z1.h[2]\n"
@@ -1856,8 +1854,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[2]\n"
       "fmla z22.h, z29.h, z1.h[2]\n"
       "fmla z26.h, z29.h, z0.h[2]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[2]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[2]\n"
       "fmla z19.h, z28.h, z2.h[2]\n"
       "fmla z23.h, z28.h, z1.h[2]\n"
@@ -1868,8 +1866,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[3]\n"
       "fmla z20.h, z29.h, z1.h[3]\n"
       "fmla z24.h, z29.h, z0.h[3]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[3]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[3]\n"
       "fmla z17.h, z28.h, z2.h[3]\n"
       "fmla z21.h, z28.h, z1.h[3]\n"
@@ -1880,8 +1878,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[3]\n"
       "fmla z22.h, z29.h, z1.h[3]\n"
       "fmla z26.h, z29.h, z0.h[3]\n"
-      "ld1h { z29.h }, p5/Z, [x10]\n"
       "fmla z11.h, z28.h, z4.h[3]\n"
+      "ld1h { z29.h }, p5/Z, [x10]\n"
       "fmla z15.h, z28.h, z3.h[3]\n"
       "fmla z19.h, z28.h, z2.h[3]\n"
       "fmla z23.h, z28.h, z1.h[3]\n"
@@ -1892,8 +1890,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[4]\n"
       "fmla z20.h, z29.h, z1.h[4]\n"
       "fmla z24.h, z29.h, z0.h[4]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[4]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[4]\n"
       "fmla z17.h, z28.h, z2.h[4]\n"
       "fmla z21.h, z28.h, z1.h[4]\n"
@@ -1904,8 +1902,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[4]\n"
       "fmla z22.h, z29.h, z1.h[4]\n"
       "fmla z26.h, z29.h, z0.h[4]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[4]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[4]\n"
       "fmla z19.h, z28.h, z2.h[4]\n"
       "fmla z23.h, z28.h, z1.h[4]\n"
@@ -1916,8 +1914,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[5]\n"
       "fmla z20.h, z29.h, z1.h[5]\n"
       "fmla z24.h, z29.h, z0.h[5]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[5]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[5]\n"
       "fmla z17.h, z28.h, z2.h[5]\n"
       "fmla z21.h, z28.h, z1.h[5]\n"
@@ -1930,8 +1928,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z22.h, z29.h, z1.h[5]\n"
       "fmla z26.h, z29.h, z0.h[5]\n"
       "fmla z11.h, z28.h, z4.h[5]\n"
-      "fmla z15.h, z28.h, z3.h[5]\n"
       "ld1h { z29.h }, p5/Z, [x10, #-8, MUL VL]\n"
+      "fmla z15.h, z28.h, z3.h[5]\n"
       "fmla z19.h, z28.h, z2.h[5]\n"
       "fmla z23.h, z28.h, z1.h[5]\n"
       "fmla z27.h, z28.h, z0.h[5]\n"
@@ -1941,8 +1939,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[6]\n"
       "fmla z20.h, z29.h, z1.h[6]\n"
       "fmla z24.h, z29.h, z0.h[6]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[6]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[6]\n"
       "fmla z17.h, z28.h, z2.h[6]\n"
       "fmla z21.h, z28.h, z1.h[6]\n"
@@ -1953,8 +1951,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z18.h, z29.h, z2.h[6]\n"
       "fmla z22.h, z29.h, z1.h[6]\n"
       "fmla z26.h, z29.h, z0.h[6]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.h, z28.h, z4.h[6]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z15.h, z28.h, z3.h[6]\n"
       "fmla z19.h, z28.h, z2.h[6]\n"
       "fmla z23.h, z28.h, z1.h[6]\n"
@@ -1965,8 +1963,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[7]\n"
       "fmla z20.h, z29.h, z1.h[7]\n"
       "fmla z24.h, z29.h, z0.h[7]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.h, z28.h, z4.h[7]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z13.h, z28.h, z3.h[7]\n"
       "fmla z17.h, z28.h, z2.h[7]\n"
       "fmla z21.h, z28.h, z1.h[7]\n"
@@ -1985,23 +1983,23 @@ void sve_hybrid_fp16_mla_6x4VL (
       "bgt 61b\n"
       "62:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
       "ld1rqh { z4.h }, p0/Z, [x22]\n"
+      "ld1h { z29.h }, p5/Z, [x10]\n"
       "fmla z8.h, z29.h, z0.h[0]\n"
       "fmla z12.h, z29.h, z1.h[0]\n"
-      "fmla z9.h, z28.h, z0.h[0]\n"
-      "fmla z13.h, z28.h, z1.h[0]\n"
+      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z16.h, z29.h, z2.h[0]\n"
       "fmla z20.h, z29.h, z3.h[0]\n"
       "fmla z24.h, z29.h, z4.h[0]\n"
-      "fmla z17.h, z28.h, z2.h[0]\n"
+      "fmla z9.h, z28.h, z0.h[0]\n"
       "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z13.h, z28.h, z1.h[0]\n"
+      "fmla z17.h, z28.h, z2.h[0]\n"
       "fmla z21.h, z28.h, z3.h[0]\n"
       "fmla z25.h, z28.h, z4.h[0]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
@@ -2019,21 +2017,21 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 63f\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
       "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z29.h, z0.h[1]\n"
       "fmla z12.h, z29.h, z1.h[1]\n"
       "fmla z16.h, z29.h, z2.h[1]\n"
       "fmla z20.h, z29.h, z3.h[1]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z29.h, z4.h[1]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z0.h[1]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z1.h[1]\n"
       "fmla z17.h, z28.h, z2.h[1]\n"
       "fmla z21.h, z28.h, z3.h[1]\n"
       "fmla z25.h, z28.h, z4.h[1]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z29.h, z0.h[1]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z29.h, z0.h[1]\n"
       "fmla z14.h, z29.h, z1.h[1]\n"
       "fmla z18.h, z29.h, z2.h[1]\n"
       "fmla z22.h, z29.h, z3.h[1]\n"
@@ -2046,21 +2044,21 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 63f\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
       "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z29.h, z0.h[2]\n"
       "fmla z12.h, z29.h, z1.h[2]\n"
       "fmla z16.h, z29.h, z2.h[2]\n"
       "fmla z20.h, z29.h, z3.h[2]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z29.h, z4.h[2]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z0.h[2]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z1.h[2]\n"
       "fmla z17.h, z28.h, z2.h[2]\n"
       "fmla z21.h, z28.h, z3.h[2]\n"
       "fmla z25.h, z28.h, z4.h[2]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z29.h, z0.h[2]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z29.h, z0.h[2]\n"
       "fmla z14.h, z29.h, z1.h[2]\n"
       "fmla z18.h, z29.h, z2.h[2]\n"
       "fmla z22.h, z29.h, z3.h[2]\n"
@@ -2073,21 +2071,21 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 63f\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
       "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z29.h, z0.h[3]\n"
       "fmla z12.h, z29.h, z1.h[3]\n"
       "fmla z16.h, z29.h, z2.h[3]\n"
       "fmla z20.h, z29.h, z3.h[3]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z29.h, z4.h[3]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z0.h[3]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z1.h[3]\n"
       "fmla z17.h, z28.h, z2.h[3]\n"
       "fmla z21.h, z28.h, z3.h[3]\n"
       "fmla z25.h, z28.h, z4.h[3]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z29.h, z0.h[3]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z29.h, z0.h[3]\n"
       "fmla z14.h, z29.h, z1.h[3]\n"
       "fmla z18.h, z29.h, z2.h[3]\n"
       "fmla z22.h, z29.h, z3.h[3]\n"
@@ -2100,21 +2098,21 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 63f\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
       "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z29.h, z0.h[4]\n"
       "fmla z12.h, z29.h, z1.h[4]\n"
       "fmla z16.h, z29.h, z2.h[4]\n"
       "fmla z20.h, z29.h, z3.h[4]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z29.h, z4.h[4]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z0.h[4]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z1.h[4]\n"
       "fmla z17.h, z28.h, z2.h[4]\n"
       "fmla z21.h, z28.h, z3.h[4]\n"
       "fmla z25.h, z28.h, z4.h[4]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z29.h, z0.h[4]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z29.h, z0.h[4]\n"
       "fmla z14.h, z29.h, z1.h[4]\n"
       "fmla z18.h, z29.h, z2.h[4]\n"
       "fmla z22.h, z29.h, z3.h[4]\n"
@@ -2127,21 +2125,21 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 63f\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
       "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z29.h, z0.h[5]\n"
       "fmla z12.h, z29.h, z1.h[5]\n"
       "fmla z16.h, z29.h, z2.h[5]\n"
       "fmla z20.h, z29.h, z3.h[5]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z29.h, z4.h[5]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z0.h[5]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z1.h[5]\n"
       "fmla z17.h, z28.h, z2.h[5]\n"
       "fmla z21.h, z28.h, z3.h[5]\n"
       "fmla z25.h, z28.h, z4.h[5]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z29.h, z0.h[5]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z29.h, z0.h[5]\n"
       "fmla z14.h, z29.h, z1.h[5]\n"
       "fmla z18.h, z29.h, z2.h[5]\n"
       "fmla z22.h, z29.h, z3.h[5]\n"
@@ -2154,21 +2152,21 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 63f\n"
       "ld1h { z29.h }, p5/Z, [x10]\n"
       "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z29.h, z0.h[6]\n"
       "fmla z12.h, z29.h, z1.h[6]\n"
       "fmla z16.h, z29.h, z2.h[6]\n"
       "fmla z20.h, z29.h, z3.h[6]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z29.h, z4.h[6]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z0.h[6]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z1.h[6]\n"
       "fmla z17.h, z28.h, z2.h[6]\n"
       "fmla z21.h, z28.h, z3.h[6]\n"
       "fmla z25.h, z28.h, z4.h[6]\n"
       "ld1h { z28.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z29.h, z0.h[6]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z29.h, z0.h[6]\n"
       "fmla z14.h, z29.h, z1.h[6]\n"
       "fmla z18.h, z29.h, z2.h[6]\n"
       "fmla z22.h, z29.h, z3.h[6]\n"
@@ -2186,8 +2184,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "fmla z16.h, z29.h, z2.h[7]\n"
       "fmla z20.h, z29.h, z3.h[7]\n"
       "fmla z24.h, z29.h, z4.h[7]\n"
-      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.h, z28.h, z0.h[7]\n"
+      "ld1h { z29.h }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.h, z28.h, z1.h[7]\n"
       "fmla z17.h, z28.h, z2.h[7]\n"
       "fmla z21.h, z28.h, z3.h[7]\n"
@@ -2210,14 +2208,14 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 58b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
+      "add x22, x23, x20, LSL #1\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z29.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z29.h }, p5/Z, [x21]\n"
       "ld1rh { z28.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z29.h\n"
       "fmin z9.h, p5/M, z9.h, z29.h\n"
@@ -2265,22 +2263,22 @@ void sve_hybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p4, [x25]\n"
-      "st1h { z17.h }, p3, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p1, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p4, [x24]\n"
-      "st1h { z21.h }, p3, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p1, [x24, #3, MUL VL]\n"
-      "st1h { z24.h }, p4, [x23]\n"
-      "st1h { z25.h }, p3, [x23, #1, MUL VL]\n"
-      "st1h { z26.h }, p2, [x23, #2, MUL VL]\n"
-      "st1h { z27.h }, p1, [x23, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p4, [x24]\n"
+      "st1h { z17.h }, p3, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p1, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p4, [x23]\n"
+      "st1h { z21.h }, p3, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p1, [x23, #3, MUL VL]\n"
+      "st1h { z24.h }, p4, [x22]\n"
+      "st1h { z25.h }, p3, [x22, #1, MUL VL]\n"
+      "st1h { z26.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z27.h }, p1, [x22, #3, MUL VL]\n"
       "65:"  // Height 5: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -2288,13 +2286,12 @@ void sve_hybrid_fp16_mla_6x4VL (
       "b 80f\n"
       "66:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0xc\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "67:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.h, x20, x11\n"
@@ -2307,18 +2304,18 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cbz x12, 68f\n"
       "ld1h { z8.h }, p5/Z, [x12]\n"
       "ld1h { z9.h }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1h { z10.h }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z11.h }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -2334,17 +2331,17 @@ void sve_hybrid_fp16_mla_6x4VL (
       "68:"  // Height 6: no bias
       "tbz %x[flags], #0, 69f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #1\n"
+      "add x23, x24, x20, LSL #1\n"
       "ld1h { z8.h }, p4/Z, [x9]\n"
+      "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "ld1h { z9.h }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1h { z10.h }, p2/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #1\n"
       "ld1h { z11.h }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #1\n"
-      "add x23, x24, x20, LSL #1\n"
-      "add x22, x23, x20, LSL #1\n"
       "ld1h { z12.h }, p4/Z, [x24]\n"
       "ld1h { z13.h }, p3/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #1\n"
-      "add x20, x21, x20, LSL #1\n"
       "ld1h { z14.h }, p2/Z, [x24, #2, MUL VL]\n"
       "ld1h { z15.h }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1h { z16.h }, p4/Z, [x23]\n"
@@ -2393,8 +2390,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "mov x28, #0x0\n"
       "71:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 72f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2425,29 +2422,29 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 75f\n"
       "74:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z1.h }, p5/Z, [x10]\n"
-      "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x8\n"
-      "cmp x27, #0x8\n"
       "ld1rqh { z7.h }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqh { z6.h }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x8\n"
       "ld1rqh { z5.h }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqh { z4.h }, p0/Z, [x23]\n"
+      "cmp x27, #0x8\n"
+      "add x26, x26, #0x10\n"
       "ld1rqh { z3.h }, p0/Z, [x22]\n"
       "ld1rqh { z2.h }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1h { z1.h }, p5/Z, [x10]\n"
+      "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z1.h, z7.h[0]\n"
       "fmla z12.h, z1.h, z6.h[0]\n"
-      "add x21, x21, #0x10\n"
       "fmla z16.h, z1.h, z5.h[0]\n"
       "fmla z20.h, z1.h, z4.h[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z24.h, z1.h, z3.h[0]\n"
       "fmla z28.h, z1.h, z2.h[0]\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x21, x21, #0x10\n"
       "fmla z9.h, z0.h, z7.h[0]\n"
       "fmla z13.h, z0.h, z6.h[0]\n"
       "fmla z17.h, z0.h, z5.h[0]\n"
@@ -2668,24 +2665,24 @@ void sve_hybrid_fp16_mla_6x4VL (
       "bgt 74b\n"
       "75:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.h, XZR, x27\n"
-      "ld1h { z7.h }, p5/Z, [x10]\n"
-      "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqh { z0.h }, p0/Z, [x26]\n"
       "ld1rqh { z1.h }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqh { z2.h }, p0/Z, [x24]\n"
       "ld1rqh { z3.h }, p0/Z, [x23]\n"
       "ld1rqh { z4.h }, p0/Z, [x22]\n"
       "ld1rqh { z5.h }, p0/Z, [x21]\n"
+      "ld1h { z7.h }, p5/Z, [x10]\n"
+      "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.h, z7.h, z0.h[0]\n"
       "fmla z12.h, z7.h, z1.h[0]\n"
-      "fmla z9.h, z6.h, z0.h[0]\n"
-      "fmla z13.h, z6.h, z1.h[0]\n"
       "fmla z16.h, z7.h, z2.h[0]\n"
       "fmla z20.h, z7.h, z3.h[0]\n"
       "fmla z24.h, z7.h, z4.h[0]\n"
       "fmla z28.h, z7.h, z5.h[0]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.h, z6.h, z0.h[0]\n"
+      "fmla z13.h, z6.h, z1.h[0]\n"
       "fmla z17.h, z6.h, z2.h[0]\n"
       "fmla z21.h, z6.h, z3.h[0]\n"
       "fmla z25.h, z6.h, z4.h[0]\n"
@@ -2707,23 +2704,23 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 76f\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z7.h, z0.h[1]\n"
       "fmla z12.h, z7.h, z1.h[1]\n"
       "fmla z16.h, z7.h, z2.h[1]\n"
       "fmla z20.h, z7.h, z3.h[1]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z7.h, z4.h[1]\n"
       "fmla z28.h, z7.h, z5.h[1]\n"
-      "fmla z9.h, z6.h, z0.h[1]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.h, z6.h, z0.h[1]\n"
       "fmla z13.h, z6.h, z1.h[1]\n"
       "fmla z17.h, z6.h, z2.h[1]\n"
       "fmla z21.h, z6.h, z3.h[1]\n"
       "fmla z25.h, z6.h, z4.h[1]\n"
       "fmla z29.h, z6.h, z5.h[1]\n"
       "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z7.h, z0.h[1]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z7.h, z0.h[1]\n"
       "fmla z14.h, z7.h, z1.h[1]\n"
       "fmla z18.h, z7.h, z2.h[1]\n"
       "fmla z22.h, z7.h, z3.h[1]\n"
@@ -2738,23 +2735,23 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 76f\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z7.h, z0.h[2]\n"
       "fmla z12.h, z7.h, z1.h[2]\n"
       "fmla z16.h, z7.h, z2.h[2]\n"
       "fmla z20.h, z7.h, z3.h[2]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z7.h, z4.h[2]\n"
       "fmla z28.h, z7.h, z5.h[2]\n"
-      "fmla z9.h, z6.h, z0.h[2]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.h, z6.h, z0.h[2]\n"
       "fmla z13.h, z6.h, z1.h[2]\n"
       "fmla z17.h, z6.h, z2.h[2]\n"
       "fmla z21.h, z6.h, z3.h[2]\n"
       "fmla z25.h, z6.h, z4.h[2]\n"
       "fmla z29.h, z6.h, z5.h[2]\n"
       "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z7.h, z0.h[2]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z7.h, z0.h[2]\n"
       "fmla z14.h, z7.h, z1.h[2]\n"
       "fmla z18.h, z7.h, z2.h[2]\n"
       "fmla z22.h, z7.h, z3.h[2]\n"
@@ -2769,23 +2766,23 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 76f\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z7.h, z0.h[3]\n"
       "fmla z12.h, z7.h, z1.h[3]\n"
       "fmla z16.h, z7.h, z2.h[3]\n"
       "fmla z20.h, z7.h, z3.h[3]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z7.h, z4.h[3]\n"
       "fmla z28.h, z7.h, z5.h[3]\n"
-      "fmla z9.h, z6.h, z0.h[3]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.h, z6.h, z0.h[3]\n"
       "fmla z13.h, z6.h, z1.h[3]\n"
       "fmla z17.h, z6.h, z2.h[3]\n"
       "fmla z21.h, z6.h, z3.h[3]\n"
       "fmla z25.h, z6.h, z4.h[3]\n"
       "fmla z29.h, z6.h, z5.h[3]\n"
       "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z7.h, z0.h[3]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z7.h, z0.h[3]\n"
       "fmla z14.h, z7.h, z1.h[3]\n"
       "fmla z18.h, z7.h, z2.h[3]\n"
       "fmla z22.h, z7.h, z3.h[3]\n"
@@ -2800,23 +2797,23 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 76f\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z7.h, z0.h[4]\n"
       "fmla z12.h, z7.h, z1.h[4]\n"
       "fmla z16.h, z7.h, z2.h[4]\n"
       "fmla z20.h, z7.h, z3.h[4]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z7.h, z4.h[4]\n"
       "fmla z28.h, z7.h, z5.h[4]\n"
-      "fmla z9.h, z6.h, z0.h[4]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.h, z6.h, z0.h[4]\n"
       "fmla z13.h, z6.h, z1.h[4]\n"
       "fmla z17.h, z6.h, z2.h[4]\n"
       "fmla z21.h, z6.h, z3.h[4]\n"
       "fmla z25.h, z6.h, z4.h[4]\n"
       "fmla z29.h, z6.h, z5.h[4]\n"
       "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z7.h, z0.h[4]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z7.h, z0.h[4]\n"
       "fmla z14.h, z7.h, z1.h[4]\n"
       "fmla z18.h, z7.h, z2.h[4]\n"
       "fmla z22.h, z7.h, z3.h[4]\n"
@@ -2831,23 +2828,23 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 76f\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z7.h, z0.h[5]\n"
       "fmla z12.h, z7.h, z1.h[5]\n"
       "fmla z16.h, z7.h, z2.h[5]\n"
       "fmla z20.h, z7.h, z3.h[5]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z7.h, z4.h[5]\n"
       "fmla z28.h, z7.h, z5.h[5]\n"
-      "fmla z9.h, z6.h, z0.h[5]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.h, z6.h, z0.h[5]\n"
       "fmla z13.h, z6.h, z1.h[5]\n"
       "fmla z17.h, z6.h, z2.h[5]\n"
       "fmla z21.h, z6.h, z3.h[5]\n"
       "fmla z25.h, z6.h, z4.h[5]\n"
       "fmla z29.h, z6.h, z5.h[5]\n"
       "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z7.h, z0.h[5]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z7.h, z0.h[5]\n"
       "fmla z14.h, z7.h, z1.h[5]\n"
       "fmla z18.h, z7.h, z2.h[5]\n"
       "fmla z22.h, z7.h, z3.h[5]\n"
@@ -2862,23 +2859,23 @@ void sve_hybrid_fp16_mla_6x4VL (
       "ble 76f\n"
       "ld1h { z7.h }, p5/Z, [x10]\n"
       "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.h, z7.h, z0.h[6]\n"
       "fmla z12.h, z7.h, z1.h[6]\n"
       "fmla z16.h, z7.h, z2.h[6]\n"
       "fmla z20.h, z7.h, z3.h[6]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.h, z7.h, z4.h[6]\n"
       "fmla z28.h, z7.h, z5.h[6]\n"
-      "fmla z9.h, z6.h, z0.h[6]\n"
       "ld1h { z7.h }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.h, z6.h, z0.h[6]\n"
       "fmla z13.h, z6.h, z1.h[6]\n"
       "fmla z17.h, z6.h, z2.h[6]\n"
       "fmla z21.h, z6.h, z3.h[6]\n"
       "fmla z25.h, z6.h, z4.h[6]\n"
       "fmla z29.h, z6.h, z5.h[6]\n"
       "ld1h { z6.h }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.h, z7.h, z0.h[6]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.h, z7.h, z0.h[6]\n"
       "fmla z14.h, z7.h, z1.h[6]\n"
       "fmla z18.h, z7.h, z2.h[6]\n"
       "fmla z22.h, z7.h, z3.h[6]\n"
@@ -2926,15 +2923,15 @@ void sve_hybrid_fp16_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 71b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #1\n"
-      "add x25, x26, x20, LSL #1\n"
+      "add x25, x9, x20, LSL #1\n"
       "add x24, x25, x20, LSL #1\n"
       "add x23, x24, x20, LSL #1\n"
       "add x22, x23, x20, LSL #1\n"
+      "add x21, x22, x20, LSL #1\n"
       "tbz %x[flags], #1, 77f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rh { z1.h }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rh { z1.h }, p5/Z, [x21]\n"
       "ld1rh { z0.h }, p5/Z, [x20]\n"
       "fmin z8.h, p5/M, z8.h, z1.h\n"
       "fmin z9.h, p5/M, z9.h, z1.h\n"
@@ -2990,26 +2987,26 @@ void sve_hybrid_fp16_mla_6x4VL (
       "st1h { z10.h }, p2, [x9, #2, MUL VL]\n"
       "st1h { z11.h }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1h { z12.h }, p4, [x26]\n"
-      "st1h { z13.h }, p3, [x26, #1, MUL VL]\n"
-      "st1h { z14.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z15.h }, p1, [x26, #3, MUL VL]\n"
-      "st1h { z16.h }, p4, [x25]\n"
-      "st1h { z17.h }, p3, [x25, #1, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z19.h }, p1, [x25, #3, MUL VL]\n"
-      "st1h { z20.h }, p4, [x24]\n"
-      "st1h { z21.h }, p3, [x24, #1, MUL VL]\n"
-      "st1h { z22.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z23.h }, p1, [x24, #3, MUL VL]\n"
-      "st1h { z24.h }, p4, [x23]\n"
-      "st1h { z25.h }, p3, [x23, #1, MUL VL]\n"
-      "st1h { z26.h }, p2, [x23, #2, MUL VL]\n"
-      "st1h { z27.h }, p1, [x23, #3, MUL VL]\n"
-      "st1h { z28.h }, p4, [x22]\n"
-      "st1h { z29.h }, p3, [x22, #1, MUL VL]\n"
-      "st1h { z30.h }, p2, [x22, #2, MUL VL]\n"
-      "st1h { z31.h }, p1, [x22, #3, MUL VL]\n"
+      "st1h { z12.h }, p4, [x25]\n"
+      "st1h { z13.h }, p3, [x25, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x25, #2, MUL VL]\n"
+      "st1h { z15.h }, p1, [x25, #3, MUL VL]\n"
+      "st1h { z16.h }, p4, [x24]\n"
+      "st1h { z17.h }, p3, [x24, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x24, #2, MUL VL]\n"
+      "st1h { z19.h }, p1, [x24, #3, MUL VL]\n"
+      "st1h { z20.h }, p4, [x23]\n"
+      "st1h { z21.h }, p3, [x23, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x23, #2, MUL VL]\n"
+      "st1h { z23.h }, p1, [x23, #3, MUL VL]\n"
+      "st1h { z24.h }, p4, [x22]\n"
+      "st1h { z25.h }, p3, [x22, #1, MUL VL]\n"
+      "st1h { z26.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z27.h }, p1, [x22, #3, MUL VL]\n"
+      "st1h { z28.h }, p4, [x21]\n"
+      "st1h { z29.h }, p3, [x21, #1, MUL VL]\n"
+      "st1h { z30.h }, p2, [x21, #2, MUL VL]\n"
+      "st1h { z31.h }, p1, [x21, #3, MUL VL]\n"
       "78:"  // Height 6: Writeback done
       "dech x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -3026,8 +3023,8 @@ void sve_hybrid_fp16_mla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL.hpp
index 3e040b6197..7936eeb11a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 4, 1> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 1> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp
index 2b836659a2..430c081288 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/a64fx.cpp
@@ -47,19 +47,18 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -104,10 +102,10 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "cmp %x[M], #0x2\n"
       "bgt 25f\n"
       "beq 13f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -140,8 +138,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -160,14 +158,14 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop
       "fmla z8.s, p4/M, z6.s, z0.s\n"
-      "ld1w { z17.s }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z17.s }, p4/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
       "add x26, x26, #0x4\n"
-      "subs x27, x27, #0x1\n"
       "fmla z10.s, p4/M, z17.s, z0.s\n"
       "fmla z11.s, p4/M, z16.s, z0.s\n"
+      "subs x27, x27, #0x1\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1w { z6.s }, p4/Z, [x10]\n"
       "ld1w { z7.s }, p4/Z, [x10, #1, MUL VL]\n"
@@ -175,19 +173,19 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "10:"  // Height 1: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "fmla z8.s, p4/M, z6.s, z0.s\n"
-      "ld1w { z17.s }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z17.s }, p4/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x10, #3, MUL VL]\n"
       "add x28, x28, #0x1\n"
-      "addvl x10, x10, #4\n"
       "cmp x28, x20\n"
       "fmla z10.s, p4/M, z17.s, z0.s\n"
       "fmla z11.s, p4/M, z16.s, z0.s\n"
+      "addvl x10, x10, #4\n"
       "bne 6b\n"
       "tbz %x[flags], #1, 11f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p4/Z, [x21]\n"
       "ld1rw { z16.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z17.s\n"
       "fmin z9.s, p4/M, z9.s, z17.s\n"
@@ -209,10 +207,10 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "bgt 2b\n"
       "b 74f\n"
       "13:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "14:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -225,22 +223,22 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "cbz x12, 15f\n"
       "ld1w { z8.s }, p4/Z, [x12]\n"
       "ld1w { z9.s }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x12, x12, #4\n"
       "b 17f\n"
       "15:"  // Height 2: no bias
       "tbz %x[flags], #0, 16f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x20]\n"
       "ld1w { z13.s }, p2/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x20, #2, MUL VL]\n"
@@ -259,8 +257,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "18:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 19f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -295,8 +293,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z10.s, p4/M, z17.s, z0.s\n"
       "fmla z14.s, p4/M, z17.s, z1.s\n"
       "fmla z11.s, p4/M, z16.s, z0.s\n"
-      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "fmla z15.s, p4/M, z16.s, z1.s\n"
+      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1w { z6.s }, p4/Z, [x10]\n"
       "ld1w { z7.s }, p4/Z, [x10, #1, MUL VL]\n"
@@ -310,19 +308,19 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "ld1w { z16.s }, p4/Z, [x10, #3, MUL VL]\n"
       "add x28, x28, #0x1\n"
-      "addvl x10, x10, #4\n"
       "cmp x28, x20\n"
       "fmla z10.s, p4/M, z17.s, z0.s\n"
       "fmla z14.s, p4/M, z17.s, z1.s\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.s, p4/M, z16.s, z0.s\n"
       "fmla z15.s, p4/M, z16.s, z1.s\n"
       "bne 18b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "tbz %x[flags], #1, 23f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p4/Z, [x21]\n"
       "ld1rw { z16.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z17.s\n"
       "fmin z9.s, p4/M, z9.s, z17.s\n"
@@ -346,20 +344,20 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
       "24:"  // Height 2: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 14b\n"
       "b 74f\n"
       "25:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "26:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -372,27 +370,27 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "cbz x12, 27f\n"
       "ld1w { z8.s }, p4/Z, [x12]\n"
       "ld1w { z9.s }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 29f\n"
       "27:"  // Height 3: no bias
       "tbz %x[flags], #0, 28f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x21]\n"
       "ld1w { z13.s }, p2/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x21, #2, MUL VL]\n"
@@ -419,8 +417,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "30:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 31f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -451,8 +449,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "add x26, x26, #0x4\n"
       "subs x27, x27, #0x1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
-      "ld1w { z21.s }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z21.s }, p4/Z, [x10, #2, MUL VL]\n"
       "add x25, x25, #0x4\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
@@ -464,11 +462,11 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z18.s, p4/M, z21.s, z2.s\n"
       "fmla z11.s, p4/M, z20.s, z0.s\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
+      "ld1w { z6.s }, p4/Z, [x10]\n"
       "fmla z15.s, p4/M, z20.s, z1.s\n"
-      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "fmla z19.s, p4/M, z20.s, z2.s\n"
+      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
-      "ld1w { z6.s }, p4/Z, [x10]\n"
       "ld1w { z7.s }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 33b\n"
       "34:"  // Height 3: Multiply loop: Main loop skip
@@ -477,13 +475,13 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z12.s, p4/M, z6.s, z1.s\n"
       "add x28, x28, #0x1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
-      "ld1w { z21.s }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z21.s }, p4/Z, [x10, #2, MUL VL]\n"
+      "cmp x28, x20\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
       "ld1w { z20.s }, p4/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
-      "cmp x28, x20\n"
       "fmla z10.s, p4/M, z21.s, z0.s\n"
       "fmla z14.s, p4/M, z21.s, z1.s\n"
       "fmla z18.s, p4/M, z21.s, z2.s\n"
@@ -492,12 +490,12 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z19.s, p4/M, z20.s, z2.s\n"
       "bne 30b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "tbz %x[flags], #1, 35f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z21.s }, p4/Z, [x21]\n"
       "ld1rw { z20.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z21.s\n"
       "fmin z9.s, p4/M, z9.s, z21.s\n"
@@ -529,24 +527,24 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p3, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p3, [x24]\n"
+      "st1w { z17.s }, p2, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p0, [x24, #3, MUL VL]\n"
       "36:"  // Height 3: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 26b\n"
       "b 74f\n"
       "37:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "38:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -559,18 +557,18 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "cbz x12, 39f\n"
       "ld1w { z8.s }, p4/Z, [x12]\n"
       "ld1w { z9.s }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -578,13 +576,13 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "39:"  // Height 4: no bias
       "tbz %x[flags], #0, 40f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x22]\n"
       "ld1w { z13.s }, p2/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x22, #2, MUL VL]\n"
@@ -619,8 +617,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "42:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 43f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -670,6 +668,7 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z14.s, p4/M, z25.s, z1.s\n"
       "fmla z18.s, p4/M, z25.s, z2.s\n"
       "fmla z22.s, p4/M, z25.s, z3.s\n"
+      "ld1w { z6.s }, p4/Z, [x10]\n"
       "fmla z11.s, p4/M, z24.s, z0.s\n"
       "fmla z15.s, p4/M, z24.s, z1.s\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
@@ -678,7 +677,6 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z23.s, p4/M, z24.s, z3.s\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
       "ld1rw { z3.s }, p4/Z, [x23]\n"
-      "ld1w { z6.s }, p4/Z, [x10]\n"
       "ld1w { z7.s }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 45b\n"
       "46:"  // Height 4: Multiply loop: Main loop skip
@@ -689,15 +687,15 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z16.s, p4/M, z6.s, z2.s\n"
       "fmla z20.s, p4/M, z6.s, z3.s\n"
       "ld1w { z25.s }, p4/Z, [x10, #2, MUL VL]\n"
+      "cmp x28, x20\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
       "fmla z21.s, p4/M, z7.s, z3.s\n"
       "ld1w { z24.s }, p4/Z, [x10, #3, MUL VL]\n"
-      "cmp x28, x20\n"
+      "addvl x10, x10, #4\n"
       "fmla z10.s, p4/M, z25.s, z0.s\n"
       "fmla z14.s, p4/M, z25.s, z1.s\n"
-      "addvl x10, x10, #4\n"
       "fmla z18.s, p4/M, z25.s, z2.s\n"
       "fmla z22.s, p4/M, z25.s, z3.s\n"
       "fmla z11.s, p4/M, z24.s, z0.s\n"
@@ -706,13 +704,13 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z23.s, p4/M, z24.s, z3.s\n"
       "bne 42b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "tbz %x[flags], #1, 47f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z25.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z25.s }, p4/Z, [x21]\n"
       "ld1rw { z24.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z25.s\n"
       "fmin z9.s, p4/M, z9.s, z25.s\n"
@@ -752,28 +750,28 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p3, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p0, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p3, [x24]\n"
-      "st1w { z21.s }, p2, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p1, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p0, [x24, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p3, [x24]\n"
+      "st1w { z17.s }, p2, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p0, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p3, [x23]\n"
+      "st1w { z21.s }, p2, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p1, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p0, [x23, #3, MUL VL]\n"
       "48:"  // Height 4: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 38b\n"
       "b 74f\n"
       "49:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "50:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -786,18 +784,18 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "cbz x12, 51f\n"
       "ld1w { z8.s }, p4/Z, [x12]\n"
       "ld1w { z9.s }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -809,16 +807,16 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "51:"  // Height 5: no bias
       "tbz %x[flags], #0, 52f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x23]\n"
       "ld1w { z13.s }, p2/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p1/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p0/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p3/Z, [x22]\n"
@@ -859,8 +857,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "54:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 55f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -903,8 +901,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "add x25, x25, #0x4\n"
       "add x24, x24, #0x4\n"
       "fmla z24.s, p4/M, z6.s, z4.s\n"
-      "ld1w { z29.s }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z29.s }, p4/Z, [x10, #2, MUL VL]\n"
       "add x23, x23, #0x4\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
@@ -922,12 +920,12 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1w { z6.s }, p4/Z, [x10]\n"
       "fmla z15.s, p4/M, z28.s, z1.s\n"
-      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "fmla z19.s, p4/M, z28.s, z2.s\n"
+      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
       "fmla z23.s, p4/M, z28.s, z3.s\n"
-      "ld1rw { z3.s }, p4/Z, [x23]\n"
       "fmla z27.s, p4/M, z28.s, z4.s\n"
+      "ld1rw { z3.s }, p4/Z, [x23]\n"
       "ld1rw { z4.s }, p4/Z, [x22]\n"
       "ld1w { z7.s }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 57b\n"
@@ -938,12 +936,12 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
       "fmla z20.s, p4/M, z6.s, z3.s\n"
+      "cmp x28, x20\n"
       "fmla z24.s, p4/M, z6.s, z4.s\n"
-      "ld1w { z29.s }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
+      "ld1w { z29.s }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
-      "cmp x28, x20\n"
       "fmla z21.s, p4/M, z7.s, z3.s\n"
       "fmla z25.s, p4/M, z7.s, z4.s\n"
       "ld1w { z28.s }, p4/Z, [x10, #3, MUL VL]\n"
@@ -960,14 +958,14 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z27.s, p4/M, z28.s, z4.s\n"
       "bne 54b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "tbz %x[flags], #1, 59f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z29.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z29.s }, p4/Z, [x21]\n"
       "ld1rw { z28.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z29.s\n"
       "fmin z9.s, p4/M, z9.s, z29.s\n"
@@ -1015,22 +1013,22 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p3, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p0, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p3, [x24]\n"
-      "st1w { z21.s }, p2, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p1, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p0, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p3, [x23]\n"
-      "st1w { z25.s }, p2, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p1, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p0, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p3, [x24]\n"
+      "st1w { z17.s }, p2, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p0, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p3, [x23]\n"
+      "st1w { z21.s }, p2, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p1, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p0, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p3, [x22]\n"
+      "st1w { z25.s }, p2, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p0, [x22, #3, MUL VL]\n"
       "60:"  // Height 5: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1038,13 +1036,12 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "b 74f\n"
       "61:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "62:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -1057,18 +1054,18 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "cbz x12, 63f\n"
       "ld1w { z8.s }, p4/Z, [x12]\n"
       "ld1w { z9.s }, p4/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p4/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p4/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1084,17 +1081,17 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "63:"  // Height 6: no bias
       "tbz %x[flags], #0, 64f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x24]\n"
       "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p3/Z, [x23]\n"
@@ -1143,8 +1140,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "66:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 67f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1232,12 +1229,12 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "fmla z16.s, p4/M, z6.s, z2.s\n"
       "fmla z20.s, p4/M, z6.s, z3.s\n"
+      "cmp x28, x20\n"
       "fmla z24.s, p4/M, z6.s, z4.s\n"
       "fmla z28.s, p4/M, z6.s, z5.s\n"
       "ld1w { z6.s }, p4/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, p4/M, z7.s, z0.s\n"
       "fmla z13.s, p4/M, z7.s, z1.s\n"
-      "cmp x28, x20\n"
       "fmla z17.s, p4/M, z7.s, z2.s\n"
       "fmla z21.s, p4/M, z7.s, z3.s\n"
       "fmla z25.s, p4/M, z7.s, z4.s\n"
@@ -1258,15 +1255,15 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "fmla z31.s, p4/M, z7.s, z5.s\n"
       "bne 66b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "tbz %x[flags], #1, 71f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p4/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p4/Z, [x21]\n"
       "ld1rw { z0.s }, p4/Z, [x20]\n"
       "fmin z8.s, p4/M, z8.s, z1.s\n"
       "fmin z9.s, p4/M, z9.s, z1.s\n"
@@ -1322,26 +1319,26 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p3, [x26]\n"
-      "st1w { z13.s }, p2, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p1, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p0, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p3, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p0, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p3, [x24]\n"
-      "st1w { z21.s }, p2, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p1, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p0, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p3, [x23]\n"
-      "st1w { z25.s }, p2, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p1, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p0, [x23, #3, MUL VL]\n"
-      "st1w { z28.s }, p3, [x22]\n"
-      "st1w { z29.s }, p2, [x22, #1, MUL VL]\n"
-      "st1w { z30.s }, p1, [x22, #2, MUL VL]\n"
-      "st1w { z31.s }, p0, [x22, #3, MUL VL]\n"
+      "st1w { z12.s }, p3, [x25]\n"
+      "st1w { z13.s }, p2, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p1, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p0, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p3, [x24]\n"
+      "st1w { z17.s }, p2, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p0, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p3, [x23]\n"
+      "st1w { z21.s }, p2, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p1, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p0, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p3, [x22]\n"
+      "st1w { z25.s }, p2, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p0, [x22, #3, MUL VL]\n"
+      "st1w { z28.s }, p3, [x21]\n"
+      "st1w { z29.s }, p2, [x21, #1, MUL VL]\n"
+      "st1w { z30.s }, p1, [x21, #2, MUL VL]\n"
+      "st1w { z31.s }, p0, [x21, #3, MUL VL]\n"
       "72:"  // Height 6: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1358,8 +1355,8 @@ void sve_hybrid_fp32_mla_6x4VL_a64fx (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "74:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp
index 99828e8f0c..da4670d7e0 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_6x4VL/generic.cpp
@@ -47,19 +47,18 @@ void sve_hybrid_fp32_mla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void sve_hybrid_fp32_mla_6x4VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -104,10 +102,10 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cmp %x[M], #0x2\n"
       "bgt 27f\n"
       "beq 14f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -140,8 +138,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -157,98 +155,98 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "fmla z8.s, z17.s, z0.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z16.s }, p5/Z, [x10]\n"
+      "fmla z8.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z10.s, z16.s, z0.s[0]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.s, z17.s, z0.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #4, MUL VL]\n"
+      "fmla z8.s, z16.s, z0.s[1]\n"
       "ld1w { z16.s }, p5/Z, [x10, #5, MUL VL]\n"
-      "fmla z8.s, z17.s, z0.s[1]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[1]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #6, MUL VL]\n"
+      "fmla z10.s, z16.s, z0.s[1]\n"
       "ld1w { z16.s }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      "fmla z10.s, z17.s, z0.s[1]\n"
       "fmla z11.s, z16.s, z0.s[1]\n"
       "ld1w { z17.s }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.s, z17.s, z0.s[2]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[2]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #-6, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #-5, MUL VL]\n"
       "fmla z10.s, z17.s, z0.s[2]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.s, z16.s, z0.s[2]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #-4, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #-3, MUL VL]\n"
       "fmla z8.s, z17.s, z0.s[3]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[3]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #-2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #-1, MUL VL]\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "fmla z10.s, z17.s, z0.s[3]\n"
       "fmla z11.s, z16.s, z0.s[3]\n"
+      "add x26, x26, #0x10\n"
       "bgt 9b\n"
       "10:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
-      "fmla z8.s, z17.s, z0.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z16.s }, p5/Z, [x10]\n"
+      "fmla z8.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[0]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z10.s, z17.s, z0.s[0]\n"
       "fmla z11.s, z16.s, z0.s[0]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z17.s, z0.s[1]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[1]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.s, z17.s, z0.s[1]\n"
       "fmla z11.s, z16.s, z0.s[1]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z17.s, z0.s[2]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[2]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.s, z17.s, z0.s[2]\n"
       "fmla z11.s, z16.s, z0.s[2]\n"
+      "addvl x10, x10, #4\n"
       "ble 11f\n"
       "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.s, z17.s, z0.s[3]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[3]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z10.s, z17.s, z0.s[3]\n"
       "fmla z11.s, z16.s, z0.s[3]\n"
+      "addvl x10, x10, #4\n"
       "11:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 6b\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z17.s\n"
       "fmin z9.s, p5/M, z9.s, z17.s\n"
@@ -270,10 +268,10 @@ void sve_hybrid_fp32_mla_6x4VL (
       "bgt 2b\n"
       "b 80f\n"
       "14:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -286,22 +284,22 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cbz x12, 16f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "addvl x12, x12, #4\n"
       "b 18f\n"
       "16:"  // Height 2: no bias
       "tbz %x[flags], #0, 17f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
@@ -320,8 +318,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "19:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -340,38 +338,38 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 23f\n"
       "22:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z1.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.s, z17.s, z1.s[0]\n"
       "fmla z12.s, z17.s, z0.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z1.s[0]\n"
       "fmla z13.s, z16.s, z0.s[0]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
       "fmla z10.s, z17.s, z1.s[0]\n"
       "fmla z14.s, z17.s, z0.s[0]\n"
       "ld1w { z17.s }, p5/Z, [x10, #4, MUL VL]\n"
+      "cmp x27, #0x4\n"
       "fmla z11.s, z16.s, z1.s[0]\n"
       "fmla z15.s, z16.s, z0.s[0]\n"
       "ld1w { z16.s }, p5/Z, [x10, #5, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       "fmla z8.s, z17.s, z1.s[1]\n"
       "fmla z12.s, z17.s, z0.s[1]\n"
       "ld1w { z17.s }, p5/Z, [x10, #6, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       "fmla z9.s, z16.s, z1.s[1]\n"
       "fmla z13.s, z16.s, z0.s[1]\n"
       "ld1w { z16.s }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
       "fmla z10.s, z17.s, z1.s[1]\n"
       "fmla z14.s, z17.s, z0.s[1]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z11.s, z16.s, z1.s[1]\n"
       "fmla z15.s, z16.s, z0.s[1]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.s, z17.s, z1.s[2]\n"
       "fmla z12.s, z17.s, z0.s[2]\n"
@@ -398,50 +396,50 @@ void sve_hybrid_fp32_mla_6x4VL (
       "bgt 22b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z17.s }, p5/Z, [x10]\n"
-      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
+      "ld1w { z17.s }, p5/Z, [x10]\n"
+      "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.s, z17.s, z0.s[0]\n"
       "fmla z12.s, z17.s, z1.s[0]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[0]\n"
       "fmla z13.s, z16.s, z1.s[0]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z10.s, z17.s, z0.s[0]\n"
       "fmla z14.s, z17.s, z1.s[0]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.s, z16.s, z0.s[0]\n"
       "fmla z15.s, z16.s, z1.s[0]\n"
       "ble 24f\n"
       "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z17.s, z0.s[1]\n"
       "fmla z12.s, z17.s, z1.s[1]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[1]\n"
       "fmla z13.s, z16.s, z1.s[1]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.s, z17.s, z0.s[1]\n"
       "fmla z14.s, z17.s, z1.s[1]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.s, z16.s, z0.s[1]\n"
       "fmla z15.s, z16.s, z1.s[1]\n"
       "ble 24f\n"
       "ld1w { z17.s }, p5/Z, [x10]\n"
       "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z17.s, z0.s[2]\n"
       "fmla z12.s, z17.s, z1.s[2]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[2]\n"
       "fmla z13.s, z16.s, z1.s[2]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x1\n"
       "fmla z10.s, z17.s, z0.s[2]\n"
       "fmla z14.s, z17.s, z1.s[2]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.s, z16.s, z0.s[2]\n"
       "fmla z15.s, z16.s, z1.s[2]\n"
       "ble 24f\n"
@@ -449,13 +447,13 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.s, z17.s, z0.s[3]\n"
       "fmla z12.s, z17.s, z1.s[3]\n"
-      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z16.s, z0.s[3]\n"
       "fmla z13.s, z16.s, z1.s[3]\n"
+      "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1w { z16.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "fmla z10.s, z17.s, z0.s[3]\n"
       "fmla z14.s, z17.s, z1.s[3]\n"
+      "addvl x10, x10, #4\n"
       "fmla z11.s, z16.s, z0.s[3]\n"
       "fmla z15.s, z16.s, z1.s[3]\n"
       "24:"  // Height 2: Multiply loop: multiply skip
@@ -464,11 +462,11 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 19b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z17.s\n"
       "fmin z9.s, p5/M, z9.s, z17.s\n"
@@ -492,20 +490,20 @@ void sve_hybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
       "26:"  // Height 2: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 15b\n"
       "b 80f\n"
       "27:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -518,27 +516,27 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cbz x12, 29f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
       "b 31f\n"
       "29:"  // Height 3: no bias
       "tbz %x[flags], #0, 30f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
@@ -565,8 +563,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -588,37 +586,37 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 36f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z21.s }, p5/Z, [x10]\n"
-      "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "ld1rqw { z0.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1w { z21.s }, p5/Z, [x10]\n"
       "fmla z8.s, z21.s, z2.s[0]\n"
       "fmla z12.s, z21.s, z1.s[0]\n"
-      "fmla z9.s, z20.s, z2.s[0]\n"
-      "fmla z13.s, z20.s, z1.s[0]\n"
+      "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z16.s, z21.s, z0.s[0]\n"
+      "fmla z9.s, z20.s, z2.s[0]\n"
       "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z13.s, z20.s, z1.s[0]\n"
       "fmla z17.s, z20.s, z0.s[0]\n"
       "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n"
+      "cmp x27, #0x4\n"
       "fmla z10.s, z21.s, z2.s[0]\n"
       "fmla z14.s, z21.s, z1.s[0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "fmla z18.s, z21.s, z0.s[0]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.s, z20.s, z2.s[0]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #4, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       "fmla z15.s, z20.s, z1.s[0]\n"
       "fmla z19.s, z20.s, z0.s[0]\n"
       "ld1w { z20.s }, p5/Z, [x10, #5, MUL VL]\n"
       "fmla z8.s, z21.s, z2.s[1]\n"
       "fmla z12.s, z21.s, z1.s[1]\n"
       "fmla z16.s, z21.s, z0.s[1]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.s, z20.s, z2.s[1]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z13.s, z20.s, z1.s[1]\n"
       "fmla z17.s, z20.s, z0.s[1]\n"
       "ld1w { z20.s }, p5/Z, [x10, #7, MUL VL]\n"
@@ -627,31 +625,31 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z14.s, z21.s, z1.s[1]\n"
       "fmla z18.s, z21.s, z0.s[1]\n"
       "fmla z11.s, z20.s, z2.s[1]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z15.s, z20.s, z1.s[1]\n"
       "fmla z19.s, z20.s, z0.s[1]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1w { z20.s }, p5/Z, [x10, #-7, MUL VL]\n"
       "fmla z8.s, z21.s, z2.s[2]\n"
       "fmla z12.s, z21.s, z1.s[2]\n"
       "fmla z16.s, z21.s, z0.s[2]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.s, z20.s, z2.s[2]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z13.s, z20.s, z1.s[2]\n"
       "fmla z17.s, z20.s, z0.s[2]\n"
       "ld1w { z20.s }, p5/Z, [x10, #-5, MUL VL]\n"
       "fmla z10.s, z21.s, z2.s[2]\n"
       "fmla z14.s, z21.s, z1.s[2]\n"
       "fmla z18.s, z21.s, z0.s[2]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.s, z20.s, z2.s[2]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z15.s, z20.s, z1.s[2]\n"
       "fmla z19.s, z20.s, z0.s[2]\n"
       "ld1w { z20.s }, p5/Z, [x10, #-3, MUL VL]\n"
       "fmla z8.s, z21.s, z2.s[3]\n"
       "fmla z12.s, z21.s, z1.s[3]\n"
       "fmla z16.s, z21.s, z0.s[3]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.s, z20.s, z2.s[3]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z13.s, z20.s, z1.s[3]\n"
       "fmla z17.s, z20.s, z0.s[3]\n"
       "ld1w { z20.s }, p5/Z, [x10, #-1, MUL VL]\n"
@@ -664,18 +662,18 @@ void sve_hybrid_fp32_mla_6x4VL (
       "bgt 35b\n"
       "36:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z21.s }, p5/Z, [x10]\n"
-      "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
+      "ld1w { z21.s }, p5/Z, [x10]\n"
       "fmla z8.s, z21.s, z0.s[0]\n"
       "fmla z12.s, z21.s, z1.s[0]\n"
-      "fmla z9.s, z20.s, z0.s[0]\n"
-      "fmla z13.s, z20.s, z1.s[0]\n"
+      "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z16.s, z21.s, z2.s[0]\n"
+      "fmla z9.s, z20.s, z0.s[0]\n"
       "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z13.s, z20.s, z1.s[0]\n"
       "fmla z17.s, z20.s, z2.s[0]\n"
       "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
@@ -688,12 +686,12 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 37f\n"
       "ld1w { z21.s }, p5/Z, [x10]\n"
       "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z21.s, z0.s[1]\n"
       "fmla z12.s, z21.s, z1.s[1]\n"
       "fmla z16.s, z21.s, z2.s[1]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z20.s, z0.s[1]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.s, z20.s, z1.s[1]\n"
       "fmla z17.s, z20.s, z2.s[1]\n"
       "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n"
@@ -707,12 +705,12 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 37f\n"
       "ld1w { z21.s }, p5/Z, [x10]\n"
       "ld1w { z20.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z21.s, z0.s[2]\n"
       "fmla z12.s, z21.s, z1.s[2]\n"
       "fmla z16.s, z21.s, z2.s[2]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z20.s, z0.s[2]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z13.s, z20.s, z1.s[2]\n"
       "fmla z17.s, z20.s, z2.s[2]\n"
       "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n"
@@ -729,8 +727,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z8.s, z21.s, z0.s[3]\n"
       "fmla z12.s, z21.s, z1.s[3]\n"
       "fmla z16.s, z21.s, z2.s[3]\n"
-      "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z20.s, z0.s[3]\n"
+      "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.s, z20.s, z1.s[3]\n"
       "fmla z17.s, z20.s, z2.s[3]\n"
       "ld1w { z20.s }, p5/Z, [x10, #3, MUL VL]\n"
@@ -747,12 +745,12 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z21.s }, p5/Z, [x21]\n"
       "ld1rw { z20.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z21.s\n"
       "fmin z9.s, p5/M, z9.s, z21.s\n"
@@ -784,24 +782,24 @@ void sve_hybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
       "39:"  // Height 3: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 28b\n"
       "b 80f\n"
       "40:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -814,18 +812,18 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cbz x12, 42f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -833,13 +831,13 @@ void sve_hybrid_fp32_mla_6x4VL (
       "42:"  // Height 4: no bias
       "tbz %x[flags], #0, 43f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
@@ -874,8 +872,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "45:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -900,25 +898,25 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 49f\n"
       "48:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z25.s }, p5/Z, [x10]\n"
-      "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z3.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z2.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "ld1rqw { z1.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "cmp x27, #0x4\n"
+      "add x26, x26, #0x10\n"
+      "ld1w { z25.s }, p5/Z, [x10]\n"
+      "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.s, z25.s, z3.s[0]\n"
       "fmla z12.s, z25.s, z2.s[0]\n"
-      "fmla z9.s, z24.s, z3.s[0]\n"
-      "fmla z13.s, z24.s, z2.s[0]\n"
       "fmla z16.s, z25.s, z1.s[0]\n"
       "fmla z20.s, z25.s, z0.s[0]\n"
       "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      "fmla z9.s, z24.s, z3.s[0]\n"
+      "fmla z13.s, z24.s, z2.s[0]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "fmla z17.s, z24.s, z1.s[0]\n"
       "fmla z21.s, z24.s, z0.s[0]\n"
       "ld1w { z24.s }, p5/Z, [x10, #3, MUL VL]\n"
@@ -947,9 +945,9 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z14.s, z25.s, z2.s[1]\n"
       "fmla z18.s, z25.s, z1.s[1]\n"
       "fmla z22.s, z25.s, z0.s[1]\n"
+      "ld1w { z25.s }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z11.s, z24.s, z3.s[1]\n"
       "fmla z15.s, z24.s, z2.s[1]\n"
-      "ld1w { z25.s }, p5/Z, [x10, #-8, MUL VL]\n"
       "fmla z19.s, z24.s, z1.s[1]\n"
       "fmla z23.s, z24.s, z0.s[1]\n"
       "ld1w { z24.s }, p5/Z, [x10, #-7, MUL VL]\n"
@@ -994,20 +992,20 @@ void sve_hybrid_fp32_mla_6x4VL (
       "bgt 48b\n"
       "49:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z25.s }, p5/Z, [x10]\n"
-      "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
       "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      "ld1w { z25.s }, p5/Z, [x10]\n"
+      "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.s, z25.s, z0.s[0]\n"
       "fmla z12.s, z25.s, z1.s[0]\n"
-      "fmla z9.s, z24.s, z0.s[0]\n"
-      "fmla z13.s, z24.s, z1.s[0]\n"
       "fmla z16.s, z25.s, z2.s[0]\n"
       "fmla z20.s, z25.s, z3.s[0]\n"
       "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.s, z24.s, z0.s[0]\n"
+      "fmla z13.s, z24.s, z1.s[0]\n"
       "fmla z17.s, z24.s, z2.s[0]\n"
       "fmla z21.s, z24.s, z3.s[0]\n"
       "ld1w { z24.s }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1023,12 +1021,12 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 50f\n"
       "ld1w { z25.s }, p5/Z, [x10]\n"
       "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z25.s, z0.s[1]\n"
       "fmla z12.s, z25.s, z1.s[1]\n"
       "fmla z16.s, z25.s, z2.s[1]\n"
       "fmla z20.s, z25.s, z3.s[1]\n"
       "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.s, z24.s, z0.s[1]\n"
       "fmla z13.s, z24.s, z1.s[1]\n"
       "fmla z17.s, z24.s, z2.s[1]\n"
@@ -1046,12 +1044,12 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 50f\n"
       "ld1w { z25.s }, p5/Z, [x10]\n"
       "ld1w { z24.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z25.s, z0.s[2]\n"
       "fmla z12.s, z25.s, z1.s[2]\n"
       "fmla z16.s, z25.s, z2.s[2]\n"
       "fmla z20.s, z25.s, z3.s[2]\n"
       "ld1w { z25.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z9.s, z24.s, z0.s[2]\n"
       "fmla z13.s, z24.s, z1.s[2]\n"
       "fmla z17.s, z24.s, z2.s[2]\n"
@@ -1094,13 +1092,13 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z25.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z25.s }, p5/Z, [x21]\n"
       "ld1rw { z24.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z25.s\n"
       "fmin z9.s, p5/M, z9.s, z25.s\n"
@@ -1140,28 +1138,28 @@ void sve_hybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
       "52:"  // Height 4: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 41b\n"
       "b 80f\n"
       "53:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "54:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1174,18 +1172,18 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cbz x12, 55f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1197,16 +1195,16 @@ void sve_hybrid_fp32_mla_6x4VL (
       "55:"  // Height 5: no bias
       "tbz %x[flags], #0, 56f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x22]\n"
@@ -1247,8 +1245,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "58:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 59f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1276,29 +1274,29 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 62f\n"
       "61:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z29.s }, p5/Z, [x10]\n"
-      "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z4.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z3.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z1.s }, p0/Z, [x23]\n"
+      "cmp x27, #0x4\n"
+      "add x26, x26, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1w { z29.s }, p5/Z, [x10]\n"
       "fmla z8.s, z29.s, z4.s[0]\n"
       "fmla z12.s, z29.s, z3.s[0]\n"
-      "fmla z9.s, z28.s, z4.s[0]\n"
+      "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z16.s, z29.s, z2.s[0]\n"
       "fmla z20.s, z29.s, z1.s[0]\n"
+      "add x25, x25, #0x10\n"
       "fmla z24.s, z29.s, z0.s[0]\n"
-      "fmla z13.s, z28.s, z3.s[0]\n"
+      "fmla z9.s, z28.s, z4.s[0]\n"
       "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      "fmla z13.s, z28.s, z3.s[0]\n"
       "fmla z17.s, z28.s, z2.s[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z21.s, z28.s, z1.s[0]\n"
       "fmla z25.s, z28.s, z0.s[0]\n"
       "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1307,8 +1305,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z18.s, z29.s, z2.s[0]\n"
       "fmla z22.s, z29.s, z1.s[0]\n"
       "fmla z26.s, z29.s, z0.s[0]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z11.s, z28.s, z4.s[0]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #4, MUL VL]\n"
       "fmla z15.s, z28.s, z3.s[0]\n"
       "fmla z19.s, z28.s, z2.s[0]\n"
       "fmla z23.s, z28.s, z1.s[0]\n"
@@ -1319,8 +1317,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z16.s, z29.s, z2.s[1]\n"
       "fmla z20.s, z29.s, z1.s[1]\n"
       "fmla z24.s, z29.s, z0.s[1]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z9.s, z28.s, z4.s[1]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #6, MUL VL]\n"
       "fmla z13.s, z28.s, z3.s[1]\n"
       "fmla z17.s, z28.s, z2.s[1]\n"
       "fmla z21.s, z28.s, z1.s[1]\n"
@@ -1333,8 +1331,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z22.s, z29.s, z1.s[1]\n"
       "fmla z26.s, z29.s, z0.s[1]\n"
       "fmla z11.s, z28.s, z4.s[1]\n"
-      "fmla z15.s, z28.s, z3.s[1]\n"
       "ld1w { z29.s }, p5/Z, [x10, #-8, MUL VL]\n"
+      "fmla z15.s, z28.s, z3.s[1]\n"
       "fmla z19.s, z28.s, z2.s[1]\n"
       "fmla z23.s, z28.s, z1.s[1]\n"
       "fmla z27.s, z28.s, z0.s[1]\n"
@@ -1344,8 +1342,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z16.s, z29.s, z2.s[2]\n"
       "fmla z20.s, z29.s, z1.s[2]\n"
       "fmla z24.s, z29.s, z0.s[2]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z9.s, z28.s, z4.s[2]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #-6, MUL VL]\n"
       "fmla z13.s, z28.s, z3.s[2]\n"
       "fmla z17.s, z28.s, z2.s[2]\n"
       "fmla z21.s, z28.s, z1.s[2]\n"
@@ -1356,8 +1354,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z18.s, z29.s, z2.s[2]\n"
       "fmla z22.s, z29.s, z1.s[2]\n"
       "fmla z26.s, z29.s, z0.s[2]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z11.s, z28.s, z4.s[2]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #-4, MUL VL]\n"
       "fmla z15.s, z28.s, z3.s[2]\n"
       "fmla z19.s, z28.s, z2.s[2]\n"
       "fmla z23.s, z28.s, z1.s[2]\n"
@@ -1368,8 +1366,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z16.s, z29.s, z2.s[3]\n"
       "fmla z20.s, z29.s, z1.s[3]\n"
       "fmla z24.s, z29.s, z0.s[3]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z9.s, z28.s, z4.s[3]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #-2, MUL VL]\n"
       "fmla z13.s, z28.s, z3.s[3]\n"
       "fmla z17.s, z28.s, z2.s[3]\n"
       "fmla z21.s, z28.s, z1.s[3]\n"
@@ -1388,23 +1386,23 @@ void sve_hybrid_fp32_mla_6x4VL (
       "bgt 61b\n"
       "62:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z29.s }, p5/Z, [x10]\n"
-      "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
       "ld1rqw { z3.s }, p0/Z, [x23]\n"
       "ld1rqw { z4.s }, p0/Z, [x22]\n"
+      "ld1w { z29.s }, p5/Z, [x10]\n"
       "fmla z8.s, z29.s, z0.s[0]\n"
       "fmla z12.s, z29.s, z1.s[0]\n"
-      "fmla z9.s, z28.s, z0.s[0]\n"
-      "fmla z13.s, z28.s, z1.s[0]\n"
+      "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z16.s, z29.s, z2.s[0]\n"
       "fmla z20.s, z29.s, z3.s[0]\n"
       "fmla z24.s, z29.s, z4.s[0]\n"
-      "fmla z17.s, z28.s, z2.s[0]\n"
+      "fmla z9.s, z28.s, z0.s[0]\n"
       "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z13.s, z28.s, z1.s[0]\n"
+      "fmla z17.s, z28.s, z2.s[0]\n"
       "fmla z21.s, z28.s, z3.s[0]\n"
       "fmla z25.s, z28.s, z4.s[0]\n"
       "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1422,21 +1420,21 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 63f\n"
       "ld1w { z29.s }, p5/Z, [x10]\n"
       "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z29.s, z0.s[1]\n"
       "fmla z12.s, z29.s, z1.s[1]\n"
       "fmla z16.s, z29.s, z2.s[1]\n"
       "fmla z20.s, z29.s, z3.s[1]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.s, z29.s, z4.s[1]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z28.s, z0.s[1]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.s, z28.s, z1.s[1]\n"
       "fmla z17.s, z28.s, z2.s[1]\n"
       "fmla z21.s, z28.s, z3.s[1]\n"
       "fmla z25.s, z28.s, z4.s[1]\n"
       "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.s, z29.s, z0.s[1]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.s, z29.s, z0.s[1]\n"
       "fmla z14.s, z29.s, z1.s[1]\n"
       "fmla z18.s, z29.s, z2.s[1]\n"
       "fmla z22.s, z29.s, z3.s[1]\n"
@@ -1449,21 +1447,21 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 63f\n"
       "ld1w { z29.s }, p5/Z, [x10]\n"
       "ld1w { z28.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z29.s, z0.s[2]\n"
       "fmla z12.s, z29.s, z1.s[2]\n"
       "fmla z16.s, z29.s, z2.s[2]\n"
       "fmla z20.s, z29.s, z3.s[2]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.s, z29.s, z4.s[2]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z28.s, z0.s[2]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.s, z28.s, z1.s[2]\n"
       "fmla z17.s, z28.s, z2.s[2]\n"
       "fmla z21.s, z28.s, z3.s[2]\n"
       "fmla z25.s, z28.s, z4.s[2]\n"
       "ld1w { z28.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.s, z29.s, z0.s[2]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.s, z29.s, z0.s[2]\n"
       "fmla z14.s, z29.s, z1.s[2]\n"
       "fmla z18.s, z29.s, z2.s[2]\n"
       "fmla z22.s, z29.s, z3.s[2]\n"
@@ -1481,8 +1479,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "fmla z16.s, z29.s, z2.s[3]\n"
       "fmla z20.s, z29.s, z3.s[3]\n"
       "fmla z24.s, z29.s, z4.s[3]\n"
-      "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z9.s, z28.s, z0.s[3]\n"
+      "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n"
       "fmla z13.s, z28.s, z1.s[3]\n"
       "fmla z17.s, z28.s, z2.s[3]\n"
       "fmla z21.s, z28.s, z3.s[3]\n"
@@ -1505,14 +1503,14 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 58b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z29.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z29.s }, p5/Z, [x21]\n"
       "ld1rw { z28.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z29.s\n"
       "fmin z9.s, p5/M, z9.s, z29.s\n"
@@ -1560,22 +1558,22 @@ void sve_hybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
       "65:"  // Height 5: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1583,13 +1581,12 @@ void sve_hybrid_fp32_mla_6x4VL (
       "b 80f\n"
       "66:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "67:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1602,18 +1599,18 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cbz x12, 68f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "mov z12.d, z8.d\n"
       "mov z13.d, z9.d\n"
-      "mov z16.d, z8.d\n"
-      "mov z17.d, z9.d\n"
-      "mov z20.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "mov z14.d, z10.d\n"
       "mov z15.d, z11.d\n"
+      "mov z16.d, z8.d\n"
+      "mov z17.d, z9.d\n"
+      "addvl x12, x12, #4\n"
       "mov z18.d, z10.d\n"
       "mov z19.d, z11.d\n"
+      "mov z20.d, z8.d\n"
       "mov z21.d, z9.d\n"
       "mov z22.d, z10.d\n"
       "mov z23.d, z11.d\n"
@@ -1629,17 +1626,17 @@ void sve_hybrid_fp32_mla_6x4VL (
       "68:"  // Height 6: no bias
       "tbz %x[flags], #0, 69f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x24]\n"
       "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x23]\n"
@@ -1688,8 +1685,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "mov x28, #0x0\n"
       "71:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 72f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1720,29 +1717,29 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 75f\n"
       "74:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z1.s }, p5/Z, [x10]\n"
-      "ld1w { z0.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z7.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqw { z6.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x4\n"
       "ld1rqw { z5.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z4.s }, p0/Z, [x23]\n"
+      "cmp x27, #0x4\n"
+      "add x26, x26, #0x10\n"
       "ld1rqw { z3.s }, p0/Z, [x22]\n"
       "ld1rqw { z2.s }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1w { z1.s }, p5/Z, [x10]\n"
+      "ld1w { z0.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.s, z1.s, z7.s[0]\n"
       "fmla z12.s, z1.s, z6.s[0]\n"
-      "add x21, x21, #0x10\n"
       "fmla z16.s, z1.s, z5.s[0]\n"
       "fmla z20.s, z1.s, z4.s[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z24.s, z1.s, z3.s[0]\n"
       "fmla z28.s, z1.s, z2.s[0]\n"
       "ld1w { z1.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x21, x21, #0x10\n"
       "fmla z9.s, z0.s, z7.s[0]\n"
       "fmla z13.s, z0.s, z6.s[0]\n"
       "fmla z17.s, z0.s, z5.s[0]\n"
@@ -1850,24 +1847,24 @@ void sve_hybrid_fp32_mla_6x4VL (
       "bgt 74b\n"
       "75:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1w { z7.s }, p5/Z, [x10]\n"
-      "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "ld1rqw { z0.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
       "ld1rqw { z3.s }, p0/Z, [x23]\n"
       "ld1rqw { z4.s }, p0/Z, [x22]\n"
       "ld1rqw { z5.s }, p0/Z, [x21]\n"
+      "ld1w { z7.s }, p5/Z, [x10]\n"
+      "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n"
       "fmla z8.s, z7.s, z0.s[0]\n"
       "fmla z12.s, z7.s, z1.s[0]\n"
-      "fmla z9.s, z6.s, z0.s[0]\n"
-      "fmla z13.s, z6.s, z1.s[0]\n"
       "fmla z16.s, z7.s, z2.s[0]\n"
       "fmla z20.s, z7.s, z3.s[0]\n"
       "fmla z24.s, z7.s, z4.s[0]\n"
       "fmla z28.s, z7.s, z5.s[0]\n"
       "ld1w { z7.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.s, z6.s, z0.s[0]\n"
+      "fmla z13.s, z6.s, z1.s[0]\n"
       "fmla z17.s, z6.s, z2.s[0]\n"
       "fmla z21.s, z6.s, z3.s[0]\n"
       "fmla z25.s, z6.s, z4.s[0]\n"
@@ -1889,23 +1886,23 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 76f\n"
       "ld1w { z7.s }, p5/Z, [x10]\n"
       "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z7.s, z0.s[1]\n"
       "fmla z12.s, z7.s, z1.s[1]\n"
       "fmla z16.s, z7.s, z2.s[1]\n"
       "fmla z20.s, z7.s, z3.s[1]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.s, z7.s, z4.s[1]\n"
       "fmla z28.s, z7.s, z5.s[1]\n"
-      "fmla z9.s, z6.s, z0.s[1]\n"
       "ld1w { z7.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.s, z6.s, z0.s[1]\n"
       "fmla z13.s, z6.s, z1.s[1]\n"
       "fmla z17.s, z6.s, z2.s[1]\n"
       "fmla z21.s, z6.s, z3.s[1]\n"
       "fmla z25.s, z6.s, z4.s[1]\n"
       "fmla z29.s, z6.s, z5.s[1]\n"
       "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.s, z7.s, z0.s[1]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.s, z7.s, z0.s[1]\n"
       "fmla z14.s, z7.s, z1.s[1]\n"
       "fmla z18.s, z7.s, z2.s[1]\n"
       "fmla z22.s, z7.s, z3.s[1]\n"
@@ -1920,23 +1917,23 @@ void sve_hybrid_fp32_mla_6x4VL (
       "ble 76f\n"
       "ld1w { z7.s }, p5/Z, [x10]\n"
       "ld1w { z6.s }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x1\n"
       "fmla z8.s, z7.s, z0.s[2]\n"
       "fmla z12.s, z7.s, z1.s[2]\n"
       "fmla z16.s, z7.s, z2.s[2]\n"
       "fmla z20.s, z7.s, z3.s[2]\n"
+      "subs x27, x27, #0x1\n"
       "fmla z24.s, z7.s, z4.s[2]\n"
       "fmla z28.s, z7.s, z5.s[2]\n"
-      "fmla z9.s, z6.s, z0.s[2]\n"
       "ld1w { z7.s }, p5/Z, [x10, #2, MUL VL]\n"
+      "fmla z9.s, z6.s, z0.s[2]\n"
       "fmla z13.s, z6.s, z1.s[2]\n"
       "fmla z17.s, z6.s, z2.s[2]\n"
       "fmla z21.s, z6.s, z3.s[2]\n"
       "fmla z25.s, z6.s, z4.s[2]\n"
       "fmla z29.s, z6.s, z5.s[2]\n"
       "ld1w { z6.s }, p5/Z, [x10, #3, MUL VL]\n"
-      "fmla z10.s, z7.s, z0.s[2]\n"
       "addvl x10, x10, #4\n"
+      "fmla z10.s, z7.s, z0.s[2]\n"
       "fmla z14.s, z7.s, z1.s[2]\n"
       "fmla z18.s, z7.s, z2.s[2]\n"
       "fmla z22.s, z7.s, z3.s[2]\n"
@@ -1984,15 +1981,15 @@ void sve_hybrid_fp32_mla_6x4VL (
       "cmp x28, x20\n"
       "bne 71b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
+      "add x25, x9, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "tbz %x[flags], #1, 77f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p5/Z, [x21]\n"
       "ld1rw { z0.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z1.s\n"
       "fmin z9.s, p5/M, z9.s, z1.s\n"
@@ -2048,26 +2045,26 @@ void sve_hybrid_fp32_mla_6x4VL (
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z12.s }, p4, [x26]\n"
-      "st1w { z13.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z14.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z15.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z20.s }, p4, [x24]\n"
-      "st1w { z21.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z22.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z23.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
-      "st1w { z28.s }, p4, [x22]\n"
-      "st1w { z29.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z30.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z31.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z12.s }, p4, [x25]\n"
+      "st1w { z13.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z14.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z15.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z20.s }, p4, [x23]\n"
+      "st1w { z21.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z22.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z23.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z28.s }, p4, [x21]\n"
+      "st1w { z29.s }, p3, [x21, #1, MUL VL]\n"
+      "st1w { z30.s }, p2, [x21, #2, MUL VL]\n"
+      "st1w { z31.s }, p1, [x21, #3, MUL VL]\n"
       "78:"  // Height 6: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -2084,8 +2081,8 @@ void sve_hybrid_fp32_mla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp
index 5073e9cd7a..06a2d34767 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL.hpp
@@ -70,7 +70,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 1, 1> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 8, 1, 1> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel=sve_hybrid_fp32_mla_8x1VL;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/a64fx.cpp
index ea3d266ec1..17eaa5641d 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/a64fx.cpp
@@ -47,19 +47,18 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -107,10 +105,10 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "cmp %x[M], #0x2\n"
       "bgt 25f\n"
       "beq 13f\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p0.s, x20, x13\n"
@@ -128,8 +126,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "mov x10, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -148,22 +146,22 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x28, x28, #0x4\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
+      "addvl x12, x12, #1\n"
       "ld1rw { z0.s }, p1/Z, [x28]\n"
       "bgt 9b\n"
       "10:"  // Height 1: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x10, x10, #0x1\n"
-      "addvl x12, x12, #1\n"
       "cmp x10, x20\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
+      "addvl x12, x12, #1\n"
       "bne 6b\n"
       "tbz %x[flags], #1, 11f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p1/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
       "ld1rw { z16.s }, p1/Z, [x20]\n"
       "fmin z24.s, p1/M, z24.s, z17.s\n"
       "fmax z24.s, p1/M, z24.s, z16.s\n"
@@ -176,23 +174,23 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "bgt 2b\n"
       "b 98f\n"
       "13:"  // Height 2
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "14:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p0.s, x20, x13\n"
       "cbz x14, 15f\n"
       "ld1w { z24.s }, p1/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "b 17f\n"
       "15:"  // Height 2: no bias
       "tbz %x[flags], #0, 16f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z24.s }, p0/Z, [x11]\n"
       "add x20, x11, x20, LSL #2\n"
+      "ld1w { z24.s }, p0/Z, [x11]\n"
       "ld1w { z25.s }, p0/Z, [x20]\n"
       "b 17f\n"
       "16:"  // Height 2: no accumulate
@@ -202,8 +200,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "mov x10, #0x0\n"
       "18:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 19f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -226,28 +224,28 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x28, x28, #0x4\n"
       "subs x9, x9, #0x1\n"
+      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "add x27, x27, #0x4\n"
+      "fmla z25.s, p1/M, z16.s, z1.s\n"
       "addvl x12, x12, #1\n"
-      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "ld1rw { z0.s }, p1/Z, [x28]\n"
-      "fmla z25.s, p1/M, z16.s, z1.s\n"
       "ld1rw { z1.s }, p1/Z, [x27]\n"
       "bgt 21b\n"
       "22:"  // Height 2: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x10, x10, #0x1\n"
-      "addvl x12, x12, #1\n"
       "cmp x10, x20\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "bne 18b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "tbz %x[flags], #1, 23f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p1/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
       "ld1rw { z16.s }, p1/Z, [x20]\n"
       "fmin z24.s, p1/M, z24.s, z17.s\n"
       "fmin z25.s, p1/M, z25.s, z17.s\n"
@@ -256,33 +254,33 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "23:"  // Height 2: No activation
       "st1w { z24.s }, p0, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p0, [x28]\n"
+      "st1w { z25.s }, p0, [x27]\n"
       "24:"  // Height 2: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 14b\n"
       "b 98f\n"
       "25:"  // Height 3
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "26:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p0.s, x20, x13\n"
       "cbz x14, 27f\n"
       "ld1w { z24.s }, p1/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "b 29f\n"
       "27:"  // Height 3: no bias
       "tbz %x[flags], #0, 28f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x11, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z24.s }, p0/Z, [x11]\n"
-      "add x20, x11, x21, LSL #2\n"
-      "ld1w { z25.s }, p0/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
+      "ld1w { z25.s }, p0/Z, [x21]\n"
       "ld1w { z26.s }, p0/Z, [x20]\n"
       "b 29f\n"
       "28:"  // Height 3: no accumulate
@@ -293,8 +291,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "mov x10, #0x0\n"
       "30:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 31f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -321,13 +319,13 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x28, x28, #0x4\n"
       "subs x9, x9, #0x1\n"
+      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "add x27, x27, #0x4\n"
       "add x26, x26, #0x4\n"
-      "addvl x12, x12, #1\n"
-      "fmla z24.s, p1/M, z16.s, z0.s\n"
-      "ld1rw { z0.s }, p1/Z, [x28]\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
       "fmla z26.s, p1/M, z16.s, z2.s\n"
+      "addvl x12, x12, #1\n"
+      "ld1rw { z0.s }, p1/Z, [x28]\n"
       "ld1rw { z1.s }, p1/Z, [x27]\n"
       "ld1rw { z2.s }, p1/Z, [x26]\n"
       "bgt 33b\n"
@@ -335,19 +333,19 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x10, x10, #0x1\n"
-      "addvl x12, x12, #1\n"
       "cmp x10, x20\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, p1/M, z16.s, z2.s\n"
       "bne 30b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
+      "add x26, x27, x20, LSL #2\n"
       "tbz %x[flags], #1, 35f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p1/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
       "ld1rw { z16.s }, p1/Z, [x20]\n"
       "fmin z24.s, p1/M, z24.s, z17.s\n"
       "fmin z25.s, p1/M, z25.s, z17.s\n"
@@ -358,37 +356,37 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "35:"  // Height 3: No activation
       "st1w { z24.s }, p0, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p0, [x28]\n"
-      "st1w { z26.s }, p0, [x27]\n"
+      "st1w { z25.s }, p0, [x27]\n"
+      "st1w { z26.s }, p0, [x26]\n"
       "36:"  // Height 3: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 26b\n"
       "b 98f\n"
       "37:"  // Height 4
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "38:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p0.s, x20, x13\n"
       "cbz x14, 39f\n"
       "ld1w { z24.s }, p1/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "b 41f\n"
       "39:"  // Height 4: no bias
       "tbz %x[flags], #0, 40f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x11, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z24.s }, p0/Z, [x11]\n"
-      "add x20, x11, x21, LSL #2\n"
-      "ld1w { z25.s }, p0/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
-      "ld1w { z26.s }, p0/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z25.s }, p0/Z, [x22]\n"
+      "ld1w { z26.s }, p0/Z, [x21]\n"
       "ld1w { z27.s }, p0/Z, [x20]\n"
       "b 41f\n"
       "40:"  // Height 4: no accumulate
@@ -400,8 +398,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "mov x10, #0x0\n"
       "42:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 43f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -432,38 +430,38 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x28, x28, #0x4\n"
       "subs x9, x9, #0x1\n"
+      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "add x27, x27, #0x4\n"
       "add x26, x26, #0x4\n"
+      "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "fmla z26.s, p1/M, z16.s, z2.s\n"
       "add x25, x25, #0x4\n"
+      "fmla z27.s, p1/M, z16.s, z3.s\n"
       "addvl x12, x12, #1\n"
-      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "ld1rw { z0.s }, p1/Z, [x28]\n"
-      "fmla z25.s, p1/M, z16.s, z1.s\n"
       "ld1rw { z1.s }, p1/Z, [x27]\n"
-      "fmla z26.s, p1/M, z16.s, z2.s\n"
       "ld1rw { z2.s }, p1/Z, [x26]\n"
-      "fmla z27.s, p1/M, z16.s, z3.s\n"
       "ld1rw { z3.s }, p1/Z, [x25]\n"
       "bgt 45b\n"
       "46:"  // Height 4: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x10, x10, #0x1\n"
-      "addvl x12, x12, #1\n"
       "cmp x10, x20\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, p1/M, z16.s, z2.s\n"
       "fmla z27.s, p1/M, z16.s, z3.s\n"
       "bne 42b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
       "tbz %x[flags], #1, 47f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p1/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
       "ld1rw { z16.s }, p1/Z, [x20]\n"
       "fmin z24.s, p1/M, z24.s, z17.s\n"
       "fmin z25.s, p1/M, z25.s, z17.s\n"
@@ -476,41 +474,41 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "47:"  // Height 4: No activation
       "st1w { z24.s }, p0, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p0, [x28]\n"
-      "st1w { z26.s }, p0, [x27]\n"
-      "st1w { z27.s }, p0, [x26]\n"
+      "st1w { z25.s }, p0, [x27]\n"
+      "st1w { z26.s }, p0, [x26]\n"
+      "st1w { z27.s }, p0, [x25]\n"
       "48:"  // Height 4: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 38b\n"
       "b 98f\n"
       "49:"  // Height 5
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "50:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p0.s, x20, x13\n"
       "cbz x14, 51f\n"
       "ld1w { z24.s }, p1/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "mov z28.d, z24.d\n"
       "b 53f\n"
       "51:"  // Height 5: no bias
       "tbz %x[flags], #0, 52f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x11, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z24.s }, p0/Z, [x11]\n"
-      "add x20, x11, x21, LSL #2\n"
-      "ld1w { z25.s }, p0/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
-      "ld1w { z26.s }, p0/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
-      "ld1w { z27.s }, p0/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z25.s }, p0/Z, [x23]\n"
+      "ld1w { z26.s }, p0/Z, [x22]\n"
+      "ld1w { z27.s }, p0/Z, [x21]\n"
       "ld1w { z28.s }, p0/Z, [x20]\n"
       "b 53f\n"
       "52:"  // Height 5: no accumulate
@@ -523,8 +521,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "mov x10, #0x0\n"
       "54:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 55f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -559,19 +557,19 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x28, x28, #0x4\n"
       "subs x9, x9, #0x1\n"
+      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "add x27, x27, #0x4\n"
       "add x26, x26, #0x4\n"
-      "add x25, x25, #0x4\n"
-      "add x24, x24, #0x4\n"
-      "addvl x12, x12, #1\n"
-      "fmla z24.s, p1/M, z16.s, z0.s\n"
-      "ld1rw { z0.s }, p1/Z, [x28]\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
       "fmla z26.s, p1/M, z16.s, z2.s\n"
+      "add x25, x25, #0x4\n"
+      "add x24, x24, #0x4\n"
       "fmla z27.s, p1/M, z16.s, z3.s\n"
+      "ld1rw { z0.s }, p1/Z, [x28]\n"
+      "addvl x12, x12, #1\n"
+      "fmla z28.s, p1/M, z16.s, z4.s\n"
       "ld1rw { z1.s }, p1/Z, [x27]\n"
       "ld1rw { z2.s }, p1/Z, [x26]\n"
-      "fmla z28.s, p1/M, z16.s, z4.s\n"
       "ld1rw { z3.s }, p1/Z, [x25]\n"
       "ld1rw { z4.s }, p1/Z, [x24]\n"
       "bgt 57b\n"
@@ -579,23 +577,23 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x10, x10, #0x1\n"
-      "addvl x12, x12, #1\n"
       "cmp x10, x20\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, p1/M, z16.s, z2.s\n"
       "fmla z27.s, p1/M, z16.s, z3.s\n"
       "fmla z28.s, p1/M, z16.s, z4.s\n"
       "bne 54b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "tbz %x[flags], #1, 59f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p1/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
       "ld1rw { z16.s }, p1/Z, [x20]\n"
       "fmin z24.s, p1/M, z24.s, z17.s\n"
       "fmin z25.s, p1/M, z25.s, z17.s\n"
@@ -610,44 +608,44 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "59:"  // Height 5: No activation
       "st1w { z24.s }, p0, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p0, [x28]\n"
-      "st1w { z26.s }, p0, [x27]\n"
-      "st1w { z27.s }, p0, [x26]\n"
-      "st1w { z28.s }, p0, [x25]\n"
+      "st1w { z25.s }, p0, [x27]\n"
+      "st1w { z26.s }, p0, [x26]\n"
+      "st1w { z27.s }, p0, [x25]\n"
+      "st1w { z28.s }, p0, [x24]\n"
       "60:"  // Height 5: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 50b\n"
       "b 98f\n"
       "61:"  // Height 6
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "62:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p0.s, x20, x13\n"
       "cbz x14, 63f\n"
       "ld1w { z24.s }, p1/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "mov z28.d, z24.d\n"
       "mov z29.d, z24.d\n"
       "b 65f\n"
       "63:"  // Height 6: no bias
       "tbz %x[flags], #0, 64f\n"
-      "ldr x22, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x24, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x11, x24, LSL #2\n"
+      "add x20, x23, x24, LSL #2\n"
       "ld1w { z24.s }, p0/Z, [x11]\n"
-      "add x20, x11, x22, LSL #2\n"
-      "ld1w { z25.s }, p0/Z, [x20]\n"
-      "add x20, x20, x22, LSL #2\n"
+      "add x22, x20, x24, LSL #2\n"
+      "add x21, x22, x24, LSL #2\n"
+      "ld1w { z25.s }, p0/Z, [x23]\n"
       "ld1w { z26.s }, p0/Z, [x20]\n"
-      "add x20, x20, x22, LSL #2\n"
-      "ld1w { z27.s }, p0/Z, [x20]\n"
-      "add x21, x20, x22, LSL #2\n"
-      "add x20, x21, x22, LSL #2\n"
+      "add x20, x21, x24, LSL #2\n"
+      "ld1w { z27.s }, p0/Z, [x22]\n"
       "ld1w { z28.s }, p0/Z, [x21]\n"
       "ld1w { z29.s }, p0/Z, [x20]\n"
       "b 65f\n"
@@ -662,8 +660,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "mov x10, #0x0\n"
       "66:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 67f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -702,48 +700,48 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x28, x28, #0x4\n"
       "subs x9, x9, #0x1\n"
+      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "add x27, x27, #0x4\n"
       "add x26, x26, #0x4\n"
+      "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "fmla z26.s, p1/M, z16.s, z2.s\n"
       "add x25, x25, #0x4\n"
       "add x24, x24, #0x4\n"
+      "fmla z27.s, p1/M, z16.s, z3.s\n"
+      "fmla z28.s, p1/M, z16.s, z4.s\n"
       "add x23, x23, #0x4\n"
       "addvl x12, x12, #1\n"
-      "fmla z24.s, p1/M, z16.s, z0.s\n"
+      "fmla z29.s, p1/M, z16.s, z5.s\n"
       "ld1rw { z0.s }, p1/Z, [x28]\n"
-      "fmla z25.s, p1/M, z16.s, z1.s\n"
       "ld1rw { z1.s }, p1/Z, [x27]\n"
-      "fmla z26.s, p1/M, z16.s, z2.s\n"
       "ld1rw { z2.s }, p1/Z, [x26]\n"
-      "fmla z27.s, p1/M, z16.s, z3.s\n"
-      "fmla z28.s, p1/M, z16.s, z4.s\n"
       "ld1rw { z3.s }, p1/Z, [x25]\n"
       "ld1rw { z4.s }, p1/Z, [x24]\n"
-      "fmla z29.s, p1/M, z16.s, z5.s\n"
       "ld1rw { z5.s }, p1/Z, [x23]\n"
       "bgt 69b\n"
       "70:"  // Height 6: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x10, x10, #0x1\n"
-      "addvl x12, x12, #1\n"
       "cmp x10, x20\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, p1/M, z16.s, z2.s\n"
       "fmla z27.s, p1/M, z16.s, z3.s\n"
       "fmla z28.s, p1/M, z16.s, z4.s\n"
       "fmla z29.s, p1/M, z16.s, z5.s\n"
       "bne 66b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "tbz %x[flags], #1, 71f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p1/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
       "ld1rw { z16.s }, p1/Z, [x20]\n"
       "fmin z24.s, p1/M, z24.s, z17.s\n"
       "fmin z25.s, p1/M, z25.s, z17.s\n"
@@ -760,29 +758,29 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "71:"  // Height 6: No activation
       "st1w { z24.s }, p0, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p0, [x28]\n"
-      "st1w { z26.s }, p0, [x27]\n"
-      "st1w { z27.s }, p0, [x26]\n"
-      "st1w { z28.s }, p0, [x25]\n"
-      "st1w { z29.s }, p0, [x24]\n"
+      "st1w { z25.s }, p0, [x27]\n"
+      "st1w { z26.s }, p0, [x26]\n"
+      "st1w { z27.s }, p0, [x25]\n"
+      "st1w { z28.s }, p0, [x24]\n"
+      "st1w { z29.s }, p0, [x23]\n"
       "72:"  // Height 6: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 62b\n"
       "b 98f\n"
       "73:"  // Height 7
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "74:"  // Height 7: Column loop
       "mov x20, #0x0\n"
       "whilelt p0.s, x20, x13\n"
       "cbz x14, 75f\n"
       "ld1w { z24.s }, p1/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "mov z28.d, z24.d\n"
       "mov z29.d, z24.d\n"
@@ -790,17 +788,17 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "b 77f\n"
       "75:"  // Height 7: no bias
       "tbz %x[flags], #0, 76f\n"
-      "ldr x23, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x24, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x11, x24, LSL #2\n"
+      "add x20, x21, x24, LSL #2\n"
       "ld1w { z24.s }, p0/Z, [x11]\n"
-      "add x20, x11, x23, LSL #2\n"
-      "ld1w { z25.s }, p0/Z, [x20]\n"
-      "add x20, x20, x23, LSL #2\n"
+      "add x23, x20, x24, LSL #2\n"
+      "add x22, x23, x24, LSL #2\n"
+      "ld1w { z25.s }, p0/Z, [x21]\n"
       "ld1w { z26.s }, p0/Z, [x20]\n"
-      "add x20, x20, x23, LSL #2\n"
-      "ld1w { z27.s }, p0/Z, [x20]\n"
-      "add x22, x20, x23, LSL #2\n"
-      "add x21, x22, x23, LSL #2\n"
-      "add x20, x21, x23, LSL #2\n"
+      "add x21, x22, x24, LSL #2\n"
+      "add x20, x21, x24, LSL #2\n"
+      "ld1w { z27.s }, p0/Z, [x23]\n"
       "ld1w { z28.s }, p0/Z, [x22]\n"
       "ld1w { z29.s }, p0/Z, [x21]\n"
       "ld1w { z30.s }, p0/Z, [x20]\n"
@@ -817,8 +815,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "mov x10, #0x0\n"
       "78:"  // Height 7: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 79f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -861,25 +859,25 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x28, x28, #0x4\n"
       "subs x9, x9, #0x1\n"
+      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "add x27, x27, #0x4\n"
       "add x26, x26, #0x4\n"
+      "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "fmla z26.s, p1/M, z16.s, z2.s\n"
       "add x25, x25, #0x4\n"
       "add x24, x24, #0x4\n"
+      "fmla z27.s, p1/M, z16.s, z3.s\n"
+      "ld1rw { z0.s }, p1/Z, [x28]\n"
       "add x23, x23, #0x4\n"
       "add x22, x22, #0x4\n"
-      "fmla z24.s, p1/M, z16.s, z0.s\n"
-      "ld1rw { z0.s }, p1/Z, [x28]\n"
+      "fmla z28.s, p1/M, z16.s, z4.s\n"
+      "fmla z29.s, p1/M, z16.s, z5.s\n"
       "addvl x12, x12, #1\n"
-      "fmla z25.s, p1/M, z16.s, z1.s\n"
-      "fmla z26.s, p1/M, z16.s, z2.s\n"
       "ld1rw { z1.s }, p1/Z, [x27]\n"
-      "fmla z27.s, p1/M, z16.s, z3.s\n"
+      "fmla z30.s, p1/M, z16.s, z6.s\n"
       "ld1rw { z2.s }, p1/Z, [x26]\n"
-      "fmla z28.s, p1/M, z16.s, z4.s\n"
       "ld1rw { z3.s }, p1/Z, [x25]\n"
-      "fmla z29.s, p1/M, z16.s, z5.s\n"
       "ld1rw { z4.s }, p1/Z, [x24]\n"
-      "fmla z30.s, p1/M, z16.s, z6.s\n"
       "ld1rw { z5.s }, p1/Z, [x23]\n"
       "ld1rw { z6.s }, p1/Z, [x22]\n"
       "bgt 81b\n"
@@ -887,10 +885,10 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x10, x10, #0x1\n"
-      "addvl x12, x12, #1\n"
       "cmp x10, x20\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, p1/M, z16.s, z2.s\n"
       "fmla z27.s, p1/M, z16.s, z3.s\n"
       "fmla z28.s, p1/M, z16.s, z4.s\n"
@@ -898,16 +896,16 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "fmla z30.s, p1/M, z16.s, z6.s\n"
       "bne 78b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "tbz %x[flags], #1, 83f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p1/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
       "ld1rw { z16.s }, p1/Z, [x20]\n"
       "fmin z24.s, p1/M, z24.s, z17.s\n"
       "fmin z25.s, p1/M, z25.s, z17.s\n"
@@ -926,12 +924,12 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "83:"  // Height 7: No activation
       "st1w { z24.s }, p0, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p0, [x28]\n"
-      "st1w { z26.s }, p0, [x27]\n"
-      "st1w { z27.s }, p0, [x26]\n"
-      "st1w { z28.s }, p0, [x25]\n"
-      "st1w { z29.s }, p0, [x24]\n"
-      "st1w { z30.s }, p0, [x23]\n"
+      "st1w { z25.s }, p0, [x27]\n"
+      "st1w { z26.s }, p0, [x26]\n"
+      "st1w { z27.s }, p0, [x25]\n"
+      "st1w { z28.s }, p0, [x24]\n"
+      "st1w { z29.s }, p0, [x23]\n"
+      "st1w { z30.s }, p0, [x22]\n"
       "84:"  // Height 7: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
@@ -939,21 +937,20 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "b 98f\n"
       "85:"  // Height 8
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x20\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x11\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "86:"  // Height 8: Column loop
       "mov x20, #0x0\n"
       "whilelt p0.s, x20, x13\n"
       "cbz x14, 87f\n"
       "ld1w { z24.s }, p1/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "mov z28.d, z24.d\n"
       "mov z29.d, z24.d\n"
@@ -962,20 +959,20 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "b 89f\n"
       "87:"  // Height 8: no bias
       "tbz %x[flags], #0, 88f\n"
-      "ldr x23, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x24, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x11, x24, LSL #2\n"
+      "add x21, x22, x24, LSL #2\n"
       "ld1w { z24.s }, p0/Z, [x11]\n"
-      "add x20, x11, x23, LSL #2\n"
-      "ld1w { z25.s }, p0/Z, [x20]\n"
-      "add x20, x20, x23, LSL #2\n"
-      "ld1w { z26.s }, p0/Z, [x20]\n"
-      "add x20, x20, x23, LSL #2\n"
-      "ld1w { z27.s }, p0/Z, [x20]\n"
-      "add x22, x20, x23, LSL #2\n"
-      "add x20, x22, x23, LSL #2\n"
-      "add x21, x20, x23, LSL #2\n"
-      "ld1w { z28.s }, p0/Z, [x22]\n"
-      "ld1w { z29.s }, p0/Z, [x20]\n"
-      "add x20, x21, x23, LSL #2\n"
+      "add x23, x21, x24, LSL #2\n"
+      "add x20, x23, x24, LSL #2\n"
+      "ld1w { z25.s }, p0/Z, [x22]\n"
+      "ld1w { z26.s }, p0/Z, [x21]\n"
+      "add x22, x20, x24, LSL #2\n"
+      "add x21, x22, x24, LSL #2\n"
+      "ld1w { z27.s }, p0/Z, [x23]\n"
+      "ld1w { z28.s }, p0/Z, [x20]\n"
+      "add x20, x21, x24, LSL #2\n"
+      "ld1w { z29.s }, p0/Z, [x22]\n"
       "ld1w { z30.s }, p0/Z, [x21]\n"
       "ld1w { z31.s }, p0/Z, [x20]\n"
       "b 89f\n"
@@ -992,8 +989,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "mov x10, #0x0\n"
       "90:"  // Height 8: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 91f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1040,28 +1037,28 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x28, x28, #0x4\n"
       "subs x9, x9, #0x1\n"
+      "fmla z24.s, p1/M, z16.s, z0.s\n"
       "add x27, x27, #0x4\n"
       "add x26, x26, #0x4\n"
+      "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "fmla z26.s, p1/M, z16.s, z2.s\n"
       "add x25, x25, #0x4\n"
       "add x24, x24, #0x4\n"
+      "fmla z27.s, p1/M, z16.s, z3.s\n"
+      "fmla z28.s, p1/M, z16.s, z4.s\n"
       "add x23, x23, #0x4\n"
       "add x22, x22, #0x4\n"
-      "fmla z24.s, p1/M, z16.s, z0.s\n"
+      "fmla z29.s, p1/M, z16.s, z5.s\n"
       "ld1rw { z0.s }, p1/Z, [x28]\n"
       "add x21, x21, #0x4\n"
       "addvl x12, x12, #1\n"
-      "fmla z25.s, p1/M, z16.s, z1.s\n"
       "ld1rw { z1.s }, p1/Z, [x27]\n"
-      "fmla z26.s, p1/M, z16.s, z2.s\n"
+      "fmla z30.s, p1/M, z16.s, z6.s\n"
+      "fmla z31.s, p1/M, z16.s, z7.s\n"
       "ld1rw { z2.s }, p1/Z, [x26]\n"
-      "fmla z27.s, p1/M, z16.s, z3.s\n"
       "ld1rw { z3.s }, p1/Z, [x25]\n"
-      "fmla z28.s, p1/M, z16.s, z4.s\n"
       "ld1rw { z4.s }, p1/Z, [x24]\n"
-      "fmla z29.s, p1/M, z16.s, z5.s\n"
       "ld1rw { z5.s }, p1/Z, [x23]\n"
-      "fmla z30.s, p1/M, z16.s, z6.s\n"
-      "fmla z31.s, p1/M, z16.s, z7.s\n"
       "ld1rw { z6.s }, p1/Z, [x22]\n"
       "ld1rw { z7.s }, p1/Z, [x21]\n"
       "bgt 93b\n"
@@ -1069,10 +1066,10 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "ld1w { z16.s }, p1/Z, [x12]\n"
       "add x10, x10, #0x1\n"
-      "addvl x12, x12, #1\n"
       "cmp x10, x20\n"
       "fmla z24.s, p1/M, z16.s, z0.s\n"
       "fmla z25.s, p1/M, z16.s, z1.s\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, p1/M, z16.s, z2.s\n"
       "fmla z27.s, p1/M, z16.s, z3.s\n"
       "fmla z28.s, p1/M, z16.s, z4.s\n"
@@ -1081,17 +1078,17 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "fmla z31.s, p1/M, z16.s, z7.s\n"
       "bne 90b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "tbz %x[flags], #1, 95f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p1/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p1/Z, [x21]\n"
       "ld1rw { z16.s }, p1/Z, [x20]\n"
       "fmin z24.s, p1/M, z24.s, z17.s\n"
       "fmin z25.s, p1/M, z25.s, z17.s\n"
@@ -1112,13 +1109,13 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "95:"  // Height 8: No activation
       "st1w { z24.s }, p0, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p0, [x28]\n"
-      "st1w { z26.s }, p0, [x27]\n"
-      "st1w { z27.s }, p0, [x26]\n"
-      "st1w { z28.s }, p0, [x25]\n"
-      "st1w { z29.s }, p0, [x24]\n"
-      "st1w { z30.s }, p0, [x23]\n"
-      "st1w { z31.s }, p0, [x22]\n"
+      "st1w { z25.s }, p0, [x27]\n"
+      "st1w { z26.s }, p0, [x26]\n"
+      "st1w { z27.s }, p0, [x25]\n"
+      "st1w { z28.s }, p0, [x24]\n"
+      "st1w { z29.s }, p0, [x23]\n"
+      "st1w { z30.s }, p0, [x22]\n"
+      "st1w { z31.s }, p0, [x21]\n"
       "96:"  // Height 8: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
@@ -1135,8 +1132,8 @@ void sve_hybrid_fp32_mla_8x1VL_a64fx (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "98:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z16", "z17", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp
index edfb3c0828..b79934094e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32_mla_8x1VL/generic.cpp
@@ -47,19 +47,18 @@ void sve_hybrid_fp32_mla_8x1VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -80,7 +79,6 @@ void sve_hybrid_fp32_mla_8x1VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -107,10 +105,10 @@ void sve_hybrid_fp32_mla_8x1VL (
       "cmp %x[M], #0x2\n"
       "bgt 27f\n"
       "beq 14f\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p1.s, x20, x13\n"
@@ -128,8 +126,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "mov x10, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -145,50 +143,50 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z19.s }, p2/Z, [x12]\n"
-      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
-      "sub x9, x9, #0x4\n"
-      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1rqw { z0.s }, p0/Z, [x28]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
+      "fmla z24.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #1, MUL VL]\n"
+      "fmla z24.s, z16.s, z0.s[1]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #2, MUL VL]\n"
+      "fmla z24.s, z16.s, z0.s[2]\n"
       "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
+      "sub x9, x9, #0x4\n"
       "cmp x9, #0x4\n"
-      "addvl x12, x12, #4\n"
-      "ld1rqw { z0.s }, p0/Z, [x28]\n"
-      "add x28, x28, #0x10\n"
-      "fmla z24.s, z19.s, z0.s[0]\n"
-      "fmla z24.s, z18.s, z0.s[1]\n"
-      "fmla z24.s, z17.s, z0.s[2]\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
+      "add x28, x28, #0x10\n"
+      "addvl x12, x12, #4\n"
       "bgt 9b\n"
       "10:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x9\n"
+      "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
-      "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "fmla z24.s, z16.s, z0.s[0]\n"
+      "addvl x12, x12, #1\n"
       "ble 11f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[1]\n"
+      "addvl x12, x12, #1\n"
       "ble 11f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[2]\n"
+      "addvl x12, x12, #1\n"
       "ble 11f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
+      "addvl x12, x12, #1\n"
       "11:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x10, x10, #0x1\n"
       "cmp x10, x20\n"
       "bne 6b\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "fmin z24.s, p2/M, z24.s, z17.s\n"
       "fmax z24.s, p2/M, z24.s, z16.s\n"
@@ -201,23 +199,23 @@ void sve_hybrid_fp32_mla_8x1VL (
       "bgt 2b\n"
       "b 106f\n"
       "14:"  // Height 2
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p1.s, x20, x13\n"
       "cbz x14, 16f\n"
       "ld1w { z24.s }, p2/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "b 18f\n"
       "16:"  // Height 2: no bias
       "tbz %x[flags], #0, 17f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z24.s }, p1/Z, [x11]\n"
       "add x20, x11, x20, LSL #2\n"
+      "ld1w { z24.s }, p1/Z, [x11]\n"
       "ld1w { z25.s }, p1/Z, [x20]\n"
       "b 18f\n"
       "17:"  // Height 2: no accumulate
@@ -227,8 +225,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "mov x10, #0x0\n"
       "19:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -247,63 +245,63 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 23f\n"
       "22:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z19.s }, p2/Z, [x12]\n"
-      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
+      "ld1rqw { z1.s }, p0/Z, [x28]\n"
+      "ld1rqw { z0.s }, p0/Z, [x27]\n"
       "sub x9, x9, #0x4\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
+      "fmla z24.s, z16.s, z1.s[0]\n"
+      "fmla z25.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #1, MUL VL]\n"
+      "fmla z24.s, z16.s, z1.s[1]\n"
+      "fmla z25.s, z16.s, z0.s[1]\n"
       "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
       "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
-      "cmp x9, #0x4\n"
-      "addvl x12, x12, #4\n"
-      "ld1rqw { z1.s }, p0/Z, [x28]\n"
-      "add x28, x28, #0x10\n"
-      "ld1rqw { z0.s }, p0/Z, [x27]\n"
-      "add x27, x27, #0x10\n"
-      "fmla z24.s, z19.s, z1.s[0]\n"
-      "fmla z25.s, z19.s, z0.s[0]\n"
-      "fmla z24.s, z18.s, z1.s[1]\n"
-      "fmla z25.s, z18.s, z0.s[1]\n"
       "fmla z24.s, z17.s, z1.s[2]\n"
       "fmla z25.s, z17.s, z0.s[2]\n"
+      "cmp x9, #0x4\n"
+      "add x28, x28, #0x10\n"
       "fmla z24.s, z16.s, z1.s[3]\n"
       "fmla z25.s, z16.s, z0.s[3]\n"
+      "add x27, x27, #0x10\n"
+      "addvl x12, x12, #4\n"
       "bgt 22b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z16.s }, p2/Z, [x12]\n"
-      "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "ld1rqw { z1.s }, p0/Z, [x27]\n"
+      "subs x9, x9, #0x1\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
       "fmla z24.s, z16.s, z0.s[0]\n"
       "fmla z25.s, z16.s, z1.s[0]\n"
+      "addvl x12, x12, #1\n"
       "ble 24f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[1]\n"
       "fmla z25.s, z16.s, z1.s[1]\n"
+      "addvl x12, x12, #1\n"
       "ble 24f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[2]\n"
       "fmla z25.s, z16.s, z1.s[2]\n"
+      "addvl x12, x12, #1\n"
       "ble 24f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
       "fmla z25.s, z16.s, z1.s[3]\n"
+      "addvl x12, x12, #1\n"
       "24:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x10, x10, #0x1\n"
       "cmp x10, x20\n"
       "bne 19b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "fmin z24.s, p2/M, z24.s, z17.s\n"
       "fmin z25.s, p2/M, z25.s, z17.s\n"
@@ -312,33 +310,33 @@ void sve_hybrid_fp32_mla_8x1VL (
       "25:"  // Height 2: No activation
       "st1w { z24.s }, p1, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p1, [x28]\n"
+      "st1w { z25.s }, p1, [x27]\n"
       "26:"  // Height 2: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 15b\n"
       "b 106f\n"
       "27:"  // Height 3
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p1.s, x20, x13\n"
       "cbz x14, 29f\n"
       "ld1w { z24.s }, p2/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "b 31f\n"
       "29:"  // Height 3: no bias
       "tbz %x[flags], #0, 30f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x11, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z24.s }, p1/Z, [x11]\n"
-      "add x20, x11, x21, LSL #2\n"
-      "ld1w { z25.s }, p1/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
+      "ld1w { z25.s }, p1/Z, [x21]\n"
       "ld1w { z26.s }, p1/Z, [x20]\n"
       "b 31f\n"
       "30:"  // Height 3: no accumulate
@@ -349,8 +347,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "mov x10, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -372,62 +370,62 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 36f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z19.s }, p2/Z, [x12]\n"
-      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
+      "ld1rqw { z2.s }, p0/Z, [x28]\n"
+      "ld1rqw { z1.s }, p0/Z, [x27]\n"
       "sub x9, x9, #0x4\n"
+      "ld1rqw { z0.s }, p0/Z, [x26]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
+      "fmla z24.s, z16.s, z2.s[0]\n"
+      "fmla z25.s, z16.s, z1.s[0]\n"
+      "fmla z26.s, z16.s, z0.s[0]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #1, MUL VL]\n"
+      "fmla z24.s, z16.s, z2.s[1]\n"
       "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
+      "fmla z25.s, z16.s, z1.s[1]\n"
+      "fmla z26.s, z16.s, z0.s[1]\n"
       "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
       "cmp x9, #0x4\n"
-      "addvl x12, x12, #4\n"
-      "ld1rqw { z2.s }, p0/Z, [x28]\n"
-      "add x28, x28, #0x10\n"
-      "ld1rqw { z1.s }, p0/Z, [x27]\n"
-      "add x27, x27, #0x10\n"
-      "ld1rqw { z0.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "fmla z24.s, z19.s, z2.s[0]\n"
-      "fmla z25.s, z19.s, z1.s[0]\n"
-      "fmla z26.s, z19.s, z0.s[0]\n"
-      "fmla z24.s, z18.s, z2.s[1]\n"
-      "fmla z25.s, z18.s, z1.s[1]\n"
-      "fmla z26.s, z18.s, z0.s[1]\n"
       "fmla z24.s, z17.s, z2.s[2]\n"
       "fmla z25.s, z17.s, z1.s[2]\n"
+      "add x28, x28, #0x10\n"
+      "add x27, x27, #0x10\n"
       "fmla z26.s, z17.s, z0.s[2]\n"
       "fmla z24.s, z16.s, z2.s[3]\n"
+      "add x26, x26, #0x10\n"
+      "addvl x12, x12, #4\n"
       "fmla z25.s, z16.s, z1.s[3]\n"
       "fmla z26.s, z16.s, z0.s[3]\n"
       "bgt 35b\n"
       "36:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z16.s }, p2/Z, [x12]\n"
-      "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "ld1rqw { z1.s }, p0/Z, [x27]\n"
+      "subs x9, x9, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
       "fmla z24.s, z16.s, z0.s[0]\n"
       "fmla z25.s, z16.s, z1.s[0]\n"
       "fmla z26.s, z16.s, z2.s[0]\n"
+      "addvl x12, x12, #1\n"
       "ble 37f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[1]\n"
       "fmla z25.s, z16.s, z1.s[1]\n"
       "fmla z26.s, z16.s, z2.s[1]\n"
+      "addvl x12, x12, #1\n"
       "ble 37f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[2]\n"
       "fmla z25.s, z16.s, z1.s[2]\n"
       "fmla z26.s, z16.s, z2.s[2]\n"
+      "addvl x12, x12, #1\n"
       "ble 37f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
       "fmla z25.s, z16.s, z1.s[3]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[3]\n"
       "37:"  // Height 3: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -435,12 +433,12 @@ void sve_hybrid_fp32_mla_8x1VL (
       "cmp x10, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
+      "add x26, x27, x20, LSL #2\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "fmin z24.s, p2/M, z24.s, z17.s\n"
       "fmin z25.s, p2/M, z25.s, z17.s\n"
@@ -451,37 +449,37 @@ void sve_hybrid_fp32_mla_8x1VL (
       "38:"  // Height 3: No activation
       "st1w { z24.s }, p1, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p1, [x28]\n"
-      "st1w { z26.s }, p1, [x27]\n"
+      "st1w { z25.s }, p1, [x27]\n"
+      "st1w { z26.s }, p1, [x26]\n"
       "39:"  // Height 3: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 28b\n"
       "b 106f\n"
       "40:"  // Height 4
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p1.s, x20, x13\n"
       "cbz x14, 42f\n"
       "ld1w { z24.s }, p2/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "b 44f\n"
       "42:"  // Height 4: no bias
       "tbz %x[flags], #0, 43f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x11, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z24.s }, p1/Z, [x11]\n"
-      "add x20, x11, x21, LSL #2\n"
-      "ld1w { z25.s }, p1/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
-      "ld1w { z26.s }, p1/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z25.s }, p1/Z, [x22]\n"
+      "ld1w { z26.s }, p1/Z, [x21]\n"
       "ld1w { z27.s }, p1/Z, [x20]\n"
       "b 44f\n"
       "43:"  // Height 4: no accumulate
@@ -493,8 +491,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "mov x10, #0x0\n"
       "45:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -519,29 +517,29 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 49f\n"
       "48:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z19.s }, p2/Z, [x12]\n"
-      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
-      "sub x9, x9, #0x4\n"
-      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
-      "cmp x9, #0x4\n"
-      "addvl x12, x12, #4\n"
       "ld1rqw { z3.s }, p0/Z, [x28]\n"
-      "add x28, x28, #0x10\n"
       "ld1rqw { z2.s }, p0/Z, [x27]\n"
-      "add x27, x27, #0x10\n"
+      "sub x9, x9, #0x4\n"
       "ld1rqw { z1.s }, p0/Z, [x26]\n"
       "ld1rqw { z0.s }, p0/Z, [x25]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
-      "fmla z24.s, z19.s, z3.s[0]\n"
-      "fmla z25.s, z19.s, z2.s[0]\n"
-      "fmla z26.s, z19.s, z1.s[0]\n"
-      "fmla z27.s, z19.s, z0.s[0]\n"
+      "cmp x9, #0x4\n"
+      "add x28, x28, #0x10\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
+      "fmla z24.s, z16.s, z3.s[0]\n"
+      "fmla z25.s, z16.s, z2.s[0]\n"
+      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
+      "fmla z26.s, z16.s, z1.s[0]\n"
+      "fmla z27.s, z16.s, z0.s[0]\n"
+      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
       "fmla z24.s, z18.s, z3.s[1]\n"
       "fmla z25.s, z18.s, z2.s[1]\n"
+      "add x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "fmla z26.s, z18.s, z1.s[1]\n"
       "fmla z27.s, z18.s, z0.s[1]\n"
+      "add x25, x25, #0x10\n"
+      "addvl x12, x12, #4\n"
       "fmla z24.s, z17.s, z3.s[2]\n"
       "fmla z25.s, z17.s, z2.s[2]\n"
       "fmla z26.s, z17.s, z1.s[2]\n"
@@ -553,38 +551,38 @@ void sve_hybrid_fp32_mla_8x1VL (
       "bgt 48b\n"
       "49:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z16.s }, p2/Z, [x12]\n"
-      "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "ld1rqw { z1.s }, p0/Z, [x27]\n"
+      "subs x9, x9, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
       "ld1rqw { z3.s }, p0/Z, [x25]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
       "fmla z24.s, z16.s, z0.s[0]\n"
       "fmla z25.s, z16.s, z1.s[0]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[0]\n"
       "fmla z27.s, z16.s, z3.s[0]\n"
       "ble 50f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[1]\n"
       "fmla z25.s, z16.s, z1.s[1]\n"
       "fmla z26.s, z16.s, z2.s[1]\n"
       "fmla z27.s, z16.s, z3.s[1]\n"
+      "addvl x12, x12, #1\n"
       "ble 50f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[2]\n"
       "fmla z25.s, z16.s, z1.s[2]\n"
       "fmla z26.s, z16.s, z2.s[2]\n"
       "fmla z27.s, z16.s, z3.s[2]\n"
+      "addvl x12, x12, #1\n"
       "ble 50f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
       "fmla z25.s, z16.s, z1.s[3]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[3]\n"
       "fmla z27.s, z16.s, z3.s[3]\n"
       "50:"  // Height 4: Multiply loop: multiply skip
@@ -593,13 +591,13 @@ void sve_hybrid_fp32_mla_8x1VL (
       "cmp x10, x20\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
+      "add x25, x26, x20, LSL #2\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "fmin z24.s, p2/M, z24.s, z17.s\n"
       "fmin z25.s, p2/M, z25.s, z17.s\n"
@@ -612,41 +610,41 @@ void sve_hybrid_fp32_mla_8x1VL (
       "51:"  // Height 4: No activation
       "st1w { z24.s }, p1, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p1, [x28]\n"
-      "st1w { z26.s }, p1, [x27]\n"
-      "st1w { z27.s }, p1, [x26]\n"
+      "st1w { z25.s }, p1, [x27]\n"
+      "st1w { z26.s }, p1, [x26]\n"
+      "st1w { z27.s }, p1, [x25]\n"
       "52:"  // Height 4: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 41b\n"
       "b 106f\n"
       "53:"  // Height 5
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "54:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p1.s, x20, x13\n"
       "cbz x14, 55f\n"
       "ld1w { z24.s }, p2/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "mov z28.d, z24.d\n"
       "b 57f\n"
       "55:"  // Height 5: no bias
       "tbz %x[flags], #0, 56f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x11, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z24.s }, p1/Z, [x11]\n"
-      "add x20, x11, x21, LSL #2\n"
-      "ld1w { z25.s }, p1/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
-      "ld1w { z26.s }, p1/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
-      "ld1w { z27.s }, p1/Z, [x20]\n"
-      "add x20, x20, x21, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z25.s }, p1/Z, [x23]\n"
+      "ld1w { z26.s }, p1/Z, [x22]\n"
+      "ld1w { z27.s }, p1/Z, [x21]\n"
       "ld1w { z28.s }, p1/Z, [x20]\n"
       "b 57f\n"
       "56:"  // Height 5: no accumulate
@@ -659,8 +657,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "mov x10, #0x0\n"
       "58:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 59f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -688,33 +686,33 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 62f\n"
       "61:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z19.s }, p2/Z, [x12]\n"
-      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
-      "sub x9, x9, #0x4\n"
-      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
-      "cmp x9, #0x4\n"
-      "addvl x12, x12, #4\n"
       "ld1rqw { z4.s }, p0/Z, [x28]\n"
-      "add x28, x28, #0x10\n"
       "ld1rqw { z3.s }, p0/Z, [x27]\n"
-      "add x27, x27, #0x10\n"
+      "sub x9, x9, #0x4\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
       "ld1rqw { z1.s }, p0/Z, [x25]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
+      "cmp x9, #0x4\n"
+      "add x28, x28, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "fmla z24.s, z19.s, z4.s[0]\n"
-      "fmla z25.s, z19.s, z3.s[0]\n"
-      "fmla z26.s, z19.s, z2.s[0]\n"
-      "fmla z27.s, z19.s, z1.s[0]\n"
-      "fmla z28.s, z19.s, z0.s[0]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
+      "fmla z24.s, z16.s, z4.s[0]\n"
+      "fmla z25.s, z16.s, z3.s[0]\n"
+      "fmla z26.s, z16.s, z2.s[0]\n"
+      "fmla z27.s, z16.s, z1.s[0]\n"
+      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
+      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
+      "fmla z28.s, z16.s, z0.s[0]\n"
       "fmla z24.s, z18.s, z4.s[1]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
+      "add x27, x27, #0x10\n"
       "fmla z25.s, z18.s, z3.s[1]\n"
       "fmla z26.s, z18.s, z2.s[1]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "fmla z27.s, z18.s, z1.s[1]\n"
       "fmla z28.s, z18.s, z0.s[1]\n"
+      "add x24, x24, #0x10\n"
+      "addvl x12, x12, #4\n"
       "fmla z24.s, z17.s, z4.s[2]\n"
       "fmla z25.s, z17.s, z3.s[2]\n"
       "fmla z26.s, z17.s, z2.s[2]\n"
@@ -728,42 +726,42 @@ void sve_hybrid_fp32_mla_8x1VL (
       "bgt 61b\n"
       "62:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z16.s }, p2/Z, [x12]\n"
-      "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "ld1rqw { z1.s }, p0/Z, [x27]\n"
+      "subs x9, x9, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
       "ld1rqw { z3.s }, p0/Z, [x25]\n"
       "ld1rqw { z4.s }, p0/Z, [x24]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
       "fmla z24.s, z16.s, z0.s[0]\n"
       "fmla z25.s, z16.s, z1.s[0]\n"
       "fmla z26.s, z16.s, z2.s[0]\n"
       "fmla z27.s, z16.s, z3.s[0]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[0]\n"
       "ble 63f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[1]\n"
       "fmla z25.s, z16.s, z1.s[1]\n"
       "fmla z26.s, z16.s, z2.s[1]\n"
       "fmla z27.s, z16.s, z3.s[1]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[1]\n"
       "ble 63f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[2]\n"
       "fmla z25.s, z16.s, z1.s[2]\n"
       "fmla z26.s, z16.s, z2.s[2]\n"
       "fmla z27.s, z16.s, z3.s[2]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[2]\n"
       "ble 63f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
       "fmla z25.s, z16.s, z1.s[3]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[3]\n"
       "fmla z27.s, z16.s, z3.s[3]\n"
       "fmla z28.s, z16.s, z4.s[3]\n"
@@ -773,14 +771,14 @@ void sve_hybrid_fp32_mla_8x1VL (
       "cmp x10, x20\n"
       "bne 58b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "fmin z24.s, p2/M, z24.s, z17.s\n"
       "fmin z25.s, p2/M, z25.s, z17.s\n"
@@ -795,44 +793,44 @@ void sve_hybrid_fp32_mla_8x1VL (
       "64:"  // Height 5: No activation
       "st1w { z24.s }, p1, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p1, [x28]\n"
-      "st1w { z26.s }, p1, [x27]\n"
-      "st1w { z27.s }, p1, [x26]\n"
-      "st1w { z28.s }, p1, [x25]\n"
+      "st1w { z25.s }, p1, [x27]\n"
+      "st1w { z26.s }, p1, [x26]\n"
+      "st1w { z27.s }, p1, [x25]\n"
+      "st1w { z28.s }, p1, [x24]\n"
       "65:"  // Height 5: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 54b\n"
       "b 106f\n"
       "66:"  // Height 6
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "67:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p1.s, x20, x13\n"
       "cbz x14, 68f\n"
       "ld1w { z24.s }, p2/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "mov z28.d, z24.d\n"
       "mov z29.d, z24.d\n"
       "b 70f\n"
       "68:"  // Height 6: no bias
       "tbz %x[flags], #0, 69f\n"
-      "ldr x22, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x24, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x11, x24, LSL #2\n"
+      "add x20, x23, x24, LSL #2\n"
       "ld1w { z24.s }, p1/Z, [x11]\n"
-      "add x20, x11, x22, LSL #2\n"
-      "ld1w { z25.s }, p1/Z, [x20]\n"
-      "add x20, x20, x22, LSL #2\n"
+      "add x22, x20, x24, LSL #2\n"
+      "add x21, x22, x24, LSL #2\n"
+      "ld1w { z25.s }, p1/Z, [x23]\n"
       "ld1w { z26.s }, p1/Z, [x20]\n"
-      "add x20, x20, x22, LSL #2\n"
-      "ld1w { z27.s }, p1/Z, [x20]\n"
-      "add x21, x20, x22, LSL #2\n"
-      "add x20, x21, x22, LSL #2\n"
+      "add x20, x21, x24, LSL #2\n"
+      "ld1w { z27.s }, p1/Z, [x22]\n"
       "ld1w { z28.s }, p1/Z, [x21]\n"
       "ld1w { z29.s }, p1/Z, [x20]\n"
       "b 70f\n"
@@ -847,8 +845,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "mov x10, #0x0\n"
       "71:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 72f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -879,33 +877,33 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 75f\n"
       "74:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z19.s }, p2/Z, [x12]\n"
-      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
-      "sub x9, x9, #0x4\n"
-      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
-      "cmp x9, #0x4\n"
-      "addvl x12, x12, #4\n"
       "ld1rqw { z5.s }, p0/Z, [x28]\n"
-      "add x28, x28, #0x10\n"
       "ld1rqw { z4.s }, p0/Z, [x27]\n"
-      "add x27, x27, #0x10\n"
+      "sub x9, x9, #0x4\n"
       "ld1rqw { z3.s }, p0/Z, [x26]\n"
       "ld1rqw { z2.s }, p0/Z, [x25]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
+      "cmp x9, #0x4\n"
+      "add x28, x28, #0x10\n"
       "ld1rqw { z1.s }, p0/Z, [x24]\n"
       "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      "add x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "ld1w { z19.s }, p2/Z, [x12]\n"
       "fmla z24.s, z19.s, z5.s[0]\n"
       "fmla z25.s, z19.s, z4.s[0]\n"
+      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
       "fmla z26.s, z19.s, z3.s[0]\n"
       "fmla z27.s, z19.s, z2.s[0]\n"
+      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
       "fmla z28.s, z19.s, z1.s[0]\n"
       "fmla z29.s, z19.s, z0.s[0]\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
       "fmla z24.s, z18.s, z5.s[1]\n"
       "fmla z25.s, z18.s, z4.s[1]\n"
+      "add x23, x23, #0x10\n"
+      "addvl x12, x12, #4\n"
       "fmla z26.s, z18.s, z3.s[1]\n"
       "fmla z27.s, z18.s, z2.s[1]\n"
       "fmla z28.s, z18.s, z1.s[1]\n"
@@ -925,17 +923,17 @@ void sve_hybrid_fp32_mla_8x1VL (
       "bgt 74b\n"
       "75:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z16.s }, p2/Z, [x12]\n"
-      "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "ld1rqw { z1.s }, p0/Z, [x27]\n"
+      "subs x9, x9, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
       "ld1rqw { z3.s }, p0/Z, [x25]\n"
       "ld1rqw { z4.s }, p0/Z, [x24]\n"
       "ld1rqw { z5.s }, p0/Z, [x23]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
       "fmla z24.s, z16.s, z0.s[0]\n"
       "fmla z25.s, z16.s, z1.s[0]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[0]\n"
       "fmla z27.s, z16.s, z3.s[0]\n"
       "fmla z28.s, z16.s, z4.s[0]\n"
@@ -943,28 +941,28 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 76f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[1]\n"
       "fmla z25.s, z16.s, z1.s[1]\n"
       "fmla z26.s, z16.s, z2.s[1]\n"
       "fmla z27.s, z16.s, z3.s[1]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[1]\n"
       "fmla z29.s, z16.s, z5.s[1]\n"
       "ble 76f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[2]\n"
       "fmla z25.s, z16.s, z1.s[2]\n"
       "fmla z26.s, z16.s, z2.s[2]\n"
       "fmla z27.s, z16.s, z3.s[2]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[2]\n"
       "fmla z29.s, z16.s, z5.s[2]\n"
       "ble 76f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
       "fmla z25.s, z16.s, z1.s[3]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[3]\n"
       "fmla z27.s, z16.s, z3.s[3]\n"
       "fmla z28.s, z16.s, z4.s[3]\n"
@@ -975,15 +973,15 @@ void sve_hybrid_fp32_mla_8x1VL (
       "cmp x10, x20\n"
       "bne 71b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "tbz %x[flags], #1, 77f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "fmin z24.s, p2/M, z24.s, z17.s\n"
       "fmin z25.s, p2/M, z25.s, z17.s\n"
@@ -1000,29 +998,29 @@ void sve_hybrid_fp32_mla_8x1VL (
       "77:"  // Height 6: No activation
       "st1w { z24.s }, p1, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p1, [x28]\n"
-      "st1w { z26.s }, p1, [x27]\n"
-      "st1w { z27.s }, p1, [x26]\n"
-      "st1w { z28.s }, p1, [x25]\n"
-      "st1w { z29.s }, p1, [x24]\n"
+      "st1w { z25.s }, p1, [x27]\n"
+      "st1w { z26.s }, p1, [x26]\n"
+      "st1w { z27.s }, p1, [x25]\n"
+      "st1w { z28.s }, p1, [x24]\n"
+      "st1w { z29.s }, p1, [x23]\n"
       "78:"  // Height 6: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
       "bgt 67b\n"
       "b 106f\n"
       "79:"  // Height 7
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
       "80:"  // Height 7: Column loop
       "mov x20, #0x0\n"
       "whilelt p1.s, x20, x13\n"
       "cbz x14, 81f\n"
       "ld1w { z24.s }, p2/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "mov z28.d, z24.d\n"
       "mov z29.d, z24.d\n"
@@ -1030,17 +1028,17 @@ void sve_hybrid_fp32_mla_8x1VL (
       "b 83f\n"
       "81:"  // Height 7: no bias
       "tbz %x[flags], #0, 82f\n"
-      "ldr x23, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x24, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x11, x24, LSL #2\n"
+      "add x20, x21, x24, LSL #2\n"
       "ld1w { z24.s }, p1/Z, [x11]\n"
-      "add x20, x11, x23, LSL #2\n"
-      "ld1w { z25.s }, p1/Z, [x20]\n"
-      "add x20, x20, x23, LSL #2\n"
+      "add x23, x20, x24, LSL #2\n"
+      "add x22, x23, x24, LSL #2\n"
+      "ld1w { z25.s }, p1/Z, [x21]\n"
       "ld1w { z26.s }, p1/Z, [x20]\n"
-      "add x20, x20, x23, LSL #2\n"
-      "ld1w { z27.s }, p1/Z, [x20]\n"
-      "add x22, x20, x23, LSL #2\n"
-      "add x21, x22, x23, LSL #2\n"
-      "add x20, x21, x23, LSL #2\n"
+      "add x21, x22, x24, LSL #2\n"
+      "add x20, x21, x24, LSL #2\n"
+      "ld1w { z27.s }, p1/Z, [x23]\n"
       "ld1w { z28.s }, p1/Z, [x22]\n"
       "ld1w { z29.s }, p1/Z, [x21]\n"
       "ld1w { z30.s }, p1/Z, [x20]\n"
@@ -1057,8 +1055,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "mov x10, #0x0\n"
       "84:"  // Height 7: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 85f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1092,37 +1090,37 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 88f\n"
       "87:"  // Height 7: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z19.s }, p2/Z, [x12]\n"
-      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
-      "sub x9, x9, #0x4\n"
-      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
-      "cmp x9, #0x4\n"
-      "addvl x12, x12, #4\n"
       "ld1rqw { z6.s }, p0/Z, [x28]\n"
-      "add x28, x28, #0x10\n"
       "ld1rqw { z5.s }, p0/Z, [x27]\n"
-      "add x27, x27, #0x10\n"
+      "sub x9, x9, #0x4\n"
       "ld1rqw { z4.s }, p0/Z, [x26]\n"
       "ld1rqw { z3.s }, p0/Z, [x25]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
+      "cmp x9, #0x4\n"
+      "add x28, x28, #0x10\n"
       "ld1rqw { z2.s }, p0/Z, [x24]\n"
       "ld1rqw { z1.s }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      "add x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x22]\n"
+      "ld1w { z19.s }, p2/Z, [x12]\n"
       "fmla z24.s, z19.s, z6.s[0]\n"
       "fmla z25.s, z19.s, z5.s[0]\n"
-      "add x22, x22, #0x10\n"
       "fmla z26.s, z19.s, z4.s[0]\n"
       "fmla z27.s, z19.s, z3.s[0]\n"
+      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
+      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
       "fmla z28.s, z19.s, z2.s[0]\n"
       "fmla z29.s, z19.s, z1.s[0]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       "fmla z30.s, z19.s, z0.s[0]\n"
       "fmla z24.s, z18.s, z6.s[1]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "fmla z25.s, z18.s, z5.s[1]\n"
       "fmla z26.s, z18.s, z4.s[1]\n"
+      "add x22, x22, #0x10\n"
+      "addvl x12, x12, #4\n"
       "fmla z27.s, z18.s, z3.s[1]\n"
       "fmla z28.s, z18.s, z2.s[1]\n"
       "fmla z29.s, z18.s, z1.s[1]\n"
@@ -1144,50 +1142,50 @@ void sve_hybrid_fp32_mla_8x1VL (
       "bgt 87b\n"
       "88:"  // Height 7: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z16.s }, p2/Z, [x12]\n"
-      "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "ld1rqw { z1.s }, p0/Z, [x27]\n"
+      "subs x9, x9, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
       "ld1rqw { z3.s }, p0/Z, [x25]\n"
       "ld1rqw { z4.s }, p0/Z, [x24]\n"
       "ld1rqw { z5.s }, p0/Z, [x23]\n"
       "ld1rqw { z6.s }, p0/Z, [x22]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
       "fmla z24.s, z16.s, z0.s[0]\n"
       "fmla z25.s, z16.s, z1.s[0]\n"
       "fmla z26.s, z16.s, z2.s[0]\n"
       "fmla z27.s, z16.s, z3.s[0]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[0]\n"
       "fmla z29.s, z16.s, z5.s[0]\n"
       "fmla z30.s, z16.s, z6.s[0]\n"
       "ble 89f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[1]\n"
       "fmla z25.s, z16.s, z1.s[1]\n"
       "fmla z26.s, z16.s, z2.s[1]\n"
       "fmla z27.s, z16.s, z3.s[1]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[1]\n"
       "fmla z29.s, z16.s, z5.s[1]\n"
       "fmla z30.s, z16.s, z6.s[1]\n"
       "ble 89f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[2]\n"
       "fmla z25.s, z16.s, z1.s[2]\n"
       "fmla z26.s, z16.s, z2.s[2]\n"
       "fmla z27.s, z16.s, z3.s[2]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[2]\n"
       "fmla z29.s, z16.s, z5.s[2]\n"
       "fmla z30.s, z16.s, z6.s[2]\n"
       "ble 89f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
       "fmla z25.s, z16.s, z1.s[3]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[3]\n"
       "fmla z27.s, z16.s, z3.s[3]\n"
       "fmla z28.s, z16.s, z4.s[3]\n"
@@ -1199,16 +1197,16 @@ void sve_hybrid_fp32_mla_8x1VL (
       "cmp x10, x20\n"
       "bne 84b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "tbz %x[flags], #1, 90f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "fmin z24.s, p2/M, z24.s, z17.s\n"
       "fmin z25.s, p2/M, z25.s, z17.s\n"
@@ -1227,12 +1225,12 @@ void sve_hybrid_fp32_mla_8x1VL (
       "90:"  // Height 7: No activation
       "st1w { z24.s }, p1, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p1, [x28]\n"
-      "st1w { z26.s }, p1, [x27]\n"
-      "st1w { z27.s }, p1, [x26]\n"
-      "st1w { z28.s }, p1, [x25]\n"
-      "st1w { z29.s }, p1, [x24]\n"
-      "st1w { z30.s }, p1, [x23]\n"
+      "st1w { z25.s }, p1, [x27]\n"
+      "st1w { z26.s }, p1, [x26]\n"
+      "st1w { z27.s }, p1, [x25]\n"
+      "st1w { z28.s }, p1, [x24]\n"
+      "st1w { z29.s }, p1, [x23]\n"
+      "st1w { z30.s }, p1, [x22]\n"
       "91:"  // Height 7: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
@@ -1240,21 +1238,20 @@ void sve_hybrid_fp32_mla_8x1VL (
       "b 106f\n"
       "92:"  // Height 8
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x20\n"
-      "ldr x14, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x14, %x[bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x11\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "93:"  // Height 8: Column loop
       "mov x20, #0x0\n"
       "whilelt p1.s, x20, x13\n"
       "cbz x14, 94f\n"
       "ld1w { z24.s }, p2/Z, [x14]\n"
-      "addvl x14, x14, #1\n"
       "mov z25.d, z24.d\n"
       "mov z26.d, z24.d\n"
+      "addvl x14, x14, #1\n"
       "mov z27.d, z24.d\n"
       "mov z28.d, z24.d\n"
       "mov z29.d, z24.d\n"
@@ -1263,20 +1260,20 @@ void sve_hybrid_fp32_mla_8x1VL (
       "b 96f\n"
       "94:"  // Height 8: no bias
       "tbz %x[flags], #0, 95f\n"
-      "ldr x23, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "ldr x24, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x11, x24, LSL #2\n"
+      "add x21, x22, x24, LSL #2\n"
       "ld1w { z24.s }, p1/Z, [x11]\n"
-      "add x20, x11, x23, LSL #2\n"
-      "ld1w { z25.s }, p1/Z, [x20]\n"
-      "add x20, x20, x23, LSL #2\n"
-      "ld1w { z26.s }, p1/Z, [x20]\n"
-      "add x20, x20, x23, LSL #2\n"
-      "ld1w { z27.s }, p1/Z, [x20]\n"
-      "add x22, x20, x23, LSL #2\n"
-      "add x20, x22, x23, LSL #2\n"
-      "add x21, x20, x23, LSL #2\n"
-      "ld1w { z28.s }, p1/Z, [x22]\n"
-      "ld1w { z29.s }, p1/Z, [x20]\n"
-      "add x20, x21, x23, LSL #2\n"
+      "add x23, x21, x24, LSL #2\n"
+      "add x20, x23, x24, LSL #2\n"
+      "ld1w { z25.s }, p1/Z, [x22]\n"
+      "ld1w { z26.s }, p1/Z, [x21]\n"
+      "add x22, x20, x24, LSL #2\n"
+      "add x21, x22, x24, LSL #2\n"
+      "ld1w { z27.s }, p1/Z, [x23]\n"
+      "ld1w { z28.s }, p1/Z, [x20]\n"
+      "add x20, x21, x24, LSL #2\n"
+      "ld1w { z29.s }, p1/Z, [x22]\n"
       "ld1w { z30.s }, p1/Z, [x21]\n"
       "ld1w { z31.s }, p1/Z, [x20]\n"
       "b 96f\n"
@@ -1293,8 +1290,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "mov x10, #0x0\n"
       "97:"  // Height 8: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w9, [x20, x10, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 98f\n"
       "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1331,37 +1328,37 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 101f\n"
       "100:"  // Height 8: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z19.s }, p2/Z, [x12]\n"
-      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
-      "sub x9, x9, #0x4\n"
-      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
-      "cmp x9, #0x4\n"
-      "addvl x12, x12, #4\n"
       "ld1rqw { z7.s }, p0/Z, [x28]\n"
-      "add x28, x28, #0x10\n"
       "ld1rqw { z6.s }, p0/Z, [x27]\n"
-      "add x27, x27, #0x10\n"
+      "sub x9, x9, #0x4\n"
       "ld1rqw { z5.s }, p0/Z, [x26]\n"
       "ld1rqw { z4.s }, p0/Z, [x25]\n"
-      "add x26, x26, #0x10\n"
-      "add x25, x25, #0x10\n"
+      "cmp x9, #0x4\n"
+      "add x28, x28, #0x10\n"
       "ld1rqw { z3.s }, p0/Z, [x24]\n"
       "ld1rqw { z2.s }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      "add x27, x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "ld1rqw { z1.s }, p0/Z, [x22]\n"
       "ld1rqw { z0.s }, p0/Z, [x21]\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1w { z19.s }, p2/Z, [x12]\n"
       "fmla z24.s, z19.s, z7.s[0]\n"
       "fmla z25.s, z19.s, z6.s[0]\n"
+      "ld1w { z18.s }, p2/Z, [x12, #1, MUL VL]\n"
       "fmla z26.s, z19.s, z5.s[0]\n"
       "fmla z27.s, z19.s, z4.s[0]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
+      "ld1w { z17.s }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x12, #3, MUL VL]\n"
       "fmla z28.s, z19.s, z3.s[0]\n"
       "fmla z29.s, z19.s, z2.s[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "fmla z30.s, z19.s, z1.s[0]\n"
       "fmla z31.s, z19.s, z0.s[0]\n"
+      "add x21, x21, #0x10\n"
+      "addvl x12, x12, #4\n"
       "fmla z24.s, z18.s, z7.s[1]\n"
       "fmla z25.s, z18.s, z6.s[1]\n"
       "fmla z26.s, z18.s, z5.s[1]\n"
@@ -1389,19 +1386,19 @@ void sve_hybrid_fp32_mla_8x1VL (
       "bgt 100b\n"
       "101:"  // Height 8: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x9\n"
-      "ld1w { z16.s }, p2/Z, [x12]\n"
-      "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "ld1rqw { z0.s }, p0/Z, [x28]\n"
       "ld1rqw { z1.s }, p0/Z, [x27]\n"
+      "subs x9, x9, #0x1\n"
       "ld1rqw { z2.s }, p0/Z, [x26]\n"
       "ld1rqw { z3.s }, p0/Z, [x25]\n"
       "ld1rqw { z4.s }, p0/Z, [x24]\n"
       "ld1rqw { z5.s }, p0/Z, [x23]\n"
       "ld1rqw { z6.s }, p0/Z, [x22]\n"
       "ld1rqw { z7.s }, p0/Z, [x21]\n"
+      "ld1w { z16.s }, p2/Z, [x12]\n"
       "fmla z24.s, z16.s, z0.s[0]\n"
       "fmla z25.s, z16.s, z1.s[0]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[0]\n"
       "fmla z27.s, z16.s, z3.s[0]\n"
       "fmla z28.s, z16.s, z4.s[0]\n"
@@ -1411,11 +1408,11 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 102f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[1]\n"
       "fmla z25.s, z16.s, z1.s[1]\n"
       "fmla z26.s, z16.s, z2.s[1]\n"
       "fmla z27.s, z16.s, z3.s[1]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[1]\n"
       "fmla z29.s, z16.s, z5.s[1]\n"
       "fmla z30.s, z16.s, z6.s[1]\n"
@@ -1423,20 +1420,20 @@ void sve_hybrid_fp32_mla_8x1VL (
       "ble 102f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
       "subs x9, x9, #0x1\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[2]\n"
       "fmla z25.s, z16.s, z1.s[2]\n"
       "fmla z26.s, z16.s, z2.s[2]\n"
       "fmla z27.s, z16.s, z3.s[2]\n"
+      "addvl x12, x12, #1\n"
       "fmla z28.s, z16.s, z4.s[2]\n"
       "fmla z29.s, z16.s, z5.s[2]\n"
       "fmla z30.s, z16.s, z6.s[2]\n"
       "fmla z31.s, z16.s, z7.s[2]\n"
       "ble 102f\n"
       "ld1w { z16.s }, p2/Z, [x12]\n"
-      "addvl x12, x12, #1\n"
       "fmla z24.s, z16.s, z0.s[3]\n"
       "fmla z25.s, z16.s, z1.s[3]\n"
+      "addvl x12, x12, #1\n"
       "fmla z26.s, z16.s, z2.s[3]\n"
       "fmla z27.s, z16.s, z3.s[3]\n"
       "fmla z28.s, z16.s, z4.s[3]\n"
@@ -1449,17 +1446,17 @@ void sve_hybrid_fp32_mla_8x1VL (
       "cmp x10, x20\n"
       "bne 97b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x28, x11, x20, LSL #2\n"
-      "add x27, x28, x20, LSL #2\n"
+      "add x27, x11, x20, LSL #2\n"
       "add x26, x27, x20, LSL #2\n"
       "add x25, x26, x20, LSL #2\n"
       "add x24, x25, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "tbz %x[flags], #1, 103f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p2/Z, [x21]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "fmin z24.s, p2/M, z24.s, z17.s\n"
       "fmin z25.s, p2/M, z25.s, z17.s\n"
@@ -1480,13 +1477,13 @@ void sve_hybrid_fp32_mla_8x1VL (
       "103:"  // Height 8: No activation
       "st1w { z24.s }, p1, [x11]\n"
       "addvl x11, x11, #1\n"
-      "st1w { z25.s }, p1, [x28]\n"
-      "st1w { z26.s }, p1, [x27]\n"
-      "st1w { z27.s }, p1, [x26]\n"
-      "st1w { z28.s }, p1, [x25]\n"
-      "st1w { z29.s }, p1, [x24]\n"
-      "st1w { z30.s }, p1, [x23]\n"
-      "st1w { z31.s }, p1, [x22]\n"
+      "st1w { z25.s }, p1, [x27]\n"
+      "st1w { z26.s }, p1, [x26]\n"
+      "st1w { z27.s }, p1, [x25]\n"
+      "st1w { z28.s }, p1, [x24]\n"
+      "st1w { z29.s }, p1, [x23]\n"
+      "st1w { z30.s }, p1, [x22]\n"
+      "st1w { z31.s }, p1, [x21]\n"
       "104:"  // Height 8: Writeback done
       "decw x13\n"
       "cmp x13, XZR\n"
@@ -1503,8 +1500,8 @@ void sve_hybrid_fp32_mla_8x1VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "106:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z16", "z17", "z18", "z19", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp
index d7436b15f4..7dc786fd66 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 4, 12, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 4, 12, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp
index be245f9ecc..b0d6f756d7 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL/generic.cpp
@@ -48,19 +48,18 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -81,7 +80,6 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -102,10 +100,10 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "cmp %x[M], #0x2\n"
       "bgt 27f\n"
       "beq 14f\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p6.s, x20, x9\n"
@@ -122,19 +120,19 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "cbz x10, 3f\n"
       "ld1w { z8.s }, p7/Z, [x10]\n"
       "ld1w { z9.s }, p7/Z, [x10, #1, MUL VL]\n"
-      "ld1w { z10.s }, p7/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z11.s }, p7/Z, [x10, #3, MUL VL]\n"
-      "ld1w { z12.s }, p7/Z, [x10, #4, MUL VL]\n"
-      "ld1w { z13.s }, p7/Z, [x10, #5, MUL VL]\n"
-      "addvl x10, x10, #6\n"
       "zip2 z14.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p7/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z11.s }, p7/Z, [x10, #3, MUL VL]\n"
       "zip2 z15.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
+      "ld1w { z12.s }, p7/Z, [x10, #4, MUL VL]\n"
+      "ld1w { z13.s }, p7/Z, [x10, #5, MUL VL]\n"
       "zip2 z16.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
       "zip2 z17.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
+      "addvl x10, x10, #6\n"
       "zip2 z18.d, z12.d, z12.d\n"
       "zip1 z12.d, z12.d, z12.d\n"
       "zip2 z19.d, z13.d, z13.d\n"
@@ -142,16 +140,16 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "b 5f\n"
       "3:"  // Height 1: no bias
       "tbz %x[flags], #0, 4f\n"
-      "ld1w { z25.s }, p6/Z, [x27]\n"
-      "ld1w { z24.s }, p5/Z, [x27, #1, MUL VL]\n"
+      "ld1w { z21.s }, p6/Z, [x27]\n"
+      "ld1w { z20.s }, p5/Z, [x27, #1, MUL VL]\n"
+      "zip1 z8.d, z21.d, z14.d\n"
+      "zip2 z14.d, z21.d, z14.d\n"
       "ld1w { z23.s }, p4/Z, [x27, #2, MUL VL]\n"
       "ld1w { z22.s }, p3/Z, [x27, #3, MUL VL]\n"
+      "zip1 z9.d, z20.d, z15.d\n"
+      "zip2 z15.d, z20.d, z15.d\n"
       "ld1w { z21.s }, p2/Z, [x27, #4, MUL VL]\n"
       "ld1w { z20.s }, p1/Z, [x27, #5, MUL VL]\n"
-      "zip1 z8.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "zip1 z9.d, z24.d, z15.d\n"
-      "zip2 z15.d, z24.d, z15.d\n"
       "zip1 z10.d, z23.d, z16.d\n"
       "zip2 z16.d, z23.d, z16.d\n"
       "zip1 z11.d, z22.d, z17.d\n"
@@ -178,8 +176,8 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "mov x26, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -195,70 +193,70 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z23.h }, p7/Z, [x28]\n"
-      "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "sub x25, x25, #0x4\n"
-      "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n"
-      "cmp x25, #0x4\n"
       "ld1rqw { z24.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       ".inst 0x658abf18  // bfcvt z24.h, p7/M, z24.s\n"
       "uzp1 z24.h, z24.h, z24.h\n"
-      ".inst 0x6477e708  // bfmmla z8.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x28, #4, MUL VL]\n"
-      ".inst 0x6476e70e  // bfmmla z14.s, z24.h, z22.h\n"
-      "ld1h { z22.h }, p7/Z, [x28, #5, MUL VL]\n"
+      "ld1h { z21.h }, p7/Z, [x28]\n"
+      "ld1h { z20.h }, p7/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x6475e708  // bfmmla z8.s, z24.h, z21.h\n"
+      ".inst 0x6474e70e  // bfmmla z14.s, z24.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n"
       ".inst 0x6475e709  // bfmmla z9.s, z24.h, z21.h\n"
-      "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n"
       ".inst 0x6474e70f  // bfmmla z15.s, z24.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x28, #4, MUL VL]\n"
+      "ld1h { z20.h }, p7/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x6475e70a  // bfmmla z10.s, z24.h, z21.h\n"
+      ".inst 0x6474e710  // bfmmla z16.s, z24.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n"
       "ld1h { z20.h }, p7/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x6477e70a  // bfmmla z10.s, z24.h, z23.h\n"
-      ".inst 0x6476e710  // bfmmla z16.s, z24.h, z22.h\n"
       ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
       ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
       "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n"
       "ld1h { z22.h }, p7/Z, [x28, #-7, MUL VL]\n"
+      "sub x25, x25, #0x4\n"
       "ld1h { z21.h }, p7/Z, [x28, #-6, MUL VL]\n"
       "ld1h { z20.h }, p7/Z, [x28, #-5, MUL VL]\n"
-      "addvl x28, x28, #-4\n"
+      "cmp x25, #0x4\n"
       ".inst 0x6477e70c  // bfmmla z12.s, z24.h, z23.h\n"
       ".inst 0x6476e712  // bfmmla z18.s, z24.h, z22.h\n"
       ".inst 0x6475e70d  // bfmmla z13.s, z24.h, z21.h\n"
+      "add x24, x24, #0x10\n"
+      "addvl x28, x28, #-4\n"
       ".inst 0x6474e713  // bfmmla z19.s, z24.h, z20.h\n"
       "bgt 9b\n"
       "10:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z23.h }, p7/Z, [x28]\n"
-      "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
+      "ld1rqw { z23.s }, p0/Z, [x24]\n"
+      ".inst 0x658abef7  // bfcvt z23.h, p7/M, z23.s\n"
+      "uzp1 z23.h, z23.h, z23.h\n"
+      "ld1h { z21.h }, p7/Z, [x28]\n"
+      "ld1h { z20.h }, p7/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x6475e6e8  // bfmmla z8.s, z23.h, z21.h\n"
+      ".inst 0x6474e6ee  // bfmmla z14.s, z23.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n"
       "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n"
-      "ld1rqw { z24.s }, p0/Z, [x24]\n"
-      ".inst 0x658abf18  // bfcvt z24.h, p7/M, z24.s\n"
-      "uzp1 z24.h, z24.h, z24.h\n"
-      ".inst 0x6477e708  // bfmmla z8.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x28, #4, MUL VL]\n"
-      ".inst 0x6476e70e  // bfmmla z14.s, z24.h, z22.h\n"
-      "ld1h { z22.h }, p7/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x6475e709  // bfmmla z9.s, z24.h, z21.h\n"
+      ".inst 0x6475e6e9  // bfmmla z9.s, z23.h, z21.h\n"
+      ".inst 0x6474e6ef  // bfmmla z15.s, z23.h, z20.h\n"
+      "ld1h { z21.h }, p7/Z, [x28, #4, MUL VL]\n"
+      "ld1h { z20.h }, p7/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x6475e6ea  // bfmmla z10.s, z23.h, z21.h\n"
+      ".inst 0x6474e6f0  // bfmmla z16.s, z23.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x6474e70f  // bfmmla z15.s, z24.h, z20.h\n"
       "ld1h { z20.h }, p7/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x6477e70a  // bfmmla z10.s, z24.h, z23.h\n"
-      ".inst 0x6476e710  // bfmmla z16.s, z24.h, z22.h\n"
-      ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
-      ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
-      "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n"
+      ".inst 0x6475e6eb  // bfmmla z11.s, z23.h, z21.h\n"
+      ".inst 0x6474e6f1  // bfmmla z17.s, z23.h, z20.h\n"
+      "ld1h { z20.h }, p7/Z, [x28, #-8, MUL VL]\n"
       "ld1h { z22.h }, p7/Z, [x28, #-7, MUL VL]\n"
+      ".inst 0x6474e6ec  // bfmmla z12.s, z23.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x28, #-6, MUL VL]\n"
       "ld1h { z20.h }, p7/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x6476e6f2  // bfmmla z18.s, z23.h, z22.h\n"
+      ".inst 0x6475e6ed  // bfmmla z13.s, z23.h, z21.h\n"
+      ".inst 0x6474e6f3  // bfmmla z19.s, z23.h, z20.h\n"
       "addvl x28, x28, #-4\n"
-      ".inst 0x6477e70c  // bfmmla z12.s, z24.h, z23.h\n"
-      ".inst 0x6476e712  // bfmmla z18.s, z24.h, z22.h\n"
-      ".inst 0x6475e70d  // bfmmla z13.s, z24.h, z21.h\n"
-      ".inst 0x6474e713  // bfmmla z19.s, z24.h, z20.h\n"
       "11:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
@@ -271,9 +269,9 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "uzp1 z12.d, z12.d, z18.d\n"
       "uzp1 z13.d, z13.d, z19.d\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z21.s }, p7/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z21.s }, p7/Z, [x21]\n"
       "ld1rw { z20.s }, p7/Z, [x20]\n"
       "fmin z8.s, p7/M, z8.s, z21.s\n"
       "fmin z9.s, p7/M, z9.s, z21.s\n"
@@ -301,10 +299,10 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "bgt 2b\n"
       "b 54f\n"
       "14:"  // Height 2
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p6.s, x20, x9\n"
@@ -321,19 +319,19 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "cbz x10, 16f\n"
       "ld1w { z8.s }, p7/Z, [x10]\n"
       "ld1w { z9.s }, p7/Z, [x10, #1, MUL VL]\n"
-      "ld1w { z10.s }, p7/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z11.s }, p7/Z, [x10, #3, MUL VL]\n"
-      "ld1w { z12.s }, p7/Z, [x10, #4, MUL VL]\n"
-      "ld1w { z13.s }, p7/Z, [x10, #5, MUL VL]\n"
-      "addvl x10, x10, #6\n"
       "zip2 z14.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p7/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z11.s }, p7/Z, [x10, #3, MUL VL]\n"
       "zip2 z15.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
+      "ld1w { z12.s }, p7/Z, [x10, #4, MUL VL]\n"
+      "ld1w { z13.s }, p7/Z, [x10, #5, MUL VL]\n"
       "zip2 z16.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
       "zip2 z17.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
+      "addvl x10, x10, #6\n"
       "zip2 z18.d, z12.d, z12.d\n"
       "zip1 z12.d, z12.d, z12.d\n"
       "zip2 z19.d, z13.d, z13.d\n"
@@ -342,25 +340,25 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "16:"  // Height 2: no bias
       "tbz %x[flags], #0, 17f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z25.s }, p6/Z, [x27]\n"
-      "ld1w { z24.s }, p5/Z, [x27, #1, MUL VL]\n"
-      "ld1w { z23.s }, p4/Z, [x27, #2, MUL VL]\n"
+      "add x20, x27, x20, LSL #2\n"
+      "ld1w { z16.s }, p6/Z, [x27]\n"
+      "ld1w { z17.s }, p5/Z, [x27, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x27, #2, MUL VL]\n"
       "ld1w { z22.s }, p3/Z, [x27, #3, MUL VL]\n"
       "ld1w { z21.s }, p2/Z, [x27, #4, MUL VL]\n"
-      "add x20, x27, x20, LSL #2\n"
       "ld1w { z20.s }, p1/Z, [x27, #5, MUL VL]\n"
       "ld1w { z14.s }, p6/Z, [x20]\n"
+      "zip1 z8.d, z16.d, z14.d\n"
+      "zip2 z14.d, z16.d, z14.d\n"
       "ld1w { z15.s }, p5/Z, [x20, #1, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x20, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z15.d\n"
+      "zip2 z15.d, z17.d, z15.d\n"
       "ld1w { z17.s }, p3/Z, [x20, #3, MUL VL]\n"
       "ld1w { z18.s }, p2/Z, [x20, #4, MUL VL]\n"
+      "zip1 z10.d, z19.d, z16.d\n"
+      "zip2 z16.d, z19.d, z16.d\n"
       "ld1w { z19.s }, p1/Z, [x20, #5, MUL VL]\n"
-      "zip1 z8.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "zip1 z9.d, z24.d, z15.d\n"
-      "zip2 z15.d, z24.d, z15.d\n"
-      "zip1 z10.d, z23.d, z16.d\n"
-      "zip2 z16.d, z23.d, z16.d\n"
       "zip1 z11.d, z22.d, z17.d\n"
       "zip2 z17.d, z22.d, z17.d\n"
       "zip1 z12.d, z21.d, z18.d\n"
@@ -385,8 +383,8 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "mov x26, #0x0\n"
       "19:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -405,77 +403,77 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "ble 23f\n"
       "22:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z23.h }, p7/Z, [x28]\n"
-      "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "sub x25, x25, #0x4\n"
-      "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z25.h }, p7/Z, [x28, #3, MUL VL]\n"
-      "cmp x25, #0x4\n"
       "ld1rqw { z24.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z20.s }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
       ".inst 0x658abf18  // bfcvt z24.h, p7/M, z24.s\n"
       ".inst 0x658abe94  // bfcvt z20.h, p7/M, z20.s\n"
       "uzp1 z24.h, z24.h, z24.h\n"
+      "ld1h { z23.h }, p7/Z, [x28]\n"
+      "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
       "uzp1 z20.h, z20.h, z20.h\n"
       "trn1 z24.d, z24.d, z20.d\n"
+      "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n"
       ".inst 0x6477e708  // bfmmla z8.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x28, #4, MUL VL]\n"
       ".inst 0x6476e70e  // bfmmla z14.s, z24.h, z22.h\n"
+      "ld1h { z23.h }, p7/Z, [x28, #4, MUL VL]\n"
       "ld1h { z22.h }, p7/Z, [x28, #5, MUL VL]\n"
       ".inst 0x6475e709  // bfmmla z9.s, z24.h, z21.h\n"
+      ".inst 0x6474e70f  // bfmmla z15.s, z24.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x6479e70f  // bfmmla z15.s, z24.h, z25.h\n"
       "ld1h { z20.h }, p7/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
       ".inst 0x6477e70a  // bfmmla z10.s, z24.h, z23.h\n"
       ".inst 0x6476e710  // bfmmla z16.s, z24.h, z22.h\n"
+      "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n"
       ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
       ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
-      "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n"
       "ld1h { z22.h }, p7/Z, [x28, #-7, MUL VL]\n"
       "ld1h { z21.h }, p7/Z, [x28, #-6, MUL VL]\n"
       "ld1h { z20.h }, p7/Z, [x28, #-5, MUL VL]\n"
-      "addvl x28, x28, #-4\n"
+      "sub x25, x25, #0x4\n"
+      "cmp x25, #0x4\n"
       ".inst 0x6477e70c  // bfmmla z12.s, z24.h, z23.h\n"
       ".inst 0x6476e712  // bfmmla z18.s, z24.h, z22.h\n"
       ".inst 0x6475e70d  // bfmmla z13.s, z24.h, z21.h\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6474e713  // bfmmla z19.s, z24.h, z20.h\n"
+      "addvl x28, x28, #-4\n"
       "bgt 22b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z23.h }, p7/Z, [x28]\n"
-      "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z25.h }, p7/Z, [x28, #3, MUL VL]\n"
       "ld1rqw { z24.s }, p0/Z, [x24]\n"
       "ld1rqw { z20.s }, p0/Z, [x23]\n"
       ".inst 0x658abf18  // bfcvt z24.h, p7/M, z24.s\n"
       ".inst 0x658abe94  // bfcvt z20.h, p7/M, z20.s\n"
       "uzp1 z24.h, z24.h, z24.h\n"
+      "ld1h { z23.h }, p7/Z, [x28]\n"
+      "ld1h { z22.h }, p7/Z, [x28, #1, MUL VL]\n"
       "uzp1 z20.h, z20.h, z20.h\n"
       "trn1 z24.d, z24.d, z20.d\n"
+      "ld1h { z21.h }, p7/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z20.h }, p7/Z, [x28, #3, MUL VL]\n"
       ".inst 0x6477e708  // bfmmla z8.s, z24.h, z23.h\n"
-      "ld1h { z23.h }, p7/Z, [x28, #4, MUL VL]\n"
       ".inst 0x6476e70e  // bfmmla z14.s, z24.h, z22.h\n"
+      "ld1h { z23.h }, p7/Z, [x28, #4, MUL VL]\n"
       "ld1h { z22.h }, p7/Z, [x28, #5, MUL VL]\n"
       ".inst 0x6475e709  // bfmmla z9.s, z24.h, z21.h\n"
+      ".inst 0x6474e70f  // bfmmla z15.s, z24.h, z20.h\n"
       "ld1h { z21.h }, p7/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x6479e70f  // bfmmla z15.s, z24.h, z25.h\n"
       "ld1h { z20.h }, p7/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
       ".inst 0x6477e70a  // bfmmla z10.s, z24.h, z23.h\n"
       ".inst 0x6476e710  // bfmmla z16.s, z24.h, z22.h\n"
+      "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n"
       ".inst 0x6475e70b  // bfmmla z11.s, z24.h, z21.h\n"
       ".inst 0x6474e711  // bfmmla z17.s, z24.h, z20.h\n"
-      "ld1h { z23.h }, p7/Z, [x28, #-8, MUL VL]\n"
       "ld1h { z22.h }, p7/Z, [x28, #-7, MUL VL]\n"
       "ld1h { z21.h }, p7/Z, [x28, #-6, MUL VL]\n"
       "ld1h { z20.h }, p7/Z, [x28, #-5, MUL VL]\n"
-      "addvl x28, x28, #-4\n"
       ".inst 0x6477e70c  // bfmmla z12.s, z24.h, z23.h\n"
       ".inst 0x6476e712  // bfmmla z18.s, z24.h, z22.h\n"
+      "addvl x28, x28, #-4\n"
       ".inst 0x6475e70d  // bfmmla z13.s, z24.h, z21.h\n"
       ".inst 0x6474e713  // bfmmla z19.s, z24.h, z20.h\n"
       "24:"  // Height 2: Multiply loop: multiply skip
@@ -486,21 +484,21 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z4.d, z8.d, z14.d\n"
       "uzp2 z8.d, z8.d, z14.d\n"
+      "add x23, x27, x20, LSL #2\n"
       "uzp1 z14.d, z9.d, z15.d\n"
       "uzp2 z9.d, z9.d, z15.d\n"
       "uzp1 z15.d, z10.d, z16.d\n"
       "uzp2 z10.d, z10.d, z16.d\n"
       "uzp1 z16.d, z11.d, z17.d\n"
       "uzp2 z11.d, z11.d, z17.d\n"
-      "add x24, x27, x20, LSL #2\n"
       "uzp1 z17.d, z12.d, z18.d\n"
       "uzp2 z12.d, z12.d, z18.d\n"
       "uzp1 z18.d, z13.d, z19.d\n"
       "uzp2 z13.d, z13.d, z19.d\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z20.s }, p7/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z20.s }, p7/Z, [x21]\n"
       "ld1rw { z19.s }, p7/Z, [x20]\n"
       "fmin z4.s, p7/M, z4.s, z20.s\n"
       "fmin z14.s, p7/M, z14.s, z20.s\n"
@@ -534,22 +532,22 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "st1w { z17.s }, p2, [x27, #4, MUL VL]\n"
       "st1w { z18.s }, p1, [x27, #5, MUL VL]\n"
       "addvl x27, x27, #6\n"
-      "st1w { z8.s }, p6, [x24]\n"
-      "st1w { z9.s }, p5, [x24, #1, MUL VL]\n"
-      "st1w { z10.s }, p4, [x24, #2, MUL VL]\n"
-      "st1w { z11.s }, p3, [x24, #3, MUL VL]\n"
-      "st1w { z12.s }, p2, [x24, #4, MUL VL]\n"
-      "st1w { z13.s }, p1, [x24, #5, MUL VL]\n"
+      "st1w { z8.s }, p6, [x23]\n"
+      "st1w { z9.s }, p5, [x23, #1, MUL VL]\n"
+      "st1w { z10.s }, p4, [x23, #2, MUL VL]\n"
+      "st1w { z11.s }, p3, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p2, [x23, #4, MUL VL]\n"
+      "st1w { z13.s }, p1, [x23, #5, MUL VL]\n"
       "26:"  // Height 2: Writeback done
       "decw x9, ALL, MUL #6\n"
       "cmp x9, XZR\n"
       "bgt 15b\n"
       "b 54f\n"
       "27:"  // Height 3
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p6.s, x20, x9\n"
@@ -566,19 +564,19 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "cbz x10, 29f\n"
       "ld1w { z8.s }, p7/Z, [x10]\n"
       "ld1w { z9.s }, p7/Z, [x10, #1, MUL VL]\n"
-      "ld1w { z10.s }, p7/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z11.s }, p7/Z, [x10, #3, MUL VL]\n"
-      "ld1w { z12.s }, p7/Z, [x10, #4, MUL VL]\n"
-      "ld1w { z13.s }, p7/Z, [x10, #5, MUL VL]\n"
-      "addvl x10, x10, #6\n"
       "zip2 z14.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p7/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z11.s }, p7/Z, [x10, #3, MUL VL]\n"
       "zip2 z15.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
+      "ld1w { z12.s }, p7/Z, [x10, #4, MUL VL]\n"
+      "ld1w { z13.s }, p7/Z, [x10, #5, MUL VL]\n"
       "zip2 z16.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
       "zip2 z17.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
+      "addvl x10, x10, #6\n"
       "zip2 z18.d, z12.d, z12.d\n"
       "zip1 z12.d, z12.d, z12.d\n"
       "zip2 z19.d, z13.d, z13.d\n"
@@ -599,38 +597,38 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "29:"  // Height 3: no bias
       "tbz %x[flags], #0, 30f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z22.s }, p6/Z, [x27]\n"
-      "ld1w { z24.s }, p5/Z, [x27, #1, MUL VL]\n"
-      "ld1w { z0.s }, p4/Z, [x27, #2, MUL VL]\n"
-      "ld1w { z2.s }, p3/Z, [x27, #3, MUL VL]\n"
-      "ld1w { z1.s }, p2/Z, [x27, #4, MUL VL]\n"
       "add x21, x27, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z16.s }, p6/Z, [x27]\n"
+      "ld1w { z17.s }, p5/Z, [x27, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x27, #2, MUL VL]\n"
+      "ld1w { z22.s }, p3/Z, [x27, #3, MUL VL]\n"
+      "ld1w { z24.s }, p2/Z, [x27, #4, MUL VL]\n"
       "ld1w { z20.s }, p1/Z, [x27, #5, MUL VL]\n"
       "ld1w { z14.s }, p6/Z, [x21]\n"
+      "zip1 z8.d, z16.d, z14.d\n"
+      "zip2 z14.d, z16.d, z14.d\n"
       "ld1w { z15.s }, p5/Z, [x21, #1, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x21, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z15.d\n"
+      "zip2 z15.d, z17.d, z15.d\n"
       "ld1w { z17.s }, p3/Z, [x21, #3, MUL VL]\n"
       "ld1w { z18.s }, p2/Z, [x21, #4, MUL VL]\n"
+      "zip1 z10.d, z19.d, z16.d\n"
+      "zip2 z16.d, z19.d, z16.d\n"
       "ld1w { z19.s }, p1/Z, [x21, #5, MUL VL]\n"
       "ld1w { z21.s }, p6/Z, [x20]\n"
-      "zip1 z8.d, z22.d, z14.d\n"
-      "zip2 z14.d, z22.d, z14.d\n"
+      "zip1 z11.d, z22.d, z17.d\n"
+      "zip2 z17.d, z22.d, z17.d\n"
       "ld1w { z22.s }, p5/Z, [x20, #1, MUL VL]\n"
       "ld1w { z23.s }, p4/Z, [x20, #2, MUL VL]\n"
-      "zip1 z9.d, z24.d, z15.d\n"
-      "zip2 z15.d, z24.d, z15.d\n"
+      "zip1 z12.d, z24.d, z18.d\n"
+      "zip2 z18.d, z24.d, z18.d\n"
       "ld1w { z24.s }, p3/Z, [x20, #3, MUL VL]\n"
       "ld1w { z25.s }, p2/Z, [x20, #4, MUL VL]\n"
-      "zip1 z10.d, z0.d, z16.d\n"
-      "zip2 z16.d, z0.d, z16.d\n"
-      "ld1w { z0.s }, p1/Z, [x20, #5, MUL VL]\n"
-      "zip1 z11.d, z2.d, z17.d\n"
-      "zip2 z17.d, z2.d, z17.d\n"
-      "zip1 z12.d, z1.d, z18.d\n"
-      "zip2 z18.d, z1.d, z18.d\n"
       "zip1 z13.d, z20.d, z19.d\n"
       "zip2 z19.d, z20.d, z19.d\n"
+      "ld1w { z0.s }, p1/Z, [x20, #5, MUL VL]\n"
       "zip1 z20.d, z21.d, z26.d\n"
       "zip2 z26.d, z21.d, z26.d\n"
       "zip1 z21.d, z22.d, z27.d\n"
@@ -673,8 +671,8 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "mov x26, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -696,52 +694,52 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "ble 36f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z3.h }, p7/Z, [x28]\n"
-      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "sub x25, x25, #0x4\n"
-      "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z1.h }, p7/Z, [x28, #3, MUL VL]\n"
-      "cmp x25, #0x4\n"
       "ld1rqw { z5.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "ld1rqw { z4.s }, p0/Z, [x22]\n"
-      "add x22, x22, #0x10\n"
       ".inst 0x658abca5  // bfcvt z5.h, p7/M, z5.s\n"
+      "ld1rqw { z4.s }, p0/Z, [x22]\n"
       ".inst 0x658abc00  // bfcvt z0.h, p7/M, z0.s\n"
       "uzp1 z5.h, z5.h, z5.h\n"
+      "ld1h { z3.h }, p7/Z, [x28]\n"
       "uzp1 z0.h, z0.h, z0.h\n"
       ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
+      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
+      "ld1h { z1.h }, p7/Z, [x28, #2, MUL VL]\n"
       "trn1 z5.d, z5.d, z0.d\n"
       "uzp1 z4.h, z4.h, z4.h\n"
+      "ld1h { z0.h }, p7/Z, [x28, #3, MUL VL]\n"
       ".inst 0x6463e4a8  // bfmmla z8.s, z5.h, z3.h\n"
-      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
-      ".inst 0x6466e4a9  // bfmmla z9.s, z5.h, z6.h\n"
-      ".inst 0x6461e4af  // bfmmla z15.s, z5.h, z1.h\n"
       ".inst 0x6463e494  // bfmmla z20.s, z4.h, z3.h\n"
+      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28, #4, MUL VL]\n"
+      "sub x25, x25, #0x4\n"
       ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
+      ".inst 0x6461e4a9  // bfmmla z9.s, z5.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x6466e495  // bfmmla z21.s, z4.h, z6.h\n"
-      ".inst 0x6461e49b  // bfmmla z27.s, z4.h, z1.h\n"
+      "cmp x25, #0x4\n"
+      ".inst 0x6461e495  // bfmmla z21.s, z4.h, z1.h\n"
+      ".inst 0x6460e4af  // bfmmla z15.s, z5.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x28, #6, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6460e49b  // bfmmla z27.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
       ".inst 0x6463e4aa  // bfmmla z10.s, z5.h, z3.h\n"
       ".inst 0x6463e496  // bfmmla z22.s, z4.h, z3.h\n"
       ".inst 0x6462e4b0  // bfmmla z16.s, z5.h, z2.h\n"
+      "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
       ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
+      "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
       ".inst 0x6460e4b1  // bfmmla z17.s, z5.h, z0.h\n"
-      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
-      "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n"
-      "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n"
       "ld1h { z1.h }, p7/Z, [x28, #-6, MUL VL]\n"
+      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x28, #-5, MUL VL]\n"
-      "addvl x28, x28, #-4\n"
       ".inst 0x6463e4ac  // bfmmla z12.s, z5.h, z3.h\n"
+      "addvl x28, x28, #-4\n"
       ".inst 0x6463e498  // bfmmla z24.s, z4.h, z3.h\n"
       ".inst 0x6462e4b2  // bfmmla z18.s, z5.h, z2.h\n"
       ".inst 0x6462e49e  // bfmmla z30.s, z4.h, z2.h\n"
@@ -752,47 +750,47 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "bgt 35b\n"
       "36:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z3.h }, p7/Z, [x28]\n"
-      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "ld1h { z6.h }, p7/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z1.h }, p7/Z, [x28, #3, MUL VL]\n"
       "ld1rqw { z5.s }, p0/Z, [x24]\n"
       "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "ld1rqw { z4.s }, p0/Z, [x22]\n"
       ".inst 0x658abca5  // bfcvt z5.h, p7/M, z5.s\n"
+      "ld1rqw { z4.s }, p0/Z, [x22]\n"
       ".inst 0x658abc00  // bfcvt z0.h, p7/M, z0.s\n"
       "uzp1 z5.h, z5.h, z5.h\n"
+      "ld1h { z3.h }, p7/Z, [x28]\n"
       "uzp1 z0.h, z0.h, z0.h\n"
       ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
+      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
+      "ld1h { z1.h }, p7/Z, [x28, #2, MUL VL]\n"
       "trn1 z5.d, z5.d, z0.d\n"
       "uzp1 z4.h, z4.h, z4.h\n"
+      "ld1h { z0.h }, p7/Z, [x28, #3, MUL VL]\n"
       ".inst 0x6463e4a8  // bfmmla z8.s, z5.h, z3.h\n"
-      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
-      ".inst 0x6466e4a9  // bfmmla z9.s, z5.h, z6.h\n"
-      ".inst 0x6461e4af  // bfmmla z15.s, z5.h, z1.h\n"
       ".inst 0x6463e494  // bfmmla z20.s, z4.h, z3.h\n"
+      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28, #4, MUL VL]\n"
       ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
+      ".inst 0x6461e4a9  // bfmmla z9.s, z5.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x6466e495  // bfmmla z21.s, z4.h, z6.h\n"
-      ".inst 0x6461e49b  // bfmmla z27.s, z4.h, z1.h\n"
+      ".inst 0x6461e495  // bfmmla z21.s, z4.h, z1.h\n"
+      ".inst 0x6460e4af  // bfmmla z15.s, z5.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x28, #6, MUL VL]\n"
+      ".inst 0x6460e49b  // bfmmla z27.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
       ".inst 0x6463e4aa  // bfmmla z10.s, z5.h, z3.h\n"
       ".inst 0x6463e496  // bfmmla z22.s, z4.h, z3.h\n"
       ".inst 0x6462e4b0  // bfmmla z16.s, z5.h, z2.h\n"
+      "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n"
       ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
       ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
+      "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n"
       ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
       ".inst 0x6460e4b1  // bfmmla z17.s, z5.h, z0.h\n"
-      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
-      "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n"
-      "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n"
       "ld1h { z1.h }, p7/Z, [x28, #-6, MUL VL]\n"
+      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x28, #-5, MUL VL]\n"
-      "addvl x28, x28, #-4\n"
       ".inst 0x6463e4ac  // bfmmla z12.s, z5.h, z3.h\n"
+      "addvl x28, x28, #-4\n"
       ".inst 0x6463e498  // bfmmla z24.s, z4.h, z3.h\n"
       ".inst 0x6462e4b2  // bfmmla z18.s, z5.h, z2.h\n"
       ".inst 0x6462e49e  // bfmmla z30.s, z4.h, z2.h\n"
@@ -806,16 +804,16 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "cmp x26, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
       "uzp1 z4.d, z8.d, z14.d\n"
       "uzp2 z8.d, z8.d, z14.d\n"
       "uzp1 z14.d, z9.d, z15.d\n"
       "uzp2 z9.d, z9.d, z15.d\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp1 z15.d, z10.d, z16.d\n"
       "uzp2 z10.d, z10.d, z16.d\n"
-      "add x24, x27, x20, LSL #2\n"
       "uzp1 z16.d, z11.d, z17.d\n"
       "uzp2 z11.d, z11.d, z17.d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z17.d, z12.d, z18.d\n"
       "uzp2 z12.d, z12.d, z18.d\n"
       "uzp1 z18.d, z13.d, z19.d\n"
@@ -827,9 +825,9 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "uzp1 z24.d, z24.d, z30.d\n"
       "uzp1 z25.d, z25.d, z31.d\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z0.s }, p7/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z0.s }, p7/Z, [x21]\n"
       "ld1rw { z19.s }, p7/Z, [x20]\n"
       "fmin z4.s, p7/M, z4.s, z0.s\n"
       "fmin z14.s, p7/M, z14.s, z0.s\n"
@@ -875,18 +873,18 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "st1w { z17.s }, p2, [x27, #4, MUL VL]\n"
       "st1w { z18.s }, p1, [x27, #5, MUL VL]\n"
       "addvl x27, x27, #6\n"
-      "st1w { z8.s }, p6, [x24]\n"
-      "st1w { z9.s }, p5, [x24, #1, MUL VL]\n"
-      "st1w { z10.s }, p4, [x24, #2, MUL VL]\n"
-      "st1w { z11.s }, p3, [x24, #3, MUL VL]\n"
-      "st1w { z12.s }, p2, [x24, #4, MUL VL]\n"
-      "st1w { z13.s }, p1, [x24, #5, MUL VL]\n"
-      "st1w { z20.s }, p6, [x23]\n"
-      "st1w { z21.s }, p5, [x23, #1, MUL VL]\n"
-      "st1w { z22.s }, p4, [x23, #2, MUL VL]\n"
-      "st1w { z23.s }, p3, [x23, #3, MUL VL]\n"
-      "st1w { z24.s }, p2, [x23, #4, MUL VL]\n"
-      "st1w { z25.s }, p1, [x23, #5, MUL VL]\n"
+      "st1w { z8.s }, p6, [x23]\n"
+      "st1w { z9.s }, p5, [x23, #1, MUL VL]\n"
+      "st1w { z10.s }, p4, [x23, #2, MUL VL]\n"
+      "st1w { z11.s }, p3, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p2, [x23, #4, MUL VL]\n"
+      "st1w { z13.s }, p1, [x23, #5, MUL VL]\n"
+      "st1w { z20.s }, p6, [x22]\n"
+      "st1w { z21.s }, p5, [x22, #1, MUL VL]\n"
+      "st1w { z22.s }, p4, [x22, #2, MUL VL]\n"
+      "st1w { z23.s }, p3, [x22, #3, MUL VL]\n"
+      "st1w { z24.s }, p2, [x22, #4, MUL VL]\n"
+      "st1w { z25.s }, p1, [x22, #5, MUL VL]\n"
       "39:"  // Height 3: Writeback done
       "decw x9, ALL, MUL #6\n"
       "cmp x9, XZR\n"
@@ -894,13 +892,12 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "b 54f\n"
       "40:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x10\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x10, %x[bias]\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p6.s, x20, x9\n"
@@ -917,19 +914,19 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "cbz x10, 42f\n"
       "ld1w { z8.s }, p7/Z, [x10]\n"
       "ld1w { z9.s }, p7/Z, [x10, #1, MUL VL]\n"
-      "ld1w { z10.s }, p7/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z11.s }, p7/Z, [x10, #3, MUL VL]\n"
-      "ld1w { z12.s }, p7/Z, [x10, #4, MUL VL]\n"
-      "ld1w { z13.s }, p7/Z, [x10, #5, MUL VL]\n"
-      "addvl x10, x10, #6\n"
       "zip2 z14.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p7/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z11.s }, p7/Z, [x10, #3, MUL VL]\n"
       "zip2 z15.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
+      "ld1w { z12.s }, p7/Z, [x10, #4, MUL VL]\n"
+      "ld1w { z13.s }, p7/Z, [x10, #5, MUL VL]\n"
       "zip2 z16.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
       "zip2 z17.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
+      "addvl x10, x10, #6\n"
       "zip2 z18.d, z12.d, z12.d\n"
       "zip1 z12.d, z12.d, z12.d\n"
       "zip2 z19.d, z13.d, z13.d\n"
@@ -950,51 +947,51 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "42:"  // Height 4: no bias
       "tbz %x[flags], #0, 43f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z22.s }, p6/Z, [x27]\n"
-      "ld1w { z24.s }, p5/Z, [x27, #1, MUL VL]\n"
-      "ld1w { z26.s }, p4/Z, [x27, #2, MUL VL]\n"
-      "ld1w { z27.s }, p3/Z, [x27, #3, MUL VL]\n"
-      "ld1w { z29.s }, p2/Z, [x27, #4, MUL VL]\n"
       "add x22, x27, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z16.s }, p6/Z, [x27]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p5/Z, [x27, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x27, #2, MUL VL]\n"
+      "ld1w { z22.s }, p3/Z, [x27, #3, MUL VL]\n"
+      "ld1w { z24.s }, p2/Z, [x27, #4, MUL VL]\n"
       "ld1w { z20.s }, p1/Z, [x27, #5, MUL VL]\n"
       "ld1w { z14.s }, p6/Z, [x22]\n"
+      "zip1 z8.d, z16.d, z14.d\n"
+      "zip2 z14.d, z16.d, z14.d\n"
       "ld1w { z15.s }, p5/Z, [x22, #1, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x22, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z15.d\n"
+      "zip2 z15.d, z17.d, z15.d\n"
       "ld1w { z17.s }, p3/Z, [x22, #3, MUL VL]\n"
       "ld1w { z18.s }, p2/Z, [x22, #4, MUL VL]\n"
+      "zip1 z10.d, z19.d, z16.d\n"
+      "zip2 z16.d, z19.d, z16.d\n"
       "ld1w { z19.s }, p1/Z, [x22, #5, MUL VL]\n"
       "ld1w { z21.s }, p6/Z, [x21]\n"
-      "zip1 z8.d, z22.d, z14.d\n"
-      "zip2 z14.d, z22.d, z14.d\n"
+      "zip1 z11.d, z22.d, z17.d\n"
+      "zip2 z17.d, z22.d, z17.d\n"
       "ld1w { z22.s }, p5/Z, [x21, #1, MUL VL]\n"
       "ld1w { z23.s }, p4/Z, [x21, #2, MUL VL]\n"
-      "zip1 z9.d, z24.d, z15.d\n"
-      "zip2 z15.d, z24.d, z15.d\n"
+      "zip1 z12.d, z24.d, z18.d\n"
+      "zip2 z18.d, z24.d, z18.d\n"
       "ld1w { z24.s }, p3/Z, [x21, #3, MUL VL]\n"
       "ld1w { z25.s }, p2/Z, [x21, #4, MUL VL]\n"
-      "zip1 z10.d, z26.d, z16.d\n"
-      "zip2 z16.d, z26.d, z16.d\n"
-      "ld1w { z0.s }, p1/Z, [x21, #5, MUL VL]\n"
-      "ld1w { z26.s }, p6/Z, [x20]\n"
-      "zip1 z11.d, z27.d, z17.d\n"
-      "zip2 z17.d, z27.d, z17.d\n"
-      "ld1w { z27.s }, p5/Z, [x20, #1, MUL VL]\n"
-      "ld1w { z28.s }, p4/Z, [x20, #2, MUL VL]\n"
-      "zip1 z12.d, z29.d, z18.d\n"
-      "zip2 z18.d, z29.d, z18.d\n"
-      "ld1w { z29.s }, p3/Z, [x20, #3, MUL VL]\n"
-      "ld1w { z30.s }, p2/Z, [x20, #4, MUL VL]\n"
       "zip1 z13.d, z20.d, z19.d\n"
       "zip2 z19.d, z20.d, z19.d\n"
-      "ld1w { z31.s }, p1/Z, [x20, #5, MUL VL]\n"
+      "ld1w { z0.s }, p1/Z, [x21, #5, MUL VL]\n"
+      "ld1w { z26.s }, p6/Z, [x20]\n"
       "zip1 z20.d, z21.d, z26.d\n"
       "zip2 z26.d, z21.d, z26.d\n"
+      "ld1w { z27.s }, p5/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z28.s }, p4/Z, [x20, #2, MUL VL]\n"
       "zip1 z21.d, z22.d, z27.d\n"
       "zip2 z27.d, z22.d, z27.d\n"
+      "ld1w { z29.s }, p3/Z, [x20, #3, MUL VL]\n"
+      "ld1w { z30.s }, p2/Z, [x20, #4, MUL VL]\n"
       "zip1 z22.d, z23.d, z28.d\n"
       "zip2 z28.d, z23.d, z28.d\n"
+      "ld1w { z31.s }, p1/Z, [x20, #5, MUL VL]\n"
       "zip1 z23.d, z24.d, z29.d\n"
       "zip2 z29.d, z24.d, z29.d\n"
       "zip1 z24.d, z25.d, z30.d\n"
@@ -1031,8 +1028,8 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "mov x26, #0x0\n"
       "45:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1057,136 +1054,136 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "ble 49f\n"
       "48:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z3.h }, p7/Z, [x28]\n"
-      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "sub x25, x25, #0x4\n"
-      "ld1h { z7.h }, p7/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z6.h }, p7/Z, [x28, #3, MUL VL]\n"
-      "cmp x25, #0x4\n"
-      "ld1rqw { z5.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqw { z1.s }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "ld1rqw { z4.s }, p0/Z, [x22]\n"
-      "ld1rqw { z0.s }, p0/Z, [x21]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
+      "ld1rqw { z7.s }, p0/Z, [x24]\n"
+      "ld1rqw { z6.s }, p0/Z, [x23]\n"
+      ".inst 0x658abce7  // bfcvt z7.h, p7/M, z7.s\n"
+      "ld1rqw { z5.s }, p0/Z, [x22]\n"
+      "ld1rqw { z4.s }, p0/Z, [x21]\n"
+      ".inst 0x658abcc6  // bfcvt z6.h, p7/M, z6.s\n"
       ".inst 0x658abca5  // bfcvt z5.h, p7/M, z5.s\n"
-      ".inst 0x658abc21  // bfcvt z1.h, p7/M, z1.s\n"
       ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
-      ".inst 0x658abc00  // bfcvt z0.h, p7/M, z0.s\n"
+      "uzp1 z7.h, z7.h, z7.h\n"
+      "ld1h { z3.h }, p7/Z, [x28]\n"
+      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
       "uzp1 z5.h, z5.h, z5.h\n"
-      "uzp1 z1.h, z1.h, z1.h\n"
+      "ld1h { z1.h }, p7/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z0.h }, p7/Z, [x28, #3, MUL VL]\n"
       "uzp1 z4.h, z4.h, z4.h\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
-      "trn1 z5.d, z5.d, z1.d\n"
-      "trn1 z4.d, z4.d, z0.d\n"
-      ".inst 0x6463e4a8  // bfmmla z8.s, z5.h, z3.h\n"
-      ".inst 0x6463e494  // bfmmla z20.s, z4.h, z3.h\n"
+      "trn1 z7.d, z7.d, z6.d\n"
+      ".inst 0x6463e4e8  // bfmmla z8.s, z7.h, z3.h\n"
+      "sub x25, x25, #0x4\n"
+      "trn1 z5.d, z5.d, z4.d\n"
+      ".inst 0x6463e4b4  // bfmmla z20.s, z5.h, z3.h\n"
+      ".inst 0x6462e4ee  // bfmmla z14.s, z7.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28, #4, MUL VL]\n"
-      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
-      ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
+      ".inst 0x6462e4ba  // bfmmla z26.s, z5.h, z2.h\n"
+      ".inst 0x6461e4e9  // bfmmla z9.s, z7.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x6467e4a9  // bfmmla z9.s, z5.h, z7.h\n"
-      ".inst 0x6467e495  // bfmmla z21.s, z4.h, z7.h\n"
-      ".inst 0x6466e4af  // bfmmla z15.s, z5.h, z6.h\n"
+      "cmp x25, #0x4\n"
+      ".inst 0x6461e4b5  // bfmmla z21.s, z5.h, z1.h\n"
+      ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x6466e49b  // bfmmla z27.s, z4.h, z6.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6460e4bb  // bfmmla z27.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x6463e4aa  // bfmmla z10.s, z5.h, z3.h\n"
-      ".inst 0x6463e496  // bfmmla z22.s, z4.h, z3.h\n"
-      ".inst 0x6462e4b0  // bfmmla z16.s, z5.h, z2.h\n"
-      ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
-      ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
-      ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
-      ".inst 0x6460e4b1  // bfmmla z17.s, z5.h, z0.h\n"
+      ".inst 0x6463e4ea  // bfmmla z10.s, z7.h, z3.h\n"
+      ".inst 0x6463e4b6  // bfmmla z22.s, z5.h, z3.h\n"
+      ".inst 0x6462e4f0  // bfmmla z16.s, z7.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6462e4bc  // bfmmla z28.s, z5.h, z2.h\n"
+      ".inst 0x6461e4eb  // bfmmla z11.s, z7.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n"
-      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
+      "add x22, x22, #0x10\n"
+      ".inst 0x6461e4b7  // bfmmla z23.s, z5.h, z1.h\n"
+      ".inst 0x6460e4f1  // bfmmla z17.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x28, #-6, MUL VL]\n"
+      "add x21, x21, #0x10\n"
+      ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x6463e4ec  // bfmmla z12.s, z7.h, z3.h\n"
       "addvl x28, x28, #-4\n"
-      ".inst 0x6463e4ac  // bfmmla z12.s, z5.h, z3.h\n"
-      ".inst 0x6463e498  // bfmmla z24.s, z4.h, z3.h\n"
-      ".inst 0x6462e4b2  // bfmmla z18.s, z5.h, z2.h\n"
-      ".inst 0x6462e49e  // bfmmla z30.s, z4.h, z2.h\n"
-      ".inst 0x6461e4ad  // bfmmla z13.s, z5.h, z1.h\n"
-      ".inst 0x6461e499  // bfmmla z25.s, z4.h, z1.h\n"
-      ".inst 0x6460e4b3  // bfmmla z19.s, z5.h, z0.h\n"
-      ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
+      ".inst 0x6463e4b8  // bfmmla z24.s, z5.h, z3.h\n"
+      ".inst 0x6462e4f2  // bfmmla z18.s, z7.h, z2.h\n"
+      ".inst 0x6462e4be  // bfmmla z30.s, z5.h, z2.h\n"
+      ".inst 0x6461e4ed  // bfmmla z13.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b9  // bfmmla z25.s, z5.h, z1.h\n"
+      ".inst 0x6460e4f3  // bfmmla z19.s, z7.h, z0.h\n"
+      ".inst 0x6460e4bf  // bfmmla z31.s, z5.h, z0.h\n"
       "bgt 48b\n"
       "49:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x25\n"
-      "ld1h { z3.h }, p7/Z, [x28]\n"
-      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
-      "ld1h { z7.h }, p7/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z6.h }, p7/Z, [x28, #3, MUL VL]\n"
-      "ld1rqw { z5.s }, p0/Z, [x24]\n"
-      "ld1rqw { z1.s }, p0/Z, [x23]\n"
-      "ld1rqw { z4.s }, p0/Z, [x22]\n"
-      "ld1rqw { z0.s }, p0/Z, [x21]\n"
+      "ld1rqw { z7.s }, p0/Z, [x24]\n"
+      "ld1rqw { z6.s }, p0/Z, [x23]\n"
+      ".inst 0x658abce7  // bfcvt z7.h, p7/M, z7.s\n"
+      "ld1rqw { z5.s }, p0/Z, [x22]\n"
+      "ld1rqw { z4.s }, p0/Z, [x21]\n"
+      ".inst 0x658abcc6  // bfcvt z6.h, p7/M, z6.s\n"
       ".inst 0x658abca5  // bfcvt z5.h, p7/M, z5.s\n"
-      ".inst 0x658abc21  // bfcvt z1.h, p7/M, z1.s\n"
       ".inst 0x658abc84  // bfcvt z4.h, p7/M, z4.s\n"
-      ".inst 0x658abc00  // bfcvt z0.h, p7/M, z0.s\n"
+      "uzp1 z7.h, z7.h, z7.h\n"
+      "ld1h { z3.h }, p7/Z, [x28]\n"
+      "ld1h { z2.h }, p7/Z, [x28, #1, MUL VL]\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
       "uzp1 z5.h, z5.h, z5.h\n"
-      "uzp1 z1.h, z1.h, z1.h\n"
+      "ld1h { z1.h }, p7/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z0.h }, p7/Z, [x28, #3, MUL VL]\n"
       "uzp1 z4.h, z4.h, z4.h\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
-      "trn1 z5.d, z5.d, z1.d\n"
-      "trn1 z4.d, z4.d, z0.d\n"
-      ".inst 0x6463e4a8  // bfmmla z8.s, z5.h, z3.h\n"
-      ".inst 0x6463e494  // bfmmla z20.s, z4.h, z3.h\n"
+      "trn1 z7.d, z7.d, z6.d\n"
+      ".inst 0x6463e4e8  // bfmmla z8.s, z7.h, z3.h\n"
+      "trn1 z5.d, z5.d, z4.d\n"
+      ".inst 0x6463e4b4  // bfmmla z20.s, z5.h, z3.h\n"
+      ".inst 0x6462e4ee  // bfmmla z14.s, z7.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28, #4, MUL VL]\n"
-      ".inst 0x6462e4ae  // bfmmla z14.s, z5.h, z2.h\n"
-      ".inst 0x6462e49a  // bfmmla z26.s, z4.h, z2.h\n"
+      ".inst 0x6462e4ba  // bfmmla z26.s, z5.h, z2.h\n"
+      ".inst 0x6461e4e9  // bfmmla z9.s, z7.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x6467e4a9  // bfmmla z9.s, z5.h, z7.h\n"
-      ".inst 0x6467e495  // bfmmla z21.s, z4.h, z7.h\n"
-      ".inst 0x6466e4af  // bfmmla z15.s, z5.h, z6.h\n"
+      ".inst 0x6461e4b5  // bfmmla z21.s, z5.h, z1.h\n"
+      ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x6466e49b  // bfmmla z27.s, z4.h, z6.h\n"
+      ".inst 0x6460e4bb  // bfmmla z27.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x6463e4aa  // bfmmla z10.s, z5.h, z3.h\n"
-      ".inst 0x6463e496  // bfmmla z22.s, z4.h, z3.h\n"
-      ".inst 0x6462e4b0  // bfmmla z16.s, z5.h, z2.h\n"
-      ".inst 0x6462e49c  // bfmmla z28.s, z4.h, z2.h\n"
-      ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
-      ".inst 0x6461e497  // bfmmla z23.s, z4.h, z1.h\n"
-      ".inst 0x6460e4b1  // bfmmla z17.s, z5.h, z0.h\n"
+      ".inst 0x6463e4ea  // bfmmla z10.s, z7.h, z3.h\n"
+      ".inst 0x6463e4b6  // bfmmla z22.s, z5.h, z3.h\n"
+      ".inst 0x6462e4f0  // bfmmla z16.s, z7.h, z2.h\n"
       "ld1h { z3.h }, p7/Z, [x28, #-8, MUL VL]\n"
+      ".inst 0x6462e4bc  // bfmmla z28.s, z5.h, z2.h\n"
+      ".inst 0x6461e4eb  // bfmmla z11.s, z7.h, z1.h\n"
       "ld1h { z2.h }, p7/Z, [x28, #-7, MUL VL]\n"
-      ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
+      ".inst 0x6461e4b7  // bfmmla z23.s, z5.h, z1.h\n"
+      ".inst 0x6460e4f1  // bfmmla z17.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p7/Z, [x28, #-6, MUL VL]\n"
+      ".inst 0x6460e4bd  // bfmmla z29.s, z5.h, z0.h\n"
       "ld1h { z0.h }, p7/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x6463e4ec  // bfmmla z12.s, z7.h, z3.h\n"
       "addvl x28, x28, #-4\n"
-      ".inst 0x6463e4ac  // bfmmla z12.s, z5.h, z3.h\n"
-      ".inst 0x6463e498  // bfmmla z24.s, z4.h, z3.h\n"
-      ".inst 0x6462e4b2  // bfmmla z18.s, z5.h, z2.h\n"
-      ".inst 0x6462e49e  // bfmmla z30.s, z4.h, z2.h\n"
-      ".inst 0x6461e4ad  // bfmmla z13.s, z5.h, z1.h\n"
-      ".inst 0x6461e499  // bfmmla z25.s, z4.h, z1.h\n"
-      ".inst 0x6460e4b3  // bfmmla z19.s, z5.h, z0.h\n"
-      ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
+      ".inst 0x6463e4b8  // bfmmla z24.s, z5.h, z3.h\n"
+      ".inst 0x6462e4f2  // bfmmla z18.s, z7.h, z2.h\n"
+      ".inst 0x6462e4be  // bfmmla z30.s, z5.h, z2.h\n"
+      ".inst 0x6461e4ed  // bfmmla z13.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b9  // bfmmla z25.s, z5.h, z1.h\n"
+      ".inst 0x6460e4f3  // bfmmla z19.s, z7.h, z0.h\n"
+      ".inst 0x6460e4bf  // bfmmla z31.s, z5.h, z0.h\n"
       "50:"  // Height 4: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x26, x26, #0x1\n"
       "cmp x26, x20\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x27, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp1 z4.d, z8.d, z14.d\n"
       "uzp2 z8.d, z8.d, z14.d\n"
       "uzp1 z14.d, z9.d, z15.d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z15.d\n"
       "uzp1 z15.d, z10.d, z16.d\n"
       "uzp2 z10.d, z10.d, z16.d\n"
-      "add x24, x27, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z16.d, z11.d, z17.d\n"
       "uzp2 z11.d, z11.d, z17.d\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 z17.d, z12.d, z18.d\n"
       "uzp2 z12.d, z12.d, z18.d\n"
       "uzp1 z18.d, z13.d, z19.d\n"
@@ -1204,9 +1201,9 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "uzp1 z30.d, z25.d, z31.d\n"
       "uzp2 z25.d, z25.d, z31.d\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p7/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p7/Z, [x21]\n"
       "ld1rw { z0.s }, p7/Z, [x20]\n"
       "fmin z4.s, p7/M, z4.s, z1.s\n"
       "fmin z14.s, p7/M, z14.s, z1.s\n"
@@ -1264,24 +1261,24 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "st1w { z17.s }, p2, [x27, #4, MUL VL]\n"
       "st1w { z18.s }, p1, [x27, #5, MUL VL]\n"
       "addvl x27, x27, #6\n"
-      "st1w { z8.s }, p6, [x24]\n"
-      "st1w { z9.s }, p5, [x24, #1, MUL VL]\n"
-      "st1w { z10.s }, p4, [x24, #2, MUL VL]\n"
-      "st1w { z11.s }, p3, [x24, #3, MUL VL]\n"
-      "st1w { z12.s }, p2, [x24, #4, MUL VL]\n"
-      "st1w { z13.s }, p1, [x24, #5, MUL VL]\n"
-      "st1w { z19.s }, p6, [x23]\n"
-      "st1w { z26.s }, p5, [x23, #1, MUL VL]\n"
-      "st1w { z27.s }, p4, [x23, #2, MUL VL]\n"
-      "st1w { z28.s }, p3, [x23, #3, MUL VL]\n"
-      "st1w { z29.s }, p2, [x23, #4, MUL VL]\n"
-      "st1w { z30.s }, p1, [x23, #5, MUL VL]\n"
-      "st1w { z20.s }, p6, [x22]\n"
-      "st1w { z21.s }, p5, [x22, #1, MUL VL]\n"
-      "st1w { z22.s }, p4, [x22, #2, MUL VL]\n"
-      "st1w { z23.s }, p3, [x22, #3, MUL VL]\n"
-      "st1w { z24.s }, p2, [x22, #4, MUL VL]\n"
-      "st1w { z25.s }, p1, [x22, #5, MUL VL]\n"
+      "st1w { z8.s }, p6, [x23]\n"
+      "st1w { z9.s }, p5, [x23, #1, MUL VL]\n"
+      "st1w { z10.s }, p4, [x23, #2, MUL VL]\n"
+      "st1w { z11.s }, p3, [x23, #3, MUL VL]\n"
+      "st1w { z12.s }, p2, [x23, #4, MUL VL]\n"
+      "st1w { z13.s }, p1, [x23, #5, MUL VL]\n"
+      "st1w { z19.s }, p6, [x22]\n"
+      "st1w { z26.s }, p5, [x22, #1, MUL VL]\n"
+      "st1w { z27.s }, p4, [x22, #2, MUL VL]\n"
+      "st1w { z28.s }, p3, [x22, #3, MUL VL]\n"
+      "st1w { z29.s }, p2, [x22, #4, MUL VL]\n"
+      "st1w { z30.s }, p1, [x22, #5, MUL VL]\n"
+      "st1w { z20.s }, p6, [x21]\n"
+      "st1w { z21.s }, p5, [x21, #1, MUL VL]\n"
+      "st1w { z22.s }, p4, [x21, #2, MUL VL]\n"
+      "st1w { z23.s }, p3, [x21, #3, MUL VL]\n"
+      "st1w { z24.s }, p2, [x21, #4, MUL VL]\n"
+      "st1w { z25.s }, p1, [x21, #5, MUL VL]\n"
       "52:"  // Height 4: Writeback done
       "decw x9, ALL, MUL #6\n"
       "cmp x9, XZR\n"
@@ -1298,8 +1295,8 @@ void sve_hybrid_fp32bf16fp32_mmla_4x6VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "54:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp
index 823d839289..230c0b77d7 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 8, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 8, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp
index 11526b6f65..d31e68993e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp
@@ -48,19 +48,18 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
-        const float *bias = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -81,7 +80,6 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
     ka.string_lengths = string_lengths;
     ka.N = N;
     ka.B_ptr = B_ptr;
-    ka.bias = bias;
     switch(act.type) {
         default:
         case Activation::Type::None:
@@ -105,10 +103,10 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cmp %x[M], #0x2\n"
       "bgt 27f\n"
       "beq 14f\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -121,26 +119,26 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cbz x12, 3f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "b 5f\n"
       "3:"  // Height 1: no bias
       "tbz %x[flags], #0, 4f\n"
-      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z16.s }, p4/Z, [x9]\n"
       "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "zip1 z8.d, z16.d, z12.d\n"
+      "zip2 z12.d, z16.d, z12.d\n"
       "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "zip1 z8.d, z19.d, z12.d\n"
-      "zip2 z12.d, z19.d, z12.d\n"
       "zip1 z9.d, z18.d, z13.d\n"
       "zip2 z13.d, z18.d, z13.d\n"
       "zip1 z10.d, z17.d, z14.d\n"
@@ -161,8 +159,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "6:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 7f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -178,52 +176,52 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "ble 10f\n"
       "9:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
       "ld1rqw { z18.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       ".inst 0x658ab652  // bfcvt z18.h, p5/M, z18.s\n"
       "uzp1 z18.h, z18.h, z18.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
       ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
+      "add x26, x26, #0x10\n"
+      "addvl x10, x10, #8\n"
       "bgt 9b\n"
       "10:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "ld1rqw { z18.s }, p0/Z, [x26]\n"
       ".inst 0x658ab652  // bfcvt z18.h, p5/M, z18.s\n"
       "uzp1 z18.h, z18.h, z18.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x6470e64c  // bfmmla z12.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
-      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
       ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
+      "addvl x10, x10, #8\n"
       "11:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -234,9 +232,9 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "uzp1 z10.d, z10.d, z14.d\n"
       "uzp1 z11.d, z11.d, z15.d\n"
       "tbz %x[flags], #1, 12f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z8.s, p5/M, z8.s, z17.s\n"
       "fmin z9.s, p5/M, z9.s, z17.s\n"
@@ -258,10 +256,10 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "bgt 2b\n"
       "b 80f\n"
       "14:"  // Height 2
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -274,34 +272,34 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cbz x12, 16f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "b 18f\n"
       "16:"  // Height 2: no bias
       "tbz %x[flags], #0, 17f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z19.s }, p4/Z, [x9]\n"
       "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
-      "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z8.d, z19.d, z12.d\n"
       "zip2 z12.d, z19.d, z12.d\n"
+      "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z9.d, z18.d, z13.d\n"
       "zip2 z13.d, z18.d, z13.d\n"
+      "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z10.d, z17.d, z14.d\n"
       "zip2 z14.d, z17.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
@@ -320,8 +318,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "19:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 20f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -340,61 +338,61 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "ble 23f\n"
       "22:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z19.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
-      "ld1rqw { z18.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqw { z16.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqw { z19.s }, p0/Z, [x26]\n"
+      "ld1rqw { z18.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab673  // bfcvt z19.h, p5/M, z19.s\n"
       ".inst 0x658ab652  // bfcvt z18.h, p5/M, z18.s\n"
-      ".inst 0x658ab610  // bfcvt z16.h, p5/M, z16.s\n"
+      "uzp1 z19.h, z19.h, z19.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "uzp1 z18.h, z18.h, z18.h\n"
-      "uzp1 z16.h, z16.h, z16.h\n"
-      "trn1 z18.d, z18.d, z16.d\n"
-      ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
+      "trn1 z19.d, z19.d, z18.d\n"
+      ".inst 0x6471e668  // bfmmla z8.s, z19.h, z17.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6473e64c  // bfmmla z12.s, z18.h, z19.h\n"
+      ".inst 0x6470e66c  // bfmmla z12.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
+      ".inst 0x6471e669  // bfmmla z9.s, z19.h, z17.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      ".inst 0x6470e66d  // bfmmla z13.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
+      ".inst 0x6471e66a  // bfmmla z10.s, z19.h, z17.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      ".inst 0x6470e66e  // bfmmla z14.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
+      ".inst 0x6471e66b  // bfmmla z11.s, z19.h, z17.h\n"
+      ".inst 0x6470e66f  // bfmmla z15.s, z19.h, z16.h\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
-      ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
       "bgt 22b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z17.h }, p5/Z, [x10]\n"
-      "ld1h { z19.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "ld1rqw { z18.s }, p0/Z, [x26]\n"
-      "ld1rqw { z16.s }, p0/Z, [x25]\n"
+      "ld1rqw { z19.s }, p0/Z, [x26]\n"
+      "ld1rqw { z18.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab673  // bfcvt z19.h, p5/M, z19.s\n"
       ".inst 0x658ab652  // bfcvt z18.h, p5/M, z18.s\n"
-      ".inst 0x658ab610  // bfcvt z16.h, p5/M, z16.s\n"
+      "uzp1 z19.h, z19.h, z19.h\n"
+      "ld1h { z17.h }, p5/Z, [x10]\n"
+      "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n"
       "uzp1 z18.h, z18.h, z18.h\n"
-      "uzp1 z16.h, z16.h, z16.h\n"
-      "trn1 z18.d, z18.d, z16.d\n"
-      ".inst 0x6471e648  // bfmmla z8.s, z18.h, z17.h\n"
+      "trn1 z19.d, z19.d, z18.d\n"
+      ".inst 0x6471e668  // bfmmla z8.s, z19.h, z17.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6473e64c  // bfmmla z12.s, z18.h, z19.h\n"
+      ".inst 0x6470e66c  // bfmmla z12.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6471e649  // bfmmla z9.s, z18.h, z17.h\n"
+      ".inst 0x6471e669  // bfmmla z9.s, z19.h, z17.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6470e64d  // bfmmla z13.s, z18.h, z16.h\n"
+      ".inst 0x6470e66d  // bfmmla z13.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6471e64a  // bfmmla z10.s, z18.h, z17.h\n"
+      ".inst 0x6471e66a  // bfmmla z10.s, z19.h, z17.h\n"
       "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6470e64e  // bfmmla z14.s, z18.h, z16.h\n"
+      ".inst 0x6470e66e  // bfmmla z14.s, z19.h, z16.h\n"
       "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n"
+      ".inst 0x6471e66b  // bfmmla z11.s, z19.h, z17.h\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6471e64b  // bfmmla z11.s, z18.h, z17.h\n"
-      ".inst 0x6470e64f  // bfmmla z15.s, z18.h, z16.h\n"
+      ".inst 0x6470e66f  // bfmmla z15.s, z19.h, z16.h\n"
       "24:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -403,17 +401,17 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z6.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
+      "add x25, x9, x20, LSL #2\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x26, x9, x20, LSL #2\n"
       "tbz %x[flags], #1, 25f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z17.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z17.s }, p5/Z, [x21]\n"
       "ld1rw { z16.s }, p5/Z, [x20]\n"
       "fmin z6.s, p5/M, z6.s, z17.s\n"
       "fmin z12.s, p5/M, z12.s, z17.s\n"
@@ -437,20 +435,20 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
       "26:"  // Height 2: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 15b\n"
       "b 80f\n"
       "27:"  // Height 3
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -463,15 +461,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cbz x12, 29f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -486,28 +484,28 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "29:"  // Height 3: no bias
       "tbz %x[flags], #0, 30f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z24.s }, p4/Z, [x9]\n"
-      "ld1w { z26.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x21, x9, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x20]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "zip1 z8.d, z24.d, z12.d\n"
-      "zip2 z12.d, z24.d, z12.d\n"
-      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z9.d, z26.d, z13.d\n"
-      "zip2 z13.d, z26.d, z13.d\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
       "zip1 z17.d, z18.d, z21.d\n"
@@ -538,8 +536,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -561,83 +559,83 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "ble 36f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
-      "ld1rqw { z27.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqw { z24.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqw { z28.s }, p0/Z, [x26]\n"
+      "ld1rqw { z27.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab79c  // bfcvt z28.h, p5/M, z28.s\n"
       "ld1rqw { z26.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       ".inst 0x658ab77b  // bfcvt z27.h, p5/M, z27.s\n"
-      ".inst 0x658ab718  // bfcvt z24.h, p5/M, z24.s\n"
+      "uzp1 z28.h, z28.h, z28.h\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
       "uzp1 z27.h, z27.h, z27.h\n"
       ".inst 0x658ab75a  // bfcvt z26.h, p5/M, z26.s\n"
-      "uzp1 z24.h, z24.h, z24.h\n"
-      "trn1 z27.d, z27.d, z24.d\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "sub x27, x27, #0x4\n"
+      "trn1 z28.d, z28.d, z27.d\n"
       "uzp1 z26.h, z26.h, z26.h\n"
-      ".inst 0x6479e768  // bfmmla z8.s, z27.h, z25.h\n"
+      ".inst 0x6479e788  // bfmmla z8.s, z28.h, z25.h\n"
+      "cmp x27, #0x4\n"
       ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e78c  // bfmmla z12.s, z28.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x647ce76c  // bfmmla z12.s, z27.h, z28.h\n"
-      ".inst 0x647ce754  // bfmmla z20.s, z26.h, z28.h\n"
+      "add x26, x26, #0x10\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
+      ".inst 0x6479e789  // bfmmla z9.s, z28.h, z25.h\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
+      ".inst 0x6478e78d  // bfmmla z13.s, z28.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
+      ".inst 0x6479e78a  // bfmmla z10.s, z28.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
+      ".inst 0x6478e78e  // bfmmla z14.s, z28.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n"
+      ".inst 0x6479e78b  // bfmmla z11.s, z28.h, z25.h\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
-      ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
+      ".inst 0x6478e78f  // bfmmla z15.s, z28.h, z24.h\n"
       ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "bgt 35b\n"
       "36:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z25.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "ld1rqw { z27.s }, p0/Z, [x26]\n"
-      "ld1rqw { z24.s }, p0/Z, [x25]\n"
+      "ld1rqw { z28.s }, p0/Z, [x26]\n"
+      "ld1rqw { z27.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab79c  // bfcvt z28.h, p5/M, z28.s\n"
       "ld1rqw { z26.s }, p0/Z, [x24]\n"
       ".inst 0x658ab77b  // bfcvt z27.h, p5/M, z27.s\n"
-      ".inst 0x658ab718  // bfcvt z24.h, p5/M, z24.s\n"
+      "uzp1 z28.h, z28.h, z28.h\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
       "uzp1 z27.h, z27.h, z27.h\n"
-      "uzp1 z24.h, z24.h, z24.h\n"
       ".inst 0x658ab75a  // bfcvt z26.h, p5/M, z26.s\n"
-      "trn1 z27.d, z27.d, z24.d\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "trn1 z28.d, z28.d, z27.d\n"
       "uzp1 z26.h, z26.h, z26.h\n"
-      ".inst 0x6479e768  // bfmmla z8.s, z27.h, z25.h\n"
-      ".inst 0x647ce76c  // bfmmla z12.s, z27.h, z28.h\n"
+      ".inst 0x6479e788  // bfmmla z8.s, z28.h, z25.h\n"
       ".inst 0x6479e750  // bfmmla z16.s, z26.h, z25.h\n"
+      ".inst 0x6478e78c  // bfmmla z12.s, z28.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x647ce754  // bfmmla z20.s, z26.h, z28.h\n"
+      ".inst 0x6478e754  // bfmmla z20.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
+      ".inst 0x6479e789  // bfmmla z9.s, z28.h, z25.h\n"
       ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
+      ".inst 0x6478e78d  // bfmmla z13.s, z28.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
       ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
+      ".inst 0x6479e78a  // bfmmla z10.s, z28.h, z25.h\n"
       ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
+      ".inst 0x6478e78e  // bfmmla z14.s, z28.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
       ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n"
+      ".inst 0x6479e78b  // bfmmla z11.s, z28.h, z25.h\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
       ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
-      ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
+      ".inst 0x6478e78f  // bfmmla z15.s, z28.h, z24.h\n"
       ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
       "37:"  // Height 3: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -645,24 +643,24 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
       "uzp1 z6.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x9, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 z16.d, z16.d, z20.d\n"
       "uzp1 z17.d, z17.d, z21.d\n"
       "uzp1 z18.d, z18.d, z22.d\n"
       "uzp1 z19.d, z19.d, z23.d\n"
       "tbz %x[flags], #1, 38f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z25.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z25.s }, p5/Z, [x21]\n"
       "ld1rw { z24.s }, p5/Z, [x20]\n"
       "fmin z6.s, p5/M, z6.s, z25.s\n"
       "fmin z12.s, p5/M, z12.s, z25.s\n"
@@ -694,24 +692,24 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x25]\n"
-      "st1w { z17.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x24]\n"
+      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
       "39:"  // Height 3: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 28b\n"
       "b 80f\n"
       "40:"  // Height 4
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -724,15 +722,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cbz x12, 42f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -747,37 +745,37 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "42:"  // Height 4: no bias
       "tbz %x[flags], #0, 43f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x22, x9, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x21]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x20]\n"
-      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x20]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
@@ -804,8 +802,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "45:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -830,110 +828,110 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "ble 49f\n"
       "48:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
-      "ld1rqw { z27.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqw { z25.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqw { z26.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqw { z24.s }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "ld1rqw { z29.s }, p0/Z, [x26]\n"
+      "ld1rqw { z28.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab7bd  // bfcvt z29.h, p5/M, z29.s\n"
+      "ld1rqw { z27.s }, p0/Z, [x24]\n"
+      "ld1rqw { z26.s }, p0/Z, [x23]\n"
+      ".inst 0x658ab79c  // bfcvt z28.h, p5/M, z28.s\n"
       ".inst 0x658ab77b  // bfcvt z27.h, p5/M, z27.s\n"
-      ".inst 0x658ab739  // bfcvt z25.h, p5/M, z25.s\n"
       ".inst 0x658ab75a  // bfcvt z26.h, p5/M, z26.s\n"
-      ".inst 0x658ab718  // bfcvt z24.h, p5/M, z24.s\n"
+      "uzp1 z29.h, z29.h, z29.h\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "uzp1 z28.h, z28.h, z28.h\n"
       "uzp1 z27.h, z27.h, z27.h\n"
-      "uzp1 z25.h, z25.h, z25.h\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "uzp1 z26.h, z26.h, z26.h\n"
-      "uzp1 z24.h, z24.h, z24.h\n"
-      "trn1 z27.d, z27.d, z25.d\n"
-      "trn1 z26.d, z26.d, z24.d\n"
-      ".inst 0x647de768  // bfmmla z8.s, z27.h, z29.h\n"
-      ".inst 0x647de750  // bfmmla z16.s, z26.h, z29.h\n"
+      "trn1 z29.d, z29.d, z28.d\n"
+      ".inst 0x6479e7a8  // bfmmla z8.s, z29.h, z25.h\n"
+      "add x26, x26, #0x10\n"
+      "trn1 z27.d, z27.d, z26.d\n"
+      ".inst 0x6479e770  // bfmmla z16.s, z27.h, z25.h\n"
+      ".inst 0x6478e7ac  // bfmmla z12.s, z29.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x647ce76c  // bfmmla z12.s, z27.h, z28.h\n"
-      ".inst 0x647ce754  // bfmmla z20.s, z26.h, z28.h\n"
+      ".inst 0x6478e774  // bfmmla z20.s, z27.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
-      ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
+      ".inst 0x6479e7a9  // bfmmla z9.s, z29.h, z25.h\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x6479e771  // bfmmla z17.s, z27.h, z25.h\n"
+      ".inst 0x6478e7ad  // bfmmla z13.s, z29.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
-      ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x6478e775  // bfmmla z21.s, z27.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
-      ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
+      ".inst 0x6479e7aa  // bfmmla z10.s, z29.h, z25.h\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6479e772  // bfmmla z18.s, z27.h, z25.h\n"
+      ".inst 0x6478e7ae  // bfmmla z14.s, z29.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
-      ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
+      ".inst 0x6478e776  // bfmmla z22.s, z27.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n"
+      ".inst 0x6479e7ab  // bfmmla z11.s, z29.h, z25.h\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
-      ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
-      ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
-      ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
+      ".inst 0x6479e773  // bfmmla z19.s, z27.h, z25.h\n"
+      ".inst 0x6478e7af  // bfmmla z15.s, z29.h, z24.h\n"
+      ".inst 0x6478e777  // bfmmla z23.s, z27.h, z24.h\n"
       "bgt 48b\n"
       "49:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z29.h }, p5/Z, [x10]\n"
-      "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "ld1rqw { z27.s }, p0/Z, [x26]\n"
-      "ld1rqw { z25.s }, p0/Z, [x25]\n"
-      "ld1rqw { z26.s }, p0/Z, [x24]\n"
-      "ld1rqw { z24.s }, p0/Z, [x23]\n"
+      "ld1rqw { z29.s }, p0/Z, [x26]\n"
+      "ld1rqw { z28.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab7bd  // bfcvt z29.h, p5/M, z29.s\n"
+      "ld1rqw { z27.s }, p0/Z, [x24]\n"
+      "ld1rqw { z26.s }, p0/Z, [x23]\n"
+      ".inst 0x658ab79c  // bfcvt z28.h, p5/M, z28.s\n"
       ".inst 0x658ab77b  // bfcvt z27.h, p5/M, z27.s\n"
-      ".inst 0x658ab739  // bfcvt z25.h, p5/M, z25.s\n"
       ".inst 0x658ab75a  // bfcvt z26.h, p5/M, z26.s\n"
-      ".inst 0x658ab718  // bfcvt z24.h, p5/M, z24.s\n"
+      "uzp1 z29.h, z29.h, z29.h\n"
+      "ld1h { z25.h }, p5/Z, [x10]\n"
+      "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "uzp1 z28.h, z28.h, z28.h\n"
       "uzp1 z27.h, z27.h, z27.h\n"
-      "uzp1 z25.h, z25.h, z25.h\n"
       "uzp1 z26.h, z26.h, z26.h\n"
-      "uzp1 z24.h, z24.h, z24.h\n"
-      "trn1 z27.d, z27.d, z25.d\n"
-      "trn1 z26.d, z26.d, z24.d\n"
-      ".inst 0x647de768  // bfmmla z8.s, z27.h, z29.h\n"
-      ".inst 0x647de750  // bfmmla z16.s, z26.h, z29.h\n"
+      "trn1 z29.d, z29.d, z28.d\n"
+      ".inst 0x6479e7a8  // bfmmla z8.s, z29.h, z25.h\n"
+      "trn1 z27.d, z27.d, z26.d\n"
+      ".inst 0x6479e770  // bfmmla z16.s, z27.h, z25.h\n"
+      ".inst 0x6478e7ac  // bfmmla z12.s, z29.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x647ce76c  // bfmmla z12.s, z27.h, z28.h\n"
-      ".inst 0x647ce754  // bfmmla z20.s, z26.h, z28.h\n"
+      ".inst 0x6478e774  // bfmmla z20.s, z27.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6479e769  // bfmmla z9.s, z27.h, z25.h\n"
-      ".inst 0x6479e751  // bfmmla z17.s, z26.h, z25.h\n"
+      ".inst 0x6479e7a9  // bfmmla z9.s, z29.h, z25.h\n"
+      ".inst 0x6479e771  // bfmmla z17.s, z27.h, z25.h\n"
+      ".inst 0x6478e7ad  // bfmmla z13.s, z29.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6478e76d  // bfmmla z13.s, z27.h, z24.h\n"
-      ".inst 0x6478e755  // bfmmla z21.s, z26.h, z24.h\n"
+      ".inst 0x6478e775  // bfmmla z21.s, z27.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6479e76a  // bfmmla z10.s, z27.h, z25.h\n"
-      ".inst 0x6479e752  // bfmmla z18.s, z26.h, z25.h\n"
+      ".inst 0x6479e7aa  // bfmmla z10.s, z29.h, z25.h\n"
+      ".inst 0x6479e772  // bfmmla z18.s, z27.h, z25.h\n"
+      ".inst 0x6478e7ae  // bfmmla z14.s, z29.h, z24.h\n"
       "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6478e76e  // bfmmla z14.s, z27.h, z24.h\n"
-      ".inst 0x6478e756  // bfmmla z22.s, z26.h, z24.h\n"
+      ".inst 0x6478e776  // bfmmla z22.s, z27.h, z24.h\n"
       "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n"
+      ".inst 0x6479e7ab  // bfmmla z11.s, z29.h, z25.h\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6479e76b  // bfmmla z11.s, z27.h, z25.h\n"
-      ".inst 0x6479e753  // bfmmla z19.s, z26.h, z25.h\n"
-      ".inst 0x6478e76f  // bfmmla z15.s, z27.h, z24.h\n"
-      ".inst 0x6478e757  // bfmmla z23.s, z26.h, z24.h\n"
+      ".inst 0x6479e773  // bfmmla z19.s, z27.h, z25.h\n"
+      ".inst 0x6478e7af  // bfmmla z15.s, z29.h, z24.h\n"
+      ".inst 0x6478e777  // bfmmla z23.s, z27.h, z24.h\n"
       "50:"  // Height 4: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z6.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
       "uzp1 z20.d, z17.d, z21.d\n"
@@ -943,9 +941,9 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
       "tbz %x[flags], #1, 51f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z24.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z24.s }, p5/Z, [x21]\n"
       "ld1rw { z23.s }, p5/Z, [x20]\n"
       "fmin z6.s, p5/M, z6.s, z24.s\n"
       "fmin z12.s, p5/M, z12.s, z24.s\n"
@@ -985,28 +983,28 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
       "52:"  // Height 4: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
       "bgt 41b\n"
       "b 80f\n"
       "53:"  // Height 5
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "54:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1019,15 +1017,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cbz x12, 55f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -1050,46 +1048,46 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "55:"  // Height 5: no bias
       "tbz %x[flags], #0, 56f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x22]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x21]\n"
-      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x20]\n"
-      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
+      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x21]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x20]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z19.d, z24.d, z23.d\n"
       "zip2 z23.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z24.d, z25.d, z28.d\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
@@ -1128,8 +1126,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "58:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 59f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1157,115 +1155,115 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "ble 62f\n"
       "61:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z6.h }, p5/Z, [x10]\n"
-      "ld1h { z5.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
-      "ld1rqw { z4.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqw { z1.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqw { z3.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "ld1rqw { z2.s }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1rqw { z6.s }, p0/Z, [x26]\n"
+      "ld1rqw { z5.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab4c6  // bfcvt z6.h, p5/M, z6.s\n"
+      "ld1rqw { z4.s }, p0/Z, [x24]\n"
+      "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      ".inst 0x658ab4a5  // bfcvt z5.h, p5/M, z5.s\n"
       ".inst 0x658ab484  // bfcvt z4.h, p5/M, z4.s\n"
-      ".inst 0x658ab421  // bfcvt z1.h, p5/M, z1.s\n"
+      "ld1rqw { z2.s }, p0/Z, [x22]\n"
       ".inst 0x658ab463  // bfcvt z3.h, p5/M, z3.s\n"
-      ".inst 0x658ab400  // bfcvt z0.h, p5/M, z0.s\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
+      "ld1h { z1.h }, p5/Z, [x10]\n"
+      "uzp1 z5.h, z5.h, z5.h\n"
       "uzp1 z4.h, z4.h, z4.h\n"
-      ".inst 0x658ab442  // bfcvt z2.h, p5/M, z2.s\n"
-      "uzp1 z1.h, z1.h, z1.h\n"
+      "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "sub x27, x27, #0x4\n"
       "uzp1 z3.h, z3.h, z3.h\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
-      "trn1 z4.d, z4.d, z1.d\n"
+      ".inst 0x658ab442  // bfcvt z2.h, p5/M, z2.s\n"
+      "cmp x27, #0x4\n"
+      "add x26, x26, #0x10\n"
+      "trn1 z6.d, z6.d, z5.d\n"
+      "trn1 z4.d, z4.d, z3.d\n"
+      ".inst 0x6461e4c8  // bfmmla z8.s, z6.h, z1.h\n"
+      "add x25, x25, #0x10\n"
       "uzp1 z2.h, z2.h, z2.h\n"
-      "trn1 z3.d, z3.d, z0.d\n"
-      ".inst 0x6466e488  // bfmmla z8.s, z4.h, z6.h\n"
-      ".inst 0x6466e458  // bfmmla z24.s, z2.h, z6.h\n"
-      ".inst 0x6465e48c  // bfmmla z12.s, z4.h, z5.h\n"
-      ".inst 0x6465e45c  // bfmmla z28.s, z2.h, z5.h\n"
-      ".inst 0x6466e470  // bfmmla z16.s, z3.h, z6.h\n"
+      ".inst 0x6461e490  // bfmmla z16.s, z4.h, z1.h\n"
+      ".inst 0x6461e458  // bfmmla z24.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6465e474  // bfmmla z20.s, z3.h, z5.h\n"
+      ".inst 0x6460e4cc  // bfmmla z12.s, z6.h, z0.h\n"
+      ".inst 0x6460e494  // bfmmla z20.s, z4.h, z0.h\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6460e45c  // bfmmla z28.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6461e489  // bfmmla z9.s, z4.h, z1.h\n"
-      ".inst 0x6461e471  // bfmmla z17.s, z3.h, z1.h\n"
+      ".inst 0x6461e4c9  // bfmmla z9.s, z6.h, z1.h\n"
+      "add x22, x22, #0x10\n"
+      ".inst 0x6461e491  // bfmmla z17.s, z4.h, z1.h\n"
       ".inst 0x6461e459  // bfmmla z25.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6460e48d  // bfmmla z13.s, z4.h, z0.h\n"
-      ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
+      ".inst 0x6460e4cd  // bfmmla z13.s, z6.h, z0.h\n"
+      ".inst 0x6460e495  // bfmmla z21.s, z4.h, z0.h\n"
       ".inst 0x6460e45d  // bfmmla z29.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6461e48a  // bfmmla z10.s, z4.h, z1.h\n"
-      ".inst 0x6461e472  // bfmmla z18.s, z3.h, z1.h\n"
+      ".inst 0x6461e4ca  // bfmmla z10.s, z6.h, z1.h\n"
+      ".inst 0x6461e492  // bfmmla z18.s, z4.h, z1.h\n"
       ".inst 0x6461e45a  // bfmmla z26.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6460e48e  // bfmmla z14.s, z4.h, z0.h\n"
-      ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
+      ".inst 0x6460e4ce  // bfmmla z14.s, z6.h, z0.h\n"
+      ".inst 0x6460e496  // bfmmla z22.s, z4.h, z0.h\n"
       ".inst 0x6460e45e  // bfmmla z30.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
+      ".inst 0x6461e4cb  // bfmmla z11.s, z6.h, z1.h\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6461e48b  // bfmmla z11.s, z4.h, z1.h\n"
-      ".inst 0x6461e473  // bfmmla z19.s, z3.h, z1.h\n"
+      ".inst 0x6461e493  // bfmmla z19.s, z4.h, z1.h\n"
       ".inst 0x6461e45b  // bfmmla z27.s, z2.h, z1.h\n"
-      ".inst 0x6460e48f  // bfmmla z15.s, z4.h, z0.h\n"
-      ".inst 0x6460e477  // bfmmla z23.s, z3.h, z0.h\n"
+      ".inst 0x6460e4cf  // bfmmla z15.s, z6.h, z0.h\n"
+      ".inst 0x6460e497  // bfmmla z23.s, z4.h, z0.h\n"
       ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
       "bgt 61b\n"
       "62:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z6.h }, p5/Z, [x10]\n"
-      "ld1h { z5.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "ld1rqw { z4.s }, p0/Z, [x26]\n"
-      "ld1rqw { z1.s }, p0/Z, [x25]\n"
-      "ld1rqw { z3.s }, p0/Z, [x24]\n"
-      "ld1rqw { z0.s }, p0/Z, [x23]\n"
-      "ld1rqw { z2.s }, p0/Z, [x22]\n"
+      "ld1rqw { z6.s }, p0/Z, [x26]\n"
+      "ld1rqw { z5.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab4c6  // bfcvt z6.h, p5/M, z6.s\n"
+      "ld1rqw { z4.s }, p0/Z, [x24]\n"
+      "ld1rqw { z3.s }, p0/Z, [x23]\n"
+      ".inst 0x658ab4a5  // bfcvt z5.h, p5/M, z5.s\n"
       ".inst 0x658ab484  // bfcvt z4.h, p5/M, z4.s\n"
-      ".inst 0x658ab421  // bfcvt z1.h, p5/M, z1.s\n"
+      "ld1rqw { z2.s }, p0/Z, [x22]\n"
       ".inst 0x658ab463  // bfcvt z3.h, p5/M, z3.s\n"
-      ".inst 0x658ab400  // bfcvt z0.h, p5/M, z0.s\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
+      "ld1h { z1.h }, p5/Z, [x10]\n"
+      "uzp1 z5.h, z5.h, z5.h\n"
       "uzp1 z4.h, z4.h, z4.h\n"
-      "uzp1 z1.h, z1.h, z1.h\n"
-      ".inst 0x658ab442  // bfcvt z2.h, p5/M, z2.s\n"
+      "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
       "uzp1 z3.h, z3.h, z3.h\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
-      "trn1 z4.d, z4.d, z1.d\n"
+      ".inst 0x658ab442  // bfcvt z2.h, p5/M, z2.s\n"
+      "trn1 z6.d, z6.d, z5.d\n"
+      "trn1 z4.d, z4.d, z3.d\n"
+      ".inst 0x6461e4c8  // bfmmla z8.s, z6.h, z1.h\n"
       "uzp1 z2.h, z2.h, z2.h\n"
-      "trn1 z3.d, z3.d, z0.d\n"
-      ".inst 0x6466e488  // bfmmla z8.s, z4.h, z6.h\n"
-      ".inst 0x6465e48c  // bfmmla z12.s, z4.h, z5.h\n"
-      ".inst 0x6466e470  // bfmmla z16.s, z3.h, z6.h\n"
-      ".inst 0x6466e458  // bfmmla z24.s, z2.h, z6.h\n"
+      ".inst 0x6461e490  // bfmmla z16.s, z4.h, z1.h\n"
+      ".inst 0x6461e458  // bfmmla z24.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6465e474  // bfmmla z20.s, z3.h, z5.h\n"
-      ".inst 0x6465e45c  // bfmmla z28.s, z2.h, z5.h\n"
+      ".inst 0x6460e4cc  // bfmmla z12.s, z6.h, z0.h\n"
+      ".inst 0x6460e494  // bfmmla z20.s, z4.h, z0.h\n"
+      ".inst 0x6460e45c  // bfmmla z28.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6461e489  // bfmmla z9.s, z4.h, z1.h\n"
-      ".inst 0x6461e471  // bfmmla z17.s, z3.h, z1.h\n"
+      ".inst 0x6461e4c9  // bfmmla z9.s, z6.h, z1.h\n"
+      ".inst 0x6461e491  // bfmmla z17.s, z4.h, z1.h\n"
       ".inst 0x6461e459  // bfmmla z25.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6460e48d  // bfmmla z13.s, z4.h, z0.h\n"
-      ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
+      ".inst 0x6460e4cd  // bfmmla z13.s, z6.h, z0.h\n"
+      ".inst 0x6460e495  // bfmmla z21.s, z4.h, z0.h\n"
       ".inst 0x6460e45d  // bfmmla z29.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6461e48a  // bfmmla z10.s, z4.h, z1.h\n"
-      ".inst 0x6461e472  // bfmmla z18.s, z3.h, z1.h\n"
+      ".inst 0x6461e4ca  // bfmmla z10.s, z6.h, z1.h\n"
+      ".inst 0x6461e492  // bfmmla z18.s, z4.h, z1.h\n"
       ".inst 0x6461e45a  // bfmmla z26.s, z2.h, z1.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6460e48e  // bfmmla z14.s, z4.h, z0.h\n"
-      ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
+      ".inst 0x6460e4ce  // bfmmla z14.s, z6.h, z0.h\n"
+      ".inst 0x6460e496  // bfmmla z22.s, z4.h, z0.h\n"
       ".inst 0x6460e45e  // bfmmla z30.s, z2.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
+      ".inst 0x6461e4cb  // bfmmla z11.s, z6.h, z1.h\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6461e48b  // bfmmla z11.s, z4.h, z1.h\n"
-      ".inst 0x6461e473  // bfmmla z19.s, z3.h, z1.h\n"
+      ".inst 0x6461e493  // bfmmla z19.s, z4.h, z1.h\n"
       ".inst 0x6461e45b  // bfmmla z27.s, z2.h, z1.h\n"
-      ".inst 0x6460e48f  // bfmmla z15.s, z4.h, z0.h\n"
-      ".inst 0x6460e477  // bfmmla z23.s, z3.h, z0.h\n"
+      ".inst 0x6460e4cf  // bfmmla z15.s, z6.h, z0.h\n"
+      ".inst 0x6460e497  // bfmmla z23.s, z4.h, z0.h\n"
       ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
       "63:"  // Height 5: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -1273,20 +1271,20 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 58b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z6.d, z8.d, z12.d\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
@@ -1298,9 +1296,9 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "uzp1 z26.d, z26.d, z30.d\n"
       "uzp1 z27.d, z27.d, z31.d\n"
       "tbz %x[flags], #1, 64f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z0.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z0.s }, p5/Z, [x21]\n"
       "ld1rw { z23.s }, p5/Z, [x20]\n"
       "fmin z6.s, p5/M, z6.s, z0.s\n"
       "fmin z12.s, p5/M, z12.s, z0.s\n"
@@ -1348,22 +1346,22 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x23]\n"
-      "st1w { z25.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x22]\n"
+      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
       "65:"  // Height 5: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1371,13 +1369,12 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "b 80f\n"
       "66:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
-      "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n"
+      "mov x12, %x[bias]\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "67:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1390,15 +1387,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "cbz x12, 68f\n"
       "ld1w { z8.s }, p5/Z, [x12]\n"
       "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n"
-      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
-      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
-      "addvl x12, x12, #4\n"
       "zip2 z12.d, z8.d, z8.d\n"
       "zip1 z8.d, z8.d, z8.d\n"
+      "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n"
+      "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n"
       "zip2 z13.d, z9.d, z9.d\n"
       "zip1 z9.d, z9.d, z9.d\n"
       "zip2 z14.d, z10.d, z10.d\n"
       "zip1 z10.d, z10.d, z10.d\n"
+      "addvl x12, x12, #4\n"
       "zip2 z15.d, z11.d, z11.d\n"
       "zip1 z11.d, z11.d, z11.d\n"
       "mov z16.d, z8.d\n"
@@ -1421,54 +1418,54 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "68:"  // Height 6: no bias
       "tbz %x[flags], #0, 69f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "ld1w { z17.s }, p4/Z, [x9]\n"
       "add x22, x23, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x24]\n"
-      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x9, #2, MUL VL]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
+      "ld1w { z12.s }, p4/Z, [x24]\n"
+      "zip1 z8.d, z17.d, z12.d\n"
+      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
+      "zip2 z12.d, z17.d, z12.d\n"
+      "zip1 z9.d, z18.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x23]\n"
+      "zip2 z13.d, z18.d, z13.d\n"
+      "zip1 z10.d, z20.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip2 z14.d, z20.d, z14.d\n"
+      "zip1 z11.d, z16.d, z15.d\n"
       "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
       "ld1w { z20.s }, p4/Z, [x22]\n"
-      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x21]\n"
-      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
       "zip2 z20.d, z17.d, z20.d\n"
-      "ld1w { z28.s }, p4/Z, [x20]\n"
-      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x21]\n"
       "zip2 z21.d, z18.d, z21.d\n"
-      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z28.s }, p4/Z, [x20]\n"
       "zip2 z23.d, z24.d, z23.d\n"
       "zip1 z24.d, z25.d, z28.d\n"
+      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
+      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip2 z29.d, z26.d, z29.d\n"
       "zip1 z26.d, z27.d, z30.d\n"
       "zip2 z30.d, z27.d, z30.d\n"
@@ -1504,8 +1501,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "71:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 72f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1536,146 +1533,146 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "ble 75f\n"
       "74:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z7.h }, p5/Z, [x10]\n"
-      "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x4\n"
-      "cmp x27, #0x4\n"
-      "ld1rqw { z5.s }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqw { z4.s }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqw { z3.s }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqw { z1.s }, p0/Z, [x23]\n"
-      "ld1rqw { z2.s }, p0/Z, [x22]\n"
-      "ld1rqw { z0.s }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1rqw { z7.s }, p0/Z, [x26]\n"
+      "ld1rqw { z6.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab4e7  // bfcvt z7.h, p5/M, z7.s\n"
+      "ld1rqw { z5.s }, p0/Z, [x24]\n"
+      "ld1rqw { z4.s }, p0/Z, [x23]\n"
+      ".inst 0x658ab4c6  // bfcvt z6.h, p5/M, z6.s\n"
       ".inst 0x658ab4a5  // bfcvt z5.h, p5/M, z5.s\n"
+      "ld1rqw { z3.s }, p0/Z, [x22]\n"
+      "ld1rqw { z2.s }, p0/Z, [x21]\n"
       ".inst 0x658ab484  // bfcvt z4.h, p5/M, z4.s\n"
-      "add x21, x21, #0x10\n"
       ".inst 0x658ab463  // bfcvt z3.h, p5/M, z3.s\n"
-      ".inst 0x658ab421  // bfcvt z1.h, p5/M, z1.s\n"
       ".inst 0x658ab442  // bfcvt z2.h, p5/M, z2.s\n"
-      ".inst 0x658ab400  // bfcvt z0.h, p5/M, z0.s\n"
+      "uzp1 z7.h, z7.h, z7.h\n"
+      "ld1h { z1.h }, p5/Z, [x10]\n"
+      "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
       "uzp1 z5.h, z5.h, z5.h\n"
+      "sub x27, x27, #0x4\n"
+      "cmp x27, #0x4\n"
       "uzp1 z4.h, z4.h, z4.h\n"
       "uzp1 z3.h, z3.h, z3.h\n"
-      "uzp1 z1.h, z1.h, z1.h\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "uzp1 z2.h, z2.h, z2.h\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
+      "trn1 z7.d, z7.d, z6.d\n"
+      ".inst 0x6461e4e8  // bfmmla z8.s, z7.h, z1.h\n"
+      "add x24, x24, #0x10\n"
       "trn1 z5.d, z5.d, z4.d\n"
-      "trn1 z3.d, z3.d, z1.d\n"
-      "trn1 z2.d, z2.d, z0.d\n"
-      ".inst 0x6467e4a8  // bfmmla z8.s, z5.h, z7.h\n"
-      ".inst 0x6467e470  // bfmmla z16.s, z3.h, z7.h\n"
-      ".inst 0x6466e4ac  // bfmmla z12.s, z5.h, z6.h\n"
-      ".inst 0x6466e474  // bfmmla z20.s, z3.h, z6.h\n"
-      ".inst 0x6467e458  // bfmmla z24.s, z2.h, z7.h\n"
+      "trn1 z3.d, z3.d, z2.d\n"
+      ".inst 0x6461e4b0  // bfmmla z16.s, z5.h, z1.h\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x6461e478  // bfmmla z24.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ec  // bfmmla z12.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6466e45c  // bfmmla z28.s, z2.h, z6.h\n"
+      "add x22, x22, #0x10\n"
+      ".inst 0x6460e4b4  // bfmmla z20.s, z5.h, z0.h\n"
+      ".inst 0x6460e47c  // bfmmla z28.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6461e4a9  // bfmmla z9.s, z5.h, z1.h\n"
-      ".inst 0x6461e471  // bfmmla z17.s, z3.h, z1.h\n"
-      ".inst 0x6461e459  // bfmmla z25.s, z2.h, z1.h\n"
+      "add x21, x21, #0x10\n"
+      ".inst 0x6461e4e9  // bfmmla z9.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b1  // bfmmla z17.s, z5.h, z1.h\n"
+      ".inst 0x6461e479  // bfmmla z25.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ed  // bfmmla z13.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6460e4ad  // bfmmla z13.s, z5.h, z0.h\n"
-      ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
-      ".inst 0x6460e45d  // bfmmla z29.s, z2.h, z0.h\n"
+      ".inst 0x6460e4b5  // bfmmla z21.s, z5.h, z0.h\n"
+      ".inst 0x6460e47d  // bfmmla z29.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6461e4aa  // bfmmla z10.s, z5.h, z1.h\n"
-      ".inst 0x6461e472  // bfmmla z18.s, z3.h, z1.h\n"
-      ".inst 0x6461e45a  // bfmmla z26.s, z2.h, z1.h\n"
+      ".inst 0x6461e4ea  // bfmmla z10.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b2  // bfmmla z18.s, z5.h, z1.h\n"
+      ".inst 0x6461e47a  // bfmmla z26.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ee  // bfmmla z14.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6460e4ae  // bfmmla z14.s, z5.h, z0.h\n"
-      ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
-      ".inst 0x6460e45e  // bfmmla z30.s, z2.h, z0.h\n"
+      ".inst 0x6460e4b6  // bfmmla z22.s, z5.h, z0.h\n"
+      ".inst 0x6460e47e  // bfmmla z30.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
-      ".inst 0x6461e473  // bfmmla z19.s, z3.h, z1.h\n"
-      ".inst 0x6461e45b  // bfmmla z27.s, z2.h, z1.h\n"
-      ".inst 0x6460e4af  // bfmmla z15.s, z5.h, z0.h\n"
-      ".inst 0x6460e477  // bfmmla z23.s, z3.h, z0.h\n"
-      ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
+      ".inst 0x6461e4eb  // bfmmla z11.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b3  // bfmmla z19.s, z5.h, z1.h\n"
+      ".inst 0x6461e47b  // bfmmla z27.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b7  // bfmmla z23.s, z5.h, z0.h\n"
+      ".inst 0x6460e47f  // bfmmla z31.s, z3.h, z0.h\n"
       "bgt 74b\n"
       "75:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.s, XZR, x27\n"
-      "ld1h { z7.h }, p5/Z, [x10]\n"
-      "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n"
-      "ld1rqw { z5.s }, p0/Z, [x26]\n"
-      "ld1rqw { z4.s }, p0/Z, [x25]\n"
-      "ld1rqw { z3.s }, p0/Z, [x24]\n"
-      "ld1rqw { z1.s }, p0/Z, [x23]\n"
-      "ld1rqw { z2.s }, p0/Z, [x22]\n"
-      "ld1rqw { z0.s }, p0/Z, [x21]\n"
+      "ld1rqw { z7.s }, p0/Z, [x26]\n"
+      "ld1rqw { z6.s }, p0/Z, [x25]\n"
+      ".inst 0x658ab4e7  // bfcvt z7.h, p5/M, z7.s\n"
+      "ld1rqw { z5.s }, p0/Z, [x24]\n"
+      "ld1rqw { z4.s }, p0/Z, [x23]\n"
+      ".inst 0x658ab4c6  // bfcvt z6.h, p5/M, z6.s\n"
       ".inst 0x658ab4a5  // bfcvt z5.h, p5/M, z5.s\n"
+      "ld1rqw { z3.s }, p0/Z, [x22]\n"
+      "ld1rqw { z2.s }, p0/Z, [x21]\n"
       ".inst 0x658ab484  // bfcvt z4.h, p5/M, z4.s\n"
       ".inst 0x658ab463  // bfcvt z3.h, p5/M, z3.s\n"
-      ".inst 0x658ab421  // bfcvt z1.h, p5/M, z1.s\n"
       ".inst 0x658ab442  // bfcvt z2.h, p5/M, z2.s\n"
-      ".inst 0x658ab400  // bfcvt z0.h, p5/M, z0.s\n"
+      "uzp1 z7.h, z7.h, z7.h\n"
+      "ld1h { z1.h }, p5/Z, [x10]\n"
+      "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n"
+      "uzp1 z6.h, z6.h, z6.h\n"
       "uzp1 z5.h, z5.h, z5.h\n"
       "uzp1 z4.h, z4.h, z4.h\n"
       "uzp1 z3.h, z3.h, z3.h\n"
-      "uzp1 z1.h, z1.h, z1.h\n"
       "uzp1 z2.h, z2.h, z2.h\n"
-      "uzp1 z0.h, z0.h, z0.h\n"
+      "trn1 z7.d, z7.d, z6.d\n"
+      ".inst 0x6461e4e8  // bfmmla z8.s, z7.h, z1.h\n"
       "trn1 z5.d, z5.d, z4.d\n"
-      "trn1 z3.d, z3.d, z1.d\n"
-      "trn1 z2.d, z2.d, z0.d\n"
-      ".inst 0x6467e4a8  // bfmmla z8.s, z5.h, z7.h\n"
-      ".inst 0x6467e470  // bfmmla z16.s, z3.h, z7.h\n"
-      ".inst 0x6466e4ac  // bfmmla z12.s, z5.h, z6.h\n"
-      ".inst 0x6466e474  // bfmmla z20.s, z3.h, z6.h\n"
-      ".inst 0x6467e458  // bfmmla z24.s, z2.h, z7.h\n"
+      "trn1 z3.d, z3.d, z2.d\n"
+      ".inst 0x6461e4b0  // bfmmla z16.s, z5.h, z1.h\n"
+      ".inst 0x6461e478  // bfmmla z24.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ec  // bfmmla z12.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x6466e45c  // bfmmla z28.s, z2.h, z6.h\n"
+      ".inst 0x6460e4b4  // bfmmla z20.s, z5.h, z0.h\n"
+      ".inst 0x6460e47c  // bfmmla z28.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x6461e4a9  // bfmmla z9.s, z5.h, z1.h\n"
-      ".inst 0x6461e471  // bfmmla z17.s, z3.h, z1.h\n"
-      ".inst 0x6461e459  // bfmmla z25.s, z2.h, z1.h\n"
+      ".inst 0x6461e4e9  // bfmmla z9.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b1  // bfmmla z17.s, z5.h, z1.h\n"
+      ".inst 0x6461e479  // bfmmla z25.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ed  // bfmmla z13.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x6460e4ad  // bfmmla z13.s, z5.h, z0.h\n"
-      ".inst 0x6460e475  // bfmmla z21.s, z3.h, z0.h\n"
-      ".inst 0x6460e45d  // bfmmla z29.s, z2.h, z0.h\n"
+      ".inst 0x6460e4b5  // bfmmla z21.s, z5.h, z0.h\n"
+      ".inst 0x6460e47d  // bfmmla z29.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x6461e4aa  // bfmmla z10.s, z5.h, z1.h\n"
-      ".inst 0x6461e472  // bfmmla z18.s, z3.h, z1.h\n"
-      ".inst 0x6461e45a  // bfmmla z26.s, z2.h, z1.h\n"
+      ".inst 0x6461e4ea  // bfmmla z10.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b2  // bfmmla z18.s, z5.h, z1.h\n"
+      ".inst 0x6461e47a  // bfmmla z26.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ee  // bfmmla z14.s, z7.h, z0.h\n"
       "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x6460e4ae  // bfmmla z14.s, z5.h, z0.h\n"
-      ".inst 0x6460e476  // bfmmla z22.s, z3.h, z0.h\n"
-      ".inst 0x6460e45e  // bfmmla z30.s, z2.h, z0.h\n"
+      ".inst 0x6460e4b6  // bfmmla z22.s, z5.h, z0.h\n"
+      ".inst 0x6460e47e  // bfmmla z30.s, z3.h, z0.h\n"
       "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x6461e4ab  // bfmmla z11.s, z5.h, z1.h\n"
-      ".inst 0x6461e473  // bfmmla z19.s, z3.h, z1.h\n"
-      ".inst 0x6461e45b  // bfmmla z27.s, z2.h, z1.h\n"
-      ".inst 0x6460e4af  // bfmmla z15.s, z5.h, z0.h\n"
-      ".inst 0x6460e477  // bfmmla z23.s, z3.h, z0.h\n"
-      ".inst 0x6460e45f  // bfmmla z31.s, z2.h, z0.h\n"
+      ".inst 0x6461e4eb  // bfmmla z11.s, z7.h, z1.h\n"
+      ".inst 0x6461e4b3  // bfmmla z19.s, z5.h, z1.h\n"
+      ".inst 0x6461e47b  // bfmmla z27.s, z3.h, z1.h\n"
+      ".inst 0x6460e4ef  // bfmmla z15.s, z7.h, z0.h\n"
+      ".inst 0x6460e4b7  // bfmmla z23.s, z5.h, z0.h\n"
+      ".inst 0x6460e47f  // bfmmla z31.s, z3.h, z0.h\n"
       "76:"  // Height 6: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 71b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x25, x9, x20, LSL #2\n"
+      "add x24, x25, x20, LSL #2\n"
       "uzp1 z6.d, z8.d, z12.d\n"
+      "add x23, x24, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x26, x9, x20, LSL #2\n"
-      "add x25, x26, x20, LSL #2\n"
-      "add x24, x25, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x23, x24, x20, LSL #2\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
@@ -1691,9 +1688,9 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "uzp1 z30.d, z27.d, z31.d\n"
       "uzp2 z27.d, z27.d, z31.d\n"
       "tbz %x[flags], #1, 77f\n"
-      "add x21, %x[args_ptr], %[offset_max]\n"
+      "add x20, %x[args_ptr], %[offset_max]\n"
+      "ld1rw { z1.s }, p5/Z, [x20]\n"
       "add x20, %x[args_ptr], %[offset_min]\n"
-      "ld1rw { z1.s }, p5/Z, [x21]\n"
       "ld1rw { z0.s }, p5/Z, [x20]\n"
       "fmin z6.s, p5/M, z6.s, z1.s\n"
       "fmin z12.s, p5/M, z12.s, z1.s\n"
@@ -1749,26 +1746,26 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x26]\n"
-      "st1w { z9.s }, p3, [x26, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x26, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x26, #3, MUL VL]\n"
-      "st1w { z15.s }, p4, [x25]\n"
-      "st1w { z20.s }, p3, [x25, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x25, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x24]\n"
-      "st1w { z17.s }, p3, [x24, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x24, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x24, #3, MUL VL]\n"
-      "st1w { z23.s }, p4, [x23]\n"
-      "st1w { z28.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z29.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z30.s }, p1, [x23, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x22]\n"
-      "st1w { z25.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z8.s }, p4, [x25]\n"
+      "st1w { z9.s }, p3, [x25, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x25, #2, MUL VL]\n"
+      "st1w { z11.s }, p1, [x25, #3, MUL VL]\n"
+      "st1w { z15.s }, p4, [x24]\n"
+      "st1w { z20.s }, p3, [x24, #1, MUL VL]\n"
+      "st1w { z21.s }, p2, [x24, #2, MUL VL]\n"
+      "st1w { z22.s }, p1, [x24, #3, MUL VL]\n"
+      "st1w { z16.s }, p4, [x23]\n"
+      "st1w { z17.s }, p3, [x23, #1, MUL VL]\n"
+      "st1w { z18.s }, p2, [x23, #2, MUL VL]\n"
+      "st1w { z19.s }, p1, [x23, #3, MUL VL]\n"
+      "st1w { z23.s }, p4, [x22]\n"
+      "st1w { z28.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z29.s }, p2, [x22, #2, MUL VL]\n"
+      "st1w { z30.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z24.s }, p4, [x21]\n"
+      "st1w { z25.s }, p3, [x21, #1, MUL VL]\n"
+      "st1w { z26.s }, p2, [x21, #2, MUL VL]\n"
+      "st1w { z27.s }, p1, [x21, #3, MUL VL]\n"
       "78:"  // Height 6: Writeback done
       "decw x11, ALL, MUL #4\n"
       "cmp x11, XZR\n"
@@ -1785,8 +1782,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL.hpp
index a8e82516a1..dc008866a1 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 4, 4, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 4, 4, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL/generic.cpp
index 7de5e09bd5..b06e0bd3c3 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_dot_4x4VL/generic.cpp
@@ -45,18 +45,18 @@ void sve_hybrid_s8qa_dot_4x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -91,20 +91,20 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "3:"  // Height 1: setup done
       "mov x26, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -120,41 +120,41 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ble 9f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z21.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z23.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x24, x24, #0x10\n"
-      "sdot z16.s, z21.b, z0.b[0]\n"
-      "ld1b { z21.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      "sdot z17.s, z26.b, z0.b[0]\n"
-      "sdot z18.s, z25.b, z0.b[0]\n"
-      "sdot z19.s, z24.b, z0.b[0]\n"
+      "ld1b { z20.b }, p2/Z, [x28]\n"
+      "sdot z16.s, z20.b, z0.b[0]\n"
+      "ld1b { z21.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "sdot z17.s, z21.b, z0.b[0]\n"
+      "sdot z18.s, z20.b, z0.b[0]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "sdot z19.s, z20.b, z0.b[0]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #4, MUL VL]\n"
       "sdot z16.s, z20.b, z0.b[1]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      "sdot z17.s, z23.b, z0.b[1]\n"
-      "ld1b { z23.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      "sdot z18.s, z22.b, z0.b[1]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      "sdot z19.s, z21.b, z0.b[1]\n"
+      "ld1b { z21.b }, p2/Z, [x28, #5, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #6, MUL VL]\n"
+      "sdot z17.s, z21.b, z0.b[1]\n"
+      "sdot z18.s, z20.b, z0.b[1]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #7, MUL VL]\n"
+      "addvl x28, x28, #16\n"
+      "sdot z19.s, z20.b, z0.b[1]\n"
+      "ld1b { z22.b }, p2/Z, [x28, #-8, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      "ld1b { z21.b }, p2/Z, [x28, #-6, MUL VL]\n"
+      "sdot z16.s, z22.b, z0.b[2]\n"
+      "sdot z17.s, z20.b, z0.b[2]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      "sdot z18.s, z21.b, z0.b[2]\n"
+      "sdot z19.s, z20.b, z0.b[2]\n"
+      "ld1b { z22.b }, p2/Z, [x28, #-4, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "ld1b { z21.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      "sdot z16.s, z20.b, z0.b[2]\n"
+      "sdot z16.s, z22.b, z0.b[3]\n"
+      "sdot z17.s, z20.b, z0.b[3]\n"
       "ld1b { z20.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      "sdot z17.s, z26.b, z0.b[2]\n"
-      "sdot z18.s, z25.b, z0.b[2]\n"
-      "sdot z19.s, z24.b, z0.b[2]\n"
-      "sdot z16.s, z23.b, z0.b[3]\n"
-      "sdot z17.s, z22.b, z0.b[3]\n"
       "sdot z18.s, z21.b, z0.b[3]\n"
       "sdot z19.s, z20.b, z0.b[3]\n"
+      "add x24, x24, #0x10\n"
       "tbnz %x[flags], #31, 8f\n"
       "sdot z11.s, z0.b, z15.b\n"
       "8:"  // Height 1: Multiply loop: unique 1: skip row sum
@@ -163,49 +163,49 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "bgt 7b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1rqb { z0.b }, p0/Z, [x24]\n"
+      "ld1b { z22.b }, p2/Z, [x28]\n"
       "subs x25, x25, #0x4\n"
+      "ld1b { z20.b }, p2/Z, [x28, #1, MUL VL]\n"
       "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "sdot z16.s, z22.b, z0.b[0]\n"
+      "sdot z17.s, z20.b, z0.b[0]\n"
       "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "sdot z16.s, z23.b, z0.b[0]\n"
-      "sdot z17.s, z22.b, z0.b[0]\n"
       "sdot z18.s, z21.b, z0.b[0]\n"
       "sdot z19.s, z20.b, z0.b[0]\n"
+      "addvl x28, x28, #4\n"
       "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
+      "ld1b { z20.b }, p2/Z, [x28]\n"
       "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "sdot z16.s, z20.b, z0.b[1]\n"
       "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z23.b, z0.b[1]\n"
       "sdot z17.s, z22.b, z0.b[1]\n"
       "sdot z18.s, z21.b, z0.b[1]\n"
       "sdot z19.s, z20.b, z0.b[1]\n"
+      "addvl x28, x28, #4\n"
       "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
+      "ld1b { z20.b }, p2/Z, [x28]\n"
       "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "sdot z16.s, z20.b, z0.b[2]\n"
       "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z23.b, z0.b[2]\n"
       "sdot z17.s, z22.b, z0.b[2]\n"
       "sdot z18.s, z21.b, z0.b[2]\n"
       "sdot z19.s, z20.b, z0.b[2]\n"
+      "addvl x28, x28, #4\n"
       "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z21.b }, p2/Z, [x28]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "sdot z16.s, z21.b, z0.b[3]\n"
+      "sdot z17.s, z20.b, z0.b[3]\n"
       "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z23.b, z0.b[3]\n"
-      "sdot z17.s, z22.b, z0.b[3]\n"
       "sdot z18.s, z21.b, z0.b[3]\n"
       "sdot z19.s, z20.b, z0.b[3]\n"
+      "addvl x28, x28, #4\n"
       "10:"  // Height 1: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 11f\n"
       "sdot z11.s, z0.b, z15.b\n"
@@ -215,35 +215,35 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "cmp x26, x20\n"
       "bne 4b\n"
       "tbnz %x[flags], #31, 12f\n"
-      "mov x21, #0x4\n"
+      "mov x20, #0x4\n"
+      "whilelt p0.s, XZR, x20\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z20.s, p2/M, z20.s\n"
       "saddv d11, p0, z11.s\n"
       "mov z11.s, z11.s[0]\n"
+      "neg z20.s, p2/M, z20.s\n"
       "mul z11.s, p2/M, z11.s, z20.s\n"
       "12:"  // Height 1: skip row sum fixup
       "add z16.s, z16.s, z11.s\n"
       "add z17.s, z17.s, z11.s\n"
       "ld1w { z23.s }, p2/Z, [x10]\n"
-      "ld1w { z20.s }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x10, #1, MUL VL]\n"
       "add z18.s, z18.s, z11.s\n"
       "add z19.s, z19.s, z11.s\n"
-      "ld1w { z22.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1w { z21.s }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x10, #3, MUL VL]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
       "add z16.s, z16.s, z23.s\n"
-      "add z17.s, z17.s, z20.s\n"
+      "add z17.s, z17.s, z22.s\n"
+      "add z18.s, z18.s, z21.s\n"
+      "add z19.s, z19.s, z20.s\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "addvl x10, x10, #4\n"
-      "add z18.s, z18.s, z22.s\n"
-      "add z19.s, z19.s, z21.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
       ".inst 0x04b47610  // sqrdmulh z16.s, z16.s, z20.s\n"
       ".inst 0x04b47631  // sqrdmulh z17.s, z17.s, z20.s\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x04b47652  // sqrdmulh z18.s, z18.s, z20.s\n"
       ".inst 0x04b47673  // sqrdmulh z19.s, z19.s, z20.s\n"
       "tbz %x[flags], #5, 13f\n"
@@ -261,19 +261,19 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "sqadd z19.s, z19.s, z20.s\n"
       "13:"  // Height 1: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z20.s }, p2/Z, [x20]\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z22.s }, p2/Z, [x20]\n"
+      "add z16.s, z16.s, z20.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z20.s\n"
+      "add z18.s, z18.s, z20.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
       "ld1rw { z21.s }, p2/Z, [x20]\n"
-      "add z16.s, z16.s, z22.s\n"
+      "add z19.s, z19.s, z20.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z22.s\n"
-      "add z18.s, z18.s, z22.s\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z22.s\n"
       "smin z16.s, p2/M, z16.s, z21.s\n"
       "smin z17.s, p2/M, z17.s, z21.s\n"
       "smin z18.s, p2/M, z18.s, z21.s\n"
@@ -281,8 +281,8 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "smax z16.s, p2/M, z16.s, z20.s\n"
       "smax z17.s, p2/M, z17.s, z20.s\n"
       "smax z18.s, p2/M, z18.s, z20.s\n"
-      "smax z19.s, p2/M, z19.s, z20.s\n"
       "uzp1 z16.h, z16.h, z17.h\n"
+      "smax z19.s, p2/M, z19.s, z20.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
       "uzp1 z16.b, z16.b, z17.b\n"
       "st1b { z16.b }, p1, [x27]\n"
@@ -300,24 +300,24 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "mov z15.b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "16:"  // Height 2: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "17:"  // Height 2: setup done
       "mov x26, #0x0\n"
       "18:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 19f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -336,45 +336,45 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ble 23f\n"
       "21:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z25.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "sdot z16.s, z24.b, z0.b[0]\n"
+      "sdot z20.s, z24.b, z1.b[0]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "sdot z17.s, z26.b, z0.b[0]\n"
+      "sdot z21.s, z26.b, z1.b[0]\n"
+      "sdot z18.s, z24.b, z0.b[0]\n"
+      "sdot z22.s, z24.b, z1.b[0]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z27.b }, p2/Z, [x28, #5, MUL VL]\n"
+      "sdot z19.s, z25.b, z0.b[0]\n"
+      "sdot z23.s, z25.b, z1.b[0]\n"
       "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "sdot z16.s, z25.b, z0.b[0]\n"
-      "sdot z20.s, z25.b, z1.b[0]\n"
       "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      "sdot z17.s, z30.b, z0.b[0]\n"
-      "sdot z21.s, z30.b, z1.b[0]\n"
-      "sdot z18.s, z29.b, z0.b[0]\n"
-      "sdot z22.s, z29.b, z1.b[0]\n"
-      "sdot z19.s, z28.b, z0.b[0]\n"
-      "sdot z23.s, z28.b, z1.b[0]\n"
       "sdot z16.s, z24.b, z0.b[1]\n"
       "sdot z20.s, z24.b, z1.b[1]\n"
       "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
       "sdot z17.s, z27.b, z0.b[1]\n"
       "sdot z21.s, z27.b, z1.b[1]\n"
+      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
       "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
       "sdot z18.s, z26.b, z0.b[1]\n"
       "sdot z22.s, z26.b, z1.b[1]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
       "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "sdot z19.s, z25.b, z0.b[1]\n"
       "sdot z23.s, z25.b, z1.b[1]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
       "sdot z16.s, z24.b, z0.b[2]\n"
       "sdot z20.s, z24.b, z1.b[2]\n"
       "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
+      "add x23, x23, #0x10\n"
       "sdot z17.s, z30.b, z0.b[2]\n"
       "sdot z21.s, z30.b, z1.b[2]\n"
       "sdot z18.s, z29.b, z0.b[2]\n"
@@ -398,34 +398,34 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "bgt 21b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z27.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1rqb { z0.b }, p0/Z, [x24]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x25, x25, #0x4\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "sdot z16.s, z24.b, z0.b[0]\n"
+      "sdot z20.s, z24.b, z1.b[0]\n"
       "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1rqb { z1.b }, p0/Z, [x23]\n"
-      "sdot z16.s, z27.b, z0.b[0]\n"
-      "sdot z20.s, z27.b, z1.b[0]\n"
       "sdot z17.s, z26.b, z0.b[0]\n"
       "sdot z21.s, z26.b, z1.b[0]\n"
       "sdot z18.s, z25.b, z0.b[0]\n"
       "sdot z22.s, z25.b, z1.b[0]\n"
+      "addvl x28, x28, #4\n"
       "sdot z19.s, z24.b, z0.b[0]\n"
       "sdot z23.s, z24.b, z1.b[0]\n"
       "ble 24f\n"
       "ld1b { z27.b }, p2/Z, [x28]\n"
       "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "sdot z16.s, z27.b, z0.b[1]\n"
       "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z27.b, z0.b[1]\n"
       "sdot z20.s, z27.b, z1.b[1]\n"
       "sdot z17.s, z26.b, z0.b[1]\n"
       "sdot z21.s, z26.b, z1.b[1]\n"
       "sdot z18.s, z25.b, z0.b[1]\n"
+      "addvl x28, x28, #4\n"
       "sdot z22.s, z25.b, z1.b[1]\n"
       "sdot z19.s, z24.b, z0.b[1]\n"
       "sdot z23.s, z24.b, z1.b[1]\n"
@@ -433,29 +433,29 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ld1b { z27.b }, p2/Z, [x28]\n"
       "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "sdot z16.s, z27.b, z0.b[2]\n"
       "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z27.b, z0.b[2]\n"
       "sdot z20.s, z27.b, z1.b[2]\n"
       "sdot z17.s, z26.b, z0.b[2]\n"
       "sdot z21.s, z26.b, z1.b[2]\n"
       "sdot z18.s, z25.b, z0.b[2]\n"
+      "addvl x28, x28, #4\n"
       "sdot z22.s, z25.b, z1.b[2]\n"
       "sdot z19.s, z24.b, z0.b[2]\n"
       "sdot z23.s, z24.b, z1.b[2]\n"
       "ble 24f\n"
-      "ld1b { z27.b }, p2/Z, [x28]\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
       "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "sdot z16.s, z24.b, z0.b[3]\n"
+      "sdot z20.s, z24.b, z1.b[3]\n"
       "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z27.b, z0.b[3]\n"
-      "sdot z20.s, z27.b, z1.b[3]\n"
       "sdot z17.s, z26.b, z0.b[3]\n"
       "sdot z21.s, z26.b, z1.b[3]\n"
       "sdot z18.s, z25.b, z0.b[3]\n"
       "sdot z22.s, z25.b, z1.b[3]\n"
+      "addvl x28, x28, #4\n"
       "sdot z19.s, z24.b, z0.b[3]\n"
       "sdot z23.s, z24.b, z1.b[3]\n"
       "24:"  // Height 2: Multiply loop: multiply skip
@@ -468,18 +468,18 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "cmp x26, x20\n"
       "bne 18b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
+      "add x23, x27, x20\n"
       "tbnz %x[flags], #31, 26f\n"
-      "mov x21, #0x4\n"
+      "mov x20, #0x4\n"
+      "whilelt p0.s, XZR, x20\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z24.s, p2/M, z24.s\n"
       "saddv d11, p0, z11.s\n"
-      "saddv d12, p0, z12.s\n"
       "mov z11.s, z11.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z24.s\n"
+      "saddv d12, p0, z12.s\n"
+      "neg z24.s, p2/M, z24.s\n"
       "mov z12.s, z12.s[0]\n"
+      "mul z11.s, p2/M, z11.s, z24.s\n"
       "mul z12.s, p2/M, z12.s, z24.s\n"
       "26:"  // Height 2: skip row sum fixup
       "add z16.s, z16.s, z11.s\n"
@@ -518,24 +518,24 @@ void sve_hybrid_s8qa_dot_4x4VL (
       ".inst 0x04b876f7  // sqrdmulh z23.s, z23.s, z24.s\n"
       "tbz %x[flags], #5, 27f\n"
       "and z24.d, z16.d, z0.d\n"
+      "asr z24.s, z24.s, #0x1f\n"
+      "sqadd z16.s, z16.s, z24.s\n"
       "and z30.d, z17.d, z0.d\n"
       "and z29.d, z18.d, z0.d\n"
       "and z28.d, z19.d, z0.d\n"
       "and z27.d, z20.d, z0.d\n"
       "and z26.d, z21.d, z0.d\n"
-      "asr z24.s, z24.s, #0x1f\n"
       "and z25.d, z22.d, z0.d\n"
+      "and z24.d, z23.d, z0.d\n"
       "asr z30.s, z30.s, #0x1f\n"
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "asr z27.s, z27.s, #0x1f\n"
-      "sqadd z16.s, z16.s, z24.s\n"
-      "and z24.d, z23.d, z0.d\n"
       "asr z26.s, z26.s, #0x1f\n"
       "asr z25.s, z25.s, #0x1f\n"
+      "asr z24.s, z24.s, #0x1f\n"
       "sqadd z17.s, z17.s, z30.s\n"
       "sqadd z18.s, z18.s, z29.s\n"
-      "asr z24.s, z24.s, #0x1f\n"
       "sqadd z19.s, z19.s, z28.s\n"
       "sqadd z20.s, z20.s, z27.s\n"
       "sqadd z21.s, z21.s, z26.s\n"
@@ -543,27 +543,27 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "sqadd z23.s, z23.s, z24.s\n"
       "27:"  // Height 2: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z24.s }, p2/Z, [x20]\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z26.s }, p2/Z, [x20]\n"
+      "add z16.s, z16.s, z24.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z24.s\n"
+      "add z18.s, z18.s, z24.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
+      "add z19.s, z19.s, z24.s\n"
+      "add z20.s, z20.s, z24.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z26.s\n"
+      "add z21.s, z21.s, z24.s\n"
+      "add z22.s, z22.s, z24.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z17.s, z17.s, z26.s\n"
-      "add z18.s, z18.s, z26.s\n"
       "ld1rw { z25.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z26.s\n"
-      "add z20.s, z20.s, z26.s\n"
+      "add z23.s, z23.s, z24.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z21.s, z21.s, z26.s\n"
-      "add z22.s, z22.s, z26.s\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z26.s\n"
       "smin z16.s, p2/M, z16.s, z25.s\n"
       "smin z17.s, p2/M, z17.s, z25.s\n"
       "smin z18.s, p2/M, z18.s, z25.s\n"
@@ -575,20 +575,20 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "smax z16.s, p2/M, z16.s, z24.s\n"
       "smax z17.s, p2/M, z17.s, z24.s\n"
       "smax z18.s, p2/M, z18.s, z24.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z19.s, p2/M, z19.s, z24.s\n"
       "smax z20.s, p2/M, z20.s, z24.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z21.s, p2/M, z21.s, z24.s\n"
       "smax z22.s, p2/M, z22.s, z24.s\n"
-      "smax z23.s, p2/M, z23.s, z24.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
       "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z17.h, z22.h, z23.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z20.b, z20.b, z17.b\n"
       "st1b { z16.b }, p1, [x27]\n"
+      "smax z23.s, p2/M, z23.s, z24.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
+      "st1b { z20.b }, p1, [x23]\n"
       "addvl x27, x27, #1\n"
-      "st1b { z20.b }, p1, [x24]\n"
       "28:"  // Height 2: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -603,16 +603,16 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "mov z15.b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "30:"  // Height 3: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
@@ -623,8 +623,8 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "mov x26, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -646,57 +646,57 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ble 37f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z3.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z31.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x22]\n"
+      "ld1b { z28.b }, p2/Z, [x28]\n"
+      "sdot z16.s, z28.b, z0.b[0]\n"
+      "sdot z20.s, z28.b, z1.b[0]\n"
+      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "sdot z24.s, z28.b, z2.b[0]\n"
+      "sdot z17.s, z30.b, z0.b[0]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "sdot z21.s, z30.b, z1.b[0]\n"
+      "sdot z25.s, z30.b, z2.b[0]\n"
+      "ld1b { z3.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "sdot z18.s, z29.b, z0.b[0]\n"
+      "sdot z22.s, z29.b, z1.b[0]\n"
+      "ld1b { z31.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1b { z30.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x22, x22, #0x10\n"
-      "sdot z16.s, z5.b, z0.b[0]\n"
-      "sdot z20.s, z5.b, z1.b[0]\n"
-      "sdot z17.s, z29.b, z0.b[0]\n"
-      "sdot z21.s, z29.b, z1.b[0]\n"
-      "sdot z18.s, z4.b, z0.b[0]\n"
-      "sdot z24.s, z5.b, z2.b[0]\n"
-      "sdot z25.s, z29.b, z2.b[0]\n"
+      "sdot z26.s, z29.b, z2.b[0]\n"
+      "sdot z19.s, z28.b, z0.b[0]\n"
       "ld1b { z29.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      "sdot z22.s, z4.b, z1.b[0]\n"
-      "sdot z26.s, z4.b, z2.b[0]\n"
-      "sdot z19.s, z28.b, z0.b[0]\n"
       "sdot z23.s, z28.b, z1.b[0]\n"
       "sdot z27.s, z28.b, z2.b[0]\n"
-      "sdot z16.s, z3.b, z0.b[1]\n"
       "ld1b { z28.b }, p2/Z, [x28, #-8, MUL VL]\n"
       "ld1b { z5.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      "sdot z16.s, z3.b, z0.b[1]\n"
       "sdot z20.s, z3.b, z1.b[1]\n"
-      "sdot z24.s, z3.b, z2.b[1]\n"
       "ld1b { z4.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z3.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      "add x23, x23, #0x10\n"
+      "sdot z24.s, z3.b, z2.b[1]\n"
       "sdot z17.s, z31.b, z0.b[1]\n"
+      "ld1b { z3.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      "add x22, x22, #0x10\n"
       "sdot z21.s, z31.b, z1.b[1]\n"
       "sdot z25.s, z31.b, z2.b[1]\n"
-      "sdot z18.s, z30.b, z0.b[1]\n"
       "ld1b { z31.b }, p2/Z, [x28, #-4, MUL VL]\n"
+      "sdot z18.s, z30.b, z0.b[1]\n"
       "sdot z22.s, z30.b, z1.b[1]\n"
       "sdot z26.s, z30.b, z2.b[1]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "sdot z19.s, z29.b, z0.b[1]\n"
+      "ld1b { z30.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "sdot z23.s, z29.b, z1.b[1]\n"
       "sdot z27.s, z29.b, z2.b[1]\n"
-      "sdot z16.s, z28.b, z0.b[2]\n"
       "ld1b { z29.b }, p2/Z, [x28, #-2, MUL VL]\n"
+      "sdot z16.s, z28.b, z0.b[2]\n"
       "sdot z20.s, z28.b, z1.b[2]\n"
       "sdot z24.s, z28.b, z2.b[2]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-1, MUL VL]\n"
       "sdot z17.s, z5.b, z0.b[2]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #-1, MUL VL]\n"
       "sdot z21.s, z5.b, z1.b[2]\n"
       "sdot z25.s, z5.b, z2.b[2]\n"
       "sdot z18.s, z4.b, z0.b[2]\n"
@@ -727,23 +727,23 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "bgt 35b\n"
       "37:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "subs x25, x25, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x22]\n"
-      "sdot z16.s, z31.b, z0.b[0]\n"
-      "sdot z20.s, z31.b, z1.b[0]\n"
+      "ld1b { z28.b }, p2/Z, [x28]\n"
+      "sdot z16.s, z28.b, z0.b[0]\n"
+      "sdot z20.s, z28.b, z1.b[0]\n"
+      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "sdot z24.s, z28.b, z2.b[0]\n"
       "sdot z17.s, z30.b, z0.b[0]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
       "sdot z21.s, z30.b, z1.b[0]\n"
+      "sdot z25.s, z30.b, z2.b[0]\n"
+      "addvl x28, x28, #4\n"
       "sdot z18.s, z29.b, z0.b[0]\n"
       "sdot z22.s, z29.b, z1.b[0]\n"
-      "sdot z24.s, z31.b, z2.b[0]\n"
-      "sdot z25.s, z30.b, z2.b[0]\n"
       "sdot z26.s, z29.b, z2.b[0]\n"
       "sdot z19.s, z28.b, z0.b[0]\n"
       "sdot z23.s, z28.b, z1.b[0]\n"
@@ -752,14 +752,14 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ld1b { z31.b }, p2/Z, [x28]\n"
       "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "sdot z16.s, z31.b, z0.b[1]\n"
       "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z31.b, z0.b[1]\n"
       "sdot z20.s, z31.b, z1.b[1]\n"
       "sdot z24.s, z31.b, z2.b[1]\n"
       "sdot z17.s, z30.b, z0.b[1]\n"
       "sdot z21.s, z30.b, z1.b[1]\n"
+      "addvl x28, x28, #4\n"
       "sdot z25.s, z30.b, z2.b[1]\n"
       "sdot z18.s, z29.b, z0.b[1]\n"
       "sdot z22.s, z29.b, z1.b[1]\n"
@@ -771,14 +771,14 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ld1b { z31.b }, p2/Z, [x28]\n"
       "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "sdot z16.s, z31.b, z0.b[2]\n"
       "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z31.b, z0.b[2]\n"
       "sdot z20.s, z31.b, z1.b[2]\n"
       "sdot z24.s, z31.b, z2.b[2]\n"
       "sdot z17.s, z30.b, z0.b[2]\n"
       "sdot z21.s, z30.b, z1.b[2]\n"
+      "addvl x28, x28, #4\n"
       "sdot z25.s, z30.b, z2.b[2]\n"
       "sdot z18.s, z29.b, z0.b[2]\n"
       "sdot z22.s, z29.b, z1.b[2]\n"
@@ -789,15 +789,15 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ble 38f\n"
       "ld1b { z31.b }, p2/Z, [x28]\n"
       "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
       "sdot z16.s, z31.b, z0.b[3]\n"
       "sdot z20.s, z31.b, z1.b[3]\n"
+      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
       "sdot z24.s, z31.b, z2.b[3]\n"
       "sdot z17.s, z30.b, z0.b[3]\n"
       "sdot z21.s, z30.b, z1.b[3]\n"
       "sdot z25.s, z30.b, z2.b[3]\n"
+      "addvl x28, x28, #4\n"
       "sdot z18.s, z29.b, z0.b[3]\n"
       "sdot z22.s, z29.b, z1.b[3]\n"
       "sdot z26.s, z29.b, z2.b[3]\n"
@@ -815,22 +815,22 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "cmp x26, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "tbnz %x[flags], #31, 40f\n"
-      "mov x21, #0x4\n"
+      "mov x20, #0x4\n"
+      "whilelt p0.s, XZR, x20\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z28.s, p2/M, z28.s\n"
       "saddv d11, p0, z11.s\n"
+      "mov z11.s, z11.s[0]\n"
       "saddv d12, p0, z12.s\n"
       "saddv d13, p0, z13.s\n"
-      "mov z11.s, z11.s[0]\n"
       "mov z12.s, z12.s[0]\n"
+      "mov z13.s, z13.s[0]\n"
+      "neg z28.s, p2/M, z28.s\n"
       "mul z11.s, p2/M, z11.s, z28.s\n"
       "mul z12.s, p2/M, z12.s, z28.s\n"
-      "mov z13.s, z13.s[0]\n"
       "mul z13.s, p2/M, z13.s, z28.s\n"
       "40:"  // Height 3: skip row sum fixup
       "add z16.s, z16.s, z11.s\n"
@@ -885,18 +885,18 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "and z30.d, z18.d, z0.d\n"
       "and z29.d, z19.d, z0.d\n"
       "and z28.d, z20.d, z0.d\n"
-      "and z3.d, z21.d, z0.d\n"
       "asr z1.s, z1.s, #0x1f\n"
       "asr z31.s, z31.s, #0x1f\n"
       "asr z30.s, z30.s, #0x1f\n"
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
-      "and z2.d, z22.d, z0.d\n"
       "sqadd z16.s, z16.s, z1.s\n"
       "sqadd z17.s, z17.s, z31.s\n"
       "sqadd z18.s, z18.s, z30.s\n"
       "sqadd z19.s, z19.s, z29.s\n"
       "sqadd z20.s, z20.s, z28.s\n"
+      "and z3.d, z21.d, z0.d\n"
+      "and z2.d, z22.d, z0.d\n"
       "and z1.d, z23.d, z0.d\n"
       "and z31.d, z24.d, z0.d\n"
       "and z30.d, z25.d, z0.d\n"
@@ -918,35 +918,35 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "sqadd z27.s, z27.s, z28.s\n"
       "41:"  // Height 3: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z28.s }, p2/Z, [x20]\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z30.s }, p2/Z, [x20]\n"
+      "add z16.s, z16.s, z28.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z28.s\n"
+      "add z18.s, z18.s, z28.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
+      "add z19.s, z19.s, z28.s\n"
+      "add z20.s, z20.s, z28.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z30.s\n"
+      "add z21.s, z21.s, z28.s\n"
+      "add z22.s, z22.s, z28.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z30.s\n"
-      "add z18.s, z18.s, z30.s\n"
+      "add z23.s, z23.s, z28.s\n"
+      "add z24.s, z24.s, z28.s\n"
       ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
       ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z30.s\n"
-      "add z20.s, z20.s, z30.s\n"
+      "add z25.s, z25.s, z28.s\n"
+      "add z26.s, z26.s, z28.s\n"
       ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z21.s, z21.s, z30.s\n"
-      "add z22.s, z22.s, z30.s\n"
       "ld1rw { z29.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z30.s\n"
-      "add z24.s, z24.s, z30.s\n"
+      "add z27.s, z27.s, z28.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z30.s\n"
-      "add z26.s, z26.s, z30.s\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z30.s\n"
       "smin z16.s, p2/M, z16.s, z29.s\n"
       "smin z17.s, p2/M, z17.s, z29.s\n"
       "smin z18.s, p2/M, z18.s, z29.s\n"
@@ -962,28 +962,28 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "smax z16.s, p2/M, z16.s, z28.s\n"
       "smax z17.s, p2/M, z17.s, z28.s\n"
       "smax z18.s, p2/M, z18.s, z28.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z19.s, p2/M, z19.s, z28.s\n"
       "smax z20.s, p2/M, z20.s, z28.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z21.s, p2/M, z21.s, z28.s\n"
       "smax z22.s, p2/M, z22.s, z28.s\n"
+      "uzp1 z20.h, z20.h, z21.h\n"
+      "st1b { z16.b }, p1, [x27]\n"
       "smax z23.s, p2/M, z23.s, z28.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z24.s, p2/M, z24.s, z28.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
       "smax z25.s, p2/M, z25.s, z28.s\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
       "smax z26.s, p2/M, z26.s, z28.s\n"
-      "smax z27.s, p2/M, z27.s, z28.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z18.h, z22.h, z23.h\n"
       "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "uzp1 z20.b, z20.b, z18.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
+      "st1b { z20.b }, p1, [x23]\n"
+      "smax z27.s, p2/M, z27.s, z28.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
+      "st1b { z24.b }, p1, [x22]\n"
       "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "st1b { z20.b }, p1, [x24]\n"
-      "st1b { z24.b }, p1, [x23]\n"
       "42:"  // Height 3: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -991,28 +991,27 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "b 58f\n"
       "43:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x4\n"
       "mov x10, %x[col_bias]\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
+      "mov z13.s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov z13.s, #0x0\n"
       "mov z14.s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
       "mov z15.b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "44:"  // Height 4: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
@@ -1027,8 +1026,8 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "mov x26, #0x0\n"
       "46:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 47f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1053,37 +1052,37 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ble 51f\n"
       "49:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x22]\n"
       "ld1rqb { z3.b }, p0/Z, [x21]\n"
+      "add x23, x23, #0x10\n"
       "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
+      "ld1b { z5.b }, p2/Z, [x28]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
       "sdot z16.s, z5.b, z0.b[0]\n"
       "sdot z20.s, z5.b, z1.b[0]\n"
-      "sdot z17.s, z10.b, z0.b[0]\n"
-      "sdot z21.s, z10.b, z1.b[0]\n"
+      "ld1b { z10.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
       "sdot z24.s, z5.b, z2.b[0]\n"
       "sdot z28.s, z5.b, z3.b[0]\n"
+      "sdot z17.s, z4.b, z0.b[0]\n"
+      "sdot z21.s, z4.b, z1.b[0]\n"
+      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
+      "sdot z25.s, z4.b, z2.b[0]\n"
+      "sdot z29.s, z4.b, z3.b[0]\n"
+      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
+      "sdot z18.s, z10.b, z0.b[0]\n"
+      "sdot z22.s, z10.b, z1.b[0]\n"
       "addvl x28, x28, #16\n"
-      "sdot z25.s, z10.b, z2.b[0]\n"
-      "sdot z29.s, z10.b, z3.b[0]\n"
-      "sdot z18.s, z4.b, z0.b[0]\n"
-      "sdot z22.s, z4.b, z1.b[0]\n"
-      "sdot z26.s, z4.b, z2.b[0]\n"
-      "sdot z30.s, z4.b, z3.b[0]\n"
       "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
+      "sdot z26.s, z10.b, z2.b[0]\n"
+      "sdot z30.s, z10.b, z3.b[0]\n"
       "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      "add x21, x21, #0x10\n"
       "sdot z19.s, z9.b, z0.b[0]\n"
       "sdot z23.s, z9.b, z1.b[0]\n"
       "sdot z27.s, z9.b, z2.b[0]\n"
@@ -1153,26 +1152,26 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "bgt 49b\n"
       "51:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z7.b }, p2/Z, [x28]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "subs x25, x25, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x22]\n"
       "ld1rqb { z3.b }, p0/Z, [x21]\n"
+      "ld1b { z7.b }, p2/Z, [x28]\n"
+      "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
       "sdot z16.s, z7.b, z0.b[0]\n"
       "sdot z20.s, z7.b, z1.b[0]\n"
-      "sdot z17.s, z6.b, z0.b[0]\n"
-      "sdot z21.s, z6.b, z1.b[0]\n"
-      "sdot z18.s, z5.b, z0.b[0]\n"
-      "sdot z22.s, z5.b, z1.b[0]\n"
+      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
       "sdot z24.s, z7.b, z2.b[0]\n"
       "sdot z28.s, z7.b, z3.b[0]\n"
+      "sdot z17.s, z6.b, z0.b[0]\n"
+      "sdot z21.s, z6.b, z1.b[0]\n"
+      "addvl x28, x28, #4\n"
       "sdot z25.s, z6.b, z2.b[0]\n"
       "sdot z29.s, z6.b, z3.b[0]\n"
+      "sdot z18.s, z5.b, z0.b[0]\n"
+      "sdot z22.s, z5.b, z1.b[0]\n"
       "sdot z26.s, z5.b, z2.b[0]\n"
       "sdot z30.s, z5.b, z3.b[0]\n"
       "sdot z19.s, z4.b, z0.b[0]\n"
@@ -1183,14 +1182,14 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ld1b { z7.b }, p2/Z, [x28]\n"
       "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "sdot z16.s, z7.b, z0.b[1]\n"
       "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z7.b, z0.b[1]\n"
       "sdot z20.s, z7.b, z1.b[1]\n"
       "sdot z24.s, z7.b, z2.b[1]\n"
       "sdot z28.s, z7.b, z3.b[1]\n"
       "sdot z17.s, z6.b, z0.b[1]\n"
+      "addvl x28, x28, #4\n"
       "sdot z21.s, z6.b, z1.b[1]\n"
       "sdot z25.s, z6.b, z2.b[1]\n"
       "sdot z29.s, z6.b, z3.b[1]\n"
@@ -1206,14 +1205,14 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ld1b { z7.b }, p2/Z, [x28]\n"
       "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "sdot z16.s, z7.b, z0.b[2]\n"
       "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "sdot z16.s, z7.b, z0.b[2]\n"
       "sdot z20.s, z7.b, z1.b[2]\n"
       "sdot z24.s, z7.b, z2.b[2]\n"
       "sdot z28.s, z7.b, z3.b[2]\n"
       "sdot z17.s, z6.b, z0.b[2]\n"
+      "addvl x28, x28, #4\n"
       "sdot z21.s, z6.b, z1.b[2]\n"
       "sdot z25.s, z6.b, z2.b[2]\n"
       "sdot z29.s, z6.b, z3.b[2]\n"
@@ -1228,15 +1227,15 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "ble 52f\n"
       "ld1b { z7.b }, p2/Z, [x28]\n"
       "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
       "sdot z16.s, z7.b, z0.b[3]\n"
       "sdot z20.s, z7.b, z1.b[3]\n"
+      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
       "sdot z24.s, z7.b, z2.b[3]\n"
       "sdot z28.s, z7.b, z3.b[3]\n"
       "sdot z17.s, z6.b, z0.b[3]\n"
       "sdot z21.s, z6.b, z1.b[3]\n"
+      "addvl x28, x28, #4\n"
       "sdot z25.s, z6.b, z2.b[3]\n"
       "sdot z29.s, z6.b, z3.b[3]\n"
       "sdot z18.s, z5.b, z0.b[3]\n"
@@ -1259,25 +1258,25 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "cmp x26, x20\n"
       "bne 46b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
+      "add x23, x27, x20\n"
       "add x22, x23, x20\n"
+      "add x21, x22, x20\n"
       "tbnz %x[flags], #31, 54f\n"
-      "mov x21, #0x4\n"
+      "mov x20, #0x4\n"
+      "whilelt p0.s, XZR, x20\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z0.s, p2/M, z0.s\n"
       "saddv d11, p0, z11.s\n"
+      "mov z11.s, z11.s[0]\n"
       "saddv d12, p0, z12.s\n"
       "saddv d13, p0, z13.s\n"
-      "saddv d14, p0, z14.s\n"
-      "mov z11.s, z11.s[0]\n"
       "mov z12.s, z12.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z0.s\n"
-      "mul z12.s, p2/M, z12.s, z0.s\n"
       "mov z13.s, z13.s[0]\n"
+      "saddv d14, p0, z14.s\n"
+      "neg z0.s, p2/M, z0.s\n"
       "mov z14.s, z14.s[0]\n"
+      "mul z11.s, p2/M, z11.s, z0.s\n"
+      "mul z12.s, p2/M, z12.s, z0.s\n"
       "mul z13.s, p2/M, z13.s, z0.s\n"
       "mul z14.s, p2/M, z14.s, z0.s\n"
       "54:"  // Height 4: skip row sum fixup
@@ -1342,32 +1341,32 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "tbz %x[flags], #5, 55f\n"
       "and z2.d, z16.d, z0.d\n"
       "and z1.d, z17.d, z0.d\n"
+      "asr z2.s, z2.s, #0x1f\n"
+      "asr z1.s, z1.s, #0x1f\n"
+      "sqadd z16.s, z16.s, z2.s\n"
+      "sqadd z17.s, z17.s, z1.s\n"
       "and z7.d, z18.d, z0.d\n"
       "and z6.d, z19.d, z0.d\n"
       "and z5.d, z20.d, z0.d\n"
       "and z4.d, z21.d, z0.d\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
       "and z3.d, z22.d, z0.d\n"
+      "and z2.d, z23.d, z0.d\n"
+      "and z1.d, z24.d, z0.d\n"
       "asr z7.s, z7.s, #0x1f\n"
       "asr z6.s, z6.s, #0x1f\n"
       "asr z5.s, z5.s, #0x1f\n"
-      "sqadd z16.s, z16.s, z2.s\n"
-      "sqadd z17.s, z17.s, z1.s\n"
-      "and z2.d, z23.d, z0.d\n"
-      "and z1.d, z24.d, z0.d\n"
       "asr z4.s, z4.s, #0x1f\n"
       "asr z3.s, z3.s, #0x1f\n"
-      "sqadd z18.s, z18.s, z7.s\n"
-      "sqadd z19.s, z19.s, z6.s\n"
       "asr z2.s, z2.s, #0x1f\n"
       "asr z1.s, z1.s, #0x1f\n"
+      "sqadd z18.s, z18.s, z7.s\n"
+      "sqadd z19.s, z19.s, z6.s\n"
       "sqadd z20.s, z20.s, z5.s\n"
       "sqadd z21.s, z21.s, z4.s\n"
       "sqadd z22.s, z22.s, z3.s\n"
-      "and z7.d, z25.d, z0.d\n"
       "sqadd z23.s, z23.s, z2.s\n"
       "sqadd z24.s, z24.s, z1.s\n"
+      "and z7.d, z25.d, z0.d\n"
       "and z6.d, z26.d, z0.d\n"
       "and z5.d, z27.d, z0.d\n"
       "and z4.d, z28.d, z0.d\n"
@@ -1390,43 +1389,43 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "sqadd z31.s, z31.s, z1.s\n"
       "55:"  // Height 4: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
       "ld1rw { z2.s }, p2/Z, [x20]\n"
+      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
+      "add z16.s, z16.s, z2.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z2.s\n"
+      "add z18.s, z18.s, z2.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
+      "add z19.s, z19.s, z2.s\n"
+      "add z20.s, z20.s, z2.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z2.s\n"
+      "add z21.s, z21.s, z2.s\n"
+      "add z22.s, z22.s, z2.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z2.s\n"
-      "add z18.s, z18.s, z2.s\n"
+      "add z23.s, z23.s, z2.s\n"
+      "add z24.s, z24.s, z2.s\n"
       ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
       ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z2.s\n"
-      "add z20.s, z20.s, z2.s\n"
+      "add z25.s, z25.s, z2.s\n"
+      "add z26.s, z26.s, z2.s\n"
       ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
       ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
-      "add z21.s, z21.s, z2.s\n"
-      "add z22.s, z22.s, z2.s\n"
+      "add z27.s, z27.s, z2.s\n"
+      "add z28.s, z28.s, z2.s\n"
       ".inst 0x4482881d  // srshl z29.s, p2/M, z29.s, z0.s\n"
       ".inst 0x4482881e  // srshl z30.s, p2/M, z30.s, z0.s\n"
-      "add z23.s, z23.s, z2.s\n"
-      "add z24.s, z24.s, z2.s\n"
+      "add z29.s, z29.s, z2.s\n"
+      "add z30.s, z30.s, z2.s\n"
       ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z25.s, z25.s, z2.s\n"
-      "add z26.s, z26.s, z2.s\n"
       "ld1rw { z1.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z2.s\n"
-      "add z28.s, z28.s, z2.s\n"
+      "add z31.s, z31.s, z2.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z29.s, z29.s, z2.s\n"
-      "add z30.s, z30.s, z2.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z31.s, z31.s, z2.s\n"
       "smin z16.s, p2/M, z16.s, z1.s\n"
       "smin z17.s, p2/M, z17.s, z1.s\n"
       "smin z18.s, p2/M, z18.s, z1.s\n"
@@ -1446,36 +1445,36 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "smax z16.s, p2/M, z16.s, z0.s\n"
       "smax z17.s, p2/M, z17.s, z0.s\n"
       "smax z18.s, p2/M, z18.s, z0.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z19.s, p2/M, z19.s, z0.s\n"
       "smax z20.s, p2/M, z20.s, z0.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z21.s, p2/M, z21.s, z0.s\n"
       "smax z22.s, p2/M, z22.s, z0.s\n"
+      "uzp1 z20.h, z20.h, z21.h\n"
+      "st1b { z16.b }, p1, [x27]\n"
       "smax z23.s, p2/M, z23.s, z0.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z24.s, p2/M, z24.s, z0.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
       "smax z25.s, p2/M, z25.s, z0.s\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
       "smax z26.s, p2/M, z26.s, z0.s\n"
+      "uzp1 z24.h, z24.h, z25.h\n"
+      "st1b { z20.b }, p1, [x23]\n"
       "smax z27.s, p2/M, z27.s, z0.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
       "smax z28.s, p2/M, z28.s, z0.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "smax z29.s, p2/M, z29.s, z0.s\n"
-      "uzp1 z17.h, z22.h, z23.h\n"
       "smax z30.s, p2/M, z30.s, z0.s\n"
-      "smax z31.s, p2/M, z31.s, z0.s\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z18.h, z26.h, z27.h\n"
       "uzp1 z28.h, z28.h, z29.h\n"
-      "uzp1 z20.b, z20.b, z17.b\n"
-      "uzp1 z17.h, z30.h, z31.h\n"
-      "st1b { z16.b }, p1, [x27]\n"
+      "st1b { z24.b }, p1, [x22]\n"
+      "smax z31.s, p2/M, z31.s, z0.s\n"
+      "uzp1 z16.h, z30.h, z31.h\n"
+      "uzp1 z28.b, z28.b, z16.b\n"
+      "st1b { z28.b }, p1, [x21]\n"
       "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z18.b\n"
-      "uzp1 z28.b, z28.b, z17.b\n"
-      "st1b { z20.b }, p1, [x24]\n"
-      "st1b { z24.b }, p1, [x23]\n"
-      "st1b { z28.b }, p1, [x22]\n"
       "56:"  // Height 4: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -1492,8 +1491,8 @@ void sve_hybrid_s8qa_dot_4x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "58:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL.hpp
index 96550f4839..4a57f89880 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 4, 8, 8> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 4, 8, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp
index b8e65e6999..0d5ea54cb7 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qa_mmla_4x4VL/generic.cpp
@@ -45,18 +45,18 @@ void sve_hybrid_s8qa_mmla_4x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -91,24 +91,24 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "3:"  // Height 1: setup done
       "mov x26, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -124,43 +124,43 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "ble 9f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z30.b }, p2/Z, [x28]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x24, x24, #0x10\n"
-      "trn1 z0.d, z1.d, z31.d\n"
-      ".inst 0x451e9810  // smmla z16.s, z0.b, z30.b\n"
-      "ld1b { z8.b }, p2/Z, [x28, #7, MUL VL]\n"
+      "trn1 z0.d, z1.d, z26.d\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      ".inst 0x45189810  // smmla z16.s, z0.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "trn2 z1.d, z1.d, z26.d\n"
+      "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45199814  // smmla z20.s, z0.b, z25.b\n"
+      ".inst 0x45189811  // smmla z17.s, z0.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x451a9815  // smmla z21.s, z0.b, z26.b\n"
+      ".inst 0x45199812  // smmla z18.s, z0.b, z25.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      "trn2 z1.d, z1.d, z31.d\n"
-      ".inst 0x451d9814  // smmla z20.s, z0.b, z29.b\n"
-      ".inst 0x451c9811  // smmla z17.s, z0.b, z28.b\n"
-      ".inst 0x451b9815  // smmla z21.s, z0.b, z27.b\n"
-      ".inst 0x451a9812  // smmla z18.s, z0.b, z26.b\n"
-      "ld1b { z31.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x45199816  // smmla z22.s, z0.b, z25.b\n"
-      ".inst 0x45189813  // smmla z19.s, z0.b, z24.b\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x45089817  // smmla z23.s, z0.b, z8.b\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x45189816  // smmla z22.s, z0.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
+      ".inst 0x451a9813  // smmla z19.s, z0.b, z26.b\n"
+      ".inst 0x45199817  // smmla z23.s, z0.b, z25.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      ".inst 0x45189830  // smmla z16.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x451a9834  // smmla z20.s, z1.b, z26.b\n"
       "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x451f9830  // smmla z16.s, z1.b, z31.b\n"
       "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
+      ".inst 0x45199831  // smmla z17.s, z1.b, z25.b\n"
+      ".inst 0x45189835  // smmla z21.s, z1.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x451e9834  // smmla z20.s, z1.b, z30.b\n"
-      ".inst 0x451d9831  // smmla z17.s, z1.b, z29.b\n"
-      ".inst 0x451c9835  // smmla z21.s, z1.b, z28.b\n"
       ".inst 0x451b9832  // smmla z18.s, z1.b, z27.b\n"
       ".inst 0x451a9836  // smmla z22.s, z1.b, z26.b\n"
       ".inst 0x45199833  // smmla z19.s, z1.b, z25.b\n"
       ".inst 0x45189837  // smmla z23.s, z1.b, z24.b\n"
+      "add x24, x24, #0x10\n"
       "tbnz %x[flags], #31, 8f\n"
       "sdot z11.s, z0.b, z15.b\n"
       "sdot z11.s, z1.b, z15.b\n"
@@ -170,45 +170,45 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "bgt 7b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
+      "ld1rqb { z1.b }, p0/Z, [x24]\n"
+      "trn1 z0.d, z1.d, z27.d\n"
       "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45189810  // smmla z16.s, z0.b, z24.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x8\n"
+      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "trn2 z1.d, z1.d, z27.d\n"
+      ".inst 0x451a9814  // smmla z20.s, z0.b, z26.b\n"
       "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
       "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
+      ".inst 0x45199811  // smmla z17.s, z0.b, z25.b\n"
+      ".inst 0x45189815  // smmla z21.s, z0.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z31.d\n"
-      ".inst 0x45189810  // smmla z16.s, z0.b, z24.b\n"
       "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #8\n"
-      "trn2 z1.d, z1.d, z31.d\n"
-      ".inst 0x451e9814  // smmla z20.s, z0.b, z30.b\n"
-      ".inst 0x451d9811  // smmla z17.s, z0.b, z29.b\n"
-      ".inst 0x451c9815  // smmla z21.s, z0.b, z28.b\n"
       ".inst 0x451b9812  // smmla z18.s, z0.b, z27.b\n"
       ".inst 0x451a9816  // smmla z22.s, z0.b, z26.b\n"
       ".inst 0x45199813  // smmla z19.s, z0.b, z25.b\n"
       ".inst 0x45189817  // smmla z23.s, z0.b, z24.b\n"
+      "addvl x28, x28, #8\n"
       "ble 10f\n"
       "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
       ".inst 0x45189830  // smmla z16.s, z1.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x45189834  // smmla z20.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45199831  // smmla z17.s, z1.b, z25.b\n"
+      ".inst 0x45189835  // smmla z21.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45199832  // smmla z18.s, z1.b, z25.b\n"
+      ".inst 0x45189836  // smmla z22.s, z1.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x451e9834  // smmla z20.s, z1.b, z30.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x451d9831  // smmla z17.s, z1.b, z29.b\n"
-      ".inst 0x451c9835  // smmla z21.s, z1.b, z28.b\n"
-      ".inst 0x451b9832  // smmla z18.s, z1.b, z27.b\n"
-      ".inst 0x451a9836  // smmla z22.s, z1.b, z26.b\n"
       ".inst 0x45199833  // smmla z19.s, z1.b, z25.b\n"
       ".inst 0x45189837  // smmla z23.s, z1.b, z24.b\n"
+      "addvl x28, x28, #8\n"
       "10:"  // Height 1: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 11f\n"
       "sdot z11.s, z0.b, z15.b\n"
@@ -225,32 +225,32 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "mov z23.d, z16.d\n"
       "tbnz %x[flags], #31, 12f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1rw { z16.s }, p2/Z, [x20]\n"
       ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
-      "ld1rw { z9.s }, p2/Z, [x20]\n"
-      "neg z9.s, p2/M, z9.s\n"
+      "neg z16.s, p2/M, z16.s\n"
       "mov z11.s, z11.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z9.s\n"
+      "mul z11.s, p2/M, z11.s, z16.s\n"
       "12:"  // Height 1: skip row sum fixup
       "add z23.s, z23.s, z11.s\n"
       "add z17.s, z17.s, z11.s\n"
       "ld1w { z22.s }, p2/Z, [x10]\n"
-      "ld1w { z24.s }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1w { z21.s }, p2/Z, [x10, #1, MUL VL]\n"
       "add z18.s, z18.s, z11.s\n"
       "add z19.s, z19.s, z11.s\n"
-      "ld1w { z21.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x10, #3, MUL VL]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
       "add z23.s, z23.s, z22.s\n"
-      "add z17.s, z17.s, z24.s\n"
+      "add z17.s, z17.s, z21.s\n"
+      "add z18.s, z18.s, z20.s\n"
+      "add z19.s, z19.s, z16.s\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "addvl x10, x10, #4\n"
-      "add z18.s, z18.s, z21.s\n"
-      "add z19.s, z19.s, z20.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
       ".inst 0x04b076f7  // sqrdmulh z23.s, z23.s, z16.s\n"
       ".inst 0x04b07631  // sqrdmulh z17.s, z17.s, z16.s\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x04b07652  // sqrdmulh z18.s, z18.s, z16.s\n"
       ".inst 0x04b07673  // sqrdmulh z19.s, z19.s, z16.s\n"
       "tbz %x[flags], #5, 13f\n"
@@ -268,19 +268,19 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "sqadd z19.s, z19.s, z16.s\n"
       "13:"  // Height 1: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z16.s }, p2/Z, [x20]\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "ld1rw { z21.s }, p2/Z, [x20]\n"
+      "add z23.s, z23.s, z16.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z16.s\n"
+      "add z18.s, z18.s, z16.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z21.s\n"
+      "add z19.s, z19.s, z16.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z21.s\n"
-      "add z18.s, z18.s, z21.s\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z21.s\n"
       "smin z23.s, p2/M, z23.s, z20.s\n"
       "smin z17.s, p2/M, z17.s, z20.s\n"
       "smin z18.s, p2/M, z18.s, z20.s\n"
@@ -288,8 +288,8 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "smax z23.s, p2/M, z23.s, z16.s\n"
       "smax z17.s, p2/M, z17.s, z16.s\n"
       "smax z18.s, p2/M, z18.s, z16.s\n"
-      "smax z19.s, p2/M, z19.s, z16.s\n"
       "uzp1 z23.h, z23.h, z17.h\n"
+      "smax z19.s, p2/M, z19.s, z16.s\n"
       "uzp1 z16.h, z18.h, z19.h\n"
       "uzp1 z23.b, z23.b, z16.b\n"
       "st1b { z23.b }, p1, [x27]\n"
@@ -307,24 +307,24 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "mov z15.b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "16:"  // Height 2: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "17:"  // Height 2: setup done
       "mov x26, #0x0\n"
       "18:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 19f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -343,45 +343,45 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "ble 23f\n"
       "21:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z25.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1b { z24.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      ".inst 0x451f9810  // smmla z16.s, z0.b, z31.b\n"
+      "ld1rqb { z26.b }, p0/Z, [x23]\n"
+      "trn1 z0.d, z1.d, z26.d\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      ".inst 0x45189810  // smmla z16.s, z0.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "trn2 z1.d, z1.d, z26.d\n"
+      "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45199814  // smmla z20.s, z0.b, z25.b\n"
+      ".inst 0x45189811  // smmla z17.s, z0.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x451a9815  // smmla z21.s, z0.b, z26.b\n"
+      ".inst 0x45199812  // smmla z18.s, z0.b, z25.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x451e9814  // smmla z20.s, z0.b, z30.b\n"
-      ".inst 0x451d9811  // smmla z17.s, z0.b, z29.b\n"
-      ".inst 0x451c9815  // smmla z21.s, z0.b, z28.b\n"
-      ".inst 0x451b9812  // smmla z18.s, z0.b, z27.b\n"
-      ".inst 0x451a9816  // smmla z22.s, z0.b, z26.b\n"
-      ".inst 0x45189813  // smmla z19.s, z0.b, z24.b\n"
+      ".inst 0x45189816  // smmla z22.s, z0.b, z24.b\n"
       "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
+      ".inst 0x451a9813  // smmla z19.s, z0.b, z26.b\n"
       ".inst 0x45199817  // smmla z23.s, z0.b, z25.b\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
       ".inst 0x45189830  // smmla z16.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x451a9834  // smmla z20.s, z1.b, z26.b\n"
+      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
       "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
+      ".inst 0x45199831  // smmla z17.s, z1.b, z25.b\n"
+      ".inst 0x45189835  // smmla z21.s, z1.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x451e9834  // smmla z20.s, z1.b, z30.b\n"
-      ".inst 0x451d9831  // smmla z17.s, z1.b, z29.b\n"
-      ".inst 0x451c9835  // smmla z21.s, z1.b, z28.b\n"
       ".inst 0x451b9832  // smmla z18.s, z1.b, z27.b\n"
       ".inst 0x451a9836  // smmla z22.s, z1.b, z26.b\n"
       ".inst 0x45199833  // smmla z19.s, z1.b, z25.b\n"
       ".inst 0x45189837  // smmla z23.s, z1.b, z24.b\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "tbnz %x[flags], #31, 22f\n"
       "sdot z11.s, z0.b, z15.b\n"
       "sdot z11.s, z1.b, z15.b\n"
@@ -391,46 +391,46 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "bgt 21b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z29.b }, p2/Z, [x28]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
+      "ld1rqb { z27.b }, p0/Z, [x23]\n"
+      "trn1 z0.d, z1.d, z27.d\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      ".inst 0x45189810  // smmla z16.s, z0.b, z24.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "subs x25, x25, #0x8\n"
+      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "trn2 z1.d, z1.d, z27.d\n"
+      ".inst 0x451a9814  // smmla z20.s, z0.b, z26.b\n"
+      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45199811  // smmla z17.s, z0.b, z25.b\n"
+      ".inst 0x45189815  // smmla z21.s, z0.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z24.d\n"
-      "trn2 z1.d, z1.d, z24.d\n"
-      ".inst 0x451d9810  // smmla z16.s, z0.b, z29.b\n"
       "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x451c9814  // smmla z20.s, z0.b, z28.b\n"
-      ".inst 0x45049811  // smmla z17.s, z0.b, z4.b\n"
-      ".inst 0x451b9815  // smmla z21.s, z0.b, z27.b\n"
-      ".inst 0x451a9812  // smmla z18.s, z0.b, z26.b\n"
-      ".inst 0x45069816  // smmla z22.s, z0.b, z6.b\n"
+      ".inst 0x451b9812  // smmla z18.s, z0.b, z27.b\n"
+      ".inst 0x451a9816  // smmla z22.s, z0.b, z26.b\n"
       ".inst 0x45199813  // smmla z19.s, z0.b, z25.b\n"
       ".inst 0x45189817  // smmla z23.s, z0.b, z24.b\n"
+      "addvl x28, x28, #8\n"
       "ble 24f\n"
       "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
       ".inst 0x45189830  // smmla z16.s, z1.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x45189834  // smmla z20.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45199831  // smmla z17.s, z1.b, z25.b\n"
+      ".inst 0x45189835  // smmla z21.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45199832  // smmla z18.s, z1.b, z25.b\n"
+      ".inst 0x45189836  // smmla z22.s, z1.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x451e9834  // smmla z20.s, z1.b, z30.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x451d9831  // smmla z17.s, z1.b, z29.b\n"
-      ".inst 0x451c9835  // smmla z21.s, z1.b, z28.b\n"
-      ".inst 0x451b9832  // smmla z18.s, z1.b, z27.b\n"
-      ".inst 0x451a9836  // smmla z22.s, z1.b, z26.b\n"
       ".inst 0x45199833  // smmla z19.s, z1.b, z25.b\n"
       ".inst 0x45189837  // smmla z23.s, z1.b, z24.b\n"
+      "addvl x28, x28, #8\n"
       "24:"  // Height 2: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 25f\n"
       "sdot z11.s, z0.b, z15.b\n"
@@ -443,18 +443,18 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "uzp1 z24.d, z16.d, z20.d\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp2 z16.d, z16.d, z20.d\n"
+      "add x23, x27, x20\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "add x23, x27, x20\n"
       "mov z23.d, z24.d\n"
       "tbnz %x[flags], #31, 26f\n"
       "add x20, %x[qp], %[b_offset]\n"
-      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
+      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
       "neg z24.s, p2/M, z24.s\n"
       "mov z12.s, z11.s[3]\n"
       "mov z11.s, z11.s[0]\n"
@@ -497,24 +497,24 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       ".inst 0x04b87673  // sqrdmulh z19.s, z19.s, z24.s\n"
       "tbz %x[flags], #5, 27f\n"
       "and z24.d, z23.d, z0.d\n"
+      "asr z24.s, z24.s, #0x1f\n"
+      "sqadd z23.s, z23.s, z24.s\n"
       "and z30.d, z20.d, z0.d\n"
       "and z29.d, z21.d, z0.d\n"
       "and z28.d, z22.d, z0.d\n"
       "and z27.d, z16.d, z0.d\n"
       "and z26.d, z17.d, z0.d\n"
-      "asr z24.s, z24.s, #0x1f\n"
       "and z25.d, z18.d, z0.d\n"
+      "and z24.d, z19.d, z0.d\n"
       "asr z30.s, z30.s, #0x1f\n"
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "asr z27.s, z27.s, #0x1f\n"
-      "sqadd z23.s, z23.s, z24.s\n"
-      "and z24.d, z19.d, z0.d\n"
       "asr z26.s, z26.s, #0x1f\n"
       "asr z25.s, z25.s, #0x1f\n"
+      "asr z24.s, z24.s, #0x1f\n"
       "sqadd z20.s, z20.s, z30.s\n"
       "sqadd z21.s, z21.s, z29.s\n"
-      "asr z24.s, z24.s, #0x1f\n"
       "sqadd z22.s, z22.s, z28.s\n"
       "sqadd z16.s, z16.s, z27.s\n"
       "sqadd z17.s, z17.s, z26.s\n"
@@ -522,27 +522,27 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "sqadd z19.s, z19.s, z24.s\n"
       "27:"  // Height 2: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z24.s }, p2/Z, [x20]\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "ld1rw { z26.s }, p2/Z, [x20]\n"
+      "add z23.s, z23.s, z24.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
+      "add z20.s, z20.s, z24.s\n"
+      "add z21.s, z21.s, z24.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
+      "add z22.s, z22.s, z24.s\n"
+      "add z16.s, z16.s, z24.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z23.s, z23.s, z26.s\n"
+      "add z17.s, z17.s, z24.s\n"
+      "add z18.s, z18.s, z24.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z20.s, z20.s, z26.s\n"
-      "add z21.s, z21.s, z26.s\n"
       "ld1rw { z25.s }, p2/Z, [x20]\n"
-      "add z22.s, z22.s, z26.s\n"
-      "add z16.s, z16.s, z26.s\n"
+      "add z19.s, z19.s, z24.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z26.s\n"
-      "add z18.s, z18.s, z26.s\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z26.s\n"
       "smin z23.s, p2/M, z23.s, z25.s\n"
       "smin z20.s, p2/M, z20.s, z25.s\n"
       "smin z21.s, p2/M, z21.s, z25.s\n"
@@ -554,20 +554,20 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "smax z23.s, p2/M, z23.s, z24.s\n"
       "smax z20.s, p2/M, z20.s, z24.s\n"
       "smax z21.s, p2/M, z21.s, z24.s\n"
+      "uzp1 z23.h, z23.h, z20.h\n"
       "smax z22.s, p2/M, z22.s, z24.s\n"
       "smax z16.s, p2/M, z16.s, z24.s\n"
+      "uzp1 z20.h, z21.h, z22.h\n"
+      "uzp1 z23.b, z23.b, z20.b\n"
       "smax z17.s, p2/M, z17.s, z24.s\n"
       "smax z18.s, p2/M, z18.s, z24.s\n"
-      "smax z19.s, p2/M, z19.s, z24.s\n"
-      "uzp1 z23.h, z23.h, z20.h\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
       "uzp1 z16.h, z16.h, z17.h\n"
+      "st1b { z23.b }, p1, [x27]\n"
+      "smax z19.s, p2/M, z19.s, z24.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
-      "uzp1 z23.b, z23.b, z20.b\n"
       "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z23.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
       "st1b { z16.b }, p1, [x23]\n"
+      "addvl x27, x27, #1\n"
       "28:"  // Height 2: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -582,16 +582,16 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "mov z15.b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "30:"  // Height 3: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
@@ -606,8 +606,8 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "mov x26, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -629,49 +629,49 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "ble 37f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
-      "add x22, x22, #0x10\n"
       "trn1 z0.d, z1.d, z2.d\n"
       "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z6.d\n"
-      "trn2 z3.d, z3.d, z6.d\n"
+      "ld1b { z4.b }, p2/Z, [x28]\n"
+      "trn1 z2.d, z3.d, z5.d\n"
+      "trn2 z3.d, z3.d, z5.d\n"
+      ".inst 0x45049810  // smmla z16.s, z0.b, z4.b\n"
+      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x45049858  // smmla z24.s, z2.b, z4.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45059814  // smmla z20.s, z0.b, z5.b\n"
+      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x4505985c  // smmla z28.s, z2.b, z5.b\n"
+      ".inst 0x45049811  // smmla z17.s, z0.b, z4.b\n"
       "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45059810  // smmla z16.s, z0.b, z5.b\n"
-      ".inst 0x450a9814  // smmla z20.s, z0.b, z10.b\n"
-      ".inst 0x45099811  // smmla z17.s, z0.b, z9.b\n"
-      ".inst 0x45089815  // smmla z21.s, z0.b, z8.b\n"
-      ".inst 0x45049812  // smmla z18.s, z0.b, z4.b\n"
-      ".inst 0x45059858  // smmla z24.s, z2.b, z5.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x450a985c  // smmla z28.s, z2.b, z10.b\n"
-      ".inst 0x45099859  // smmla z25.s, z2.b, z9.b\n"
-      ".inst 0x4508985d  // smmla z29.s, z2.b, z8.b\n"
-      ".inst 0x4504985a  // smmla z26.s, z2.b, z4.b\n"
-      ".inst 0x45079816  // smmla z22.s, z0.b, z7.b\n"
+      ".inst 0x45049859  // smmla z25.s, z2.b, z4.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x4507985e  // smmla z30.s, z2.b, z7.b\n"
-      ".inst 0x45069813  // smmla z19.s, z0.b, z6.b\n"
+      ".inst 0x45099815  // smmla z21.s, z0.b, z9.b\n"
+      ".inst 0x4509985d  // smmla z29.s, z2.b, z9.b\n"
       "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      ".inst 0x45089812  // smmla z18.s, z0.b, z8.b\n"
+      ".inst 0x4508985a  // smmla z26.s, z2.b, z8.b\n"
       "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x4506985b  // smmla z27.s, z2.b, z6.b\n"
-      ".inst 0x45059817  // smmla z23.s, z0.b, z5.b\n"
       "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x45079816  // smmla z22.s, z0.b, z7.b\n"
+      ".inst 0x4507985e  // smmla z30.s, z2.b, z7.b\n"
       "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x4505985f  // smmla z31.s, z2.b, z5.b\n"
-      ".inst 0x45049830  // smmla z16.s, z1.b, z4.b\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45069813  // smmla z19.s, z0.b, z6.b\n"
+      ".inst 0x4506985b  // smmla z27.s, z2.b, z6.b\n"
       "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x45059817  // smmla z23.s, z0.b, z5.b\n"
+      ".inst 0x4505985f  // smmla z31.s, z2.b, z5.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
+      "add x22, x22, #0x10\n"
+      ".inst 0x45049830  // smmla z16.s, z1.b, z4.b\n"
       ".inst 0x45049878  // smmla z24.s, z3.b, z4.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
       ".inst 0x450a9834  // smmla z20.s, z1.b, z10.b\n"
@@ -699,32 +699,32 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "bgt 35b\n"
       "37:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "trn1 z0.d, z1.d, z2.d\n"
       "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z5.d\n"
-      "trn2 z3.d, z3.d, z5.d\n"
+      "ld1b { z5.b }, p2/Z, [x28]\n"
+      "trn1 z2.d, z3.d, z4.d\n"
+      "trn2 z3.d, z3.d, z4.d\n"
+      ".inst 0x45059810  // smmla z16.s, z0.b, z5.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x45059858  // smmla z24.s, z2.b, z5.b\n"
+      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "subs x25, x25, #0x8\n"
+      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45049814  // smmla z20.s, z0.b, z4.b\n"
+      ".inst 0x4504985c  // smmla z28.s, z2.b, z4.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45049810  // smmla z16.s, z0.b, z4.b\n"
-      ".inst 0x450a9814  // smmla z20.s, z0.b, z10.b\n"
-      ".inst 0x45099811  // smmla z17.s, z0.b, z9.b\n"
-      ".inst 0x45089815  // smmla z21.s, z0.b, z8.b\n"
-      ".inst 0x45079812  // smmla z18.s, z0.b, z7.b\n"
-      ".inst 0x45049858  // smmla z24.s, z2.b, z4.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x450a985c  // smmla z28.s, z2.b, z10.b\n"
-      "addvl x28, x28, #8\n"
+      ".inst 0x45099811  // smmla z17.s, z0.b, z9.b\n"
       ".inst 0x45099859  // smmla z25.s, z2.b, z9.b\n"
+      ".inst 0x45089815  // smmla z21.s, z0.b, z8.b\n"
       ".inst 0x4508985d  // smmla z29.s, z2.b, z8.b\n"
+      "addvl x28, x28, #8\n"
+      ".inst 0x45079812  // smmla z18.s, z0.b, z7.b\n"
       ".inst 0x4507985a  // smmla z26.s, z2.b, z7.b\n"
       ".inst 0x45069816  // smmla z22.s, z0.b, z6.b\n"
       ".inst 0x4506985e  // smmla z30.s, z2.b, z6.b\n"
@@ -734,24 +734,24 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       ".inst 0x4504985f  // smmla z31.s, z2.b, z4.b\n"
       "ble 38f\n"
       "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
+      ".inst 0x45049830  // smmla z16.s, z1.b, z4.b\n"
+      ".inst 0x45049878  // smmla z24.s, z3.b, z4.b\n"
+      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45059834  // smmla z20.s, z1.b, z5.b\n"
+      ".inst 0x4505987c  // smmla z28.s, z3.b, z5.b\n"
       "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
       "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x45049830  // smmla z16.s, z1.b, z4.b\n"
-      ".inst 0x45049878  // smmla z24.s, z3.b, z4.b\n"
+      ".inst 0x45049831  // smmla z17.s, z1.b, z4.b\n"
+      ".inst 0x45049879  // smmla z25.s, z3.b, z4.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x450a9834  // smmla z20.s, z1.b, z10.b\n"
-      ".inst 0x450a987c  // smmla z28.s, z3.b, z10.b\n"
-      ".inst 0x45099831  // smmla z17.s, z1.b, z9.b\n"
-      ".inst 0x45099879  // smmla z25.s, z3.b, z9.b\n"
-      "addvl x28, x28, #8\n"
       ".inst 0x45089835  // smmla z21.s, z1.b, z8.b\n"
       ".inst 0x4508987d  // smmla z29.s, z3.b, z8.b\n"
       ".inst 0x45079832  // smmla z18.s, z1.b, z7.b\n"
       ".inst 0x4507987a  // smmla z26.s, z3.b, z7.b\n"
+      "addvl x28, x28, #8\n"
       ".inst 0x45069836  // smmla z22.s, z1.b, z6.b\n"
       ".inst 0x4506987e  // smmla z30.s, z3.b, z6.b\n"
       ".inst 0x45059833  // smmla z19.s, z1.b, z5.b\n"
@@ -771,15 +771,15 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z0.d, z16.d, z20.d\n"
+      "add x23, x27, x20\n"
       "uzp2 z16.d, z16.d, z20.d\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
+      "add x22, x23, x20\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "add x23, x27, x20\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "add x22, x23, x20\n"
       "uzp1 z24.d, z24.d, z28.d\n"
       "uzp1 z25.d, z25.d, z29.d\n"
       "uzp1 z26.d, z26.d, z30.d\n"
@@ -787,14 +787,14 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "mov z31.d, z0.d\n"
       "tbnz %x[flags], #31, 40f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1rw { z23.s }, p2/Z, [x20]\n"
       ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
       ".inst 0x4491a9ad  // addp z13.s, p2/m, z13.s, z13.s\n"
-      "ld1rw { z23.s }, p2/Z, [x20]\n"
       "neg z23.s, p2/M, z23.s\n"
       "mov z12.s, z11.s[3]\n"
       "mov z11.s, z11.s[0]\n"
-      "mov z13.s, z13.s[0]\n"
       "mul z11.s, p2/M, z11.s, z23.s\n"
+      "mov z13.s, z13.s[0]\n"
       "mul z12.s, p2/M, z12.s, z23.s\n"
       "mul z13.s, p2/M, z13.s, z23.s\n"
       "40:"  // Height 3: skip row sum fixup
@@ -850,18 +850,18 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "and z29.d, z21.d, z0.d\n"
       "and z28.d, z22.d, z0.d\n"
       "and z23.d, z16.d, z0.d\n"
-      "and z3.d, z17.d, z0.d\n"
       "asr z1.s, z1.s, #0x1f\n"
       "asr z30.s, z30.s, #0x1f\n"
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "asr z23.s, z23.s, #0x1f\n"
-      "and z2.d, z18.d, z0.d\n"
       "sqadd z31.s, z31.s, z1.s\n"
       "sqadd z20.s, z20.s, z30.s\n"
       "sqadd z21.s, z21.s, z29.s\n"
       "sqadd z22.s, z22.s, z28.s\n"
       "sqadd z16.s, z16.s, z23.s\n"
+      "and z3.d, z17.d, z0.d\n"
+      "and z2.d, z18.d, z0.d\n"
       "and z1.d, z19.d, z0.d\n"
       "and z30.d, z24.d, z0.d\n"
       "and z29.d, z25.d, z0.d\n"
@@ -883,35 +883,35 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "sqadd z27.s, z27.s, z23.s\n"
       "41:"  // Height 3: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z23.s }, p2/Z, [x20]\n"
       ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
-      "ld1rw { z29.s }, p2/Z, [x20]\n"
+      "add z31.s, z31.s, z23.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
+      "add z20.s, z20.s, z23.s\n"
+      "add z21.s, z21.s, z23.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
+      "add z22.s, z22.s, z23.s\n"
+      "add z16.s, z16.s, z23.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z31.s, z31.s, z29.s\n"
+      "add z17.s, z17.s, z23.s\n"
+      "add z18.s, z18.s, z23.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z20.s, z20.s, z29.s\n"
-      "add z21.s, z21.s, z29.s\n"
+      "add z19.s, z19.s, z23.s\n"
+      "add z24.s, z24.s, z23.s\n"
       ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
       ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z22.s, z22.s, z29.s\n"
-      "add z16.s, z16.s, z29.s\n"
+      "add z25.s, z25.s, z23.s\n"
+      "add z26.s, z26.s, z23.s\n"
       ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z17.s, z17.s, z29.s\n"
-      "add z18.s, z18.s, z29.s\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z29.s\n"
-      "add z24.s, z24.s, z29.s\n"
+      "add z27.s, z27.s, z23.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z29.s\n"
-      "add z26.s, z26.s, z29.s\n"
       "ld1rw { z23.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z29.s\n"
       "smin z31.s, p2/M, z31.s, z28.s\n"
       "smin z20.s, p2/M, z20.s, z28.s\n"
       "smin z21.s, p2/M, z21.s, z28.s\n"
@@ -927,28 +927,28 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "smax z31.s, p2/M, z31.s, z23.s\n"
       "smax z20.s, p2/M, z20.s, z23.s\n"
       "smax z21.s, p2/M, z21.s, z23.s\n"
+      "uzp1 z31.h, z31.h, z20.h\n"
       "smax z22.s, p2/M, z22.s, z23.s\n"
       "smax z16.s, p2/M, z16.s, z23.s\n"
+      "uzp1 z20.h, z21.h, z22.h\n"
+      "uzp1 z31.b, z31.b, z20.b\n"
       "smax z17.s, p2/M, z17.s, z23.s\n"
       "smax z18.s, p2/M, z18.s, z23.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
+      "st1b { z31.b }, p1, [x27]\n"
       "smax z19.s, p2/M, z19.s, z23.s\n"
-      "uzp1 z31.h, z31.h, z20.h\n"
       "smax z24.s, p2/M, z24.s, z23.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z25.s, p2/M, z25.s, z23.s\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
       "smax z26.s, p2/M, z26.s, z23.s\n"
-      "smax z27.s, p2/M, z27.s, z23.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
       "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z31.b, z31.b, z20.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "st1b { z31.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
       "st1b { z16.b }, p1, [x23]\n"
+      "smax z27.s, p2/M, z27.s, z23.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "st1b { z24.b }, p1, [x22]\n"
+      "addvl x27, x27, #1\n"
       "42:"  // Height 3: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -956,28 +956,27 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "b 58f\n"
       "43:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x4\n"
       "mov x10, %x[col_bias]\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
+      "mov z13.s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov z13.s, #0x0\n"
       "mov z14.s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
       "mov z15.b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "44:"  // Height 4: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
@@ -992,8 +991,8 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "mov x26, #0x0\n"
       "46:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 47f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1018,56 +1017,56 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "ble 51f\n"
       "49:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x22]\n"
-      "ld1rqb { z6.b }, p0/Z, [x21]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
       "trn1 z0.d, z1.d, z2.d\n"
+      "ld1rqb { z3.b }, p0/Z, [x22]\n"
+      "ld1rqb { z5.b }, p0/Z, [x21]\n"
       "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z6.d\n"
-      "trn2 z3.d, z3.d, z6.d\n"
-      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45059810  // smmla z16.s, z0.b, z5.b\n"
+      "trn1 z2.d, z3.d, z5.d\n"
+      "ld1b { z4.b }, p2/Z, [x28]\n"
+      "trn2 z3.d, z3.d, z5.d\n"
+      ".inst 0x45049810  // smmla z16.s, z0.b, z4.b\n"
+      ".inst 0x45049858  // smmla z24.s, z2.b, z4.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
       ".inst 0x45049814  // smmla z20.s, z0.b, z4.b\n"
-      ".inst 0x45079811  // smmla z17.s, z0.b, z7.b\n"
-      ".inst 0x45099815  // smmla z21.s, z0.b, z9.b\n"
-      ".inst 0x45089812  // smmla z18.s, z0.b, z8.b\n"
-      ".inst 0x45059858  // smmla z24.s, z2.b, z5.b\n"
+      ".inst 0x4504985c  // smmla z28.s, z2.b, z4.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
+      ".inst 0x45059811  // smmla z17.s, z0.b, z5.b\n"
+      ".inst 0x45059859  // smmla z25.s, z2.b, z5.b\n"
+      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
+      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
+      ".inst 0x45049815  // smmla z21.s, z0.b, z4.b\n"
+      ".inst 0x4504985d  // smmla z29.s, z2.b, z4.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x4504985c  // smmla z28.s, z2.b, z4.b\n"
-      ".inst 0x45079859  // smmla z25.s, z2.b, z7.b\n"
-      ".inst 0x4509985d  // smmla z29.s, z2.b, z9.b\n"
-      ".inst 0x4508985a  // smmla z26.s, z2.b, z8.b\n"
-      ".inst 0x450a9816  // smmla z22.s, z0.b, z10.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x450a985e  // smmla z30.s, z2.b, z10.b\n"
-      ".inst 0x45069813  // smmla z19.s, z0.b, z6.b\n"
+      ".inst 0x45089812  // smmla z18.s, z0.b, z8.b\n"
+      ".inst 0x4508985a  // smmla z26.s, z2.b, z8.b\n"
+      ".inst 0x45079816  // smmla z22.s, z0.b, z7.b\n"
       "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
       "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x4506985b  // smmla z27.s, z2.b, z6.b\n"
-      ".inst 0x45059817  // smmla z23.s, z0.b, z5.b\n"
+      ".inst 0x4507985e  // smmla z30.s, z2.b, z7.b\n"
+      ".inst 0x45069813  // smmla z19.s, z0.b, z6.b\n"
       "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
       "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
+      ".inst 0x4506985b  // smmla z27.s, z2.b, z6.b\n"
+      ".inst 0x45059817  // smmla z23.s, z0.b, z5.b\n"
+      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x4505985f  // smmla z31.s, z2.b, z5.b\n"
       ".inst 0x45049830  // smmla z16.s, z1.b, z4.b\n"
-      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x45049878  // smmla z24.s, z3.b, z4.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
       ".inst 0x450a9834  // smmla z20.s, z1.b, z10.b\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x450a987c  // smmla z28.s, z3.b, z10.b\n"
       ".inst 0x45099831  // smmla z17.s, z1.b, z9.b\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x45099879  // smmla z25.s, z3.b, z9.b\n"
       ".inst 0x45089835  // smmla z21.s, z1.b, z8.b\n"
       ".inst 0x4508987d  // smmla z29.s, z3.b, z8.b\n"
@@ -1090,60 +1089,60 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "bgt 49b\n"
       "51:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z6.b }, p2/Z, [x28]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z7.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
+      "trn1 z0.d, z1.d, z2.d\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "ld1rqb { z5.b }, p0/Z, [x21]\n"
-      "trn1 z0.d, z1.d, z2.d\n"
       "trn2 z1.d, z1.d, z2.d\n"
       "trn1 z2.d, z3.d, z5.d\n"
+      "ld1b { z4.b }, p2/Z, [x28]\n"
       "trn2 z3.d, z3.d, z5.d\n"
+      ".inst 0x45049810  // smmla z16.s, z0.b, z4.b\n"
+      ".inst 0x45049858  // smmla z24.s, z2.b, z4.b\n"
+      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "subs x25, x25, #0x8\n"
+      ".inst 0x45059814  // smmla z20.s, z0.b, z5.b\n"
+      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
+      ".inst 0x4505985c  // smmla z28.s, z2.b, z5.b\n"
+      ".inst 0x45049811  // smmla z17.s, z0.b, z4.b\n"
+      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45069810  // smmla z16.s, z0.b, z6.b\n"
-      ".inst 0x45049814  // smmla z20.s, z0.b, z4.b\n"
-      ".inst 0x45079811  // smmla z17.s, z0.b, z7.b\n"
-      ".inst 0x45099815  // smmla z21.s, z0.b, z9.b\n"
-      ".inst 0x45089812  // smmla z18.s, z0.b, z8.b\n"
-      ".inst 0x45069858  // smmla z24.s, z2.b, z6.b\n"
-      "ld1b { z6.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x4504985c  // smmla z28.s, z2.b, z4.b\n"
+      ".inst 0x45049859  // smmla z25.s, z2.b, z4.b\n"
+      ".inst 0x45089815  // smmla z21.s, z0.b, z8.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
+      ".inst 0x4508985d  // smmla z29.s, z2.b, z8.b\n"
+      ".inst 0x45079812  // smmla z18.s, z0.b, z7.b\n"
       "addvl x28, x28, #8\n"
-      ".inst 0x45079859  // smmla z25.s, z2.b, z7.b\n"
-      ".inst 0x4509985d  // smmla z29.s, z2.b, z9.b\n"
-      ".inst 0x4508985a  // smmla z26.s, z2.b, z8.b\n"
-      ".inst 0x450a9816  // smmla z22.s, z0.b, z10.b\n"
-      ".inst 0x450a985e  // smmla z30.s, z2.b, z10.b\n"
+      ".inst 0x4507985a  // smmla z26.s, z2.b, z7.b\n"
+      ".inst 0x45069816  // smmla z22.s, z0.b, z6.b\n"
+      ".inst 0x4506985e  // smmla z30.s, z2.b, z6.b\n"
       ".inst 0x45059813  // smmla z19.s, z0.b, z5.b\n"
       ".inst 0x4505985b  // smmla z27.s, z2.b, z5.b\n"
-      ".inst 0x45069817  // smmla z23.s, z0.b, z6.b\n"
-      ".inst 0x4506985f  // smmla z31.s, z2.b, z6.b\n"
+      ".inst 0x45049817  // smmla z23.s, z0.b, z4.b\n"
+      ".inst 0x4504985f  // smmla z31.s, z2.b, z4.b\n"
       "ble 52f\n"
       "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
+      ".inst 0x45049830  // smmla z16.s, z1.b, z4.b\n"
+      ".inst 0x45049878  // smmla z24.s, z3.b, z4.b\n"
+      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45059834  // smmla z20.s, z1.b, z5.b\n"
+      ".inst 0x4505987c  // smmla z28.s, z3.b, z5.b\n"
       "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
       "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x45049830  // smmla z16.s, z1.b, z4.b\n"
-      ".inst 0x45049878  // smmla z24.s, z3.b, z4.b\n"
+      ".inst 0x45049831  // smmla z17.s, z1.b, z4.b\n"
+      ".inst 0x45049879  // smmla z25.s, z3.b, z4.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x450a9834  // smmla z20.s, z1.b, z10.b\n"
-      ".inst 0x450a987c  // smmla z28.s, z3.b, z10.b\n"
-      ".inst 0x45099831  // smmla z17.s, z1.b, z9.b\n"
-      ".inst 0x45099879  // smmla z25.s, z3.b, z9.b\n"
-      "addvl x28, x28, #8\n"
       ".inst 0x45089835  // smmla z21.s, z1.b, z8.b\n"
       ".inst 0x4508987d  // smmla z29.s, z3.b, z8.b\n"
       ".inst 0x45079832  // smmla z18.s, z1.b, z7.b\n"
       ".inst 0x4507987a  // smmla z26.s, z3.b, z7.b\n"
+      "addvl x28, x28, #8\n"
       ".inst 0x45069836  // smmla z22.s, z1.b, z6.b\n"
       ".inst 0x4506987e  // smmla z30.s, z3.b, z6.b\n"
       ".inst 0x45059833  // smmla z19.s, z1.b, z5.b\n"
@@ -1163,16 +1162,16 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "bne 46b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z0.d, z16.d, z20.d\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "uzp2 z16.d, z16.d, z20.d\n"
       "uzp1 z20.d, z17.d, z21.d\n"
+      "add x21, x22, x20\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "add x23, x27, x20\n"
-      "add x22, x23, x20\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "add x21, x22, x20\n"
       "uzp1 z23.d, z24.d, z28.d\n"
       "uzp2 z24.d, z24.d, z28.d\n"
       "uzp1 z28.d, z25.d, z29.d\n"
@@ -1184,15 +1183,15 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "mov z31.d, z0.d\n"
       "tbnz %x[flags], #31, 54f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1rw { z0.s }, p2/Z, [x20]\n"
       ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
       ".inst 0x4491a9ad  // addp z13.s, p2/m, z13.s, z13.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
       "neg z0.s, p2/M, z0.s\n"
       "mov z12.s, z11.s[3]\n"
       "mov z11.s, z11.s[0]\n"
+      "mul z11.s, p2/M, z11.s, z0.s\n"
       "mov z14.s, z13.s[3]\n"
       "mov z13.s, z13.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z0.s\n"
       "mul z12.s, p2/M, z12.s, z0.s\n"
       "mul z13.s, p2/M, z13.s, z0.s\n"
       "mul z14.s, p2/M, z14.s, z0.s\n"
@@ -1258,32 +1257,32 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "tbz %x[flags], #5, 55f\n"
       "and z2.d, z31.d, z0.d\n"
       "and z1.d, z20.d, z0.d\n"
+      "asr z2.s, z2.s, #0x1f\n"
+      "asr z1.s, z1.s, #0x1f\n"
+      "sqadd z31.s, z31.s, z2.s\n"
+      "sqadd z20.s, z20.s, z1.s\n"
       "and z7.d, z21.d, z0.d\n"
       "and z6.d, z22.d, z0.d\n"
       "and z5.d, z16.d, z0.d\n"
       "and z4.d, z17.d, z0.d\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
       "and z3.d, z18.d, z0.d\n"
+      "and z2.d, z19.d, z0.d\n"
+      "and z1.d, z23.d, z0.d\n"
       "asr z7.s, z7.s, #0x1f\n"
       "asr z6.s, z6.s, #0x1f\n"
       "asr z5.s, z5.s, #0x1f\n"
-      "sqadd z31.s, z31.s, z2.s\n"
-      "sqadd z20.s, z20.s, z1.s\n"
-      "and z2.d, z19.d, z0.d\n"
-      "and z1.d, z23.d, z0.d\n"
       "asr z4.s, z4.s, #0x1f\n"
       "asr z3.s, z3.s, #0x1f\n"
-      "sqadd z21.s, z21.s, z7.s\n"
-      "sqadd z22.s, z22.s, z6.s\n"
       "asr z2.s, z2.s, #0x1f\n"
       "asr z1.s, z1.s, #0x1f\n"
+      "sqadd z21.s, z21.s, z7.s\n"
+      "sqadd z22.s, z22.s, z6.s\n"
       "sqadd z16.s, z16.s, z5.s\n"
       "sqadd z17.s, z17.s, z4.s\n"
       "sqadd z18.s, z18.s, z3.s\n"
-      "and z7.d, z28.d, z0.d\n"
       "sqadd z19.s, z19.s, z2.s\n"
       "sqadd z23.s, z23.s, z1.s\n"
+      "and z7.d, z28.d, z0.d\n"
       "and z6.d, z29.d, z0.d\n"
       "and z5.d, z30.d, z0.d\n"
       "and z4.d, z24.d, z0.d\n"
@@ -1306,43 +1305,43 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "sqadd z27.s, z27.s, z1.s\n"
       "55:"  // Height 4: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
       "ld1rw { z2.s }, p2/Z, [x20]\n"
+      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
+      "add z31.s, z31.s, z2.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
+      "add z20.s, z20.s, z2.s\n"
+      "add z21.s, z21.s, z2.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
+      "add z22.s, z22.s, z2.s\n"
+      "add z16.s, z16.s, z2.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z31.s, z31.s, z2.s\n"
+      "add z17.s, z17.s, z2.s\n"
+      "add z18.s, z18.s, z2.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "add z20.s, z20.s, z2.s\n"
-      "add z21.s, z21.s, z2.s\n"
+      "add z19.s, z19.s, z2.s\n"
+      "add z23.s, z23.s, z2.s\n"
       ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
       ".inst 0x4482881d  // srshl z29.s, p2/M, z29.s, z0.s\n"
-      "add z22.s, z22.s, z2.s\n"
-      "add z16.s, z16.s, z2.s\n"
+      "add z28.s, z28.s, z2.s\n"
+      "add z29.s, z29.s, z2.s\n"
       ".inst 0x4482881e  // srshl z30.s, p2/M, z30.s, z0.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z2.s\n"
-      "add z18.s, z18.s, z2.s\n"
+      "add z30.s, z30.s, z2.s\n"
+      "add z24.s, z24.s, z2.s\n"
       ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
       ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z2.s\n"
-      "add z23.s, z23.s, z2.s\n"
+      "add z25.s, z25.s, z2.s\n"
+      "add z26.s, z26.s, z2.s\n"
       ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z28.s, z28.s, z2.s\n"
-      "add z29.s, z29.s, z2.s\n"
       "ld1rw { z1.s }, p2/Z, [x20]\n"
-      "add z30.s, z30.s, z2.s\n"
-      "add z24.s, z24.s, z2.s\n"
+      "add z27.s, z27.s, z2.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z2.s\n"
-      "add z26.s, z26.s, z2.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z2.s\n"
       "smin z31.s, p2/M, z31.s, z1.s\n"
       "smin z20.s, p2/M, z20.s, z1.s\n"
       "smin z21.s, p2/M, z21.s, z1.s\n"
@@ -1362,36 +1361,36 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "smax z31.s, p2/M, z31.s, z0.s\n"
       "smax z20.s, p2/M, z20.s, z0.s\n"
       "smax z21.s, p2/M, z21.s, z0.s\n"
+      "uzp1 z31.h, z31.h, z20.h\n"
       "smax z22.s, p2/M, z22.s, z0.s\n"
       "smax z16.s, p2/M, z16.s, z0.s\n"
+      "uzp1 z20.h, z21.h, z22.h\n"
+      "uzp1 z31.b, z31.b, z20.b\n"
       "smax z17.s, p2/M, z17.s, z0.s\n"
       "smax z18.s, p2/M, z18.s, z0.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
+      "st1b { z31.b }, p1, [x27]\n"
       "smax z19.s, p2/M, z19.s, z0.s\n"
-      "uzp1 z31.h, z31.h, z20.h\n"
       "smax z23.s, p2/M, z23.s, z0.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z28.s, p2/M, z28.s, z0.s\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
       "smax z29.s, p2/M, z29.s, z0.s\n"
+      "uzp1 z23.h, z23.h, z28.h\n"
+      "st1b { z16.b }, p1, [x23]\n"
       "smax z30.s, p2/M, z30.s, z0.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z24.s, p2/M, z24.s, z0.s\n"
+      "uzp1 z16.h, z29.h, z30.h\n"
+      "uzp1 z23.b, z23.b, z16.b\n"
       "smax z25.s, p2/M, z25.s, z0.s\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
       "smax z26.s, p2/M, z26.s, z0.s\n"
-      "smax z27.s, p2/M, z27.s, z0.s\n"
-      "uzp1 z23.h, z23.h, z28.h\n"
-      "uzp1 z31.b, z31.b, z20.b\n"
-      "uzp1 z18.h, z29.h, z30.h\n"
       "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "st1b { z31.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "uzp1 z23.b, z23.b, z18.b\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "st1b { z16.b }, p1, [x23]\n"
       "st1b { z23.b }, p1, [x22]\n"
+      "smax z27.s, p2/M, z27.s, z0.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "st1b { z24.b }, p1, [x21]\n"
+      "addvl x27, x27, #1\n"
       "56:"  // Height 4: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -1408,8 +1407,8 @@ void sve_hybrid_s8qa_mmla_4x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "58:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL.hpp
index bd34f29894..6116e0cefb 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 4, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp
index 49930b57f7..b3fb963111 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_dot_6x4VL/generic.cpp
@@ -47,18 +47,18 @@ void sve_hybrid_s8qs_dot_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -98,22 +98,22 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "3:"  // Height 1: setup done
       "mov x28, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -129,103 +129,103 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "ble 8f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "sdot z8.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9]\n"
+      "sdot z8.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[0]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "sdot z10.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "sdot z10.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z11.s, z16.b, z0.b[0]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #5, MUL VL]\n"
-      "sdot z8.s, z17.b, z0.b[1]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "sdot z8.s, z16.b, z0.b[1]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[1]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      "sdot z10.s, z17.b, z0.b[1]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "sdot z10.s, z16.b, z0.b[1]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
       "sdot z11.s, z16.b, z0.b[1]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-8, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-7, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-6, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[2]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-5, MUL VL]\n"
       "sdot z10.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-4, MUL VL]\n"
       "sdot z11.s, z16.b, z0.b[2]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-3, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[3]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[3]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-1, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
       "sdot z10.s, z17.b, z0.b[3]\n"
       "sdot z11.s, z16.b, z0.b[3]\n"
+      "add x26, x26, #0x10\n"
       "bgt 7b\n"
       "8:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
-      "sdot z8.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9]\n"
+      "sdot z8.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[0]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z10.s, z17.b, z0.b[0]\n"
       "sdot z11.s, z16.b, z0.b[0]\n"
+      "addvl x9, x9, #4\n"
       "ble 9f\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[1]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[1]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b[1]\n"
       "sdot z11.s, z16.b, z0.b[1]\n"
+      "addvl x9, x9, #4\n"
       "ble 9f\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[2]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b[2]\n"
       "sdot z11.s, z16.b, z0.b[2]\n"
+      "addvl x9, x9, #4\n"
       "ble 9f\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[3]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[3]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z10.s, z17.b, z0.b[3]\n"
       "sdot z11.s, z16.b, z0.b[3]\n"
+      "addvl x9, x9, #4\n"
       "9:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 4b\n"
-      "ld1w { z19.s }, p2/Z, [x14]\n"
-      "ld1w { z18.s }, p2/Z, [x14, #1, MUL VL]\n"
+      "ld1w { z17.s }, p2/Z, [x14]\n"
+      "ld1w { z16.s }, p2/Z, [x14, #1, MUL VL]\n"
+      "add z8.s, z8.s, z17.s\n"
+      "add z9.s, z9.s, z16.s\n"
       "ld1w { z17.s }, p2/Z, [x14, #2, MUL VL]\n"
       "ld1w { z16.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
-      "add z8.s, z8.s, z19.s\n"
-      "add z9.s, z9.s, z18.s\n"
       "add z10.s, z10.s, z17.s\n"
       "add z11.s, z11.s, z16.s\n"
+      "addvl x14, x14, #4\n"
       "tbz %x[flags], #4, 10f\n"
       "ld1w { z0.s }, p2/Z, [x12]\n"
       "ld1w { z4.s }, p2/Z, [x13]\n"
@@ -269,19 +269,19 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sqadd z11.s, z11.s, z16.s\n"
       "12:"  // Height 1: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z16.s }, p2/Z, [x20]\n"
       ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
-      "ld1rw { z18.s }, p2/Z, [x20]\n"
+      "add z8.s, z8.s, z16.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
+      "add z9.s, z9.s, z16.s\n"
+      "add z10.s, z10.s, z16.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
       "ld1rw { z17.s }, p2/Z, [x20]\n"
-      "add z8.s, z8.s, z18.s\n"
+      "add z11.s, z11.s, z16.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z9.s, z9.s, z18.s\n"
-      "add z10.s, z10.s, z18.s\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
-      "add z11.s, z11.s, z18.s\n"
       "smin z8.s, p2/M, z8.s, z17.s\n"
       "smin z9.s, p2/M, z9.s, z17.s\n"
       "smin z10.s, p2/M, z10.s, z17.s\n"
@@ -289,41 +289,41 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "smax z8.s, p2/M, z8.s, z16.s\n"
       "smax z9.s, p2/M, z9.s, z16.s\n"
       "smax z10.s, p2/M, z10.s, z16.s\n"
-      "smax z11.s, p2/M, z11.s, z16.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
+      "smax z11.s, p2/M, z11.s, z16.s\n"
       "uzp1 z16.h, z10.h, z11.h\n"
       "uzp1 z8.b, z8.b, z16.b\n"
-      "st1b { z8.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "st1b { z8.b }, p1, [x11]\n"
+      "addvl x11, x11, #1\n"
       "13:"  // Height 1: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 2b\n"
       "b 80f\n"
       "14:"  // Height 2
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "16:"  // Height 2: setup done
       "mov x28, #0x0\n"
       "17:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 18f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -342,57 +342,57 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "ble 21f\n"
       "20:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z1.b[0]\n"
       "sdot z12.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z1.b[0]\n"
       "sdot z13.s, z16.b, z0.b[0]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z10.s, z17.b, z1.b[0]\n"
       "sdot z14.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "cmp x27, #0x10\n"
       "sdot z11.s, z16.b, z1.b[0]\n"
       "sdot z15.s, z16.b, z0.b[0]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       "sdot z8.s, z17.b, z1.b[1]\n"
       "sdot z12.s, z17.b, z0.b[1]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       "sdot z9.s, z16.b, z1.b[1]\n"
       "sdot z13.s, z16.b, z0.b[1]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
+      "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
       "sdot z10.s, z17.b, z1.b[1]\n"
       "sdot z14.s, z17.b, z0.b[1]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-8, MUL VL]\n"
       "sdot z11.s, z16.b, z1.b[1]\n"
       "sdot z15.s, z16.b, z0.b[1]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-7, MUL VL]\n"
       "sdot z8.s, z17.b, z1.b[2]\n"
       "sdot z12.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-6, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-6, MUL VL]\n"
       "sdot z9.s, z16.b, z1.b[2]\n"
       "sdot z13.s, z16.b, z0.b[2]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-5, MUL VL]\n"
       "sdot z10.s, z17.b, z1.b[2]\n"
       "sdot z14.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-4, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-4, MUL VL]\n"
       "sdot z11.s, z16.b, z1.b[2]\n"
       "sdot z15.s, z16.b, z0.b[2]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-3, MUL VL]\n"
       "sdot z8.s, z17.b, z1.b[3]\n"
       "sdot z12.s, z17.b, z0.b[3]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-2, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-2, MUL VL]\n"
       "sdot z9.s, z16.b, z1.b[3]\n"
       "sdot z13.s, z16.b, z0.b[3]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-1, MUL VL]\n"
       "sdot z10.s, z17.b, z1.b[3]\n"
       "sdot z14.s, z17.b, z0.b[3]\n"
       "sdot z11.s, z16.b, z1.b[3]\n"
@@ -400,64 +400,64 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "bgt 20b\n"
       "21:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[0]\n"
       "sdot z12.s, z17.b, z1.b[0]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[0]\n"
       "sdot z13.s, z16.b, z1.b[0]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z10.s, z17.b, z0.b[0]\n"
       "sdot z14.s, z17.b, z1.b[0]\n"
+      "addvl x9, x9, #4\n"
       "sdot z11.s, z16.b, z0.b[0]\n"
       "sdot z15.s, z16.b, z1.b[0]\n"
       "ble 22f\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[1]\n"
       "sdot z12.s, z17.b, z1.b[1]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[1]\n"
       "sdot z13.s, z16.b, z1.b[1]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b[1]\n"
       "sdot z14.s, z17.b, z1.b[1]\n"
+      "addvl x9, x9, #4\n"
       "sdot z11.s, z16.b, z0.b[1]\n"
       "sdot z15.s, z16.b, z1.b[1]\n"
       "ble 22f\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[2]\n"
       "sdot z12.s, z17.b, z1.b[2]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[2]\n"
       "sdot z13.s, z16.b, z1.b[2]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b[2]\n"
       "sdot z14.s, z17.b, z1.b[2]\n"
+      "addvl x9, x9, #4\n"
       "sdot z11.s, z16.b, z0.b[2]\n"
       "sdot z15.s, z16.b, z1.b[2]\n"
       "ble 22f\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[3]\n"
       "sdot z12.s, z17.b, z1.b[3]\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[3]\n"
       "sdot z13.s, z16.b, z1.b[3]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z10.s, z17.b, z0.b[3]\n"
       "sdot z14.s, z17.b, z1.b[3]\n"
+      "addvl x9, x9, #4\n"
       "sdot z11.s, z16.b, z0.b[3]\n"
       "sdot z15.s, z16.b, z1.b[3]\n"
       "22:"  // Height 2: Multiply loop: multiply skip
@@ -467,17 +467,17 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "bne 17b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "ld1w { z19.s }, p2/Z, [x14]\n"
+      "add x26, x11, x20\n"
+      "add z8.s, z8.s, z19.s\n"
       "ld1w { z18.s }, p2/Z, [x14, #1, MUL VL]\n"
       "ld1w { z17.s }, p2/Z, [x14, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
-      "add x26, x9, x20\n"
-      "add z8.s, z8.s, z19.s\n"
-      "add z12.s, z12.s, z19.s\n"
       "add z9.s, z9.s, z18.s\n"
       "add z10.s, z10.s, z17.s\n"
-      "add z13.s, z13.s, z18.s\n"
+      "ld1w { z16.s }, p2/Z, [x14, #3, MUL VL]\n"
       "add z11.s, z11.s, z16.s\n"
+      "add z12.s, z12.s, z19.s\n"
+      "addvl x14, x14, #4\n"
+      "add z13.s, z13.s, z18.s\n"
       "add z14.s, z14.s, z17.s\n"
       "add z15.s, z15.s, z16.s\n"
       "tbz %x[flags], #4, 23f\n"
@@ -522,11 +522,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z17.s, z17.s, #0x1f\n"
       "asr z16.s, z16.s, #0x1f\n"
       "sqadd z8.s, z8.s, z19.s\n"
-      "and z19.d, z12.d, z0.d\n"
       "sqadd z9.s, z9.s, z18.s\n"
-      "and z18.d, z13.d, z1.d\n"
       "sqadd z10.s, z10.s, z17.s\n"
       "sqadd z11.s, z11.s, z16.s\n"
+      "and z19.d, z12.d, z0.d\n"
+      "and z18.d, z13.d, z1.d\n"
       "and z17.d, z14.d, z2.d\n"
       "and z16.d, z15.d, z3.d\n"
       "asr z19.s, z19.s, #0x1f\n"
@@ -539,73 +539,73 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sqadd z15.s, z15.s, z16.s\n"
       "25:"  // Height 2: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
-      "ld1rw { z18.s }, p2/Z, [x20]\n"
+      "add z8.s, z8.s, z17.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
+      "add z9.s, z9.s, z17.s\n"
+      "add z10.s, z10.s, z17.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x4482880c  // srshl z12.s, p2/M, z12.s, z0.s\n"
+      "add z11.s, z11.s, z17.s\n"
+      "add z12.s, z12.s, z17.s\n"
       ".inst 0x4482882d  // srshl z13.s, p2/M, z13.s, z1.s\n"
       ".inst 0x4482884e  // srshl z14.s, p2/M, z14.s, z2.s\n"
-      "add z8.s, z8.s, z18.s\n"
+      "add z13.s, z13.s, z17.s\n"
+      "add z14.s, z14.s, z17.s\n"
       ".inst 0x4482886f  // srshl z15.s, p2/M, z15.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z9.s, z9.s, z18.s\n"
-      "add z10.s, z10.s, z18.s\n"
-      "ld1rw { z17.s }, p2/Z, [x20]\n"
-      "add z11.s, z11.s, z18.s\n"
-      "add z12.s, z12.s, z18.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z13.s, z13.s, z18.s\n"
-      "add z14.s, z14.s, z18.s\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
-      "add z15.s, z15.s, z18.s\n"
-      "smin z8.s, p2/M, z8.s, z17.s\n"
-      "smin z9.s, p2/M, z9.s, z17.s\n"
-      "smin z10.s, p2/M, z10.s, z17.s\n"
-      "smin z11.s, p2/M, z11.s, z17.s\n"
-      "smin z12.s, p2/M, z12.s, z17.s\n"
-      "smin z13.s, p2/M, z13.s, z17.s\n"
-      "smin z14.s, p2/M, z14.s, z17.s\n"
-      "smin z15.s, p2/M, z15.s, z17.s\n"
-      "smax z8.s, p2/M, z8.s, z16.s\n"
-      "smax z9.s, p2/M, z9.s, z16.s\n"
-      "smax z10.s, p2/M, z10.s, z16.s\n"
-      "smax z11.s, p2/M, z11.s, z16.s\n"
-      "smax z12.s, p2/M, z12.s, z16.s\n"
-      "smax z13.s, p2/M, z13.s, z16.s\n"
-      "smax z14.s, p2/M, z14.s, z16.s\n"
-      "smax z15.s, p2/M, z15.s, z16.s\n"
+      "add z15.s, z15.s, z17.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
+      "smin z8.s, p2/M, z8.s, z16.s\n"
+      "smin z9.s, p2/M, z9.s, z16.s\n"
+      "smin z10.s, p2/M, z10.s, z16.s\n"
+      "smin z11.s, p2/M, z11.s, z16.s\n"
+      "smin z12.s, p2/M, z12.s, z16.s\n"
+      "smin z13.s, p2/M, z13.s, z16.s\n"
+      "smin z14.s, p2/M, z14.s, z16.s\n"
+      "smin z15.s, p2/M, z15.s, z16.s\n"
+      "smax z8.s, p2/M, z8.s, z17.s\n"
+      "smax z9.s, p2/M, z9.s, z17.s\n"
+      "smax z10.s, p2/M, z10.s, z17.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
-      "uzp1 z17.h, z10.h, z11.h\n"
+      "smax z11.s, p2/M, z11.s, z17.s\n"
+      "smax z12.s, p2/M, z12.s, z17.s\n"
+      "uzp1 z16.h, z10.h, z11.h\n"
+      "uzp1 z8.b, z8.b, z16.b\n"
+      "smax z13.s, p2/M, z13.s, z17.s\n"
+      "smax z14.s, p2/M, z14.s, z17.s\n"
       "uzp1 z12.h, z12.h, z13.h\n"
+      "st1b { z8.b }, p1, [x11]\n"
+      "smax z15.s, p2/M, z15.s, z17.s\n"
       "uzp1 z16.h, z14.h, z15.h\n"
-      "uzp1 z8.b, z8.b, z17.b\n"
       "uzp1 z12.b, z12.b, z16.b\n"
-      "st1b { z8.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
       "st1b { z12.b }, p1, [x26]\n"
+      "addvl x11, x11, #1\n"
       "26:"  // Height 2: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 15b\n"
       "b 80f\n"
       "27:"  // Height 3
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "mov z16.s, #0x0\n"
@@ -616,8 +616,8 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "mov x28, #0x0\n"
       "30:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 31f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -639,73 +639,73 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "ble 34f\n"
       "33:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z21.b }, p2/Z, [x10]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1b { z21.b }, p2/Z, [x9]\n"
       "sdot z8.s, z21.b, z2.b[0]\n"
       "sdot z12.s, z21.b, z1.b[0]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #1, MUL VL]\n"
+      "sdot z16.s, z21.b, z0.b[0]\n"
       "sdot z9.s, z20.b, z2.b[0]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[0]\n"
-      "sdot z16.s, z21.b, z0.b[0]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z17.s, z20.b, z0.b[0]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "cmp x27, #0x10\n"
       "sdot z10.s, z21.b, z2.b[0]\n"
       "sdot z14.s, z21.b, z1.b[0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "sdot z18.s, z21.b, z0.b[0]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #4, MUL VL]\n"
       "sdot z11.s, z20.b, z2.b[0]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       "sdot z15.s, z20.b, z1.b[0]\n"
       "sdot z19.s, z20.b, z0.b[0]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #5, MUL VL]\n"
       "sdot z8.s, z21.b, z2.b[1]\n"
       "sdot z12.s, z21.b, z1.b[1]\n"
       "sdot z16.s, z21.b, z0.b[1]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #6, MUL VL]\n"
       "sdot z9.s, z20.b, z2.b[1]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #6, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[1]\n"
       "sdot z17.s, z20.b, z0.b[1]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
+      "ld1b { z20.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
       "sdot z10.s, z21.b, z2.b[1]\n"
       "sdot z14.s, z21.b, z1.b[1]\n"
       "sdot z18.s, z21.b, z0.b[1]\n"
       "sdot z11.s, z20.b, z2.b[1]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #-8, MUL VL]\n"
       "sdot z15.s, z20.b, z1.b[1]\n"
       "sdot z19.s, z20.b, z0.b[1]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #-7, MUL VL]\n"
       "sdot z8.s, z21.b, z2.b[2]\n"
       "sdot z12.s, z21.b, z1.b[2]\n"
       "sdot z16.s, z21.b, z0.b[2]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #-6, MUL VL]\n"
       "sdot z9.s, z20.b, z2.b[2]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #-6, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[2]\n"
       "sdot z17.s, z20.b, z0.b[2]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #-5, MUL VL]\n"
       "sdot z10.s, z21.b, z2.b[2]\n"
       "sdot z14.s, z21.b, z1.b[2]\n"
       "sdot z18.s, z21.b, z0.b[2]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #-4, MUL VL]\n"
       "sdot z11.s, z20.b, z2.b[2]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #-4, MUL VL]\n"
       "sdot z15.s, z20.b, z1.b[2]\n"
       "sdot z19.s, z20.b, z0.b[2]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #-3, MUL VL]\n"
       "sdot z8.s, z21.b, z2.b[3]\n"
       "sdot z12.s, z21.b, z1.b[3]\n"
       "sdot z16.s, z21.b, z0.b[3]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #-2, MUL VL]\n"
       "sdot z9.s, z20.b, z2.b[3]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #-2, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[3]\n"
       "sdot z17.s, z20.b, z0.b[3]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #-1, MUL VL]\n"
       "sdot z10.s, z21.b, z2.b[3]\n"
       "sdot z14.s, z21.b, z1.b[3]\n"
       "sdot z18.s, z21.b, z0.b[3]\n"
@@ -715,21 +715,21 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "bgt 33b\n"
       "34:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z21.b }, p2/Z, [x10]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
+      "ld1b { z21.b }, p2/Z, [x9]\n"
       "sdot z8.s, z21.b, z0.b[0]\n"
       "sdot z12.s, z21.b, z1.b[0]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #1, MUL VL]\n"
+      "sdot z16.s, z21.b, z2.b[0]\n"
       "sdot z9.s, z20.b, z0.b[0]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[0]\n"
-      "sdot z16.s, z21.b, z2.b[0]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z17.s, z20.b, z2.b[0]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z20.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z21.b, z0.b[0]\n"
       "sdot z14.s, z21.b, z1.b[0]\n"
       "sdot z18.s, z21.b, z2.b[0]\n"
@@ -737,18 +737,18 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z15.s, z20.b, z1.b[0]\n"
       "sdot z19.s, z20.b, z2.b[0]\n"
       "ble 35f\n"
-      "ld1b { z21.b }, p2/Z, [x10]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z21.b }, p2/Z, [x9]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z21.b, z0.b[1]\n"
       "sdot z12.s, z21.b, z1.b[1]\n"
       "sdot z16.s, z21.b, z2.b[1]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z20.b, z0.b[1]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z13.s, z20.b, z1.b[1]\n"
       "sdot z17.s, z20.b, z2.b[1]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z20.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z21.b, z0.b[1]\n"
       "sdot z14.s, z21.b, z1.b[1]\n"
       "sdot z18.s, z21.b, z2.b[1]\n"
@@ -756,18 +756,18 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z15.s, z20.b, z1.b[1]\n"
       "sdot z19.s, z20.b, z2.b[1]\n"
       "ble 35f\n"
-      "ld1b { z21.b }, p2/Z, [x10]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z21.b }, p2/Z, [x9]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z21.b, z0.b[2]\n"
       "sdot z12.s, z21.b, z1.b[2]\n"
       "sdot z16.s, z21.b, z2.b[2]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z20.b, z0.b[2]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z13.s, z20.b, z1.b[2]\n"
       "sdot z17.s, z20.b, z2.b[2]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z20.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z21.b, z0.b[2]\n"
       "sdot z14.s, z21.b, z1.b[2]\n"
       "sdot z18.s, z21.b, z2.b[2]\n"
@@ -775,17 +775,17 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z15.s, z20.b, z1.b[2]\n"
       "sdot z19.s, z20.b, z2.b[2]\n"
       "ble 35f\n"
-      "ld1b { z21.b }, p2/Z, [x10]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z21.b }, p2/Z, [x9]\n"
+      "ld1b { z20.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z21.b, z0.b[3]\n"
       "sdot z12.s, z21.b, z1.b[3]\n"
       "sdot z16.s, z21.b, z2.b[3]\n"
-      "ld1b { z21.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z20.b, z0.b[3]\n"
+      "ld1b { z21.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[3]\n"
       "sdot z17.s, z20.b, z2.b[3]\n"
-      "ld1b { z20.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z20.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z21.b, z0.b[3]\n"
       "sdot z14.s, z21.b, z1.b[3]\n"
       "sdot z18.s, z21.b, z2.b[3]\n"
@@ -799,17 +799,17 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "bne 30b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "ld1w { z23.s }, p2/Z, [x14]\n"
+      "add x26, x11, x20\n"
+      "add x25, x26, x20\n"
       "ld1w { z22.s }, p2/Z, [x14, #1, MUL VL]\n"
       "ld1w { z21.s }, p2/Z, [x14, #2, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
       "add z8.s, z8.s, z23.s\n"
-      "add z12.s, z12.s, z23.s\n"
       "add z9.s, z9.s, z22.s\n"
+      "ld1w { z20.s }, p2/Z, [x14, #3, MUL VL]\n"
       "add z10.s, z10.s, z21.s\n"
       "add z11.s, z11.s, z20.s\n"
+      "addvl x14, x14, #4\n"
+      "add z12.s, z12.s, z23.s\n"
       "add z13.s, z13.s, z22.s\n"
       "add z14.s, z14.s, z21.s\n"
       "add z15.s, z15.s, z20.s\n"
@@ -863,11 +863,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z21.s, z21.s, #0x1f\n"
       "asr z20.s, z20.s, #0x1f\n"
       "sqadd z8.s, z8.s, z23.s\n"
-      "and z23.d, z12.d, z0.d\n"
       "sqadd z9.s, z9.s, z22.s\n"
-      "and z22.d, z13.d, z1.d\n"
       "sqadd z10.s, z10.s, z21.s\n"
       "sqadd z11.s, z11.s, z20.s\n"
+      "and z23.d, z12.d, z0.d\n"
+      "and z22.d, z13.d, z1.d\n"
       "and z21.d, z14.d, z2.d\n"
       "and z20.d, z15.d, z3.d\n"
       "asr z23.s, z23.s, #0x1f\n"
@@ -875,11 +875,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z21.s, z21.s, #0x1f\n"
       "asr z20.s, z20.s, #0x1f\n"
       "sqadd z12.s, z12.s, z23.s\n"
-      "and z23.d, z16.d, z0.d\n"
       "sqadd z13.s, z13.s, z22.s\n"
-      "and z22.d, z17.d, z1.d\n"
       "sqadd z14.s, z14.s, z21.s\n"
       "sqadd z15.s, z15.s, z20.s\n"
+      "and z23.d, z16.d, z0.d\n"
+      "and z22.d, z17.d, z1.d\n"
       "and z21.d, z18.d, z2.d\n"
       "and z20.d, z19.d, z3.d\n"
       "asr z23.s, z23.s, #0x1f\n"
@@ -892,93 +892,93 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sqadd z19.s, z19.s, z20.s\n"
       "38:"  // Height 3: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z21.s }, p2/Z, [x20]\n"
       ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
-      "ld1rw { z22.s }, p2/Z, [x20]\n"
+      "add z8.s, z8.s, z21.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
+      "add z9.s, z9.s, z21.s\n"
+      "add z10.s, z10.s, z21.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x4482880c  // srshl z12.s, p2/M, z12.s, z0.s\n"
+      "add z11.s, z11.s, z21.s\n"
+      "add z12.s, z12.s, z21.s\n"
       ".inst 0x4482882d  // srshl z13.s, p2/M, z13.s, z1.s\n"
       ".inst 0x4482884e  // srshl z14.s, p2/M, z14.s, z2.s\n"
-      "add z8.s, z8.s, z22.s\n"
+      "add z13.s, z13.s, z21.s\n"
+      "add z14.s, z14.s, z21.s\n"
       ".inst 0x4482886f  // srshl z15.s, p2/M, z15.s, z3.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "add z9.s, z9.s, z22.s\n"
-      "add z10.s, z10.s, z22.s\n"
+      "add z15.s, z15.s, z21.s\n"
+      "add z16.s, z16.s, z21.s\n"
       ".inst 0x44828831  // srshl z17.s, p2/M, z17.s, z1.s\n"
       ".inst 0x44828852  // srshl z18.s, p2/M, z18.s, z2.s\n"
-      "add z11.s, z11.s, z22.s\n"
-      "add z12.s, z12.s, z22.s\n"
+      "add z17.s, z17.s, z21.s\n"
+      "add z18.s, z18.s, z21.s\n"
       ".inst 0x44828873  // srshl z19.s, p2/M, z19.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z13.s, z13.s, z22.s\n"
-      "add z14.s, z14.s, z22.s\n"
-      "ld1rw { z21.s }, p2/Z, [x20]\n"
-      "add z15.s, z15.s, z22.s\n"
-      "add z16.s, z16.s, z22.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z22.s\n"
-      "add z18.s, z18.s, z22.s\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z22.s\n"
-      "smin z8.s, p2/M, z8.s, z21.s\n"
-      "smin z9.s, p2/M, z9.s, z21.s\n"
-      "smin z10.s, p2/M, z10.s, z21.s\n"
-      "smin z11.s, p2/M, z11.s, z21.s\n"
-      "smin z12.s, p2/M, z12.s, z21.s\n"
-      "smin z13.s, p2/M, z13.s, z21.s\n"
-      "smin z14.s, p2/M, z14.s, z21.s\n"
-      "smin z15.s, p2/M, z15.s, z21.s\n"
-      "smin z16.s, p2/M, z16.s, z21.s\n"
-      "smin z17.s, p2/M, z17.s, z21.s\n"
-      "smin z18.s, p2/M, z18.s, z21.s\n"
-      "smin z19.s, p2/M, z19.s, z21.s\n"
-      "smax z8.s, p2/M, z8.s, z20.s\n"
-      "smax z9.s, p2/M, z9.s, z20.s\n"
-      "smax z10.s, p2/M, z10.s, z20.s\n"
-      "smax z11.s, p2/M, z11.s, z20.s\n"
-      "smax z12.s, p2/M, z12.s, z20.s\n"
-      "smax z13.s, p2/M, z13.s, z20.s\n"
-      "smax z14.s, p2/M, z14.s, z20.s\n"
-      "smax z15.s, p2/M, z15.s, z20.s\n"
+      "add z19.s, z19.s, z21.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z21.s }, p2/Z, [x20]\n"
+      "smin z8.s, p2/M, z8.s, z20.s\n"
+      "smin z9.s, p2/M, z9.s, z20.s\n"
+      "smin z10.s, p2/M, z10.s, z20.s\n"
+      "smin z11.s, p2/M, z11.s, z20.s\n"
+      "smin z12.s, p2/M, z12.s, z20.s\n"
+      "smin z13.s, p2/M, z13.s, z20.s\n"
+      "smin z14.s, p2/M, z14.s, z20.s\n"
+      "smin z15.s, p2/M, z15.s, z20.s\n"
+      "smin z16.s, p2/M, z16.s, z20.s\n"
+      "smin z17.s, p2/M, z17.s, z20.s\n"
+      "smin z18.s, p2/M, z18.s, z20.s\n"
+      "smin z19.s, p2/M, z19.s, z20.s\n"
+      "smax z8.s, p2/M, z8.s, z21.s\n"
+      "smax z9.s, p2/M, z9.s, z21.s\n"
+      "smax z10.s, p2/M, z10.s, z21.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
-      "smax z16.s, p2/M, z16.s, z20.s\n"
-      "smax z17.s, p2/M, z17.s, z20.s\n"
-      "uzp1 z21.h, z10.h, z11.h\n"
-      "smax z18.s, p2/M, z18.s, z20.s\n"
-      "smax z19.s, p2/M, z19.s, z20.s\n"
+      "smax z11.s, p2/M, z11.s, z21.s\n"
+      "smax z12.s, p2/M, z12.s, z21.s\n"
+      "uzp1 z20.h, z10.h, z11.h\n"
+      "uzp1 z8.b, z8.b, z20.b\n"
+      "smax z13.s, p2/M, z13.s, z21.s\n"
+      "smax z14.s, p2/M, z14.s, z21.s\n"
       "uzp1 z12.h, z12.h, z13.h\n"
+      "st1b { z8.b }, p1, [x11]\n"
+      "smax z15.s, p2/M, z15.s, z21.s\n"
+      "smax z16.s, p2/M, z16.s, z21.s\n"
       "uzp1 z20.h, z14.h, z15.h\n"
+      "uzp1 z12.b, z12.b, z20.b\n"
+      "smax z17.s, p2/M, z17.s, z21.s\n"
+      "smax z18.s, p2/M, z18.s, z21.s\n"
       "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z8.b, z8.b, z21.b\n"
+      "st1b { z12.b }, p1, [x26]\n"
+      "smax z19.s, p2/M, z19.s, z21.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
-      "uzp1 z12.b, z12.b, z20.b\n"
-      "st1b { z8.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
       "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z12.b }, p1, [x26]\n"
       "st1b { z16.b }, p1, [x25]\n"
+      "addvl x11, x11, #1\n"
       "39:"  // Height 3: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 28b\n"
       "b 80f\n"
       "40:"  // Height 4
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "mov z16.s, #0x0\n"
@@ -993,8 +993,8 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "mov x28, #0x0\n"
       "43:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 44f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1019,89 +1019,89 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "ble 47f\n"
       "46:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p2/Z, [x10]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z25.b, z3.b[0]\n"
       "sdot z12.s, z25.b, z2.b[0]\n"
-      "sdot z9.s, z24.b, z3.b[0]\n"
-      "sdot z13.s, z24.b, z2.b[0]\n"
       "sdot z16.s, z25.b, z1.b[0]\n"
       "sdot z20.s, z25.b, z0.b[0]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      "sdot z9.s, z24.b, z3.b[0]\n"
+      "sdot z13.s, z24.b, z2.b[0]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "sdot z17.s, z24.b, z1.b[0]\n"
       "sdot z21.s, z24.b, z0.b[0]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z10.s, z25.b, z3.b[0]\n"
       "sdot z14.s, z25.b, z2.b[0]\n"
       "sdot z18.s, z25.b, z1.b[0]\n"
       "sdot z22.s, z25.b, z0.b[0]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
       "sdot z11.s, z24.b, z3.b[0]\n"
       "sdot z15.s, z24.b, z2.b[0]\n"
       "sdot z19.s, z24.b, z1.b[0]\n"
       "sdot z23.s, z24.b, z0.b[0]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
       "sdot z8.s, z25.b, z3.b[1]\n"
       "sdot z12.s, z25.b, z2.b[1]\n"
       "sdot z16.s, z25.b, z1.b[1]\n"
       "sdot z20.s, z25.b, z0.b[1]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
       "sdot z9.s, z24.b, z3.b[1]\n"
       "sdot z13.s, z24.b, z2.b[1]\n"
       "sdot z17.s, z24.b, z1.b[1]\n"
       "sdot z21.s, z24.b, z0.b[1]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
+      "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
       "sdot z10.s, z25.b, z3.b[1]\n"
       "sdot z14.s, z25.b, z2.b[1]\n"
       "sdot z18.s, z25.b, z1.b[1]\n"
       "sdot z22.s, z25.b, z0.b[1]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-8, MUL VL]\n"
       "sdot z11.s, z24.b, z3.b[1]\n"
       "sdot z15.s, z24.b, z2.b[1]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-8, MUL VL]\n"
       "sdot z19.s, z24.b, z1.b[1]\n"
       "sdot z23.s, z24.b, z0.b[1]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-7, MUL VL]\n"
       "sdot z8.s, z25.b, z3.b[2]\n"
       "sdot z12.s, z25.b, z2.b[2]\n"
       "sdot z16.s, z25.b, z1.b[2]\n"
       "sdot z20.s, z25.b, z0.b[2]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-6, MUL VL]\n"
       "sdot z9.s, z24.b, z3.b[2]\n"
       "sdot z13.s, z24.b, z2.b[2]\n"
       "sdot z17.s, z24.b, z1.b[2]\n"
       "sdot z21.s, z24.b, z0.b[2]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-5, MUL VL]\n"
       "sdot z10.s, z25.b, z3.b[2]\n"
       "sdot z14.s, z25.b, z2.b[2]\n"
       "sdot z18.s, z25.b, z1.b[2]\n"
       "sdot z22.s, z25.b, z0.b[2]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-4, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-4, MUL VL]\n"
       "sdot z11.s, z24.b, z3.b[2]\n"
       "sdot z15.s, z24.b, z2.b[2]\n"
       "sdot z19.s, z24.b, z1.b[2]\n"
       "sdot z23.s, z24.b, z0.b[2]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-3, MUL VL]\n"
       "sdot z8.s, z25.b, z3.b[3]\n"
       "sdot z12.s, z25.b, z2.b[3]\n"
       "sdot z16.s, z25.b, z1.b[3]\n"
       "sdot z20.s, z25.b, z0.b[3]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-2, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-2, MUL VL]\n"
       "sdot z9.s, z24.b, z3.b[3]\n"
       "sdot z13.s, z24.b, z2.b[3]\n"
       "sdot z17.s, z24.b, z1.b[3]\n"
       "sdot z21.s, z24.b, z0.b[3]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-1, MUL VL]\n"
       "sdot z10.s, z25.b, z3.b[3]\n"
       "sdot z14.s, z25.b, z2.b[3]\n"
       "sdot z18.s, z25.b, z1.b[3]\n"
@@ -1113,24 +1113,24 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "bgt 46b\n"
       "47:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p2/Z, [x10]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z25.b, z0.b[0]\n"
       "sdot z12.s, z25.b, z1.b[0]\n"
-      "sdot z9.s, z24.b, z0.b[0]\n"
-      "sdot z13.s, z24.b, z1.b[0]\n"
       "sdot z16.s, z25.b, z2.b[0]\n"
       "sdot z20.s, z25.b, z3.b[0]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "sdot z9.s, z24.b, z0.b[0]\n"
+      "sdot z13.s, z24.b, z1.b[0]\n"
       "sdot z17.s, z24.b, z2.b[0]\n"
       "sdot z21.s, z24.b, z3.b[0]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z25.b, z0.b[0]\n"
       "sdot z14.s, z25.b, z1.b[0]\n"
       "sdot z18.s, z25.b, z2.b[0]\n"
@@ -1140,20 +1140,20 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z19.s, z24.b, z2.b[0]\n"
       "sdot z23.s, z24.b, z3.b[0]\n"
       "ble 48f\n"
-      "ld1b { z25.b }, p2/Z, [x10]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z25.b, z0.b[1]\n"
       "sdot z12.s, z25.b, z1.b[1]\n"
       "sdot z16.s, z25.b, z2.b[1]\n"
       "sdot z20.s, z25.b, z3.b[1]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z9.s, z24.b, z0.b[1]\n"
       "sdot z13.s, z24.b, z1.b[1]\n"
       "sdot z17.s, z24.b, z2.b[1]\n"
       "sdot z21.s, z24.b, z3.b[1]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z25.b, z0.b[1]\n"
       "sdot z14.s, z25.b, z1.b[1]\n"
       "sdot z18.s, z25.b, z2.b[1]\n"
@@ -1163,20 +1163,20 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z19.s, z24.b, z2.b[1]\n"
       "sdot z23.s, z24.b, z3.b[1]\n"
       "ble 48f\n"
-      "ld1b { z25.b }, p2/Z, [x10]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z25.b, z0.b[2]\n"
       "sdot z12.s, z25.b, z1.b[2]\n"
       "sdot z16.s, z25.b, z2.b[2]\n"
       "sdot z20.s, z25.b, z3.b[2]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z9.s, z24.b, z0.b[2]\n"
       "sdot z13.s, z24.b, z1.b[2]\n"
       "sdot z17.s, z24.b, z2.b[2]\n"
       "sdot z21.s, z24.b, z3.b[2]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z25.b, z0.b[2]\n"
       "sdot z14.s, z25.b, z1.b[2]\n"
       "sdot z18.s, z25.b, z2.b[2]\n"
@@ -1186,19 +1186,19 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z19.s, z24.b, z2.b[2]\n"
       "sdot z23.s, z24.b, z3.b[2]\n"
       "ble 48f\n"
-      "ld1b { z25.b }, p2/Z, [x10]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z25.b, z0.b[3]\n"
       "sdot z12.s, z25.b, z1.b[3]\n"
       "sdot z16.s, z25.b, z2.b[3]\n"
       "sdot z20.s, z25.b, z3.b[3]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z9.s, z24.b, z0.b[3]\n"
       "sdot z13.s, z24.b, z1.b[3]\n"
       "sdot z17.s, z24.b, z2.b[3]\n"
       "sdot z21.s, z24.b, z3.b[3]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z25.b, z0.b[3]\n"
       "sdot z14.s, z25.b, z1.b[3]\n"
       "sdot z18.s, z25.b, z2.b[3]\n"
@@ -1214,18 +1214,18 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "bne 43b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "ld1w { z27.s }, p2/Z, [x14]\n"
+      "add x26, x11, x20\n"
+      "add x25, x26, x20\n"
       "ld1w { z26.s }, p2/Z, [x14, #1, MUL VL]\n"
       "ld1w { z25.s }, p2/Z, [x14, #2, MUL VL]\n"
-      "ld1w { z24.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
-      "add z8.s, z8.s, z27.s\n"
-      "add z12.s, z12.s, z27.s\n"
       "add x24, x25, x20\n"
+      "add z8.s, z8.s, z27.s\n"
+      "ld1w { z24.s }, p2/Z, [x14, #3, MUL VL]\n"
       "add z9.s, z9.s, z26.s\n"
       "add z10.s, z10.s, z25.s\n"
+      "addvl x14, x14, #4\n"
       "add z11.s, z11.s, z24.s\n"
+      "add z12.s, z12.s, z27.s\n"
       "add z13.s, z13.s, z26.s\n"
       "add z14.s, z14.s, z25.s\n"
       "add z15.s, z15.s, z24.s\n"
@@ -1287,11 +1287,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z25.s, z25.s, #0x1f\n"
       "asr z24.s, z24.s, #0x1f\n"
       "sqadd z8.s, z8.s, z27.s\n"
-      "and z27.d, z12.d, z0.d\n"
       "sqadd z9.s, z9.s, z26.s\n"
-      "and z26.d, z13.d, z1.d\n"
       "sqadd z10.s, z10.s, z25.s\n"
       "sqadd z11.s, z11.s, z24.s\n"
+      "and z27.d, z12.d, z0.d\n"
+      "and z26.d, z13.d, z1.d\n"
       "and z25.d, z14.d, z2.d\n"
       "and z24.d, z15.d, z3.d\n"
       "asr z27.s, z27.s, #0x1f\n"
@@ -1299,11 +1299,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z25.s, z25.s, #0x1f\n"
       "asr z24.s, z24.s, #0x1f\n"
       "sqadd z12.s, z12.s, z27.s\n"
-      "and z27.d, z16.d, z0.d\n"
       "sqadd z13.s, z13.s, z26.s\n"
-      "and z26.d, z17.d, z1.d\n"
       "sqadd z14.s, z14.s, z25.s\n"
       "sqadd z15.s, z15.s, z24.s\n"
+      "and z27.d, z16.d, z0.d\n"
+      "and z26.d, z17.d, z1.d\n"
       "and z25.d, z18.d, z2.d\n"
       "and z24.d, z19.d, z3.d\n"
       "asr z27.s, z27.s, #0x1f\n"
@@ -1311,11 +1311,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z25.s, z25.s, #0x1f\n"
       "asr z24.s, z24.s, #0x1f\n"
       "sqadd z16.s, z16.s, z27.s\n"
-      "and z27.d, z20.d, z0.d\n"
       "sqadd z17.s, z17.s, z26.s\n"
-      "and z26.d, z21.d, z1.d\n"
       "sqadd z18.s, z18.s, z25.s\n"
       "sqadd z19.s, z19.s, z24.s\n"
+      "and z27.d, z20.d, z0.d\n"
+      "and z26.d, z21.d, z1.d\n"
       "and z25.d, z22.d, z2.d\n"
       "and z24.d, z23.d, z3.d\n"
       "asr z27.s, z27.s, #0x1f\n"
@@ -1328,43 +1328,43 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sqadd z23.s, z23.s, z24.s\n"
       "51:"  // Height 4: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
       "ld1rw { z25.s }, p2/Z, [x20]\n"
+      ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
+      "add z8.s, z8.s, z25.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
+      "add z9.s, z9.s, z25.s\n"
+      "add z10.s, z10.s, z25.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x4482880c  // srshl z12.s, p2/M, z12.s, z0.s\n"
+      "add z11.s, z11.s, z25.s\n"
+      "add z12.s, z12.s, z25.s\n"
       ".inst 0x4482882d  // srshl z13.s, p2/M, z13.s, z1.s\n"
       ".inst 0x4482884e  // srshl z14.s, p2/M, z14.s, z2.s\n"
-      "add z8.s, z8.s, z25.s\n"
+      "add z13.s, z13.s, z25.s\n"
+      "add z14.s, z14.s, z25.s\n"
       ".inst 0x4482886f  // srshl z15.s, p2/M, z15.s, z3.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "add z9.s, z9.s, z25.s\n"
-      "add z10.s, z10.s, z25.s\n"
+      "add z15.s, z15.s, z25.s\n"
+      "add z16.s, z16.s, z25.s\n"
       ".inst 0x44828831  // srshl z17.s, p2/M, z17.s, z1.s\n"
       ".inst 0x44828852  // srshl z18.s, p2/M, z18.s, z2.s\n"
-      "add z11.s, z11.s, z25.s\n"
-      "add z12.s, z12.s, z25.s\n"
+      "add z17.s, z17.s, z25.s\n"
+      "add z18.s, z18.s, z25.s\n"
       ".inst 0x44828873  // srshl z19.s, p2/M, z19.s, z3.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      "add z13.s, z13.s, z25.s\n"
-      "add z14.s, z14.s, z25.s\n"
+      "add z19.s, z19.s, z25.s\n"
+      "add z20.s, z20.s, z25.s\n"
       ".inst 0x44828835  // srshl z21.s, p2/M, z21.s, z1.s\n"
       ".inst 0x44828856  // srshl z22.s, p2/M, z22.s, z2.s\n"
-      "add z15.s, z15.s, z25.s\n"
-      "add z16.s, z16.s, z25.s\n"
+      "add z21.s, z21.s, z25.s\n"
+      "add z22.s, z22.s, z25.s\n"
       ".inst 0x44828877  // srshl z23.s, p2/M, z23.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z17.s, z17.s, z25.s\n"
-      "add z18.s, z18.s, z25.s\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z25.s\n"
-      "add z20.s, z20.s, z25.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z21.s, z21.s, z25.s\n"
-      "add z22.s, z22.s, z25.s\n"
-      "ld1rw { z26.s }, p2/Z, [x20]\n"
       "add z23.s, z23.s, z25.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z25.s }, p2/Z, [x20]\n"
       "smin z8.s, p2/M, z8.s, z24.s\n"
       "smin z9.s, p2/M, z9.s, z24.s\n"
       "smin z10.s, p2/M, z10.s, z24.s\n"
@@ -1381,60 +1381,60 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "smin z21.s, p2/M, z21.s, z24.s\n"
       "smin z22.s, p2/M, z22.s, z24.s\n"
       "smin z23.s, p2/M, z23.s, z24.s\n"
-      "smax z8.s, p2/M, z8.s, z26.s\n"
-      "smax z9.s, p2/M, z9.s, z26.s\n"
-      "smax z10.s, p2/M, z10.s, z26.s\n"
-      "smax z11.s, p2/M, z11.s, z26.s\n"
-      "smax z12.s, p2/M, z12.s, z26.s\n"
-      "smax z13.s, p2/M, z13.s, z26.s\n"
-      "smax z14.s, p2/M, z14.s, z26.s\n"
-      "smax z15.s, p2/M, z15.s, z26.s\n"
+      "smax z8.s, p2/M, z8.s, z25.s\n"
+      "smax z9.s, p2/M, z9.s, z25.s\n"
+      "smax z10.s, p2/M, z10.s, z25.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
-      "smax z16.s, p2/M, z16.s, z26.s\n"
-      "smax z17.s, p2/M, z17.s, z26.s\n"
-      "uzp1 z25.h, z10.h, z11.h\n"
-      "smax z18.s, p2/M, z18.s, z26.s\n"
-      "smax z19.s, p2/M, z19.s, z26.s\n"
+      "smax z11.s, p2/M, z11.s, z25.s\n"
+      "smax z12.s, p2/M, z12.s, z25.s\n"
+      "uzp1 z24.h, z10.h, z11.h\n"
+      "uzp1 z8.b, z8.b, z24.b\n"
+      "smax z13.s, p2/M, z13.s, z25.s\n"
+      "smax z14.s, p2/M, z14.s, z25.s\n"
       "uzp1 z12.h, z12.h, z13.h\n"
-      "smax z20.s, p2/M, z20.s, z26.s\n"
-      "smax z21.s, p2/M, z21.s, z26.s\n"
+      "st1b { z8.b }, p1, [x11]\n"
+      "smax z15.s, p2/M, z15.s, z25.s\n"
+      "smax z16.s, p2/M, z16.s, z25.s\n"
       "uzp1 z24.h, z14.h, z15.h\n"
-      "smax z22.s, p2/M, z22.s, z26.s\n"
-      "smax z23.s, p2/M, z23.s, z26.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z8.b, z8.b, z25.b\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
       "uzp1 z12.b, z12.b, z24.b\n"
-      "uzp1 z17.h, z22.h, z23.h\n"
-      "st1b { z8.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z20.b, z20.b, z17.b\n"
+      "smax z17.s, p2/M, z17.s, z25.s\n"
+      "smax z18.s, p2/M, z18.s, z25.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "st1b { z12.b }, p1, [x26]\n"
+      "smax z19.s, p2/M, z19.s, z25.s\n"
+      "smax z20.s, p2/M, z20.s, z25.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
+      "smax z21.s, p2/M, z21.s, z25.s\n"
+      "smax z22.s, p2/M, z22.s, z25.s\n"
+      "uzp1 z20.h, z20.h, z21.h\n"
       "st1b { z16.b }, p1, [x25]\n"
+      "smax z23.s, p2/M, z23.s, z25.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
       "st1b { z20.b }, p1, [x24]\n"
+      "addvl x11, x11, #1\n"
       "52:"  // Height 4: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 41b\n"
       "b 80f\n"
       "53:"  // Height 5
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "54:"  // Height 5: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "mov z16.s, #0x0\n"
@@ -1453,8 +1453,8 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "mov x28, #0x0\n"
       "56:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 57f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1482,105 +1482,105 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "ble 60f\n"
       "59:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p2/Z, [x10]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z4.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1b { z29.b }, p2/Z, [x9]\n"
       "sdot z8.s, z29.b, z4.b[0]\n"
       "sdot z12.s, z29.b, z3.b[0]\n"
-      "sdot z9.s, z28.b, z4.b[0]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z16.s, z29.b, z2.b[0]\n"
       "sdot z20.s, z29.b, z1.b[0]\n"
+      "add x25, x25, #0x10\n"
       "sdot z24.s, z29.b, z0.b[0]\n"
+      "sdot z9.s, z28.b, z4.b[0]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       "sdot z13.s, z28.b, z3.b[0]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z17.s, z28.b, z2.b[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "sdot z21.s, z28.b, z1.b[0]\n"
       "sdot z25.s, z28.b, z0.b[0]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z10.s, z29.b, z4.b[0]\n"
       "sdot z14.s, z29.b, z3.b[0]\n"
       "sdot z18.s, z29.b, z2.b[0]\n"
       "sdot z22.s, z29.b, z1.b[0]\n"
       "sdot z26.s, z29.b, z0.b[0]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #4, MUL VL]\n"
       "sdot z11.s, z28.b, z4.b[0]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #4, MUL VL]\n"
       "sdot z15.s, z28.b, z3.b[0]\n"
       "sdot z19.s, z28.b, z2.b[0]\n"
       "sdot z23.s, z28.b, z1.b[0]\n"
       "sdot z27.s, z28.b, z0.b[0]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #5, MUL VL]\n"
       "sdot z8.s, z29.b, z4.b[1]\n"
       "sdot z12.s, z29.b, z3.b[1]\n"
       "sdot z16.s, z29.b, z2.b[1]\n"
       "sdot z20.s, z29.b, z1.b[1]\n"
       "sdot z24.s, z29.b, z0.b[1]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #6, MUL VL]\n"
       "sdot z9.s, z28.b, z4.b[1]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #6, MUL VL]\n"
       "sdot z13.s, z28.b, z3.b[1]\n"
       "sdot z17.s, z28.b, z2.b[1]\n"
       "sdot z21.s, z28.b, z1.b[1]\n"
       "sdot z25.s, z28.b, z0.b[1]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
+      "ld1b { z28.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
       "sdot z10.s, z29.b, z4.b[1]\n"
       "sdot z14.s, z29.b, z3.b[1]\n"
       "sdot z18.s, z29.b, z2.b[1]\n"
       "sdot z22.s, z29.b, z1.b[1]\n"
       "sdot z26.s, z29.b, z0.b[1]\n"
       "sdot z11.s, z28.b, z4.b[1]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #-8, MUL VL]\n"
       "sdot z15.s, z28.b, z3.b[1]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #-8, MUL VL]\n"
       "sdot z19.s, z28.b, z2.b[1]\n"
       "sdot z23.s, z28.b, z1.b[1]\n"
       "sdot z27.s, z28.b, z0.b[1]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #-7, MUL VL]\n"
       "sdot z8.s, z29.b, z4.b[2]\n"
       "sdot z12.s, z29.b, z3.b[2]\n"
       "sdot z16.s, z29.b, z2.b[2]\n"
       "sdot z20.s, z29.b, z1.b[2]\n"
       "sdot z24.s, z29.b, z0.b[2]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #-6, MUL VL]\n"
       "sdot z9.s, z28.b, z4.b[2]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #-6, MUL VL]\n"
       "sdot z13.s, z28.b, z3.b[2]\n"
       "sdot z17.s, z28.b, z2.b[2]\n"
       "sdot z21.s, z28.b, z1.b[2]\n"
       "sdot z25.s, z28.b, z0.b[2]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #-5, MUL VL]\n"
       "sdot z10.s, z29.b, z4.b[2]\n"
       "sdot z14.s, z29.b, z3.b[2]\n"
       "sdot z18.s, z29.b, z2.b[2]\n"
       "sdot z22.s, z29.b, z1.b[2]\n"
       "sdot z26.s, z29.b, z0.b[2]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #-4, MUL VL]\n"
       "sdot z11.s, z28.b, z4.b[2]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #-4, MUL VL]\n"
       "sdot z15.s, z28.b, z3.b[2]\n"
       "sdot z19.s, z28.b, z2.b[2]\n"
       "sdot z23.s, z28.b, z1.b[2]\n"
       "sdot z27.s, z28.b, z0.b[2]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #-3, MUL VL]\n"
       "sdot z8.s, z29.b, z4.b[3]\n"
       "sdot z12.s, z29.b, z3.b[3]\n"
       "sdot z16.s, z29.b, z2.b[3]\n"
       "sdot z20.s, z29.b, z1.b[3]\n"
       "sdot z24.s, z29.b, z0.b[3]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #-2, MUL VL]\n"
       "sdot z9.s, z28.b, z4.b[3]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #-2, MUL VL]\n"
       "sdot z13.s, z28.b, z3.b[3]\n"
       "sdot z17.s, z28.b, z2.b[3]\n"
       "sdot z21.s, z28.b, z1.b[3]\n"
       "sdot z25.s, z28.b, z0.b[3]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #-1, MUL VL]\n"
       "sdot z10.s, z29.b, z4.b[3]\n"
       "sdot z14.s, z29.b, z3.b[3]\n"
       "sdot z18.s, z29.b, z2.b[3]\n"
@@ -1594,27 +1594,27 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "bgt 59b\n"
       "60:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p2/Z, [x10]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
       "ld1rqb { z4.b }, p0/Z, [x22]\n"
+      "ld1b { z29.b }, p2/Z, [x9]\n"
       "sdot z8.s, z29.b, z0.b[0]\n"
       "sdot z12.s, z29.b, z1.b[0]\n"
-      "sdot z9.s, z28.b, z0.b[0]\n"
-      "sdot z13.s, z28.b, z1.b[0]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z16.s, z29.b, z2.b[0]\n"
       "sdot z20.s, z29.b, z3.b[0]\n"
       "sdot z24.s, z29.b, z4.b[0]\n"
+      "sdot z9.s, z28.b, z0.b[0]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "sdot z13.s, z28.b, z1.b[0]\n"
       "sdot z17.s, z28.b, z2.b[0]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z21.s, z28.b, z3.b[0]\n"
       "sdot z25.s, z28.b, z4.b[0]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z28.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z29.b, z0.b[0]\n"
       "sdot z14.s, z29.b, z1.b[0]\n"
       "sdot z18.s, z29.b, z2.b[0]\n"
@@ -1626,23 +1626,23 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z23.s, z28.b, z3.b[0]\n"
       "sdot z27.s, z28.b, z4.b[0]\n"
       "ble 61f\n"
-      "ld1b { z29.b }, p2/Z, [x10]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z29.b }, p2/Z, [x9]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z29.b, z0.b[1]\n"
       "sdot z12.s, z29.b, z1.b[1]\n"
       "sdot z16.s, z29.b, z2.b[1]\n"
       "sdot z20.s, z29.b, z3.b[1]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z24.s, z29.b, z4.b[1]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z28.b, z0.b[1]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z13.s, z28.b, z1.b[1]\n"
       "sdot z17.s, z28.b, z2.b[1]\n"
       "sdot z21.s, z28.b, z3.b[1]\n"
       "sdot z25.s, z28.b, z4.b[1]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z29.b, z0.b[1]\n"
-      "addvl x10, x10, #4\n"
       "sdot z14.s, z29.b, z1.b[1]\n"
       "sdot z18.s, z29.b, z2.b[1]\n"
       "sdot z22.s, z29.b, z3.b[1]\n"
@@ -1653,23 +1653,23 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z23.s, z28.b, z3.b[1]\n"
       "sdot z27.s, z28.b, z4.b[1]\n"
       "ble 61f\n"
-      "ld1b { z29.b }, p2/Z, [x10]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z29.b }, p2/Z, [x9]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z29.b, z0.b[2]\n"
       "sdot z12.s, z29.b, z1.b[2]\n"
       "sdot z16.s, z29.b, z2.b[2]\n"
       "sdot z20.s, z29.b, z3.b[2]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z24.s, z29.b, z4.b[2]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z28.b, z0.b[2]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z13.s, z28.b, z1.b[2]\n"
       "sdot z17.s, z28.b, z2.b[2]\n"
       "sdot z21.s, z28.b, z3.b[2]\n"
       "sdot z25.s, z28.b, z4.b[2]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z29.b, z0.b[2]\n"
-      "addvl x10, x10, #4\n"
       "sdot z14.s, z29.b, z1.b[2]\n"
       "sdot z18.s, z29.b, z2.b[2]\n"
       "sdot z22.s, z29.b, z3.b[2]\n"
@@ -1680,21 +1680,21 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z23.s, z28.b, z3.b[2]\n"
       "sdot z27.s, z28.b, z4.b[2]\n"
       "ble 61f\n"
-      "ld1b { z29.b }, p2/Z, [x10]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z29.b }, p2/Z, [x9]\n"
+      "ld1b { z28.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z29.b, z0.b[3]\n"
       "sdot z12.s, z29.b, z1.b[3]\n"
       "sdot z16.s, z29.b, z2.b[3]\n"
       "sdot z20.s, z29.b, z3.b[3]\n"
       "sdot z24.s, z29.b, z4.b[3]\n"
-      "ld1b { z29.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z28.b, z0.b[3]\n"
+      "ld1b { z29.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z13.s, z28.b, z1.b[3]\n"
       "sdot z17.s, z28.b, z2.b[3]\n"
       "sdot z21.s, z28.b, z3.b[3]\n"
       "sdot z25.s, z28.b, z4.b[3]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z28.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z29.b, z0.b[3]\n"
       "sdot z14.s, z29.b, z1.b[3]\n"
       "sdot z18.s, z29.b, z2.b[3]\n"
@@ -1711,20 +1711,20 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 56b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x26, x11, x20\n"
       "ld1w { z31.s }, p2/Z, [x14]\n"
+      "add x25, x26, x20\n"
       "ld1w { z30.s }, p2/Z, [x14, #1, MUL VL]\n"
       "ld1w { z29.s }, p2/Z, [x14, #2, MUL VL]\n"
-      "ld1w { z28.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
-      "add z8.s, z8.s, z31.s\n"
-      "add z12.s, z12.s, z31.s\n"
       "add x24, x25, x20\n"
       "add x23, x24, x20\n"
+      "ld1w { z28.s }, p2/Z, [x14, #3, MUL VL]\n"
+      "add z8.s, z8.s, z31.s\n"
       "add z9.s, z9.s, z30.s\n"
+      "addvl x14, x14, #4\n"
       "add z10.s, z10.s, z29.s\n"
       "add z11.s, z11.s, z28.s\n"
+      "add z12.s, z12.s, z31.s\n"
       "add z13.s, z13.s, z30.s\n"
       "add z14.s, z14.s, z29.s\n"
       "add z15.s, z15.s, z28.s\n"
@@ -1794,11 +1794,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "sqadd z8.s, z8.s, z31.s\n"
-      "and z31.d, z12.d, z0.d\n"
       "sqadd z9.s, z9.s, z30.s\n"
-      "and z30.d, z13.d, z1.d\n"
       "sqadd z10.s, z10.s, z29.s\n"
       "sqadd z11.s, z11.s, z28.s\n"
+      "and z31.d, z12.d, z0.d\n"
+      "and z30.d, z13.d, z1.d\n"
       "and z29.d, z14.d, z2.d\n"
       "and z28.d, z15.d, z3.d\n"
       "asr z31.s, z31.s, #0x1f\n"
@@ -1806,11 +1806,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "sqadd z12.s, z12.s, z31.s\n"
-      "and z31.d, z16.d, z0.d\n"
       "sqadd z13.s, z13.s, z30.s\n"
-      "and z30.d, z17.d, z1.d\n"
       "sqadd z14.s, z14.s, z29.s\n"
       "sqadd z15.s, z15.s, z28.s\n"
+      "and z31.d, z16.d, z0.d\n"
+      "and z30.d, z17.d, z1.d\n"
       "and z29.d, z18.d, z2.d\n"
       "and z28.d, z19.d, z3.d\n"
       "asr z31.s, z31.s, #0x1f\n"
@@ -1818,11 +1818,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "sqadd z16.s, z16.s, z31.s\n"
-      "and z31.d, z20.d, z0.d\n"
       "sqadd z17.s, z17.s, z30.s\n"
-      "and z30.d, z21.d, z1.d\n"
       "sqadd z18.s, z18.s, z29.s\n"
       "sqadd z19.s, z19.s, z28.s\n"
+      "and z31.d, z20.d, z0.d\n"
+      "and z30.d, z21.d, z1.d\n"
       "and z29.d, z22.d, z2.d\n"
       "and z28.d, z23.d, z3.d\n"
       "asr z31.s, z31.s, #0x1f\n"
@@ -1830,11 +1830,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "sqadd z20.s, z20.s, z31.s\n"
-      "and z31.d, z24.d, z0.d\n"
       "sqadd z21.s, z21.s, z30.s\n"
-      "and z30.d, z25.d, z1.d\n"
       "sqadd z22.s, z22.s, z29.s\n"
       "sqadd z23.s, z23.s, z28.s\n"
+      "and z31.d, z24.d, z0.d\n"
+      "and z30.d, z25.d, z1.d\n"
       "and z29.d, z26.d, z2.d\n"
       "and z28.d, z27.d, z3.d\n"
       "asr z31.s, z31.s, #0x1f\n"
@@ -1847,51 +1847,51 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sqadd z27.s, z27.s, z28.s\n"
       "64:"  // Height 5: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
       "ld1rw { z29.s }, p2/Z, [x20]\n"
+      ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
+      "add z8.s, z8.s, z29.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
+      "add z9.s, z9.s, z29.s\n"
+      "add z10.s, z10.s, z29.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x4482880c  // srshl z12.s, p2/M, z12.s, z0.s\n"
+      "add z11.s, z11.s, z29.s\n"
+      "add z12.s, z12.s, z29.s\n"
       ".inst 0x4482882d  // srshl z13.s, p2/M, z13.s, z1.s\n"
       ".inst 0x4482884e  // srshl z14.s, p2/M, z14.s, z2.s\n"
-      "add z8.s, z8.s, z29.s\n"
+      "add z13.s, z13.s, z29.s\n"
+      "add z14.s, z14.s, z29.s\n"
       ".inst 0x4482886f  // srshl z15.s, p2/M, z15.s, z3.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "add z9.s, z9.s, z29.s\n"
-      "add z10.s, z10.s, z29.s\n"
+      "add z15.s, z15.s, z29.s\n"
+      "add z16.s, z16.s, z29.s\n"
       ".inst 0x44828831  // srshl z17.s, p2/M, z17.s, z1.s\n"
       ".inst 0x44828852  // srshl z18.s, p2/M, z18.s, z2.s\n"
-      "add z11.s, z11.s, z29.s\n"
-      "add z12.s, z12.s, z29.s\n"
+      "add z17.s, z17.s, z29.s\n"
+      "add z18.s, z18.s, z29.s\n"
       ".inst 0x44828873  // srshl z19.s, p2/M, z19.s, z3.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      "add z13.s, z13.s, z29.s\n"
-      "add z14.s, z14.s, z29.s\n"
+      "add z19.s, z19.s, z29.s\n"
+      "add z20.s, z20.s, z29.s\n"
       ".inst 0x44828835  // srshl z21.s, p2/M, z21.s, z1.s\n"
       ".inst 0x44828856  // srshl z22.s, p2/M, z22.s, z2.s\n"
-      "add z15.s, z15.s, z29.s\n"
-      "add z16.s, z16.s, z29.s\n"
+      "add z21.s, z21.s, z29.s\n"
+      "add z22.s, z22.s, z29.s\n"
       ".inst 0x44828877  // srshl z23.s, p2/M, z23.s, z3.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z29.s\n"
-      "add z18.s, z18.s, z29.s\n"
+      "add z23.s, z23.s, z29.s\n"
+      "add z24.s, z24.s, z29.s\n"
       ".inst 0x44828839  // srshl z25.s, p2/M, z25.s, z1.s\n"
       ".inst 0x4482885a  // srshl z26.s, p2/M, z26.s, z2.s\n"
-      "add z19.s, z19.s, z29.s\n"
-      "add z20.s, z20.s, z29.s\n"
+      "add z25.s, z25.s, z29.s\n"
+      "add z26.s, z26.s, z29.s\n"
       ".inst 0x4482887b  // srshl z27.s, p2/M, z27.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z21.s, z21.s, z29.s\n"
-      "add z22.s, z22.s, z29.s\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z29.s\n"
-      "add z24.s, z24.s, z29.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z29.s\n"
-      "add z26.s, z26.s, z29.s\n"
-      "ld1rw { z30.s }, p2/Z, [x20]\n"
       "add z27.s, z27.s, z29.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z29.s }, p2/Z, [x20]\n"
       "smin z8.s, p2/M, z8.s, z28.s\n"
       "smin z9.s, p2/M, z9.s, z28.s\n"
       "smin z10.s, p2/M, z10.s, z28.s\n"
@@ -1912,72 +1912,71 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "smin z25.s, p2/M, z25.s, z28.s\n"
       "smin z26.s, p2/M, z26.s, z28.s\n"
       "smin z27.s, p2/M, z27.s, z28.s\n"
-      "smax z8.s, p2/M, z8.s, z30.s\n"
-      "smax z9.s, p2/M, z9.s, z30.s\n"
-      "smax z10.s, p2/M, z10.s, z30.s\n"
-      "smax z11.s, p2/M, z11.s, z30.s\n"
-      "smax z12.s, p2/M, z12.s, z30.s\n"
-      "smax z13.s, p2/M, z13.s, z30.s\n"
-      "smax z14.s, p2/M, z14.s, z30.s\n"
-      "smax z15.s, p2/M, z15.s, z30.s\n"
+      "smax z8.s, p2/M, z8.s, z29.s\n"
+      "smax z9.s, p2/M, z9.s, z29.s\n"
+      "smax z10.s, p2/M, z10.s, z29.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
-      "smax z16.s, p2/M, z16.s, z30.s\n"
-      "smax z17.s, p2/M, z17.s, z30.s\n"
-      "uzp1 z29.h, z10.h, z11.h\n"
-      "smax z18.s, p2/M, z18.s, z30.s\n"
-      "smax z19.s, p2/M, z19.s, z30.s\n"
+      "smax z11.s, p2/M, z11.s, z29.s\n"
+      "smax z12.s, p2/M, z12.s, z29.s\n"
+      "uzp1 z28.h, z10.h, z11.h\n"
+      "uzp1 z8.b, z8.b, z28.b\n"
+      "smax z13.s, p2/M, z13.s, z29.s\n"
+      "smax z14.s, p2/M, z14.s, z29.s\n"
       "uzp1 z12.h, z12.h, z13.h\n"
-      "smax z20.s, p2/M, z20.s, z30.s\n"
-      "smax z21.s, p2/M, z21.s, z30.s\n"
+      "st1b { z8.b }, p1, [x11]\n"
+      "smax z15.s, p2/M, z15.s, z29.s\n"
+      "smax z16.s, p2/M, z16.s, z29.s\n"
       "uzp1 z28.h, z14.h, z15.h\n"
-      "smax z22.s, p2/M, z22.s, z30.s\n"
-      "smax z23.s, p2/M, z23.s, z30.s\n"
+      "uzp1 z12.b, z12.b, z28.b\n"
+      "smax z17.s, p2/M, z17.s, z29.s\n"
+      "smax z18.s, p2/M, z18.s, z29.s\n"
       "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z8.b, z8.b, z29.b\n"
-      "smax z24.s, p2/M, z24.s, z30.s\n"
-      "smax z25.s, p2/M, z25.s, z30.s\n"
+      "st1b { z12.b }, p1, [x26]\n"
+      "smax z19.s, p2/M, z19.s, z29.s\n"
+      "smax z20.s, p2/M, z20.s, z29.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
-      "smax z26.s, p2/M, z26.s, z30.s\n"
-      "smax z27.s, p2/M, z27.s, z30.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z12.b, z12.b, z28.b\n"
-      "uzp1 z18.h, z22.h, z23.h\n"
-      "st1b { z8.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
       "uzp1 z16.b, z16.b, z17.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "st1b { z12.b }, p1, [x26]\n"
-      "uzp1 z20.b, z20.b, z18.b\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
+      "smax z21.s, p2/M, z21.s, z29.s\n"
+      "smax z22.s, p2/M, z22.s, z29.s\n"
+      "uzp1 z20.h, z20.h, z21.h\n"
       "st1b { z16.b }, p1, [x25]\n"
+      "smax z23.s, p2/M, z23.s, z29.s\n"
+      "smax z24.s, p2/M, z24.s, z29.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
+      "smax z25.s, p2/M, z25.s, z29.s\n"
+      "smax z26.s, p2/M, z26.s, z29.s\n"
+      "uzp1 z24.h, z24.h, z25.h\n"
       "st1b { z20.b }, p1, [x24]\n"
+      "smax z27.s, p2/M, z27.s, z29.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "st1b { z24.b }, p1, [x23]\n"
+      "addvl x11, x11, #1\n"
       "65:"  // Height 5: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 54b\n"
       "b 80f\n"
       "66:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x6\n"
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "67:"  // Height 6: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "mov z16.s, #0x0\n"
@@ -2000,8 +1999,8 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "mov x28, #0x0\n"
       "69:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 70f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -2032,121 +2031,121 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "ble 73f\n"
       "72:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p2/Z, [x10]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z7.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z6.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z5.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "ld1rqb { z2.b }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1b { z1.b }, p2/Z, [x9]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z1.b, z7.b[0]\n"
       "sdot z12.s, z1.b, z6.b[0]\n"
-      "add x21, x21, #0x10\n"
       "sdot z16.s, z1.b, z5.b[0]\n"
       "sdot z20.s, z1.b, z4.b[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "sdot z24.s, z1.b, z3.b[0]\n"
       "sdot z28.s, z1.b, z2.b[0]\n"
-      "ld1b { z1.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z1.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "add x21, x21, #0x10\n"
       "sdot z9.s, z0.b, z7.b[0]\n"
       "sdot z13.s, z0.b, z6.b[0]\n"
       "sdot z17.s, z0.b, z5.b[0]\n"
       "sdot z21.s, z0.b, z4.b[0]\n"
       "sdot z25.s, z0.b, z3.b[0]\n"
       "sdot z29.s, z0.b, z2.b[0]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
       "sdot z10.s, z1.b, z7.b[0]\n"
       "sdot z14.s, z1.b, z6.b[0]\n"
       "sdot z18.s, z1.b, z5.b[0]\n"
       "sdot z22.s, z1.b, z4.b[0]\n"
       "sdot z26.s, z1.b, z3.b[0]\n"
       "sdot z30.s, z1.b, z2.b[0]\n"
-      "ld1b { z1.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z1.b }, p2/Z, [x9, #4, MUL VL]\n"
       "sdot z11.s, z0.b, z7.b[0]\n"
       "sdot z15.s, z0.b, z6.b[0]\n"
       "sdot z19.s, z0.b, z5.b[0]\n"
       "sdot z23.s, z0.b, z4.b[0]\n"
       "sdot z27.s, z0.b, z3.b[0]\n"
       "sdot z31.s, z0.b, z2.b[0]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
       "sdot z8.s, z1.b, z7.b[1]\n"
       "sdot z12.s, z1.b, z6.b[1]\n"
       "sdot z16.s, z1.b, z5.b[1]\n"
       "sdot z20.s, z1.b, z4.b[1]\n"
       "sdot z24.s, z1.b, z3.b[1]\n"
       "sdot z28.s, z1.b, z2.b[1]\n"
-      "ld1b { z1.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z1.b }, p2/Z, [x9, #6, MUL VL]\n"
       "sdot z9.s, z0.b, z7.b[1]\n"
       "sdot z13.s, z0.b, z6.b[1]\n"
       "sdot z17.s, z0.b, z5.b[1]\n"
       "sdot z21.s, z0.b, z4.b[1]\n"
       "sdot z25.s, z0.b, z3.b[1]\n"
       "sdot z29.s, z0.b, z2.b[1]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
+      "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
       "sdot z10.s, z1.b, z7.b[1]\n"
       "sdot z14.s, z1.b, z6.b[1]\n"
       "sdot z18.s, z1.b, z5.b[1]\n"
       "sdot z22.s, z1.b, z4.b[1]\n"
       "sdot z26.s, z1.b, z3.b[1]\n"
       "sdot z30.s, z1.b, z2.b[1]\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-8, MUL VL]\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-8, MUL VL]\n"
       "sdot z11.s, z0.b, z7.b[1]\n"
       "sdot z15.s, z0.b, z6.b[1]\n"
       "sdot z19.s, z0.b, z5.b[1]\n"
       "sdot z23.s, z0.b, z4.b[1]\n"
       "sdot z27.s, z0.b, z3.b[1]\n"
       "sdot z31.s, z0.b, z2.b[1]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-7, MUL VL]\n"
       "sdot z8.s, z1.b, z7.b[2]\n"
       "sdot z12.s, z1.b, z6.b[2]\n"
       "sdot z16.s, z1.b, z5.b[2]\n"
       "sdot z20.s, z1.b, z4.b[2]\n"
       "sdot z24.s, z1.b, z3.b[2]\n"
       "sdot z28.s, z1.b, z2.b[2]\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-6, MUL VL]\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-6, MUL VL]\n"
       "sdot z9.s, z0.b, z7.b[2]\n"
       "sdot z13.s, z0.b, z6.b[2]\n"
       "sdot z17.s, z0.b, z5.b[2]\n"
       "sdot z21.s, z0.b, z4.b[2]\n"
       "sdot z25.s, z0.b, z3.b[2]\n"
       "sdot z29.s, z0.b, z2.b[2]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-5, MUL VL]\n"
       "sdot z10.s, z1.b, z7.b[2]\n"
       "sdot z14.s, z1.b, z6.b[2]\n"
       "sdot z18.s, z1.b, z5.b[2]\n"
       "sdot z22.s, z1.b, z4.b[2]\n"
       "sdot z26.s, z1.b, z3.b[2]\n"
       "sdot z30.s, z1.b, z2.b[2]\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-4, MUL VL]\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-4, MUL VL]\n"
       "sdot z11.s, z0.b, z7.b[2]\n"
       "sdot z15.s, z0.b, z6.b[2]\n"
       "sdot z19.s, z0.b, z5.b[2]\n"
       "sdot z23.s, z0.b, z4.b[2]\n"
       "sdot z27.s, z0.b, z3.b[2]\n"
       "sdot z31.s, z0.b, z2.b[2]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-3, MUL VL]\n"
       "sdot z8.s, z1.b, z7.b[3]\n"
       "sdot z12.s, z1.b, z6.b[3]\n"
       "sdot z16.s, z1.b, z5.b[3]\n"
       "sdot z20.s, z1.b, z4.b[3]\n"
       "sdot z24.s, z1.b, z3.b[3]\n"
       "sdot z28.s, z1.b, z2.b[3]\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-2, MUL VL]\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-2, MUL VL]\n"
       "sdot z9.s, z0.b, z7.b[3]\n"
       "sdot z13.s, z0.b, z6.b[3]\n"
       "sdot z17.s, z0.b, z5.b[3]\n"
       "sdot z21.s, z0.b, z4.b[3]\n"
       "sdot z25.s, z0.b, z3.b[3]\n"
       "sdot z29.s, z0.b, z2.b[3]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-1, MUL VL]\n"
       "sdot z10.s, z1.b, z7.b[3]\n"
       "sdot z14.s, z1.b, z6.b[3]\n"
       "sdot z18.s, z1.b, z5.b[3]\n"
@@ -2162,30 +2161,30 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "bgt 72b\n"
       "73:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z7.b }, p2/Z, [x10]\n"
-      "ld1b { z6.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
       "ld1rqb { z4.b }, p0/Z, [x22]\n"
       "ld1rqb { z5.b }, p0/Z, [x21]\n"
+      "ld1b { z7.b }, p2/Z, [x9]\n"
+      "ld1b { z6.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z7.b, z0.b[0]\n"
       "sdot z12.s, z7.b, z1.b[0]\n"
-      "sdot z9.s, z6.b, z0.b[0]\n"
-      "sdot z13.s, z6.b, z1.b[0]\n"
       "sdot z16.s, z7.b, z2.b[0]\n"
       "sdot z20.s, z7.b, z3.b[0]\n"
       "sdot z24.s, z7.b, z4.b[0]\n"
       "sdot z28.s, z7.b, z5.b[0]\n"
-      "ld1b { z7.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "sdot z9.s, z6.b, z0.b[0]\n"
+      "sdot z13.s, z6.b, z1.b[0]\n"
       "sdot z17.s, z6.b, z2.b[0]\n"
       "sdot z21.s, z6.b, z3.b[0]\n"
       "sdot z25.s, z6.b, z4.b[0]\n"
       "sdot z29.s, z6.b, z5.b[0]\n"
-      "ld1b { z6.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z6.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z7.b, z0.b[0]\n"
       "sdot z14.s, z7.b, z1.b[0]\n"
       "sdot z18.s, z7.b, z2.b[0]\n"
@@ -2199,25 +2198,25 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z27.s, z6.b, z4.b[0]\n"
       "sdot z31.s, z6.b, z5.b[0]\n"
       "ble 74f\n"
-      "ld1b { z7.b }, p2/Z, [x10]\n"
-      "ld1b { z6.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z7.b }, p2/Z, [x9]\n"
+      "ld1b { z6.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z7.b, z0.b[1]\n"
       "sdot z12.s, z7.b, z1.b[1]\n"
       "sdot z16.s, z7.b, z2.b[1]\n"
       "sdot z20.s, z7.b, z3.b[1]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z24.s, z7.b, z4.b[1]\n"
       "sdot z28.s, z7.b, z5.b[1]\n"
+      "ld1b { z7.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z9.s, z6.b, z0.b[1]\n"
-      "ld1b { z7.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z13.s, z6.b, z1.b[1]\n"
       "sdot z17.s, z6.b, z2.b[1]\n"
       "sdot z21.s, z6.b, z3.b[1]\n"
       "sdot z25.s, z6.b, z4.b[1]\n"
       "sdot z29.s, z6.b, z5.b[1]\n"
-      "ld1b { z6.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z6.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z7.b, z0.b[1]\n"
-      "addvl x10, x10, #4\n"
       "sdot z14.s, z7.b, z1.b[1]\n"
       "sdot z18.s, z7.b, z2.b[1]\n"
       "sdot z22.s, z7.b, z3.b[1]\n"
@@ -2230,25 +2229,25 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z27.s, z6.b, z4.b[1]\n"
       "sdot z31.s, z6.b, z5.b[1]\n"
       "ble 74f\n"
-      "ld1b { z7.b }, p2/Z, [x10]\n"
-      "ld1b { z6.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
+      "ld1b { z7.b }, p2/Z, [x9]\n"
+      "ld1b { z6.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z7.b, z0.b[2]\n"
       "sdot z12.s, z7.b, z1.b[2]\n"
       "sdot z16.s, z7.b, z2.b[2]\n"
       "sdot z20.s, z7.b, z3.b[2]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z24.s, z7.b, z4.b[2]\n"
       "sdot z28.s, z7.b, z5.b[2]\n"
+      "ld1b { z7.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z9.s, z6.b, z0.b[2]\n"
-      "ld1b { z7.b }, p2/Z, [x10, #2, MUL VL]\n"
       "sdot z13.s, z6.b, z1.b[2]\n"
       "sdot z17.s, z6.b, z2.b[2]\n"
       "sdot z21.s, z6.b, z3.b[2]\n"
       "sdot z25.s, z6.b, z4.b[2]\n"
       "sdot z29.s, z6.b, z5.b[2]\n"
-      "ld1b { z6.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z6.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z7.b, z0.b[2]\n"
-      "addvl x10, x10, #4\n"
       "sdot z14.s, z7.b, z1.b[2]\n"
       "sdot z18.s, z7.b, z2.b[2]\n"
       "sdot z22.s, z7.b, z3.b[2]\n"
@@ -2261,23 +2260,23 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sdot z27.s, z6.b, z4.b[2]\n"
       "sdot z31.s, z6.b, z5.b[2]\n"
       "ble 74f\n"
-      "ld1b { z7.b }, p2/Z, [x10]\n"
-      "ld1b { z6.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x9]\n"
+      "ld1b { z6.b }, p2/Z, [x9, #1, MUL VL]\n"
       "sdot z8.s, z7.b, z0.b[3]\n"
       "sdot z12.s, z7.b, z1.b[3]\n"
       "sdot z16.s, z7.b, z2.b[3]\n"
       "sdot z20.s, z7.b, z3.b[3]\n"
       "sdot z24.s, z7.b, z4.b[3]\n"
       "sdot z28.s, z7.b, z5.b[3]\n"
-      "ld1b { z7.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sdot z9.s, z6.b, z0.b[3]\n"
       "sdot z13.s, z6.b, z1.b[3]\n"
       "sdot z17.s, z6.b, z2.b[3]\n"
       "sdot z21.s, z6.b, z3.b[3]\n"
       "sdot z25.s, z6.b, z4.b[3]\n"
       "sdot z29.s, z6.b, z5.b[3]\n"
-      "ld1b { z6.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "ld1b { z6.b }, p2/Z, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "sdot z10.s, z7.b, z0.b[3]\n"
       "sdot z14.s, z7.b, z1.b[3]\n"
       "sdot z18.s, z7.b, z2.b[3]\n"
@@ -2296,21 +2295,21 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 69b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x26, x11, x20\n"
+      "add x25, x26, x20\n"
       "ld1w { z3.s }, p2/Z, [x14]\n"
       "ld1w { z2.s }, p2/Z, [x14, #1, MUL VL]\n"
       "ld1w { z1.s }, p2/Z, [x14, #2, MUL VL]\n"
-      "ld1w { z0.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
-      "add x26, x9, x20\n"
-      "add x25, x26, x20\n"
-      "add z8.s, z8.s, z3.s\n"
-      "add z12.s, z12.s, z3.s\n"
       "add x24, x25, x20\n"
       "add x23, x24, x20\n"
+      "ld1w { z0.s }, p2/Z, [x14, #3, MUL VL]\n"
+      "add x22, x23, x20\n"
+      "add z8.s, z8.s, z3.s\n"
       "add z9.s, z9.s, z2.s\n"
       "add z10.s, z10.s, z1.s\n"
-      "add x22, x23, x20\n"
       "add z11.s, z11.s, z0.s\n"
+      "addvl x14, x14, #4\n"
+      "add z12.s, z12.s, z3.s\n"
       "add z13.s, z13.s, z2.s\n"
       "add z14.s, z14.s, z1.s\n"
       "add z15.s, z15.s, z0.s\n"
@@ -2388,11 +2387,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z8.s, z8.s, z7.s\n"
-      "and z7.d, z12.d, z0.d\n"
       "sqadd z9.s, z9.s, z6.s\n"
-      "and z6.d, z13.d, z1.d\n"
       "sqadd z10.s, z10.s, z5.s\n"
       "sqadd z11.s, z11.s, z4.s\n"
+      "and z7.d, z12.d, z0.d\n"
+      "and z6.d, z13.d, z1.d\n"
       "and z5.d, z14.d, z2.d\n"
       "and z4.d, z15.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2400,11 +2399,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z12.s, z12.s, z7.s\n"
-      "and z7.d, z16.d, z0.d\n"
       "sqadd z13.s, z13.s, z6.s\n"
-      "and z6.d, z17.d, z1.d\n"
       "sqadd z14.s, z14.s, z5.s\n"
       "sqadd z15.s, z15.s, z4.s\n"
+      "and z7.d, z16.d, z0.d\n"
+      "and z6.d, z17.d, z1.d\n"
       "and z5.d, z18.d, z2.d\n"
       "and z4.d, z19.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2412,11 +2411,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z16.s, z16.s, z7.s\n"
-      "and z7.d, z20.d, z0.d\n"
       "sqadd z17.s, z17.s, z6.s\n"
-      "and z6.d, z21.d, z1.d\n"
       "sqadd z18.s, z18.s, z5.s\n"
       "sqadd z19.s, z19.s, z4.s\n"
+      "and z7.d, z20.d, z0.d\n"
+      "and z6.d, z21.d, z1.d\n"
       "and z5.d, z22.d, z2.d\n"
       "and z4.d, z23.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2424,11 +2423,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z20.s, z20.s, z7.s\n"
-      "and z7.d, z24.d, z0.d\n"
       "sqadd z21.s, z21.s, z6.s\n"
-      "and z6.d, z25.d, z1.d\n"
       "sqadd z22.s, z22.s, z5.s\n"
       "sqadd z23.s, z23.s, z4.s\n"
+      "and z7.d, z24.d, z0.d\n"
+      "and z6.d, z25.d, z1.d\n"
       "and z5.d, z26.d, z2.d\n"
       "and z4.d, z27.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2436,11 +2435,11 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z24.s, z24.s, z7.s\n"
-      "and z7.d, z28.d, z0.d\n"
       "sqadd z25.s, z25.s, z6.s\n"
-      "and z6.d, z29.d, z1.d\n"
       "sqadd z26.s, z26.s, z5.s\n"
       "sqadd z27.s, z27.s, z4.s\n"
+      "and z7.d, z28.d, z0.d\n"
+      "and z6.d, z29.d, z1.d\n"
       "and z5.d, z30.d, z2.d\n"
       "and z4.d, z31.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2453,59 +2452,59 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "sqadd z31.s, z31.s, z4.s\n"
       "77:"  // Height 6: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
       "ld1rw { z4.s }, p2/Z, [x20]\n"
+      ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
+      "add z8.s, z8.s, z4.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
+      "add z9.s, z9.s, z4.s\n"
+      "add z10.s, z10.s, z4.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x4482880c  // srshl z12.s, p2/M, z12.s, z0.s\n"
+      "add z11.s, z11.s, z4.s\n"
+      "add z12.s, z12.s, z4.s\n"
       ".inst 0x4482882d  // srshl z13.s, p2/M, z13.s, z1.s\n"
       ".inst 0x4482884e  // srshl z14.s, p2/M, z14.s, z2.s\n"
-      "add z8.s, z8.s, z4.s\n"
+      "add z13.s, z13.s, z4.s\n"
+      "add z14.s, z14.s, z4.s\n"
       ".inst 0x4482886f  // srshl z15.s, p2/M, z15.s, z3.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "add z9.s, z9.s, z4.s\n"
-      "add z10.s, z10.s, z4.s\n"
+      "add z15.s, z15.s, z4.s\n"
+      "add z16.s, z16.s, z4.s\n"
       ".inst 0x44828831  // srshl z17.s, p2/M, z17.s, z1.s\n"
       ".inst 0x44828852  // srshl z18.s, p2/M, z18.s, z2.s\n"
-      "add z11.s, z11.s, z4.s\n"
-      "add z12.s, z12.s, z4.s\n"
+      "add z17.s, z17.s, z4.s\n"
+      "add z18.s, z18.s, z4.s\n"
       ".inst 0x44828873  // srshl z19.s, p2/M, z19.s, z3.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      "add z13.s, z13.s, z4.s\n"
-      "add z14.s, z14.s, z4.s\n"
+      "add z19.s, z19.s, z4.s\n"
+      "add z20.s, z20.s, z4.s\n"
       ".inst 0x44828835  // srshl z21.s, p2/M, z21.s, z1.s\n"
       ".inst 0x44828856  // srshl z22.s, p2/M, z22.s, z2.s\n"
-      "add z15.s, z15.s, z4.s\n"
-      "add z16.s, z16.s, z4.s\n"
+      "add z21.s, z21.s, z4.s\n"
+      "add z22.s, z22.s, z4.s\n"
       ".inst 0x44828877  // srshl z23.s, p2/M, z23.s, z3.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z4.s\n"
-      "add z18.s, z18.s, z4.s\n"
+      "add z23.s, z23.s, z4.s\n"
+      "add z24.s, z24.s, z4.s\n"
       ".inst 0x44828839  // srshl z25.s, p2/M, z25.s, z1.s\n"
       ".inst 0x4482885a  // srshl z26.s, p2/M, z26.s, z2.s\n"
-      "add z19.s, z19.s, z4.s\n"
-      "add z20.s, z20.s, z4.s\n"
+      "add z25.s, z25.s, z4.s\n"
+      "add z26.s, z26.s, z4.s\n"
       ".inst 0x4482887b  // srshl z27.s, p2/M, z27.s, z3.s\n"
       ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
-      "add z21.s, z21.s, z4.s\n"
-      "add z22.s, z22.s, z4.s\n"
+      "add z27.s, z27.s, z4.s\n"
+      "add z28.s, z28.s, z4.s\n"
       ".inst 0x4482883d  // srshl z29.s, p2/M, z29.s, z1.s\n"
       ".inst 0x4482885e  // srshl z30.s, p2/M, z30.s, z2.s\n"
-      "add z23.s, z23.s, z4.s\n"
-      "add z24.s, z24.s, z4.s\n"
+      "add z29.s, z29.s, z4.s\n"
+      "add z30.s, z30.s, z4.s\n"
       ".inst 0x4482887f  // srshl z31.s, p2/M, z31.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z25.s, z25.s, z4.s\n"
-      "add z26.s, z26.s, z4.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z4.s\n"
-      "add z28.s, z28.s, z4.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z29.s, z29.s, z4.s\n"
-      "add z30.s, z30.s, z4.s\n"
-      "ld1rw { z2.s }, p2/Z, [x20]\n"
       "add z31.s, z31.s, z4.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z1.s }, p2/Z, [x20]\n"
       "smin z8.s, p2/M, z8.s, z0.s\n"
       "smin z9.s, p2/M, z9.s, z0.s\n"
       "smin z10.s, p2/M, z10.s, z0.s\n"
@@ -2530,58 +2529,58 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "smin z29.s, p2/M, z29.s, z0.s\n"
       "smin z30.s, p2/M, z30.s, z0.s\n"
       "smin z31.s, p2/M, z31.s, z0.s\n"
-      "smax z8.s, p2/M, z8.s, z2.s\n"
-      "smax z9.s, p2/M, z9.s, z2.s\n"
-      "smax z10.s, p2/M, z10.s, z2.s\n"
-      "smax z11.s, p2/M, z11.s, z2.s\n"
-      "smax z12.s, p2/M, z12.s, z2.s\n"
-      "smax z13.s, p2/M, z13.s, z2.s\n"
-      "smax z14.s, p2/M, z14.s, z2.s\n"
-      "smax z15.s, p2/M, z15.s, z2.s\n"
+      "smax z8.s, p2/M, z8.s, z1.s\n"
+      "smax z9.s, p2/M, z9.s, z1.s\n"
+      "smax z10.s, p2/M, z10.s, z1.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
-      "smax z16.s, p2/M, z16.s, z2.s\n"
-      "smax z17.s, p2/M, z17.s, z2.s\n"
-      "uzp1 z1.h, z10.h, z11.h\n"
-      "smax z18.s, p2/M, z18.s, z2.s\n"
-      "smax z19.s, p2/M, z19.s, z2.s\n"
+      "smax z11.s, p2/M, z11.s, z1.s\n"
+      "smax z12.s, p2/M, z12.s, z1.s\n"
+      "uzp1 z0.h, z10.h, z11.h\n"
+      "uzp1 z8.b, z8.b, z0.b\n"
+      "smax z13.s, p2/M, z13.s, z1.s\n"
+      "smax z14.s, p2/M, z14.s, z1.s\n"
       "uzp1 z12.h, z12.h, z13.h\n"
-      "smax z20.s, p2/M, z20.s, z2.s\n"
-      "smax z21.s, p2/M, z21.s, z2.s\n"
+      "st1b { z8.b }, p1, [x11]\n"
+      "smax z15.s, p2/M, z15.s, z1.s\n"
+      "smax z16.s, p2/M, z16.s, z1.s\n"
       "uzp1 z0.h, z14.h, z15.h\n"
-      "smax z22.s, p2/M, z22.s, z2.s\n"
-      "smax z23.s, p2/M, z23.s, z2.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z8.b, z8.b, z1.b\n"
-      "smax z24.s, p2/M, z24.s, z2.s\n"
-      "smax z25.s, p2/M, z25.s, z2.s\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
-      "smax z26.s, p2/M, z26.s, z2.s\n"
-      "smax z27.s, p2/M, z27.s, z2.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
       "uzp1 z12.b, z12.b, z0.b\n"
-      "smax z28.s, p2/M, z28.s, z2.s\n"
-      "smax z29.s, p2/M, z29.s, z2.s\n"
-      "uzp1 z17.h, z22.h, z23.h\n"
-      "st1b { z8.b }, p1, [x9]\n"
-      "smax z30.s, p2/M, z30.s, z2.s\n"
-      "smax z31.s, p2/M, z31.s, z2.s\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z18.h, z26.h, z27.h\n"
+      "smax z17.s, p2/M, z17.s, z1.s\n"
+      "smax z18.s, p2/M, z18.s, z1.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "st1b { z12.b }, p1, [x26]\n"
-      "addvl x9, x9, #1\n"
-      "uzp1 z28.h, z28.h, z29.h\n"
-      "uzp1 z20.b, z20.b, z17.b\n"
-      "uzp1 z17.h, z30.h, z31.h\n"
+      "smax z19.s, p2/M, z19.s, z1.s\n"
+      "smax z20.s, p2/M, z20.s, z1.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
+      "smax z21.s, p2/M, z21.s, z1.s\n"
+      "smax z22.s, p2/M, z22.s, z1.s\n"
+      "uzp1 z20.h, z20.h, z21.h\n"
       "st1b { z16.b }, p1, [x25]\n"
-      "uzp1 z24.b, z24.b, z18.b\n"
-      "uzp1 z28.b, z28.b, z17.b\n"
+      "smax z23.s, p2/M, z23.s, z1.s\n"
+      "smax z24.s, p2/M, z24.s, z1.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
+      "smax z25.s, p2/M, z25.s, z1.s\n"
+      "smax z26.s, p2/M, z26.s, z1.s\n"
+      "uzp1 z24.h, z24.h, z25.h\n"
       "st1b { z20.b }, p1, [x24]\n"
+      "smax z27.s, p2/M, z27.s, z1.s\n"
+      "smax z28.s, p2/M, z28.s, z1.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
+      "smax z29.s, p2/M, z29.s, z1.s\n"
+      "smax z30.s, p2/M, z30.s, z1.s\n"
+      "uzp1 z28.h, z28.h, z29.h\n"
       "st1b { z24.b }, p1, [x23]\n"
+      "smax z31.s, p2/M, z31.s, z1.s\n"
+      "uzp1 z16.h, z30.h, z31.h\n"
+      "uzp1 z28.b, z28.b, z16.b\n"
       "st1b { z28.b }, p1, [x22]\n"
+      "addvl x11, x11, #1\n"
       "78:"  // Height 6: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 67b\n"
       "subs %x[M], %x[M], #0x6\n"
       "beq 80f\n"
@@ -2595,8 +2594,8 @@ void sve_hybrid_s8qs_dot_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL.hpp
index 73e4bd32b9..11ff5b2f15 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 8, 8> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 8, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp
index 759e3e2f3d..1a1201310e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8qs_mmla_6x4VL/generic.cpp
@@ -47,18 +47,18 @@ void sve_hybrid_s8qs_mmla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -98,26 +98,26 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "3:"  // Height 1: setup done
       "mov x28, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -133,87 +133,87 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "ble 8f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z19.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "trn1 z18.d, z19.d, z22.d\n"
-      "trn2 z19.d, z19.d, z22.d\n"
+      "ld1rqb { z20.b }, p0/Z, [x26]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45119a48  // smmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45109a4c  // smmla z12.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45119a49  // smmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45109a4d  // smmla z13.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x45119a4a  // smmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45109a4e  // smmla z14.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
+      "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
+      "trn2 z20.d, z20.d, z19.d\n"
       ".inst 0x45119a4b  // smmla z11.s, z18.b, z17.b\n"
       ".inst 0x45109a4f  // smmla z15.s, z18.b, z16.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45119a68  // smmla z8.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45109a6c  // smmla z12.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45119a69  // smmla z9.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45109a6d  // smmla z13.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45119a6a  // smmla z10.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45109a6e  // smmla z14.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45119a6b  // smmla z11.s, z19.b, z17.b\n"
-      ".inst 0x45109a6f  // smmla z15.s, z19.b, z16.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-8, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-7, MUL VL]\n"
+      ".inst 0x45119a88  // smmla z8.s, z20.b, z17.b\n"
+      ".inst 0x45109a8c  // smmla z12.s, z20.b, z16.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-5, MUL VL]\n"
+      ".inst 0x45119a89  // smmla z9.s, z20.b, z17.b\n"
+      ".inst 0x45109a8d  // smmla z13.s, z20.b, z16.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-3, MUL VL]\n"
+      ".inst 0x45119a8a  // smmla z10.s, z20.b, z17.b\n"
+      ".inst 0x45109a8e  // smmla z14.s, z20.b, z16.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-1, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45119a8b  // smmla z11.s, z20.b, z17.b\n"
+      ".inst 0x45109a8f  // smmla z15.s, z20.b, z16.b\n"
+      "add x26, x26, #0x10\n"
       "bgt 7b\n"
       "8:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45119a48  // smmla z8.s, z18.b, z17.b\n"
-      "ld1b { z7.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45109a4c  // smmla z12.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
-      ".inst 0x45079a49  // smmla z9.s, z18.b, z7.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
+      ".inst 0x45119a49  // smmla z9.s, z18.b, z17.b\n"
       ".inst 0x45109a4d  // smmla z13.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x45119a4a  // smmla z10.s, z18.b, z17.b\n"
-      "ld1b { z26.b }, p2/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45109a4e  // smmla z14.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x451a9a4b  // smmla z11.s, z18.b, z26.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
+      "trn2 z1.d, z1.d, z19.d\n"
+      ".inst 0x45119a4b  // smmla z11.s, z18.b, z17.b\n"
       ".inst 0x45109a4f  // smmla z15.s, z18.b, z16.b\n"
+      "addvl x9, x9, #8\n"
       "ble 9f\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45119828  // smmla z8.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4510982c  // smmla z12.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45119829  // smmla z9.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #4, MUL VL]\n"
       ".inst 0x4510982d  // smmla z13.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x4511982a  // smmla z10.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #6, MUL VL]\n"
       ".inst 0x4510982e  // smmla z14.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
       ".inst 0x4511982b  // smmla z11.s, z1.b, z17.b\n"
       ".inst 0x4510982f  // smmla z15.s, z1.b, z16.b\n"
+      "addvl x9, x9, #8\n"
       "9:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -227,10 +227,10 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "uzp1 z11.d, z11.d, z15.d\n"
       "ld1w { z17.s }, p2/Z, [x14, #2, MUL VL]\n"
       "ld1w { z16.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
       "mov z15.d, z8.d\n"
-      "add z9.s, z9.s, z18.s\n"
       "add z15.s, z15.s, z19.s\n"
+      "addvl x14, x14, #4\n"
+      "add z9.s, z9.s, z18.s\n"
       "add z10.s, z10.s, z17.s\n"
       "add z11.s, z11.s, z16.s\n"
       "tbz %x[flags], #4, 10f\n"
@@ -276,61 +276,61 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "sqadd z11.s, z11.s, z16.s\n"
       "12:"  // Height 1: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x4482880f  // srshl z15.s, p2/M, z15.s, z0.s\n"
       "ld1rw { z17.s }, p2/Z, [x20]\n"
+      ".inst 0x4482880f  // srshl z15.s, p2/M, z15.s, z0.s\n"
+      "add z15.s, z15.s, z17.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
+      "add z9.s, z9.s, z17.s\n"
+      "add z10.s, z10.s, z17.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
-      "add z15.s, z15.s, z17.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z9.s, z9.s, z17.s\n"
-      "add z10.s, z10.s, z17.s\n"
-      "ld1rw { z28.s }, p2/Z, [x20]\n"
       "add z11.s, z11.s, z17.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z31.s }, p2/Z, [x20]\n"
       "smin z15.s, p2/M, z15.s, z16.s\n"
       "smin z9.s, p2/M, z9.s, z16.s\n"
       "smin z10.s, p2/M, z10.s, z16.s\n"
       "smin z11.s, p2/M, z11.s, z16.s\n"
-      "smax z15.s, p2/M, z15.s, z28.s\n"
-      "smax z9.s, p2/M, z9.s, z28.s\n"
-      "smax z10.s, p2/M, z10.s, z28.s\n"
-      "smax z11.s, p2/M, z11.s, z28.s\n"
+      "smax z15.s, p2/M, z15.s, z31.s\n"
+      "smax z9.s, p2/M, z9.s, z31.s\n"
+      "smax z10.s, p2/M, z10.s, z31.s\n"
       "uzp1 z15.h, z15.h, z9.h\n"
+      "smax z11.s, p2/M, z11.s, z31.s\n"
       "uzp1 z16.h, z10.h, z11.h\n"
       "uzp1 z15.b, z15.b, z16.b\n"
-      "st1b { z15.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
+      "st1b { z15.b }, p1, [x11]\n"
+      "addvl x11, x11, #1\n"
       "13:"  // Height 1: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 2b\n"
       "b 80f\n"
       "14:"  // Height 2
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "15:"  // Height 2: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "16:"  // Height 2: setup done
       "mov x28, #0x0\n"
       "17:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 18f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -349,90 +349,90 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "ble 21f\n"
       "20:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z19.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z18.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z16.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "trn1 z2.d, z18.d, z16.d\n"
-      "trn2 z18.d, z18.d, z16.d\n"
-      ".inst 0x45119848  // smmla z8.s, z2.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4513984c  // smmla z12.s, z2.b, z19.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45119849  // smmla z9.s, z2.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4510984d  // smmla z13.s, z2.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4511984a  // smmla z10.s, z2.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4510984e  // smmla z14.s, z2.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x4511984b  // smmla z11.s, z2.b, z17.b\n"
-      ".inst 0x4510984f  // smmla z15.s, z2.b, z16.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1rqb { z20.b }, p0/Z, [x26]\n"
+      "ld1rqb { z19.b }, p0/Z, [x25]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45119a48  // smmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x45109a4c  // smmla z12.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45119a49  // smmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x45109a4d  // smmla z13.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x45119a4a  // smmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x45109a4e  // smmla z14.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
+      "trn2 z20.d, z20.d, z19.d\n"
       ".inst 0x45119a4b  // smmla z11.s, z18.b, z17.b\n"
       ".inst 0x45109a4f  // smmla z15.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-8, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-7, MUL VL]\n"
+      ".inst 0x45119a88  // smmla z8.s, z20.b, z17.b\n"
+      ".inst 0x45109a8c  // smmla z12.s, z20.b, z16.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-5, MUL VL]\n"
+      ".inst 0x45119a89  // smmla z9.s, z20.b, z17.b\n"
+      ".inst 0x45109a8d  // smmla z13.s, z20.b, z16.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-3, MUL VL]\n"
+      ".inst 0x45119a8a  // smmla z10.s, z20.b, z17.b\n"
+      ".inst 0x45109a8e  // smmla z14.s, z20.b, z16.b\n"
+      "ld1b { z17.b }, p2/Z, [x9, #-2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #-1, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45119a8b  // smmla z11.s, z20.b, z17.b\n"
+      ".inst 0x45109a8f  // smmla z15.s, z20.b, z16.b\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "bgt 20b\n"
       "21:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "ld1rqb { z19.b }, p0/Z, [x25]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45119a48  // smmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45109a4c  // smmla z12.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45119a49  // smmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45109a4d  // smmla z13.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x45119a4a  // smmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45109a4e  // smmla z14.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45119a4b  // smmla z11.s, z18.b, z17.b\n"
       ".inst 0x45109a4f  // smmla z15.s, z18.b, z16.b\n"
+      "addvl x9, x9, #8\n"
       "ble 22f\n"
-      "ld1b { z17.b }, p2/Z, [x10]\n"
-      "ld1b { z16.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45119828  // smmla z8.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4510982c  // smmla z12.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45119829  // smmla z9.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #4, MUL VL]\n"
       ".inst 0x4510982d  // smmla z13.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z17.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x4511982a  // smmla z10.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p2/Z, [x10, #6, MUL VL]\n"
       ".inst 0x4510982e  // smmla z14.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z17.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x9, #7, MUL VL]\n"
       ".inst 0x4511982b  // smmla z11.s, z1.b, z17.b\n"
       ".inst 0x4510982f  // smmla z15.s, z1.b, z16.b\n"
+      "addvl x9, x9, #8\n"
       "22:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -449,16 +449,16 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
       "ld1w { z16.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
+      "add x26, x11, x20\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x26, x9, x20\n"
+      "addvl x14, x14, #4\n"
       "mov z15.d, z20.d\n"
+      "add z15.s, z15.s, z19.s\n"
       "add z12.s, z12.s, z18.s\n"
       "add z13.s, z13.s, z17.s\n"
-      "add z8.s, z8.s, z19.s\n"
-      "add z15.s, z15.s, z19.s\n"
       "add z14.s, z14.s, z16.s\n"
+      "add z8.s, z8.s, z19.s\n"
       "add z9.s, z9.s, z18.s\n"
       "add z10.s, z10.s, z17.s\n"
       "add z11.s, z11.s, z16.s\n"
@@ -495,99 +495,99 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       ".inst 0x04a6754a  // sqrdmulh z10.s, z10.s, z6.s\n"
       ".inst 0x04a7756b  // sqrdmulh z11.s, z11.s, z7.s\n"
       "tbz %x[flags], #5, 25f\n"
-      "and z18.d, z15.d, z0.d\n"
-      "and z19.d, z12.d, z1.d\n"
+      "and z19.d, z15.d, z0.d\n"
+      "and z18.d, z12.d, z1.d\n"
       "and z17.d, z13.d, z2.d\n"
       "and z16.d, z14.d, z3.d\n"
-      "asr z18.s, z18.s, #0x1f\n"
       "asr z19.s, z19.s, #0x1f\n"
+      "asr z18.s, z18.s, #0x1f\n"
       "asr z17.s, z17.s, #0x1f\n"
       "asr z16.s, z16.s, #0x1f\n"
-      "sqadd z15.s, z15.s, z18.s\n"
-      "and z18.d, z8.d, z0.d\n"
-      "sqadd z12.s, z12.s, z19.s\n"
-      "and z19.d, z9.d, z1.d\n"
+      "sqadd z15.s, z15.s, z19.s\n"
+      "sqadd z12.s, z12.s, z18.s\n"
       "sqadd z13.s, z13.s, z17.s\n"
       "sqadd z14.s, z14.s, z16.s\n"
+      "and z18.d, z8.d, z0.d\n"
+      "and z24.d, z9.d, z1.d\n"
       "and z17.d, z10.d, z2.d\n"
       "and z16.d, z11.d, z3.d\n"
       "asr z18.s, z18.s, #0x1f\n"
-      "asr z19.s, z19.s, #0x1f\n"
+      "asr z24.s, z24.s, #0x1f\n"
       "asr z17.s, z17.s, #0x1f\n"
       "asr z16.s, z16.s, #0x1f\n"
       "sqadd z8.s, z8.s, z18.s\n"
-      "sqadd z9.s, z9.s, z19.s\n"
+      "sqadd z9.s, z9.s, z24.s\n"
       "sqadd z10.s, z10.s, z17.s\n"
       "sqadd z11.s, z11.s, z16.s\n"
       "25:"  // Height 2: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
       ".inst 0x4482880f  // srshl z15.s, p2/M, z15.s, z0.s\n"
-      "ld1rw { z18.s }, p2/Z, [x20]\n"
+      "add z15.s, z15.s, z17.s\n"
       ".inst 0x4482882c  // srshl z12.s, p2/M, z12.s, z1.s\n"
       ".inst 0x4482884d  // srshl z13.s, p2/M, z13.s, z2.s\n"
+      "add z12.s, z12.s, z17.s\n"
+      "add z13.s, z13.s, z17.s\n"
       ".inst 0x4482886e  // srshl z14.s, p2/M, z14.s, z3.s\n"
       ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
+      "add z14.s, z14.s, z17.s\n"
+      "add z8.s, z8.s, z17.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
-      "add z15.s, z15.s, z18.s\n"
+      "add z9.s, z9.s, z17.s\n"
+      "add z10.s, z10.s, z17.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z12.s, z12.s, z18.s\n"
-      "add z13.s, z13.s, z18.s\n"
-      "ld1rw { z17.s }, p2/Z, [x20]\n"
-      "add z14.s, z14.s, z18.s\n"
-      "add z8.s, z8.s, z18.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z9.s, z9.s, z18.s\n"
-      "add z10.s, z10.s, z18.s\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
-      "add z11.s, z11.s, z18.s\n"
-      "smin z15.s, p2/M, z15.s, z17.s\n"
-      "smin z12.s, p2/M, z12.s, z17.s\n"
-      "smin z13.s, p2/M, z13.s, z17.s\n"
-      "smin z14.s, p2/M, z14.s, z17.s\n"
-      "smin z8.s, p2/M, z8.s, z17.s\n"
-      "smin z9.s, p2/M, z9.s, z17.s\n"
-      "smin z10.s, p2/M, z10.s, z17.s\n"
-      "smin z11.s, p2/M, z11.s, z17.s\n"
-      "smax z15.s, p2/M, z15.s, z16.s\n"
-      "smax z12.s, p2/M, z12.s, z16.s\n"
-      "smax z13.s, p2/M, z13.s, z16.s\n"
-      "smax z14.s, p2/M, z14.s, z16.s\n"
-      "smax z8.s, p2/M, z8.s, z16.s\n"
-      "smax z9.s, p2/M, z9.s, z16.s\n"
-      "smax z10.s, p2/M, z10.s, z16.s\n"
-      "smax z11.s, p2/M, z11.s, z16.s\n"
+      "add z11.s, z11.s, z17.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z17.s }, p2/Z, [x20]\n"
+      "smin z15.s, p2/M, z15.s, z16.s\n"
+      "smin z12.s, p2/M, z12.s, z16.s\n"
+      "smin z13.s, p2/M, z13.s, z16.s\n"
+      "smin z14.s, p2/M, z14.s, z16.s\n"
+      "smin z8.s, p2/M, z8.s, z16.s\n"
+      "smin z9.s, p2/M, z9.s, z16.s\n"
+      "smin z10.s, p2/M, z10.s, z16.s\n"
+      "smin z11.s, p2/M, z11.s, z16.s\n"
+      "smax z15.s, p2/M, z15.s, z17.s\n"
+      "smax z12.s, p2/M, z12.s, z17.s\n"
+      "smax z13.s, p2/M, z13.s, z17.s\n"
       "uzp1 z15.h, z15.h, z12.h\n"
-      "uzp1 z17.h, z13.h, z14.h\n"
+      "smax z14.s, p2/M, z14.s, z17.s\n"
+      "smax z8.s, p2/M, z8.s, z17.s\n"
+      "uzp1 z16.h, z13.h, z14.h\n"
+      "uzp1 z15.b, z15.b, z16.b\n"
+      "smax z9.s, p2/M, z9.s, z17.s\n"
+      "smax z10.s, p2/M, z10.s, z17.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
+      "st1b { z15.b }, p1, [x11]\n"
+      "smax z11.s, p2/M, z11.s, z17.s\n"
       "uzp1 z16.h, z10.h, z11.h\n"
-      "uzp1 z15.b, z15.b, z17.b\n"
       "uzp1 z8.b, z8.b, z16.b\n"
-      "st1b { z15.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
       "st1b { z8.b }, p1, [x26]\n"
+      "addvl x11, x11, #1\n"
       "26:"  // Height 2: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 15b\n"
       "b 80f\n"
       "27:"  // Height 3
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "28:"  // Height 3: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "mov z16.s, #0x0\n"
@@ -602,8 +602,8 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "mov x28, #0x0\n"
       "30:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 31f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -625,125 +625,125 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "ble 34f\n"
       "33:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z28.b }, p2/Z, [x10]\n"
-      "ld1b { z25.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z30.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z24.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z29.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1rqb { z28.b }, p0/Z, [x24]\n"
       "trn1 z27.d, z30.d, z24.d\n"
       "trn2 z30.d, z30.d, z24.d\n"
-      "trn1 z26.d, z29.d, z31.d\n"
-      "trn2 z29.d, z29.d, z31.d\n"
-      ".inst 0x451c9b68  // smmla z8.s, z27.b, z28.b\n"
-      ".inst 0x45199b6c  // smmla z12.s, z27.b, z25.b\n"
-      ".inst 0x451c9b50  // smmla z16.s, z26.b, z28.b\n"
-      "ld1b { z4.b }, p2/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45199b54  // smmla z20.s, z26.b, z25.b\n"
-      "ld1b { z28.b }, p2/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45049b69  // smmla z9.s, z27.b, z4.b\n"
-      ".inst 0x45049b51  // smmla z17.s, z26.b, z4.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x451c9b6d  // smmla z13.s, z27.b, z28.b\n"
-      ".inst 0x451c9b55  // smmla z21.s, z26.b, z28.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "trn1 z26.d, z28.d, z29.d\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+      ".inst 0x45199b68  // smmla z8.s, z27.b, z25.b\n"
+      ".inst 0x45199b50  // smmla z16.s, z26.b, z25.b\n"
+      ".inst 0x45189b6c  // smmla z12.s, z27.b, z24.b\n"
+      ".inst 0x45189b54  // smmla z20.s, z26.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+      ".inst 0x45199b69  // smmla z9.s, z27.b, z25.b\n"
+      ".inst 0x45199b51  // smmla z17.s, z26.b, z25.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "trn2 z28.d, z28.d, z29.d\n"
+      ".inst 0x45189b6d  // smmla z13.s, z27.b, z24.b\n"
+      ".inst 0x45189b55  // smmla z21.s, z26.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x45199b6a  // smmla z10.s, z27.b, z25.b\n"
       ".inst 0x45199b52  // smmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "cmp x27, #0x10\n"
       ".inst 0x45189b6e  // smmla z14.s, z27.b, z24.b\n"
       ".inst 0x45189b56  // smmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
+      "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
       ".inst 0x45199b6b  // smmla z11.s, z27.b, z25.b\n"
       ".inst 0x45199b53  // smmla z19.s, z26.b, z25.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-8, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x45189b6f  // smmla z15.s, z27.b, z24.b\n"
       ".inst 0x45189b57  // smmla z23.s, z26.b, z24.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-7, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x45199bc8  // smmla z8.s, z30.b, z25.b\n"
-      ".inst 0x45199bb0  // smmla z16.s, z29.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-6, MUL VL]\n"
+      ".inst 0x45199b90  // smmla z16.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-6, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x45189bcc  // smmla z12.s, z30.b, z24.b\n"
-      ".inst 0x45189bb4  // smmla z20.s, z29.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      ".inst 0x45189b94  // smmla z20.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-5, MUL VL]\n"
       ".inst 0x45199bc9  // smmla z9.s, z30.b, z25.b\n"
-      ".inst 0x45199bb1  // smmla z17.s, z29.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-4, MUL VL]\n"
+      ".inst 0x45199b91  // smmla z17.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-4, MUL VL]\n"
       ".inst 0x45189bcd  // smmla z13.s, z30.b, z24.b\n"
-      ".inst 0x45189bb5  // smmla z21.s, z29.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      ".inst 0x45189b95  // smmla z21.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-3, MUL VL]\n"
       ".inst 0x45199bca  // smmla z10.s, z30.b, z25.b\n"
-      ".inst 0x45199bb2  // smmla z18.s, z29.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-2, MUL VL]\n"
+      ".inst 0x45199b92  // smmla z18.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-2, MUL VL]\n"
       ".inst 0x45189bce  // smmla z14.s, z30.b, z24.b\n"
-      ".inst 0x45189bb6  // smmla z22.s, z29.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      ".inst 0x45189b96  // smmla z22.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-1, MUL VL]\n"
       ".inst 0x45199bcb  // smmla z11.s, z30.b, z25.b\n"
-      ".inst 0x45199bb3  // smmla z19.s, z29.b, z25.b\n"
+      ".inst 0x45199b93  // smmla z19.s, z28.b, z25.b\n"
       ".inst 0x45189bcf  // smmla z15.s, z30.b, z24.b\n"
-      ".inst 0x45189bb7  // smmla z23.s, z29.b, z24.b\n"
+      ".inst 0x45189b97  // smmla z23.s, z28.b, z24.b\n"
       "bgt 33b\n"
       "34:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p2/Z, [x10]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "ld1rqb { z24.b }, p0/Z, [x25]\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
       "trn1 z27.d, z1.d, z24.d\n"
       "trn2 z1.d, z1.d, z24.d\n"
-      "trn1 z26.d, z3.d, z29.d\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "trn1 z26.d, z3.d, z28.d\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45199b68  // smmla z8.s, z27.b, z25.b\n"
-      ".inst 0x451c9b6c  // smmla z12.s, z27.b, z28.b\n"
-      "trn2 z3.d, z3.d, z29.d\n"
       ".inst 0x45199b50  // smmla z16.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x451c9b54  // smmla z20.s, z26.b, z28.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
+      ".inst 0x45189b6c  // smmla z12.s, z27.b, z24.b\n"
+      ".inst 0x45189b54  // smmla z20.s, z26.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45199b69  // smmla z9.s, z27.b, z25.b\n"
       ".inst 0x45199b51  // smmla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x45189b6d  // smmla z13.s, z27.b, z24.b\n"
       ".inst 0x45189b55  // smmla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+      "trn2 z3.d, z3.d, z28.d\n"
       ".inst 0x45199b6a  // smmla z10.s, z27.b, z25.b\n"
       ".inst 0x45199b52  // smmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
       ".inst 0x45189b6e  // smmla z14.s, z27.b, z24.b\n"
       ".inst 0x45189b56  // smmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #8\n"
       ".inst 0x45199b6b  // smmla z11.s, z27.b, z25.b\n"
       ".inst 0x45199b53  // smmla z19.s, z26.b, z25.b\n"
       ".inst 0x45189b6f  // smmla z15.s, z27.b, z24.b\n"
       ".inst 0x45189b57  // smmla z23.s, z26.b, z24.b\n"
       "ble 35f\n"
-      "ld1b { z25.b }, p2/Z, [x10]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45199828  // smmla z8.s, z1.b, z25.b\n"
       ".inst 0x45199870  // smmla z16.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4518982c  // smmla z12.s, z1.b, z24.b\n"
       ".inst 0x45189874  // smmla z20.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45199829  // smmla z9.s, z1.b, z25.b\n"
       ".inst 0x45199871  // smmla z17.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
       ".inst 0x4518982d  // smmla z13.s, z1.b, z24.b\n"
       ".inst 0x45189875  // smmla z21.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x4519982a  // smmla z10.s, z1.b, z25.b\n"
       ".inst 0x45199872  // smmla z18.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
       ".inst 0x4518982e  // smmla z14.s, z1.b, z24.b\n"
       ".inst 0x45189876  // smmla z22.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #8\n"
       ".inst 0x4519982b  // smmla z11.s, z1.b, z25.b\n"
       ".inst 0x45199873  // smmla z19.s, z3.b, z25.b\n"
       ".inst 0x4518982f  // smmla z15.s, z1.b, z24.b\n"
@@ -764,20 +764,20 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
       "ld1w { z24.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
+      "add x26, x11, x20\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x26, x9, x20\n"
+      "add x25, x26, x20\n"
+      "addvl x14, x14, #4\n"
       "uzp1 z16.d, z16.d, z20.d\n"
       "uzp1 z17.d, z17.d, z21.d\n"
-      "add x25, x26, x20\n"
       "uzp1 z18.d, z18.d, z22.d\n"
       "uzp1 z19.d, z19.d, z23.d\n"
       "mov z23.d, z28.d\n"
+      "add z23.s, z23.s, z27.s\n"
       "add z12.s, z12.s, z26.s\n"
       "add z13.s, z13.s, z25.s\n"
       "add z14.s, z14.s, z24.s\n"
-      "add z23.s, z23.s, z27.s\n"
       "add z8.s, z8.s, z27.s\n"
       "add z9.s, z9.s, z26.s\n"
       "add z10.s, z10.s, z25.s\n"
@@ -832,11 +832,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z21.s, z21.s, #0x1f\n"
       "asr z20.s, z20.s, #0x1f\n"
       "sqadd z23.s, z23.s, z24.s\n"
-      "and z24.d, z8.d, z0.d\n"
       "sqadd z12.s, z12.s, z22.s\n"
-      "and z22.d, z9.d, z1.d\n"
       "sqadd z13.s, z13.s, z21.s\n"
       "sqadd z14.s, z14.s, z20.s\n"
+      "and z24.d, z8.d, z0.d\n"
+      "and z22.d, z9.d, z1.d\n"
       "and z21.d, z10.d, z2.d\n"
       "and z20.d, z11.d, z3.d\n"
       "asr z24.s, z24.s, #0x1f\n"
@@ -844,11 +844,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z21.s, z21.s, #0x1f\n"
       "asr z20.s, z20.s, #0x1f\n"
       "sqadd z8.s, z8.s, z24.s\n"
-      "and z24.d, z16.d, z0.d\n"
       "sqadd z9.s, z9.s, z22.s\n"
-      "and z22.d, z17.d, z1.d\n"
       "sqadd z10.s, z10.s, z21.s\n"
       "sqadd z11.s, z11.s, z20.s\n"
+      "and z24.d, z16.d, z0.d\n"
+      "and z22.d, z17.d, z1.d\n"
       "and z21.d, z18.d, z2.d\n"
       "and z20.d, z19.d, z3.d\n"
       "asr z24.s, z24.s, #0x1f\n"
@@ -861,93 +861,93 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "sqadd z19.s, z19.s, z20.s\n"
       "38:"  // Height 3: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z21.s }, p2/Z, [x20]\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "ld1rw { z22.s }, p2/Z, [x20]\n"
+      "add z23.s, z23.s, z21.s\n"
       ".inst 0x4482882c  // srshl z12.s, p2/M, z12.s, z1.s\n"
       ".inst 0x4482884d  // srshl z13.s, p2/M, z13.s, z2.s\n"
+      "add z12.s, z12.s, z21.s\n"
+      "add z13.s, z13.s, z21.s\n"
       ".inst 0x4482886e  // srshl z14.s, p2/M, z14.s, z3.s\n"
       ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
+      "add z14.s, z14.s, z21.s\n"
+      "add z8.s, z8.s, z21.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
-      "add z23.s, z23.s, z22.s\n"
+      "add z9.s, z9.s, z21.s\n"
+      "add z10.s, z10.s, z21.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "add z12.s, z12.s, z22.s\n"
-      "add z13.s, z13.s, z22.s\n"
+      "add z11.s, z11.s, z21.s\n"
+      "add z16.s, z16.s, z21.s\n"
       ".inst 0x44828831  // srshl z17.s, p2/M, z17.s, z1.s\n"
       ".inst 0x44828852  // srshl z18.s, p2/M, z18.s, z2.s\n"
-      "add z14.s, z14.s, z22.s\n"
-      "add z8.s, z8.s, z22.s\n"
+      "add z17.s, z17.s, z21.s\n"
+      "add z18.s, z18.s, z21.s\n"
       ".inst 0x44828873  // srshl z19.s, p2/M, z19.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z9.s, z9.s, z22.s\n"
-      "add z10.s, z10.s, z22.s\n"
-      "ld1rw { z21.s }, p2/Z, [x20]\n"
-      "add z11.s, z11.s, z22.s\n"
-      "add z16.s, z16.s, z22.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z22.s\n"
-      "add z18.s, z18.s, z22.s\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z22.s\n"
-      "smin z23.s, p2/M, z23.s, z21.s\n"
-      "smin z12.s, p2/M, z12.s, z21.s\n"
-      "smin z13.s, p2/M, z13.s, z21.s\n"
-      "smin z14.s, p2/M, z14.s, z21.s\n"
-      "smin z8.s, p2/M, z8.s, z21.s\n"
-      "smin z9.s, p2/M, z9.s, z21.s\n"
-      "smin z10.s, p2/M, z10.s, z21.s\n"
-      "smin z11.s, p2/M, z11.s, z21.s\n"
-      "smin z16.s, p2/M, z16.s, z21.s\n"
-      "smin z17.s, p2/M, z17.s, z21.s\n"
-      "smin z18.s, p2/M, z18.s, z21.s\n"
-      "smin z19.s, p2/M, z19.s, z21.s\n"
-      "smax z23.s, p2/M, z23.s, z20.s\n"
-      "smax z12.s, p2/M, z12.s, z20.s\n"
-      "smax z13.s, p2/M, z13.s, z20.s\n"
-      "smax z14.s, p2/M, z14.s, z20.s\n"
-      "smax z8.s, p2/M, z8.s, z20.s\n"
-      "smax z9.s, p2/M, z9.s, z20.s\n"
-      "smax z10.s, p2/M, z10.s, z20.s\n"
-      "smax z11.s, p2/M, z11.s, z20.s\n"
+      "add z19.s, z19.s, z21.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z21.s }, p2/Z, [x20]\n"
+      "smin z23.s, p2/M, z23.s, z20.s\n"
+      "smin z12.s, p2/M, z12.s, z20.s\n"
+      "smin z13.s, p2/M, z13.s, z20.s\n"
+      "smin z14.s, p2/M, z14.s, z20.s\n"
+      "smin z8.s, p2/M, z8.s, z20.s\n"
+      "smin z9.s, p2/M, z9.s, z20.s\n"
+      "smin z10.s, p2/M, z10.s, z20.s\n"
+      "smin z11.s, p2/M, z11.s, z20.s\n"
+      "smin z16.s, p2/M, z16.s, z20.s\n"
+      "smin z17.s, p2/M, z17.s, z20.s\n"
+      "smin z18.s, p2/M, z18.s, z20.s\n"
+      "smin z19.s, p2/M, z19.s, z20.s\n"
+      "smax z23.s, p2/M, z23.s, z21.s\n"
+      "smax z12.s, p2/M, z12.s, z21.s\n"
+      "smax z13.s, p2/M, z13.s, z21.s\n"
       "uzp1 z23.h, z23.h, z12.h\n"
-      "smax z16.s, p2/M, z16.s, z20.s\n"
-      "smax z17.s, p2/M, z17.s, z20.s\n"
-      "uzp1 z21.h, z13.h, z14.h\n"
-      "smax z18.s, p2/M, z18.s, z20.s\n"
-      "smax z19.s, p2/M, z19.s, z20.s\n"
+      "smax z14.s, p2/M, z14.s, z21.s\n"
+      "smax z8.s, p2/M, z8.s, z21.s\n"
+      "uzp1 z20.h, z13.h, z14.h\n"
+      "uzp1 z23.b, z23.b, z20.b\n"
+      "smax z9.s, p2/M, z9.s, z21.s\n"
+      "smax z10.s, p2/M, z10.s, z21.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
+      "st1b { z23.b }, p1, [x11]\n"
+      "smax z11.s, p2/M, z11.s, z21.s\n"
+      "smax z16.s, p2/M, z16.s, z21.s\n"
       "uzp1 z20.h, z10.h, z11.h\n"
+      "uzp1 z8.b, z8.b, z20.b\n"
+      "smax z17.s, p2/M, z17.s, z21.s\n"
+      "smax z18.s, p2/M, z18.s, z21.s\n"
       "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z23.b, z23.b, z21.b\n"
+      "st1b { z8.b }, p1, [x26]\n"
+      "smax z19.s, p2/M, z19.s, z21.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
-      "uzp1 z8.b, z8.b, z20.b\n"
-      "st1b { z23.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
       "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z8.b }, p1, [x26]\n"
       "st1b { z16.b }, p1, [x25]\n"
+      "addvl x11, x11, #1\n"
       "39:"  // Height 3: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 28b\n"
       "b 80f\n"
       "40:"  // Height 4
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "41:"  // Height 4: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "mov z16.s, #0x0\n"
@@ -962,8 +962,8 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "mov x28, #0x0\n"
       "43:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 44f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -988,128 +988,128 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "ble 47f\n"
       "46:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z31.b }, p2/Z, [x10]\n"
-      "ld1b { z30.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z29.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqb { z30.b }, p0/Z, [x26]\n"
+      "ld1rqb { z24.b }, p0/Z, [x25]\n"
+      "trn1 z29.d, z30.d, z24.d\n"
       "ld1rqb { z28.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "trn1 z27.d, z29.d, z25.d\n"
-      "trn2 z29.d, z29.d, z25.d\n"
-      "trn1 z26.d, z28.d, z24.d\n"
-      "trn2 z28.d, z28.d, z24.d\n"
-      ".inst 0x451f9b68  // smmla z8.s, z27.b, z31.b\n"
-      ".inst 0x451e9b6c  // smmla z12.s, z27.b, z30.b\n"
-      ".inst 0x451f9b50  // smmla z16.s, z26.b, z31.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x451e9b54  // smmla z20.s, z26.b, z30.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45199b69  // smmla z9.s, z27.b, z25.b\n"
+      "ld1rqb { z27.b }, p0/Z, [x23]\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "trn1 z26.d, z28.d, z27.d\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+      ".inst 0x45199ba8  // smmla z8.s, z29.b, z25.b\n"
+      ".inst 0x45199b50  // smmla z16.s, z26.b, z25.b\n"
+      ".inst 0x45189bac  // smmla z12.s, z29.b, z24.b\n"
+      ".inst 0x45189b54  // smmla z20.s, z26.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+      ".inst 0x45199ba9  // smmla z9.s, z29.b, z25.b\n"
       ".inst 0x45199b51  // smmla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45189b6d  // smmla z13.s, z27.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "trn2 z28.d, z28.d, z27.d\n"
+      ".inst 0x45189bad  // smmla z13.s, z29.b, z24.b\n"
       ".inst 0x45189b55  // smmla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45199b6a  // smmla z10.s, z27.b, z25.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
+      ".inst 0x45199baa  // smmla z10.s, z29.b, z25.b\n"
       ".inst 0x45199b52  // smmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45189b6e  // smmla z14.s, z27.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45189bae  // smmla z14.s, z29.b, z24.b\n"
       ".inst 0x45189b56  // smmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x45199b6b  // smmla z11.s, z27.b, z25.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
+      ".inst 0x45199bab  // smmla z11.s, z29.b, z25.b\n"
       ".inst 0x45199b53  // smmla z19.s, z26.b, z25.b\n"
-      ".inst 0x45189b6f  // smmla z15.s, z27.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-8, MUL VL]\n"
+      "add x26, x26, #0x10\n"
+      ".inst 0x45189baf  // smmla z15.s, z29.b, z24.b\n"
       ".inst 0x45189b57  // smmla z23.s, z26.b, z24.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45199ba8  // smmla z8.s, z29.b, z25.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-7, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x45199bc8  // smmla z8.s, z30.b, z25.b\n"
       ".inst 0x45199b90  // smmla z16.s, z28.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45189bac  // smmla z12.s, z29.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-6, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45189bcc  // smmla z12.s, z30.b, z24.b\n"
       ".inst 0x45189b94  // smmla z20.s, z28.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45199ba9  // smmla z9.s, z29.b, z25.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-5, MUL VL]\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x45199bc9  // smmla z9.s, z30.b, z25.b\n"
       ".inst 0x45199b91  // smmla z17.s, z28.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45189bad  // smmla z13.s, z29.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-4, MUL VL]\n"
+      ".inst 0x45189bcd  // smmla z13.s, z30.b, z24.b\n"
       ".inst 0x45189b95  // smmla z21.s, z28.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45199baa  // smmla z10.s, z29.b, z25.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-3, MUL VL]\n"
+      ".inst 0x45199bca  // smmla z10.s, z30.b, z25.b\n"
       ".inst 0x45199b92  // smmla z18.s, z28.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45189bae  // smmla z14.s, z29.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #-2, MUL VL]\n"
+      ".inst 0x45189bce  // smmla z14.s, z30.b, z24.b\n"
       ".inst 0x45189b96  // smmla z22.s, z28.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45199bab  // smmla z11.s, z29.b, z25.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #-1, MUL VL]\n"
+      ".inst 0x45199bcb  // smmla z11.s, z30.b, z25.b\n"
       ".inst 0x45199b93  // smmla z19.s, z28.b, z25.b\n"
-      ".inst 0x45189baf  // smmla z15.s, z29.b, z24.b\n"
+      ".inst 0x45189bcf  // smmla z15.s, z30.b, z24.b\n"
       ".inst 0x45189b97  // smmla z23.s, z28.b, z24.b\n"
       "bgt 46b\n"
       "47:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p2/Z, [x10]\n"
-      "ld1b { z28.b }, p2/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
+      "ld1rqb { z24.b }, p0/Z, [x25]\n"
+      "trn1 z28.d, z1.d, z24.d\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "trn1 z27.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      "trn1 z26.d, z3.d, z24.d\n"
-      ".inst 0x451d9b68  // smmla z8.s, z27.b, z29.b\n"
-      ".inst 0x451c9b6c  // smmla z12.s, z27.b, z28.b\n"
-      "trn2 z3.d, z3.d, z24.d\n"
-      ".inst 0x451d9b50  // smmla z16.s, z26.b, z29.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x451c9b54  // smmla z20.s, z26.b, z28.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45199b69  // smmla z9.s, z27.b, z25.b\n"
+      "ld1rqb { z27.b }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z24.d\n"
+      "trn1 z26.d, z3.d, z27.d\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
+      ".inst 0x45199b88  // smmla z8.s, z28.b, z25.b\n"
+      ".inst 0x45199b50  // smmla z16.s, z26.b, z25.b\n"
+      ".inst 0x45189b8c  // smmla z12.s, z28.b, z24.b\n"
+      ".inst 0x45189b54  // smmla z20.s, z26.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
+      ".inst 0x45199b89  // smmla z9.s, z28.b, z25.b\n"
       ".inst 0x45199b51  // smmla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45189b6d  // smmla z13.s, z27.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
+      ".inst 0x45189b8d  // smmla z13.s, z28.b, z24.b\n"
       ".inst 0x45189b55  // smmla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45199b6a  // smmla z10.s, z27.b, z25.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
+      "trn2 z3.d, z3.d, z27.d\n"
+      ".inst 0x45199b8a  // smmla z10.s, z28.b, z25.b\n"
       ".inst 0x45199b52  // smmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45189b6e  // smmla z14.s, z27.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
+      ".inst 0x45189b8e  // smmla z14.s, z28.b, z24.b\n"
       ".inst 0x45189b56  // smmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x45199b6b  // smmla z11.s, z27.b, z25.b\n"
+      "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #8\n"
+      ".inst 0x45199b8b  // smmla z11.s, z28.b, z25.b\n"
       ".inst 0x45199b53  // smmla z19.s, z26.b, z25.b\n"
-      ".inst 0x45189b6f  // smmla z15.s, z27.b, z24.b\n"
+      ".inst 0x45189b8f  // smmla z15.s, z28.b, z24.b\n"
       ".inst 0x45189b57  // smmla z23.s, z26.b, z24.b\n"
       "ble 48f\n"
-      "ld1b { z25.b }, p2/Z, [x10]\n"
-      "ld1b { z24.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45199828  // smmla z8.s, z1.b, z25.b\n"
       ".inst 0x45199870  // smmla z16.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4518982c  // smmla z12.s, z1.b, z24.b\n"
       ".inst 0x45189874  // smmla z20.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45199829  // smmla z9.s, z1.b, z25.b\n"
       ".inst 0x45199871  // smmla z17.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #4, MUL VL]\n"
       ".inst 0x4518982d  // smmla z13.s, z1.b, z24.b\n"
       ".inst 0x45189875  // smmla z21.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x4519982a  // smmla z10.s, z1.b, z25.b\n"
       ".inst 0x45199872  // smmla z18.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x9, #6, MUL VL]\n"
       ".inst 0x4518982e  // smmla z14.s, z1.b, z24.b\n"
       ".inst 0x45189876  // smmla z22.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z24.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #8\n"
       ".inst 0x4519982b  // smmla z11.s, z1.b, z25.b\n"
       ".inst 0x45199873  // smmla z19.s, z3.b, z25.b\n"
       ".inst 0x4518982f  // smmla z15.s, z1.b, z24.b\n"
@@ -1130,25 +1130,25 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
       "ld1w { z24.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
+      "add x26, x11, x20\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x26, x9, x20\n"
+      "add x25, x26, x20\n"
+      "add x24, x25, x20\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x25, x26, x20\n"
+      "addvl x14, x14, #4\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "add x24, x25, x20\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
       "mov z23.d, z28.d\n"
+      "add z23.s, z23.s, z27.s\n"
       "add z12.s, z12.s, z26.s\n"
       "add z13.s, z13.s, z25.s\n"
       "add z14.s, z14.s, z24.s\n"
-      "add z23.s, z23.s, z27.s\n"
       "add z8.s, z8.s, z27.s\n"
       "add z9.s, z9.s, z26.s\n"
       "add z10.s, z10.s, z25.s\n"
@@ -1211,11 +1211,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z25.s, z25.s, #0x1f\n"
       "asr z24.s, z24.s, #0x1f\n"
       "sqadd z23.s, z23.s, z27.s\n"
-      "and z27.d, z8.d, z0.d\n"
       "sqadd z12.s, z12.s, z26.s\n"
-      "and z26.d, z9.d, z1.d\n"
       "sqadd z13.s, z13.s, z25.s\n"
       "sqadd z14.s, z14.s, z24.s\n"
+      "and z27.d, z8.d, z0.d\n"
+      "and z26.d, z9.d, z1.d\n"
       "and z25.d, z10.d, z2.d\n"
       "and z24.d, z11.d, z3.d\n"
       "asr z27.s, z27.s, #0x1f\n"
@@ -1223,11 +1223,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z25.s, z25.s, #0x1f\n"
       "asr z24.s, z24.s, #0x1f\n"
       "sqadd z8.s, z8.s, z27.s\n"
-      "and z27.d, z15.d, z0.d\n"
       "sqadd z9.s, z9.s, z26.s\n"
-      "and z26.d, z20.d, z1.d\n"
       "sqadd z10.s, z10.s, z25.s\n"
       "sqadd z11.s, z11.s, z24.s\n"
+      "and z27.d, z15.d, z0.d\n"
+      "and z26.d, z20.d, z1.d\n"
       "and z25.d, z21.d, z2.d\n"
       "and z24.d, z22.d, z3.d\n"
       "asr z27.s, z27.s, #0x1f\n"
@@ -1235,11 +1235,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z25.s, z25.s, #0x1f\n"
       "asr z24.s, z24.s, #0x1f\n"
       "sqadd z15.s, z15.s, z27.s\n"
-      "and z27.d, z16.d, z0.d\n"
       "sqadd z20.s, z20.s, z26.s\n"
-      "and z26.d, z17.d, z1.d\n"
       "sqadd z21.s, z21.s, z25.s\n"
       "sqadd z22.s, z22.s, z24.s\n"
+      "and z27.d, z16.d, z0.d\n"
+      "and z26.d, z17.d, z1.d\n"
       "and z25.d, z18.d, z2.d\n"
       "and z24.d, z19.d, z3.d\n"
       "asr z27.s, z27.s, #0x1f\n"
@@ -1252,43 +1252,43 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "sqadd z19.s, z19.s, z24.s\n"
       "51:"  // Height 4: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
       "ld1rw { z25.s }, p2/Z, [x20]\n"
+      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
+      "add z23.s, z23.s, z25.s\n"
       ".inst 0x4482882c  // srshl z12.s, p2/M, z12.s, z1.s\n"
       ".inst 0x4482884d  // srshl z13.s, p2/M, z13.s, z2.s\n"
+      "add z12.s, z12.s, z25.s\n"
+      "add z13.s, z13.s, z25.s\n"
       ".inst 0x4482886e  // srshl z14.s, p2/M, z14.s, z3.s\n"
       ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
+      "add z14.s, z14.s, z25.s\n"
+      "add z8.s, z8.s, z25.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
-      "add z23.s, z23.s, z25.s\n"
+      "add z9.s, z9.s, z25.s\n"
+      "add z10.s, z10.s, z25.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x4482880f  // srshl z15.s, p2/M, z15.s, z0.s\n"
-      "add z12.s, z12.s, z25.s\n"
-      "add z13.s, z13.s, z25.s\n"
+      "add z11.s, z11.s, z25.s\n"
+      "add z15.s, z15.s, z25.s\n"
       ".inst 0x44828834  // srshl z20.s, p2/M, z20.s, z1.s\n"
       ".inst 0x44828855  // srshl z21.s, p2/M, z21.s, z2.s\n"
-      "add z14.s, z14.s, z25.s\n"
-      "add z8.s, z8.s, z25.s\n"
+      "add z20.s, z20.s, z25.s\n"
+      "add z21.s, z21.s, z25.s\n"
       ".inst 0x44828876  // srshl z22.s, p2/M, z22.s, z3.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "add z9.s, z9.s, z25.s\n"
-      "add z10.s, z10.s, z25.s\n"
+      "add z22.s, z22.s, z25.s\n"
+      "add z16.s, z16.s, z25.s\n"
       ".inst 0x44828831  // srshl z17.s, p2/M, z17.s, z1.s\n"
       ".inst 0x44828852  // srshl z18.s, p2/M, z18.s, z2.s\n"
-      "add z11.s, z11.s, z25.s\n"
-      "add z15.s, z15.s, z25.s\n"
+      "add z17.s, z17.s, z25.s\n"
+      "add z18.s, z18.s, z25.s\n"
       ".inst 0x44828873  // srshl z19.s, p2/M, z19.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z20.s, z20.s, z25.s\n"
-      "add z21.s, z21.s, z25.s\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add z22.s, z22.s, z25.s\n"
-      "add z16.s, z16.s, z25.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z25.s\n"
-      "add z18.s, z18.s, z25.s\n"
-      "ld1rw { z26.s }, p2/Z, [x20]\n"
       "add z19.s, z19.s, z25.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z25.s }, p2/Z, [x20]\n"
       "smin z23.s, p2/M, z23.s, z24.s\n"
       "smin z12.s, p2/M, z12.s, z24.s\n"
       "smin z13.s, p2/M, z13.s, z24.s\n"
@@ -1305,60 +1305,60 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "smin z17.s, p2/M, z17.s, z24.s\n"
       "smin z18.s, p2/M, z18.s, z24.s\n"
       "smin z19.s, p2/M, z19.s, z24.s\n"
-      "smax z23.s, p2/M, z23.s, z26.s\n"
-      "smax z12.s, p2/M, z12.s, z26.s\n"
-      "smax z13.s, p2/M, z13.s, z26.s\n"
-      "smax z14.s, p2/M, z14.s, z26.s\n"
-      "smax z8.s, p2/M, z8.s, z26.s\n"
-      "smax z9.s, p2/M, z9.s, z26.s\n"
-      "smax z10.s, p2/M, z10.s, z26.s\n"
-      "smax z11.s, p2/M, z11.s, z26.s\n"
+      "smax z23.s, p2/M, z23.s, z25.s\n"
+      "smax z12.s, p2/M, z12.s, z25.s\n"
+      "smax z13.s, p2/M, z13.s, z25.s\n"
       "uzp1 z23.h, z23.h, z12.h\n"
-      "smax z15.s, p2/M, z15.s, z26.s\n"
-      "smax z20.s, p2/M, z20.s, z26.s\n"
-      "uzp1 z25.h, z13.h, z14.h\n"
-      "smax z21.s, p2/M, z21.s, z26.s\n"
-      "smax z22.s, p2/M, z22.s, z26.s\n"
+      "smax z14.s, p2/M, z14.s, z25.s\n"
+      "smax z8.s, p2/M, z8.s, z25.s\n"
+      "uzp1 z24.h, z13.h, z14.h\n"
+      "uzp1 z23.b, z23.b, z24.b\n"
+      "smax z9.s, p2/M, z9.s, z25.s\n"
+      "smax z10.s, p2/M, z10.s, z25.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
-      "smax z16.s, p2/M, z16.s, z26.s\n"
-      "smax z17.s, p2/M, z17.s, z26.s\n"
-      "uzp1 z24.h, z10.h, z11.h\n"
-      "smax z18.s, p2/M, z18.s, z26.s\n"
-      "smax z19.s, p2/M, z19.s, z26.s\n"
+      "st1b { z23.b }, p1, [x11]\n"
+      "smax z11.s, p2/M, z11.s, z25.s\n"
+      "smax z15.s, p2/M, z15.s, z25.s\n"
+      "uzp1 z23.h, z10.h, z11.h\n"
+      "uzp1 z8.b, z8.b, z23.b\n"
+      "smax z20.s, p2/M, z20.s, z25.s\n"
+      "smax z21.s, p2/M, z21.s, z25.s\n"
       "uzp1 z15.h, z15.h, z20.h\n"
-      "uzp1 z23.b, z23.b, z25.b\n"
+      "st1b { z8.b }, p1, [x26]\n"
+      "smax z22.s, p2/M, z22.s, z25.s\n"
+      "smax z16.s, p2/M, z16.s, z25.s\n"
       "uzp1 z20.h, z21.h, z22.h\n"
+      "uzp1 z15.b, z15.b, z20.b\n"
+      "smax z17.s, p2/M, z17.s, z25.s\n"
+      "smax z18.s, p2/M, z18.s, z25.s\n"
       "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z8.b, z8.b, z24.b\n"
+      "st1b { z15.b }, p1, [x25]\n"
+      "smax z19.s, p2/M, z19.s, z25.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
-      "st1b { z23.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
-      "uzp1 z15.b, z15.b, z20.b\n"
       "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z8.b }, p1, [x26]\n"
-      "st1b { z15.b }, p1, [x25]\n"
       "st1b { z16.b }, p1, [x24]\n"
+      "addvl x11, x11, #1\n"
       "52:"  // Height 4: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 41b\n"
       "b 80f\n"
       "53:"  // Height 5
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "54:"  // Height 5: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "mov z16.s, #0x0\n"
@@ -1381,8 +1381,8 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "mov x28, #0x0\n"
       "56:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 57f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1410,165 +1410,165 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "ble 60f\n"
       "59:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p2/Z, [x10]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z6.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqb { z1.b }, p0/Z, [x25]\n"
       "ld1rqb { z7.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
+      "trn1 z5.d, z6.d, z1.d\n"
+      "trn2 z6.d, z6.d, z1.d\n"
+      "ld1rqb { z4.b }, p0/Z, [x22]\n"
       "trn1 z3.d, z7.d, z2.d\n"
       "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
-      "ld1b { z0.b }, p2/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45019888  // smmla z8.s, z4.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9]\n"
+      "trn1 z2.d, z4.d, z0.d\n"
+      "trn2 z4.d, z4.d, z0.d\n"
+      "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
+      ".inst 0x450198a8  // smmla z8.s, z5.b, z1.b\n"
       ".inst 0x45019870  // smmla z16.s, z3.b, z1.b\n"
       ".inst 0x45019858  // smmla z24.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4500988c  // smmla z12.s, z4.b, z0.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
+      ".inst 0x450098ac  // smmla z12.s, z5.b, z0.b\n"
       ".inst 0x45009874  // smmla z20.s, z3.b, z0.b\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4500985c  // smmla z28.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45019889  // smmla z9.s, z4.b, z1.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
+      ".inst 0x450198a9  // smmla z9.s, z5.b, z1.b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x45019871  // smmla z17.s, z3.b, z1.b\n"
       ".inst 0x45019859  // smmla z25.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4500988d  // smmla z13.s, z4.b, z0.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #4, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x450098ad  // smmla z13.s, z5.b, z0.b\n"
       ".inst 0x45009875  // smmla z21.s, z3.b, z0.b\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4500985d  // smmla z29.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4501988a  // smmla z10.s, z4.b, z1.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
+      ".inst 0x450198aa  // smmla z10.s, z5.b, z1.b\n"
       ".inst 0x45019872  // smmla z18.s, z3.b, z1.b\n"
       ".inst 0x4501985a  // smmla z26.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4500988e  // smmla z14.s, z4.b, z0.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #6, MUL VL]\n"
+      ".inst 0x450098ae  // smmla z14.s, z5.b, z0.b\n"
       ".inst 0x45009876  // smmla z22.s, z3.b, z0.b\n"
       ".inst 0x4500985e  // smmla z30.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x4501988b  // smmla z11.s, z4.b, z1.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
+      ".inst 0x450198ab  // smmla z11.s, z5.b, z1.b\n"
       ".inst 0x45019873  // smmla z19.s, z3.b, z1.b\n"
       ".inst 0x4501985b  // smmla z27.s, z2.b, z1.b\n"
-      ".inst 0x4500988f  // smmla z15.s, z4.b, z0.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-8, MUL VL]\n"
+      ".inst 0x450098af  // smmla z15.s, z5.b, z0.b\n"
       ".inst 0x45009877  // smmla z23.s, z3.b, z0.b\n"
       ".inst 0x4500985f  // smmla z31.s, z2.b, z0.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-7, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-7, MUL VL]\n"
       ".inst 0x450198c8  // smmla z8.s, z6.b, z1.b\n"
       ".inst 0x450198f0  // smmla z16.s, z7.b, z1.b\n"
-      ".inst 0x450198b8  // smmla z24.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-6, MUL VL]\n"
+      ".inst 0x45019898  // smmla z24.s, z4.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-6, MUL VL]\n"
       ".inst 0x450098cc  // smmla z12.s, z6.b, z0.b\n"
       ".inst 0x450098f4  // smmla z20.s, z7.b, z0.b\n"
-      ".inst 0x450098bc  // smmla z28.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-5, MUL VL]\n"
+      ".inst 0x4500989c  // smmla z28.s, z4.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-5, MUL VL]\n"
       ".inst 0x450198c9  // smmla z9.s, z6.b, z1.b\n"
       ".inst 0x450198f1  // smmla z17.s, z7.b, z1.b\n"
-      ".inst 0x450198b9  // smmla z25.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-4, MUL VL]\n"
+      ".inst 0x45019899  // smmla z25.s, z4.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-4, MUL VL]\n"
       ".inst 0x450098cd  // smmla z13.s, z6.b, z0.b\n"
       ".inst 0x450098f5  // smmla z21.s, z7.b, z0.b\n"
-      ".inst 0x450098bd  // smmla z29.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      ".inst 0x4500989d  // smmla z29.s, z4.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-3, MUL VL]\n"
       ".inst 0x450198ca  // smmla z10.s, z6.b, z1.b\n"
       ".inst 0x450198f2  // smmla z18.s, z7.b, z1.b\n"
-      ".inst 0x450198ba  // smmla z26.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-2, MUL VL]\n"
+      ".inst 0x4501989a  // smmla z26.s, z4.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-2, MUL VL]\n"
       ".inst 0x450098ce  // smmla z14.s, z6.b, z0.b\n"
       ".inst 0x450098f6  // smmla z22.s, z7.b, z0.b\n"
-      ".inst 0x450098be  // smmla z30.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      ".inst 0x4500989e  // smmla z30.s, z4.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-1, MUL VL]\n"
       ".inst 0x450198cb  // smmla z11.s, z6.b, z1.b\n"
       ".inst 0x450198f3  // smmla z19.s, z7.b, z1.b\n"
-      ".inst 0x450198bb  // smmla z27.s, z5.b, z1.b\n"
+      ".inst 0x4501989b  // smmla z27.s, z4.b, z1.b\n"
       ".inst 0x450098cf  // smmla z15.s, z6.b, z0.b\n"
       ".inst 0x450098f7  // smmla z23.s, z7.b, z0.b\n"
-      ".inst 0x450098bf  // smmla z31.s, z5.b, z0.b\n"
+      ".inst 0x4500989f  // smmla z31.s, z4.b, z0.b\n"
       "bgt 59b\n"
       "60:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z2.b }, p2/Z, [x10]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z6.b }, p0/Z, [x25]\n"
+      "ld1rqb { z4.b }, p0/Z, [x25]\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "ld1rqb { z2.b }, p0/Z, [x23]\n"
+      "trn1 z7.d, z1.d, z4.d\n"
+      "trn2 z1.d, z1.d, z4.d\n"
       "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
+      "ld1b { z2.b }, p2/Z, [x9]\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
-      "ld1b { z0.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x450298e8  // smmla z8.s, z7.b, z2.b\n"
       ".inst 0x450298d0  // smmla z16.s, z6.b, z2.b\n"
       ".inst 0x45029898  // smmla z24.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x450098ec  // smmla z12.s, z7.b, z0.b\n"
       ".inst 0x450098d4  // smmla z20.s, z6.b, z0.b\n"
       ".inst 0x4500989c  // smmla z28.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x450298e9  // smmla z9.s, z7.b, z2.b\n"
       ".inst 0x450298d1  // smmla z17.s, z6.b, z2.b\n"
       ".inst 0x45029899  // smmla z25.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9, #4, MUL VL]\n"
       ".inst 0x450098ed  // smmla z13.s, z7.b, z0.b\n"
       ".inst 0x450098d5  // smmla z21.s, z6.b, z0.b\n"
       ".inst 0x4500989d  // smmla z29.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x450298ea  // smmla z10.s, z7.b, z2.b\n"
       ".inst 0x450298d2  // smmla z18.s, z6.b, z2.b\n"
       ".inst 0x4502989a  // smmla z26.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9, #6, MUL VL]\n"
       ".inst 0x450098ee  // smmla z14.s, z7.b, z0.b\n"
       ".inst 0x450098d6  // smmla z22.s, z6.b, z0.b\n"
       ".inst 0x4500989e  // smmla z30.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
       ".inst 0x450298eb  // smmla z11.s, z7.b, z2.b\n"
+      "addvl x9, x9, #8\n"
       ".inst 0x450298d3  // smmla z19.s, z6.b, z2.b\n"
       ".inst 0x4502989b  // smmla z27.s, z4.b, z2.b\n"
       ".inst 0x450098ef  // smmla z15.s, z7.b, z0.b\n"
       ".inst 0x450098d7  // smmla z23.s, z6.b, z0.b\n"
       ".inst 0x4500989f  // smmla z31.s, z4.b, z0.b\n"
       "ble 61f\n"
-      "ld1b { z2.b }, p2/Z, [x10]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45029828  // smmla z8.s, z1.b, z2.b\n"
       ".inst 0x45029870  // smmla z16.s, z3.b, z2.b\n"
       ".inst 0x450298b8  // smmla z24.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4500982c  // smmla z12.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p2/Z, [x9, #2, MUL VL]\n"
       ".inst 0x45009874  // smmla z20.s, z3.b, z0.b\n"
       ".inst 0x450098bc  // smmla z28.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45029829  // smmla z9.s, z1.b, z2.b\n"
       ".inst 0x45029871  // smmla z17.s, z3.b, z2.b\n"
       ".inst 0x450298b9  // smmla z25.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #4, MUL VL]\n"
       ".inst 0x4500982d  // smmla z13.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p2/Z, [x9, #4, MUL VL]\n"
       ".inst 0x45009875  // smmla z21.s, z3.b, z0.b\n"
       ".inst 0x450098bd  // smmla z29.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x4502982a  // smmla z10.s, z1.b, z2.b\n"
       ".inst 0x45029872  // smmla z18.s, z3.b, z2.b\n"
       ".inst 0x450298ba  // smmla z26.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #6, MUL VL]\n"
       ".inst 0x4500982e  // smmla z14.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p2/Z, [x9, #6, MUL VL]\n"
       ".inst 0x45009876  // smmla z22.s, z3.b, z0.b\n"
       ".inst 0x450098be  // smmla z30.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #8\n"
       ".inst 0x4502982b  // smmla z11.s, z1.b, z2.b\n"
       ".inst 0x45029873  // smmla z19.s, z3.b, z2.b\n"
       ".inst 0x450298bb  // smmla z27.s, z5.b, z2.b\n"
@@ -1583,27 +1583,27 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z4.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
-      "ld1w { z3.s }, p2/Z, [x14]\n"
+      "add x26, x11, x20\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
+      "ld1w { z3.s }, p2/Z, [x14]\n"
       "ld1w { z2.s }, p2/Z, [x14, #1, MUL VL]\n"
-      "ld1w { z1.s }, p2/Z, [x14, #2, MUL VL]\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
+      "ld1w { z1.s }, p2/Z, [x14, #2, MUL VL]\n"
       "ld1w { z0.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "add x26, x9, x20\n"
+      "add x25, x26, x20\n"
+      "add x24, x25, x20\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x25, x26, x20\n"
+      "add x23, x24, x20\n"
+      "addvl x14, x14, #4\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "add x24, x25, x20\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "add x23, x24, x20\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
       "uzp1 z24.d, z24.d, z28.d\n"
@@ -1611,10 +1611,10 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "uzp1 z26.d, z26.d, z30.d\n"
       "uzp1 z27.d, z27.d, z31.d\n"
       "mov z31.d, z4.d\n"
+      "add z31.s, z31.s, z3.s\n"
       "add z12.s, z12.s, z2.s\n"
       "add z13.s, z13.s, z1.s\n"
       "add z14.s, z14.s, z0.s\n"
-      "add z31.s, z31.s, z3.s\n"
       "add z8.s, z8.s, z3.s\n"
       "add z9.s, z9.s, z2.s\n"
       "add z10.s, z10.s, z1.s\n"
@@ -1685,11 +1685,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z28.s, z28.s, #0x1f\n"
       "asr z23.s, z23.s, #0x1f\n"
       "sqadd z31.s, z31.s, z30.s\n"
-      "and z30.d, z8.d, z0.d\n"
       "sqadd z12.s, z12.s, z29.s\n"
-      "and z29.d, z9.d, z1.d\n"
       "sqadd z13.s, z13.s, z28.s\n"
       "sqadd z14.s, z14.s, z23.s\n"
+      "and z30.d, z8.d, z0.d\n"
+      "and z29.d, z9.d, z1.d\n"
       "and z28.d, z10.d, z2.d\n"
       "and z23.d, z11.d, z3.d\n"
       "asr z30.s, z30.s, #0x1f\n"
@@ -1697,11 +1697,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z28.s, z28.s, #0x1f\n"
       "asr z23.s, z23.s, #0x1f\n"
       "sqadd z8.s, z8.s, z30.s\n"
-      "and z30.d, z15.d, z0.d\n"
       "sqadd z9.s, z9.s, z29.s\n"
-      "and z29.d, z20.d, z1.d\n"
       "sqadd z10.s, z10.s, z28.s\n"
       "sqadd z11.s, z11.s, z23.s\n"
+      "and z30.d, z15.d, z0.d\n"
+      "and z29.d, z20.d, z1.d\n"
       "and z28.d, z21.d, z2.d\n"
       "and z23.d, z22.d, z3.d\n"
       "asr z30.s, z30.s, #0x1f\n"
@@ -1709,11 +1709,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z28.s, z28.s, #0x1f\n"
       "asr z23.s, z23.s, #0x1f\n"
       "sqadd z15.s, z15.s, z30.s\n"
-      "and z30.d, z16.d, z0.d\n"
       "sqadd z20.s, z20.s, z29.s\n"
-      "and z29.d, z17.d, z1.d\n"
       "sqadd z21.s, z21.s, z28.s\n"
       "sqadd z22.s, z22.s, z23.s\n"
+      "and z30.d, z16.d, z0.d\n"
+      "and z29.d, z17.d, z1.d\n"
       "and z28.d, z18.d, z2.d\n"
       "and z23.d, z19.d, z3.d\n"
       "asr z30.s, z30.s, #0x1f\n"
@@ -1721,11 +1721,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z28.s, z28.s, #0x1f\n"
       "asr z23.s, z23.s, #0x1f\n"
       "sqadd z16.s, z16.s, z30.s\n"
-      "and z30.d, z24.d, z0.d\n"
       "sqadd z17.s, z17.s, z29.s\n"
-      "and z29.d, z25.d, z1.d\n"
       "sqadd z18.s, z18.s, z28.s\n"
       "sqadd z19.s, z19.s, z23.s\n"
+      "and z30.d, z24.d, z0.d\n"
+      "and z29.d, z25.d, z1.d\n"
       "and z28.d, z26.d, z2.d\n"
       "and z23.d, z27.d, z3.d\n"
       "asr z30.s, z30.s, #0x1f\n"
@@ -1738,51 +1738,51 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "sqadd z27.s, z27.s, z23.s\n"
       "64:"  // Height 5: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
+      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
+      "add z31.s, z31.s, z28.s\n"
       ".inst 0x4482882c  // srshl z12.s, p2/M, z12.s, z1.s\n"
       ".inst 0x4482884d  // srshl z13.s, p2/M, z13.s, z2.s\n"
+      "add z12.s, z12.s, z28.s\n"
+      "add z13.s, z13.s, z28.s\n"
       ".inst 0x4482886e  // srshl z14.s, p2/M, z14.s, z3.s\n"
       ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
+      "add z14.s, z14.s, z28.s\n"
+      "add z8.s, z8.s, z28.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
-      "add z31.s, z31.s, z28.s\n"
+      "add z9.s, z9.s, z28.s\n"
+      "add z10.s, z10.s, z28.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x4482880f  // srshl z15.s, p2/M, z15.s, z0.s\n"
-      "add z12.s, z12.s, z28.s\n"
-      "add z13.s, z13.s, z28.s\n"
+      "add z11.s, z11.s, z28.s\n"
+      "add z15.s, z15.s, z28.s\n"
       ".inst 0x44828834  // srshl z20.s, p2/M, z20.s, z1.s\n"
       ".inst 0x44828855  // srshl z21.s, p2/M, z21.s, z2.s\n"
-      "add z14.s, z14.s, z28.s\n"
-      "add z8.s, z8.s, z28.s\n"
+      "add z20.s, z20.s, z28.s\n"
+      "add z21.s, z21.s, z28.s\n"
       ".inst 0x44828876  // srshl z22.s, p2/M, z22.s, z3.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "add z9.s, z9.s, z28.s\n"
-      "add z10.s, z10.s, z28.s\n"
+      "add z22.s, z22.s, z28.s\n"
+      "add z16.s, z16.s, z28.s\n"
       ".inst 0x44828831  // srshl z17.s, p2/M, z17.s, z1.s\n"
       ".inst 0x44828852  // srshl z18.s, p2/M, z18.s, z2.s\n"
-      "add z11.s, z11.s, z28.s\n"
-      "add z15.s, z15.s, z28.s\n"
+      "add z17.s, z17.s, z28.s\n"
+      "add z18.s, z18.s, z28.s\n"
       ".inst 0x44828873  // srshl z19.s, p2/M, z19.s, z3.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z20.s, z20.s, z28.s\n"
-      "add z21.s, z21.s, z28.s\n"
+      "add z19.s, z19.s, z28.s\n"
+      "add z24.s, z24.s, z28.s\n"
       ".inst 0x44828839  // srshl z25.s, p2/M, z25.s, z1.s\n"
       ".inst 0x4482885a  // srshl z26.s, p2/M, z26.s, z2.s\n"
-      "add z22.s, z22.s, z28.s\n"
-      "add z16.s, z16.s, z28.s\n"
+      "add z25.s, z25.s, z28.s\n"
+      "add z26.s, z26.s, z28.s\n"
       ".inst 0x4482887b  // srshl z27.s, p2/M, z27.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z17.s, z17.s, z28.s\n"
-      "add z18.s, z18.s, z28.s\n"
       "ld1rw { z23.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z28.s\n"
-      "add z24.s, z24.s, z28.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z28.s\n"
-      "add z26.s, z26.s, z28.s\n"
-      "ld1rw { z29.s }, p2/Z, [x20]\n"
       "add z27.s, z27.s, z28.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z28.s }, p2/Z, [x20]\n"
       "smin z31.s, p2/M, z31.s, z23.s\n"
       "smin z12.s, p2/M, z12.s, z23.s\n"
       "smin z13.s, p2/M, z13.s, z23.s\n"
@@ -1803,72 +1803,71 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "smin z25.s, p2/M, z25.s, z23.s\n"
       "smin z26.s, p2/M, z26.s, z23.s\n"
       "smin z27.s, p2/M, z27.s, z23.s\n"
-      "smax z31.s, p2/M, z31.s, z29.s\n"
-      "smax z12.s, p2/M, z12.s, z29.s\n"
-      "smax z13.s, p2/M, z13.s, z29.s\n"
-      "smax z14.s, p2/M, z14.s, z29.s\n"
-      "smax z8.s, p2/M, z8.s, z29.s\n"
-      "smax z9.s, p2/M, z9.s, z29.s\n"
-      "smax z10.s, p2/M, z10.s, z29.s\n"
-      "smax z11.s, p2/M, z11.s, z29.s\n"
+      "smax z31.s, p2/M, z31.s, z28.s\n"
+      "smax z12.s, p2/M, z12.s, z28.s\n"
+      "smax z13.s, p2/M, z13.s, z28.s\n"
       "uzp1 z31.h, z31.h, z12.h\n"
-      "smax z15.s, p2/M, z15.s, z29.s\n"
-      "smax z20.s, p2/M, z20.s, z29.s\n"
-      "uzp1 z28.h, z13.h, z14.h\n"
-      "smax z21.s, p2/M, z21.s, z29.s\n"
-      "smax z22.s, p2/M, z22.s, z29.s\n"
+      "smax z14.s, p2/M, z14.s, z28.s\n"
+      "smax z8.s, p2/M, z8.s, z28.s\n"
+      "uzp1 z23.h, z13.h, z14.h\n"
+      "uzp1 z31.b, z31.b, z23.b\n"
+      "smax z9.s, p2/M, z9.s, z28.s\n"
+      "smax z10.s, p2/M, z10.s, z28.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
-      "smax z16.s, p2/M, z16.s, z29.s\n"
-      "smax z17.s, p2/M, z17.s, z29.s\n"
+      "st1b { z31.b }, p1, [x11]\n"
+      "smax z11.s, p2/M, z11.s, z28.s\n"
+      "smax z15.s, p2/M, z15.s, z28.s\n"
       "uzp1 z23.h, z10.h, z11.h\n"
-      "smax z18.s, p2/M, z18.s, z29.s\n"
-      "smax z19.s, p2/M, z19.s, z29.s\n"
+      "uzp1 z8.b, z8.b, z23.b\n"
+      "smax z20.s, p2/M, z20.s, z28.s\n"
+      "smax z21.s, p2/M, z21.s, z28.s\n"
       "uzp1 z15.h, z15.h, z20.h\n"
-      "uzp1 z31.b, z31.b, z28.b\n"
-      "smax z24.s, p2/M, z24.s, z29.s\n"
-      "smax z25.s, p2/M, z25.s, z29.s\n"
+      "st1b { z8.b }, p1, [x26]\n"
+      "smax z22.s, p2/M, z22.s, z28.s\n"
+      "smax z16.s, p2/M, z16.s, z28.s\n"
       "uzp1 z20.h, z21.h, z22.h\n"
-      "smax z26.s, p2/M, z26.s, z29.s\n"
-      "smax z27.s, p2/M, z27.s, z29.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z8.b, z8.b, z23.b\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
-      "st1b { z31.b }, p1, [x9]\n"
-      "addvl x9, x9, #1\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
       "uzp1 z15.b, z15.b, z20.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "st1b { z8.b }, p1, [x26]\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
+      "smax z17.s, p2/M, z17.s, z28.s\n"
+      "smax z18.s, p2/M, z18.s, z28.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "st1b { z15.b }, p1, [x25]\n"
+      "smax z19.s, p2/M, z19.s, z28.s\n"
+      "smax z24.s, p2/M, z24.s, z28.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
+      "smax z25.s, p2/M, z25.s, z28.s\n"
+      "smax z26.s, p2/M, z26.s, z28.s\n"
+      "uzp1 z24.h, z24.h, z25.h\n"
       "st1b { z16.b }, p1, [x24]\n"
+      "smax z27.s, p2/M, z27.s, z28.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "st1b { z24.b }, p1, [x23]\n"
+      "addvl x11, x11, #1\n"
       "65:"  // Height 5: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 54b\n"
       "b 80f\n"
       "66:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x6\n"
       "mov x14, %x[col_bias]\n"
       "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n"
       "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x10, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x11, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
+      "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n"
       "67:"  // Height 6: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x10\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
       "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
-      "whilelt p1.b, x20, x11\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
       "mov z16.s, #0x0\n"
@@ -1891,8 +1890,8 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "mov x28, #0x0\n"
       "69:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 70f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1923,168 +1922,168 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "ble 73f\n"
       "72:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p2/Z, [x10]\n"
+      "ld1rqb { z7.b }, p0/Z, [x26]\n"
+      "ld1rqb { z0.b }, p0/Z, [x25]\n"
+      "trn1 z6.d, z7.d, z0.d\n"
+      "ld1rqb { z5.b }, p0/Z, [x24]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "trn2 z7.d, z7.d, z0.d\n"
+      "trn1 z4.d, z5.d, z1.d\n"
+      "ld1rqb { z3.b }, p0/Z, [x22]\n"
+      "ld1rqb { z0.b }, p0/Z, [x21]\n"
+      "trn2 z5.d, z5.d, z1.d\n"
+      "trn1 z2.d, z3.d, z0.d\n"
+      "trn2 z3.d, z3.d, z0.d\n"
+      "ld1b { z1.b }, p2/Z, [x9]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
+      ".inst 0x450198c8  // smmla z8.s, z6.b, z1.b\n"
+      ".inst 0x45019890  // smmla z16.s, z4.b, z1.b\n"
+      ".inst 0x45019858  // smmla z24.s, z2.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #2, MUL VL]\n"
       "sub x27, x27, #0x10\n"
+      ".inst 0x450098cc  // smmla z12.s, z6.b, z0.b\n"
+      ".inst 0x45009894  // smmla z20.s, z4.b, z0.b\n"
       "cmp x27, #0x10\n"
-      "ld1rqb { z6.b }, p0/Z, [x26]\n"
       "add x26, x26, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x25]\n"
+      ".inst 0x4500985c  // smmla z28.s, z2.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
+      ".inst 0x450198c9  // smmla z9.s, z6.b, z1.b\n"
       "add x25, x25, #0x10\n"
-      "ld1rqb { z7.b }, p0/Z, [x24]\n"
+      ".inst 0x45019891  // smmla z17.s, z4.b, z1.b\n"
+      ".inst 0x45019859  // smmla z25.s, z2.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #4, MUL VL]\n"
       "add x24, x24, #0x10\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "ld1rqb { z0.b }, p0/Z, [x21]\n"
+      ".inst 0x450098cd  // smmla z13.s, z6.b, z0.b\n"
+      ".inst 0x45009895  // smmla z21.s, z4.b, z0.b\n"
       "add x23, x23, #0x10\n"
       "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
-      "add x21, x21, #0x10\n"
-      "trn1 z3.d, z7.d, z2.d\n"
-      "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
-      "ld1b { z0.b }, p2/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45019888  // smmla z8.s, z4.b, z1.b\n"
-      ".inst 0x45019870  // smmla z16.s, z3.b, z1.b\n"
-      ".inst 0x45019858  // smmla z24.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4500988c  // smmla z12.s, z4.b, z0.b\n"
-      ".inst 0x45009874  // smmla z20.s, z3.b, z0.b\n"
-      ".inst 0x4500985c  // smmla z28.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45019889  // smmla z9.s, z4.b, z1.b\n"
-      ".inst 0x45019871  // smmla z17.s, z3.b, z1.b\n"
-      ".inst 0x45019859  // smmla z25.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4500988d  // smmla z13.s, z4.b, z0.b\n"
-      ".inst 0x45009875  // smmla z21.s, z3.b, z0.b\n"
       ".inst 0x4500985d  // smmla z29.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4501988a  // smmla z10.s, z4.b, z1.b\n"
-      ".inst 0x45019872  // smmla z18.s, z3.b, z1.b\n"
-      ".inst 0x4501985a  // smmla z26.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4500988e  // smmla z14.s, z4.b, z0.b\n"
-      ".inst 0x45009876  // smmla z22.s, z3.b, z0.b\n"
-      ".inst 0x4500985e  // smmla z30.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x4501988b  // smmla z11.s, z4.b, z1.b\n"
-      ".inst 0x45019873  // smmla z19.s, z3.b, z1.b\n"
-      ".inst 0x4501985b  // smmla z27.s, z2.b, z1.b\n"
-      ".inst 0x4500988f  // smmla z15.s, z4.b, z0.b\n"
-      ".inst 0x45009877  // smmla z23.s, z3.b, z0.b\n"
-      ".inst 0x4500985f  // smmla z31.s, z2.b, z0.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x450198c8  // smmla z8.s, z6.b, z1.b\n"
-      ".inst 0x450198f0  // smmla z16.s, z7.b, z1.b\n"
-      ".inst 0x450198b8  // smmla z24.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x450098cc  // smmla z12.s, z6.b, z0.b\n"
-      ".inst 0x450098f4  // smmla z20.s, z7.b, z0.b\n"
-      ".inst 0x450098bc  // smmla z28.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x450198c9  // smmla z9.s, z6.b, z1.b\n"
-      ".inst 0x450198f1  // smmla z17.s, z7.b, z1.b\n"
-      ".inst 0x450198b9  // smmla z25.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x450098cd  // smmla z13.s, z6.b, z0.b\n"
-      ".inst 0x450098f5  // smmla z21.s, z7.b, z0.b\n"
-      ".inst 0x450098bd  // smmla z29.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x450198ca  // smmla z10.s, z6.b, z1.b\n"
-      ".inst 0x450198f2  // smmla z18.s, z7.b, z1.b\n"
-      ".inst 0x450198ba  // smmla z26.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p2/Z, [x10, #-2, MUL VL]\n"
+      "add x21, x21, #0x10\n"
+      ".inst 0x45019892  // smmla z18.s, z4.b, z1.b\n"
+      ".inst 0x4501985a  // smmla z26.s, z2.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #6, MUL VL]\n"
       ".inst 0x450098ce  // smmla z14.s, z6.b, z0.b\n"
-      ".inst 0x450098f6  // smmla z22.s, z7.b, z0.b\n"
-      ".inst 0x450098be  // smmla z30.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #-1, MUL VL]\n"
+      ".inst 0x45009896  // smmla z22.s, z4.b, z0.b\n"
+      ".inst 0x4500985e  // smmla z30.s, z2.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #16\n"
       ".inst 0x450198cb  // smmla z11.s, z6.b, z1.b\n"
-      ".inst 0x450198f3  // smmla z19.s, z7.b, z1.b\n"
-      ".inst 0x450198bb  // smmla z27.s, z5.b, z1.b\n"
+      ".inst 0x45019893  // smmla z19.s, z4.b, z1.b\n"
+      ".inst 0x4501985b  // smmla z27.s, z2.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-8, MUL VL]\n"
       ".inst 0x450098cf  // smmla z15.s, z6.b, z0.b\n"
-      ".inst 0x450098f7  // smmla z23.s, z7.b, z0.b\n"
-      ".inst 0x450098bf  // smmla z31.s, z5.b, z0.b\n"
+      ".inst 0x45009897  // smmla z23.s, z4.b, z0.b\n"
+      ".inst 0x4500985f  // smmla z31.s, z2.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-7, MUL VL]\n"
+      ".inst 0x450198e8  // smmla z8.s, z7.b, z1.b\n"
+      ".inst 0x450198b0  // smmla z16.s, z5.b, z1.b\n"
+      ".inst 0x45019878  // smmla z24.s, z3.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-6, MUL VL]\n"
+      ".inst 0x450098ec  // smmla z12.s, z7.b, z0.b\n"
+      ".inst 0x450098b4  // smmla z20.s, z5.b, z0.b\n"
+      ".inst 0x4500987c  // smmla z28.s, z3.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-5, MUL VL]\n"
+      ".inst 0x450198e9  // smmla z9.s, z7.b, z1.b\n"
+      ".inst 0x450198b1  // smmla z17.s, z5.b, z1.b\n"
+      ".inst 0x45019879  // smmla z25.s, z3.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-4, MUL VL]\n"
+      ".inst 0x450098ed  // smmla z13.s, z7.b, z0.b\n"
+      ".inst 0x450098b5  // smmla z21.s, z5.b, z0.b\n"
+      ".inst 0x4500987d  // smmla z29.s, z3.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-3, MUL VL]\n"
+      ".inst 0x450198ea  // smmla z10.s, z7.b, z1.b\n"
+      ".inst 0x450198b2  // smmla z18.s, z5.b, z1.b\n"
+      ".inst 0x4501987a  // smmla z26.s, z3.b, z1.b\n"
+      "ld1b { z1.b }, p2/Z, [x9, #-2, MUL VL]\n"
+      ".inst 0x450098ee  // smmla z14.s, z7.b, z0.b\n"
+      ".inst 0x450098b6  // smmla z22.s, z5.b, z0.b\n"
+      ".inst 0x4500987e  // smmla z30.s, z3.b, z0.b\n"
+      "ld1b { z0.b }, p2/Z, [x9, #-1, MUL VL]\n"
+      ".inst 0x450198eb  // smmla z11.s, z7.b, z1.b\n"
+      ".inst 0x450198b3  // smmla z19.s, z5.b, z1.b\n"
+      ".inst 0x4501987b  // smmla z27.s, z3.b, z1.b\n"
+      ".inst 0x450098ef  // smmla z15.s, z7.b, z0.b\n"
+      ".inst 0x450098b7  // smmla z23.s, z5.b, z0.b\n"
+      ".inst 0x4500987f  // smmla z31.s, z3.b, z0.b\n"
       "bgt 72b\n"
       "73:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z2.b }, p2/Z, [x10]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z6.b }, p0/Z, [x25]\n"
+      "ld1rqb { z0.b }, p0/Z, [x25]\n"
+      "trn1 z7.d, z1.d, z0.d\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "ld1rqb { z2.b }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z0.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
       "ld1rqb { z5.b }, p0/Z, [x22]\n"
       "ld1rqb { z0.b }, p0/Z, [x21]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
-      "ld1b { z0.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x450298e8  // smmla z8.s, z7.b, z2.b\n"
       ".inst 0x450298d0  // smmla z16.s, z6.b, z2.b\n"
       ".inst 0x45029898  // smmla z24.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9, #2, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x450098ec  // smmla z12.s, z7.b, z0.b\n"
       ".inst 0x450098d4  // smmla z20.s, z6.b, z0.b\n"
       ".inst 0x4500989c  // smmla z28.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x450298e9  // smmla z9.s, z7.b, z2.b\n"
       ".inst 0x450298d1  // smmla z17.s, z6.b, z2.b\n"
       ".inst 0x45029899  // smmla z25.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #4, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9, #4, MUL VL]\n"
       ".inst 0x450098ed  // smmla z13.s, z7.b, z0.b\n"
       ".inst 0x450098d5  // smmla z21.s, z6.b, z0.b\n"
       ".inst 0x4500989d  // smmla z29.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x450298ea  // smmla z10.s, z7.b, z2.b\n"
       ".inst 0x450298d2  // smmla z18.s, z6.b, z2.b\n"
       ".inst 0x4502989a  // smmla z26.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #6, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9, #6, MUL VL]\n"
       ".inst 0x450098ee  // smmla z14.s, z7.b, z0.b\n"
       ".inst 0x450098d6  // smmla z22.s, z6.b, z0.b\n"
       ".inst 0x4500989e  // smmla z30.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
       ".inst 0x450298eb  // smmla z11.s, z7.b, z2.b\n"
+      "addvl x9, x9, #8\n"
       ".inst 0x450298d3  // smmla z19.s, z6.b, z2.b\n"
       ".inst 0x4502989b  // smmla z27.s, z4.b, z2.b\n"
       ".inst 0x450098ef  // smmla z15.s, z7.b, z0.b\n"
       ".inst 0x450098d7  // smmla z23.s, z6.b, z0.b\n"
       ".inst 0x4500989f  // smmla z31.s, z4.b, z0.b\n"
       "ble 74f\n"
-      "ld1b { z2.b }, p2/Z, [x10]\n"
-      "ld1b { z0.b }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z2.b }, p2/Z, [x9]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #1, MUL VL]\n"
       ".inst 0x45029828  // smmla z8.s, z1.b, z2.b\n"
       ".inst 0x45029870  // smmla z16.s, z3.b, z2.b\n"
       ".inst 0x450298b8  // smmla z24.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4500982c  // smmla z12.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p2/Z, [x9, #2, MUL VL]\n"
       ".inst 0x45009874  // smmla z20.s, z3.b, z0.b\n"
       ".inst 0x450098bc  // smmla z28.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #3, MUL VL]\n"
       ".inst 0x45029829  // smmla z9.s, z1.b, z2.b\n"
       ".inst 0x45029871  // smmla z17.s, z3.b, z2.b\n"
       ".inst 0x450298b9  // smmla z25.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #4, MUL VL]\n"
       ".inst 0x4500982d  // smmla z13.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p2/Z, [x9, #4, MUL VL]\n"
       ".inst 0x45009875  // smmla z21.s, z3.b, z0.b\n"
       ".inst 0x450098bd  // smmla z29.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #5, MUL VL]\n"
+      "ld1b { z0.b }, p2/Z, [x9, #5, MUL VL]\n"
       ".inst 0x4502982a  // smmla z10.s, z1.b, z2.b\n"
       ".inst 0x45029872  // smmla z18.s, z3.b, z2.b\n"
       ".inst 0x450298ba  // smmla z26.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x10, #6, MUL VL]\n"
       ".inst 0x4500982e  // smmla z14.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p2/Z, [x9, #6, MUL VL]\n"
       ".inst 0x45009876  // smmla z22.s, z3.b, z0.b\n"
       ".inst 0x450098be  // smmla z30.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p2/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "ld1b { z0.b }, p2/Z, [x9, #7, MUL VL]\n"
+      "addvl x9, x9, #8\n"
       ".inst 0x4502982b  // smmla z11.s, z1.b, z2.b\n"
       ".inst 0x45029873  // smmla z19.s, z3.b, z2.b\n"
       ".inst 0x450298bb  // smmla z27.s, z5.b, z2.b\n"
@@ -2098,31 +2097,31 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "bne 69b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z4.d, z8.d, z12.d\n"
+      "add x26, x11, x20\n"
       "uzp2 z8.d, z8.d, z12.d\n"
-      "ld1w { z3.s }, p2/Z, [x14]\n"
       "uzp1 z12.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
-      "ld1w { z2.s }, p2/Z, [x14, #1, MUL VL]\n"
-      "ld1w { z1.s }, p2/Z, [x14, #2, MUL VL]\n"
+      "add x25, x26, x20\n"
+      "ld1w { z3.s }, p2/Z, [x14]\n"
       "uzp1 z13.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "ld1w { z0.s }, p2/Z, [x14, #3, MUL VL]\n"
-      "addvl x14, x14, #4\n"
-      "add x26, x9, x20\n"
+      "ld1w { z2.s }, p2/Z, [x14, #1, MUL VL]\n"
+      "ld1w { z1.s }, p2/Z, [x14, #2, MUL VL]\n"
       "uzp1 z14.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
+      "ld1w { z0.s }, p2/Z, [x14, #3, MUL VL]\n"
+      "add x24, x25, x20\n"
       "uzp1 z15.d, z16.d, z20.d\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "add x25, x26, x20\n"
+      "add x23, x24, x20\n"
+      "add x22, x23, x20\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "add x24, x25, x20\n"
+      "addvl x14, x14, #4\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "add x23, x24, x20\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "add x22, x23, x20\n"
       "uzp1 z23.d, z24.d, z28.d\n"
       "uzp2 z24.d, z24.d, z28.d\n"
       "uzp1 z28.d, z25.d, z29.d\n"
@@ -2132,10 +2131,10 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "uzp1 z30.d, z27.d, z31.d\n"
       "uzp2 z27.d, z27.d, z31.d\n"
       "mov z31.d, z4.d\n"
+      "add z31.s, z31.s, z3.s\n"
       "add z12.s, z12.s, z2.s\n"
       "add z13.s, z13.s, z1.s\n"
       "add z14.s, z14.s, z0.s\n"
-      "add z31.s, z31.s, z3.s\n"
       "add z8.s, z8.s, z3.s\n"
       "add z9.s, z9.s, z2.s\n"
       "add z10.s, z10.s, z1.s\n"
@@ -2214,11 +2213,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z31.s, z31.s, z7.s\n"
-      "and z7.d, z8.d, z0.d\n"
       "sqadd z12.s, z12.s, z6.s\n"
-      "and z6.d, z9.d, z1.d\n"
       "sqadd z13.s, z13.s, z5.s\n"
       "sqadd z14.s, z14.s, z4.s\n"
+      "and z7.d, z8.d, z0.d\n"
+      "and z6.d, z9.d, z1.d\n"
       "and z5.d, z10.d, z2.d\n"
       "and z4.d, z11.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2226,11 +2225,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z8.s, z8.s, z7.s\n"
-      "and z7.d, z15.d, z0.d\n"
       "sqadd z9.s, z9.s, z6.s\n"
-      "and z6.d, z20.d, z1.d\n"
       "sqadd z10.s, z10.s, z5.s\n"
       "sqadd z11.s, z11.s, z4.s\n"
+      "and z7.d, z15.d, z0.d\n"
+      "and z6.d, z20.d, z1.d\n"
       "and z5.d, z21.d, z2.d\n"
       "and z4.d, z22.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2238,11 +2237,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z15.s, z15.s, z7.s\n"
-      "and z7.d, z16.d, z0.d\n"
       "sqadd z20.s, z20.s, z6.s\n"
-      "and z6.d, z17.d, z1.d\n"
       "sqadd z21.s, z21.s, z5.s\n"
       "sqadd z22.s, z22.s, z4.s\n"
+      "and z7.d, z16.d, z0.d\n"
+      "and z6.d, z17.d, z1.d\n"
       "and z5.d, z18.d, z2.d\n"
       "and z4.d, z19.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2250,11 +2249,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z16.s, z16.s, z7.s\n"
-      "and z7.d, z23.d, z0.d\n"
       "sqadd z17.s, z17.s, z6.s\n"
-      "and z6.d, z28.d, z1.d\n"
       "sqadd z18.s, z18.s, z5.s\n"
       "sqadd z19.s, z19.s, z4.s\n"
+      "and z7.d, z23.d, z0.d\n"
+      "and z6.d, z28.d, z1.d\n"
       "and z5.d, z29.d, z2.d\n"
       "and z4.d, z30.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2262,11 +2261,11 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "asr z5.s, z5.s, #0x1f\n"
       "asr z4.s, z4.s, #0x1f\n"
       "sqadd z23.s, z23.s, z7.s\n"
-      "and z7.d, z24.d, z0.d\n"
       "sqadd z28.s, z28.s, z6.s\n"
-      "and z6.d, z25.d, z1.d\n"
       "sqadd z29.s, z29.s, z5.s\n"
       "sqadd z30.s, z30.s, z4.s\n"
+      "and z7.d, z24.d, z0.d\n"
+      "and z6.d, z25.d, z1.d\n"
       "and z5.d, z26.d, z2.d\n"
       "and z4.d, z27.d, z3.d\n"
       "asr z7.s, z7.s, #0x1f\n"
@@ -2279,59 +2278,59 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "sqadd z27.s, z27.s, z4.s\n"
       "77:"  // Height 6: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
       "ld1rw { z4.s }, p2/Z, [x20]\n"
+      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
+      "add z31.s, z31.s, z4.s\n"
       ".inst 0x4482882c  // srshl z12.s, p2/M, z12.s, z1.s\n"
       ".inst 0x4482884d  // srshl z13.s, p2/M, z13.s, z2.s\n"
+      "add z12.s, z12.s, z4.s\n"
+      "add z13.s, z13.s, z4.s\n"
       ".inst 0x4482886e  // srshl z14.s, p2/M, z14.s, z3.s\n"
       ".inst 0x44828808  // srshl z8.s, p2/M, z8.s, z0.s\n"
+      "add z14.s, z14.s, z4.s\n"
+      "add z8.s, z8.s, z4.s\n"
       ".inst 0x44828829  // srshl z9.s, p2/M, z9.s, z1.s\n"
       ".inst 0x4482884a  // srshl z10.s, p2/M, z10.s, z2.s\n"
-      "add z31.s, z31.s, z4.s\n"
+      "add z9.s, z9.s, z4.s\n"
+      "add z10.s, z10.s, z4.s\n"
       ".inst 0x4482886b  // srshl z11.s, p2/M, z11.s, z3.s\n"
       ".inst 0x4482880f  // srshl z15.s, p2/M, z15.s, z0.s\n"
-      "add z12.s, z12.s, z4.s\n"
-      "add z13.s, z13.s, z4.s\n"
+      "add z11.s, z11.s, z4.s\n"
+      "add z15.s, z15.s, z4.s\n"
       ".inst 0x44828834  // srshl z20.s, p2/M, z20.s, z1.s\n"
       ".inst 0x44828855  // srshl z21.s, p2/M, z21.s, z2.s\n"
-      "add z14.s, z14.s, z4.s\n"
-      "add z8.s, z8.s, z4.s\n"
+      "add z20.s, z20.s, z4.s\n"
+      "add z21.s, z21.s, z4.s\n"
       ".inst 0x44828876  // srshl z22.s, p2/M, z22.s, z3.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "add z9.s, z9.s, z4.s\n"
-      "add z10.s, z10.s, z4.s\n"
+      "add z22.s, z22.s, z4.s\n"
+      "add z16.s, z16.s, z4.s\n"
       ".inst 0x44828831  // srshl z17.s, p2/M, z17.s, z1.s\n"
       ".inst 0x44828852  // srshl z18.s, p2/M, z18.s, z2.s\n"
-      "add z11.s, z11.s, z4.s\n"
-      "add z15.s, z15.s, z4.s\n"
+      "add z17.s, z17.s, z4.s\n"
+      "add z18.s, z18.s, z4.s\n"
       ".inst 0x44828873  // srshl z19.s, p2/M, z19.s, z3.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "add z20.s, z20.s, z4.s\n"
-      "add z21.s, z21.s, z4.s\n"
+      "add z19.s, z19.s, z4.s\n"
+      "add z23.s, z23.s, z4.s\n"
       ".inst 0x4482883c  // srshl z28.s, p2/M, z28.s, z1.s\n"
       ".inst 0x4482885d  // srshl z29.s, p2/M, z29.s, z2.s\n"
-      "add z22.s, z22.s, z4.s\n"
-      "add z16.s, z16.s, z4.s\n"
+      "add z28.s, z28.s, z4.s\n"
+      "add z29.s, z29.s, z4.s\n"
       ".inst 0x4482887e  // srshl z30.s, p2/M, z30.s, z3.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z4.s\n"
-      "add z18.s, z18.s, z4.s\n"
+      "add z30.s, z30.s, z4.s\n"
+      "add z24.s, z24.s, z4.s\n"
       ".inst 0x44828839  // srshl z25.s, p2/M, z25.s, z1.s\n"
       ".inst 0x4482885a  // srshl z26.s, p2/M, z26.s, z2.s\n"
-      "add z19.s, z19.s, z4.s\n"
-      "add z23.s, z23.s, z4.s\n"
+      "add z25.s, z25.s, z4.s\n"
+      "add z26.s, z26.s, z4.s\n"
       ".inst 0x4482887b  // srshl z27.s, p2/M, z27.s, z3.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z28.s, z28.s, z4.s\n"
-      "add z29.s, z29.s, z4.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z30.s, z30.s, z4.s\n"
-      "add z24.s, z24.s, z4.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z4.s\n"
-      "add z26.s, z26.s, z4.s\n"
-      "ld1rw { z2.s }, p2/Z, [x20]\n"
       "add z27.s, z27.s, z4.s\n"
+      "add x20, %x[qp], %[minval]\n"
+      "ld1rw { z1.s }, p2/Z, [x20]\n"
       "smin z31.s, p2/M, z31.s, z0.s\n"
       "smin z12.s, p2/M, z12.s, z0.s\n"
       "smin z13.s, p2/M, z13.s, z0.s\n"
@@ -2356,58 +2355,58 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "smin z25.s, p2/M, z25.s, z0.s\n"
       "smin z26.s, p2/M, z26.s, z0.s\n"
       "smin z27.s, p2/M, z27.s, z0.s\n"
-      "smax z31.s, p2/M, z31.s, z2.s\n"
-      "smax z12.s, p2/M, z12.s, z2.s\n"
-      "smax z13.s, p2/M, z13.s, z2.s\n"
-      "smax z14.s, p2/M, z14.s, z2.s\n"
-      "smax z8.s, p2/M, z8.s, z2.s\n"
-      "smax z9.s, p2/M, z9.s, z2.s\n"
-      "smax z10.s, p2/M, z10.s, z2.s\n"
-      "smax z11.s, p2/M, z11.s, z2.s\n"
+      "smax z31.s, p2/M, z31.s, z1.s\n"
+      "smax z12.s, p2/M, z12.s, z1.s\n"
+      "smax z13.s, p2/M, z13.s, z1.s\n"
       "uzp1 z31.h, z31.h, z12.h\n"
-      "smax z15.s, p2/M, z15.s, z2.s\n"
-      "smax z20.s, p2/M, z20.s, z2.s\n"
-      "uzp1 z1.h, z13.h, z14.h\n"
-      "smax z21.s, p2/M, z21.s, z2.s\n"
-      "smax z22.s, p2/M, z22.s, z2.s\n"
+      "smax z14.s, p2/M, z14.s, z1.s\n"
+      "smax z8.s, p2/M, z8.s, z1.s\n"
+      "uzp1 z0.h, z13.h, z14.h\n"
+      "uzp1 z31.b, z31.b, z0.b\n"
+      "smax z9.s, p2/M, z9.s, z1.s\n"
+      "smax z10.s, p2/M, z10.s, z1.s\n"
       "uzp1 z8.h, z8.h, z9.h\n"
-      "smax z16.s, p2/M, z16.s, z2.s\n"
-      "smax z17.s, p2/M, z17.s, z2.s\n"
-      "uzp1 z0.h, z10.h, z11.h\n"
-      "smax z18.s, p2/M, z18.s, z2.s\n"
-      "smax z19.s, p2/M, z19.s, z2.s\n"
+      "st1b { z31.b }, p1, [x11]\n"
+      "smax z11.s, p2/M, z11.s, z1.s\n"
+      "smax z15.s, p2/M, z15.s, z1.s\n"
+      "uzp1 z31.h, z10.h, z11.h\n"
+      "uzp1 z8.b, z8.b, z31.b\n"
+      "smax z20.s, p2/M, z20.s, z1.s\n"
+      "smax z21.s, p2/M, z21.s, z1.s\n"
       "uzp1 z15.h, z15.h, z20.h\n"
-      "uzp1 z31.b, z31.b, z1.b\n"
-      "smax z23.s, p2/M, z23.s, z2.s\n"
-      "smax z28.s, p2/M, z28.s, z2.s\n"
+      "st1b { z8.b }, p1, [x26]\n"
+      "smax z22.s, p2/M, z22.s, z1.s\n"
+      "smax z16.s, p2/M, z16.s, z1.s\n"
       "uzp1 z20.h, z21.h, z22.h\n"
-      "smax z29.s, p2/M, z29.s, z2.s\n"
-      "smax z30.s, p2/M, z30.s, z2.s\n"
+      "uzp1 z15.b, z15.b, z20.b\n"
+      "smax z17.s, p2/M, z17.s, z1.s\n"
+      "smax z18.s, p2/M, z18.s, z1.s\n"
       "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z8.b, z8.b, z0.b\n"
-      "smax z24.s, p2/M, z24.s, z2.s\n"
-      "smax z25.s, p2/M, z25.s, z2.s\n"
+      "st1b { z15.b }, p1, [x25]\n"
+      "smax z19.s, p2/M, z19.s, z1.s\n"
+      "smax z23.s, p2/M, z23.s, z1.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
-      "st1b { z31.b }, p1, [x9]\n"
-      "smax z26.s, p2/M, z26.s, z2.s\n"
-      "smax z27.s, p2/M, z27.s, z2.s\n"
-      "uzp1 z23.h, z23.h, z28.h\n"
-      "uzp1 z15.b, z15.b, z20.b\n"
-      "uzp1 z18.h, z29.h, z30.h\n"
-      "st1b { z8.b }, p1, [x26]\n"
-      "addvl x9, x9, #1\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
       "uzp1 z16.b, z16.b, z17.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "st1b { z15.b }, p1, [x25]\n"
-      "uzp1 z23.b, z23.b, z18.b\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
+      "smax z28.s, p2/M, z28.s, z1.s\n"
+      "smax z29.s, p2/M, z29.s, z1.s\n"
+      "uzp1 z23.h, z23.h, z28.h\n"
       "st1b { z16.b }, p1, [x24]\n"
+      "smax z30.s, p2/M, z30.s, z1.s\n"
+      "smax z24.s, p2/M, z24.s, z1.s\n"
+      "uzp1 z16.h, z29.h, z30.h\n"
+      "uzp1 z23.b, z23.b, z16.b\n"
+      "smax z25.s, p2/M, z25.s, z1.s\n"
+      "smax z26.s, p2/M, z26.s, z1.s\n"
+      "uzp1 z24.h, z24.h, z25.h\n"
       "st1b { z23.b }, p1, [x23]\n"
+      "smax z27.s, p2/M, z27.s, z1.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "st1b { z24.b }, p1, [x22]\n"
+      "addvl x11, x11, #1\n"
       "78:"  // Height 6: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
+      "decw x10, ALL, MUL #4\n"
+      "cmp x10, XZR\n"
       "bgt 67b\n"
       "subs %x[M], %x[M], #0x6\n"
       "beq 80f\n"
@@ -2421,8 +2420,8 @@ void sve_hybrid_s8qs_mmla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "80:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL.hpp
index dc377f56e8..a6abb8d354 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 4, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp
index f94f76564b..92a350c8a8 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/a64fx.cpp
@@ -44,18 +44,18 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -89,7 +89,7 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "beq 11f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -114,8 +114,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "5:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 6f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -134,14 +134,14 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "ble 9f\n"
       "8:"  // Height 1: Multiply loop: Main loop
       "sdot z8.s, z6.b, z0.b\n"
-      "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z7.b, z0.b\n"
+      "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
       "add x26, x26, #0x4\n"
-      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b\n"
       "sdot z11.s, z16.b, z0.b\n"
+      "subs x27, x27, #0x4\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1b { z6.b }, p4/Z, [x10]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
@@ -149,14 +149,14 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "9:"  // Height 1: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "sdot z8.s, z6.b, z0.b\n"
-      "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z7.b, z0.b\n"
+      "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n"
       "add x28, x28, #0x1\n"
-      "addvl x10, x10, #4\n"
       "cmp x28, x20\n"
       "sdot z10.s, z17.b, z0.b\n"
       "sdot z11.s, z16.b, z0.b\n"
+      "addvl x10, x10, #4\n"
       "bne 5b\n"
       "st1w { z8.s }, p3, [x9]\n"
       "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
@@ -171,7 +171,7 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "11:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "12:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -183,11 +183,11 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 13f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x20]\n"
       "ld1w { z13.s }, p2/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x20, #2, MUL VL]\n"
@@ -206,8 +206,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "15:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 16f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -242,8 +242,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z10.s, z17.b, z0.b\n"
       "sdot z14.s, z17.b, z1.b\n"
       "sdot z11.s, z16.b, z0.b\n"
-      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "sdot z15.s, z16.b, z1.b\n"
+      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1b { z6.b }, p4/Z, [x10]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
@@ -257,18 +257,18 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z13.s, z7.b, z1.b\n"
       "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n"
       "add x28, x28, #0x1\n"
-      "addvl x10, x10, #4\n"
       "cmp x28, x20\n"
       "sdot z10.s, z17.b, z0.b\n"
       "sdot z14.s, z17.b, z1.b\n"
+      "addvl x10, x10, #4\n"
       "sdot z11.s, z16.b, z0.b\n"
       "sdot z15.s, z16.b, z1.b\n"
       "bne 15b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "st1w { z8.s }, p3, [x9]\n"
       "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p3, [x20]\n"
@@ -283,7 +283,7 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "21:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "22:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -295,12 +295,12 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 23f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x21]\n"
       "ld1w { z13.s }, p2/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x21, #2, MUL VL]\n"
@@ -327,8 +327,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "25:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 26f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -359,8 +359,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "add x26, x26, #0x4\n"
       "subs x27, x27, #0x4\n"
       "sdot z16.s, z6.b, z2.b\n"
-      "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z7.b, z0.b\n"
+      "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n"
       "add x25, x25, #0x4\n"
       "sdot z13.s, z7.b, z1.b\n"
       "sdot z17.s, z7.b, z2.b\n"
@@ -372,11 +372,11 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z18.s, z21.b, z2.b\n"
       "sdot z11.s, z20.b, z0.b\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
+      "ld1b { z6.b }, p4/Z, [x10]\n"
       "sdot z15.s, z20.b, z1.b\n"
-      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "sdot z19.s, z20.b, z2.b\n"
+      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
-      "ld1b { z6.b }, p4/Z, [x10]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 28b\n"
       "29:"  // Height 3: Multiply loop: Main loop skip
@@ -385,13 +385,13 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z12.s, z6.b, z1.b\n"
       "add x28, x28, #0x1\n"
       "sdot z16.s, z6.b, z2.b\n"
-      "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z7.b, z0.b\n"
+      "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n"
+      "cmp x28, x20\n"
       "sdot z13.s, z7.b, z1.b\n"
       "sdot z17.s, z7.b, z2.b\n"
       "ld1b { z20.b }, p4/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
-      "cmp x28, x20\n"
       "sdot z10.s, z21.b, z0.b\n"
       "sdot z14.s, z21.b, z1.b\n"
       "sdot z18.s, z21.b, z2.b\n"
@@ -400,11 +400,11 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z19.s, z20.b, z2.b\n"
       "bne 25b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "st1w { z8.s }, p3, [x9]\n"
       "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p3, [x21]\n"
@@ -423,7 +423,7 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "31:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "32:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -435,13 +435,13 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 33f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x22]\n"
       "ld1w { z13.s }, p2/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x22, #2, MUL VL]\n"
@@ -476,8 +476,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "35:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 36f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -527,6 +527,7 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z14.s, z25.b, z1.b\n"
       "sdot z18.s, z25.b, z2.b\n"
       "sdot z22.s, z25.b, z3.b\n"
+      "ld1b { z6.b }, p4/Z, [x10]\n"
       "sdot z11.s, z24.b, z0.b\n"
       "sdot z15.s, z24.b, z1.b\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
@@ -535,7 +536,6 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z23.s, z24.b, z3.b\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
       "ld1rw { z3.s }, p4/Z, [x23]\n"
-      "ld1b { z6.b }, p4/Z, [x10]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 38b\n"
       "39:"  // Height 4: Multiply loop: Main loop skip
@@ -546,15 +546,15 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z16.s, z6.b, z2.b\n"
       "sdot z20.s, z6.b, z3.b\n"
       "ld1b { z25.b }, p4/Z, [x10, #2, MUL VL]\n"
+      "cmp x28, x20\n"
       "sdot z9.s, z7.b, z0.b\n"
       "sdot z13.s, z7.b, z1.b\n"
       "sdot z17.s, z7.b, z2.b\n"
       "sdot z21.s, z7.b, z3.b\n"
       "ld1b { z24.b }, p4/Z, [x10, #3, MUL VL]\n"
-      "cmp x28, x20\n"
+      "addvl x10, x10, #4\n"
       "sdot z10.s, z25.b, z0.b\n"
       "sdot z14.s, z25.b, z1.b\n"
-      "addvl x10, x10, #4\n"
       "sdot z18.s, z25.b, z2.b\n"
       "sdot z22.s, z25.b, z3.b\n"
       "sdot z11.s, z24.b, z0.b\n"
@@ -563,12 +563,12 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z23.s, z24.b, z3.b\n"
       "bne 35b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p3, [x9]\n"
-      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "add x22, x9, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "st1w { z8.s }, p3, [x9]\n"
       "add x20, x21, x20, LSL #2\n"
+      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
+      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p3, [x22]\n"
@@ -591,7 +591,7 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "41:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "42:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -603,16 +603,16 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 43f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x23]\n"
       "ld1w { z13.s }, p2/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p1/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p0/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p3/Z, [x22]\n"
@@ -653,8 +653,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "45:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -697,8 +697,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "add x25, x25, #0x4\n"
       "add x24, x24, #0x4\n"
       "sdot z24.s, z6.b, z4.b\n"
-      "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z7.b, z0.b\n"
+      "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n"
       "add x23, x23, #0x4\n"
       "sdot z13.s, z7.b, z1.b\n"
       "sdot z17.s, z7.b, z2.b\n"
@@ -716,12 +716,12 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1b { z6.b }, p4/Z, [x10]\n"
       "sdot z15.s, z28.b, z1.b\n"
-      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "sdot z19.s, z28.b, z2.b\n"
+      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
       "sdot z23.s, z28.b, z3.b\n"
-      "ld1rw { z3.s }, p4/Z, [x23]\n"
       "sdot z27.s, z28.b, z4.b\n"
+      "ld1rw { z3.s }, p4/Z, [x23]\n"
       "ld1rw { z4.s }, p4/Z, [x22]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 48b\n"
@@ -732,12 +732,12 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "sdot z16.s, z6.b, z2.b\n"
       "sdot z20.s, z6.b, z3.b\n"
+      "cmp x28, x20\n"
       "sdot z24.s, z6.b, z4.b\n"
-      "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z7.b, z0.b\n"
+      "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n"
       "sdot z13.s, z7.b, z1.b\n"
       "sdot z17.s, z7.b, z2.b\n"
-      "cmp x28, x20\n"
       "sdot z21.s, z7.b, z3.b\n"
       "sdot z25.s, z7.b, z4.b\n"
       "ld1b { z28.b }, p4/Z, [x10, #3, MUL VL]\n"
@@ -754,15 +754,15 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z27.s, z28.b, z4.b\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p3, [x9]\n"
-      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "st1w { z8.s }, p3, [x9]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
+      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z12.s }, p3, [x23]\n"
       "st1w { z13.s }, p2, [x23, #1, MUL VL]\n"
       "st1w { z14.s }, p1, [x23, #2, MUL VL]\n"
@@ -786,12 +786,11 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "b 62f\n"
       "51:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "52:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -803,17 +802,17 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 53f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x24]\n"
       "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p3/Z, [x23]\n"
@@ -862,8 +861,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "55:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 56f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -951,12 +950,12 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "sdot z16.s, z6.b, z2.b\n"
       "sdot z20.s, z6.b, z3.b\n"
+      "cmp x28, x20\n"
       "sdot z24.s, z6.b, z4.b\n"
       "sdot z28.s, z6.b, z5.b\n"
       "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z7.b, z0.b\n"
       "sdot z13.s, z7.b, z1.b\n"
-      "cmp x28, x20\n"
       "sdot z17.s, z7.b, z2.b\n"
       "sdot z21.s, z7.b, z3.b\n"
       "sdot z25.s, z7.b, z4.b\n"
@@ -977,17 +976,17 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "sdot z31.s, z7.b, z5.b\n"
       "bne 55b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p3, [x9]\n"
-      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "st1w { z8.s }, p3, [x9]\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
+      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
+      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "add x21, x22, x20, LSL #2\n"
       "st1w { z12.s }, p3, [x24]\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z13.s }, p2, [x24, #1, MUL VL]\n"
       "st1w { z14.s }, p1, [x24, #2, MUL VL]\n"
       "st1w { z15.s }, p0, [x24, #3, MUL VL]\n"
@@ -1023,8 +1022,8 @@ void sve_hybrid_s8s32_dot_6x4VL_a64fx (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "62:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/generic.cpp
index 70362ae888..f9b84e26fc 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_dot_6x4VL/generic.cpp
@@ -44,18 +44,18 @@ void sve_hybrid_s8s32_dot_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -89,7 +89,7 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "beq 12f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -114,8 +114,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "mov x28, #0x0\n"
       "5:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 6f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -131,89 +131,89 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 9f\n"
       "8:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "sdot z8.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10]\n"
+      "sdot z8.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "sdot z10.s, z16.b, z0.b[0]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "sdot z10.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "sdot z11.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "sdot z8.s, z16.b, z0.b[1]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      "sdot z8.s, z17.b, z0.b[1]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[1]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #6, MUL VL]\n"
+      "sdot z10.s, z16.b, z0.b[1]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      "sdot z10.s, z17.b, z0.b[1]\n"
       "sdot z11.s, z16.b, z0.b[1]\n"
       "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[2]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n"
       "sdot z10.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "sdot z11.s, z16.b, z0.b[2]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-3, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[3]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[3]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
       "sdot z10.s, z17.b, z0.b[3]\n"
       "sdot z11.s, z16.b, z0.b[3]\n"
+      "add x26, x26, #0x10\n"
       "bgt 8b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
-      "sdot z8.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10]\n"
+      "sdot z8.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[0]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "sdot z10.s, z17.b, z0.b[0]\n"
       "sdot z11.s, z16.b, z0.b[0]\n"
+      "addvl x10, x10, #4\n"
       "ble 10f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z17.b, z0.b[1]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[1]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b[1]\n"
       "sdot z11.s, z16.b, z0.b[1]\n"
+      "addvl x10, x10, #4\n"
       "ble 10f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[2]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b[2]\n"
       "sdot z11.s, z16.b, z0.b[2]\n"
+      "addvl x10, x10, #4\n"
       "ble 10f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[3]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[3]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "sdot z10.s, z17.b, z0.b[3]\n"
       "sdot z11.s, z16.b, z0.b[3]\n"
+      "addvl x10, x10, #4\n"
       "10:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -232,7 +232,7 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "12:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "13:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -244,11 +244,11 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 14f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
@@ -267,8 +267,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "mov x28, #0x0\n"
       "16:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 17f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -287,38 +287,38 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 20f\n"
       "19:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z1.b[0]\n"
       "sdot z12.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z1.b[0]\n"
       "sdot z13.s, z16.b, z0.b[0]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
       "sdot z10.s, z17.b, z1.b[0]\n"
       "sdot z14.s, z17.b, z0.b[0]\n"
       "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "cmp x27, #0x10\n"
       "sdot z11.s, z16.b, z1.b[0]\n"
       "sdot z15.s, z16.b, z0.b[0]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       "sdot z8.s, z17.b, z1.b[1]\n"
       "sdot z12.s, z17.b, z0.b[1]\n"
       "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       "sdot z9.s, z16.b, z1.b[1]\n"
       "sdot z13.s, z16.b, z0.b[1]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
       "sdot z10.s, z17.b, z1.b[1]\n"
       "sdot z14.s, z17.b, z0.b[1]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "sdot z11.s, z16.b, z1.b[1]\n"
       "sdot z15.s, z16.b, z0.b[1]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
       "sdot z8.s, z17.b, z1.b[2]\n"
       "sdot z12.s, z17.b, z0.b[2]\n"
@@ -345,50 +345,50 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "bgt 19b\n"
       "20:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[0]\n"
       "sdot z12.s, z17.b, z1.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[0]\n"
       "sdot z13.s, z16.b, z1.b[0]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "sdot z10.s, z17.b, z0.b[0]\n"
       "sdot z14.s, z17.b, z1.b[0]\n"
+      "addvl x10, x10, #4\n"
       "sdot z11.s, z16.b, z0.b[0]\n"
       "sdot z15.s, z16.b, z1.b[0]\n"
       "ble 21f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z17.b, z0.b[1]\n"
       "sdot z12.s, z17.b, z1.b[1]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[1]\n"
       "sdot z13.s, z16.b, z1.b[1]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b[1]\n"
       "sdot z14.s, z17.b, z1.b[1]\n"
+      "addvl x10, x10, #4\n"
       "sdot z11.s, z16.b, z0.b[1]\n"
       "sdot z15.s, z16.b, z1.b[1]\n"
       "ble 21f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z17.b, z0.b[2]\n"
       "sdot z12.s, z17.b, z1.b[2]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[2]\n"
       "sdot z13.s, z16.b, z1.b[2]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x4\n"
       "sdot z10.s, z17.b, z0.b[2]\n"
       "sdot z14.s, z17.b, z1.b[2]\n"
+      "addvl x10, x10, #4\n"
       "sdot z11.s, z16.b, z0.b[2]\n"
       "sdot z15.s, z16.b, z1.b[2]\n"
       "ble 21f\n"
@@ -396,13 +396,13 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z8.s, z17.b, z0.b[3]\n"
       "sdot z12.s, z17.b, z1.b[3]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z16.b, z0.b[3]\n"
       "sdot z13.s, z16.b, z1.b[3]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "sdot z10.s, z17.b, z0.b[3]\n"
       "sdot z14.s, z17.b, z1.b[3]\n"
+      "addvl x10, x10, #4\n"
       "sdot z11.s, z16.b, z0.b[3]\n"
       "sdot z15.s, z16.b, z1.b[3]\n"
       "21:"  // Height 2: Multiply loop: multiply skip
@@ -411,10 +411,10 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 16b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "st1w { z8.s }, p4, [x9]\n"
       "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p4, [x20]\n"
@@ -429,7 +429,7 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "23:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "24:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -441,12 +441,12 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 25f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
@@ -473,8 +473,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "mov x28, #0x0\n"
       "27:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 28f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -496,37 +496,37 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 31f\n"
       "30:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z21.b }, p5/Z, [x10]\n"
-      "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1b { z21.b }, p5/Z, [x10]\n"
       "sdot z8.s, z21.b, z2.b[0]\n"
       "sdot z12.s, z21.b, z1.b[0]\n"
-      "sdot z9.s, z20.b, z2.b[0]\n"
-      "sdot z13.s, z20.b, z1.b[0]\n"
+      "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z16.s, z21.b, z0.b[0]\n"
+      "sdot z9.s, z20.b, z2.b[0]\n"
       "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "sdot z13.s, z20.b, z1.b[0]\n"
       "sdot z17.s, z20.b, z0.b[0]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
+      "cmp x27, #0x10\n"
       "sdot z10.s, z21.b, z2.b[0]\n"
       "sdot z14.s, z21.b, z1.b[0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "sdot z18.s, z21.b, z0.b[0]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #4, MUL VL]\n"
       "sdot z11.s, z20.b, z2.b[0]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       "sdot z15.s, z20.b, z1.b[0]\n"
       "sdot z19.s, z20.b, z0.b[0]\n"
       "ld1b { z20.b }, p5/Z, [x10, #5, MUL VL]\n"
       "sdot z8.s, z21.b, z2.b[1]\n"
       "sdot z12.s, z21.b, z1.b[1]\n"
       "sdot z16.s, z21.b, z0.b[1]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #6, MUL VL]\n"
       "sdot z9.s, z20.b, z2.b[1]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #6, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[1]\n"
       "sdot z17.s, z20.b, z0.b[1]\n"
       "ld1b { z20.b }, p5/Z, [x10, #7, MUL VL]\n"
@@ -535,31 +535,31 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z14.s, z21.b, z1.b[1]\n"
       "sdot z18.s, z21.b, z0.b[1]\n"
       "sdot z11.s, z20.b, z2.b[1]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "sdot z15.s, z20.b, z1.b[1]\n"
       "sdot z19.s, z20.b, z0.b[1]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z20.b }, p5/Z, [x10, #-7, MUL VL]\n"
       "sdot z8.s, z21.b, z2.b[2]\n"
       "sdot z12.s, z21.b, z1.b[2]\n"
       "sdot z16.s, z21.b, z0.b[2]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "sdot z9.s, z20.b, z2.b[2]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[2]\n"
       "sdot z17.s, z20.b, z0.b[2]\n"
       "ld1b { z20.b }, p5/Z, [x10, #-5, MUL VL]\n"
       "sdot z10.s, z21.b, z2.b[2]\n"
       "sdot z14.s, z21.b, z1.b[2]\n"
       "sdot z18.s, z21.b, z0.b[2]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "sdot z11.s, z20.b, z2.b[2]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "sdot z15.s, z20.b, z1.b[2]\n"
       "sdot z19.s, z20.b, z0.b[2]\n"
       "ld1b { z20.b }, p5/Z, [x10, #-3, MUL VL]\n"
       "sdot z8.s, z21.b, z2.b[3]\n"
       "sdot z12.s, z21.b, z1.b[3]\n"
       "sdot z16.s, z21.b, z0.b[3]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "sdot z9.s, z20.b, z2.b[3]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[3]\n"
       "sdot z17.s, z20.b, z0.b[3]\n"
       "ld1b { z20.b }, p5/Z, [x10, #-1, MUL VL]\n"
@@ -572,18 +572,18 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "bgt 30b\n"
       "31:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z21.b }, p5/Z, [x10]\n"
-      "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
+      "ld1b { z21.b }, p5/Z, [x10]\n"
       "sdot z8.s, z21.b, z0.b[0]\n"
       "sdot z12.s, z21.b, z1.b[0]\n"
-      "sdot z9.s, z20.b, z0.b[0]\n"
-      "sdot z13.s, z20.b, z1.b[0]\n"
+      "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z16.s, z21.b, z2.b[0]\n"
+      "sdot z9.s, z20.b, z0.b[0]\n"
       "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "sdot z13.s, z20.b, z1.b[0]\n"
       "sdot z17.s, z20.b, z2.b[0]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
@@ -596,12 +596,12 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 32f\n"
       "ld1b { z21.b }, p5/Z, [x10]\n"
       "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z21.b, z0.b[1]\n"
       "sdot z12.s, z21.b, z1.b[1]\n"
       "sdot z16.s, z21.b, z2.b[1]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z20.b, z0.b[1]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z13.s, z20.b, z1.b[1]\n"
       "sdot z17.s, z20.b, z2.b[1]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -615,12 +615,12 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 32f\n"
       "ld1b { z21.b }, p5/Z, [x10]\n"
       "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z21.b, z0.b[2]\n"
       "sdot z12.s, z21.b, z1.b[2]\n"
       "sdot z16.s, z21.b, z2.b[2]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z20.b, z0.b[2]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z13.s, z20.b, z1.b[2]\n"
       "sdot z17.s, z20.b, z2.b[2]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -637,8 +637,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z8.s, z21.b, z0.b[3]\n"
       "sdot z12.s, z21.b, z1.b[3]\n"
       "sdot z16.s, z21.b, z2.b[3]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z20.b, z0.b[3]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z13.s, z20.b, z1.b[3]\n"
       "sdot z17.s, z20.b, z2.b[3]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -655,11 +655,11 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 27b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "st1w { z8.s }, p4, [x9]\n"
       "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p4, [x21]\n"
@@ -678,7 +678,7 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "34:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "35:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -690,13 +690,13 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 36f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
@@ -731,8 +731,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "mov x28, #0x0\n"
       "38:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 39f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -757,25 +757,25 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 42f\n"
       "41:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z8.s, z25.b, z3.b[0]\n"
       "sdot z12.s, z25.b, z2.b[0]\n"
-      "sdot z9.s, z24.b, z3.b[0]\n"
-      "sdot z13.s, z24.b, z2.b[0]\n"
       "sdot z16.s, z25.b, z1.b[0]\n"
       "sdot z20.s, z25.b, z0.b[0]\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      "sdot z9.s, z24.b, z3.b[0]\n"
+      "sdot z13.s, z24.b, z2.b[0]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "sdot z17.s, z24.b, z1.b[0]\n"
       "sdot z21.s, z24.b, z0.b[0]\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -804,9 +804,9 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z14.s, z25.b, z2.b[1]\n"
       "sdot z18.s, z25.b, z1.b[1]\n"
       "sdot z22.s, z25.b, z0.b[1]\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "sdot z11.s, z24.b, z3.b[1]\n"
       "sdot z15.s, z24.b, z2.b[1]\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "sdot z19.s, z24.b, z1.b[1]\n"
       "sdot z23.s, z24.b, z0.b[1]\n"
       "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
@@ -851,20 +851,20 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "bgt 41b\n"
       "42:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z8.s, z25.b, z0.b[0]\n"
       "sdot z12.s, z25.b, z1.b[0]\n"
-      "sdot z9.s, z24.b, z0.b[0]\n"
-      "sdot z13.s, z24.b, z1.b[0]\n"
       "sdot z16.s, z25.b, z2.b[0]\n"
       "sdot z20.s, z25.b, z3.b[0]\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "sdot z9.s, z24.b, z0.b[0]\n"
+      "sdot z13.s, z24.b, z1.b[0]\n"
       "sdot z17.s, z24.b, z2.b[0]\n"
       "sdot z21.s, z24.b, z3.b[0]\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -880,12 +880,12 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 43f\n"
       "ld1b { z25.b }, p5/Z, [x10]\n"
       "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z25.b, z0.b[1]\n"
       "sdot z12.s, z25.b, z1.b[1]\n"
       "sdot z16.s, z25.b, z2.b[1]\n"
       "sdot z20.s, z25.b, z3.b[1]\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z9.s, z24.b, z0.b[1]\n"
       "sdot z13.s, z24.b, z1.b[1]\n"
       "sdot z17.s, z24.b, z2.b[1]\n"
@@ -903,12 +903,12 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 43f\n"
       "ld1b { z25.b }, p5/Z, [x10]\n"
       "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z25.b, z0.b[2]\n"
       "sdot z12.s, z25.b, z1.b[2]\n"
       "sdot z16.s, z25.b, z2.b[2]\n"
       "sdot z20.s, z25.b, z3.b[2]\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z9.s, z24.b, z0.b[2]\n"
       "sdot z13.s, z24.b, z1.b[2]\n"
       "sdot z17.s, z24.b, z2.b[2]\n"
@@ -951,12 +951,12 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 38b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p4, [x9]\n"
-      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "add x22, x9, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "st1w { z8.s }, p4, [x9]\n"
       "add x20, x21, x20, LSL #2\n"
+      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p4, [x22]\n"
@@ -979,7 +979,7 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "45:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "46:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -991,16 +991,16 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 47f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x22]\n"
@@ -1041,8 +1041,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "mov x28, #0x0\n"
       "49:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 50f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1070,29 +1070,29 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 53f\n"
       "52:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z4.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1b { z29.b }, p5/Z, [x10]\n"
       "sdot z8.s, z29.b, z4.b[0]\n"
       "sdot z12.s, z29.b, z3.b[0]\n"
-      "sdot z9.s, z28.b, z4.b[0]\n"
+      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z16.s, z29.b, z2.b[0]\n"
       "sdot z20.s, z29.b, z1.b[0]\n"
+      "add x25, x25, #0x10\n"
       "sdot z24.s, z29.b, z0.b[0]\n"
-      "sdot z13.s, z28.b, z3.b[0]\n"
+      "sdot z9.s, z28.b, z4.b[0]\n"
       "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      "sdot z13.s, z28.b, z3.b[0]\n"
       "sdot z17.s, z28.b, z2.b[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "sdot z21.s, z28.b, z1.b[0]\n"
       "sdot z25.s, z28.b, z0.b[0]\n"
       "ld1b { z28.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1101,8 +1101,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z18.s, z29.b, z2.b[0]\n"
       "sdot z22.s, z29.b, z1.b[0]\n"
       "sdot z26.s, z29.b, z0.b[0]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #4, MUL VL]\n"
       "sdot z11.s, z28.b, z4.b[0]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #4, MUL VL]\n"
       "sdot z15.s, z28.b, z3.b[0]\n"
       "sdot z19.s, z28.b, z2.b[0]\n"
       "sdot z23.s, z28.b, z1.b[0]\n"
@@ -1113,8 +1113,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z16.s, z29.b, z2.b[1]\n"
       "sdot z20.s, z29.b, z1.b[1]\n"
       "sdot z24.s, z29.b, z0.b[1]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #6, MUL VL]\n"
       "sdot z9.s, z28.b, z4.b[1]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #6, MUL VL]\n"
       "sdot z13.s, z28.b, z3.b[1]\n"
       "sdot z17.s, z28.b, z2.b[1]\n"
       "sdot z21.s, z28.b, z1.b[1]\n"
@@ -1127,8 +1127,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z22.s, z29.b, z1.b[1]\n"
       "sdot z26.s, z29.b, z0.b[1]\n"
       "sdot z11.s, z28.b, z4.b[1]\n"
-      "sdot z15.s, z28.b, z3.b[1]\n"
       "ld1b { z29.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      "sdot z15.s, z28.b, z3.b[1]\n"
       "sdot z19.s, z28.b, z2.b[1]\n"
       "sdot z23.s, z28.b, z1.b[1]\n"
       "sdot z27.s, z28.b, z0.b[1]\n"
@@ -1138,8 +1138,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z16.s, z29.b, z2.b[2]\n"
       "sdot z20.s, z29.b, z1.b[2]\n"
       "sdot z24.s, z29.b, z0.b[2]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "sdot z9.s, z28.b, z4.b[2]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "sdot z13.s, z28.b, z3.b[2]\n"
       "sdot z17.s, z28.b, z2.b[2]\n"
       "sdot z21.s, z28.b, z1.b[2]\n"
@@ -1150,8 +1150,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z18.s, z29.b, z2.b[2]\n"
       "sdot z22.s, z29.b, z1.b[2]\n"
       "sdot z26.s, z29.b, z0.b[2]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "sdot z11.s, z28.b, z4.b[2]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "sdot z15.s, z28.b, z3.b[2]\n"
       "sdot z19.s, z28.b, z2.b[2]\n"
       "sdot z23.s, z28.b, z1.b[2]\n"
@@ -1162,8 +1162,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z16.s, z29.b, z2.b[3]\n"
       "sdot z20.s, z29.b, z1.b[3]\n"
       "sdot z24.s, z29.b, z0.b[3]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "sdot z9.s, z28.b, z4.b[3]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "sdot z13.s, z28.b, z3.b[3]\n"
       "sdot z17.s, z28.b, z2.b[3]\n"
       "sdot z21.s, z28.b, z1.b[3]\n"
@@ -1182,23 +1182,23 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "bgt 52b\n"
       "53:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
       "ld1rqb { z4.b }, p0/Z, [x22]\n"
+      "ld1b { z29.b }, p5/Z, [x10]\n"
       "sdot z8.s, z29.b, z0.b[0]\n"
       "sdot z12.s, z29.b, z1.b[0]\n"
-      "sdot z9.s, z28.b, z0.b[0]\n"
-      "sdot z13.s, z28.b, z1.b[0]\n"
+      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z16.s, z29.b, z2.b[0]\n"
       "sdot z20.s, z29.b, z3.b[0]\n"
       "sdot z24.s, z29.b, z4.b[0]\n"
-      "sdot z17.s, z28.b, z2.b[0]\n"
+      "sdot z9.s, z28.b, z0.b[0]\n"
       "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "sdot z13.s, z28.b, z1.b[0]\n"
+      "sdot z17.s, z28.b, z2.b[0]\n"
       "sdot z21.s, z28.b, z3.b[0]\n"
       "sdot z25.s, z28.b, z4.b[0]\n"
       "ld1b { z28.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1216,21 +1216,21 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 54f\n"
       "ld1b { z29.b }, p5/Z, [x10]\n"
       "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z29.b, z0.b[1]\n"
       "sdot z12.s, z29.b, z1.b[1]\n"
       "sdot z16.s, z29.b, z2.b[1]\n"
       "sdot z20.s, z29.b, z3.b[1]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z24.s, z29.b, z4.b[1]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z28.b, z0.b[1]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z13.s, z28.b, z1.b[1]\n"
       "sdot z17.s, z28.b, z2.b[1]\n"
       "sdot z21.s, z28.b, z3.b[1]\n"
       "sdot z25.s, z28.b, z4.b[1]\n"
       "ld1b { z28.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "sdot z10.s, z29.b, z0.b[1]\n"
       "addvl x10, x10, #4\n"
+      "sdot z10.s, z29.b, z0.b[1]\n"
       "sdot z14.s, z29.b, z1.b[1]\n"
       "sdot z18.s, z29.b, z2.b[1]\n"
       "sdot z22.s, z29.b, z3.b[1]\n"
@@ -1243,21 +1243,21 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 54f\n"
       "ld1b { z29.b }, p5/Z, [x10]\n"
       "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z29.b, z0.b[2]\n"
       "sdot z12.s, z29.b, z1.b[2]\n"
       "sdot z16.s, z29.b, z2.b[2]\n"
       "sdot z20.s, z29.b, z3.b[2]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z24.s, z29.b, z4.b[2]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z28.b, z0.b[2]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z13.s, z28.b, z1.b[2]\n"
       "sdot z17.s, z28.b, z2.b[2]\n"
       "sdot z21.s, z28.b, z3.b[2]\n"
       "sdot z25.s, z28.b, z4.b[2]\n"
       "ld1b { z28.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "sdot z10.s, z29.b, z0.b[2]\n"
       "addvl x10, x10, #4\n"
+      "sdot z10.s, z29.b, z0.b[2]\n"
       "sdot z14.s, z29.b, z1.b[2]\n"
       "sdot z18.s, z29.b, z2.b[2]\n"
       "sdot z22.s, z29.b, z3.b[2]\n"
@@ -1275,8 +1275,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "sdot z16.s, z29.b, z2.b[3]\n"
       "sdot z20.s, z29.b, z3.b[3]\n"
       "sdot z24.s, z29.b, z4.b[3]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z9.s, z28.b, z0.b[3]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "sdot z13.s, z28.b, z1.b[3]\n"
       "sdot z17.s, z28.b, z2.b[3]\n"
       "sdot z21.s, z28.b, z3.b[3]\n"
@@ -1299,15 +1299,15 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 49b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p4, [x9]\n"
-      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "st1w { z8.s }, p4, [x9]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z12.s }, p4, [x23]\n"
       "st1w { z13.s }, p3, [x23, #1, MUL VL]\n"
       "st1w { z14.s }, p2, [x23, #2, MUL VL]\n"
@@ -1331,12 +1331,11 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "b 68f\n"
       "56:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "57:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1348,17 +1347,17 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 58f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x24]\n"
       "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x23]\n"
@@ -1407,8 +1406,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "mov x28, #0x0\n"
       "60:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 61f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1439,29 +1438,29 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 64f\n"
       "63:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p5/Z, [x10]\n"
-      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z7.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z6.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z5.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "ld1rqb { z2.b }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1b { z1.b }, p5/Z, [x10]\n"
+      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z8.s, z1.b, z7.b[0]\n"
       "sdot z12.s, z1.b, z6.b[0]\n"
-      "add x21, x21, #0x10\n"
       "sdot z16.s, z1.b, z5.b[0]\n"
       "sdot z20.s, z1.b, z4.b[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "sdot z24.s, z1.b, z3.b[0]\n"
       "sdot z28.s, z1.b, z2.b[0]\n"
       "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x21, x21, #0x10\n"
       "sdot z9.s, z0.b, z7.b[0]\n"
       "sdot z13.s, z0.b, z6.b[0]\n"
       "sdot z17.s, z0.b, z5.b[0]\n"
@@ -1569,24 +1568,24 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "bgt 63b\n"
       "64:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z7.b }, p5/Z, [x10]\n"
-      "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
       "ld1rqb { z4.b }, p0/Z, [x22]\n"
       "ld1rqb { z5.b }, p0/Z, [x21]\n"
+      "ld1b { z7.b }, p5/Z, [x10]\n"
+      "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n"
       "sdot z8.s, z7.b, z0.b[0]\n"
       "sdot z12.s, z7.b, z1.b[0]\n"
-      "sdot z9.s, z6.b, z0.b[0]\n"
-      "sdot z13.s, z6.b, z1.b[0]\n"
       "sdot z16.s, z7.b, z2.b[0]\n"
       "sdot z20.s, z7.b, z3.b[0]\n"
       "sdot z24.s, z7.b, z4.b[0]\n"
       "sdot z28.s, z7.b, z5.b[0]\n"
       "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "sdot z9.s, z6.b, z0.b[0]\n"
+      "sdot z13.s, z6.b, z1.b[0]\n"
       "sdot z17.s, z6.b, z2.b[0]\n"
       "sdot z21.s, z6.b, z3.b[0]\n"
       "sdot z25.s, z6.b, z4.b[0]\n"
@@ -1608,23 +1607,23 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 65f\n"
       "ld1b { z7.b }, p5/Z, [x10]\n"
       "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z7.b, z0.b[1]\n"
       "sdot z12.s, z7.b, z1.b[1]\n"
       "sdot z16.s, z7.b, z2.b[1]\n"
       "sdot z20.s, z7.b, z3.b[1]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z24.s, z7.b, z4.b[1]\n"
       "sdot z28.s, z7.b, z5.b[1]\n"
-      "sdot z9.s, z6.b, z0.b[1]\n"
       "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "sdot z9.s, z6.b, z0.b[1]\n"
       "sdot z13.s, z6.b, z1.b[1]\n"
       "sdot z17.s, z6.b, z2.b[1]\n"
       "sdot z21.s, z6.b, z3.b[1]\n"
       "sdot z25.s, z6.b, z4.b[1]\n"
       "sdot z29.s, z6.b, z5.b[1]\n"
       "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "sdot z10.s, z7.b, z0.b[1]\n"
       "addvl x10, x10, #4\n"
+      "sdot z10.s, z7.b, z0.b[1]\n"
       "sdot z14.s, z7.b, z1.b[1]\n"
       "sdot z18.s, z7.b, z2.b[1]\n"
       "sdot z22.s, z7.b, z3.b[1]\n"
@@ -1639,23 +1638,23 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "ble 65f\n"
       "ld1b { z7.b }, p5/Z, [x10]\n"
       "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "sdot z8.s, z7.b, z0.b[2]\n"
       "sdot z12.s, z7.b, z1.b[2]\n"
       "sdot z16.s, z7.b, z2.b[2]\n"
       "sdot z20.s, z7.b, z3.b[2]\n"
+      "subs x27, x27, #0x4\n"
       "sdot z24.s, z7.b, z4.b[2]\n"
       "sdot z28.s, z7.b, z5.b[2]\n"
-      "sdot z9.s, z6.b, z0.b[2]\n"
       "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "sdot z9.s, z6.b, z0.b[2]\n"
       "sdot z13.s, z6.b, z1.b[2]\n"
       "sdot z17.s, z6.b, z2.b[2]\n"
       "sdot z21.s, z6.b, z3.b[2]\n"
       "sdot z25.s, z6.b, z4.b[2]\n"
       "sdot z29.s, z6.b, z5.b[2]\n"
       "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "sdot z10.s, z7.b, z0.b[2]\n"
       "addvl x10, x10, #4\n"
+      "sdot z10.s, z7.b, z0.b[2]\n"
       "sdot z14.s, z7.b, z1.b[2]\n"
       "sdot z18.s, z7.b, z2.b[2]\n"
       "sdot z22.s, z7.b, z3.b[2]\n"
@@ -1703,17 +1702,17 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 60b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p4, [x9]\n"
-      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "st1w { z8.s }, p4, [x9]\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
+      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
+      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "add x21, x22, x20, LSL #2\n"
       "st1w { z12.s }, p4, [x24]\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z13.s }, p3, [x24, #1, MUL VL]\n"
       "st1w { z14.s }, p2, [x24, #2, MUL VL]\n"
       "st1w { z15.s }, p1, [x24, #3, MUL VL]\n"
@@ -1749,8 +1748,8 @@ void sve_hybrid_s8s32_dot_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "68:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL.hpp
index fceaeb119c..8135172b54 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL.hpp
@@ -70,7 +70,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 8, 8> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 8, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL/generic.cpp
index d257cc69de..abe4b92faf 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_s8s32_mmla_6x4VL/generic.cpp
@@ -44,18 +44,18 @@ void sve_hybrid_s8s32_mmla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -89,7 +89,7 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "beq 12f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -100,14 +100,14 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "incw x20\n"
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 3f\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z19.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "zip1 z9.d, z19.d, z13.d\n"
-      "zip2 z13.d, z19.d, z13.d\n"
+      "zip1 z9.d, z18.d, z13.d\n"
+      "zip2 z13.d, z18.d, z13.d\n"
       "zip1 z10.d, z17.d, z14.d\n"
       "zip2 z14.d, z17.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
@@ -126,8 +126,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "5:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 6f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -143,87 +143,87 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "ble 9f\n"
       "8:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z16.b }, p5/Z, [x10]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z19.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "trn1 z18.d, z19.d, z20.d\n"
-      "trn2 z19.d, z19.d, z20.d\n"
-      ".inst 0x45109a48  // smmla z8.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45119a4c  // smmla z12.s, z18.b, z17.b\n"
-      "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45109a49  // smmla z9.s, z18.b, z16.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45149a4d  // smmla z13.s, z18.b, z20.b\n"
+      "ld1rqb { z20.b }, p0/Z, [x26]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x45119a48  // smmla z8.s, z18.b, z17.b\n"
+      ".inst 0x45109a4c  // smmla z12.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
+      ".inst 0x45119a49  // smmla z9.s, z18.b, z17.b\n"
+      ".inst 0x45109a4d  // smmla z13.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45019a4a  // smmla z10.s, z18.b, z1.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
+      ".inst 0x45119a4a  // smmla z10.s, z18.b, z17.b\n"
       ".inst 0x45109a4e  // smmla z14.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
+      "trn2 z20.d, z20.d, z19.d\n"
       ".inst 0x45119a4b  // smmla z11.s, z18.b, z17.b\n"
       ".inst 0x45109a4f  // smmla z15.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45109a68  // smmla z8.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45119a6c  // smmla z12.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45109a69  // smmla z9.s, z19.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
+      ".inst 0x45119a88  // smmla z8.s, z20.b, z17.b\n"
+      ".inst 0x45109a8c  // smmla z12.s, z20.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n"
+      ".inst 0x45119a89  // smmla z9.s, z20.b, z17.b\n"
+      ".inst 0x45109a8d  // smmla z13.s, z20.b, z16.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45119a6d  // smmla z13.s, z19.b, z17.b\n"
-      "ld1b { z3.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45109a6a  // smmla z10.s, z19.b, z16.b\n"
+      "ld1b { z7.b }, p5/Z, [x10, #-3, MUL VL]\n"
+      ".inst 0x45109a8a  // smmla z10.s, z20.b, z16.b\n"
+      ".inst 0x45079a8e  // smmla z14.s, z20.b, z7.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45039a6e  // smmla z14.s, z19.b, z3.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45119a6b  // smmla z11.s, z19.b, z17.b\n"
-      ".inst 0x45109a6f  // smmla z15.s, z19.b, z16.b\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45119a8b  // smmla z11.s, z20.b, z17.b\n"
+      ".inst 0x45109a8f  // smmla z15.s, z20.b, z16.b\n"
+      "add x26, x26, #0x10\n"
       "bgt 8b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45119a48  // smmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45109a4c  // smmla z12.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45119a49  // smmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45109a4d  // smmla z13.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45119a4a  // smmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45109a4e  // smmla z14.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "subs x27, x27, #0x8\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45119a4b  // smmla z11.s, z18.b, z17.b\n"
       ".inst 0x45109a4f  // smmla z15.s, z18.b, z16.b\n"
+      "addvl x10, x10, #8\n"
       "ble 10f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45119828  // smmla z8.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4510982c  // smmla z12.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45119829  // smmla z9.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x4510982d  // smmla z13.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x4511982a  // smmla z10.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x4510982e  // smmla z14.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x4511982b  // smmla z11.s, z1.b, z17.b\n"
       ".inst 0x4510982f  // smmla z15.s, z1.b, z16.b\n"
+      "addvl x10, x10, #8\n"
       "10:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -231,9 +231,9 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "bne 5b\n"
       "uzp1 z8.d, z8.d, z12.d\n"
       "uzp1 z9.d, z9.d, z13.d\n"
+      "st1w { z8.s }, p4, [x9]\n"
       "uzp1 z10.d, z10.d, z14.d\n"
       "uzp1 z11.d, z11.d, z15.d\n"
-      "st1w { z8.s }, p4, [x9]\n"
       "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
@@ -246,7 +246,7 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "12:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "13:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -258,19 +258,19 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 14f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z18.s }, p4/Z, [x9]\n"
-      "ld1w { z24.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z2.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
+      "zip1 z8.d, z18.d, z12.d\n"
+      "zip2 z12.d, z18.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
+      "zip1 z9.d, z2.d, z13.d\n"
+      "zip2 z13.d, z2.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z8.d, z18.d, z12.d\n"
-      "zip2 z12.d, z18.d, z12.d\n"
-      "zip1 z9.d, z24.d, z13.d\n"
-      "zip2 z13.d, z24.d, z13.d\n"
       "zip1 z10.d, z17.d, z14.d\n"
       "zip2 z14.d, z17.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
@@ -289,8 +289,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "16:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 17f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -309,109 +309,109 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "ble 20f\n"
       "19:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
+      "ld1rqb { z20.b }, p0/Z, [x26]\n"
+      "ld1rqb { z19.b }, p0/Z, [x25]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z19.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "trn1 z18.d, z19.d, z25.d\n"
-      "trn2 z19.d, z19.d, z25.d\n"
       ".inst 0x45119a48  // smmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45109a4c  // smmla z12.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45119a49  // smmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45109a4d  // smmla z13.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45119a4a  // smmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45109a4e  // smmla z14.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
+      "trn2 z20.d, z20.d, z19.d\n"
       ".inst 0x45119a4b  // smmla z11.s, z18.b, z17.b\n"
       ".inst 0x45109a4f  // smmla z15.s, z18.b, z16.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45119a68  // smmla z8.s, z19.b, z17.b\n"
+      ".inst 0x45119a88  // smmla z8.s, z20.b, z17.b\n"
+      ".inst 0x45109a8c  // smmla z12.s, z20.b, z16.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45109a6c  // smmla z12.s, z19.b, z16.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45119a69  // smmla z9.s, z19.b, z17.b\n"
+      ".inst 0x45119a89  // smmla z9.s, z20.b, z17.b\n"
+      ".inst 0x45109a8d  // smmla z13.s, z20.b, z16.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45109a6d  // smmla z13.s, z19.b, z16.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45119a6a  // smmla z10.s, z19.b, z17.b\n"
+      ".inst 0x45119a8a  // smmla z10.s, z20.b, z17.b\n"
+      ".inst 0x45109a8e  // smmla z14.s, z20.b, z16.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45109a6e  // smmla z14.s, z19.b, z16.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45119a6b  // smmla z11.s, z19.b, z17.b\n"
-      ".inst 0x45109a6f  // smmla z15.s, z19.b, z16.b\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45119a8b  // smmla z11.s, z20.b, z17.b\n"
+      ".inst 0x45109a8f  // smmla z15.s, z20.b, z16.b\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "bgt 19b\n"
       "20:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "ld1rqb { z19.b }, p0/Z, [x25]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45119a48  // smmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45109a4c  // smmla z12.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45119a49  // smmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45109a4d  // smmla z13.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45119a4a  // smmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45109a4e  // smmla z14.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "subs x27, x27, #0x8\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45119a4b  // smmla z11.s, z18.b, z17.b\n"
       ".inst 0x45109a4f  // smmla z15.s, z18.b, z16.b\n"
+      "addvl x10, x10, #8\n"
       "ble 21f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45119828  // smmla z8.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4510982c  // smmla z12.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45119829  // smmla z9.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x4510982d  // smmla z13.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x4511982a  // smmla z10.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x4510982e  // smmla z14.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x4511982b  // smmla z11.s, z1.b, z17.b\n"
       ".inst 0x4510982f  // smmla z15.s, z1.b, z16.b\n"
+      "addvl x10, x10, #8\n"
       "21:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 16b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z17.d, z8.d, z12.d\n"
+      "add x20, x9, x20, LSL #2\n"
+      "uzp1 z16.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z16.d, z9.d, z13.d\n"
+      "uzp1 z17.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z12.d, z10.d, z14.d\n"
+      "st1w { z16.s }, p4, [x9]\n"
+      "uzp1 z16.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x20, x9, x20, LSL #2\n"
-      "uzp1 z26.d, z11.d, z15.d\n"
+      "st1w { z17.s }, p3, [x9, #1, MUL VL]\n"
+      "uzp1 z2.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z17.s }, p4, [x9]\n"
-      "st1w { z16.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z12.s }, p2, [x9, #2, MUL VL]\n"
-      "st1w { z26.s }, p1, [x9, #3, MUL VL]\n"
+      "st1w { z16.s }, p2, [x9, #2, MUL VL]\n"
+      "st1w { z2.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z8.s }, p4, [x20]\n"
       "st1w { z9.s }, p3, [x20, #1, MUL VL]\n"
@@ -425,7 +425,7 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "23:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "24:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -437,28 +437,28 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 25f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z24.s }, p4/Z, [x9]\n"
-      "ld1w { z26.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x21, x9, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x20]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "zip1 z8.d, z24.d, z12.d\n"
-      "zip2 z12.d, z24.d, z12.d\n"
-      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z9.d, z26.d, z13.d\n"
-      "zip2 z13.d, z26.d, z13.d\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
       "zip1 z17.d, z18.d, z21.d\n"
@@ -489,8 +489,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "27:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 28f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -512,92 +512,92 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "ble 31f\n"
       "30:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z27.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
+      "ld1rqb { z30.b }, p0/Z, [x26]\n"
       "ld1rqb { z24.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z26.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "trn1 z6.d, z27.d, z24.d\n"
-      "trn2 z27.d, z27.d, z24.d\n"
-      "trn1 z30.d, z26.d, z29.d\n"
-      "trn2 z26.d, z26.d, z29.d\n"
-      ".inst 0x451998c8  // smmla z8.s, z6.b, z25.b\n"
-      ".inst 0x451c98cc  // smmla z12.s, z6.b, z28.b\n"
-      ".inst 0x45199bd0  // smmla z16.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x451c9bd4  // smmla z20.s, z30.b, z28.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x451998c9  // smmla z9.s, z6.b, z25.b\n"
-      ".inst 0x45199bd1  // smmla z17.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x451898cd  // smmla z13.s, z6.b, z24.b\n"
-      ".inst 0x45189bd5  // smmla z21.s, z30.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x451998ca  // smmla z10.s, z6.b, z25.b\n"
-      ".inst 0x45199bd2  // smmla z18.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x451898ce  // smmla z14.s, z6.b, z24.b\n"
-      ".inst 0x45189bd6  // smmla z22.s, z30.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x451998cb  // smmla z11.s, z6.b, z25.b\n"
-      ".inst 0x45199bd3  // smmla z19.s, z30.b, z25.b\n"
-      ".inst 0x451898cf  // smmla z15.s, z6.b, z24.b\n"
-      ".inst 0x45189bd7  // smmla z23.s, z30.b, z24.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
+      "ld1rqb { z28.b }, p0/Z, [x24]\n"
+      "trn1 z27.d, z30.d, z24.d\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "trn1 z26.d, z28.d, z29.d\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45199b68  // smmla z8.s, z27.b, z25.b\n"
       ".inst 0x45199b50  // smmla z16.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x45189b6c  // smmla z12.s, z27.b, z24.b\n"
       ".inst 0x45189b54  // smmla z20.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45199b69  // smmla z9.s, z27.b, z25.b\n"
       ".inst 0x45199b51  // smmla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n"
+      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "trn2 z28.d, z28.d, z29.d\n"
       ".inst 0x45189b6d  // smmla z13.s, z27.b, z24.b\n"
       ".inst 0x45189b55  // smmla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x45199b6a  // smmla z10.s, z27.b, z25.b\n"
       ".inst 0x45199b52  // smmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n"
+      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
+      "cmp x27, #0x10\n"
       ".inst 0x45189b6e  // smmla z14.s, z27.b, z24.b\n"
       ".inst 0x45189b56  // smmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
+      "addvl x10, x10, #16\n"
       ".inst 0x45199b6b  // smmla z11.s, z27.b, z25.b\n"
       ".inst 0x45199b53  // smmla z19.s, z26.b, z25.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x45189b6f  // smmla z15.s, z27.b, z24.b\n"
       ".inst 0x45189b57  // smmla z23.s, z26.b, z24.b\n"
+      "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x45199bc8  // smmla z8.s, z30.b, z25.b\n"
+      ".inst 0x45199b90  // smmla z16.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45189bcc  // smmla z12.s, z30.b, z24.b\n"
+      ".inst 0x45189b94  // smmla z20.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n"
+      ".inst 0x45199bc9  // smmla z9.s, z30.b, z25.b\n"
+      ".inst 0x45199b91  // smmla z17.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n"
+      ".inst 0x45189bcd  // smmla z13.s, z30.b, z24.b\n"
+      ".inst 0x45189b95  // smmla z21.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n"
+      ".inst 0x45199bca  // smmla z10.s, z30.b, z25.b\n"
+      ".inst 0x45199b92  // smmla z18.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n"
+      ".inst 0x45189bce  // smmla z14.s, z30.b, z24.b\n"
+      ".inst 0x45189b96  // smmla z22.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n"
+      ".inst 0x45199bcb  // smmla z11.s, z30.b, z25.b\n"
+      ".inst 0x45199b93  // smmla z19.s, z28.b, z25.b\n"
+      ".inst 0x45189bcf  // smmla z15.s, z30.b, z24.b\n"
+      ".inst 0x45189b97  // smmla z23.s, z28.b, z24.b\n"
       "bgt 30b\n"
       "31:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "ld1rqb { z24.b }, p0/Z, [x25]\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
       "trn1 z27.d, z1.d, z24.d\n"
       "trn2 z1.d, z1.d, z24.d\n"
-      "trn1 z26.d, z3.d, z29.d\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "trn1 z26.d, z3.d, z28.d\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45199b68  // smmla z8.s, z27.b, z25.b\n"
-      ".inst 0x451c9b6c  // smmla z12.s, z27.b, z28.b\n"
-      "trn2 z3.d, z3.d, z29.d\n"
       ".inst 0x45199b50  // smmla z16.s, z26.b, z25.b\n"
+      ".inst 0x45189b6c  // smmla z12.s, z27.b, z24.b\n"
+      ".inst 0x45189b54  // smmla z20.s, z26.b, z24.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x451c9b54  // smmla z20.s, z26.b, z28.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45199b69  // smmla z9.s, z27.b, z25.b\n"
       ".inst 0x45199b51  // smmla z17.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x45189b6d  // smmla z13.s, z27.b, z24.b\n"
       ".inst 0x45189b55  // smmla z21.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
+      "trn2 z3.d, z3.d, z28.d\n"
       ".inst 0x45199b6a  // smmla z10.s, z27.b, z25.b\n"
       ".inst 0x45199b52  // smmla z18.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
@@ -614,9 +614,9 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45199828  // smmla z8.s, z1.b, z25.b\n"
       ".inst 0x45199870  // smmla z16.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4518982c  // smmla z12.s, z1.b, z24.b\n"
       ".inst 0x45189874  // smmla z20.s, z3.b, z24.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45199829  // smmla z9.s, z1.b, z25.b\n"
       ".inst 0x45199871  // smmla z17.s, z3.b, z25.b\n"
@@ -641,26 +641,26 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 27b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z27.d, z8.d, z12.d\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "uzp1 z25.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z26.d, z9.d, z13.d\n"
+      "uzp1 z24.d, z9.d, z13.d\n"
+      "st1w { z25.s }, p4, [x9]\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z25.d, z10.d, z14.d\n"
+      "st1w { z24.s }, p3, [x9, #1, MUL VL]\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "uzp1 z24.d, z11.d, z15.d\n"
+      "st1w { z25.s }, p2, [x9, #2, MUL VL]\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z27.s }, p4, [x9]\n"
       "uzp1 z16.d, z16.d, z20.d\n"
-      "uzp1 z17.d, z17.d, z21.d\n"
-      "st1w { z26.s }, p3, [x9, #1, MUL VL]\n"
-      "uzp1 z18.d, z18.d, z22.d\n"
-      "uzp1 z19.d, z19.d, z23.d\n"
-      "st1w { z25.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z24.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
+      "uzp1 z17.d, z17.d, z21.d\n"
+      "uzp1 z18.d, z18.d, z22.d\n"
       "st1w { z8.s }, p4, [x21]\n"
+      "uzp1 z19.d, z19.d, z23.d\n"
       "st1w { z9.s }, p3, [x21, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x21, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x21, #3, MUL VL]\n"
@@ -676,7 +676,7 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "34:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "35:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -688,37 +688,37 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 36f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x22, x9, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x21]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x20]\n"
-      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x20]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
@@ -745,8 +745,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "38:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 39f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -771,114 +771,114 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "ble 42f\n"
       "41:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z31.b }, p5/Z, [x10]\n"
-      "ld1b { z30.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z29.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqb { z30.b }, p0/Z, [x26]\n"
+      "ld1rqb { z24.b }, p0/Z, [x25]\n"
+      "trn1 z29.d, z30.d, z24.d\n"
       "ld1rqb { z28.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "trn1 z27.d, z29.d, z25.d\n"
-      "trn2 z29.d, z29.d, z25.d\n"
-      "trn1 z26.d, z28.d, z24.d\n"
-      "trn2 z28.d, z28.d, z24.d\n"
-      ".inst 0x451f9b68  // smmla z8.s, z27.b, z31.b\n"
-      ".inst 0x451e9b6c  // smmla z12.s, z27.b, z30.b\n"
-      ".inst 0x451f9b50  // smmla z16.s, z26.b, z31.b\n"
+      "ld1rqb { z27.b }, p0/Z, [x23]\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "trn1 z26.d, z28.d, z27.d\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x45199ba8  // smmla z8.s, z29.b, z25.b\n"
+      ".inst 0x45199b50  // smmla z16.s, z26.b, z25.b\n"
+      ".inst 0x45189bac  // smmla z12.s, z29.b, z24.b\n"
+      ".inst 0x45189b54  // smmla z20.s, z26.b, z24.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x451e9b54  // smmla z20.s, z26.b, z30.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45199b69  // smmla z9.s, z27.b, z25.b\n"
+      ".inst 0x45199ba9  // smmla z9.s, z29.b, z25.b\n"
       ".inst 0x45199b51  // smmla z17.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45189b6d  // smmla z13.s, z27.b, z24.b\n"
+      "trn2 z28.d, z28.d, z27.d\n"
+      ".inst 0x45189bad  // smmla z13.s, z29.b, z24.b\n"
       ".inst 0x45189b55  // smmla z21.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45199b6a  // smmla z10.s, z27.b, z25.b\n"
+      "sub x27, x27, #0x10\n"
+      ".inst 0x45199baa  // smmla z10.s, z29.b, z25.b\n"
       ".inst 0x45199b52  // smmla z18.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45189b6e  // smmla z14.s, z27.b, z24.b\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45189bae  // smmla z14.s, z29.b, z24.b\n"
       ".inst 0x45189b56  // smmla z22.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x45199b6b  // smmla z11.s, z27.b, z25.b\n"
+      ".inst 0x45199bab  // smmla z11.s, z29.b, z25.b\n"
       ".inst 0x45199b53  // smmla z19.s, z26.b, z25.b\n"
-      ".inst 0x45189b6f  // smmla z15.s, z27.b, z24.b\n"
-      ".inst 0x45189b57  // smmla z23.s, z26.b, z24.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      "add x26, x26, #0x10\n"
+      ".inst 0x45189baf  // smmla z15.s, z29.b, z24.b\n"
+      ".inst 0x45189b57  // smmla z23.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45199ba8  // smmla z8.s, z29.b, z25.b\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x45199bc8  // smmla z8.s, z30.b, z25.b\n"
       ".inst 0x45199b90  // smmla z16.s, z28.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45189bac  // smmla z12.s, z29.b, z24.b\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45189bcc  // smmla z12.s, z30.b, z24.b\n"
       ".inst 0x45189b94  // smmla z20.s, z28.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45199ba9  // smmla z9.s, z29.b, z25.b\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x45199bc9  // smmla z9.s, z30.b, z25.b\n"
       ".inst 0x45199b91  // smmla z17.s, z28.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45189bad  // smmla z13.s, z29.b, z24.b\n"
+      ".inst 0x45189bcd  // smmla z13.s, z30.b, z24.b\n"
       ".inst 0x45189b95  // smmla z21.s, z28.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45199baa  // smmla z10.s, z29.b, z25.b\n"
+      ".inst 0x45199bca  // smmla z10.s, z30.b, z25.b\n"
       ".inst 0x45199b92  // smmla z18.s, z28.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45189bae  // smmla z14.s, z29.b, z24.b\n"
+      ".inst 0x45189bce  // smmla z14.s, z30.b, z24.b\n"
       ".inst 0x45189b96  // smmla z22.s, z28.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45199bab  // smmla z11.s, z29.b, z25.b\n"
+      ".inst 0x45199bcb  // smmla z11.s, z30.b, z25.b\n"
       ".inst 0x45199b93  // smmla z19.s, z28.b, z25.b\n"
-      ".inst 0x45189baf  // smmla z15.s, z29.b, z24.b\n"
+      ".inst 0x45189bcf  // smmla z15.s, z30.b, z24.b\n"
       ".inst 0x45189b97  // smmla z23.s, z28.b, z24.b\n"
       "bgt 41b\n"
       "42:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
+      "ld1rqb { z24.b }, p0/Z, [x25]\n"
+      "trn1 z28.d, z1.d, z24.d\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "trn1 z27.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      "trn1 z26.d, z3.d, z24.d\n"
-      ".inst 0x451d9b68  // smmla z8.s, z27.b, z29.b\n"
-      ".inst 0x451c9b6c  // smmla z12.s, z27.b, z28.b\n"
-      "trn2 z3.d, z3.d, z24.d\n"
-      ".inst 0x451d9b50  // smmla z16.s, z26.b, z29.b\n"
+      "ld1rqb { z27.b }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z24.d\n"
+      "trn1 z26.d, z3.d, z27.d\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x45199b88  // smmla z8.s, z28.b, z25.b\n"
+      ".inst 0x45199b50  // smmla z16.s, z26.b, z25.b\n"
+      ".inst 0x45189b8c  // smmla z12.s, z28.b, z24.b\n"
+      ".inst 0x45189b54  // smmla z20.s, z26.b, z24.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x451c9b54  // smmla z20.s, z26.b, z28.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45199b69  // smmla z9.s, z27.b, z25.b\n"
+      ".inst 0x45199b89  // smmla z9.s, z28.b, z25.b\n"
       ".inst 0x45199b51  // smmla z17.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45189b6d  // smmla z13.s, z27.b, z24.b\n"
+      "subs x27, x27, #0x8\n"
+      ".inst 0x45189b8d  // smmla z13.s, z28.b, z24.b\n"
       ".inst 0x45189b55  // smmla z21.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45199b6a  // smmla z10.s, z27.b, z25.b\n"
+      "trn2 z3.d, z3.d, z27.d\n"
+      ".inst 0x45199b8a  // smmla z10.s, z28.b, z25.b\n"
       ".inst 0x45199b52  // smmla z18.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45189b6e  // smmla z14.s, z27.b, z24.b\n"
+      ".inst 0x45189b8e  // smmla z14.s, z28.b, z24.b\n"
       ".inst 0x45189b56  // smmla z22.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x45199b6b  // smmla z11.s, z27.b, z25.b\n"
+      ".inst 0x45199b8b  // smmla z11.s, z28.b, z25.b\n"
       ".inst 0x45199b53  // smmla z19.s, z26.b, z25.b\n"
-      ".inst 0x45189b6f  // smmla z15.s, z27.b, z24.b\n"
+      ".inst 0x45189b8f  // smmla z15.s, z28.b, z24.b\n"
       ".inst 0x45189b57  // smmla z23.s, z26.b, z24.b\n"
       "ble 43f\n"
       "ld1b { z25.b }, p5/Z, [x10]\n"
       "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45199828  // smmla z8.s, z1.b, z25.b\n"
       ".inst 0x45199870  // smmla z16.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4518982c  // smmla z12.s, z1.b, z24.b\n"
       ".inst 0x45189874  // smmla z20.s, z3.b, z24.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45199829  // smmla z9.s, z1.b, z25.b\n"
       ".inst 0x45199871  // smmla z17.s, z3.b, z25.b\n"
@@ -903,33 +903,33 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 38b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp1 z25.d, z8.d, z12.d\n"
+      "add x20, x21, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z24.d, z9.d, z13.d\n"
+      "st1w { z25.s }, p4, [x9]\n"
       "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z27.d, z10.d, z14.d\n"
+      "uzp1 z25.d, z10.d, z14.d\n"
+      "st1w { z24.s }, p3, [x9, #1, MUL VL]\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "uzp1 z26.d, z11.d, z15.d\n"
+      "uzp1 z24.d, z11.d, z15.d\n"
+      "st1w { z25.s }, p2, [x9, #2, MUL VL]\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z25.s }, p4, [x9]\n"
       "uzp1 z25.d, z16.d, z20.d\n"
+      "st1w { z24.s }, p1, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z24.s }, p3, [x9, #1, MUL VL]\n"
       "uzp1 z24.d, z17.d, z21.d\n"
+      "st1w { z8.s }, p4, [x22]\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z27.s }, p2, [x9, #2, MUL VL]\n"
       "uzp1 z21.d, z18.d, z22.d\n"
+      "st1w { z9.s }, p3, [x22, #1, MUL VL]\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z26.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
       "uzp1 z20.d, z19.d, z23.d\n"
-      "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x22]\n"
-      "st1w { z9.s }, p3, [x22, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x22, #2, MUL VL]\n"
+      "uzp2 z19.d, z19.d, z23.d\n"
       "st1w { z11.s }, p1, [x22, #3, MUL VL]\n"
       "st1w { z25.s }, p4, [x21]\n"
       "st1w { z24.s }, p3, [x21, #1, MUL VL]\n"
@@ -947,7 +947,7 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "45:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "46:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -959,46 +959,46 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 47f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x22]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x21]\n"
-      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x20]\n"
-      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
+      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x21]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x20]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z19.d, z24.d, z23.d\n"
       "zip2 z23.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z24.d, z25.d, z28.d\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
@@ -1037,8 +1037,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "49:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 50f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1066,103 +1066,102 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "ble 53f\n"
       "52:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p5/Z, [x10]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z6.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqb { z1.b }, p0/Z, [x25]\n"
       "ld1rqb { z7.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
+      "trn1 z5.d, z6.d, z1.d\n"
+      "trn2 z6.d, z6.d, z1.d\n"
+      "ld1rqb { z4.b }, p0/Z, [x22]\n"
       "trn1 z3.d, z7.d, z2.d\n"
       "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
+      "ld1b { z1.b }, p5/Z, [x10]\n"
+      "trn1 z2.d, z4.d, z0.d\n"
+      "trn2 z4.d, z4.d, z0.d\n"
       "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45019888  // smmla z8.s, z4.b, z1.b\n"
+      ".inst 0x450198a8  // smmla z8.s, z5.b, z1.b\n"
       ".inst 0x45019870  // smmla z16.s, z3.b, z1.b\n"
       ".inst 0x45019858  // smmla z24.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4500988c  // smmla z12.s, z4.b, z0.b\n"
+      "sub x27, x27, #0x10\n"
+      ".inst 0x450098ac  // smmla z12.s, z5.b, z0.b\n"
       ".inst 0x45009874  // smmla z20.s, z3.b, z0.b\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4500985c  // smmla z28.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45019889  // smmla z9.s, z4.b, z1.b\n"
+      ".inst 0x450198a9  // smmla z9.s, z5.b, z1.b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x45019871  // smmla z17.s, z3.b, z1.b\n"
       ".inst 0x45019859  // smmla z25.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4500988d  // smmla z13.s, z4.b, z0.b\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x450098ad  // smmla z13.s, z5.b, z0.b\n"
       ".inst 0x45009875  // smmla z21.s, z3.b, z0.b\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4500985d  // smmla z29.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4501988a  // smmla z10.s, z4.b, z1.b\n"
+      ".inst 0x450198aa  // smmla z10.s, z5.b, z1.b\n"
       ".inst 0x45019872  // smmla z18.s, z3.b, z1.b\n"
       ".inst 0x4501985a  // smmla z26.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4500988e  // smmla z14.s, z4.b, z0.b\n"
+      ".inst 0x450098ae  // smmla z14.s, z5.b, z0.b\n"
       ".inst 0x45009876  // smmla z22.s, z3.b, z0.b\n"
       ".inst 0x4500985e  // smmla z30.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x4501988b  // smmla z11.s, z4.b, z1.b\n"
+      ".inst 0x450198ab  // smmla z11.s, z5.b, z1.b\n"
       ".inst 0x45019873  // smmla z19.s, z3.b, z1.b\n"
       ".inst 0x4501985b  // smmla z27.s, z2.b, z1.b\n"
-      ".inst 0x4500988f  // smmla z15.s, z4.b, z0.b\n"
+      "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      ".inst 0x450098af  // smmla z15.s, z5.b, z0.b\n"
       ".inst 0x45009877  // smmla z23.s, z3.b, z0.b\n"
       ".inst 0x4500985f  // smmla z31.s, z2.b, z0.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z0.b }, p5/Z, [x10, #-7, MUL VL]\n"
       ".inst 0x450198c8  // smmla z8.s, z6.b, z1.b\n"
       ".inst 0x450198f0  // smmla z16.s, z7.b, z1.b\n"
-      ".inst 0x450198b8  // smmla z24.s, z5.b, z1.b\n"
+      ".inst 0x45019898  // smmla z24.s, z4.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x450098cc  // smmla z12.s, z6.b, z0.b\n"
       ".inst 0x450098f4  // smmla z20.s, z7.b, z0.b\n"
-      ".inst 0x450098bc  // smmla z28.s, z5.b, z0.b\n"
+      ".inst 0x4500989c  // smmla z28.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-5, MUL VL]\n"
       ".inst 0x450198c9  // smmla z9.s, z6.b, z1.b\n"
       ".inst 0x450198f1  // smmla z17.s, z7.b, z1.b\n"
-      ".inst 0x450198b9  // smmla z25.s, z5.b, z1.b\n"
+      ".inst 0x45019899  // smmla z25.s, z4.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x450098cd  // smmla z13.s, z6.b, z0.b\n"
       ".inst 0x450098f5  // smmla z21.s, z7.b, z0.b\n"
-      ".inst 0x450098bd  // smmla z29.s, z5.b, z0.b\n"
+      ".inst 0x4500989d  // smmla z29.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-3, MUL VL]\n"
       ".inst 0x450198ca  // smmla z10.s, z6.b, z1.b\n"
       ".inst 0x450198f2  // smmla z18.s, z7.b, z1.b\n"
-      ".inst 0x450198ba  // smmla z26.s, z5.b, z1.b\n"
+      ".inst 0x4501989a  // smmla z26.s, z4.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x450098ce  // smmla z14.s, z6.b, z0.b\n"
       ".inst 0x450098f6  // smmla z22.s, z7.b, z0.b\n"
-      ".inst 0x450098be  // smmla z30.s, z5.b, z0.b\n"
+      ".inst 0x4500989e  // smmla z30.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-1, MUL VL]\n"
       ".inst 0x450198cb  // smmla z11.s, z6.b, z1.b\n"
       ".inst 0x450198f3  // smmla z19.s, z7.b, z1.b\n"
-      ".inst 0x450198bb  // smmla z27.s, z5.b, z1.b\n"
+      ".inst 0x4501989b  // smmla z27.s, z4.b, z1.b\n"
       ".inst 0x450098cf  // smmla z15.s, z6.b, z0.b\n"
       ".inst 0x450098f7  // smmla z23.s, z7.b, z0.b\n"
-      ".inst 0x450098bf  // smmla z31.s, z5.b, z0.b\n"
+      ".inst 0x4500989f  // smmla z31.s, z4.b, z0.b\n"
       "bgt 52b\n"
       "53:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z2.b }, p5/Z, [x10]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z6.b }, p0/Z, [x25]\n"
+      "ld1rqb { z4.b }, p0/Z, [x25]\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "ld1rqb { z2.b }, p0/Z, [x23]\n"
+      "trn1 z7.d, z1.d, z4.d\n"
+      "trn2 z1.d, z1.d, z4.d\n"
       "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
+      "ld1b { z2.b }, p5/Z, [x10]\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
       "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
@@ -1170,6 +1169,7 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       ".inst 0x450298d0  // smmla z16.s, z6.b, z2.b\n"
       ".inst 0x45029898  // smmla z24.s, z4.b, z2.b\n"
       "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x450098ec  // smmla z12.s, z7.b, z0.b\n"
       ".inst 0x450098d4  // smmla z20.s, z6.b, z0.b\n"
       ".inst 0x4500989c  // smmla z28.s, z4.b, z0.b\n"
@@ -1190,8 +1190,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       ".inst 0x450098d6  // smmla z22.s, z6.b, z0.b\n"
       ".inst 0x4500989e  // smmla z30.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x450298eb  // smmla z11.s, z7.b, z2.b\n"
+      "addvl x10, x10, #8\n"
       ".inst 0x450298d3  // smmla z19.s, z6.b, z2.b\n"
       ".inst 0x4502989b  // smmla z27.s, z4.b, z2.b\n"
       ".inst 0x450098ef  // smmla z15.s, z7.b, z0.b\n"
@@ -1203,24 +1203,24 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       ".inst 0x45029828  // smmla z8.s, z1.b, z2.b\n"
       ".inst 0x45029870  // smmla z16.s, z3.b, z2.b\n"
       ".inst 0x450298b8  // smmla z24.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4500982c  // smmla z12.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45009874  // smmla z20.s, z3.b, z0.b\n"
       ".inst 0x450098bc  // smmla z28.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45029829  // smmla z9.s, z1.b, z2.b\n"
       ".inst 0x45029871  // smmla z17.s, z3.b, z2.b\n"
       ".inst 0x450298b9  // smmla z25.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x4500982d  // smmla z13.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45009875  // smmla z21.s, z3.b, z0.b\n"
       ".inst 0x450098bd  // smmla z29.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x4502982a  // smmla z10.s, z1.b, z2.b\n"
       ".inst 0x45029872  // smmla z18.s, z3.b, z2.b\n"
       ".inst 0x450298ba  // smmla z26.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x4500982e  // smmla z14.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45009876  // smmla z22.s, z3.b, z0.b\n"
       ".inst 0x450098be  // smmla z30.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
@@ -1237,39 +1237,39 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 49b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z1.d, z8.d, z12.d\n"
-      "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z0.d, z9.d, z13.d\n"
-      "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z3.d, z10.d, z14.d\n"
-      "uzp2 z10.d, z10.d, z14.d\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "uzp1 z2.d, z8.d, z12.d\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "uzp2 z8.d, z8.d, z12.d\n"
+      "uzp1 z1.d, z9.d, z13.d\n"
+      "uzp2 z9.d, z9.d, z13.d\n"
+      "uzp1 z0.d, z10.d, z14.d\n"
+      "st1w { z2.s }, p4, [x9]\n"
+      "uzp2 z10.d, z10.d, z14.d\n"
       "uzp1 z2.d, z11.d, z15.d\n"
+      "st1w { z1.s }, p3, [x9, #1, MUL VL]\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z1.s }, p4, [x9]\n"
-      "add x20, x21, x20, LSL #2\n"
       "uzp1 z1.d, z16.d, z20.d\n"
+      "st1w { z0.s }, p2, [x9, #2, MUL VL]\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z0.s }, p3, [x9, #1, MUL VL]\n"
       "uzp1 z0.d, z17.d, z21.d\n"
+      "st1w { z2.s }, p1, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z3.s }, p2, [x9, #2, MUL VL]\n"
       "uzp1 z21.d, z18.d, z22.d\n"
+      "st1w { z8.s }, p4, [x23]\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z2.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
       "uzp1 z20.d, z19.d, z23.d\n"
+      "st1w { z9.s }, p3, [x23, #1, MUL VL]\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x23]\n"
       "uzp1 z24.d, z24.d, z28.d\n"
+      "st1w { z10.s }, p2, [x23, #2, MUL VL]\n"
       "uzp1 z25.d, z25.d, z29.d\n"
-      "st1w { z9.s }, p3, [x23, #1, MUL VL]\n"
       "uzp1 z26.d, z26.d, z30.d\n"
-      "uzp1 z27.d, z27.d, z31.d\n"
-      "st1w { z10.s }, p2, [x23, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x23, #3, MUL VL]\n"
+      "uzp1 z27.d, z27.d, z31.d\n"
       "st1w { z1.s }, p4, [x22]\n"
       "st1w { z0.s }, p3, [x22, #1, MUL VL]\n"
       "st1w { z21.s }, p2, [x22, #2, MUL VL]\n"
@@ -1289,12 +1289,11 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "b 68f\n"
       "56:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "57:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1306,54 +1305,54 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 58f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "ld1w { z17.s }, p4/Z, [x9]\n"
       "add x22, x23, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x24]\n"
-      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x9, #2, MUL VL]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
+      "ld1w { z12.s }, p4/Z, [x24]\n"
+      "zip1 z8.d, z17.d, z12.d\n"
+      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
+      "zip2 z12.d, z17.d, z12.d\n"
+      "zip1 z9.d, z18.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x23]\n"
+      "zip2 z13.d, z18.d, z13.d\n"
+      "zip1 z10.d, z20.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip2 z14.d, z20.d, z14.d\n"
+      "zip1 z11.d, z16.d, z15.d\n"
       "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
       "ld1w { z20.s }, p4/Z, [x22]\n"
-      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x21]\n"
-      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
       "zip2 z20.d, z17.d, z20.d\n"
-      "ld1w { z28.s }, p4/Z, [x20]\n"
-      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x21]\n"
       "zip2 z21.d, z18.d, z21.d\n"
-      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z28.s }, p4/Z, [x20]\n"
       "zip2 z23.d, z24.d, z23.d\n"
       "zip1 z24.d, z25.d, z28.d\n"
+      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
+      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip2 z29.d, z26.d, z29.d\n"
       "zip1 z26.d, z27.d, z30.d\n"
       "zip2 z30.d, z27.d, z30.d\n"
@@ -1389,8 +1388,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "60:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 61f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1421,113 +1420,113 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "ble 64f\n"
       "63:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p5/Z, [x10]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z6.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z7.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
+      "ld1rqb { z7.b }, p0/Z, [x26]\n"
+      "ld1rqb { z0.b }, p0/Z, [x25]\n"
+      "trn1 z6.d, z7.d, z0.d\n"
+      "ld1rqb { z5.b }, p0/Z, [x24]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "trn2 z7.d, z7.d, z0.d\n"
+      "trn1 z4.d, z5.d, z1.d\n"
+      "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "ld1rqb { z0.b }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
-      "add x21, x21, #0x10\n"
-      "trn1 z3.d, z7.d, z2.d\n"
-      "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
+      "trn2 z5.d, z5.d, z1.d\n"
+      "trn1 z2.d, z3.d, z0.d\n"
+      "trn2 z3.d, z3.d, z0.d\n"
+      "ld1b { z1.b }, p5/Z, [x10]\n"
       "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45019888  // smmla z8.s, z4.b, z1.b\n"
-      ".inst 0x45019870  // smmla z16.s, z3.b, z1.b\n"
+      ".inst 0x450198c8  // smmla z8.s, z6.b, z1.b\n"
+      ".inst 0x45019890  // smmla z16.s, z4.b, z1.b\n"
       ".inst 0x45019858  // smmla z24.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4500988c  // smmla z12.s, z4.b, z0.b\n"
-      ".inst 0x45009874  // smmla z20.s, z3.b, z0.b\n"
+      "sub x27, x27, #0x10\n"
+      ".inst 0x450098cc  // smmla z12.s, z6.b, z0.b\n"
+      ".inst 0x45009894  // smmla z20.s, z4.b, z0.b\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x4500985c  // smmla z28.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45019889  // smmla z9.s, z4.b, z1.b\n"
-      ".inst 0x45019871  // smmla z17.s, z3.b, z1.b\n"
+      ".inst 0x450198c9  // smmla z9.s, z6.b, z1.b\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x45019891  // smmla z17.s, z4.b, z1.b\n"
       ".inst 0x45019859  // smmla z25.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4500988d  // smmla z13.s, z4.b, z0.b\n"
-      ".inst 0x45009875  // smmla z21.s, z3.b, z0.b\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x450098cd  // smmla z13.s, z6.b, z0.b\n"
+      ".inst 0x45009895  // smmla z21.s, z4.b, z0.b\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x4500985d  // smmla z29.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4501988a  // smmla z10.s, z4.b, z1.b\n"
-      ".inst 0x45019872  // smmla z18.s, z3.b, z1.b\n"
+      ".inst 0x450198ca  // smmla z10.s, z6.b, z1.b\n"
+      "add x21, x21, #0x10\n"
+      ".inst 0x45019892  // smmla z18.s, z4.b, z1.b\n"
       ".inst 0x4501985a  // smmla z26.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4500988e  // smmla z14.s, z4.b, z0.b\n"
-      ".inst 0x45009876  // smmla z22.s, z3.b, z0.b\n"
+      ".inst 0x450098ce  // smmla z14.s, z6.b, z0.b\n"
+      ".inst 0x45009896  // smmla z22.s, z4.b, z0.b\n"
       ".inst 0x4500985e  // smmla z30.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x4501988b  // smmla z11.s, z4.b, z1.b\n"
-      ".inst 0x45019873  // smmla z19.s, z3.b, z1.b\n"
+      ".inst 0x450198cb  // smmla z11.s, z6.b, z1.b\n"
+      ".inst 0x45019893  // smmla z19.s, z4.b, z1.b\n"
       ".inst 0x4501985b  // smmla z27.s, z2.b, z1.b\n"
-      ".inst 0x4500988f  // smmla z15.s, z4.b, z0.b\n"
-      ".inst 0x45009877  // smmla z23.s, z3.b, z0.b\n"
-      ".inst 0x4500985f  // smmla z31.s, z2.b, z0.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      ".inst 0x450098cf  // smmla z15.s, z6.b, z0.b\n"
+      ".inst 0x45009897  // smmla z23.s, z4.b, z0.b\n"
+      ".inst 0x4500985f  // smmla z31.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x450198c8  // smmla z8.s, z6.b, z1.b\n"
-      ".inst 0x450198f0  // smmla z16.s, z7.b, z1.b\n"
-      ".inst 0x450198b8  // smmla z24.s, z5.b, z1.b\n"
+      ".inst 0x450198e8  // smmla z8.s, z7.b, z1.b\n"
+      ".inst 0x450198b0  // smmla z16.s, z5.b, z1.b\n"
+      ".inst 0x45019878  // smmla z24.s, z3.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x450098cc  // smmla z12.s, z6.b, z0.b\n"
-      ".inst 0x450098f4  // smmla z20.s, z7.b, z0.b\n"
-      ".inst 0x450098bc  // smmla z28.s, z5.b, z0.b\n"
+      ".inst 0x450098ec  // smmla z12.s, z7.b, z0.b\n"
+      ".inst 0x450098b4  // smmla z20.s, z5.b, z0.b\n"
+      ".inst 0x4500987c  // smmla z28.s, z3.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x450198c9  // smmla z9.s, z6.b, z1.b\n"
-      ".inst 0x450198f1  // smmla z17.s, z7.b, z1.b\n"
-      ".inst 0x450198b9  // smmla z25.s, z5.b, z1.b\n"
+      ".inst 0x450198e9  // smmla z9.s, z7.b, z1.b\n"
+      ".inst 0x450198b1  // smmla z17.s, z5.b, z1.b\n"
+      ".inst 0x45019879  // smmla z25.s, z3.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x450098cd  // smmla z13.s, z6.b, z0.b\n"
-      ".inst 0x450098f5  // smmla z21.s, z7.b, z0.b\n"
-      ".inst 0x450098bd  // smmla z29.s, z5.b, z0.b\n"
+      ".inst 0x450098ed  // smmla z13.s, z7.b, z0.b\n"
+      ".inst 0x450098b5  // smmla z21.s, z5.b, z0.b\n"
+      ".inst 0x4500987d  // smmla z29.s, z3.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x450198ca  // smmla z10.s, z6.b, z1.b\n"
-      ".inst 0x450198f2  // smmla z18.s, z7.b, z1.b\n"
-      ".inst 0x450198ba  // smmla z26.s, z5.b, z1.b\n"
+      ".inst 0x450198ea  // smmla z10.s, z7.b, z1.b\n"
+      ".inst 0x450198b2  // smmla z18.s, z5.b, z1.b\n"
+      ".inst 0x4501987a  // smmla z26.s, z3.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x450098ce  // smmla z14.s, z6.b, z0.b\n"
-      ".inst 0x450098f6  // smmla z22.s, z7.b, z0.b\n"
-      ".inst 0x450098be  // smmla z30.s, z5.b, z0.b\n"
+      ".inst 0x450098ee  // smmla z14.s, z7.b, z0.b\n"
+      ".inst 0x450098b6  // smmla z22.s, z5.b, z0.b\n"
+      ".inst 0x4500987e  // smmla z30.s, z3.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x450198cb  // smmla z11.s, z6.b, z1.b\n"
-      ".inst 0x450198f3  // smmla z19.s, z7.b, z1.b\n"
-      ".inst 0x450198bb  // smmla z27.s, z5.b, z1.b\n"
-      ".inst 0x450098cf  // smmla z15.s, z6.b, z0.b\n"
-      ".inst 0x450098f7  // smmla z23.s, z7.b, z0.b\n"
-      ".inst 0x450098bf  // smmla z31.s, z5.b, z0.b\n"
+      ".inst 0x450198eb  // smmla z11.s, z7.b, z1.b\n"
+      ".inst 0x450198b3  // smmla z19.s, z5.b, z1.b\n"
+      ".inst 0x4501987b  // smmla z27.s, z3.b, z1.b\n"
+      ".inst 0x450098ef  // smmla z15.s, z7.b, z0.b\n"
+      ".inst 0x450098b7  // smmla z23.s, z5.b, z0.b\n"
+      ".inst 0x4500987f  // smmla z31.s, z3.b, z0.b\n"
       "bgt 63b\n"
       "64:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z2.b }, p5/Z, [x10]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z6.b }, p0/Z, [x25]\n"
+      "ld1rqb { z0.b }, p0/Z, [x25]\n"
+      "trn1 z7.d, z1.d, z0.d\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "ld1rqb { z2.b }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z0.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
       "ld1rqb { z5.b }, p0/Z, [x22]\n"
       "ld1rqb { z0.b }, p0/Z, [x21]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
+      "ld1b { z2.b }, p5/Z, [x10]\n"
       "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x450298e8  // smmla z8.s, z7.b, z2.b\n"
       ".inst 0x450298d0  // smmla z16.s, z6.b, z2.b\n"
       ".inst 0x45029898  // smmla z24.s, z4.b, z2.b\n"
       "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x450098ec  // smmla z12.s, z7.b, z0.b\n"
       ".inst 0x450098d4  // smmla z20.s, z6.b, z0.b\n"
       ".inst 0x4500989c  // smmla z28.s, z4.b, z0.b\n"
@@ -1548,8 +1547,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       ".inst 0x450098d6  // smmla z22.s, z6.b, z0.b\n"
       ".inst 0x4500989e  // smmla z30.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x450298eb  // smmla z11.s, z7.b, z2.b\n"
+      "addvl x10, x10, #8\n"
       ".inst 0x450298d3  // smmla z19.s, z6.b, z2.b\n"
       ".inst 0x4502989b  // smmla z27.s, z4.b, z2.b\n"
       ".inst 0x450098ef  // smmla z15.s, z7.b, z0.b\n"
@@ -1561,24 +1560,24 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       ".inst 0x45029828  // smmla z8.s, z1.b, z2.b\n"
       ".inst 0x45029870  // smmla z16.s, z3.b, z2.b\n"
       ".inst 0x450298b8  // smmla z24.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x4500982c  // smmla z12.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45009874  // smmla z20.s, z3.b, z0.b\n"
       ".inst 0x450098bc  // smmla z28.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45029829  // smmla z9.s, z1.b, z2.b\n"
       ".inst 0x45029871  // smmla z17.s, z3.b, z2.b\n"
       ".inst 0x450298b9  // smmla z25.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x4500982d  // smmla z13.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45009875  // smmla z21.s, z3.b, z0.b\n"
       ".inst 0x450098bd  // smmla z29.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x4502982a  // smmla z10.s, z1.b, z2.b\n"
       ".inst 0x45029872  // smmla z18.s, z3.b, z2.b\n"
       ".inst 0x450298ba  // smmla z26.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x4500982e  // smmla z14.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45009876  // smmla z22.s, z3.b, z0.b\n"
       ".inst 0x450098be  // smmla z30.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
@@ -1595,46 +1594,46 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 60b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 z0.d, z8.d, z12.d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
+      "add x20, x21, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
+      "st1w { z0.s }, p4, [x9]\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
+      "st1w { z12.s }, p3, [x9, #1, MUL VL]\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z0.s }, p4, [x9]\n"
-      "add x21, x22, x20, LSL #2\n"
       "uzp1 z15.d, z16.d, z20.d\n"
+      "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z12.s }, p3, [x9, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "uzp1 z20.d, z17.d, z21.d\n"
+      "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "uzp1 z21.d, z18.d, z22.d\n"
+      "st1w { z8.s }, p4, [x24]\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
       "uzp1 z22.d, z19.d, z23.d\n"
+      "st1w { z9.s }, p3, [x24, #1, MUL VL]\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x24]\n"
       "uzp1 z23.d, z24.d, z28.d\n"
+      "st1w { z10.s }, p2, [x24, #2, MUL VL]\n"
       "uzp2 z24.d, z24.d, z28.d\n"
-      "st1w { z9.s }, p3, [x24, #1, MUL VL]\n"
       "uzp1 z28.d, z25.d, z29.d\n"
+      "st1w { z11.s }, p1, [x24, #3, MUL VL]\n"
       "uzp2 z25.d, z25.d, z29.d\n"
-      "st1w { z10.s }, p2, [x24, #2, MUL VL]\n"
       "uzp1 z29.d, z26.d, z30.d\n"
+      "st1w { z15.s }, p4, [x23]\n"
       "uzp2 z26.d, z26.d, z30.d\n"
-      "st1w { z11.s }, p1, [x24, #3, MUL VL]\n"
       "uzp1 z30.d, z27.d, z31.d\n"
-      "uzp2 z27.d, z27.d, z31.d\n"
-      "st1w { z15.s }, p4, [x23]\n"
       "st1w { z20.s }, p3, [x23, #1, MUL VL]\n"
+      "uzp2 z27.d, z27.d, z31.d\n"
       "st1w { z21.s }, p2, [x23, #2, MUL VL]\n"
       "st1w { z22.s }, p1, [x23, #3, MUL VL]\n"
       "st1w { z16.s }, p4, [x22]\n"
@@ -1665,8 +1664,8 @@ void sve_hybrid_s8s32_mmla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "68:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp
index 8d508f94f0..839ff6f0af 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 4, 4, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 4, 4, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp
index 7b598bac57..85e933fd46 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_dot_4x4VL/generic.cpp
@@ -45,18 +45,18 @@ void sve_hybrid_u8qa_dot_4x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -91,20 +91,20 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "3:"  // Height 1: setup done
       "mov x26, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -120,41 +120,41 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ble 9f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z21.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z23.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x24, x24, #0x10\n"
-      "udot z16.s, z21.b, z0.b[0]\n"
-      "ld1b { z21.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      "udot z17.s, z26.b, z0.b[0]\n"
-      "udot z18.s, z25.b, z0.b[0]\n"
-      "udot z19.s, z24.b, z0.b[0]\n"
+      "ld1b { z20.b }, p2/Z, [x28]\n"
+      "udot z16.s, z20.b, z0.b[0]\n"
+      "ld1b { z21.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "udot z17.s, z21.b, z0.b[0]\n"
+      "udot z18.s, z20.b, z0.b[0]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "udot z19.s, z20.b, z0.b[0]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #4, MUL VL]\n"
       "udot z16.s, z20.b, z0.b[1]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      "udot z17.s, z23.b, z0.b[1]\n"
-      "ld1b { z23.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      "udot z18.s, z22.b, z0.b[1]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      "udot z19.s, z21.b, z0.b[1]\n"
+      "ld1b { z21.b }, p2/Z, [x28, #5, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #6, MUL VL]\n"
+      "udot z17.s, z21.b, z0.b[1]\n"
+      "udot z18.s, z20.b, z0.b[1]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #7, MUL VL]\n"
+      "addvl x28, x28, #16\n"
+      "udot z19.s, z20.b, z0.b[1]\n"
+      "ld1b { z22.b }, p2/Z, [x28, #-8, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      "ld1b { z21.b }, p2/Z, [x28, #-6, MUL VL]\n"
+      "udot z16.s, z22.b, z0.b[2]\n"
+      "udot z17.s, z20.b, z0.b[2]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      "udot z18.s, z21.b, z0.b[2]\n"
+      "udot z19.s, z20.b, z0.b[2]\n"
+      "ld1b { z22.b }, p2/Z, [x28, #-4, MUL VL]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "ld1b { z21.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      "udot z16.s, z20.b, z0.b[2]\n"
+      "udot z16.s, z22.b, z0.b[3]\n"
+      "udot z17.s, z20.b, z0.b[3]\n"
       "ld1b { z20.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      "udot z17.s, z26.b, z0.b[2]\n"
-      "udot z18.s, z25.b, z0.b[2]\n"
-      "udot z19.s, z24.b, z0.b[2]\n"
-      "udot z16.s, z23.b, z0.b[3]\n"
-      "udot z17.s, z22.b, z0.b[3]\n"
       "udot z18.s, z21.b, z0.b[3]\n"
       "udot z19.s, z20.b, z0.b[3]\n"
+      "add x24, x24, #0x10\n"
       "tbnz %x[flags], #31, 8f\n"
       "udot z11.s, z0.b, z15.b\n"
       "8:"  // Height 1: Multiply loop: unique 1: skip row sum
@@ -163,49 +163,49 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "bgt 7b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1rqb { z0.b }, p0/Z, [x24]\n"
+      "ld1b { z22.b }, p2/Z, [x28]\n"
       "subs x25, x25, #0x4\n"
+      "ld1b { z20.b }, p2/Z, [x28, #1, MUL VL]\n"
       "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "udot z16.s, z22.b, z0.b[0]\n"
+      "udot z17.s, z20.b, z0.b[0]\n"
       "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "udot z16.s, z23.b, z0.b[0]\n"
-      "udot z17.s, z22.b, z0.b[0]\n"
       "udot z18.s, z21.b, z0.b[0]\n"
       "udot z19.s, z20.b, z0.b[0]\n"
+      "addvl x28, x28, #4\n"
       "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
+      "ld1b { z20.b }, p2/Z, [x28]\n"
       "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "udot z16.s, z20.b, z0.b[1]\n"
       "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z23.b, z0.b[1]\n"
       "udot z17.s, z22.b, z0.b[1]\n"
       "udot z18.s, z21.b, z0.b[1]\n"
       "udot z19.s, z20.b, z0.b[1]\n"
+      "addvl x28, x28, #4\n"
       "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
+      "ld1b { z20.b }, p2/Z, [x28]\n"
       "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "udot z16.s, z20.b, z0.b[2]\n"
       "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z23.b, z0.b[2]\n"
       "udot z17.s, z22.b, z0.b[2]\n"
       "udot z18.s, z21.b, z0.b[2]\n"
       "udot z19.s, z20.b, z0.b[2]\n"
+      "addvl x28, x28, #4\n"
       "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z21.b }, p2/Z, [x28]\n"
+      "ld1b { z20.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "udot z16.s, z21.b, z0.b[3]\n"
+      "udot z17.s, z20.b, z0.b[3]\n"
       "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z23.b, z0.b[3]\n"
-      "udot z17.s, z22.b, z0.b[3]\n"
       "udot z18.s, z21.b, z0.b[3]\n"
       "udot z19.s, z20.b, z0.b[3]\n"
+      "addvl x28, x28, #4\n"
       "10:"  // Height 1: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 11f\n"
       "udot z11.s, z0.b, z15.b\n"
@@ -215,35 +215,35 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "cmp x26, x20\n"
       "bne 4b\n"
       "tbnz %x[flags], #31, 12f\n"
-      "mov x21, #0x4\n"
+      "mov x20, #0x4\n"
+      "whilelt p0.s, XZR, x20\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z20.s, p2/M, z20.s\n"
       "uaddv d11, p0, z11.s\n"
       "mov z11.s, z11.s[0]\n"
+      "neg z20.s, p2/M, z20.s\n"
       "mul z11.s, p2/M, z11.s, z20.s\n"
       "12:"  // Height 1: skip row sum fixup
       "add z16.s, z16.s, z11.s\n"
       "add z17.s, z17.s, z11.s\n"
       "ld1w { z23.s }, p2/Z, [x10]\n"
-      "ld1w { z20.s }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x10, #1, MUL VL]\n"
       "add z18.s, z18.s, z11.s\n"
       "add z19.s, z19.s, z11.s\n"
-      "ld1w { z22.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1w { z21.s }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x10, #3, MUL VL]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
       "add z16.s, z16.s, z23.s\n"
-      "add z17.s, z17.s, z20.s\n"
+      "add z17.s, z17.s, z22.s\n"
+      "add z18.s, z18.s, z21.s\n"
+      "add z19.s, z19.s, z20.s\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "addvl x10, x10, #4\n"
-      "add z18.s, z18.s, z22.s\n"
-      "add z19.s, z19.s, z21.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
       ".inst 0x04b47610  // sqrdmulh z16.s, z16.s, z20.s\n"
       ".inst 0x04b47631  // sqrdmulh z17.s, z17.s, z20.s\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x04b47652  // sqrdmulh z18.s, z18.s, z20.s\n"
       ".inst 0x04b47673  // sqrdmulh z19.s, z19.s, z20.s\n"
       "tbz %x[flags], #5, 13f\n"
@@ -261,19 +261,19 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "sqadd z19.s, z19.s, z20.s\n"
       "13:"  // Height 1: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z20.s }, p2/Z, [x20]\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z22.s }, p2/Z, [x20]\n"
+      "add z16.s, z16.s, z20.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z20.s\n"
+      "add z18.s, z18.s, z20.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
       "ld1rw { z21.s }, p2/Z, [x20]\n"
-      "add z16.s, z16.s, z22.s\n"
+      "add z19.s, z19.s, z20.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z22.s\n"
-      "add z18.s, z18.s, z22.s\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z22.s\n"
       "smin z16.s, p2/M, z16.s, z21.s\n"
       "smin z17.s, p2/M, z17.s, z21.s\n"
       "smin z18.s, p2/M, z18.s, z21.s\n"
@@ -281,8 +281,8 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "smax z16.s, p2/M, z16.s, z20.s\n"
       "smax z17.s, p2/M, z17.s, z20.s\n"
       "smax z18.s, p2/M, z18.s, z20.s\n"
-      "smax z19.s, p2/M, z19.s, z20.s\n"
       "uzp1 z16.h, z16.h, z17.h\n"
+      "smax z19.s, p2/M, z19.s, z20.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
       "uzp1 z16.b, z16.b, z17.b\n"
       "st1b { z16.b }, p1, [x27]\n"
@@ -300,24 +300,24 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "mov z15.b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "16:"  // Height 2: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "17:"  // Height 2: setup done
       "mov x26, #0x0\n"
       "18:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 19f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -336,45 +336,45 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ble 23f\n"
       "21:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z25.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "udot z16.s, z24.b, z0.b[0]\n"
+      "udot z20.s, z24.b, z1.b[0]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "udot z17.s, z26.b, z0.b[0]\n"
+      "udot z21.s, z26.b, z1.b[0]\n"
+      "udot z18.s, z24.b, z0.b[0]\n"
+      "udot z22.s, z24.b, z1.b[0]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z27.b }, p2/Z, [x28, #5, MUL VL]\n"
+      "udot z19.s, z25.b, z0.b[0]\n"
+      "udot z23.s, z25.b, z1.b[0]\n"
       "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "udot z16.s, z25.b, z0.b[0]\n"
-      "udot z20.s, z25.b, z1.b[0]\n"
       "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      "udot z17.s, z30.b, z0.b[0]\n"
-      "udot z21.s, z30.b, z1.b[0]\n"
-      "udot z18.s, z29.b, z0.b[0]\n"
-      "udot z22.s, z29.b, z1.b[0]\n"
-      "udot z19.s, z28.b, z0.b[0]\n"
-      "udot z23.s, z28.b, z1.b[0]\n"
       "udot z16.s, z24.b, z0.b[1]\n"
       "udot z20.s, z24.b, z1.b[1]\n"
       "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
       "udot z17.s, z27.b, z0.b[1]\n"
       "udot z21.s, z27.b, z1.b[1]\n"
+      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
       "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
       "udot z18.s, z26.b, z0.b[1]\n"
       "udot z22.s, z26.b, z1.b[1]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
       "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "udot z19.s, z25.b, z0.b[1]\n"
       "udot z23.s, z25.b, z1.b[1]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
       "udot z16.s, z24.b, z0.b[2]\n"
       "udot z20.s, z24.b, z1.b[2]\n"
       "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
+      "add x23, x23, #0x10\n"
       "udot z17.s, z30.b, z0.b[2]\n"
       "udot z21.s, z30.b, z1.b[2]\n"
       "udot z18.s, z29.b, z0.b[2]\n"
@@ -398,34 +398,34 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "bgt 21b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z27.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1rqb { z0.b }, p0/Z, [x24]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "subs x25, x25, #0x4\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "udot z16.s, z24.b, z0.b[0]\n"
+      "udot z20.s, z24.b, z1.b[0]\n"
       "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1rqb { z1.b }, p0/Z, [x23]\n"
-      "udot z16.s, z27.b, z0.b[0]\n"
-      "udot z20.s, z27.b, z1.b[0]\n"
       "udot z17.s, z26.b, z0.b[0]\n"
       "udot z21.s, z26.b, z1.b[0]\n"
       "udot z18.s, z25.b, z0.b[0]\n"
       "udot z22.s, z25.b, z1.b[0]\n"
+      "addvl x28, x28, #4\n"
       "udot z19.s, z24.b, z0.b[0]\n"
       "udot z23.s, z24.b, z1.b[0]\n"
       "ble 24f\n"
       "ld1b { z27.b }, p2/Z, [x28]\n"
       "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "udot z16.s, z27.b, z0.b[1]\n"
       "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z27.b, z0.b[1]\n"
       "udot z20.s, z27.b, z1.b[1]\n"
       "udot z17.s, z26.b, z0.b[1]\n"
       "udot z21.s, z26.b, z1.b[1]\n"
       "udot z18.s, z25.b, z0.b[1]\n"
+      "addvl x28, x28, #4\n"
       "udot z22.s, z25.b, z1.b[1]\n"
       "udot z19.s, z24.b, z0.b[1]\n"
       "udot z23.s, z24.b, z1.b[1]\n"
@@ -433,29 +433,29 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ld1b { z27.b }, p2/Z, [x28]\n"
       "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "udot z16.s, z27.b, z0.b[2]\n"
       "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z27.b, z0.b[2]\n"
       "udot z20.s, z27.b, z1.b[2]\n"
       "udot z17.s, z26.b, z0.b[2]\n"
       "udot z21.s, z26.b, z1.b[2]\n"
       "udot z18.s, z25.b, z0.b[2]\n"
+      "addvl x28, x28, #4\n"
       "udot z22.s, z25.b, z1.b[2]\n"
       "udot z19.s, z24.b, z0.b[2]\n"
       "udot z23.s, z24.b, z1.b[2]\n"
       "ble 24f\n"
-      "ld1b { z27.b }, p2/Z, [x28]\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
       "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "udot z16.s, z24.b, z0.b[3]\n"
+      "udot z20.s, z24.b, z1.b[3]\n"
       "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z27.b, z0.b[3]\n"
-      "udot z20.s, z27.b, z1.b[3]\n"
       "udot z17.s, z26.b, z0.b[3]\n"
       "udot z21.s, z26.b, z1.b[3]\n"
       "udot z18.s, z25.b, z0.b[3]\n"
       "udot z22.s, z25.b, z1.b[3]\n"
+      "addvl x28, x28, #4\n"
       "udot z19.s, z24.b, z0.b[3]\n"
       "udot z23.s, z24.b, z1.b[3]\n"
       "24:"  // Height 2: Multiply loop: multiply skip
@@ -468,18 +468,18 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "cmp x26, x20\n"
       "bne 18b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
+      "add x23, x27, x20\n"
       "tbnz %x[flags], #31, 26f\n"
-      "mov x21, #0x4\n"
+      "mov x20, #0x4\n"
+      "whilelt p0.s, XZR, x20\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z24.s, p2/M, z24.s\n"
       "uaddv d11, p0, z11.s\n"
-      "uaddv d12, p0, z12.s\n"
       "mov z11.s, z11.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z24.s\n"
+      "uaddv d12, p0, z12.s\n"
+      "neg z24.s, p2/M, z24.s\n"
       "mov z12.s, z12.s[0]\n"
+      "mul z11.s, p2/M, z11.s, z24.s\n"
       "mul z12.s, p2/M, z12.s, z24.s\n"
       "26:"  // Height 2: skip row sum fixup
       "add z16.s, z16.s, z11.s\n"
@@ -518,24 +518,24 @@ void sve_hybrid_u8qa_dot_4x4VL (
       ".inst 0x04b876f7  // sqrdmulh z23.s, z23.s, z24.s\n"
       "tbz %x[flags], #5, 27f\n"
       "and z24.d, z16.d, z0.d\n"
+      "asr z24.s, z24.s, #0x1f\n"
+      "sqadd z16.s, z16.s, z24.s\n"
       "and z30.d, z17.d, z0.d\n"
       "and z29.d, z18.d, z0.d\n"
       "and z28.d, z19.d, z0.d\n"
       "and z27.d, z20.d, z0.d\n"
       "and z26.d, z21.d, z0.d\n"
-      "asr z24.s, z24.s, #0x1f\n"
       "and z25.d, z22.d, z0.d\n"
+      "and z24.d, z23.d, z0.d\n"
       "asr z30.s, z30.s, #0x1f\n"
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "asr z27.s, z27.s, #0x1f\n"
-      "sqadd z16.s, z16.s, z24.s\n"
-      "and z24.d, z23.d, z0.d\n"
       "asr z26.s, z26.s, #0x1f\n"
       "asr z25.s, z25.s, #0x1f\n"
+      "asr z24.s, z24.s, #0x1f\n"
       "sqadd z17.s, z17.s, z30.s\n"
       "sqadd z18.s, z18.s, z29.s\n"
-      "asr z24.s, z24.s, #0x1f\n"
       "sqadd z19.s, z19.s, z28.s\n"
       "sqadd z20.s, z20.s, z27.s\n"
       "sqadd z21.s, z21.s, z26.s\n"
@@ -543,27 +543,27 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "sqadd z23.s, z23.s, z24.s\n"
       "27:"  // Height 2: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z24.s }, p2/Z, [x20]\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z26.s }, p2/Z, [x20]\n"
+      "add z16.s, z16.s, z24.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z24.s\n"
+      "add z18.s, z18.s, z24.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
+      "add z19.s, z19.s, z24.s\n"
+      "add z20.s, z20.s, z24.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z26.s\n"
+      "add z21.s, z21.s, z24.s\n"
+      "add z22.s, z22.s, z24.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z17.s, z17.s, z26.s\n"
-      "add z18.s, z18.s, z26.s\n"
       "ld1rw { z25.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z26.s\n"
-      "add z20.s, z20.s, z26.s\n"
+      "add z23.s, z23.s, z24.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z21.s, z21.s, z26.s\n"
-      "add z22.s, z22.s, z26.s\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z26.s\n"
       "smin z16.s, p2/M, z16.s, z25.s\n"
       "smin z17.s, p2/M, z17.s, z25.s\n"
       "smin z18.s, p2/M, z18.s, z25.s\n"
@@ -575,20 +575,20 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "smax z16.s, p2/M, z16.s, z24.s\n"
       "smax z17.s, p2/M, z17.s, z24.s\n"
       "smax z18.s, p2/M, z18.s, z24.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z19.s, p2/M, z19.s, z24.s\n"
       "smax z20.s, p2/M, z20.s, z24.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z21.s, p2/M, z21.s, z24.s\n"
       "smax z22.s, p2/M, z22.s, z24.s\n"
-      "smax z23.s, p2/M, z23.s, z24.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
       "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z17.h, z22.h, z23.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z20.b, z20.b, z17.b\n"
       "st1b { z16.b }, p1, [x27]\n"
+      "smax z23.s, p2/M, z23.s, z24.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
+      "st1b { z20.b }, p1, [x23]\n"
       "addvl x27, x27, #1\n"
-      "st1b { z20.b }, p1, [x24]\n"
       "28:"  // Height 2: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -603,16 +603,16 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "mov z15.b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "30:"  // Height 3: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
@@ -623,8 +623,8 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "mov x26, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -646,57 +646,57 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ble 37f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z3.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z31.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x22]\n"
+      "ld1b { z28.b }, p2/Z, [x28]\n"
+      "udot z16.s, z28.b, z0.b[0]\n"
+      "udot z20.s, z28.b, z1.b[0]\n"
+      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "udot z24.s, z28.b, z2.b[0]\n"
+      "udot z17.s, z30.b, z0.b[0]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "udot z21.s, z30.b, z1.b[0]\n"
+      "udot z25.s, z30.b, z2.b[0]\n"
+      "ld1b { z3.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "udot z18.s, z29.b, z0.b[0]\n"
+      "udot z22.s, z29.b, z1.b[0]\n"
+      "ld1b { z31.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1b { z30.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x22, x22, #0x10\n"
-      "udot z16.s, z5.b, z0.b[0]\n"
-      "udot z20.s, z5.b, z1.b[0]\n"
-      "udot z17.s, z29.b, z0.b[0]\n"
-      "udot z21.s, z29.b, z1.b[0]\n"
-      "udot z18.s, z4.b, z0.b[0]\n"
-      "udot z24.s, z5.b, z2.b[0]\n"
-      "udot z25.s, z29.b, z2.b[0]\n"
+      "udot z26.s, z29.b, z2.b[0]\n"
+      "udot z19.s, z28.b, z0.b[0]\n"
       "ld1b { z29.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      "udot z22.s, z4.b, z1.b[0]\n"
-      "udot z26.s, z4.b, z2.b[0]\n"
-      "udot z19.s, z28.b, z0.b[0]\n"
       "udot z23.s, z28.b, z1.b[0]\n"
       "udot z27.s, z28.b, z2.b[0]\n"
-      "udot z16.s, z3.b, z0.b[1]\n"
       "ld1b { z28.b }, p2/Z, [x28, #-8, MUL VL]\n"
       "ld1b { z5.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      "udot z16.s, z3.b, z0.b[1]\n"
       "udot z20.s, z3.b, z1.b[1]\n"
-      "udot z24.s, z3.b, z2.b[1]\n"
       "ld1b { z4.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z3.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      "add x23, x23, #0x10\n"
+      "udot z24.s, z3.b, z2.b[1]\n"
       "udot z17.s, z31.b, z0.b[1]\n"
+      "ld1b { z3.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      "add x22, x22, #0x10\n"
       "udot z21.s, z31.b, z1.b[1]\n"
       "udot z25.s, z31.b, z2.b[1]\n"
-      "udot z18.s, z30.b, z0.b[1]\n"
       "ld1b { z31.b }, p2/Z, [x28, #-4, MUL VL]\n"
+      "udot z18.s, z30.b, z0.b[1]\n"
       "udot z22.s, z30.b, z1.b[1]\n"
       "udot z26.s, z30.b, z2.b[1]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "udot z19.s, z29.b, z0.b[1]\n"
+      "ld1b { z30.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "udot z23.s, z29.b, z1.b[1]\n"
       "udot z27.s, z29.b, z2.b[1]\n"
-      "udot z16.s, z28.b, z0.b[2]\n"
       "ld1b { z29.b }, p2/Z, [x28, #-2, MUL VL]\n"
+      "udot z16.s, z28.b, z0.b[2]\n"
       "udot z20.s, z28.b, z1.b[2]\n"
       "udot z24.s, z28.b, z2.b[2]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-1, MUL VL]\n"
       "udot z17.s, z5.b, z0.b[2]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #-1, MUL VL]\n"
       "udot z21.s, z5.b, z1.b[2]\n"
       "udot z25.s, z5.b, z2.b[2]\n"
       "udot z18.s, z4.b, z0.b[2]\n"
@@ -727,23 +727,23 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "bgt 35b\n"
       "37:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "subs x25, x25, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x22]\n"
-      "udot z16.s, z31.b, z0.b[0]\n"
-      "udot z20.s, z31.b, z1.b[0]\n"
+      "ld1b { z28.b }, p2/Z, [x28]\n"
+      "udot z16.s, z28.b, z0.b[0]\n"
+      "udot z20.s, z28.b, z1.b[0]\n"
+      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "udot z24.s, z28.b, z2.b[0]\n"
       "udot z17.s, z30.b, z0.b[0]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
       "udot z21.s, z30.b, z1.b[0]\n"
+      "udot z25.s, z30.b, z2.b[0]\n"
+      "addvl x28, x28, #4\n"
       "udot z18.s, z29.b, z0.b[0]\n"
       "udot z22.s, z29.b, z1.b[0]\n"
-      "udot z24.s, z31.b, z2.b[0]\n"
-      "udot z25.s, z30.b, z2.b[0]\n"
       "udot z26.s, z29.b, z2.b[0]\n"
       "udot z19.s, z28.b, z0.b[0]\n"
       "udot z23.s, z28.b, z1.b[0]\n"
@@ -752,14 +752,14 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ld1b { z31.b }, p2/Z, [x28]\n"
       "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "udot z16.s, z31.b, z0.b[1]\n"
       "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z31.b, z0.b[1]\n"
       "udot z20.s, z31.b, z1.b[1]\n"
       "udot z24.s, z31.b, z2.b[1]\n"
       "udot z17.s, z30.b, z0.b[1]\n"
       "udot z21.s, z30.b, z1.b[1]\n"
+      "addvl x28, x28, #4\n"
       "udot z25.s, z30.b, z2.b[1]\n"
       "udot z18.s, z29.b, z0.b[1]\n"
       "udot z22.s, z29.b, z1.b[1]\n"
@@ -771,14 +771,14 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ld1b { z31.b }, p2/Z, [x28]\n"
       "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "udot z16.s, z31.b, z0.b[2]\n"
       "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z31.b, z0.b[2]\n"
       "udot z20.s, z31.b, z1.b[2]\n"
       "udot z24.s, z31.b, z2.b[2]\n"
       "udot z17.s, z30.b, z0.b[2]\n"
       "udot z21.s, z30.b, z1.b[2]\n"
+      "addvl x28, x28, #4\n"
       "udot z25.s, z30.b, z2.b[2]\n"
       "udot z18.s, z29.b, z0.b[2]\n"
       "udot z22.s, z29.b, z1.b[2]\n"
@@ -789,15 +789,15 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ble 38f\n"
       "ld1b { z31.b }, p2/Z, [x28]\n"
       "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
       "udot z16.s, z31.b, z0.b[3]\n"
       "udot z20.s, z31.b, z1.b[3]\n"
+      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
       "udot z24.s, z31.b, z2.b[3]\n"
       "udot z17.s, z30.b, z0.b[3]\n"
       "udot z21.s, z30.b, z1.b[3]\n"
       "udot z25.s, z30.b, z2.b[3]\n"
+      "addvl x28, x28, #4\n"
       "udot z18.s, z29.b, z0.b[3]\n"
       "udot z22.s, z29.b, z1.b[3]\n"
       "udot z26.s, z29.b, z2.b[3]\n"
@@ -815,22 +815,22 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "cmp x26, x20\n"
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "tbnz %x[flags], #31, 40f\n"
-      "mov x21, #0x4\n"
+      "mov x20, #0x4\n"
+      "whilelt p0.s, XZR, x20\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z28.s, p2/M, z28.s\n"
       "uaddv d11, p0, z11.s\n"
+      "mov z11.s, z11.s[0]\n"
       "uaddv d12, p0, z12.s\n"
       "uaddv d13, p0, z13.s\n"
-      "mov z11.s, z11.s[0]\n"
       "mov z12.s, z12.s[0]\n"
+      "mov z13.s, z13.s[0]\n"
+      "neg z28.s, p2/M, z28.s\n"
       "mul z11.s, p2/M, z11.s, z28.s\n"
       "mul z12.s, p2/M, z12.s, z28.s\n"
-      "mov z13.s, z13.s[0]\n"
       "mul z13.s, p2/M, z13.s, z28.s\n"
       "40:"  // Height 3: skip row sum fixup
       "add z16.s, z16.s, z11.s\n"
@@ -885,18 +885,18 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "and z30.d, z18.d, z0.d\n"
       "and z29.d, z19.d, z0.d\n"
       "and z28.d, z20.d, z0.d\n"
-      "and z3.d, z21.d, z0.d\n"
       "asr z1.s, z1.s, #0x1f\n"
       "asr z31.s, z31.s, #0x1f\n"
       "asr z30.s, z30.s, #0x1f\n"
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
-      "and z2.d, z22.d, z0.d\n"
       "sqadd z16.s, z16.s, z1.s\n"
       "sqadd z17.s, z17.s, z31.s\n"
       "sqadd z18.s, z18.s, z30.s\n"
       "sqadd z19.s, z19.s, z29.s\n"
       "sqadd z20.s, z20.s, z28.s\n"
+      "and z3.d, z21.d, z0.d\n"
+      "and z2.d, z22.d, z0.d\n"
       "and z1.d, z23.d, z0.d\n"
       "and z31.d, z24.d, z0.d\n"
       "and z30.d, z25.d, z0.d\n"
@@ -918,35 +918,35 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "sqadd z27.s, z27.s, z28.s\n"
       "41:"  // Height 3: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z28.s }, p2/Z, [x20]\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z30.s }, p2/Z, [x20]\n"
+      "add z16.s, z16.s, z28.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z28.s\n"
+      "add z18.s, z18.s, z28.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
+      "add z19.s, z19.s, z28.s\n"
+      "add z20.s, z20.s, z28.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z30.s\n"
+      "add z21.s, z21.s, z28.s\n"
+      "add z22.s, z22.s, z28.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z30.s\n"
-      "add z18.s, z18.s, z30.s\n"
+      "add z23.s, z23.s, z28.s\n"
+      "add z24.s, z24.s, z28.s\n"
       ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
       ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z30.s\n"
-      "add z20.s, z20.s, z30.s\n"
+      "add z25.s, z25.s, z28.s\n"
+      "add z26.s, z26.s, z28.s\n"
       ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z21.s, z21.s, z30.s\n"
-      "add z22.s, z22.s, z30.s\n"
       "ld1rw { z29.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z30.s\n"
-      "add z24.s, z24.s, z30.s\n"
+      "add z27.s, z27.s, z28.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z30.s\n"
-      "add z26.s, z26.s, z30.s\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z30.s\n"
       "smin z16.s, p2/M, z16.s, z29.s\n"
       "smin z17.s, p2/M, z17.s, z29.s\n"
       "smin z18.s, p2/M, z18.s, z29.s\n"
@@ -962,28 +962,28 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "smax z16.s, p2/M, z16.s, z28.s\n"
       "smax z17.s, p2/M, z17.s, z28.s\n"
       "smax z18.s, p2/M, z18.s, z28.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z19.s, p2/M, z19.s, z28.s\n"
       "smax z20.s, p2/M, z20.s, z28.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z21.s, p2/M, z21.s, z28.s\n"
       "smax z22.s, p2/M, z22.s, z28.s\n"
+      "uzp1 z20.h, z20.h, z21.h\n"
+      "st1b { z16.b }, p1, [x27]\n"
       "smax z23.s, p2/M, z23.s, z28.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z24.s, p2/M, z24.s, z28.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
       "smax z25.s, p2/M, z25.s, z28.s\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
       "smax z26.s, p2/M, z26.s, z28.s\n"
-      "smax z27.s, p2/M, z27.s, z28.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z18.h, z22.h, z23.h\n"
       "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "uzp1 z20.b, z20.b, z18.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
+      "st1b { z20.b }, p1, [x23]\n"
+      "smax z27.s, p2/M, z27.s, z28.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
+      "st1b { z24.b }, p1, [x22]\n"
       "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "st1b { z20.b }, p1, [x24]\n"
-      "st1b { z24.b }, p1, [x23]\n"
       "42:"  // Height 3: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -991,28 +991,27 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "b 58f\n"
       "43:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x4\n"
       "mov x10, %x[col_bias]\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
+      "mov z13.s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov z13.s, #0x0\n"
       "mov z14.s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
       "mov z15.b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "44:"  // Height 4: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
@@ -1027,8 +1026,8 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "mov x26, #0x0\n"
       "46:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 47f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1053,37 +1052,37 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ble 51f\n"
       "49:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
       "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x22]\n"
       "ld1rqb { z3.b }, p0/Z, [x21]\n"
+      "add x23, x23, #0x10\n"
       "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
+      "ld1b { z5.b }, p2/Z, [x28]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
       "udot z16.s, z5.b, z0.b[0]\n"
       "udot z20.s, z5.b, z1.b[0]\n"
-      "udot z17.s, z10.b, z0.b[0]\n"
-      "udot z21.s, z10.b, z1.b[0]\n"
+      "ld1b { z10.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
       "udot z24.s, z5.b, z2.b[0]\n"
       "udot z28.s, z5.b, z3.b[0]\n"
+      "udot z17.s, z4.b, z0.b[0]\n"
+      "udot z21.s, z4.b, z1.b[0]\n"
+      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
+      "udot z25.s, z4.b, z2.b[0]\n"
+      "udot z29.s, z4.b, z3.b[0]\n"
+      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
+      "udot z18.s, z10.b, z0.b[0]\n"
+      "udot z22.s, z10.b, z1.b[0]\n"
       "addvl x28, x28, #16\n"
-      "udot z25.s, z10.b, z2.b[0]\n"
-      "udot z29.s, z10.b, z3.b[0]\n"
-      "udot z18.s, z4.b, z0.b[0]\n"
-      "udot z22.s, z4.b, z1.b[0]\n"
-      "udot z26.s, z4.b, z2.b[0]\n"
-      "udot z30.s, z4.b, z3.b[0]\n"
       "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
+      "udot z26.s, z10.b, z2.b[0]\n"
+      "udot z30.s, z10.b, z3.b[0]\n"
       "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      "add x21, x21, #0x10\n"
       "udot z19.s, z9.b, z0.b[0]\n"
       "udot z23.s, z9.b, z1.b[0]\n"
       "udot z27.s, z9.b, z2.b[0]\n"
@@ -1153,26 +1152,26 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "bgt 49b\n"
       "51:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z7.b }, p2/Z, [x28]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "subs x25, x25, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x22]\n"
       "ld1rqb { z3.b }, p0/Z, [x21]\n"
+      "ld1b { z7.b }, p2/Z, [x28]\n"
+      "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
       "udot z16.s, z7.b, z0.b[0]\n"
       "udot z20.s, z7.b, z1.b[0]\n"
-      "udot z17.s, z6.b, z0.b[0]\n"
-      "udot z21.s, z6.b, z1.b[0]\n"
-      "udot z18.s, z5.b, z0.b[0]\n"
-      "udot z22.s, z5.b, z1.b[0]\n"
+      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
       "udot z24.s, z7.b, z2.b[0]\n"
       "udot z28.s, z7.b, z3.b[0]\n"
+      "udot z17.s, z6.b, z0.b[0]\n"
+      "udot z21.s, z6.b, z1.b[0]\n"
+      "addvl x28, x28, #4\n"
       "udot z25.s, z6.b, z2.b[0]\n"
       "udot z29.s, z6.b, z3.b[0]\n"
+      "udot z18.s, z5.b, z0.b[0]\n"
+      "udot z22.s, z5.b, z1.b[0]\n"
       "udot z26.s, z5.b, z2.b[0]\n"
       "udot z30.s, z5.b, z3.b[0]\n"
       "udot z19.s, z4.b, z0.b[0]\n"
@@ -1183,14 +1182,14 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ld1b { z7.b }, p2/Z, [x28]\n"
       "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "udot z16.s, z7.b, z0.b[1]\n"
       "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z7.b, z0.b[1]\n"
       "udot z20.s, z7.b, z1.b[1]\n"
       "udot z24.s, z7.b, z2.b[1]\n"
       "udot z28.s, z7.b, z3.b[1]\n"
       "udot z17.s, z6.b, z0.b[1]\n"
+      "addvl x28, x28, #4\n"
       "udot z21.s, z6.b, z1.b[1]\n"
       "udot z25.s, z6.b, z2.b[1]\n"
       "udot z29.s, z6.b, z3.b[1]\n"
@@ -1206,14 +1205,14 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ld1b { z7.b }, p2/Z, [x28]\n"
       "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x4\n"
+      "udot z16.s, z7.b, z0.b[2]\n"
       "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "udot z16.s, z7.b, z0.b[2]\n"
       "udot z20.s, z7.b, z1.b[2]\n"
       "udot z24.s, z7.b, z2.b[2]\n"
       "udot z28.s, z7.b, z3.b[2]\n"
       "udot z17.s, z6.b, z0.b[2]\n"
+      "addvl x28, x28, #4\n"
       "udot z21.s, z6.b, z1.b[2]\n"
       "udot z25.s, z6.b, z2.b[2]\n"
       "udot z29.s, z6.b, z3.b[2]\n"
@@ -1228,15 +1227,15 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "ble 52f\n"
       "ld1b { z7.b }, p2/Z, [x28]\n"
       "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
       "udot z16.s, z7.b, z0.b[3]\n"
       "udot z20.s, z7.b, z1.b[3]\n"
+      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
       "udot z24.s, z7.b, z2.b[3]\n"
       "udot z28.s, z7.b, z3.b[3]\n"
       "udot z17.s, z6.b, z0.b[3]\n"
       "udot z21.s, z6.b, z1.b[3]\n"
+      "addvl x28, x28, #4\n"
       "udot z25.s, z6.b, z2.b[3]\n"
       "udot z29.s, z6.b, z3.b[3]\n"
       "udot z18.s, z5.b, z0.b[3]\n"
@@ -1259,25 +1258,25 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "cmp x26, x20\n"
       "bne 46b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
+      "add x23, x27, x20\n"
       "add x22, x23, x20\n"
+      "add x21, x22, x20\n"
       "tbnz %x[flags], #31, 54f\n"
-      "mov x21, #0x4\n"
+      "mov x20, #0x4\n"
+      "whilelt p0.s, XZR, x20\n"
       "add x20, %x[qp], %[b_offset]\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z0.s, p2/M, z0.s\n"
       "uaddv d11, p0, z11.s\n"
+      "mov z11.s, z11.s[0]\n"
       "uaddv d12, p0, z12.s\n"
       "uaddv d13, p0, z13.s\n"
-      "uaddv d14, p0, z14.s\n"
-      "mov z11.s, z11.s[0]\n"
       "mov z12.s, z12.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z0.s\n"
-      "mul z12.s, p2/M, z12.s, z0.s\n"
       "mov z13.s, z13.s[0]\n"
+      "uaddv d14, p0, z14.s\n"
+      "neg z0.s, p2/M, z0.s\n"
       "mov z14.s, z14.s[0]\n"
+      "mul z11.s, p2/M, z11.s, z0.s\n"
+      "mul z12.s, p2/M, z12.s, z0.s\n"
       "mul z13.s, p2/M, z13.s, z0.s\n"
       "mul z14.s, p2/M, z14.s, z0.s\n"
       "54:"  // Height 4: skip row sum fixup
@@ -1342,32 +1341,32 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "tbz %x[flags], #5, 55f\n"
       "and z2.d, z16.d, z0.d\n"
       "and z1.d, z17.d, z0.d\n"
+      "asr z2.s, z2.s, #0x1f\n"
+      "asr z1.s, z1.s, #0x1f\n"
+      "sqadd z16.s, z16.s, z2.s\n"
+      "sqadd z17.s, z17.s, z1.s\n"
       "and z7.d, z18.d, z0.d\n"
       "and z6.d, z19.d, z0.d\n"
       "and z5.d, z20.d, z0.d\n"
       "and z4.d, z21.d, z0.d\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
       "and z3.d, z22.d, z0.d\n"
+      "and z2.d, z23.d, z0.d\n"
+      "and z1.d, z24.d, z0.d\n"
       "asr z7.s, z7.s, #0x1f\n"
       "asr z6.s, z6.s, #0x1f\n"
       "asr z5.s, z5.s, #0x1f\n"
-      "sqadd z16.s, z16.s, z2.s\n"
-      "sqadd z17.s, z17.s, z1.s\n"
-      "and z2.d, z23.d, z0.d\n"
-      "and z1.d, z24.d, z0.d\n"
       "asr z4.s, z4.s, #0x1f\n"
       "asr z3.s, z3.s, #0x1f\n"
-      "sqadd z18.s, z18.s, z7.s\n"
-      "sqadd z19.s, z19.s, z6.s\n"
       "asr z2.s, z2.s, #0x1f\n"
       "asr z1.s, z1.s, #0x1f\n"
+      "sqadd z18.s, z18.s, z7.s\n"
+      "sqadd z19.s, z19.s, z6.s\n"
       "sqadd z20.s, z20.s, z5.s\n"
       "sqadd z21.s, z21.s, z4.s\n"
       "sqadd z22.s, z22.s, z3.s\n"
-      "and z7.d, z25.d, z0.d\n"
       "sqadd z23.s, z23.s, z2.s\n"
       "sqadd z24.s, z24.s, z1.s\n"
+      "and z7.d, z25.d, z0.d\n"
       "and z6.d, z26.d, z0.d\n"
       "and z5.d, z27.d, z0.d\n"
       "and z4.d, z28.d, z0.d\n"
@@ -1390,43 +1389,43 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "sqadd z31.s, z31.s, z1.s\n"
       "55:"  // Height 4: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
       "ld1rw { z2.s }, p2/Z, [x20]\n"
+      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
+      "add z16.s, z16.s, z2.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z2.s\n"
+      "add z18.s, z18.s, z2.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
+      "add z19.s, z19.s, z2.s\n"
+      "add z20.s, z20.s, z2.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z2.s\n"
+      "add z21.s, z21.s, z2.s\n"
+      "add z22.s, z22.s, z2.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z2.s\n"
-      "add z18.s, z18.s, z2.s\n"
+      "add z23.s, z23.s, z2.s\n"
+      "add z24.s, z24.s, z2.s\n"
       ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
       ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z2.s\n"
-      "add z20.s, z20.s, z2.s\n"
+      "add z25.s, z25.s, z2.s\n"
+      "add z26.s, z26.s, z2.s\n"
       ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
       ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
-      "add z21.s, z21.s, z2.s\n"
-      "add z22.s, z22.s, z2.s\n"
+      "add z27.s, z27.s, z2.s\n"
+      "add z28.s, z28.s, z2.s\n"
       ".inst 0x4482881d  // srshl z29.s, p2/M, z29.s, z0.s\n"
       ".inst 0x4482881e  // srshl z30.s, p2/M, z30.s, z0.s\n"
-      "add z23.s, z23.s, z2.s\n"
-      "add z24.s, z24.s, z2.s\n"
+      "add z29.s, z29.s, z2.s\n"
+      "add z30.s, z30.s, z2.s\n"
       ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z25.s, z25.s, z2.s\n"
-      "add z26.s, z26.s, z2.s\n"
       "ld1rw { z1.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z2.s\n"
-      "add z28.s, z28.s, z2.s\n"
+      "add z31.s, z31.s, z2.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z29.s, z29.s, z2.s\n"
-      "add z30.s, z30.s, z2.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z31.s, z31.s, z2.s\n"
       "smin z16.s, p2/M, z16.s, z1.s\n"
       "smin z17.s, p2/M, z17.s, z1.s\n"
       "smin z18.s, p2/M, z18.s, z1.s\n"
@@ -1446,36 +1445,36 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "smax z16.s, p2/M, z16.s, z0.s\n"
       "smax z17.s, p2/M, z17.s, z0.s\n"
       "smax z18.s, p2/M, z18.s, z0.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z19.s, p2/M, z19.s, z0.s\n"
       "smax z20.s, p2/M, z20.s, z0.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z21.s, p2/M, z21.s, z0.s\n"
       "smax z22.s, p2/M, z22.s, z0.s\n"
+      "uzp1 z20.h, z20.h, z21.h\n"
+      "st1b { z16.b }, p1, [x27]\n"
       "smax z23.s, p2/M, z23.s, z0.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z24.s, p2/M, z24.s, z0.s\n"
+      "uzp1 z16.h, z22.h, z23.h\n"
+      "uzp1 z20.b, z20.b, z16.b\n"
       "smax z25.s, p2/M, z25.s, z0.s\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
       "smax z26.s, p2/M, z26.s, z0.s\n"
+      "uzp1 z24.h, z24.h, z25.h\n"
+      "st1b { z20.b }, p1, [x23]\n"
       "smax z27.s, p2/M, z27.s, z0.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
       "smax z28.s, p2/M, z28.s, z0.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "smax z29.s, p2/M, z29.s, z0.s\n"
-      "uzp1 z17.h, z22.h, z23.h\n"
       "smax z30.s, p2/M, z30.s, z0.s\n"
-      "smax z31.s, p2/M, z31.s, z0.s\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z18.h, z26.h, z27.h\n"
       "uzp1 z28.h, z28.h, z29.h\n"
-      "uzp1 z20.b, z20.b, z17.b\n"
-      "uzp1 z17.h, z30.h, z31.h\n"
-      "st1b { z16.b }, p1, [x27]\n"
+      "st1b { z24.b }, p1, [x22]\n"
+      "smax z31.s, p2/M, z31.s, z0.s\n"
+      "uzp1 z16.h, z30.h, z31.h\n"
+      "uzp1 z28.b, z28.b, z16.b\n"
+      "st1b { z28.b }, p1, [x21]\n"
       "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z18.b\n"
-      "uzp1 z28.b, z28.b, z17.b\n"
-      "st1b { z20.b }, p1, [x24]\n"
-      "st1b { z24.b }, p1, [x23]\n"
-      "st1b { z28.b }, p1, [x22]\n"
       "56:"  // Height 4: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -1492,8 +1491,8 @@ void sve_hybrid_u8qa_dot_4x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "58:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL.hpp
index ab37e6ad5b..e37ade4a00 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL.hpp
@@ -70,7 +70,7 @@ public:
         return false;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 4, 8, 8> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 4, 8, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp
index 212e178065..7b67ccd545 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp
@@ -45,18 +45,18 @@ void sve_hybrid_u8qa_mmla_4x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -91,24 +91,24 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "3:"  // Height 1: setup done
       "mov x26, #0x0\n"
       "4:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 5f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -124,43 +124,43 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "ble 9f\n"
       "7:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z30.b }, p2/Z, [x28]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x24, x24, #0x10\n"
-      "trn1 z0.d, z1.d, z31.d\n"
-      ".inst 0x45de9810  // ummla z16.s, z0.b, z30.b\n"
-      "ld1b { z8.b }, p2/Z, [x28, #7, MUL VL]\n"
+      "trn1 z0.d, z1.d, z26.d\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "trn2 z1.d, z1.d, z26.d\n"
+      "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45d99814  // ummla z20.s, z0.b, z25.b\n"
+      ".inst 0x45d89811  // ummla z17.s, z0.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45da9815  // ummla z21.s, z0.b, z26.b\n"
+      ".inst 0x45d99812  // ummla z18.s, z0.b, z25.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      "trn2 z1.d, z1.d, z31.d\n"
-      ".inst 0x45dd9814  // ummla z20.s, z0.b, z29.b\n"
-      ".inst 0x45dc9811  // ummla z17.s, z0.b, z28.b\n"
-      ".inst 0x45db9815  // ummla z21.s, z0.b, z27.b\n"
-      ".inst 0x45da9812  // ummla z18.s, z0.b, z26.b\n"
-      "ld1b { z31.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x45d99816  // ummla z22.s, z0.b, z25.b\n"
-      ".inst 0x45d89813  // ummla z19.s, z0.b, z24.b\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x45c89817  // ummla z23.s, z0.b, z8.b\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x45d89816  // ummla z22.s, z0.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
+      ".inst 0x45da9813  // ummla z19.s, z0.b, z26.b\n"
+      ".inst 0x45d99817  // ummla z23.s, z0.b, z25.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      ".inst 0x45d89830  // ummla z16.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x45da9834  // ummla z20.s, z1.b, z26.b\n"
       "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x45df9830  // ummla z16.s, z1.b, z31.b\n"
       "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
+      ".inst 0x45d99831  // ummla z17.s, z1.b, z25.b\n"
+      ".inst 0x45d89835  // ummla z21.s, z1.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x45de9834  // ummla z20.s, z1.b, z30.b\n"
-      ".inst 0x45dd9831  // ummla z17.s, z1.b, z29.b\n"
-      ".inst 0x45dc9835  // ummla z21.s, z1.b, z28.b\n"
       ".inst 0x45db9832  // ummla z18.s, z1.b, z27.b\n"
       ".inst 0x45da9836  // ummla z22.s, z1.b, z26.b\n"
       ".inst 0x45d99833  // ummla z19.s, z1.b, z25.b\n"
       ".inst 0x45d89837  // ummla z23.s, z1.b, z24.b\n"
+      "add x24, x24, #0x10\n"
       "tbnz %x[flags], #31, 8f\n"
       "udot z11.s, z0.b, z15.b\n"
       "udot z11.s, z1.b, z15.b\n"
@@ -170,45 +170,45 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "bgt 7b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
+      "ld1rqb { z1.b }, p0/Z, [x24]\n"
+      "trn1 z0.d, z1.d, z27.d\n"
       "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
       "subs x25, x25, #0x8\n"
+      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "trn2 z1.d, z1.d, z27.d\n"
+      ".inst 0x45da9814  // ummla z20.s, z0.b, z26.b\n"
       "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
       "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
+      ".inst 0x45d99811  // ummla z17.s, z0.b, z25.b\n"
+      ".inst 0x45d89815  // ummla z21.s, z0.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z31.d\n"
-      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
       "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #8\n"
-      "trn2 z1.d, z1.d, z31.d\n"
-      ".inst 0x45de9814  // ummla z20.s, z0.b, z30.b\n"
-      ".inst 0x45dd9811  // ummla z17.s, z0.b, z29.b\n"
-      ".inst 0x45dc9815  // ummla z21.s, z0.b, z28.b\n"
       ".inst 0x45db9812  // ummla z18.s, z0.b, z27.b\n"
       ".inst 0x45da9816  // ummla z22.s, z0.b, z26.b\n"
       ".inst 0x45d99813  // ummla z19.s, z0.b, z25.b\n"
       ".inst 0x45d89817  // ummla z23.s, z0.b, z24.b\n"
+      "addvl x28, x28, #8\n"
       "ble 10f\n"
       "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
       ".inst 0x45d89830  // ummla z16.s, z1.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x45d89834  // ummla z20.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45d99831  // ummla z17.s, z1.b, z25.b\n"
+      ".inst 0x45d89835  // ummla z21.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45d99832  // ummla z18.s, z1.b, z25.b\n"
+      ".inst 0x45d89836  // ummla z22.s, z1.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x45de9834  // ummla z20.s, z1.b, z30.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x45dd9831  // ummla z17.s, z1.b, z29.b\n"
-      ".inst 0x45dc9835  // ummla z21.s, z1.b, z28.b\n"
-      ".inst 0x45db9832  // ummla z18.s, z1.b, z27.b\n"
-      ".inst 0x45da9836  // ummla z22.s, z1.b, z26.b\n"
       ".inst 0x45d99833  // ummla z19.s, z1.b, z25.b\n"
       ".inst 0x45d89837  // ummla z23.s, z1.b, z24.b\n"
+      "addvl x28, x28, #8\n"
       "10:"  // Height 1: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 11f\n"
       "udot z11.s, z0.b, z15.b\n"
@@ -225,32 +225,32 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "mov z23.d, z16.d\n"
       "tbnz %x[flags], #31, 12f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1rw { z16.s }, p2/Z, [x20]\n"
       ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
-      "ld1rw { z9.s }, p2/Z, [x20]\n"
-      "neg z9.s, p2/M, z9.s\n"
+      "neg z16.s, p2/M, z16.s\n"
       "mov z11.s, z11.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z9.s\n"
+      "mul z11.s, p2/M, z11.s, z16.s\n"
       "12:"  // Height 1: skip row sum fixup
       "add z23.s, z23.s, z11.s\n"
       "add z17.s, z17.s, z11.s\n"
       "ld1w { z22.s }, p2/Z, [x10]\n"
-      "ld1w { z24.s }, p2/Z, [x10, #1, MUL VL]\n"
+      "ld1w { z21.s }, p2/Z, [x10, #1, MUL VL]\n"
       "add z18.s, z18.s, z11.s\n"
       "add z19.s, z19.s, z11.s\n"
-      "ld1w { z21.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x10, #3, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x10, #3, MUL VL]\n"
       "add x20, %x[qp], %[per_layer_mul]\n"
       "orr %x[flags], %x[flags], #0x80000000\n"
       "add z23.s, z23.s, z22.s\n"
-      "add z17.s, z17.s, z24.s\n"
+      "add z17.s, z17.s, z21.s\n"
+      "add z18.s, z18.s, z20.s\n"
+      "add z19.s, z19.s, z16.s\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
       "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "addvl x10, x10, #4\n"
-      "add z18.s, z18.s, z21.s\n"
-      "add z19.s, z19.s, z20.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
       ".inst 0x04b076f7  // sqrdmulh z23.s, z23.s, z16.s\n"
       ".inst 0x04b07631  // sqrdmulh z17.s, z17.s, z16.s\n"
+      "addvl x10, x10, #4\n"
       ".inst 0x04b07652  // sqrdmulh z18.s, z18.s, z16.s\n"
       ".inst 0x04b07673  // sqrdmulh z19.s, z19.s, z16.s\n"
       "tbz %x[flags], #5, 13f\n"
@@ -268,19 +268,19 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "sqadd z19.s, z19.s, z16.s\n"
       "13:"  // Height 1: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z16.s }, p2/Z, [x20]\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "ld1rw { z21.s }, p2/Z, [x20]\n"
+      "add z23.s, z23.s, z16.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
+      "add z17.s, z17.s, z16.s\n"
+      "add z18.s, z18.s, z16.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
       "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z21.s\n"
+      "add z19.s, z19.s, z16.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z21.s\n"
-      "add z18.s, z18.s, z21.s\n"
       "ld1rw { z16.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z21.s\n"
       "smin z23.s, p2/M, z23.s, z20.s\n"
       "smin z17.s, p2/M, z17.s, z20.s\n"
       "smin z18.s, p2/M, z18.s, z20.s\n"
@@ -288,8 +288,8 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "smax z23.s, p2/M, z23.s, z16.s\n"
       "smax z17.s, p2/M, z17.s, z16.s\n"
       "smax z18.s, p2/M, z18.s, z16.s\n"
-      "smax z19.s, p2/M, z19.s, z16.s\n"
       "uzp1 z23.h, z23.h, z17.h\n"
+      "smax z19.s, p2/M, z19.s, z16.s\n"
       "uzp1 z16.h, z18.h, z19.h\n"
       "uzp1 z23.b, z23.b, z16.b\n"
       "st1b { z23.b }, p1, [x27]\n"
@@ -307,24 +307,24 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "mov z15.b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "16:"  // Height 2: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "17:"  // Height 2: setup done
       "mov x26, #0x0\n"
       "18:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 19f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -343,45 +343,45 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "ble 23f\n"
       "21:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z25.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1b { z24.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      ".inst 0x45df9810  // ummla z16.s, z0.b, z31.b\n"
+      "ld1rqb { z26.b }, p0/Z, [x23]\n"
+      "trn1 z0.d, z1.d, z26.d\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "trn2 z1.d, z1.d, z26.d\n"
+      "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45d99814  // ummla z20.s, z0.b, z25.b\n"
+      ".inst 0x45d89811  // ummla z17.s, z0.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45da9815  // ummla z21.s, z0.b, z26.b\n"
+      ".inst 0x45d99812  // ummla z18.s, z0.b, z25.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x45de9814  // ummla z20.s, z0.b, z30.b\n"
-      ".inst 0x45dd9811  // ummla z17.s, z0.b, z29.b\n"
-      ".inst 0x45dc9815  // ummla z21.s, z0.b, z28.b\n"
-      ".inst 0x45db9812  // ummla z18.s, z0.b, z27.b\n"
-      ".inst 0x45da9816  // ummla z22.s, z0.b, z26.b\n"
-      ".inst 0x45d89813  // ummla z19.s, z0.b, z24.b\n"
+      ".inst 0x45d89816  // ummla z22.s, z0.b, z24.b\n"
       "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
+      ".inst 0x45da9813  // ummla z19.s, z0.b, z26.b\n"
       ".inst 0x45d99817  // ummla z23.s, z0.b, z25.b\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
       ".inst 0x45d89830  // ummla z16.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x45da9834  // ummla z20.s, z1.b, z26.b\n"
+      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
       "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
+      ".inst 0x45d99831  // ummla z17.s, z1.b, z25.b\n"
+      ".inst 0x45d89835  // ummla z21.s, z1.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x45de9834  // ummla z20.s, z1.b, z30.b\n"
-      ".inst 0x45dd9831  // ummla z17.s, z1.b, z29.b\n"
-      ".inst 0x45dc9835  // ummla z21.s, z1.b, z28.b\n"
       ".inst 0x45db9832  // ummla z18.s, z1.b, z27.b\n"
       ".inst 0x45da9836  // ummla z22.s, z1.b, z26.b\n"
       ".inst 0x45d99833  // ummla z19.s, z1.b, z25.b\n"
       ".inst 0x45d89837  // ummla z23.s, z1.b, z24.b\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "tbnz %x[flags], #31, 22f\n"
       "udot z11.s, z0.b, z15.b\n"
       "udot z11.s, z1.b, z15.b\n"
@@ -391,46 +391,46 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "bgt 21b\n"
       "23:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z29.b }, p2/Z, [x28]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
+      "ld1rqb { z27.b }, p0/Z, [x23]\n"
+      "trn1 z0.d, z1.d, z27.d\n"
+      "ld1b { z24.b }, p2/Z, [x28]\n"
+      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
+      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "subs x25, x25, #0x8\n"
+      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "trn2 z1.d, z1.d, z27.d\n"
+      ".inst 0x45da9814  // ummla z20.s, z0.b, z26.b\n"
+      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45d99811  // ummla z17.s, z0.b, z25.b\n"
+      ".inst 0x45d89815  // ummla z21.s, z0.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z24.d\n"
-      "trn2 z1.d, z1.d, z24.d\n"
-      ".inst 0x45dd9810  // ummla z16.s, z0.b, z29.b\n"
       "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x45dc9814  // ummla z20.s, z0.b, z28.b\n"
-      ".inst 0x45c49811  // ummla z17.s, z0.b, z4.b\n"
-      ".inst 0x45db9815  // ummla z21.s, z0.b, z27.b\n"
-      ".inst 0x45da9812  // ummla z18.s, z0.b, z26.b\n"
-      ".inst 0x45c69816  // ummla z22.s, z0.b, z6.b\n"
+      ".inst 0x45db9812  // ummla z18.s, z0.b, z27.b\n"
+      ".inst 0x45da9816  // ummla z22.s, z0.b, z26.b\n"
       ".inst 0x45d99813  // ummla z19.s, z0.b, z25.b\n"
       ".inst 0x45d89817  // ummla z23.s, z0.b, z24.b\n"
+      "addvl x28, x28, #8\n"
       "ble 24f\n"
       "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
       ".inst 0x45d89830  // ummla z16.s, z1.b, z24.b\n"
+      "ld1b { z24.b }, p2/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x45d89834  // ummla z20.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45d99831  // ummla z17.s, z1.b, z25.b\n"
+      ".inst 0x45d89835  // ummla z21.s, z1.b, z24.b\n"
+      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45d99832  // ummla z18.s, z1.b, z25.b\n"
+      ".inst 0x45d89836  // ummla z22.s, z1.b, z24.b\n"
       "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x45de9834  // ummla z20.s, z1.b, z30.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x45dd9831  // ummla z17.s, z1.b, z29.b\n"
-      ".inst 0x45dc9835  // ummla z21.s, z1.b, z28.b\n"
-      ".inst 0x45db9832  // ummla z18.s, z1.b, z27.b\n"
-      ".inst 0x45da9836  // ummla z22.s, z1.b, z26.b\n"
       ".inst 0x45d99833  // ummla z19.s, z1.b, z25.b\n"
       ".inst 0x45d89837  // ummla z23.s, z1.b, z24.b\n"
+      "addvl x28, x28, #8\n"
       "24:"  // Height 2: Multiply loop: multiply skip
       "tbnz %x[flags], #31, 25f\n"
       "udot z11.s, z0.b, z15.b\n"
@@ -443,18 +443,18 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "uzp1 z24.d, z16.d, z20.d\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp2 z16.d, z16.d, z20.d\n"
+      "add x23, x27, x20\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "add x23, x27, x20\n"
       "mov z23.d, z24.d\n"
       "tbnz %x[flags], #31, 26f\n"
       "add x20, %x[qp], %[b_offset]\n"
-      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
+      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
       "neg z24.s, p2/M, z24.s\n"
       "mov z12.s, z11.s[3]\n"
       "mov z11.s, z11.s[0]\n"
@@ -497,24 +497,24 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       ".inst 0x04b87673  // sqrdmulh z19.s, z19.s, z24.s\n"
       "tbz %x[flags], #5, 27f\n"
       "and z24.d, z23.d, z0.d\n"
+      "asr z24.s, z24.s, #0x1f\n"
+      "sqadd z23.s, z23.s, z24.s\n"
       "and z30.d, z20.d, z0.d\n"
       "and z29.d, z21.d, z0.d\n"
       "and z28.d, z22.d, z0.d\n"
       "and z27.d, z16.d, z0.d\n"
       "and z26.d, z17.d, z0.d\n"
-      "asr z24.s, z24.s, #0x1f\n"
       "and z25.d, z18.d, z0.d\n"
+      "and z24.d, z19.d, z0.d\n"
       "asr z30.s, z30.s, #0x1f\n"
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "asr z27.s, z27.s, #0x1f\n"
-      "sqadd z23.s, z23.s, z24.s\n"
-      "and z24.d, z19.d, z0.d\n"
       "asr z26.s, z26.s, #0x1f\n"
       "asr z25.s, z25.s, #0x1f\n"
+      "asr z24.s, z24.s, #0x1f\n"
       "sqadd z20.s, z20.s, z30.s\n"
       "sqadd z21.s, z21.s, z29.s\n"
-      "asr z24.s, z24.s, #0x1f\n"
       "sqadd z22.s, z22.s, z28.s\n"
       "sqadd z16.s, z16.s, z27.s\n"
       "sqadd z17.s, z17.s, z26.s\n"
@@ -522,27 +522,27 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "sqadd z19.s, z19.s, z24.s\n"
       "27:"  // Height 2: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z24.s }, p2/Z, [x20]\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "ld1rw { z26.s }, p2/Z, [x20]\n"
+      "add z23.s, z23.s, z24.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
+      "add z20.s, z20.s, z24.s\n"
+      "add z21.s, z21.s, z24.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
+      "add z22.s, z22.s, z24.s\n"
+      "add z16.s, z16.s, z24.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z23.s, z23.s, z26.s\n"
+      "add z17.s, z17.s, z24.s\n"
+      "add z18.s, z18.s, z24.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z20.s, z20.s, z26.s\n"
-      "add z21.s, z21.s, z26.s\n"
       "ld1rw { z25.s }, p2/Z, [x20]\n"
-      "add z22.s, z22.s, z26.s\n"
-      "add z16.s, z16.s, z26.s\n"
+      "add z19.s, z19.s, z24.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z26.s\n"
-      "add z18.s, z18.s, z26.s\n"
       "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z26.s\n"
       "smin z23.s, p2/M, z23.s, z25.s\n"
       "smin z20.s, p2/M, z20.s, z25.s\n"
       "smin z21.s, p2/M, z21.s, z25.s\n"
@@ -554,20 +554,20 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "smax z23.s, p2/M, z23.s, z24.s\n"
       "smax z20.s, p2/M, z20.s, z24.s\n"
       "smax z21.s, p2/M, z21.s, z24.s\n"
+      "uzp1 z23.h, z23.h, z20.h\n"
       "smax z22.s, p2/M, z22.s, z24.s\n"
       "smax z16.s, p2/M, z16.s, z24.s\n"
+      "uzp1 z20.h, z21.h, z22.h\n"
+      "uzp1 z23.b, z23.b, z20.b\n"
       "smax z17.s, p2/M, z17.s, z24.s\n"
       "smax z18.s, p2/M, z18.s, z24.s\n"
-      "smax z19.s, p2/M, z19.s, z24.s\n"
-      "uzp1 z23.h, z23.h, z20.h\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
       "uzp1 z16.h, z16.h, z17.h\n"
+      "st1b { z23.b }, p1, [x27]\n"
+      "smax z19.s, p2/M, z19.s, z24.s\n"
       "uzp1 z17.h, z18.h, z19.h\n"
-      "uzp1 z23.b, z23.b, z20.b\n"
       "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z23.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
       "st1b { z16.b }, p1, [x23]\n"
+      "addvl x27, x27, #1\n"
       "28:"  // Height 2: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -582,16 +582,16 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "mov z15.b, #0x1\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
       "30:"  // Height 3: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
@@ -606,8 +606,8 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "mov x26, #0x0\n"
       "32:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 33f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -629,49 +629,49 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "ble 37f\n"
       "35:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
-      "add x22, x22, #0x10\n"
       "trn1 z0.d, z1.d, z2.d\n"
       "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z6.d\n"
-      "trn2 z3.d, z3.d, z6.d\n"
+      "ld1b { z4.b }, p2/Z, [x28]\n"
+      "trn1 z2.d, z3.d, z5.d\n"
+      "trn2 z3.d, z3.d, z5.d\n"
+      ".inst 0x45c49810  // ummla z16.s, z0.b, z4.b\n"
+      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x45c49858  // ummla z24.s, z2.b, z4.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45c59814  // ummla z20.s, z0.b, z5.b\n"
+      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45c5985c  // ummla z28.s, z2.b, z5.b\n"
+      ".inst 0x45c49811  // ummla z17.s, z0.b, z4.b\n"
       "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45c59810  // ummla z16.s, z0.b, z5.b\n"
-      ".inst 0x45ca9814  // ummla z20.s, z0.b, z10.b\n"
-      ".inst 0x45c99811  // ummla z17.s, z0.b, z9.b\n"
-      ".inst 0x45c89815  // ummla z21.s, z0.b, z8.b\n"
-      ".inst 0x45c49812  // ummla z18.s, z0.b, z4.b\n"
-      ".inst 0x45c59858  // ummla z24.s, z2.b, z5.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x45ca985c  // ummla z28.s, z2.b, z10.b\n"
-      ".inst 0x45c99859  // ummla z25.s, z2.b, z9.b\n"
-      ".inst 0x45c8985d  // ummla z29.s, z2.b, z8.b\n"
-      ".inst 0x45c4985a  // ummla z26.s, z2.b, z4.b\n"
-      ".inst 0x45c79816  // ummla z22.s, z0.b, z7.b\n"
+      ".inst 0x45c49859  // ummla z25.s, z2.b, z4.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x45c7985e  // ummla z30.s, z2.b, z7.b\n"
-      ".inst 0x45c69813  // ummla z19.s, z0.b, z6.b\n"
+      ".inst 0x45c99815  // ummla z21.s, z0.b, z9.b\n"
+      ".inst 0x45c9985d  // ummla z29.s, z2.b, z9.b\n"
       "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
+      ".inst 0x45c89812  // ummla z18.s, z0.b, z8.b\n"
+      ".inst 0x45c8985a  // ummla z26.s, z2.b, z8.b\n"
       "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x45c6985b  // ummla z27.s, z2.b, z6.b\n"
-      ".inst 0x45c59817  // ummla z23.s, z0.b, z5.b\n"
       "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
+      ".inst 0x45c79816  // ummla z22.s, z0.b, z7.b\n"
+      ".inst 0x45c7985e  // ummla z30.s, z2.b, z7.b\n"
       "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x45c5985f  // ummla z31.s, z2.b, z5.b\n"
-      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45c69813  // ummla z19.s, z0.b, z6.b\n"
+      ".inst 0x45c6985b  // ummla z27.s, z2.b, z6.b\n"
       "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x45c59817  // ummla z23.s, z0.b, z5.b\n"
+      ".inst 0x45c5985f  // ummla z31.s, z2.b, z5.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
+      "add x22, x22, #0x10\n"
+      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
       ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
       ".inst 0x45ca9834  // ummla z20.s, z1.b, z10.b\n"
@@ -699,32 +699,32 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "bgt 35b\n"
       "37:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "trn1 z0.d, z1.d, z2.d\n"
       "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z5.d\n"
-      "trn2 z3.d, z3.d, z5.d\n"
+      "ld1b { z5.b }, p2/Z, [x28]\n"
+      "trn1 z2.d, z3.d, z4.d\n"
+      "trn2 z3.d, z3.d, z4.d\n"
+      ".inst 0x45c59810  // ummla z16.s, z0.b, z5.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
+      ".inst 0x45c59858  // ummla z24.s, z2.b, z5.b\n"
+      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "subs x25, x25, #0x8\n"
+      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
+      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
+      ".inst 0x45c49814  // ummla z20.s, z0.b, z4.b\n"
+      ".inst 0x45c4985c  // ummla z28.s, z2.b, z4.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45c49810  // ummla z16.s, z0.b, z4.b\n"
-      ".inst 0x45ca9814  // ummla z20.s, z0.b, z10.b\n"
-      ".inst 0x45c99811  // ummla z17.s, z0.b, z9.b\n"
-      ".inst 0x45c89815  // ummla z21.s, z0.b, z8.b\n"
-      ".inst 0x45c79812  // ummla z18.s, z0.b, z7.b\n"
-      ".inst 0x45c49858  // ummla z24.s, z2.b, z4.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x45ca985c  // ummla z28.s, z2.b, z10.b\n"
-      "addvl x28, x28, #8\n"
+      ".inst 0x45c99811  // ummla z17.s, z0.b, z9.b\n"
       ".inst 0x45c99859  // ummla z25.s, z2.b, z9.b\n"
+      ".inst 0x45c89815  // ummla z21.s, z0.b, z8.b\n"
       ".inst 0x45c8985d  // ummla z29.s, z2.b, z8.b\n"
+      "addvl x28, x28, #8\n"
+      ".inst 0x45c79812  // ummla z18.s, z0.b, z7.b\n"
       ".inst 0x45c7985a  // ummla z26.s, z2.b, z7.b\n"
       ".inst 0x45c69816  // ummla z22.s, z0.b, z6.b\n"
       ".inst 0x45c6985e  // ummla z30.s, z2.b, z6.b\n"
@@ -734,24 +734,24 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       ".inst 0x45c4985f  // ummla z31.s, z2.b, z4.b\n"
       "ble 38f\n"
       "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
+      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
+      ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
+      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45c59834  // ummla z20.s, z1.b, z5.b\n"
+      ".inst 0x45c5987c  // ummla z28.s, z3.b, z5.b\n"
       "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
       "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
-      ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
+      ".inst 0x45c49831  // ummla z17.s, z1.b, z4.b\n"
+      ".inst 0x45c49879  // ummla z25.s, z3.b, z4.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x45ca9834  // ummla z20.s, z1.b, z10.b\n"
-      ".inst 0x45ca987c  // ummla z28.s, z3.b, z10.b\n"
-      ".inst 0x45c99831  // ummla z17.s, z1.b, z9.b\n"
-      ".inst 0x45c99879  // ummla z25.s, z3.b, z9.b\n"
-      "addvl x28, x28, #8\n"
       ".inst 0x45c89835  // ummla z21.s, z1.b, z8.b\n"
       ".inst 0x45c8987d  // ummla z29.s, z3.b, z8.b\n"
       ".inst 0x45c79832  // ummla z18.s, z1.b, z7.b\n"
       ".inst 0x45c7987a  // ummla z26.s, z3.b, z7.b\n"
+      "addvl x28, x28, #8\n"
       ".inst 0x45c69836  // ummla z22.s, z1.b, z6.b\n"
       ".inst 0x45c6987e  // ummla z30.s, z3.b, z6.b\n"
       ".inst 0x45c59833  // ummla z19.s, z1.b, z5.b\n"
@@ -771,15 +771,15 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "bne 32b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z0.d, z16.d, z20.d\n"
+      "add x23, x27, x20\n"
       "uzp2 z16.d, z16.d, z20.d\n"
       "uzp1 z20.d, z17.d, z21.d\n"
       "uzp2 z17.d, z17.d, z21.d\n"
+      "add x22, x23, x20\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "add x23, x27, x20\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "add x22, x23, x20\n"
       "uzp1 z24.d, z24.d, z28.d\n"
       "uzp1 z25.d, z25.d, z29.d\n"
       "uzp1 z26.d, z26.d, z30.d\n"
@@ -787,14 +787,14 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "mov z31.d, z0.d\n"
       "tbnz %x[flags], #31, 40f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1rw { z23.s }, p2/Z, [x20]\n"
       ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
       ".inst 0x4491a9ad  // addp z13.s, p2/m, z13.s, z13.s\n"
-      "ld1rw { z23.s }, p2/Z, [x20]\n"
       "neg z23.s, p2/M, z23.s\n"
       "mov z12.s, z11.s[3]\n"
       "mov z11.s, z11.s[0]\n"
-      "mov z13.s, z13.s[0]\n"
       "mul z11.s, p2/M, z11.s, z23.s\n"
+      "mov z13.s, z13.s[0]\n"
       "mul z12.s, p2/M, z12.s, z23.s\n"
       "mul z13.s, p2/M, z13.s, z23.s\n"
       "40:"  // Height 3: skip row sum fixup
@@ -850,18 +850,18 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "and z29.d, z21.d, z0.d\n"
       "and z28.d, z22.d, z0.d\n"
       "and z23.d, z16.d, z0.d\n"
-      "and z3.d, z17.d, z0.d\n"
       "asr z1.s, z1.s, #0x1f\n"
       "asr z30.s, z30.s, #0x1f\n"
       "asr z29.s, z29.s, #0x1f\n"
       "asr z28.s, z28.s, #0x1f\n"
       "asr z23.s, z23.s, #0x1f\n"
-      "and z2.d, z18.d, z0.d\n"
       "sqadd z31.s, z31.s, z1.s\n"
       "sqadd z20.s, z20.s, z30.s\n"
       "sqadd z21.s, z21.s, z29.s\n"
       "sqadd z22.s, z22.s, z28.s\n"
       "sqadd z16.s, z16.s, z23.s\n"
+      "and z3.d, z17.d, z0.d\n"
+      "and z2.d, z18.d, z0.d\n"
       "and z1.d, z19.d, z0.d\n"
       "and z30.d, z24.d, z0.d\n"
       "and z29.d, z25.d, z0.d\n"
@@ -883,35 +883,35 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "sqadd z27.s, z27.s, z23.s\n"
       "41:"  // Height 3: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
+      "ld1rw { z23.s }, p2/Z, [x20]\n"
       ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
-      "ld1rw { z29.s }, p2/Z, [x20]\n"
+      "add z31.s, z31.s, z23.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
+      "add z20.s, z20.s, z23.s\n"
+      "add z21.s, z21.s, z23.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
+      "add z22.s, z22.s, z23.s\n"
+      "add z16.s, z16.s, z23.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z31.s, z31.s, z29.s\n"
+      "add z17.s, z17.s, z23.s\n"
+      "add z18.s, z18.s, z23.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z20.s, z20.s, z29.s\n"
-      "add z21.s, z21.s, z29.s\n"
+      "add z19.s, z19.s, z23.s\n"
+      "add z24.s, z24.s, z23.s\n"
       ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
       ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z22.s, z22.s, z29.s\n"
-      "add z16.s, z16.s, z29.s\n"
+      "add z25.s, z25.s, z23.s\n"
+      "add z26.s, z26.s, z23.s\n"
       ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z17.s, z17.s, z29.s\n"
-      "add z18.s, z18.s, z29.s\n"
       "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z29.s\n"
-      "add z24.s, z24.s, z29.s\n"
+      "add z27.s, z27.s, z23.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z29.s\n"
-      "add z26.s, z26.s, z29.s\n"
       "ld1rw { z23.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z29.s\n"
       "smin z31.s, p2/M, z31.s, z28.s\n"
       "smin z20.s, p2/M, z20.s, z28.s\n"
       "smin z21.s, p2/M, z21.s, z28.s\n"
@@ -927,28 +927,28 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "smax z31.s, p2/M, z31.s, z23.s\n"
       "smax z20.s, p2/M, z20.s, z23.s\n"
       "smax z21.s, p2/M, z21.s, z23.s\n"
+      "uzp1 z31.h, z31.h, z20.h\n"
       "smax z22.s, p2/M, z22.s, z23.s\n"
       "smax z16.s, p2/M, z16.s, z23.s\n"
+      "uzp1 z20.h, z21.h, z22.h\n"
+      "uzp1 z31.b, z31.b, z20.b\n"
       "smax z17.s, p2/M, z17.s, z23.s\n"
       "smax z18.s, p2/M, z18.s, z23.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
+      "st1b { z31.b }, p1, [x27]\n"
       "smax z19.s, p2/M, z19.s, z23.s\n"
-      "uzp1 z31.h, z31.h, z20.h\n"
       "smax z24.s, p2/M, z24.s, z23.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z25.s, p2/M, z25.s, z23.s\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
       "smax z26.s, p2/M, z26.s, z23.s\n"
-      "smax z27.s, p2/M, z27.s, z23.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
       "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z31.b, z31.b, z20.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "st1b { z31.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
       "st1b { z16.b }, p1, [x23]\n"
+      "smax z27.s, p2/M, z27.s, z23.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "st1b { z24.b }, p1, [x22]\n"
+      "addvl x27, x27, #1\n"
       "42:"  // Height 3: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -956,28 +956,27 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "b 58f\n"
       "43:"  // Height 4
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x4\n"
       "mov x10, %x[col_bias]\n"
       "mov z11.s, #0x0\n"
       "mov z12.s, #0x0\n"
+      "mov z13.s, #0x0\n"
       "bic %x[flags], %x[flags], #0x80000000\n"
       "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov z13.s, #0x0\n"
       "mov z14.s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
       "mov z15.b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+      "mov x27, %x[output_ptr]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "44:"  // Height 4: Column loop
       "mov x20, #0x0\n"
+      "whilelt p1.b, x20, x9\n"
       "mov z16.s, #0x0\n"
       "mov z17.s, #0x0\n"
       "mov z18.s, #0x0\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
       "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
@@ -992,8 +991,8 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "mov x26, #0x0\n"
       "46:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w25, [x20, x26, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 47f\n"
       "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1018,56 +1017,56 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "ble 51f\n"
       "49:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x22]\n"
-      "ld1rqb { z6.b }, p0/Z, [x21]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
       "trn1 z0.d, z1.d, z2.d\n"
+      "ld1rqb { z3.b }, p0/Z, [x22]\n"
+      "ld1rqb { z5.b }, p0/Z, [x21]\n"
       "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z6.d\n"
-      "trn2 z3.d, z3.d, z6.d\n"
-      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45c59810  // ummla z16.s, z0.b, z5.b\n"
+      "trn1 z2.d, z3.d, z5.d\n"
+      "ld1b { z4.b }, p2/Z, [x28]\n"
+      "trn2 z3.d, z3.d, z5.d\n"
+      ".inst 0x45c49810  // ummla z16.s, z0.b, z4.b\n"
+      ".inst 0x45c49858  // ummla z24.s, z2.b, z4.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
       ".inst 0x45c49814  // ummla z20.s, z0.b, z4.b\n"
-      ".inst 0x45c79811  // ummla z17.s, z0.b, z7.b\n"
-      ".inst 0x45c99815  // ummla z21.s, z0.b, z9.b\n"
-      ".inst 0x45c89812  // ummla z18.s, z0.b, z8.b\n"
-      ".inst 0x45c59858  // ummla z24.s, z2.b, z5.b\n"
+      ".inst 0x45c4985c  // ummla z28.s, z2.b, z4.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
+      ".inst 0x45c59811  // ummla z17.s, z0.b, z5.b\n"
+      ".inst 0x45c59859  // ummla z25.s, z2.b, z5.b\n"
+      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
+      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
+      ".inst 0x45c49815  // ummla z21.s, z0.b, z4.b\n"
+      ".inst 0x45c4985d  // ummla z29.s, z2.b, z4.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #16\n"
-      ".inst 0x45c4985c  // ummla z28.s, z2.b, z4.b\n"
-      ".inst 0x45c79859  // ummla z25.s, z2.b, z7.b\n"
-      ".inst 0x45c9985d  // ummla z29.s, z2.b, z9.b\n"
-      ".inst 0x45c8985a  // ummla z26.s, z2.b, z8.b\n"
-      ".inst 0x45ca9816  // ummla z22.s, z0.b, z10.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x45ca985e  // ummla z30.s, z2.b, z10.b\n"
-      ".inst 0x45c69813  // ummla z19.s, z0.b, z6.b\n"
+      ".inst 0x45c89812  // ummla z18.s, z0.b, z8.b\n"
+      ".inst 0x45c8985a  // ummla z26.s, z2.b, z8.b\n"
+      ".inst 0x45c79816  // ummla z22.s, z0.b, z7.b\n"
       "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
       "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x45c6985b  // ummla z27.s, z2.b, z6.b\n"
-      ".inst 0x45c59817  // ummla z23.s, z0.b, z5.b\n"
+      ".inst 0x45c7985e  // ummla z30.s, z2.b, z7.b\n"
+      ".inst 0x45c69813  // ummla z19.s, z0.b, z6.b\n"
       "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
       "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
+      ".inst 0x45c6985b  // ummla z27.s, z2.b, z6.b\n"
+      ".inst 0x45c59817  // ummla z23.s, z0.b, z5.b\n"
+      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       ".inst 0x45c5985f  // ummla z31.s, z2.b, z5.b\n"
       ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
-      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
       "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
+      "add x23, x23, #0x10\n"
       ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
       "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
       ".inst 0x45ca9834  // ummla z20.s, z1.b, z10.b\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x45ca987c  // ummla z28.s, z3.b, z10.b\n"
       ".inst 0x45c99831  // ummla z17.s, z1.b, z9.b\n"
+      "add x21, x21, #0x10\n"
       ".inst 0x45c99879  // ummla z25.s, z3.b, z9.b\n"
       ".inst 0x45c89835  // ummla z21.s, z1.b, z8.b\n"
       ".inst 0x45c8987d  // ummla z29.s, z3.b, z8.b\n"
@@ -1090,60 +1089,60 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "bgt 49b\n"
       "51:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x25\n"
-      "ld1b { z6.b }, p2/Z, [x28]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z7.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
+      "trn1 z0.d, z1.d, z2.d\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "ld1rqb { z5.b }, p0/Z, [x21]\n"
-      "trn1 z0.d, z1.d, z2.d\n"
       "trn2 z1.d, z1.d, z2.d\n"
       "trn1 z2.d, z3.d, z5.d\n"
+      "ld1b { z4.b }, p2/Z, [x28]\n"
       "trn2 z3.d, z3.d, z5.d\n"
+      ".inst 0x45c49810  // ummla z16.s, z0.b, z4.b\n"
+      ".inst 0x45c49858  // ummla z24.s, z2.b, z4.b\n"
+      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
+      "subs x25, x25, #0x8\n"
+      ".inst 0x45c59814  // ummla z20.s, z0.b, z5.b\n"
+      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
+      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
+      ".inst 0x45c5985c  // ummla z28.s, z2.b, z5.b\n"
+      ".inst 0x45c49811  // ummla z17.s, z0.b, z4.b\n"
+      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
       "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45c69810  // ummla z16.s, z0.b, z6.b\n"
-      ".inst 0x45c49814  // ummla z20.s, z0.b, z4.b\n"
-      ".inst 0x45c79811  // ummla z17.s, z0.b, z7.b\n"
-      ".inst 0x45c99815  // ummla z21.s, z0.b, z9.b\n"
-      ".inst 0x45c89812  // ummla z18.s, z0.b, z8.b\n"
-      ".inst 0x45c69858  // ummla z24.s, z2.b, z6.b\n"
-      "ld1b { z6.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x45c4985c  // ummla z28.s, z2.b, z4.b\n"
+      ".inst 0x45c49859  // ummla z25.s, z2.b, z4.b\n"
+      ".inst 0x45c89815  // ummla z21.s, z0.b, z8.b\n"
+      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
+      ".inst 0x45c8985d  // ummla z29.s, z2.b, z8.b\n"
+      ".inst 0x45c79812  // ummla z18.s, z0.b, z7.b\n"
       "addvl x28, x28, #8\n"
-      ".inst 0x45c79859  // ummla z25.s, z2.b, z7.b\n"
-      ".inst 0x45c9985d  // ummla z29.s, z2.b, z9.b\n"
-      ".inst 0x45c8985a  // ummla z26.s, z2.b, z8.b\n"
-      ".inst 0x45ca9816  // ummla z22.s, z0.b, z10.b\n"
-      ".inst 0x45ca985e  // ummla z30.s, z2.b, z10.b\n"
+      ".inst 0x45c7985a  // ummla z26.s, z2.b, z7.b\n"
+      ".inst 0x45c69816  // ummla z22.s, z0.b, z6.b\n"
+      ".inst 0x45c6985e  // ummla z30.s, z2.b, z6.b\n"
       ".inst 0x45c59813  // ummla z19.s, z0.b, z5.b\n"
       ".inst 0x45c5985b  // ummla z27.s, z2.b, z5.b\n"
-      ".inst 0x45c69817  // ummla z23.s, z0.b, z6.b\n"
-      ".inst 0x45c6985f  // ummla z31.s, z2.b, z6.b\n"
+      ".inst 0x45c49817  // ummla z23.s, z0.b, z4.b\n"
+      ".inst 0x45c4985f  // ummla z31.s, z2.b, z4.b\n"
       "ble 52f\n"
       "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
+      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
+      ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
+      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
       "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
+      ".inst 0x45c59834  // ummla z20.s, z1.b, z5.b\n"
+      ".inst 0x45c5987c  // ummla z28.s, z3.b, z5.b\n"
       "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
       "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
-      ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
+      ".inst 0x45c49831  // ummla z17.s, z1.b, z4.b\n"
+      ".inst 0x45c49879  // ummla z25.s, z3.b, z4.b\n"
       "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
       "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x45ca9834  // ummla z20.s, z1.b, z10.b\n"
-      ".inst 0x45ca987c  // ummla z28.s, z3.b, z10.b\n"
-      ".inst 0x45c99831  // ummla z17.s, z1.b, z9.b\n"
-      ".inst 0x45c99879  // ummla z25.s, z3.b, z9.b\n"
-      "addvl x28, x28, #8\n"
       ".inst 0x45c89835  // ummla z21.s, z1.b, z8.b\n"
       ".inst 0x45c8987d  // ummla z29.s, z3.b, z8.b\n"
       ".inst 0x45c79832  // ummla z18.s, z1.b, z7.b\n"
       ".inst 0x45c7987a  // ummla z26.s, z3.b, z7.b\n"
+      "addvl x28, x28, #8\n"
       ".inst 0x45c69836  // ummla z22.s, z1.b, z6.b\n"
       ".inst 0x45c6987e  // ummla z30.s, z3.b, z6.b\n"
       ".inst 0x45c59833  // ummla z19.s, z1.b, z5.b\n"
@@ -1163,16 +1162,16 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "bne 46b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
       "uzp1 z0.d, z16.d, z20.d\n"
+      "add x23, x27, x20\n"
+      "add x22, x23, x20\n"
       "uzp2 z16.d, z16.d, z20.d\n"
       "uzp1 z20.d, z17.d, z21.d\n"
+      "add x21, x22, x20\n"
       "uzp2 z17.d, z17.d, z21.d\n"
       "uzp1 z21.d, z18.d, z22.d\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "add x23, x27, x20\n"
-      "add x22, x23, x20\n"
       "uzp1 z22.d, z19.d, z23.d\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "add x21, x22, x20\n"
       "uzp1 z23.d, z24.d, z28.d\n"
       "uzp2 z24.d, z24.d, z28.d\n"
       "uzp1 z28.d, z25.d, z29.d\n"
@@ -1184,15 +1183,15 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "mov z31.d, z0.d\n"
       "tbnz %x[flags], #31, 54f\n"
       "add x20, %x[qp], %[b_offset]\n"
+      "ld1rw { z0.s }, p2/Z, [x20]\n"
       ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
       ".inst 0x4491a9ad  // addp z13.s, p2/m, z13.s, z13.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
       "neg z0.s, p2/M, z0.s\n"
       "mov z12.s, z11.s[3]\n"
       "mov z11.s, z11.s[0]\n"
+      "mul z11.s, p2/M, z11.s, z0.s\n"
       "mov z14.s, z13.s[3]\n"
       "mov z13.s, z13.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z0.s\n"
       "mul z12.s, p2/M, z12.s, z0.s\n"
       "mul z13.s, p2/M, z13.s, z0.s\n"
       "mul z14.s, p2/M, z14.s, z0.s\n"
@@ -1258,32 +1257,32 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "tbz %x[flags], #5, 55f\n"
       "and z2.d, z31.d, z0.d\n"
       "and z1.d, z20.d, z0.d\n"
+      "asr z2.s, z2.s, #0x1f\n"
+      "asr z1.s, z1.s, #0x1f\n"
+      "sqadd z31.s, z31.s, z2.s\n"
+      "sqadd z20.s, z20.s, z1.s\n"
       "and z7.d, z21.d, z0.d\n"
       "and z6.d, z22.d, z0.d\n"
       "and z5.d, z16.d, z0.d\n"
       "and z4.d, z17.d, z0.d\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
       "and z3.d, z18.d, z0.d\n"
+      "and z2.d, z19.d, z0.d\n"
+      "and z1.d, z23.d, z0.d\n"
       "asr z7.s, z7.s, #0x1f\n"
       "asr z6.s, z6.s, #0x1f\n"
       "asr z5.s, z5.s, #0x1f\n"
-      "sqadd z31.s, z31.s, z2.s\n"
-      "sqadd z20.s, z20.s, z1.s\n"
-      "and z2.d, z19.d, z0.d\n"
-      "and z1.d, z23.d, z0.d\n"
       "asr z4.s, z4.s, #0x1f\n"
       "asr z3.s, z3.s, #0x1f\n"
-      "sqadd z21.s, z21.s, z7.s\n"
-      "sqadd z22.s, z22.s, z6.s\n"
       "asr z2.s, z2.s, #0x1f\n"
       "asr z1.s, z1.s, #0x1f\n"
+      "sqadd z21.s, z21.s, z7.s\n"
+      "sqadd z22.s, z22.s, z6.s\n"
       "sqadd z16.s, z16.s, z5.s\n"
       "sqadd z17.s, z17.s, z4.s\n"
       "sqadd z18.s, z18.s, z3.s\n"
-      "and z7.d, z28.d, z0.d\n"
       "sqadd z19.s, z19.s, z2.s\n"
       "sqadd z23.s, z23.s, z1.s\n"
+      "and z7.d, z28.d, z0.d\n"
       "and z6.d, z29.d, z0.d\n"
       "and z5.d, z30.d, z0.d\n"
       "and z4.d, z24.d, z0.d\n"
@@ -1306,43 +1305,43 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "sqadd z27.s, z27.s, z1.s\n"
       "55:"  // Height 4: no shift correction
       "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
       "ld1rw { z2.s }, p2/Z, [x20]\n"
+      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
+      "add z31.s, z31.s, z2.s\n"
       ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
       ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
+      "add z20.s, z20.s, z2.s\n"
+      "add z21.s, z21.s, z2.s\n"
       ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
       ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
+      "add z22.s, z22.s, z2.s\n"
+      "add z16.s, z16.s, z2.s\n"
       ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
       ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z31.s, z31.s, z2.s\n"
+      "add z17.s, z17.s, z2.s\n"
+      "add z18.s, z18.s, z2.s\n"
       ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
       ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "add z20.s, z20.s, z2.s\n"
-      "add z21.s, z21.s, z2.s\n"
+      "add z19.s, z19.s, z2.s\n"
+      "add z23.s, z23.s, z2.s\n"
       ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
       ".inst 0x4482881d  // srshl z29.s, p2/M, z29.s, z0.s\n"
-      "add z22.s, z22.s, z2.s\n"
-      "add z16.s, z16.s, z2.s\n"
+      "add z28.s, z28.s, z2.s\n"
+      "add z29.s, z29.s, z2.s\n"
       ".inst 0x4482881e  // srshl z30.s, p2/M, z30.s, z0.s\n"
       ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z2.s\n"
-      "add z18.s, z18.s, z2.s\n"
+      "add z30.s, z30.s, z2.s\n"
+      "add z24.s, z24.s, z2.s\n"
       ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
       ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z2.s\n"
-      "add z23.s, z23.s, z2.s\n"
+      "add z25.s, z25.s, z2.s\n"
+      "add z26.s, z26.s, z2.s\n"
       ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
       "add x20, %x[qp], %[maxval]\n"
-      "add z28.s, z28.s, z2.s\n"
-      "add z29.s, z29.s, z2.s\n"
       "ld1rw { z1.s }, p2/Z, [x20]\n"
-      "add z30.s, z30.s, z2.s\n"
-      "add z24.s, z24.s, z2.s\n"
+      "add z27.s, z27.s, z2.s\n"
       "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z2.s\n"
-      "add z26.s, z26.s, z2.s\n"
       "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z2.s\n"
       "smin z31.s, p2/M, z31.s, z1.s\n"
       "smin z20.s, p2/M, z20.s, z1.s\n"
       "smin z21.s, p2/M, z21.s, z1.s\n"
@@ -1362,36 +1361,36 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "smax z31.s, p2/M, z31.s, z0.s\n"
       "smax z20.s, p2/M, z20.s, z0.s\n"
       "smax z21.s, p2/M, z21.s, z0.s\n"
+      "uzp1 z31.h, z31.h, z20.h\n"
       "smax z22.s, p2/M, z22.s, z0.s\n"
       "smax z16.s, p2/M, z16.s, z0.s\n"
+      "uzp1 z20.h, z21.h, z22.h\n"
+      "uzp1 z31.b, z31.b, z20.b\n"
       "smax z17.s, p2/M, z17.s, z0.s\n"
       "smax z18.s, p2/M, z18.s, z0.s\n"
+      "uzp1 z16.h, z16.h, z17.h\n"
+      "st1b { z31.b }, p1, [x27]\n"
       "smax z19.s, p2/M, z19.s, z0.s\n"
-      "uzp1 z31.h, z31.h, z20.h\n"
       "smax z23.s, p2/M, z23.s, z0.s\n"
+      "uzp1 z17.h, z18.h, z19.h\n"
+      "uzp1 z16.b, z16.b, z17.b\n"
       "smax z28.s, p2/M, z28.s, z0.s\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
       "smax z29.s, p2/M, z29.s, z0.s\n"
+      "uzp1 z23.h, z23.h, z28.h\n"
+      "st1b { z16.b }, p1, [x23]\n"
       "smax z30.s, p2/M, z30.s, z0.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
       "smax z24.s, p2/M, z24.s, z0.s\n"
+      "uzp1 z16.h, z29.h, z30.h\n"
+      "uzp1 z23.b, z23.b, z16.b\n"
       "smax z25.s, p2/M, z25.s, z0.s\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
       "smax z26.s, p2/M, z26.s, z0.s\n"
-      "smax z27.s, p2/M, z27.s, z0.s\n"
-      "uzp1 z23.h, z23.h, z28.h\n"
-      "uzp1 z31.b, z31.b, z20.b\n"
-      "uzp1 z18.h, z29.h, z30.h\n"
       "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "st1b { z31.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "uzp1 z23.b, z23.b, z18.b\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "st1b { z16.b }, p1, [x23]\n"
       "st1b { z23.b }, p1, [x22]\n"
+      "smax z27.s, p2/M, z27.s, z0.s\n"
+      "uzp1 z16.h, z26.h, z27.h\n"
+      "uzp1 z24.b, z24.b, z16.b\n"
       "st1b { z24.b }, p1, [x21]\n"
+      "addvl x27, x27, #1\n"
       "56:"  // Height 4: Writeback done
       "decw x9, ALL, MUL #4\n"
       "cmp x9, XZR\n"
@@ -1408,8 +1407,8 @@ void sve_hybrid_u8qa_mmla_4x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "58:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
       : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_dot_4x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_dot_4x4VL.hpp
deleted file mode 100644
index ebf1883ccf..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_dot_4x4VL.hpp
+++ /dev/null
@@ -1,99 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include "../std_transforms_sve.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    unsigned int, const unsigned int *, \
-    IndirectInputArg<uint8_t>, \
-    size_t, size_t, \
-    const int8_t *, \
-    IndirectOutputArg<uint8_t>, \
-    const Requantize32 *, const int32_t *, unsigned int
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void sve_hybrid_u8s8qa_dot_4x4VL( ARGLIST );
-
-class cls_sve_hybrid_u8s8qa_dot_4x4VL
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef uint8_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 4;
-    }
-
-    static unsigned int out_width()
-    {
-        return get_vector_length<int32_t>() * 4;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 4;
-    }
-
-    static constexpr bool supports_accumulate()
-    {
-        return false;
-    }
-
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 4, 4, 4> transforms = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 29.89 };
-                case CPUModel::A510:
-                    return { 17.12 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=sve_hybrid_u8s8qa_dot_4x4VL;
-    cls_sve_hybrid_u8s8qa_dot_4x4VL(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_dot_4x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_dot_4x4VL/generic.cpp
deleted file mode 100644
index dc3b7ef3ec..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_dot_4x4VL/generic.cpp
+++ /dev/null
@@ -1,1502 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include "arm_gemm.hpp"
-#include "../../utils.hpp"
-
-#include <cassert>
-#include <cstdint>
-
-namespace arm_gemm {
-
-void sve_hybrid_u8s8qa_dot_4x4VL (
-    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg,
-    size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<uint8_t> output_arg,
-    const Requantize32 *qp, const int32_t *col_bias, unsigned int
-)
-{
-    struct KernelArgs {
-        unsigned int num_strings = {};
-        const unsigned int *string_lengths = {};
-        size_t N = {};
-        const int8_t *B_ptr = {};
-        size_t output_offset = {};
-        size_t input_initial_col = {};
-        size_t input_offset = {};
-        void *output_ptr = {};
-    } ka;
-
-    unsigned long flags=0;
-    void *input_ptr;
-
-    if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
-        ka.output_offset=output_arg.indirect.offset;
-        flags |= 0x4;
-    } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
-        ka.output_offset=output_arg.direct.stride;
-    }
-
-    if (A_arg.is_indirect) {
-        input_ptr=(void *)(A_arg.indirect.ptr);
-        ka.input_offset=A_arg.indirect.start_row;
-        ka.input_initial_col=A_arg.indirect.start_col;
-        flags |= 0x8;
-    } else {
-        assert(num_strings==1);
-        input_ptr=(void *)(A_arg.direct.base);
-        ka.input_offset=A_arg.direct.stride;
-    }
-    ka.num_strings = num_strings;
-    ka.string_lengths = string_lengths;
-    ka.N = N;
-    ka.B_ptr = B_ptr;
-    if (qp->c_offset > qp->minval) {
-        flags |= 0x20;
-    }
-    __asm__ __volatile__(
-      "ptrue p2.b\n"
-      "1:"  // Row loop
-      "cmp %x[M], #0x4\n"
-      "bge 43f\n"
-      "cmp %x[M], #0x2\n"
-      "bgt 29f\n"
-      "beq 15f\n"
-      "mov x10, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z15.b, #0x1\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "2:"  // Height 1: Column loop
-      "mov x20, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
-      "3:"  // Height 1: setup done
-      "mov x26, #0x0\n"
-      "4:"  // Height 1: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 5f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "cbnz x26, 6f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "b 6f\n"
-      "5:"  // Height 1: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "6:"  // Height 1: input setup done
-      "cmp x25, #0x10\n"
-      "ble 9f\n"
-      "7:"  // Height 1: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z21.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z23.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x24, x24, #0x10\n"
-      ".inst 0x44a01eb0  // sudot z16.s, z21.b, z0.b[0]\n"
-      "ld1b { z21.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      ".inst 0x44a01f51  // sudot z17.s, z26.b, z0.b[0]\n"
-      ".inst 0x44a01f32  // sudot z18.s, z25.b, z0.b[0]\n"
-      ".inst 0x44a01f13  // sudot z19.s, z24.b, z0.b[0]\n"
-      ".inst 0x44a81e90  // sudot z16.s, z20.b, z0.b[1]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      ".inst 0x44a81ef1  // sudot z17.s, z23.b, z0.b[1]\n"
-      "ld1b { z23.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x44a81ed2  // sudot z18.s, z22.b, z0.b[1]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      ".inst 0x44a81eb3  // sudot z19.s, z21.b, z0.b[1]\n"
-      "ld1b { z21.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      ".inst 0x44b01e90  // sudot z16.s, z20.b, z0.b[2]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x44b01f51  // sudot z17.s, z26.b, z0.b[2]\n"
-      ".inst 0x44b01f32  // sudot z18.s, z25.b, z0.b[2]\n"
-      ".inst 0x44b01f13  // sudot z19.s, z24.b, z0.b[2]\n"
-      ".inst 0x44b81ef0  // sudot z16.s, z23.b, z0.b[3]\n"
-      ".inst 0x44b81ed1  // sudot z17.s, z22.b, z0.b[3]\n"
-      ".inst 0x44b81eb2  // sudot z18.s, z21.b, z0.b[3]\n"
-      ".inst 0x44b81e93  // sudot z19.s, z20.b, z0.b[3]\n"
-      "tbnz %x[flags], #31, 8f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "8:"  // Height 1: Multiply loop: unique 1: skip row sum
-      "sub x25, x25, #0x10\n"
-      "cmp x25, #0x10\n"
-      "bgt 7b\n"
-      "9:"  // Height 1: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      ".inst 0x44a01ef0  // sudot z16.s, z23.b, z0.b[0]\n"
-      ".inst 0x44a01ed1  // sudot z17.s, z22.b, z0.b[0]\n"
-      ".inst 0x44a01eb2  // sudot z18.s, z21.b, z0.b[0]\n"
-      ".inst 0x44a01e93  // sudot z19.s, z20.b, z0.b[0]\n"
-      "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44a81ef0  // sudot z16.s, z23.b, z0.b[1]\n"
-      ".inst 0x44a81ed1  // sudot z17.s, z22.b, z0.b[1]\n"
-      ".inst 0x44a81eb2  // sudot z18.s, z21.b, z0.b[1]\n"
-      ".inst 0x44a81e93  // sudot z19.s, z20.b, z0.b[1]\n"
-      "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44b01ef0  // sudot z16.s, z23.b, z0.b[2]\n"
-      ".inst 0x44b01ed1  // sudot z17.s, z22.b, z0.b[2]\n"
-      ".inst 0x44b01eb2  // sudot z18.s, z21.b, z0.b[2]\n"
-      ".inst 0x44b01e93  // sudot z19.s, z20.b, z0.b[2]\n"
-      "ble 10f\n"
-      "ld1b { z23.b }, p2/Z, [x28]\n"
-      "ld1b { z22.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z21.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z20.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44b81ef0  // sudot z16.s, z23.b, z0.b[3]\n"
-      ".inst 0x44b81ed1  // sudot z17.s, z22.b, z0.b[3]\n"
-      ".inst 0x44b81eb2  // sudot z18.s, z21.b, z0.b[3]\n"
-      ".inst 0x44b81e93  // sudot z19.s, z20.b, z0.b[3]\n"
-      "10:"  // Height 1: Multiply loop: multiply skip
-      "tbnz %x[flags], #31, 11f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "11:"  // Height 1: Multiply loop: unique 2: skip row sum
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 4b\n"
-      "tbnz %x[flags], #31, 12f\n"
-      "mov x21, #0x4\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z20.s, p2/M, z20.s\n"
-      "saddv d11, p0, z11.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z20.s\n"
-      "12:"  // Height 1: skip row sum fixup
-      "add z16.s, z16.s, z11.s\n"
-      "add z17.s, z17.s, z11.s\n"
-      "ld1w { z23.s }, p2/Z, [x10]\n"
-      "ld1w { z20.s }, p2/Z, [x10, #1, MUL VL]\n"
-      "add z18.s, z18.s, z11.s\n"
-      "add z19.s, z19.s, z11.s\n"
-      "ld1w { z22.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x10, #3, MUL VL]\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "add z16.s, z16.s, z23.s\n"
-      "add z17.s, z17.s, z20.s\n"
-      "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "addvl x10, x10, #4\n"
-      "add z18.s, z18.s, z22.s\n"
-      "add z19.s, z19.s, z21.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      ".inst 0x04b47610  // sqrdmulh z16.s, z16.s, z20.s\n"
-      ".inst 0x04b47631  // sqrdmulh z17.s, z17.s, z20.s\n"
-      ".inst 0x04b47652  // sqrdmulh z18.s, z18.s, z20.s\n"
-      ".inst 0x04b47673  // sqrdmulh z19.s, z19.s, z20.s\n"
-      "tbz %x[flags], #5, 13f\n"
-      "and z23.d, z16.d, z0.d\n"
-      "and z22.d, z17.d, z0.d\n"
-      "and z21.d, z18.d, z0.d\n"
-      "and z20.d, z19.d, z0.d\n"
-      "asr z23.s, z23.s, #0x1f\n"
-      "asr z22.s, z22.s, #0x1f\n"
-      "asr z21.s, z21.s, #0x1f\n"
-      "asr z20.s, z20.s, #0x1f\n"
-      "sqadd z16.s, z16.s, z23.s\n"
-      "sqadd z17.s, z17.s, z22.s\n"
-      "sqadd z18.s, z18.s, z21.s\n"
-      "sqadd z19.s, z19.s, z20.s\n"
-      "13:"  // Height 1: no shift correction
-      "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z22.s }, p2/Z, [x20]\n"
-      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
-      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1rw { z21.s }, p2/Z, [x20]\n"
-      "add z16.s, z16.s, z22.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z22.s\n"
-      "add z18.s, z18.s, z22.s\n"
-      "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z22.s\n"
-      "smin z16.s, p2/M, z16.s, z21.s\n"
-      "smin z17.s, p2/M, z17.s, z21.s\n"
-      "smin z18.s, p2/M, z18.s, z21.s\n"
-      "smin z19.s, p2/M, z19.s, z21.s\n"
-      "smax z16.s, p2/M, z16.s, z20.s\n"
-      "smax z17.s, p2/M, z17.s, z20.s\n"
-      "smax z18.s, p2/M, z18.s, z20.s\n"
-      "smax z19.s, p2/M, z19.s, z20.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "14:"  // Height 1: Writeback done
-      "decw x9, ALL, MUL #4\n"
-      "cmp x9, XZR\n"
-      "bgt 2b\n"
-      "b 58f\n"
-      "15:"  // Height 2
-      "mov x10, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "mov z15.b, #0x1\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "16:"  // Height 2: Column loop
-      "mov x20, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "17:"  // Height 2: setup done
-      "mov x26, #0x0\n"
-      "18:"  // Height 2: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 19f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "cbnz x26, 20f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "b 20f\n"
-      "19:"  // Height 2: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "20:"  // Height 2: input setup done
-      "cmp x25, #0x10\n"
-      "ble 23f\n"
-      "21:"  // Height 2: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z25.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1rqb { z1.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x44a01f30  // sudot z16.s, z25.b, z0.b[0]\n"
-      ".inst 0x44a11f34  // sudot z20.s, z25.b, z1.b[0]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      ".inst 0x44a01fd1  // sudot z17.s, z30.b, z0.b[0]\n"
-      ".inst 0x44a11fd5  // sudot z21.s, z30.b, z1.b[0]\n"
-      ".inst 0x44a01fb2  // sudot z18.s, z29.b, z0.b[0]\n"
-      ".inst 0x44a11fb6  // sudot z22.s, z29.b, z1.b[0]\n"
-      ".inst 0x44a01f93  // sudot z19.s, z28.b, z0.b[0]\n"
-      ".inst 0x44a11f97  // sudot z23.s, z28.b, z1.b[0]\n"
-      ".inst 0x44a81f10  // sudot z16.s, z24.b, z0.b[1]\n"
-      ".inst 0x44a91f14  // sudot z20.s, z24.b, z1.b[1]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      ".inst 0x44a81f71  // sudot z17.s, z27.b, z0.b[1]\n"
-      ".inst 0x44a91f75  // sudot z21.s, z27.b, z1.b[1]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      ".inst 0x44a81f52  // sudot z18.s, z26.b, z0.b[1]\n"
-      ".inst 0x44a91f56  // sudot z22.s, z26.b, z1.b[1]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      ".inst 0x44a81f33  // sudot z19.s, z25.b, z0.b[1]\n"
-      ".inst 0x44a91f37  // sudot z23.s, z25.b, z1.b[1]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      ".inst 0x44b01f10  // sudot z16.s, z24.b, z0.b[2]\n"
-      ".inst 0x44b11f14  // sudot z20.s, z24.b, z1.b[2]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x44b01fd1  // sudot z17.s, z30.b, z0.b[2]\n"
-      ".inst 0x44b11fd5  // sudot z21.s, z30.b, z1.b[2]\n"
-      ".inst 0x44b01fb2  // sudot z18.s, z29.b, z0.b[2]\n"
-      ".inst 0x44b11fb6  // sudot z22.s, z29.b, z1.b[2]\n"
-      ".inst 0x44b01f93  // sudot z19.s, z28.b, z0.b[2]\n"
-      ".inst 0x44b11f97  // sudot z23.s, z28.b, z1.b[2]\n"
-      ".inst 0x44b81f70  // sudot z16.s, z27.b, z0.b[3]\n"
-      ".inst 0x44b91f74  // sudot z20.s, z27.b, z1.b[3]\n"
-      ".inst 0x44b81f51  // sudot z17.s, z26.b, z0.b[3]\n"
-      ".inst 0x44b91f55  // sudot z21.s, z26.b, z1.b[3]\n"
-      ".inst 0x44b81f32  // sudot z18.s, z25.b, z0.b[3]\n"
-      ".inst 0x44b91f36  // sudot z22.s, z25.b, z1.b[3]\n"
-      ".inst 0x44b81f13  // sudot z19.s, z24.b, z0.b[3]\n"
-      ".inst 0x44b91f17  // sudot z23.s, z24.b, z1.b[3]\n"
-      "tbnz %x[flags], #31, 22f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z12.s, z1.b, z15.b\n"
-      "22:"  // Height 2: Multiply loop: unique 3: skip row sum
-      "sub x25, x25, #0x10\n"
-      "cmp x25, #0x10\n"
-      "bgt 21b\n"
-      "23:"  // Height 2: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z27.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1rqb { z1.b }, p0/Z, [x23]\n"
-      ".inst 0x44a01f70  // sudot z16.s, z27.b, z0.b[0]\n"
-      ".inst 0x44a11f74  // sudot z20.s, z27.b, z1.b[0]\n"
-      ".inst 0x44a01f51  // sudot z17.s, z26.b, z0.b[0]\n"
-      ".inst 0x44a11f55  // sudot z21.s, z26.b, z1.b[0]\n"
-      ".inst 0x44a01f32  // sudot z18.s, z25.b, z0.b[0]\n"
-      ".inst 0x44a11f36  // sudot z22.s, z25.b, z1.b[0]\n"
-      ".inst 0x44a01f13  // sudot z19.s, z24.b, z0.b[0]\n"
-      ".inst 0x44a11f17  // sudot z23.s, z24.b, z1.b[0]\n"
-      "ble 24f\n"
-      "ld1b { z27.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44a81f70  // sudot z16.s, z27.b, z0.b[1]\n"
-      ".inst 0x44a91f74  // sudot z20.s, z27.b, z1.b[1]\n"
-      ".inst 0x44a81f51  // sudot z17.s, z26.b, z0.b[1]\n"
-      ".inst 0x44a91f55  // sudot z21.s, z26.b, z1.b[1]\n"
-      ".inst 0x44a81f32  // sudot z18.s, z25.b, z0.b[1]\n"
-      ".inst 0x44a91f36  // sudot z22.s, z25.b, z1.b[1]\n"
-      ".inst 0x44a81f13  // sudot z19.s, z24.b, z0.b[1]\n"
-      ".inst 0x44a91f17  // sudot z23.s, z24.b, z1.b[1]\n"
-      "ble 24f\n"
-      "ld1b { z27.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44b01f70  // sudot z16.s, z27.b, z0.b[2]\n"
-      ".inst 0x44b11f74  // sudot z20.s, z27.b, z1.b[2]\n"
-      ".inst 0x44b01f51  // sudot z17.s, z26.b, z0.b[2]\n"
-      ".inst 0x44b11f55  // sudot z21.s, z26.b, z1.b[2]\n"
-      ".inst 0x44b01f32  // sudot z18.s, z25.b, z0.b[2]\n"
-      ".inst 0x44b11f36  // sudot z22.s, z25.b, z1.b[2]\n"
-      ".inst 0x44b01f13  // sudot z19.s, z24.b, z0.b[2]\n"
-      ".inst 0x44b11f17  // sudot z23.s, z24.b, z1.b[2]\n"
-      "ble 24f\n"
-      "ld1b { z27.b }, p2/Z, [x28]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44b81f70  // sudot z16.s, z27.b, z0.b[3]\n"
-      ".inst 0x44b91f74  // sudot z20.s, z27.b, z1.b[3]\n"
-      ".inst 0x44b81f51  // sudot z17.s, z26.b, z0.b[3]\n"
-      ".inst 0x44b91f55  // sudot z21.s, z26.b, z1.b[3]\n"
-      ".inst 0x44b81f32  // sudot z18.s, z25.b, z0.b[3]\n"
-      ".inst 0x44b91f36  // sudot z22.s, z25.b, z1.b[3]\n"
-      ".inst 0x44b81f13  // sudot z19.s, z24.b, z0.b[3]\n"
-      ".inst 0x44b91f17  // sudot z23.s, z24.b, z1.b[3]\n"
-      "24:"  // Height 2: Multiply loop: multiply skip
-      "tbnz %x[flags], #31, 25f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z12.s, z1.b, z15.b\n"
-      "25:"  // Height 2: Multiply loop: unique 4: skip row sum
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 18b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
-      "tbnz %x[flags], #31, 26f\n"
-      "mov x21, #0x4\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z24.s, p2/M, z24.s\n"
-      "saddv d11, p0, z11.s\n"
-      "saddv d12, p0, z12.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z24.s\n"
-      "mov z12.s, z12.s[0]\n"
-      "mul z12.s, p2/M, z12.s, z24.s\n"
-      "26:"  // Height 2: skip row sum fixup
-      "add z16.s, z16.s, z11.s\n"
-      "add z17.s, z17.s, z11.s\n"
-      "ld1w { z28.s }, p2/Z, [x10]\n"
-      "ld1w { z27.s }, p2/Z, [x10, #1, MUL VL]\n"
-      "add z18.s, z18.s, z11.s\n"
-      "add z19.s, z19.s, z11.s\n"
-      "ld1w { z26.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x10, #3, MUL VL]\n"
-      "add z20.s, z20.s, z12.s\n"
-      "add z21.s, z21.s, z12.s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "add z22.s, z22.s, z12.s\n"
-      "add z23.s, z23.s, z12.s\n"
-      "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add z16.s, z16.s, z28.s\n"
-      "add z17.s, z17.s, z27.s\n"
-      "addvl x10, x10, #4\n"
-      "add z18.s, z18.s, z26.s\n"
-      "add z19.s, z19.s, z25.s\n"
-      "add z20.s, z20.s, z28.s\n"
-      "add z21.s, z21.s, z27.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z22.s, z22.s, z26.s\n"
-      "add z23.s, z23.s, z25.s\n"
-      ".inst 0x04b87610  // sqrdmulh z16.s, z16.s, z24.s\n"
-      ".inst 0x04b87631  // sqrdmulh z17.s, z17.s, z24.s\n"
-      ".inst 0x04b87652  // sqrdmulh z18.s, z18.s, z24.s\n"
-      ".inst 0x04b87673  // sqrdmulh z19.s, z19.s, z24.s\n"
-      ".inst 0x04b87694  // sqrdmulh z20.s, z20.s, z24.s\n"
-      ".inst 0x04b876b5  // sqrdmulh z21.s, z21.s, z24.s\n"
-      ".inst 0x04b876d6  // sqrdmulh z22.s, z22.s, z24.s\n"
-      ".inst 0x04b876f7  // sqrdmulh z23.s, z23.s, z24.s\n"
-      "tbz %x[flags], #5, 27f\n"
-      "and z24.d, z16.d, z0.d\n"
-      "and z30.d, z17.d, z0.d\n"
-      "and z29.d, z18.d, z0.d\n"
-      "and z28.d, z19.d, z0.d\n"
-      "and z27.d, z20.d, z0.d\n"
-      "and z26.d, z21.d, z0.d\n"
-      "asr z24.s, z24.s, #0x1f\n"
-      "and z25.d, z22.d, z0.d\n"
-      "asr z30.s, z30.s, #0x1f\n"
-      "asr z29.s, z29.s, #0x1f\n"
-      "asr z28.s, z28.s, #0x1f\n"
-      "asr z27.s, z27.s, #0x1f\n"
-      "sqadd z16.s, z16.s, z24.s\n"
-      "and z24.d, z23.d, z0.d\n"
-      "asr z26.s, z26.s, #0x1f\n"
-      "asr z25.s, z25.s, #0x1f\n"
-      "sqadd z17.s, z17.s, z30.s\n"
-      "sqadd z18.s, z18.s, z29.s\n"
-      "asr z24.s, z24.s, #0x1f\n"
-      "sqadd z19.s, z19.s, z28.s\n"
-      "sqadd z20.s, z20.s, z27.s\n"
-      "sqadd z21.s, z21.s, z26.s\n"
-      "sqadd z22.s, z22.s, z25.s\n"
-      "sqadd z23.s, z23.s, z24.s\n"
-      "27:"  // Height 2: no shift correction
-      "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z26.s }, p2/Z, [x20]\n"
-      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
-      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
-      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
-      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z26.s\n"
-      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "add z17.s, z17.s, z26.s\n"
-      "add z18.s, z18.s, z26.s\n"
-      "ld1rw { z25.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z26.s\n"
-      "add z20.s, z20.s, z26.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z21.s, z21.s, z26.s\n"
-      "add z22.s, z22.s, z26.s\n"
-      "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z26.s\n"
-      "smin z16.s, p2/M, z16.s, z25.s\n"
-      "smin z17.s, p2/M, z17.s, z25.s\n"
-      "smin z18.s, p2/M, z18.s, z25.s\n"
-      "smin z19.s, p2/M, z19.s, z25.s\n"
-      "smin z20.s, p2/M, z20.s, z25.s\n"
-      "smin z21.s, p2/M, z21.s, z25.s\n"
-      "smin z22.s, p2/M, z22.s, z25.s\n"
-      "smin z23.s, p2/M, z23.s, z25.s\n"
-      "smax z16.s, p2/M, z16.s, z24.s\n"
-      "smax z17.s, p2/M, z17.s, z24.s\n"
-      "smax z18.s, p2/M, z18.s, z24.s\n"
-      "smax z19.s, p2/M, z19.s, z24.s\n"
-      "smax z20.s, p2/M, z20.s, z24.s\n"
-      "smax z21.s, p2/M, z21.s, z24.s\n"
-      "smax z22.s, p2/M, z22.s, z24.s\n"
-      "smax z23.s, p2/M, z23.s, z24.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z17.h, z22.h, z23.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z20.b, z20.b, z17.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "st1b { z20.b }, p1, [x24]\n"
-      "28:"  // Height 2: Writeback done
-      "decw x9, ALL, MUL #4\n"
-      "cmp x9, XZR\n"
-      "bgt 16b\n"
-      "b 58f\n"
-      "29:"  // Height 3
-      "mov x10, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "mov z13.s, #0x0\n"
-      "mov z15.b, #0x1\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "30:"  // Height 3: Column loop
-      "mov x20, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "mov z24.s, #0x0\n"
-      "mov z25.s, #0x0\n"
-      "mov z26.s, #0x0\n"
-      "mov z27.s, #0x0\n"
-      "31:"  // Height 3: setup done
-      "mov x26, #0x0\n"
-      "32:"  // Height 3: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 33f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "ldr x22, [x20, #0x10]\n"
-      "cbnz x26, 34f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "b 34f\n"
-      "33:"  // Height 3: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "34:"  // Height 3: input setup done
-      "cmp x25, #0x10\n"
-      "ble 37f\n"
-      "35:"  // Height 3: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z3.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z31.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1rqb { z1.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1rqb { z2.b }, p0/Z, [x22]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x22, x22, #0x10\n"
-      ".inst 0x44a01cb0  // sudot z16.s, z5.b, z0.b[0]\n"
-      ".inst 0x44a11cb4  // sudot z20.s, z5.b, z1.b[0]\n"
-      ".inst 0x44a01fb1  // sudot z17.s, z29.b, z0.b[0]\n"
-      ".inst 0x44a11fb5  // sudot z21.s, z29.b, z1.b[0]\n"
-      ".inst 0x44a01c92  // sudot z18.s, z4.b, z0.b[0]\n"
-      ".inst 0x44a21cb8  // sudot z24.s, z5.b, z2.b[0]\n"
-      ".inst 0x44a21fb9  // sudot z25.s, z29.b, z2.b[0]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      ".inst 0x44a11c96  // sudot z22.s, z4.b, z1.b[0]\n"
-      ".inst 0x44a21c9a  // sudot z26.s, z4.b, z2.b[0]\n"
-      ".inst 0x44a01f93  // sudot z19.s, z28.b, z0.b[0]\n"
-      ".inst 0x44a11f97  // sudot z23.s, z28.b, z1.b[0]\n"
-      ".inst 0x44a21f9b  // sudot z27.s, z28.b, z2.b[0]\n"
-      ".inst 0x44a81c70  // sudot z16.s, z3.b, z0.b[1]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      "ld1b { z5.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      ".inst 0x44a91c74  // sudot z20.s, z3.b, z1.b[1]\n"
-      ".inst 0x44aa1c78  // sudot z24.s, z3.b, z2.b[1]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z3.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      ".inst 0x44a81ff1  // sudot z17.s, z31.b, z0.b[1]\n"
-      ".inst 0x44a91ff5  // sudot z21.s, z31.b, z1.b[1]\n"
-      ".inst 0x44aa1ff9  // sudot z25.s, z31.b, z2.b[1]\n"
-      ".inst 0x44a81fd2  // sudot z18.s, z30.b, z0.b[1]\n"
-      "ld1b { z31.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x44a91fd6  // sudot z22.s, z30.b, z1.b[1]\n"
-      ".inst 0x44aa1fda  // sudot z26.s, z30.b, z2.b[1]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      ".inst 0x44a81fb3  // sudot z19.s, z29.b, z0.b[1]\n"
-      ".inst 0x44a91fb7  // sudot z23.s, z29.b, z1.b[1]\n"
-      ".inst 0x44aa1fbb  // sudot z27.s, z29.b, z2.b[1]\n"
-      ".inst 0x44b01f90  // sudot z16.s, z28.b, z0.b[2]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      ".inst 0x44b11f94  // sudot z20.s, z28.b, z1.b[2]\n"
-      ".inst 0x44b21f98  // sudot z24.s, z28.b, z2.b[2]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x44b01cb1  // sudot z17.s, z5.b, z0.b[2]\n"
-      ".inst 0x44b11cb5  // sudot z21.s, z5.b, z1.b[2]\n"
-      ".inst 0x44b21cb9  // sudot z25.s, z5.b, z2.b[2]\n"
-      ".inst 0x44b01c92  // sudot z18.s, z4.b, z0.b[2]\n"
-      ".inst 0x44b11c96  // sudot z22.s, z4.b, z1.b[2]\n"
-      ".inst 0x44b21c9a  // sudot z26.s, z4.b, z2.b[2]\n"
-      ".inst 0x44b01c73  // sudot z19.s, z3.b, z0.b[2]\n"
-      ".inst 0x44b11c77  // sudot z23.s, z3.b, z1.b[2]\n"
-      ".inst 0x44b21c7b  // sudot z27.s, z3.b, z2.b[2]\n"
-      ".inst 0x44b81ff0  // sudot z16.s, z31.b, z0.b[3]\n"
-      ".inst 0x44b91ff4  // sudot z20.s, z31.b, z1.b[3]\n"
-      ".inst 0x44ba1ff8  // sudot z24.s, z31.b, z2.b[3]\n"
-      ".inst 0x44b81fd1  // sudot z17.s, z30.b, z0.b[3]\n"
-      ".inst 0x44b91fd5  // sudot z21.s, z30.b, z1.b[3]\n"
-      ".inst 0x44ba1fd9  // sudot z25.s, z30.b, z2.b[3]\n"
-      ".inst 0x44b81fb2  // sudot z18.s, z29.b, z0.b[3]\n"
-      ".inst 0x44b91fb6  // sudot z22.s, z29.b, z1.b[3]\n"
-      ".inst 0x44ba1fba  // sudot z26.s, z29.b, z2.b[3]\n"
-      ".inst 0x44b81f93  // sudot z19.s, z28.b, z0.b[3]\n"
-      ".inst 0x44b91f97  // sudot z23.s, z28.b, z1.b[3]\n"
-      ".inst 0x44ba1f9b  // sudot z27.s, z28.b, z2.b[3]\n"
-      "tbnz %x[flags], #31, 36f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z12.s, z1.b, z15.b\n"
-      "udot z13.s, z2.b, z15.b\n"
-      "36:"  // Height 3: Multiply loop: unique 5: skip row sum
-      "sub x25, x25, #0x10\n"
-      "cmp x25, #0x10\n"
-      "bgt 35b\n"
-      "37:"  // Height 3: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1rqb { z1.b }, p0/Z, [x23]\n"
-      "ld1rqb { z2.b }, p0/Z, [x22]\n"
-      ".inst 0x44a01ff0  // sudot z16.s, z31.b, z0.b[0]\n"
-      ".inst 0x44a11ff4  // sudot z20.s, z31.b, z1.b[0]\n"
-      ".inst 0x44a01fd1  // sudot z17.s, z30.b, z0.b[0]\n"
-      ".inst 0x44a11fd5  // sudot z21.s, z30.b, z1.b[0]\n"
-      ".inst 0x44a01fb2  // sudot z18.s, z29.b, z0.b[0]\n"
-      ".inst 0x44a11fb6  // sudot z22.s, z29.b, z1.b[0]\n"
-      ".inst 0x44a21ff8  // sudot z24.s, z31.b, z2.b[0]\n"
-      ".inst 0x44a21fd9  // sudot z25.s, z30.b, z2.b[0]\n"
-      ".inst 0x44a21fba  // sudot z26.s, z29.b, z2.b[0]\n"
-      ".inst 0x44a01f93  // sudot z19.s, z28.b, z0.b[0]\n"
-      ".inst 0x44a11f97  // sudot z23.s, z28.b, z1.b[0]\n"
-      ".inst 0x44a21f9b  // sudot z27.s, z28.b, z2.b[0]\n"
-      "ble 38f\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44a81ff0  // sudot z16.s, z31.b, z0.b[1]\n"
-      ".inst 0x44a91ff4  // sudot z20.s, z31.b, z1.b[1]\n"
-      ".inst 0x44aa1ff8  // sudot z24.s, z31.b, z2.b[1]\n"
-      ".inst 0x44a81fd1  // sudot z17.s, z30.b, z0.b[1]\n"
-      ".inst 0x44a91fd5  // sudot z21.s, z30.b, z1.b[1]\n"
-      ".inst 0x44aa1fd9  // sudot z25.s, z30.b, z2.b[1]\n"
-      ".inst 0x44a81fb2  // sudot z18.s, z29.b, z0.b[1]\n"
-      ".inst 0x44a91fb6  // sudot z22.s, z29.b, z1.b[1]\n"
-      ".inst 0x44aa1fba  // sudot z26.s, z29.b, z2.b[1]\n"
-      ".inst 0x44a81f93  // sudot z19.s, z28.b, z0.b[1]\n"
-      ".inst 0x44a91f97  // sudot z23.s, z28.b, z1.b[1]\n"
-      ".inst 0x44aa1f9b  // sudot z27.s, z28.b, z2.b[1]\n"
-      "ble 38f\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44b01ff0  // sudot z16.s, z31.b, z0.b[2]\n"
-      ".inst 0x44b11ff4  // sudot z20.s, z31.b, z1.b[2]\n"
-      ".inst 0x44b21ff8  // sudot z24.s, z31.b, z2.b[2]\n"
-      ".inst 0x44b01fd1  // sudot z17.s, z30.b, z0.b[2]\n"
-      ".inst 0x44b11fd5  // sudot z21.s, z30.b, z1.b[2]\n"
-      ".inst 0x44b21fd9  // sudot z25.s, z30.b, z2.b[2]\n"
-      ".inst 0x44b01fb2  // sudot z18.s, z29.b, z0.b[2]\n"
-      ".inst 0x44b11fb6  // sudot z22.s, z29.b, z1.b[2]\n"
-      ".inst 0x44b21fba  // sudot z26.s, z29.b, z2.b[2]\n"
-      ".inst 0x44b01f93  // sudot z19.s, z28.b, z0.b[2]\n"
-      ".inst 0x44b11f97  // sudot z23.s, z28.b, z1.b[2]\n"
-      ".inst 0x44b21f9b  // sudot z27.s, z28.b, z2.b[2]\n"
-      "ble 38f\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44b81ff0  // sudot z16.s, z31.b, z0.b[3]\n"
-      ".inst 0x44b91ff4  // sudot z20.s, z31.b, z1.b[3]\n"
-      ".inst 0x44ba1ff8  // sudot z24.s, z31.b, z2.b[3]\n"
-      ".inst 0x44b81fd1  // sudot z17.s, z30.b, z0.b[3]\n"
-      ".inst 0x44b91fd5  // sudot z21.s, z30.b, z1.b[3]\n"
-      ".inst 0x44ba1fd9  // sudot z25.s, z30.b, z2.b[3]\n"
-      ".inst 0x44b81fb2  // sudot z18.s, z29.b, z0.b[3]\n"
-      ".inst 0x44b91fb6  // sudot z22.s, z29.b, z1.b[3]\n"
-      ".inst 0x44ba1fba  // sudot z26.s, z29.b, z2.b[3]\n"
-      ".inst 0x44b81f93  // sudot z19.s, z28.b, z0.b[3]\n"
-      ".inst 0x44b91f97  // sudot z23.s, z28.b, z1.b[3]\n"
-      ".inst 0x44ba1f9b  // sudot z27.s, z28.b, z2.b[3]\n"
-      "38:"  // Height 3: Multiply loop: multiply skip
-      "tbnz %x[flags], #31, 39f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z12.s, z1.b, z15.b\n"
-      "udot z13.s, z2.b, z15.b\n"
-      "39:"  // Height 3: Multiply loop: unique 6: skip row sum
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 32b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
-      "tbnz %x[flags], #31, 40f\n"
-      "mov x21, #0x4\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z28.s, p2/M, z28.s\n"
-      "saddv d11, p0, z11.s\n"
-      "saddv d12, p0, z12.s\n"
-      "saddv d13, p0, z13.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "mov z12.s, z12.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z28.s\n"
-      "mul z12.s, p2/M, z12.s, z28.s\n"
-      "mov z13.s, z13.s[0]\n"
-      "mul z13.s, p2/M, z13.s, z28.s\n"
-      "40:"  // Height 3: skip row sum fixup
-      "add z16.s, z16.s, z11.s\n"
-      "add z17.s, z17.s, z11.s\n"
-      "ld1w { z0.s }, p2/Z, [x10]\n"
-      "ld1w { z31.s }, p2/Z, [x10, #1, MUL VL]\n"
-      "add z18.s, z18.s, z11.s\n"
-      "add z19.s, z19.s, z11.s\n"
-      "ld1w { z30.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z29.s }, p2/Z, [x10, #3, MUL VL]\n"
-      "add z20.s, z20.s, z12.s\n"
-      "add z21.s, z21.s, z12.s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "add z22.s, z22.s, z12.s\n"
-      "add z23.s, z23.s, z12.s\n"
-      "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add z24.s, z24.s, z13.s\n"
-      "add z25.s, z25.s, z13.s\n"
-      "addvl x10, x10, #4\n"
-      "add z26.s, z26.s, z13.s\n"
-      "add z27.s, z27.s, z13.s\n"
-      "add z16.s, z16.s, z0.s\n"
-      "add z17.s, z17.s, z31.s\n"
-      "add z18.s, z18.s, z30.s\n"
-      "add z19.s, z19.s, z29.s\n"
-      "add z20.s, z20.s, z0.s\n"
-      "add z21.s, z21.s, z31.s\n"
-      "add z22.s, z22.s, z30.s\n"
-      "add z23.s, z23.s, z29.s\n"
-      "add z24.s, z24.s, z0.s\n"
-      "add z25.s, z25.s, z31.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z26.s, z26.s, z30.s\n"
-      "add z27.s, z27.s, z29.s\n"
-      ".inst 0x04bc7610  // sqrdmulh z16.s, z16.s, z28.s\n"
-      ".inst 0x04bc7631  // sqrdmulh z17.s, z17.s, z28.s\n"
-      ".inst 0x04bc7652  // sqrdmulh z18.s, z18.s, z28.s\n"
-      ".inst 0x04bc7673  // sqrdmulh z19.s, z19.s, z28.s\n"
-      ".inst 0x04bc7694  // sqrdmulh z20.s, z20.s, z28.s\n"
-      ".inst 0x04bc76b5  // sqrdmulh z21.s, z21.s, z28.s\n"
-      ".inst 0x04bc76d6  // sqrdmulh z22.s, z22.s, z28.s\n"
-      ".inst 0x04bc76f7  // sqrdmulh z23.s, z23.s, z28.s\n"
-      ".inst 0x04bc7718  // sqrdmulh z24.s, z24.s, z28.s\n"
-      ".inst 0x04bc7739  // sqrdmulh z25.s, z25.s, z28.s\n"
-      ".inst 0x04bc775a  // sqrdmulh z26.s, z26.s, z28.s\n"
-      ".inst 0x04bc777b  // sqrdmulh z27.s, z27.s, z28.s\n"
-      "tbz %x[flags], #5, 41f\n"
-      "and z1.d, z16.d, z0.d\n"
-      "and z31.d, z17.d, z0.d\n"
-      "and z30.d, z18.d, z0.d\n"
-      "and z29.d, z19.d, z0.d\n"
-      "and z28.d, z20.d, z0.d\n"
-      "and z3.d, z21.d, z0.d\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "asr z31.s, z31.s, #0x1f\n"
-      "asr z30.s, z30.s, #0x1f\n"
-      "asr z29.s, z29.s, #0x1f\n"
-      "asr z28.s, z28.s, #0x1f\n"
-      "and z2.d, z22.d, z0.d\n"
-      "sqadd z16.s, z16.s, z1.s\n"
-      "sqadd z17.s, z17.s, z31.s\n"
-      "sqadd z18.s, z18.s, z30.s\n"
-      "sqadd z19.s, z19.s, z29.s\n"
-      "sqadd z20.s, z20.s, z28.s\n"
-      "and z1.d, z23.d, z0.d\n"
-      "and z31.d, z24.d, z0.d\n"
-      "and z30.d, z25.d, z0.d\n"
-      "and z29.d, z26.d, z0.d\n"
-      "and z28.d, z27.d, z0.d\n"
-      "asr z3.s, z3.s, #0x1f\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "asr z31.s, z31.s, #0x1f\n"
-      "asr z30.s, z30.s, #0x1f\n"
-      "asr z29.s, z29.s, #0x1f\n"
-      "asr z28.s, z28.s, #0x1f\n"
-      "sqadd z21.s, z21.s, z3.s\n"
-      "sqadd z22.s, z22.s, z2.s\n"
-      "sqadd z23.s, z23.s, z1.s\n"
-      "sqadd z24.s, z24.s, z31.s\n"
-      "sqadd z25.s, z25.s, z30.s\n"
-      "sqadd z26.s, z26.s, z29.s\n"
-      "sqadd z27.s, z27.s, z28.s\n"
-      "41:"  // Height 3: no shift correction
-      "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z30.s }, p2/Z, [x20]\n"
-      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
-      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
-      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
-      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z30.s\n"
-      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z30.s\n"
-      "add z18.s, z18.s, z30.s\n"
-      ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
-      ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z30.s\n"
-      "add z20.s, z20.s, z30.s\n"
-      ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "add z21.s, z21.s, z30.s\n"
-      "add z22.s, z22.s, z30.s\n"
-      "ld1rw { z29.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z30.s\n"
-      "add z24.s, z24.s, z30.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z30.s\n"
-      "add z26.s, z26.s, z30.s\n"
-      "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z30.s\n"
-      "smin z16.s, p2/M, z16.s, z29.s\n"
-      "smin z17.s, p2/M, z17.s, z29.s\n"
-      "smin z18.s, p2/M, z18.s, z29.s\n"
-      "smin z19.s, p2/M, z19.s, z29.s\n"
-      "smin z20.s, p2/M, z20.s, z29.s\n"
-      "smin z21.s, p2/M, z21.s, z29.s\n"
-      "smin z22.s, p2/M, z22.s, z29.s\n"
-      "smin z23.s, p2/M, z23.s, z29.s\n"
-      "smin z24.s, p2/M, z24.s, z29.s\n"
-      "smin z25.s, p2/M, z25.s, z29.s\n"
-      "smin z26.s, p2/M, z26.s, z29.s\n"
-      "smin z27.s, p2/M, z27.s, z29.s\n"
-      "smax z16.s, p2/M, z16.s, z28.s\n"
-      "smax z17.s, p2/M, z17.s, z28.s\n"
-      "smax z18.s, p2/M, z18.s, z28.s\n"
-      "smax z19.s, p2/M, z19.s, z28.s\n"
-      "smax z20.s, p2/M, z20.s, z28.s\n"
-      "smax z21.s, p2/M, z21.s, z28.s\n"
-      "smax z22.s, p2/M, z22.s, z28.s\n"
-      "smax z23.s, p2/M, z23.s, z28.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "smax z24.s, p2/M, z24.s, z28.s\n"
-      "smax z25.s, p2/M, z25.s, z28.s\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
-      "smax z26.s, p2/M, z26.s, z28.s\n"
-      "smax z27.s, p2/M, z27.s, z28.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
-      "uzp1 z18.h, z22.h, z23.h\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "uzp1 z20.b, z20.b, z18.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "st1b { z20.b }, p1, [x24]\n"
-      "st1b { z24.b }, p1, [x23]\n"
-      "42:"  // Height 3: Writeback done
-      "decw x9, ALL, MUL #4\n"
-      "cmp x9, XZR\n"
-      "bgt 30b\n"
-      "b 58f\n"
-      "43:"  // Height 4
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x20, #0x4\n"
-      "mov x10, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
-      "mov z15.b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "44:"  // Height 4: Column loop
-      "mov x20, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "mov z24.s, #0x0\n"
-      "mov z25.s, #0x0\n"
-      "mov z26.s, #0x0\n"
-      "mov z27.s, #0x0\n"
-      "mov z28.s, #0x0\n"
-      "mov z29.s, #0x0\n"
-      "mov z30.s, #0x0\n"
-      "mov z31.s, #0x0\n"
-      "45:"  // Height 4: setup done
-      "mov x26, #0x0\n"
-      "46:"  // Height 4: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 47f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "ldr x22, [x20, #0x10]\n"
-      "ldr x21, [x20, #0x18]\n"
-      "cbnz x26, 48f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "add x21, x21, x20\n"
-      "b 48f\n"
-      "47:"  // Height 4: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "add x21, x22, x21\n"
-      "48:"  // Height 4: input setup done
-      "cmp x25, #0x10\n"
-      "ble 51f\n"
-      "49:"  // Height 4: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1rqb { z1.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1rqb { z2.b }, p0/Z, [x22]\n"
-      "ld1rqb { z3.b }, p0/Z, [x21]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x44a01cb0  // sudot z16.s, z5.b, z0.b[0]\n"
-      ".inst 0x44a11cb4  // sudot z20.s, z5.b, z1.b[0]\n"
-      ".inst 0x44a01d51  // sudot z17.s, z10.b, z0.b[0]\n"
-      ".inst 0x44a11d55  // sudot z21.s, z10.b, z1.b[0]\n"
-      ".inst 0x44a21cb8  // sudot z24.s, z5.b, z2.b[0]\n"
-      ".inst 0x44a31cbc  // sudot z28.s, z5.b, z3.b[0]\n"
-      "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      ".inst 0x44a21d59  // sudot z25.s, z10.b, z2.b[0]\n"
-      ".inst 0x44a31d5d  // sudot z29.s, z10.b, z3.b[0]\n"
-      ".inst 0x44a01c92  // sudot z18.s, z4.b, z0.b[0]\n"
-      ".inst 0x44a11c96  // sudot z22.s, z4.b, z1.b[0]\n"
-      ".inst 0x44a21c9a  // sudot z26.s, z4.b, z2.b[0]\n"
-      ".inst 0x44a31c9e  // sudot z30.s, z4.b, z3.b[0]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      ".inst 0x44a01d33  // sudot z19.s, z9.b, z0.b[0]\n"
-      ".inst 0x44a11d37  // sudot z23.s, z9.b, z1.b[0]\n"
-      ".inst 0x44a21d3b  // sudot z27.s, z9.b, z2.b[0]\n"
-      ".inst 0x44a31d3f  // sudot z31.s, z9.b, z3.b[0]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x44a81d10  // sudot z16.s, z8.b, z0.b[1]\n"
-      ".inst 0x44a91d14  // sudot z20.s, z8.b, z1.b[1]\n"
-      ".inst 0x44aa1d18  // sudot z24.s, z8.b, z2.b[1]\n"
-      ".inst 0x44ab1d1c  // sudot z28.s, z8.b, z3.b[1]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      ".inst 0x44a81cf1  // sudot z17.s, z7.b, z0.b[1]\n"
-      ".inst 0x44a91cf5  // sudot z21.s, z7.b, z1.b[1]\n"
-      ".inst 0x44aa1cf9  // sudot z25.s, z7.b, z2.b[1]\n"
-      ".inst 0x44ab1cfd  // sudot z29.s, z7.b, z3.b[1]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x44a81cd2  // sudot z18.s, z6.b, z0.b[1]\n"
-      ".inst 0x44a91cd6  // sudot z22.s, z6.b, z1.b[1]\n"
-      ".inst 0x44aa1cda  // sudot z26.s, z6.b, z2.b[1]\n"
-      ".inst 0x44ab1cde  // sudot z30.s, z6.b, z3.b[1]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      ".inst 0x44a81cb3  // sudot z19.s, z5.b, z0.b[1]\n"
-      ".inst 0x44a91cb7  // sudot z23.s, z5.b, z1.b[1]\n"
-      ".inst 0x44aa1cbb  // sudot z27.s, z5.b, z2.b[1]\n"
-      ".inst 0x44ab1cbf  // sudot z31.s, z5.b, z3.b[1]\n"
-      "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      ".inst 0x44b01c90  // sudot z16.s, z4.b, z0.b[2]\n"
-      ".inst 0x44b11c94  // sudot z20.s, z4.b, z1.b[2]\n"
-      ".inst 0x44b21c98  // sudot z24.s, z4.b, z2.b[2]\n"
-      ".inst 0x44b31c9c  // sudot z28.s, z4.b, z3.b[2]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x44b01d51  // sudot z17.s, z10.b, z0.b[2]\n"
-      ".inst 0x44b11d55  // sudot z21.s, z10.b, z1.b[2]\n"
-      ".inst 0x44b21d59  // sudot z25.s, z10.b, z2.b[2]\n"
-      ".inst 0x44b31d5d  // sudot z29.s, z10.b, z3.b[2]\n"
-      ".inst 0x44b01d32  // sudot z18.s, z9.b, z0.b[2]\n"
-      ".inst 0x44b11d36  // sudot z22.s, z9.b, z1.b[2]\n"
-      ".inst 0x44b21d3a  // sudot z26.s, z9.b, z2.b[2]\n"
-      ".inst 0x44b31d3e  // sudot z30.s, z9.b, z3.b[2]\n"
-      ".inst 0x44b01d13  // sudot z19.s, z8.b, z0.b[2]\n"
-      ".inst 0x44b11d17  // sudot z23.s, z8.b, z1.b[2]\n"
-      ".inst 0x44b21d1b  // sudot z27.s, z8.b, z2.b[2]\n"
-      ".inst 0x44b31d1f  // sudot z31.s, z8.b, z3.b[2]\n"
-      ".inst 0x44b81cf0  // sudot z16.s, z7.b, z0.b[3]\n"
-      ".inst 0x44b91cf4  // sudot z20.s, z7.b, z1.b[3]\n"
-      ".inst 0x44ba1cf8  // sudot z24.s, z7.b, z2.b[3]\n"
-      ".inst 0x44bb1cfc  // sudot z28.s, z7.b, z3.b[3]\n"
-      ".inst 0x44b81cd1  // sudot z17.s, z6.b, z0.b[3]\n"
-      ".inst 0x44b91cd5  // sudot z21.s, z6.b, z1.b[3]\n"
-      ".inst 0x44ba1cd9  // sudot z25.s, z6.b, z2.b[3]\n"
-      ".inst 0x44bb1cdd  // sudot z29.s, z6.b, z3.b[3]\n"
-      ".inst 0x44b81cb2  // sudot z18.s, z5.b, z0.b[3]\n"
-      ".inst 0x44b91cb6  // sudot z22.s, z5.b, z1.b[3]\n"
-      ".inst 0x44ba1cba  // sudot z26.s, z5.b, z2.b[3]\n"
-      ".inst 0x44bb1cbe  // sudot z30.s, z5.b, z3.b[3]\n"
-      ".inst 0x44b81c93  // sudot z19.s, z4.b, z0.b[3]\n"
-      ".inst 0x44b91c97  // sudot z23.s, z4.b, z1.b[3]\n"
-      ".inst 0x44ba1c9b  // sudot z27.s, z4.b, z2.b[3]\n"
-      ".inst 0x44bb1c9f  // sudot z31.s, z4.b, z3.b[3]\n"
-      "tbnz %x[flags], #31, 50f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z12.s, z1.b, z15.b\n"
-      "udot z13.s, z2.b, z15.b\n"
-      "udot z14.s, z3.b, z15.b\n"
-      "50:"  // Height 4: Multiply loop: unique 7: skip row sum
-      "sub x25, x25, #0x10\n"
-      "cmp x25, #0x10\n"
-      "bgt 49b\n"
-      "51:"  // Height 4: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z7.b }, p2/Z, [x28]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "ld1rqb { z1.b }, p0/Z, [x23]\n"
-      "ld1rqb { z2.b }, p0/Z, [x22]\n"
-      "ld1rqb { z3.b }, p0/Z, [x21]\n"
-      ".inst 0x44a01cf0  // sudot z16.s, z7.b, z0.b[0]\n"
-      ".inst 0x44a11cf4  // sudot z20.s, z7.b, z1.b[0]\n"
-      ".inst 0x44a01cd1  // sudot z17.s, z6.b, z0.b[0]\n"
-      ".inst 0x44a11cd5  // sudot z21.s, z6.b, z1.b[0]\n"
-      ".inst 0x44a01cb2  // sudot z18.s, z5.b, z0.b[0]\n"
-      ".inst 0x44a11cb6  // sudot z22.s, z5.b, z1.b[0]\n"
-      ".inst 0x44a21cf8  // sudot z24.s, z7.b, z2.b[0]\n"
-      ".inst 0x44a31cfc  // sudot z28.s, z7.b, z3.b[0]\n"
-      ".inst 0x44a21cd9  // sudot z25.s, z6.b, z2.b[0]\n"
-      ".inst 0x44a31cdd  // sudot z29.s, z6.b, z3.b[0]\n"
-      ".inst 0x44a21cba  // sudot z26.s, z5.b, z2.b[0]\n"
-      ".inst 0x44a31cbe  // sudot z30.s, z5.b, z3.b[0]\n"
-      ".inst 0x44a01c93  // sudot z19.s, z4.b, z0.b[0]\n"
-      ".inst 0x44a11c97  // sudot z23.s, z4.b, z1.b[0]\n"
-      ".inst 0x44a21c9b  // sudot z27.s, z4.b, z2.b[0]\n"
-      ".inst 0x44a31c9f  // sudot z31.s, z4.b, z3.b[0]\n"
-      "ble 52f\n"
-      "ld1b { z7.b }, p2/Z, [x28]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44a81cf0  // sudot z16.s, z7.b, z0.b[1]\n"
-      ".inst 0x44a91cf4  // sudot z20.s, z7.b, z1.b[1]\n"
-      ".inst 0x44aa1cf8  // sudot z24.s, z7.b, z2.b[1]\n"
-      ".inst 0x44ab1cfc  // sudot z28.s, z7.b, z3.b[1]\n"
-      ".inst 0x44a81cd1  // sudot z17.s, z6.b, z0.b[1]\n"
-      ".inst 0x44a91cd5  // sudot z21.s, z6.b, z1.b[1]\n"
-      ".inst 0x44aa1cd9  // sudot z25.s, z6.b, z2.b[1]\n"
-      ".inst 0x44ab1cdd  // sudot z29.s, z6.b, z3.b[1]\n"
-      ".inst 0x44a81cb2  // sudot z18.s, z5.b, z0.b[1]\n"
-      ".inst 0x44a91cb6  // sudot z22.s, z5.b, z1.b[1]\n"
-      ".inst 0x44aa1cba  // sudot z26.s, z5.b, z2.b[1]\n"
-      ".inst 0x44ab1cbe  // sudot z30.s, z5.b, z3.b[1]\n"
-      ".inst 0x44a81c93  // sudot z19.s, z4.b, z0.b[1]\n"
-      ".inst 0x44a91c97  // sudot z23.s, z4.b, z1.b[1]\n"
-      ".inst 0x44aa1c9b  // sudot z27.s, z4.b, z2.b[1]\n"
-      ".inst 0x44ab1c9f  // sudot z31.s, z4.b, z3.b[1]\n"
-      "ble 52f\n"
-      "ld1b { z7.b }, p2/Z, [x28]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x4\n"
-      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44b01cf0  // sudot z16.s, z7.b, z0.b[2]\n"
-      ".inst 0x44b11cf4  // sudot z20.s, z7.b, z1.b[2]\n"
-      ".inst 0x44b21cf8  // sudot z24.s, z7.b, z2.b[2]\n"
-      ".inst 0x44b31cfc  // sudot z28.s, z7.b, z3.b[2]\n"
-      ".inst 0x44b01cd1  // sudot z17.s, z6.b, z0.b[2]\n"
-      ".inst 0x44b11cd5  // sudot z21.s, z6.b, z1.b[2]\n"
-      ".inst 0x44b21cd9  // sudot z25.s, z6.b, z2.b[2]\n"
-      ".inst 0x44b31cdd  // sudot z29.s, z6.b, z3.b[2]\n"
-      ".inst 0x44b01cb2  // sudot z18.s, z5.b, z0.b[2]\n"
-      ".inst 0x44b11cb6  // sudot z22.s, z5.b, z1.b[2]\n"
-      ".inst 0x44b21cba  // sudot z26.s, z5.b, z2.b[2]\n"
-      ".inst 0x44b31cbe  // sudot z30.s, z5.b, z3.b[2]\n"
-      ".inst 0x44b01c93  // sudot z19.s, z4.b, z0.b[2]\n"
-      ".inst 0x44b11c97  // sudot z23.s, z4.b, z1.b[2]\n"
-      ".inst 0x44b21c9b  // sudot z27.s, z4.b, z2.b[2]\n"
-      ".inst 0x44b31c9f  // sudot z31.s, z4.b, z3.b[2]\n"
-      "ble 52f\n"
-      "ld1b { z7.b }, p2/Z, [x28]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      ".inst 0x44b81cf0  // sudot z16.s, z7.b, z0.b[3]\n"
-      ".inst 0x44b91cf4  // sudot z20.s, z7.b, z1.b[3]\n"
-      ".inst 0x44ba1cf8  // sudot z24.s, z7.b, z2.b[3]\n"
-      ".inst 0x44bb1cfc  // sudot z28.s, z7.b, z3.b[3]\n"
-      ".inst 0x44b81cd1  // sudot z17.s, z6.b, z0.b[3]\n"
-      ".inst 0x44b91cd5  // sudot z21.s, z6.b, z1.b[3]\n"
-      ".inst 0x44ba1cd9  // sudot z25.s, z6.b, z2.b[3]\n"
-      ".inst 0x44bb1cdd  // sudot z29.s, z6.b, z3.b[3]\n"
-      ".inst 0x44b81cb2  // sudot z18.s, z5.b, z0.b[3]\n"
-      ".inst 0x44b91cb6  // sudot z22.s, z5.b, z1.b[3]\n"
-      ".inst 0x44ba1cba  // sudot z26.s, z5.b, z2.b[3]\n"
-      ".inst 0x44bb1cbe  // sudot z30.s, z5.b, z3.b[3]\n"
-      ".inst 0x44b81c93  // sudot z19.s, z4.b, z0.b[3]\n"
-      ".inst 0x44b91c97  // sudot z23.s, z4.b, z1.b[3]\n"
-      ".inst 0x44ba1c9b  // sudot z27.s, z4.b, z2.b[3]\n"
-      ".inst 0x44bb1c9f  // sudot z31.s, z4.b, z3.b[3]\n"
-      "52:"  // Height 4: Multiply loop: multiply skip
-      "tbnz %x[flags], #31, 53f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z12.s, z1.b, z15.b\n"
-      "udot z13.s, z2.b, z15.b\n"
-      "udot z14.s, z3.b, z15.b\n"
-      "53:"  // Height 4: Multiply loop: unique 8: skip row sum
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 46b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "add x24, x27, x20\n"
-      "add x23, x24, x20\n"
-      "add x22, x23, x20\n"
-      "tbnz %x[flags], #31, 54f\n"
-      "mov x21, #0x4\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "neg z0.s, p2/M, z0.s\n"
-      "saddv d11, p0, z11.s\n"
-      "saddv d12, p0, z12.s\n"
-      "saddv d13, p0, z13.s\n"
-      "saddv d14, p0, z14.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "mov z12.s, z12.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z0.s\n"
-      "mul z12.s, p2/M, z12.s, z0.s\n"
-      "mov z13.s, z13.s[0]\n"
-      "mov z14.s, z14.s[0]\n"
-      "mul z13.s, p2/M, z13.s, z0.s\n"
-      "mul z14.s, p2/M, z14.s, z0.s\n"
-      "54:"  // Height 4: skip row sum fixup
-      "add z16.s, z16.s, z11.s\n"
-      "add z17.s, z17.s, z11.s\n"
-      "ld1w { z4.s }, p2/Z, [x10]\n"
-      "ld1w { z0.s }, p2/Z, [x10, #1, MUL VL]\n"
-      "add z18.s, z18.s, z11.s\n"
-      "add z19.s, z19.s, z11.s\n"
-      "ld1w { z3.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z2.s }, p2/Z, [x10, #3, MUL VL]\n"
-      "add z20.s, z20.s, z12.s\n"
-      "add z21.s, z21.s, z12.s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "add z22.s, z22.s, z12.s\n"
-      "add z23.s, z23.s, z12.s\n"
-      "ld1rw { z1.s }, p2/Z, [x20]\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add z24.s, z24.s, z13.s\n"
-      "add z25.s, z25.s, z13.s\n"
-      "addvl x10, x10, #4\n"
-      "add z26.s, z26.s, z13.s\n"
-      "add z27.s, z27.s, z13.s\n"
-      "add z28.s, z28.s, z14.s\n"
-      "add z29.s, z29.s, z14.s\n"
-      "add z30.s, z30.s, z14.s\n"
-      "add z31.s, z31.s, z14.s\n"
-      "add z16.s, z16.s, z4.s\n"
-      "add z17.s, z17.s, z0.s\n"
-      "add z18.s, z18.s, z3.s\n"
-      "add z19.s, z19.s, z2.s\n"
-      "add z20.s, z20.s, z4.s\n"
-      "add z21.s, z21.s, z0.s\n"
-      "add z22.s, z22.s, z3.s\n"
-      "add z23.s, z23.s, z2.s\n"
-      "add z24.s, z24.s, z4.s\n"
-      "add z25.s, z25.s, z0.s\n"
-      "add z26.s, z26.s, z3.s\n"
-      "add z27.s, z27.s, z2.s\n"
-      "add z28.s, z28.s, z4.s\n"
-      "add z29.s, z29.s, z0.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z30.s, z30.s, z3.s\n"
-      "add z31.s, z31.s, z2.s\n"
-      ".inst 0x04a17610  // sqrdmulh z16.s, z16.s, z1.s\n"
-      ".inst 0x04a17631  // sqrdmulh z17.s, z17.s, z1.s\n"
-      ".inst 0x04a17652  // sqrdmulh z18.s, z18.s, z1.s\n"
-      ".inst 0x04a17673  // sqrdmulh z19.s, z19.s, z1.s\n"
-      ".inst 0x04a17694  // sqrdmulh z20.s, z20.s, z1.s\n"
-      ".inst 0x04a176b5  // sqrdmulh z21.s, z21.s, z1.s\n"
-      ".inst 0x04a176d6  // sqrdmulh z22.s, z22.s, z1.s\n"
-      ".inst 0x04a176f7  // sqrdmulh z23.s, z23.s, z1.s\n"
-      ".inst 0x04a17718  // sqrdmulh z24.s, z24.s, z1.s\n"
-      ".inst 0x04a17739  // sqrdmulh z25.s, z25.s, z1.s\n"
-      ".inst 0x04a1775a  // sqrdmulh z26.s, z26.s, z1.s\n"
-      ".inst 0x04a1777b  // sqrdmulh z27.s, z27.s, z1.s\n"
-      ".inst 0x04a1779c  // sqrdmulh z28.s, z28.s, z1.s\n"
-      ".inst 0x04a177bd  // sqrdmulh z29.s, z29.s, z1.s\n"
-      ".inst 0x04a177de  // sqrdmulh z30.s, z30.s, z1.s\n"
-      ".inst 0x04a177ff  // sqrdmulh z31.s, z31.s, z1.s\n"
-      "tbz %x[flags], #5, 55f\n"
-      "and z2.d, z16.d, z0.d\n"
-      "and z1.d, z17.d, z0.d\n"
-      "and z7.d, z18.d, z0.d\n"
-      "and z6.d, z19.d, z0.d\n"
-      "and z5.d, z20.d, z0.d\n"
-      "and z4.d, z21.d, z0.d\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "and z3.d, z22.d, z0.d\n"
-      "asr z7.s, z7.s, #0x1f\n"
-      "asr z6.s, z6.s, #0x1f\n"
-      "asr z5.s, z5.s, #0x1f\n"
-      "sqadd z16.s, z16.s, z2.s\n"
-      "sqadd z17.s, z17.s, z1.s\n"
-      "and z2.d, z23.d, z0.d\n"
-      "and z1.d, z24.d, z0.d\n"
-      "asr z4.s, z4.s, #0x1f\n"
-      "asr z3.s, z3.s, #0x1f\n"
-      "sqadd z18.s, z18.s, z7.s\n"
-      "sqadd z19.s, z19.s, z6.s\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "sqadd z20.s, z20.s, z5.s\n"
-      "sqadd z21.s, z21.s, z4.s\n"
-      "sqadd z22.s, z22.s, z3.s\n"
-      "and z7.d, z25.d, z0.d\n"
-      "sqadd z23.s, z23.s, z2.s\n"
-      "sqadd z24.s, z24.s, z1.s\n"
-      "and z6.d, z26.d, z0.d\n"
-      "and z5.d, z27.d, z0.d\n"
-      "and z4.d, z28.d, z0.d\n"
-      "and z3.d, z29.d, z0.d\n"
-      "and z2.d, z30.d, z0.d\n"
-      "and z1.d, z31.d, z0.d\n"
-      "asr z7.s, z7.s, #0x1f\n"
-      "asr z6.s, z6.s, #0x1f\n"
-      "asr z5.s, z5.s, #0x1f\n"
-      "asr z4.s, z4.s, #0x1f\n"
-      "asr z3.s, z3.s, #0x1f\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "sqadd z25.s, z25.s, z7.s\n"
-      "sqadd z26.s, z26.s, z6.s\n"
-      "sqadd z27.s, z27.s, z5.s\n"
-      "sqadd z28.s, z28.s, z4.s\n"
-      "sqadd z29.s, z29.s, z3.s\n"
-      "sqadd z30.s, z30.s, z2.s\n"
-      "sqadd z31.s, z31.s, z1.s\n"
-      "55:"  // Height 4: no shift correction
-      "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      "ld1rw { z2.s }, p2/Z, [x20]\n"
-      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
-      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
-      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
-      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      "add z16.s, z16.s, z2.s\n"
-      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z2.s\n"
-      "add z18.s, z18.s, z2.s\n"
-      ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
-      ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z2.s\n"
-      "add z20.s, z20.s, z2.s\n"
-      ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
-      ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
-      "add z21.s, z21.s, z2.s\n"
-      "add z22.s, z22.s, z2.s\n"
-      ".inst 0x4482881d  // srshl z29.s, p2/M, z29.s, z0.s\n"
-      ".inst 0x4482881e  // srshl z30.s, p2/M, z30.s, z0.s\n"
-      "add z23.s, z23.s, z2.s\n"
-      "add z24.s, z24.s, z2.s\n"
-      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "add z25.s, z25.s, z2.s\n"
-      "add z26.s, z26.s, z2.s\n"
-      "ld1rw { z1.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z2.s\n"
-      "add z28.s, z28.s, z2.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z29.s, z29.s, z2.s\n"
-      "add z30.s, z30.s, z2.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z31.s, z31.s, z2.s\n"
-      "smin z16.s, p2/M, z16.s, z1.s\n"
-      "smin z17.s, p2/M, z17.s, z1.s\n"
-      "smin z18.s, p2/M, z18.s, z1.s\n"
-      "smin z19.s, p2/M, z19.s, z1.s\n"
-      "smin z20.s, p2/M, z20.s, z1.s\n"
-      "smin z21.s, p2/M, z21.s, z1.s\n"
-      "smin z22.s, p2/M, z22.s, z1.s\n"
-      "smin z23.s, p2/M, z23.s, z1.s\n"
-      "smin z24.s, p2/M, z24.s, z1.s\n"
-      "smin z25.s, p2/M, z25.s, z1.s\n"
-      "smin z26.s, p2/M, z26.s, z1.s\n"
-      "smin z27.s, p2/M, z27.s, z1.s\n"
-      "smin z28.s, p2/M, z28.s, z1.s\n"
-      "smin z29.s, p2/M, z29.s, z1.s\n"
-      "smin z30.s, p2/M, z30.s, z1.s\n"
-      "smin z31.s, p2/M, z31.s, z1.s\n"
-      "smax z16.s, p2/M, z16.s, z0.s\n"
-      "smax z17.s, p2/M, z17.s, z0.s\n"
-      "smax z18.s, p2/M, z18.s, z0.s\n"
-      "smax z19.s, p2/M, z19.s, z0.s\n"
-      "smax z20.s, p2/M, z20.s, z0.s\n"
-      "smax z21.s, p2/M, z21.s, z0.s\n"
-      "smax z22.s, p2/M, z22.s, z0.s\n"
-      "smax z23.s, p2/M, z23.s, z0.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "smax z24.s, p2/M, z24.s, z0.s\n"
-      "smax z25.s, p2/M, z25.s, z0.s\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
-      "smax z26.s, p2/M, z26.s, z0.s\n"
-      "smax z27.s, p2/M, z27.s, z0.s\n"
-      "uzp1 z20.h, z20.h, z21.h\n"
-      "smax z28.s, p2/M, z28.s, z0.s\n"
-      "smax z29.s, p2/M, z29.s, z0.s\n"
-      "uzp1 z17.h, z22.h, z23.h\n"
-      "smax z30.s, p2/M, z30.s, z0.s\n"
-      "smax z31.s, p2/M, z31.s, z0.s\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "uzp1 z18.h, z26.h, z27.h\n"
-      "uzp1 z28.h, z28.h, z29.h\n"
-      "uzp1 z20.b, z20.b, z17.b\n"
-      "uzp1 z17.h, z30.h, z31.h\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z18.b\n"
-      "uzp1 z28.b, z28.b, z17.b\n"
-      "st1b { z20.b }, p1, [x24]\n"
-      "st1b { z24.b }, p1, [x23]\n"
-      "st1b { z28.b }, p1, [x22]\n"
-      "56:"  // Height 4: Writeback done
-      "decw x9, ALL, MUL #4\n"
-      "cmp x9, XZR\n"
-      "bgt 44b\n"
-      "subs %x[M], %x[M], #0x4\n"
-      "beq 58f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "tbz %x[flags], #3, 57f\n"
-      "add x21, x21, #0x4\n"
-      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "b 1b\n"
-      "57:"  // Update direct input
-      "mov x20, #0x4\n"
-      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
-      "b 1b\n"
-      "58:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
-      : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
-    );
-}
-
-} // namespace arm_gemm
-#endif  // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_mmla_4x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_mmla_4x4VL.hpp
deleted file mode 100644
index b073731751..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_mmla_4x4VL.hpp
+++ /dev/null
@@ -1,99 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include "../std_transforms_sve.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    unsigned int, const unsigned int *, \
-    IndirectInputArg<uint8_t>, \
-    size_t, size_t, \
-    const int8_t *, \
-    IndirectOutputArg<uint8_t>, \
-    const Requantize32 *, const int32_t *, unsigned int
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void sve_hybrid_u8s8qa_mmla_4x4VL( ARGLIST );
-
-class cls_sve_hybrid_u8s8qa_mmla_4x4VL
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef uint8_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 4;
-    }
-
-    static unsigned int out_width()
-    {
-        return get_vector_length<int32_t>() * 4;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 8;
-    }
-
-    static constexpr bool supports_accumulate()
-    {
-        return false;
-    }
-
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 4, 8, 8> transforms = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 47.37 };
-                case CPUModel::A510:
-                    return { 20.88 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=sve_hybrid_u8s8qa_mmla_4x4VL;
-    cls_sve_hybrid_u8s8qa_mmla_4x4VL(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_mmla_4x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_mmla_4x4VL/generic.cpp
deleted file mode 100644
index 01bdac2967..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8qa_mmla_4x4VL/generic.cpp
+++ /dev/null
@@ -1,1418 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include "arm_gemm.hpp"
-#include "../../utils.hpp"
-
-#include <cassert>
-#include <cstdint>
-
-namespace arm_gemm {
-
-void sve_hybrid_u8s8qa_mmla_4x4VL (
-    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg,
-    size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<uint8_t> output_arg,
-    const Requantize32 *qp, const int32_t *col_bias, unsigned int
-)
-{
-    struct KernelArgs {
-        unsigned int num_strings = {};
-        const unsigned int *string_lengths = {};
-        size_t N = {};
-        const int8_t *B_ptr = {};
-        size_t output_offset = {};
-        size_t input_initial_col = {};
-        size_t input_offset = {};
-        void *output_ptr = {};
-    } ka;
-
-    unsigned long flags=0;
-    void *input_ptr;
-
-    if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
-        ka.output_offset=output_arg.indirect.offset;
-        flags |= 0x4;
-    } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
-        ka.output_offset=output_arg.direct.stride;
-    }
-
-    if (A_arg.is_indirect) {
-        input_ptr=(void *)(A_arg.indirect.ptr);
-        ka.input_offset=A_arg.indirect.start_row;
-        ka.input_initial_col=A_arg.indirect.start_col;
-        flags |= 0x8;
-    } else {
-        assert(num_strings==1);
-        input_ptr=(void *)(A_arg.direct.base);
-        ka.input_offset=A_arg.direct.stride;
-    }
-    ka.num_strings = num_strings;
-    ka.string_lengths = string_lengths;
-    ka.N = N;
-    ka.B_ptr = B_ptr;
-    if (qp->c_offset > qp->minval) {
-        flags |= 0x20;
-    }
-    __asm__ __volatile__(
-      "ptrue p2.b\n"
-      "1:"  // Row loop
-      "cmp %x[M], #0x4\n"
-      "bge 43f\n"
-      "cmp %x[M], #0x2\n"
-      "bgt 29f\n"
-      "beq 15f\n"
-      "mov x10, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z15.b, #0x1\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "2:"  // Height 1: Column loop
-      "mov x20, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "3:"  // Height 1: setup done
-      "mov x26, #0x0\n"
-      "4:"  // Height 1: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 5f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "cbnz x26, 6f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "b 6f\n"
-      "5:"  // Height 1: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "6:"  // Height 1: input setup done
-      "cmp x25, #0x10\n"
-      "ble 9f\n"
-      "7:"  // Height 1: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z30.b }, p2/Z, [x28]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "add x24, x24, #0x10\n"
-      "trn1 z0.d, z1.d, z31.d\n"
-      ".inst 0x459e9810  // usmmla z16.s, z0.b, z30.b\n"
-      "ld1b { z8.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      "trn2 z1.d, z1.d, z31.d\n"
-      ".inst 0x459d9814  // usmmla z20.s, z0.b, z29.b\n"
-      ".inst 0x459c9811  // usmmla z17.s, z0.b, z28.b\n"
-      ".inst 0x459b9815  // usmmla z21.s, z0.b, z27.b\n"
-      ".inst 0x459a9812  // usmmla z18.s, z0.b, z26.b\n"
-      "ld1b { z31.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x45999816  // usmmla z22.s, z0.b, z25.b\n"
-      ".inst 0x45989813  // usmmla z19.s, z0.b, z24.b\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x45889817  // usmmla z23.s, z0.b, z8.b\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x459f9830  // usmmla z16.s, z1.b, z31.b\n"
-      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x459e9834  // usmmla z20.s, z1.b, z30.b\n"
-      ".inst 0x459d9831  // usmmla z17.s, z1.b, z29.b\n"
-      ".inst 0x459c9835  // usmmla z21.s, z1.b, z28.b\n"
-      ".inst 0x459b9832  // usmmla z18.s, z1.b, z27.b\n"
-      ".inst 0x459a9836  // usmmla z22.s, z1.b, z26.b\n"
-      ".inst 0x45999833  // usmmla z19.s, z1.b, z25.b\n"
-      ".inst 0x45989837  // usmmla z23.s, z1.b, z24.b\n"
-      "tbnz %x[flags], #31, 8f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z11.s, z1.b, z15.b\n"
-      "8:"  // Height 1: Multiply loop: unique 1: skip row sum
-      "sub x25, x25, #0x10\n"
-      "cmp x25, #0x10\n"
-      "bgt 7b\n"
-      "9:"  // Height 1: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z31.d\n"
-      ".inst 0x45989810  // usmmla z16.s, z0.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #8\n"
-      "trn2 z1.d, z1.d, z31.d\n"
-      ".inst 0x459e9814  // usmmla z20.s, z0.b, z30.b\n"
-      ".inst 0x459d9811  // usmmla z17.s, z0.b, z29.b\n"
-      ".inst 0x459c9815  // usmmla z21.s, z0.b, z28.b\n"
-      ".inst 0x459b9812  // usmmla z18.s, z0.b, z27.b\n"
-      ".inst 0x459a9816  // usmmla z22.s, z0.b, z26.b\n"
-      ".inst 0x45999813  // usmmla z19.s, z0.b, z25.b\n"
-      ".inst 0x45989817  // usmmla z23.s, z0.b, z24.b\n"
-      "ble 10f\n"
-      "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x45989830  // usmmla z16.s, z1.b, z24.b\n"
-      "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x459e9834  // usmmla z20.s, z1.b, z30.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x459d9831  // usmmla z17.s, z1.b, z29.b\n"
-      ".inst 0x459c9835  // usmmla z21.s, z1.b, z28.b\n"
-      ".inst 0x459b9832  // usmmla z18.s, z1.b, z27.b\n"
-      ".inst 0x459a9836  // usmmla z22.s, z1.b, z26.b\n"
-      ".inst 0x45999833  // usmmla z19.s, z1.b, z25.b\n"
-      ".inst 0x45989837  // usmmla z23.s, z1.b, z24.b\n"
-      "10:"  // Height 1: Multiply loop: multiply skip
-      "tbnz %x[flags], #31, 11f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z11.s, z1.b, z15.b\n"
-      "11:"  // Height 1: Multiply loop: unique 2: skip row sum
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 4b\n"
-      "uzp1 z16.d, z16.d, z20.d\n"
-      "uzp1 z17.d, z17.d, z21.d\n"
-      "uzp1 z18.d, z18.d, z22.d\n"
-      "uzp1 z19.d, z19.d, z23.d\n"
-      "mov z23.d, z16.d\n"
-      "tbnz %x[flags], #31, 12f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
-      "ld1rw { z9.s }, p2/Z, [x20]\n"
-      "neg z9.s, p2/M, z9.s\n"
-      "mov z11.s, z11.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z9.s\n"
-      "12:"  // Height 1: skip row sum fixup
-      "add z23.s, z23.s, z11.s\n"
-      "add z17.s, z17.s, z11.s\n"
-      "ld1w { z22.s }, p2/Z, [x10]\n"
-      "ld1w { z24.s }, p2/Z, [x10, #1, MUL VL]\n"
-      "add z18.s, z18.s, z11.s\n"
-      "add z19.s, z19.s, z11.s\n"
-      "ld1w { z21.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x10, #3, MUL VL]\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "add z23.s, z23.s, z22.s\n"
-      "add z17.s, z17.s, z24.s\n"
-      "ld1rw { z16.s }, p2/Z, [x20]\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "addvl x10, x10, #4\n"
-      "add z18.s, z18.s, z21.s\n"
-      "add z19.s, z19.s, z20.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      ".inst 0x04b076f7  // sqrdmulh z23.s, z23.s, z16.s\n"
-      ".inst 0x04b07631  // sqrdmulh z17.s, z17.s, z16.s\n"
-      ".inst 0x04b07652  // sqrdmulh z18.s, z18.s, z16.s\n"
-      ".inst 0x04b07673  // sqrdmulh z19.s, z19.s, z16.s\n"
-      "tbz %x[flags], #5, 13f\n"
-      "and z22.d, z23.d, z0.d\n"
-      "and z21.d, z17.d, z0.d\n"
-      "and z20.d, z18.d, z0.d\n"
-      "and z16.d, z19.d, z0.d\n"
-      "asr z22.s, z22.s, #0x1f\n"
-      "asr z21.s, z21.s, #0x1f\n"
-      "asr z20.s, z20.s, #0x1f\n"
-      "asr z16.s, z16.s, #0x1f\n"
-      "sqadd z23.s, z23.s, z22.s\n"
-      "sqadd z17.s, z17.s, z21.s\n"
-      "sqadd z18.s, z18.s, z20.s\n"
-      "sqadd z19.s, z19.s, z16.s\n"
-      "13:"  // Height 1: no shift correction
-      "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "ld1rw { z21.s }, p2/Z, [x20]\n"
-      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
-      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "ld1rw { z20.s }, p2/Z, [x20]\n"
-      "add z23.s, z23.s, z21.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z21.s\n"
-      "add z18.s, z18.s, z21.s\n"
-      "ld1rw { z16.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z21.s\n"
-      "smin z23.s, p2/M, z23.s, z20.s\n"
-      "smin z17.s, p2/M, z17.s, z20.s\n"
-      "smin z18.s, p2/M, z18.s, z20.s\n"
-      "smin z19.s, p2/M, z19.s, z20.s\n"
-      "smax z23.s, p2/M, z23.s, z16.s\n"
-      "smax z17.s, p2/M, z17.s, z16.s\n"
-      "smax z18.s, p2/M, z18.s, z16.s\n"
-      "smax z19.s, p2/M, z19.s, z16.s\n"
-      "uzp1 z23.h, z23.h, z17.h\n"
-      "uzp1 z16.h, z18.h, z19.h\n"
-      "uzp1 z23.b, z23.b, z16.b\n"
-      "st1b { z23.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "14:"  // Height 1: Writeback done
-      "decw x9, ALL, MUL #4\n"
-      "cmp x9, XZR\n"
-      "bgt 2b\n"
-      "b 58f\n"
-      "15:"  // Height 2
-      "mov x10, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "mov z15.b, #0x1\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "16:"  // Height 2: Column loop
-      "mov x20, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "17:"  // Height 2: setup done
-      "mov x26, #0x0\n"
-      "18:"  // Height 2: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 19f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "cbnz x26, 20f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "b 20f\n"
-      "19:"  // Height 2: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "20:"  // Height 2: input setup done
-      "cmp x25, #0x10\n"
-      "ble 23f\n"
-      "21:"  // Height 2: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z31.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z25.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1b { z24.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      ".inst 0x459f9810  // usmmla z16.s, z0.b, z31.b\n"
-      "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      ".inst 0x459e9814  // usmmla z20.s, z0.b, z30.b\n"
-      ".inst 0x459d9811  // usmmla z17.s, z0.b, z29.b\n"
-      ".inst 0x459c9815  // usmmla z21.s, z0.b, z28.b\n"
-      ".inst 0x459b9812  // usmmla z18.s, z0.b, z27.b\n"
-      ".inst 0x459a9816  // usmmla z22.s, z0.b, z26.b\n"
-      ".inst 0x45989813  // usmmla z19.s, z0.b, z24.b\n"
-      "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x45999817  // usmmla z23.s, z0.b, z25.b\n"
-      "ld1b { z30.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x45989830  // usmmla z16.s, z1.b, z24.b\n"
-      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x459e9834  // usmmla z20.s, z1.b, z30.b\n"
-      ".inst 0x459d9831  // usmmla z17.s, z1.b, z29.b\n"
-      ".inst 0x459c9835  // usmmla z21.s, z1.b, z28.b\n"
-      ".inst 0x459b9832  // usmmla z18.s, z1.b, z27.b\n"
-      ".inst 0x459a9836  // usmmla z22.s, z1.b, z26.b\n"
-      ".inst 0x45999833  // usmmla z19.s, z1.b, z25.b\n"
-      ".inst 0x45989837  // usmmla z23.s, z1.b, z24.b\n"
-      "tbnz %x[flags], #31, 22f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z11.s, z1.b, z15.b\n"
-      "22:"  // Height 2: Multiply loop: unique 3: skip row sum
-      "sub x25, x25, #0x10\n"
-      "cmp x25, #0x10\n"
-      "bgt 21b\n"
-      "23:"  // Height 2: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z29.b }, p2/Z, [x28]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "trn1 z0.d, z1.d, z24.d\n"
-      "trn2 z1.d, z1.d, z24.d\n"
-      ".inst 0x459d9810  // usmmla z16.s, z0.b, z29.b\n"
-      "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x459c9814  // usmmla z20.s, z0.b, z28.b\n"
-      ".inst 0x45849811  // usmmla z17.s, z0.b, z4.b\n"
-      ".inst 0x459b9815  // usmmla z21.s, z0.b, z27.b\n"
-      ".inst 0x459a9812  // usmmla z18.s, z0.b, z26.b\n"
-      ".inst 0x45869816  // usmmla z22.s, z0.b, z6.b\n"
-      ".inst 0x45999813  // usmmla z19.s, z0.b, z25.b\n"
-      ".inst 0x45989817  // usmmla z23.s, z0.b, z24.b\n"
-      "ble 24f\n"
-      "ld1b { z24.b }, p2/Z, [x28]\n"
-      "ld1b { z30.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z28.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x45989830  // usmmla z16.s, z1.b, z24.b\n"
-      "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x459e9834  // usmmla z20.s, z1.b, z30.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x459d9831  // usmmla z17.s, z1.b, z29.b\n"
-      ".inst 0x459c9835  // usmmla z21.s, z1.b, z28.b\n"
-      ".inst 0x459b9832  // usmmla z18.s, z1.b, z27.b\n"
-      ".inst 0x459a9836  // usmmla z22.s, z1.b, z26.b\n"
-      ".inst 0x45999833  // usmmla z19.s, z1.b, z25.b\n"
-      ".inst 0x45989837  // usmmla z23.s, z1.b, z24.b\n"
-      "24:"  // Height 2: Multiply loop: multiply skip
-      "tbnz %x[flags], #31, 25f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z11.s, z1.b, z15.b\n"
-      "25:"  // Height 2: Multiply loop: unique 4: skip row sum
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 18b\n"
-      "uzp1 z24.d, z16.d, z20.d\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp2 z16.d, z16.d, z20.d\n"
-      "uzp1 z20.d, z17.d, z21.d\n"
-      "uzp2 z17.d, z17.d, z21.d\n"
-      "uzp1 z21.d, z18.d, z22.d\n"
-      "uzp2 z18.d, z18.d, z22.d\n"
-      "uzp1 z22.d, z19.d, z23.d\n"
-      "uzp2 z19.d, z19.d, z23.d\n"
-      "add x23, x27, x20\n"
-      "mov z23.d, z24.d\n"
-      "tbnz %x[flags], #31, 26f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
-      "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "neg z24.s, p2/M, z24.s\n"
-      "mov z12.s, z11.s[3]\n"
-      "mov z11.s, z11.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z24.s\n"
-      "mul z12.s, p2/M, z12.s, z24.s\n"
-      "26:"  // Height 2: skip row sum fixup
-      "add z23.s, z23.s, z11.s\n"
-      "add z20.s, z20.s, z11.s\n"
-      "ld1w { z28.s }, p2/Z, [x10]\n"
-      "ld1w { z27.s }, p2/Z, [x10, #1, MUL VL]\n"
-      "add z21.s, z21.s, z11.s\n"
-      "add z22.s, z22.s, z11.s\n"
-      "ld1w { z26.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x10, #3, MUL VL]\n"
-      "add z16.s, z16.s, z12.s\n"
-      "add z17.s, z17.s, z12.s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "add z18.s, z18.s, z12.s\n"
-      "add z19.s, z19.s, z12.s\n"
-      "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add z23.s, z23.s, z28.s\n"
-      "add z20.s, z20.s, z27.s\n"
-      "addvl x10, x10, #4\n"
-      "add z21.s, z21.s, z26.s\n"
-      "add z22.s, z22.s, z25.s\n"
-      "add z16.s, z16.s, z28.s\n"
-      "add z17.s, z17.s, z27.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z18.s, z18.s, z26.s\n"
-      "add z19.s, z19.s, z25.s\n"
-      ".inst 0x04b876f7  // sqrdmulh z23.s, z23.s, z24.s\n"
-      ".inst 0x04b87694  // sqrdmulh z20.s, z20.s, z24.s\n"
-      ".inst 0x04b876b5  // sqrdmulh z21.s, z21.s, z24.s\n"
-      ".inst 0x04b876d6  // sqrdmulh z22.s, z22.s, z24.s\n"
-      ".inst 0x04b87610  // sqrdmulh z16.s, z16.s, z24.s\n"
-      ".inst 0x04b87631  // sqrdmulh z17.s, z17.s, z24.s\n"
-      ".inst 0x04b87652  // sqrdmulh z18.s, z18.s, z24.s\n"
-      ".inst 0x04b87673  // sqrdmulh z19.s, z19.s, z24.s\n"
-      "tbz %x[flags], #5, 27f\n"
-      "and z24.d, z23.d, z0.d\n"
-      "and z30.d, z20.d, z0.d\n"
-      "and z29.d, z21.d, z0.d\n"
-      "and z28.d, z22.d, z0.d\n"
-      "and z27.d, z16.d, z0.d\n"
-      "and z26.d, z17.d, z0.d\n"
-      "asr z24.s, z24.s, #0x1f\n"
-      "and z25.d, z18.d, z0.d\n"
-      "asr z30.s, z30.s, #0x1f\n"
-      "asr z29.s, z29.s, #0x1f\n"
-      "asr z28.s, z28.s, #0x1f\n"
-      "asr z27.s, z27.s, #0x1f\n"
-      "sqadd z23.s, z23.s, z24.s\n"
-      "and z24.d, z19.d, z0.d\n"
-      "asr z26.s, z26.s, #0x1f\n"
-      "asr z25.s, z25.s, #0x1f\n"
-      "sqadd z20.s, z20.s, z30.s\n"
-      "sqadd z21.s, z21.s, z29.s\n"
-      "asr z24.s, z24.s, #0x1f\n"
-      "sqadd z22.s, z22.s, z28.s\n"
-      "sqadd z16.s, z16.s, z27.s\n"
-      "sqadd z17.s, z17.s, z26.s\n"
-      "sqadd z18.s, z18.s, z25.s\n"
-      "sqadd z19.s, z19.s, z24.s\n"
-      "27:"  // Height 2: no shift correction
-      "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "ld1rw { z26.s }, p2/Z, [x20]\n"
-      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
-      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
-      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z23.s, z23.s, z26.s\n"
-      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "add z20.s, z20.s, z26.s\n"
-      "add z21.s, z21.s, z26.s\n"
-      "ld1rw { z25.s }, p2/Z, [x20]\n"
-      "add z22.s, z22.s, z26.s\n"
-      "add z16.s, z16.s, z26.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z17.s, z17.s, z26.s\n"
-      "add z18.s, z18.s, z26.s\n"
-      "ld1rw { z24.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z26.s\n"
-      "smin z23.s, p2/M, z23.s, z25.s\n"
-      "smin z20.s, p2/M, z20.s, z25.s\n"
-      "smin z21.s, p2/M, z21.s, z25.s\n"
-      "smin z22.s, p2/M, z22.s, z25.s\n"
-      "smin z16.s, p2/M, z16.s, z25.s\n"
-      "smin z17.s, p2/M, z17.s, z25.s\n"
-      "smin z18.s, p2/M, z18.s, z25.s\n"
-      "smin z19.s, p2/M, z19.s, z25.s\n"
-      "smax z23.s, p2/M, z23.s, z24.s\n"
-      "smax z20.s, p2/M, z20.s, z24.s\n"
-      "smax z21.s, p2/M, z21.s, z24.s\n"
-      "smax z22.s, p2/M, z22.s, z24.s\n"
-      "smax z16.s, p2/M, z16.s, z24.s\n"
-      "smax z17.s, p2/M, z17.s, z24.s\n"
-      "smax z18.s, p2/M, z18.s, z24.s\n"
-      "smax z19.s, p2/M, z19.s, z24.s\n"
-      "uzp1 z23.h, z23.h, z20.h\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
-      "uzp1 z23.b, z23.b, z20.b\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "st1b { z23.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "st1b { z16.b }, p1, [x23]\n"
-      "28:"  // Height 2: Writeback done
-      "decw x9, ALL, MUL #4\n"
-      "cmp x9, XZR\n"
-      "bgt 16b\n"
-      "b 58f\n"
-      "29:"  // Height 3
-      "mov x10, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "mov z13.s, #0x0\n"
-      "mov z15.b, #0x1\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "30:"  // Height 3: Column loop
-      "mov x20, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "mov z24.s, #0x0\n"
-      "mov z25.s, #0x0\n"
-      "mov z26.s, #0x0\n"
-      "mov z27.s, #0x0\n"
-      "mov z28.s, #0x0\n"
-      "mov z29.s, #0x0\n"
-      "mov z30.s, #0x0\n"
-      "mov z31.s, #0x0\n"
-      "31:"  // Height 3: setup done
-      "mov x26, #0x0\n"
-      "32:"  // Height 3: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 33f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "ldr x22, [x20, #0x10]\n"
-      "cbnz x26, 34f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "b 34f\n"
-      "33:"  // Height 3: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "34:"  // Height 3: input setup done
-      "cmp x25, #0x10\n"
-      "ble 37f\n"
-      "35:"  // Height 3: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x22]\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z0.d, z1.d, z2.d\n"
-      "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z6.d\n"
-      "trn2 z3.d, z3.d, z6.d\n"
-      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45859810  // usmmla z16.s, z0.b, z5.b\n"
-      ".inst 0x458a9814  // usmmla z20.s, z0.b, z10.b\n"
-      ".inst 0x45899811  // usmmla z17.s, z0.b, z9.b\n"
-      ".inst 0x45889815  // usmmla z21.s, z0.b, z8.b\n"
-      ".inst 0x45849812  // usmmla z18.s, z0.b, z4.b\n"
-      ".inst 0x45859858  // usmmla z24.s, z2.b, z5.b\n"
-      "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      ".inst 0x458a985c  // usmmla z28.s, z2.b, z10.b\n"
-      ".inst 0x45899859  // usmmla z25.s, z2.b, z9.b\n"
-      ".inst 0x4588985d  // usmmla z29.s, z2.b, z8.b\n"
-      ".inst 0x4584985a  // usmmla z26.s, z2.b, z4.b\n"
-      ".inst 0x45879816  // usmmla z22.s, z0.b, z7.b\n"
-      "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x4587985e  // usmmla z30.s, z2.b, z7.b\n"
-      ".inst 0x45869813  // usmmla z19.s, z0.b, z6.b\n"
-      "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x4586985b  // usmmla z27.s, z2.b, z6.b\n"
-      ".inst 0x45859817  // usmmla z23.s, z0.b, z5.b\n"
-      "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x4585985f  // usmmla z31.s, z2.b, z5.b\n"
-      ".inst 0x45849830  // usmmla z16.s, z1.b, z4.b\n"
-      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      ".inst 0x45849878  // usmmla z24.s, z3.b, z4.b\n"
-      "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x458a9834  // usmmla z20.s, z1.b, z10.b\n"
-      ".inst 0x458a987c  // usmmla z28.s, z3.b, z10.b\n"
-      ".inst 0x45899831  // usmmla z17.s, z1.b, z9.b\n"
-      ".inst 0x45899879  // usmmla z25.s, z3.b, z9.b\n"
-      ".inst 0x45889835  // usmmla z21.s, z1.b, z8.b\n"
-      ".inst 0x4588987d  // usmmla z29.s, z3.b, z8.b\n"
-      ".inst 0x45879832  // usmmla z18.s, z1.b, z7.b\n"
-      ".inst 0x4587987a  // usmmla z26.s, z3.b, z7.b\n"
-      ".inst 0x45869836  // usmmla z22.s, z1.b, z6.b\n"
-      ".inst 0x4586987e  // usmmla z30.s, z3.b, z6.b\n"
-      ".inst 0x45859833  // usmmla z19.s, z1.b, z5.b\n"
-      ".inst 0x4585987b  // usmmla z27.s, z3.b, z5.b\n"
-      ".inst 0x45849837  // usmmla z23.s, z1.b, z4.b\n"
-      ".inst 0x4584987f  // usmmla z31.s, z3.b, z4.b\n"
-      "tbnz %x[flags], #31, 36f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z13.s, z2.b, z15.b\n"
-      "udot z11.s, z1.b, z15.b\n"
-      "udot z13.s, z3.b, z15.b\n"
-      "36:"  // Height 3: Multiply loop: unique 5: skip row sum
-      "sub x25, x25, #0x10\n"
-      "cmp x25, #0x10\n"
-      "bgt 35b\n"
-      "37:"  // Height 3: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z3.b }, p0/Z, [x22]\n"
-      "trn1 z0.d, z1.d, z2.d\n"
-      "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z5.d\n"
-      "trn2 z3.d, z3.d, z5.d\n"
-      "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45849810  // usmmla z16.s, z0.b, z4.b\n"
-      ".inst 0x458a9814  // usmmla z20.s, z0.b, z10.b\n"
-      ".inst 0x45899811  // usmmla z17.s, z0.b, z9.b\n"
-      ".inst 0x45889815  // usmmla z21.s, z0.b, z8.b\n"
-      ".inst 0x45879812  // usmmla z18.s, z0.b, z7.b\n"
-      ".inst 0x45849858  // usmmla z24.s, z2.b, z4.b\n"
-      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x458a985c  // usmmla z28.s, z2.b, z10.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x45899859  // usmmla z25.s, z2.b, z9.b\n"
-      ".inst 0x4588985d  // usmmla z29.s, z2.b, z8.b\n"
-      ".inst 0x4587985a  // usmmla z26.s, z2.b, z7.b\n"
-      ".inst 0x45869816  // usmmla z22.s, z0.b, z6.b\n"
-      ".inst 0x4586985e  // usmmla z30.s, z2.b, z6.b\n"
-      ".inst 0x45859813  // usmmla z19.s, z0.b, z5.b\n"
-      ".inst 0x4585985b  // usmmla z27.s, z2.b, z5.b\n"
-      ".inst 0x45849817  // usmmla z23.s, z0.b, z4.b\n"
-      ".inst 0x4584985f  // usmmla z31.s, z2.b, z4.b\n"
-      "ble 38f\n"
-      "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x45849830  // usmmla z16.s, z1.b, z4.b\n"
-      ".inst 0x45849878  // usmmla z24.s, z3.b, z4.b\n"
-      "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x458a9834  // usmmla z20.s, z1.b, z10.b\n"
-      ".inst 0x458a987c  // usmmla z28.s, z3.b, z10.b\n"
-      ".inst 0x45899831  // usmmla z17.s, z1.b, z9.b\n"
-      ".inst 0x45899879  // usmmla z25.s, z3.b, z9.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x45889835  // usmmla z21.s, z1.b, z8.b\n"
-      ".inst 0x4588987d  // usmmla z29.s, z3.b, z8.b\n"
-      ".inst 0x45879832  // usmmla z18.s, z1.b, z7.b\n"
-      ".inst 0x4587987a  // usmmla z26.s, z3.b, z7.b\n"
-      ".inst 0x45869836  // usmmla z22.s, z1.b, z6.b\n"
-      ".inst 0x4586987e  // usmmla z30.s, z3.b, z6.b\n"
-      ".inst 0x45859833  // usmmla z19.s, z1.b, z5.b\n"
-      ".inst 0x4585987b  // usmmla z27.s, z3.b, z5.b\n"
-      ".inst 0x45849837  // usmmla z23.s, z1.b, z4.b\n"
-      ".inst 0x4584987f  // usmmla z31.s, z3.b, z4.b\n"
-      "38:"  // Height 3: Multiply loop: multiply skip
-      "tbnz %x[flags], #31, 39f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z13.s, z2.b, z15.b\n"
-      "udot z11.s, z1.b, z15.b\n"
-      "udot z13.s, z3.b, z15.b\n"
-      "39:"  // Height 3: Multiply loop: unique 6: skip row sum
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 32b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z0.d, z16.d, z20.d\n"
-      "uzp2 z16.d, z16.d, z20.d\n"
-      "uzp1 z20.d, z17.d, z21.d\n"
-      "uzp2 z17.d, z17.d, z21.d\n"
-      "uzp1 z21.d, z18.d, z22.d\n"
-      "uzp2 z18.d, z18.d, z22.d\n"
-      "add x23, x27, x20\n"
-      "uzp1 z22.d, z19.d, z23.d\n"
-      "uzp2 z19.d, z19.d, z23.d\n"
-      "add x22, x23, x20\n"
-      "uzp1 z24.d, z24.d, z28.d\n"
-      "uzp1 z25.d, z25.d, z29.d\n"
-      "uzp1 z26.d, z26.d, z30.d\n"
-      "uzp1 z27.d, z27.d, z31.d\n"
-      "mov z31.d, z0.d\n"
-      "tbnz %x[flags], #31, 40f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
-      ".inst 0x4491a9ad  // addp z13.s, p2/m, z13.s, z13.s\n"
-      "ld1rw { z23.s }, p2/Z, [x20]\n"
-      "neg z23.s, p2/M, z23.s\n"
-      "mov z12.s, z11.s[3]\n"
-      "mov z11.s, z11.s[0]\n"
-      "mov z13.s, z13.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z23.s\n"
-      "mul z12.s, p2/M, z12.s, z23.s\n"
-      "mul z13.s, p2/M, z13.s, z23.s\n"
-      "40:"  // Height 3: skip row sum fixup
-      "add z31.s, z31.s, z11.s\n"
-      "add z20.s, z20.s, z11.s\n"
-      "ld1w { z0.s }, p2/Z, [x10]\n"
-      "ld1w { z30.s }, p2/Z, [x10, #1, MUL VL]\n"
-      "add z21.s, z21.s, z11.s\n"
-      "add z22.s, z22.s, z11.s\n"
-      "ld1w { z29.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z28.s }, p2/Z, [x10, #3, MUL VL]\n"
-      "add z16.s, z16.s, z12.s\n"
-      "add z17.s, z17.s, z12.s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "add z18.s, z18.s, z12.s\n"
-      "add z19.s, z19.s, z12.s\n"
-      "ld1rw { z23.s }, p2/Z, [x20]\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add z24.s, z24.s, z13.s\n"
-      "add z25.s, z25.s, z13.s\n"
-      "addvl x10, x10, #4\n"
-      "add z26.s, z26.s, z13.s\n"
-      "add z27.s, z27.s, z13.s\n"
-      "add z31.s, z31.s, z0.s\n"
-      "add z20.s, z20.s, z30.s\n"
-      "add z21.s, z21.s, z29.s\n"
-      "add z22.s, z22.s, z28.s\n"
-      "add z16.s, z16.s, z0.s\n"
-      "add z17.s, z17.s, z30.s\n"
-      "add z18.s, z18.s, z29.s\n"
-      "add z19.s, z19.s, z28.s\n"
-      "add z24.s, z24.s, z0.s\n"
-      "add z25.s, z25.s, z30.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z26.s, z26.s, z29.s\n"
-      "add z27.s, z27.s, z28.s\n"
-      ".inst 0x04b777ff  // sqrdmulh z31.s, z31.s, z23.s\n"
-      ".inst 0x04b77694  // sqrdmulh z20.s, z20.s, z23.s\n"
-      ".inst 0x04b776b5  // sqrdmulh z21.s, z21.s, z23.s\n"
-      ".inst 0x04b776d6  // sqrdmulh z22.s, z22.s, z23.s\n"
-      ".inst 0x04b77610  // sqrdmulh z16.s, z16.s, z23.s\n"
-      ".inst 0x04b77631  // sqrdmulh z17.s, z17.s, z23.s\n"
-      ".inst 0x04b77652  // sqrdmulh z18.s, z18.s, z23.s\n"
-      ".inst 0x04b77673  // sqrdmulh z19.s, z19.s, z23.s\n"
-      ".inst 0x04b77718  // sqrdmulh z24.s, z24.s, z23.s\n"
-      ".inst 0x04b77739  // sqrdmulh z25.s, z25.s, z23.s\n"
-      ".inst 0x04b7775a  // sqrdmulh z26.s, z26.s, z23.s\n"
-      ".inst 0x04b7777b  // sqrdmulh z27.s, z27.s, z23.s\n"
-      "tbz %x[flags], #5, 41f\n"
-      "and z1.d, z31.d, z0.d\n"
-      "and z30.d, z20.d, z0.d\n"
-      "and z29.d, z21.d, z0.d\n"
-      "and z28.d, z22.d, z0.d\n"
-      "and z23.d, z16.d, z0.d\n"
-      "and z3.d, z17.d, z0.d\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "asr z30.s, z30.s, #0x1f\n"
-      "asr z29.s, z29.s, #0x1f\n"
-      "asr z28.s, z28.s, #0x1f\n"
-      "asr z23.s, z23.s, #0x1f\n"
-      "and z2.d, z18.d, z0.d\n"
-      "sqadd z31.s, z31.s, z1.s\n"
-      "sqadd z20.s, z20.s, z30.s\n"
-      "sqadd z21.s, z21.s, z29.s\n"
-      "sqadd z22.s, z22.s, z28.s\n"
-      "sqadd z16.s, z16.s, z23.s\n"
-      "and z1.d, z19.d, z0.d\n"
-      "and z30.d, z24.d, z0.d\n"
-      "and z29.d, z25.d, z0.d\n"
-      "and z28.d, z26.d, z0.d\n"
-      "and z23.d, z27.d, z0.d\n"
-      "asr z3.s, z3.s, #0x1f\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "asr z30.s, z30.s, #0x1f\n"
-      "asr z29.s, z29.s, #0x1f\n"
-      "asr z28.s, z28.s, #0x1f\n"
-      "asr z23.s, z23.s, #0x1f\n"
-      "sqadd z17.s, z17.s, z3.s\n"
-      "sqadd z18.s, z18.s, z2.s\n"
-      "sqadd z19.s, z19.s, z1.s\n"
-      "sqadd z24.s, z24.s, z30.s\n"
-      "sqadd z25.s, z25.s, z29.s\n"
-      "sqadd z26.s, z26.s, z28.s\n"
-      "sqadd z27.s, z27.s, z23.s\n"
-      "41:"  // Height 3: no shift correction
-      "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
-      "ld1rw { z29.s }, p2/Z, [x20]\n"
-      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
-      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
-      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z31.s, z31.s, z29.s\n"
-      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
-      ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z20.s, z20.s, z29.s\n"
-      "add z21.s, z21.s, z29.s\n"
-      ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
-      ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z22.s, z22.s, z29.s\n"
-      "add z16.s, z16.s, z29.s\n"
-      ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "add z17.s, z17.s, z29.s\n"
-      "add z18.s, z18.s, z29.s\n"
-      "ld1rw { z28.s }, p2/Z, [x20]\n"
-      "add z19.s, z19.s, z29.s\n"
-      "add z24.s, z24.s, z29.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z29.s\n"
-      "add z26.s, z26.s, z29.s\n"
-      "ld1rw { z23.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z29.s\n"
-      "smin z31.s, p2/M, z31.s, z28.s\n"
-      "smin z20.s, p2/M, z20.s, z28.s\n"
-      "smin z21.s, p2/M, z21.s, z28.s\n"
-      "smin z22.s, p2/M, z22.s, z28.s\n"
-      "smin z16.s, p2/M, z16.s, z28.s\n"
-      "smin z17.s, p2/M, z17.s, z28.s\n"
-      "smin z18.s, p2/M, z18.s, z28.s\n"
-      "smin z19.s, p2/M, z19.s, z28.s\n"
-      "smin z24.s, p2/M, z24.s, z28.s\n"
-      "smin z25.s, p2/M, z25.s, z28.s\n"
-      "smin z26.s, p2/M, z26.s, z28.s\n"
-      "smin z27.s, p2/M, z27.s, z28.s\n"
-      "smax z31.s, p2/M, z31.s, z23.s\n"
-      "smax z20.s, p2/M, z20.s, z23.s\n"
-      "smax z21.s, p2/M, z21.s, z23.s\n"
-      "smax z22.s, p2/M, z22.s, z23.s\n"
-      "smax z16.s, p2/M, z16.s, z23.s\n"
-      "smax z17.s, p2/M, z17.s, z23.s\n"
-      "smax z18.s, p2/M, z18.s, z23.s\n"
-      "smax z19.s, p2/M, z19.s, z23.s\n"
-      "uzp1 z31.h, z31.h, z20.h\n"
-      "smax z24.s, p2/M, z24.s, z23.s\n"
-      "smax z25.s, p2/M, z25.s, z23.s\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
-      "smax z26.s, p2/M, z26.s, z23.s\n"
-      "smax z27.s, p2/M, z27.s, z23.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "uzp1 z18.h, z18.h, z19.h\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z31.b, z31.b, z20.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "uzp1 z16.b, z16.b, z18.b\n"
-      "st1b { z31.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "st1b { z16.b }, p1, [x23]\n"
-      "st1b { z24.b }, p1, [x22]\n"
-      "42:"  // Height 3: Writeback done
-      "decw x9, ALL, MUL #4\n"
-      "cmp x9, XZR\n"
-      "bgt 30b\n"
-      "b 58f\n"
-      "43:"  // Height 4
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x20, #0x4\n"
-      "mov x10, %x[col_bias]\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "bic %x[flags], %x[flags], #0x80000000\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x27\n"
-      "mov z15.b, #0x1\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "44:"  // Height 4: Column loop
-      "mov x20, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "whilelt p1.b, x20, x9\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "mov z24.s, #0x0\n"
-      "mov z25.s, #0x0\n"
-      "mov z26.s, #0x0\n"
-      "mov z27.s, #0x0\n"
-      "mov z28.s, #0x0\n"
-      "mov z29.s, #0x0\n"
-      "mov z30.s, #0x0\n"
-      "mov z31.s, #0x0\n"
-      "45:"  // Height 4: setup done
-      "mov x26, #0x0\n"
-      "46:"  // Height 4: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w25, [x20, x26, LSL #0x2]\n"
-      "tbz %x[flags], #3, 47f\n"
-      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x24, [x20, #0x0]\n"
-      "ldr x23, [x20, #0x8]\n"
-      "ldr x22, [x20, #0x10]\n"
-      "ldr x21, [x20, #0x18]\n"
-      "cbnz x26, 48f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "add x21, x21, x20\n"
-      "b 48f\n"
-      "47:"  // Height 4: setup direct input
-      "mov x24, %x[input_ptr]\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "add x21, x22, x21\n"
-      "48:"  // Height 4: input setup done
-      "cmp x25, #0x10\n"
-      "ble 51f\n"
-      "49:"  // Height 4: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z5.b }, p2/Z, [x28]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x22]\n"
-      "ld1rqb { z6.b }, p0/Z, [x21]\n"
-      "add x22, x22, #0x10\n"
-      "add x21, x21, #0x10\n"
-      "trn1 z0.d, z1.d, z2.d\n"
-      "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z6.d\n"
-      "trn2 z3.d, z3.d, z6.d\n"
-      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45859810  // usmmla z16.s, z0.b, z5.b\n"
-      ".inst 0x45849814  // usmmla z20.s, z0.b, z4.b\n"
-      ".inst 0x45879811  // usmmla z17.s, z0.b, z7.b\n"
-      ".inst 0x45899815  // usmmla z21.s, z0.b, z9.b\n"
-      ".inst 0x45889812  // usmmla z18.s, z0.b, z8.b\n"
-      ".inst 0x45859858  // usmmla z24.s, z2.b, z5.b\n"
-      "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
-      "addvl x28, x28, #16\n"
-      ".inst 0x4584985c  // usmmla z28.s, z2.b, z4.b\n"
-      ".inst 0x45879859  // usmmla z25.s, z2.b, z7.b\n"
-      ".inst 0x4589985d  // usmmla z29.s, z2.b, z9.b\n"
-      ".inst 0x4588985a  // usmmla z26.s, z2.b, z8.b\n"
-      ".inst 0x458a9816  // usmmla z22.s, z0.b, z10.b\n"
-      "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
-      ".inst 0x458a985e  // usmmla z30.s, z2.b, z10.b\n"
-      ".inst 0x45869813  // usmmla z19.s, z0.b, z6.b\n"
-      "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
-      ".inst 0x4586985b  // usmmla z27.s, z2.b, z6.b\n"
-      ".inst 0x45859817  // usmmla z23.s, z0.b, z5.b\n"
-      "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
-      ".inst 0x4585985f  // usmmla z31.s, z2.b, z5.b\n"
-      ".inst 0x45849830  // usmmla z16.s, z1.b, z4.b\n"
-      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
-      "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
-      ".inst 0x45849878  // usmmla z24.s, z3.b, z4.b\n"
-      "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x458a9834  // usmmla z20.s, z1.b, z10.b\n"
-      ".inst 0x458a987c  // usmmla z28.s, z3.b, z10.b\n"
-      ".inst 0x45899831  // usmmla z17.s, z1.b, z9.b\n"
-      ".inst 0x45899879  // usmmla z25.s, z3.b, z9.b\n"
-      ".inst 0x45889835  // usmmla z21.s, z1.b, z8.b\n"
-      ".inst 0x4588987d  // usmmla z29.s, z3.b, z8.b\n"
-      ".inst 0x45879832  // usmmla z18.s, z1.b, z7.b\n"
-      ".inst 0x4587987a  // usmmla z26.s, z3.b, z7.b\n"
-      ".inst 0x45869836  // usmmla z22.s, z1.b, z6.b\n"
-      ".inst 0x4586987e  // usmmla z30.s, z3.b, z6.b\n"
-      ".inst 0x45859833  // usmmla z19.s, z1.b, z5.b\n"
-      ".inst 0x4585987b  // usmmla z27.s, z3.b, z5.b\n"
-      ".inst 0x45849837  // usmmla z23.s, z1.b, z4.b\n"
-      ".inst 0x4584987f  // usmmla z31.s, z3.b, z4.b\n"
-      "tbnz %x[flags], #31, 50f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z13.s, z2.b, z15.b\n"
-      "udot z11.s, z1.b, z15.b\n"
-      "udot z13.s, z3.b, z15.b\n"
-      "50:"  // Height 4: Multiply loop: unique 7: skip row sum
-      "sub x25, x25, #0x10\n"
-      "cmp x25, #0x10\n"
-      "bgt 49b\n"
-      "51:"  // Height 4: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x25\n"
-      "ld1b { z6.b }, p2/Z, [x28]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "subs x25, x25, #0x8\n"
-      "ld1b { z7.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z3.b }, p0/Z, [x22]\n"
-      "ld1rqb { z5.b }, p0/Z, [x21]\n"
-      "trn1 z0.d, z1.d, z2.d\n"
-      "trn2 z1.d, z1.d, z2.d\n"
-      "trn1 z2.d, z3.d, z5.d\n"
-      "trn2 z3.d, z3.d, z5.d\n"
-      "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
-      ".inst 0x45869810  // usmmla z16.s, z0.b, z6.b\n"
-      ".inst 0x45849814  // usmmla z20.s, z0.b, z4.b\n"
-      ".inst 0x45879811  // usmmla z17.s, z0.b, z7.b\n"
-      ".inst 0x45899815  // usmmla z21.s, z0.b, z9.b\n"
-      ".inst 0x45889812  // usmmla z18.s, z0.b, z8.b\n"
-      ".inst 0x45869858  // usmmla z24.s, z2.b, z6.b\n"
-      "ld1b { z6.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x4584985c  // usmmla z28.s, z2.b, z4.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x45879859  // usmmla z25.s, z2.b, z7.b\n"
-      ".inst 0x4589985d  // usmmla z29.s, z2.b, z9.b\n"
-      ".inst 0x4588985a  // usmmla z26.s, z2.b, z8.b\n"
-      ".inst 0x458a9816  // usmmla z22.s, z0.b, z10.b\n"
-      ".inst 0x458a985e  // usmmla z30.s, z2.b, z10.b\n"
-      ".inst 0x45859813  // usmmla z19.s, z0.b, z5.b\n"
-      ".inst 0x4585985b  // usmmla z27.s, z2.b, z5.b\n"
-      ".inst 0x45869817  // usmmla z23.s, z0.b, z6.b\n"
-      ".inst 0x4586985f  // usmmla z31.s, z2.b, z6.b\n"
-      "ble 52f\n"
-      "ld1b { z4.b }, p2/Z, [x28]\n"
-      "ld1b { z10.b }, p2/Z, [x28, #1, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
-      ".inst 0x45849830  // usmmla z16.s, z1.b, z4.b\n"
-      ".inst 0x45849878  // usmmla z24.s, z3.b, z4.b\n"
-      "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
-      ".inst 0x458a9834  // usmmla z20.s, z1.b, z10.b\n"
-      ".inst 0x458a987c  // usmmla z28.s, z3.b, z10.b\n"
-      ".inst 0x45899831  // usmmla z17.s, z1.b, z9.b\n"
-      ".inst 0x45899879  // usmmla z25.s, z3.b, z9.b\n"
-      "addvl x28, x28, #8\n"
-      ".inst 0x45889835  // usmmla z21.s, z1.b, z8.b\n"
-      ".inst 0x4588987d  // usmmla z29.s, z3.b, z8.b\n"
-      ".inst 0x45879832  // usmmla z18.s, z1.b, z7.b\n"
-      ".inst 0x4587987a  // usmmla z26.s, z3.b, z7.b\n"
-      ".inst 0x45869836  // usmmla z22.s, z1.b, z6.b\n"
-      ".inst 0x4586987e  // usmmla z30.s, z3.b, z6.b\n"
-      ".inst 0x45859833  // usmmla z19.s, z1.b, z5.b\n"
-      ".inst 0x4585987b  // usmmla z27.s, z3.b, z5.b\n"
-      ".inst 0x45849837  // usmmla z23.s, z1.b, z4.b\n"
-      ".inst 0x4584987f  // usmmla z31.s, z3.b, z4.b\n"
-      "52:"  // Height 4: Multiply loop: multiply skip
-      "tbnz %x[flags], #31, 53f\n"
-      "udot z11.s, z0.b, z15.b\n"
-      "udot z13.s, z2.b, z15.b\n"
-      "udot z11.s, z1.b, z15.b\n"
-      "udot z13.s, z3.b, z15.b\n"
-      "53:"  // Height 4: Multiply loop: unique 8: skip row sum
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x26, x26, #0x1\n"
-      "cmp x26, x20\n"
-      "bne 46b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z0.d, z16.d, z20.d\n"
-      "uzp2 z16.d, z16.d, z20.d\n"
-      "uzp1 z20.d, z17.d, z21.d\n"
-      "uzp2 z17.d, z17.d, z21.d\n"
-      "uzp1 z21.d, z18.d, z22.d\n"
-      "uzp2 z18.d, z18.d, z22.d\n"
-      "add x23, x27, x20\n"
-      "add x22, x23, x20\n"
-      "uzp1 z22.d, z19.d, z23.d\n"
-      "uzp2 z19.d, z19.d, z23.d\n"
-      "add x21, x22, x20\n"
-      "uzp1 z23.d, z24.d, z28.d\n"
-      "uzp2 z24.d, z24.d, z28.d\n"
-      "uzp1 z28.d, z25.d, z29.d\n"
-      "uzp2 z25.d, z25.d, z29.d\n"
-      "uzp1 z29.d, z26.d, z30.d\n"
-      "uzp2 z26.d, z26.d, z30.d\n"
-      "uzp1 z30.d, z27.d, z31.d\n"
-      "uzp2 z27.d, z27.d, z31.d\n"
-      "mov z31.d, z0.d\n"
-      "tbnz %x[flags], #31, 54f\n"
-      "add x20, %x[qp], %[b_offset]\n"
-      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
-      ".inst 0x4491a9ad  // addp z13.s, p2/m, z13.s, z13.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "neg z0.s, p2/M, z0.s\n"
-      "mov z12.s, z11.s[3]\n"
-      "mov z11.s, z11.s[0]\n"
-      "mov z14.s, z13.s[3]\n"
-      "mov z13.s, z13.s[0]\n"
-      "mul z11.s, p2/M, z11.s, z0.s\n"
-      "mul z12.s, p2/M, z12.s, z0.s\n"
-      "mul z13.s, p2/M, z13.s, z0.s\n"
-      "mul z14.s, p2/M, z14.s, z0.s\n"
-      "54:"  // Height 4: skip row sum fixup
-      "add z31.s, z31.s, z11.s\n"
-      "add z20.s, z20.s, z11.s\n"
-      "ld1w { z4.s }, p2/Z, [x10]\n"
-      "ld1w { z0.s }, p2/Z, [x10, #1, MUL VL]\n"
-      "add z21.s, z21.s, z11.s\n"
-      "add z22.s, z22.s, z11.s\n"
-      "ld1w { z3.s }, p2/Z, [x10, #2, MUL VL]\n"
-      "ld1w { z2.s }, p2/Z, [x10, #3, MUL VL]\n"
-      "add z16.s, z16.s, z12.s\n"
-      "add z17.s, z17.s, z12.s\n"
-      "add x20, %x[qp], %[per_layer_mul]\n"
-      "orr %x[flags], %x[flags], #0x80000000\n"
-      "add z18.s, z18.s, z12.s\n"
-      "add z19.s, z19.s, z12.s\n"
-      "ld1rw { z1.s }, p2/Z, [x20]\n"
-      "add x20, %x[qp], %[per_layer_right_shift]\n"
-      "add z23.s, z23.s, z13.s\n"
-      "add z28.s, z28.s, z13.s\n"
-      "addvl x10, x10, #4\n"
-      "add z29.s, z29.s, z13.s\n"
-      "add z30.s, z30.s, z13.s\n"
-      "add z24.s, z24.s, z14.s\n"
-      "add z25.s, z25.s, z14.s\n"
-      "add z26.s, z26.s, z14.s\n"
-      "add z27.s, z27.s, z14.s\n"
-      "add z31.s, z31.s, z4.s\n"
-      "add z20.s, z20.s, z0.s\n"
-      "add z21.s, z21.s, z3.s\n"
-      "add z22.s, z22.s, z2.s\n"
-      "add z16.s, z16.s, z4.s\n"
-      "add z17.s, z17.s, z0.s\n"
-      "add z18.s, z18.s, z3.s\n"
-      "add z19.s, z19.s, z2.s\n"
-      "add z23.s, z23.s, z4.s\n"
-      "add z28.s, z28.s, z0.s\n"
-      "add z29.s, z29.s, z3.s\n"
-      "add z30.s, z30.s, z2.s\n"
-      "add z24.s, z24.s, z4.s\n"
-      "add z25.s, z25.s, z0.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z26.s, z26.s, z3.s\n"
-      "add z27.s, z27.s, z2.s\n"
-      ".inst 0x04a177ff  // sqrdmulh z31.s, z31.s, z1.s\n"
-      ".inst 0x04a17694  // sqrdmulh z20.s, z20.s, z1.s\n"
-      ".inst 0x04a176b5  // sqrdmulh z21.s, z21.s, z1.s\n"
-      ".inst 0x04a176d6  // sqrdmulh z22.s, z22.s, z1.s\n"
-      ".inst 0x04a17610  // sqrdmulh z16.s, z16.s, z1.s\n"
-      ".inst 0x04a17631  // sqrdmulh z17.s, z17.s, z1.s\n"
-      ".inst 0x04a17652  // sqrdmulh z18.s, z18.s, z1.s\n"
-      ".inst 0x04a17673  // sqrdmulh z19.s, z19.s, z1.s\n"
-      ".inst 0x04a176f7  // sqrdmulh z23.s, z23.s, z1.s\n"
-      ".inst 0x04a1779c  // sqrdmulh z28.s, z28.s, z1.s\n"
-      ".inst 0x04a177bd  // sqrdmulh z29.s, z29.s, z1.s\n"
-      ".inst 0x04a177de  // sqrdmulh z30.s, z30.s, z1.s\n"
-      ".inst 0x04a17718  // sqrdmulh z24.s, z24.s, z1.s\n"
-      ".inst 0x04a17739  // sqrdmulh z25.s, z25.s, z1.s\n"
-      ".inst 0x04a1775a  // sqrdmulh z26.s, z26.s, z1.s\n"
-      ".inst 0x04a1777b  // sqrdmulh z27.s, z27.s, z1.s\n"
-      "tbz %x[flags], #5, 55f\n"
-      "and z2.d, z31.d, z0.d\n"
-      "and z1.d, z20.d, z0.d\n"
-      "and z7.d, z21.d, z0.d\n"
-      "and z6.d, z22.d, z0.d\n"
-      "and z5.d, z16.d, z0.d\n"
-      "and z4.d, z17.d, z0.d\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "and z3.d, z18.d, z0.d\n"
-      "asr z7.s, z7.s, #0x1f\n"
-      "asr z6.s, z6.s, #0x1f\n"
-      "asr z5.s, z5.s, #0x1f\n"
-      "sqadd z31.s, z31.s, z2.s\n"
-      "sqadd z20.s, z20.s, z1.s\n"
-      "and z2.d, z19.d, z0.d\n"
-      "and z1.d, z23.d, z0.d\n"
-      "asr z4.s, z4.s, #0x1f\n"
-      "asr z3.s, z3.s, #0x1f\n"
-      "sqadd z21.s, z21.s, z7.s\n"
-      "sqadd z22.s, z22.s, z6.s\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "sqadd z16.s, z16.s, z5.s\n"
-      "sqadd z17.s, z17.s, z4.s\n"
-      "sqadd z18.s, z18.s, z3.s\n"
-      "and z7.d, z28.d, z0.d\n"
-      "sqadd z19.s, z19.s, z2.s\n"
-      "sqadd z23.s, z23.s, z1.s\n"
-      "and z6.d, z29.d, z0.d\n"
-      "and z5.d, z30.d, z0.d\n"
-      "and z4.d, z24.d, z0.d\n"
-      "and z3.d, z25.d, z0.d\n"
-      "and z2.d, z26.d, z0.d\n"
-      "and z1.d, z27.d, z0.d\n"
-      "asr z7.s, z7.s, #0x1f\n"
-      "asr z6.s, z6.s, #0x1f\n"
-      "asr z5.s, z5.s, #0x1f\n"
-      "asr z4.s, z4.s, #0x1f\n"
-      "asr z3.s, z3.s, #0x1f\n"
-      "asr z2.s, z2.s, #0x1f\n"
-      "asr z1.s, z1.s, #0x1f\n"
-      "sqadd z28.s, z28.s, z7.s\n"
-      "sqadd z29.s, z29.s, z6.s\n"
-      "sqadd z30.s, z30.s, z5.s\n"
-      "sqadd z24.s, z24.s, z4.s\n"
-      "sqadd z25.s, z25.s, z3.s\n"
-      "sqadd z26.s, z26.s, z2.s\n"
-      "sqadd z27.s, z27.s, z1.s\n"
-      "55:"  // Height 4: no shift correction
-      "add x20, %x[qp], %[c_offset]\n"
-      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
-      "ld1rw { z2.s }, p2/Z, [x20]\n"
-      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
-      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
-      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
-      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
-      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
-      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
-      "add z31.s, z31.s, z2.s\n"
-      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
-      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
-      "add z20.s, z20.s, z2.s\n"
-      "add z21.s, z21.s, z2.s\n"
-      ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
-      ".inst 0x4482881d  // srshl z29.s, p2/M, z29.s, z0.s\n"
-      "add z22.s, z22.s, z2.s\n"
-      "add z16.s, z16.s, z2.s\n"
-      ".inst 0x4482881e  // srshl z30.s, p2/M, z30.s, z0.s\n"
-      ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
-      "add z17.s, z17.s, z2.s\n"
-      "add z18.s, z18.s, z2.s\n"
-      ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
-      ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
-      "add z19.s, z19.s, z2.s\n"
-      "add z23.s, z23.s, z2.s\n"
-      ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
-      "add x20, %x[qp], %[maxval]\n"
-      "add z28.s, z28.s, z2.s\n"
-      "add z29.s, z29.s, z2.s\n"
-      "ld1rw { z1.s }, p2/Z, [x20]\n"
-      "add z30.s, z30.s, z2.s\n"
-      "add z24.s, z24.s, z2.s\n"
-      "add x20, %x[qp], %[minval]\n"
-      "add z25.s, z25.s, z2.s\n"
-      "add z26.s, z26.s, z2.s\n"
-      "ld1rw { z0.s }, p2/Z, [x20]\n"
-      "add z27.s, z27.s, z2.s\n"
-      "smin z31.s, p2/M, z31.s, z1.s\n"
-      "smin z20.s, p2/M, z20.s, z1.s\n"
-      "smin z21.s, p2/M, z21.s, z1.s\n"
-      "smin z22.s, p2/M, z22.s, z1.s\n"
-      "smin z16.s, p2/M, z16.s, z1.s\n"
-      "smin z17.s, p2/M, z17.s, z1.s\n"
-      "smin z18.s, p2/M, z18.s, z1.s\n"
-      "smin z19.s, p2/M, z19.s, z1.s\n"
-      "smin z23.s, p2/M, z23.s, z1.s\n"
-      "smin z28.s, p2/M, z28.s, z1.s\n"
-      "smin z29.s, p2/M, z29.s, z1.s\n"
-      "smin z30.s, p2/M, z30.s, z1.s\n"
-      "smin z24.s, p2/M, z24.s, z1.s\n"
-      "smin z25.s, p2/M, z25.s, z1.s\n"
-      "smin z26.s, p2/M, z26.s, z1.s\n"
-      "smin z27.s, p2/M, z27.s, z1.s\n"
-      "smax z31.s, p2/M, z31.s, z0.s\n"
-      "smax z20.s, p2/M, z20.s, z0.s\n"
-      "smax z21.s, p2/M, z21.s, z0.s\n"
-      "smax z22.s, p2/M, z22.s, z0.s\n"
-      "smax z16.s, p2/M, z16.s, z0.s\n"
-      "smax z17.s, p2/M, z17.s, z0.s\n"
-      "smax z18.s, p2/M, z18.s, z0.s\n"
-      "smax z19.s, p2/M, z19.s, z0.s\n"
-      "uzp1 z31.h, z31.h, z20.h\n"
-      "smax z23.s, p2/M, z23.s, z0.s\n"
-      "smax z28.s, p2/M, z28.s, z0.s\n"
-      "uzp1 z20.h, z21.h, z22.h\n"
-      "smax z29.s, p2/M, z29.s, z0.s\n"
-      "smax z30.s, p2/M, z30.s, z0.s\n"
-      "uzp1 z16.h, z16.h, z17.h\n"
-      "smax z24.s, p2/M, z24.s, z0.s\n"
-      "smax z25.s, p2/M, z25.s, z0.s\n"
-      "uzp1 z17.h, z18.h, z19.h\n"
-      "smax z26.s, p2/M, z26.s, z0.s\n"
-      "smax z27.s, p2/M, z27.s, z0.s\n"
-      "uzp1 z23.h, z23.h, z28.h\n"
-      "uzp1 z31.b, z31.b, z20.b\n"
-      "uzp1 z18.h, z29.h, z30.h\n"
-      "uzp1 z24.h, z24.h, z25.h\n"
-      "uzp1 z16.b, z16.b, z17.b\n"
-      "uzp1 z17.h, z26.h, z27.h\n"
-      "st1b { z31.b }, p1, [x27]\n"
-      "addvl x27, x27, #1\n"
-      "uzp1 z23.b, z23.b, z18.b\n"
-      "uzp1 z24.b, z24.b, z17.b\n"
-      "st1b { z16.b }, p1, [x23]\n"
-      "st1b { z23.b }, p1, [x22]\n"
-      "st1b { z24.b }, p1, [x21]\n"
-      "56:"  // Height 4: Writeback done
-      "decw x9, ALL, MUL #4\n"
-      "cmp x9, XZR\n"
-      "bgt 44b\n"
-      "subs %x[M], %x[M], #0x4\n"
-      "beq 58f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "tbz %x[flags], #3, 57f\n"
-      "add x21, x21, #0x4\n"
-      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "b 1b\n"
-      "57:"  // Update direct input
-      "mov x20, #0x4\n"
-      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
-      "b 1b\n"
-      "58:"  // Exit
-      : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
-      : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
-    );
-}
-
-} // namespace arm_gemm
-#endif  // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8s32_mmla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8s32_mmla_6x4VL.hpp
deleted file mode 100644
index 7a8ee8ecb8..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8s32_mmla_6x4VL.hpp
+++ /dev/null
@@ -1,112 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include "../std_transforms_sve.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    unsigned int, const unsigned int *, \
-    IndirectInputArg<uint8_t>, \
-    size_t, size_t, \
-    const int8_t *, \
-    IndirectOutputArg<int32_t>, \
-    const int32_t *, Activation, bool
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void sve_hybrid_u8s8s32_mmla_6x4VL( ARGLIST );
-
-class cls_sve_hybrid_u8s8s32_mmla_6x4VL
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef int32_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 6;
-    }
-
-    static unsigned int out_width()
-    {
-        return get_vector_length<int32_t>() * 4;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 8;
-    }
-
-    static constexpr bool supports_accumulate()
-    {
-        return true;
-    }
-
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 8, 8> transforms = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-        if (std::is_same<T, uint32_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 54.45 };
-                case CPUModel::A510:
-                    return { 24.22 };
-                case CPUModel::V1:
-                    return { 105.16 };
-            }
-        }
-
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 54.90, 15.69, 0.62 };
-                case CPUModel::A510:
-                    return { 26.80, 3.89, 0.47 };
-                case CPUModel::V1:
-                    return { 75.14, 15.87, 0.83 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=sve_hybrid_u8s8s32_mmla_6x4VL;
-    cls_sve_hybrid_u8s8s32_mmla_6x4VL(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8s32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8s32_mmla_6x4VL/generic.cpp
deleted file mode 100644
index 14299e80d6..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8s8s32_mmla_6x4VL/generic.cpp
+++ /dev/null
@@ -1,1675 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include "arm_gemm.hpp"
-#include "../../utils.hpp"
-
-#include <cassert>
-
-namespace arm_gemm {
-
-void sve_hybrid_u8s8s32_mmla_6x4VL (
-    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg,
-    size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<int32_t> output_arg,
-    const int32_t *, Activation, bool accumulate
-)
-{
-    struct KernelArgs {
-        unsigned int num_strings = {};
-        const unsigned int *string_lengths = {};
-        size_t N = {};
-        const int8_t *B_ptr = {};
-        size_t output_offset = {};
-        size_t input_initial_col = {};
-        size_t input_offset = {};
-        void *output_ptr = {};
-    } ka;
-
-    unsigned long flags=0;
-    void *input_ptr;
-
-    if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
-        ka.output_offset=output_arg.indirect.offset;
-        flags |= 0x4;
-    } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
-        ka.output_offset=output_arg.direct.stride;
-    }
-
-    if (A_arg.is_indirect) {
-        input_ptr=(void *)(A_arg.indirect.ptr);
-        ka.input_offset=A_arg.indirect.start_row;
-        ka.input_initial_col=A_arg.indirect.start_col;
-        flags |= 0x8;
-    } else {
-        assert(num_strings==1);
-        input_ptr=(void *)(A_arg.direct.base);
-        ka.input_offset=A_arg.direct.stride;
-    }
-    if (accumulate) {
-        flags |= 0x1;
-    }
-    ka.num_strings = num_strings;
-    ka.string_lengths = string_lengths;
-    ka.N = N;
-    ka.B_ptr = B_ptr;
-    __asm__ __volatile__(
-      "ptrue p5.b\n"
-      "1:"  // Row loop
-      "cmp %x[M], #0x6\n"
-      "bge 56f\n"
-      "cmp %x[M], #0x4\n"
-      "bgt 45f\n"
-      "beq 34f\n"
-      "cmp %x[M], #0x2\n"
-      "bgt 23f\n"
-      "beq 12f\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "2:"  // Height 1: Column loop
-      "mov x20, #0x0\n"
-      "whilelt p4.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p3.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p2.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p1.s, x20, x11\n"
-      "tbz %x[flags], #0, 3f\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z19.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "zip1 z9.d, z19.d, z13.d\n"
-      "zip2 z13.d, z19.d, z13.d\n"
-      "zip1 z10.d, z17.d, z14.d\n"
-      "zip2 z14.d, z17.d, z14.d\n"
-      "zip1 z11.d, z16.d, z15.d\n"
-      "zip2 z15.d, z16.d, z15.d\n"
-      "b 4f\n"
-      "3:"  // Height 1: no accumulate
-      "mov z8.s, #0x0\n"
-      "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "mov z15.s, #0x0\n"
-      "4:"  // Height 1: setup done
-      "mov x28, #0x0\n"
-      "5:"  // Height 1: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 6f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "cbnz x28, 7f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "b 7f\n"
-      "6:"  // Height 1: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "7:"  // Height 1: input setup done
-      "cmp x27, #0x10\n"
-      "ble 9f\n"
-      "8:"  // Height 1: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z16.b }, p5/Z, [x10]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z19.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "trn1 z18.d, z19.d, z20.d\n"
-      "trn2 z19.d, z19.d, z20.d\n"
-      ".inst 0x45909a48  // usmmla z8.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45919a4c  // usmmla z12.s, z18.b, z17.b\n"
-      "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45909a49  // usmmla z9.s, z18.b, z16.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45949a4d  // usmmla z13.s, z18.b, z20.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45819a4a  // usmmla z10.s, z18.b, z1.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45909a4e  // usmmla z14.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x45919a4b  // usmmla z11.s, z18.b, z17.b\n"
-      ".inst 0x45909a4f  // usmmla z15.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45909a68  // usmmla z8.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45919a6c  // usmmla z12.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45909a69  // usmmla z9.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45919a6d  // usmmla z13.s, z19.b, z17.b\n"
-      "ld1b { z3.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45909a6a  // usmmla z10.s, z19.b, z16.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45839a6e  // usmmla z14.s, z19.b, z3.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45919a6b  // usmmla z11.s, z19.b, z17.b\n"
-      ".inst 0x45909a6f  // usmmla z15.s, z19.b, z16.b\n"
-      "bgt 8b\n"
-      "9:"  // Height 1: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
-      "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "trn1 z18.d, z1.d, z19.d\n"
-      ".inst 0x45919a48  // usmmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45909a4c  // usmmla z12.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
-      ".inst 0x45919a49  // usmmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45909a4d  // usmmla z13.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45919a4a  // usmmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45909a4e  // usmmla z14.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x45919a4b  // usmmla z11.s, z18.b, z17.b\n"
-      ".inst 0x45909a4f  // usmmla z15.s, z18.b, z16.b\n"
-      "ble 10f\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45919828  // usmmla z8.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4590982c  // usmmla z12.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45919829  // usmmla z9.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4590982d  // usmmla z13.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4591982a  // usmmla z10.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4590982e  // usmmla z14.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x4591982b  // usmmla z11.s, z1.b, z17.b\n"
-      ".inst 0x4590982f  // usmmla z15.s, z1.b, z16.b\n"
-      "10:"  // Height 1: Multiply loop: multiply skip
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 5b\n"
-      "uzp1 z8.d, z8.d, z12.d\n"
-      "uzp1 z9.d, z9.d, z13.d\n"
-      "uzp1 z10.d, z10.d, z14.d\n"
-      "uzp1 z11.d, z11.d, z15.d\n"
-      "st1w { z8.s }, p4, [x9]\n"
-      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
-      "11:"  // Height 1: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
-      "bgt 2b\n"
-      "b 68f\n"
-      "12:"  // Height 2
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "13:"  // Height 2: Column loop
-      "mov x20, #0x0\n"
-      "whilelt p4.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p3.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p2.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p1.s, x20, x11\n"
-      "tbz %x[flags], #0, 14f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z18.s }, p4/Z, [x9]\n"
-      "ld1w { z24.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x20]\n"
-      "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z8.d, z18.d, z12.d\n"
-      "zip2 z12.d, z18.d, z12.d\n"
-      "zip1 z9.d, z24.d, z13.d\n"
-      "zip2 z13.d, z24.d, z13.d\n"
-      "zip1 z10.d, z17.d, z14.d\n"
-      "zip2 z14.d, z17.d, z14.d\n"
-      "zip1 z11.d, z16.d, z15.d\n"
-      "zip2 z15.d, z16.d, z15.d\n"
-      "b 15f\n"
-      "14:"  // Height 2: no accumulate
-      "mov z8.s, #0x0\n"
-      "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "mov z15.s, #0x0\n"
-      "15:"  // Height 2: setup done
-      "mov x28, #0x0\n"
-      "16:"  // Height 2: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 17f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "cbnz x28, 18f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "b 18f\n"
-      "17:"  // Height 2: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "18:"  // Height 2: input setup done
-      "cmp x27, #0x10\n"
-      "ble 20f\n"
-      "19:"  // Height 2: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z19.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "trn1 z18.d, z19.d, z25.d\n"
-      "trn2 z19.d, z19.d, z25.d\n"
-      ".inst 0x45919a48  // usmmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45909a4c  // usmmla z12.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45919a49  // usmmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45909a4d  // usmmla z13.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45919a4a  // usmmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45909a4e  // usmmla z14.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x45919a4b  // usmmla z11.s, z18.b, z17.b\n"
-      ".inst 0x45909a4f  // usmmla z15.s, z18.b, z16.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45919a68  // usmmla z8.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45909a6c  // usmmla z12.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45919a69  // usmmla z9.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45909a6d  // usmmla z13.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45919a6a  // usmmla z10.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45909a6e  // usmmla z14.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45919a6b  // usmmla z11.s, z19.b, z17.b\n"
-      ".inst 0x45909a6f  // usmmla z15.s, z19.b, z16.b\n"
-      "bgt 19b\n"
-      "20:"  // Height 2: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
-      "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z19.b }, p0/Z, [x25]\n"
-      "trn1 z18.d, z1.d, z19.d\n"
-      ".inst 0x45919a48  // usmmla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45909a4c  // usmmla z12.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
-      ".inst 0x45919a49  // usmmla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45909a4d  // usmmla z13.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45919a4a  // usmmla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45909a4e  // usmmla z14.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x45919a4b  // usmmla z11.s, z18.b, z17.b\n"
-      ".inst 0x45909a4f  // usmmla z15.s, z18.b, z16.b\n"
-      "ble 21f\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45919828  // usmmla z8.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4590982c  // usmmla z12.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45919829  // usmmla z9.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4590982d  // usmmla z13.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4591982a  // usmmla z10.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4590982e  // usmmla z14.s, z1.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x4591982b  // usmmla z11.s, z1.b, z17.b\n"
-      ".inst 0x4590982f  // usmmla z15.s, z1.b, z16.b\n"
-      "21:"  // Height 2: Multiply loop: multiply skip
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 16b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z17.d, z8.d, z12.d\n"
-      "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z16.d, z9.d, z13.d\n"
-      "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z12.d, z10.d, z14.d\n"
-      "uzp2 z10.d, z10.d, z14.d\n"
-      "add x20, x9, x20, LSL #2\n"
-      "uzp1 z26.d, z11.d, z15.d\n"
-      "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z17.s }, p4, [x9]\n"
-      "st1w { z16.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z12.s }, p2, [x9, #2, MUL VL]\n"
-      "st1w { z26.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x20]\n"
-      "st1w { z9.s }, p3, [x20, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x20, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x20, #3, MUL VL]\n"
-      "22:"  // Height 2: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
-      "bgt 13b\n"
-      "b 68f\n"
-      "23:"  // Height 3
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "24:"  // Height 3: Column loop
-      "mov x20, #0x0\n"
-      "whilelt p4.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p3.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p2.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p1.s, x20, x11\n"
-      "tbz %x[flags], #0, 25f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z24.s }, p4/Z, [x9]\n"
-      "ld1w { z26.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x21]\n"
-      "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "ld1w { z17.s }, p4/Z, [x20]\n"
-      "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "zip1 z8.d, z24.d, z12.d\n"
-      "zip2 z12.d, z24.d, z12.d\n"
-      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z9.d, z26.d, z13.d\n"
-      "zip2 z13.d, z26.d, z13.d\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "zip1 z11.d, z16.d, z15.d\n"
-      "zip2 z15.d, z16.d, z15.d\n"
-      "zip1 z16.d, z17.d, z20.d\n"
-      "zip2 z20.d, z17.d, z20.d\n"
-      "zip1 z17.d, z18.d, z21.d\n"
-      "zip2 z21.d, z18.d, z21.d\n"
-      "zip1 z18.d, z19.d, z22.d\n"
-      "zip2 z22.d, z19.d, z22.d\n"
-      "zip1 z19.d, z24.d, z23.d\n"
-      "zip2 z23.d, z24.d, z23.d\n"
-      "b 26f\n"
-      "25:"  // Height 3: no accumulate
-      "mov z8.s, #0x0\n"
-      "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "mov z15.s, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "26:"  // Height 3: setup done
-      "mov x28, #0x0\n"
-      "27:"  // Height 3: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 28f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "cbnz x28, 29f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "b 29f\n"
-      "28:"  // Height 3: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "29:"  // Height 3: input setup done
-      "cmp x27, #0x10\n"
-      "ble 31f\n"
-      "30:"  // Height 3: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z27.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z24.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z26.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "trn1 z6.d, z27.d, z24.d\n"
-      "trn2 z27.d, z27.d, z24.d\n"
-      "trn1 z30.d, z26.d, z29.d\n"
-      "trn2 z26.d, z26.d, z29.d\n"
-      ".inst 0x459998c8  // usmmla z8.s, z6.b, z25.b\n"
-      ".inst 0x459c98cc  // usmmla z12.s, z6.b, z28.b\n"
-      ".inst 0x45999bd0  // usmmla z16.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x459c9bd4  // usmmla z20.s, z30.b, z28.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x459998c9  // usmmla z9.s, z6.b, z25.b\n"
-      ".inst 0x45999bd1  // usmmla z17.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x459898cd  // usmmla z13.s, z6.b, z24.b\n"
-      ".inst 0x45989bd5  // usmmla z21.s, z30.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x459998ca  // usmmla z10.s, z6.b, z25.b\n"
-      ".inst 0x45999bd2  // usmmla z18.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x459898ce  // usmmla z14.s, z6.b, z24.b\n"
-      ".inst 0x45989bd6  // usmmla z22.s, z30.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x459998cb  // usmmla z11.s, z6.b, z25.b\n"
-      ".inst 0x45999bd3  // usmmla z19.s, z30.b, z25.b\n"
-      ".inst 0x459898cf  // usmmla z15.s, z6.b, z24.b\n"
-      ".inst 0x45989bd7  // usmmla z23.s, z30.b, z24.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45999b68  // usmmla z8.s, z27.b, z25.b\n"
-      ".inst 0x45999b50  // usmmla z16.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45989b6c  // usmmla z12.s, z27.b, z24.b\n"
-      ".inst 0x45989b54  // usmmla z20.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45999b69  // usmmla z9.s, z27.b, z25.b\n"
-      ".inst 0x45999b51  // usmmla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45989b6d  // usmmla z13.s, z27.b, z24.b\n"
-      ".inst 0x45989b55  // usmmla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45999b6a  // usmmla z10.s, z27.b, z25.b\n"
-      ".inst 0x45999b52  // usmmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45989b6e  // usmmla z14.s, z27.b, z24.b\n"
-      ".inst 0x45989b56  // usmmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45999b6b  // usmmla z11.s, z27.b, z25.b\n"
-      ".inst 0x45999b53  // usmmla z19.s, z26.b, z25.b\n"
-      ".inst 0x45989b6f  // usmmla z15.s, z27.b, z24.b\n"
-      ".inst 0x45989b57  // usmmla z23.s, z26.b, z24.b\n"
-      "bgt 30b\n"
-      "31:"  // Height 3: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
-      "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z24.b }, p0/Z, [x25]\n"
-      "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "trn1 z27.d, z1.d, z24.d\n"
-      "trn2 z1.d, z1.d, z24.d\n"
-      "trn1 z26.d, z3.d, z29.d\n"
-      ".inst 0x45999b68  // usmmla z8.s, z27.b, z25.b\n"
-      ".inst 0x459c9b6c  // usmmla z12.s, z27.b, z28.b\n"
-      "trn2 z3.d, z3.d, z29.d\n"
-      ".inst 0x45999b50  // usmmla z16.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x459c9b54  // usmmla z20.s, z26.b, z28.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45999b69  // usmmla z9.s, z27.b, z25.b\n"
-      ".inst 0x45999b51  // usmmla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45989b6d  // usmmla z13.s, z27.b, z24.b\n"
-      ".inst 0x45989b55  // usmmla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45999b6a  // usmmla z10.s, z27.b, z25.b\n"
-      ".inst 0x45999b52  // usmmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45989b6e  // usmmla z14.s, z27.b, z24.b\n"
-      ".inst 0x45989b56  // usmmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x45999b6b  // usmmla z11.s, z27.b, z25.b\n"
-      ".inst 0x45999b53  // usmmla z19.s, z26.b, z25.b\n"
-      ".inst 0x45989b6f  // usmmla z15.s, z27.b, z24.b\n"
-      ".inst 0x45989b57  // usmmla z23.s, z26.b, z24.b\n"
-      "ble 32f\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45999828  // usmmla z8.s, z1.b, z25.b\n"
-      ".inst 0x45999870  // usmmla z16.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4598982c  // usmmla z12.s, z1.b, z24.b\n"
-      ".inst 0x45989874  // usmmla z20.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45999829  // usmmla z9.s, z1.b, z25.b\n"
-      ".inst 0x45999871  // usmmla z17.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4598982d  // usmmla z13.s, z1.b, z24.b\n"
-      ".inst 0x45989875  // usmmla z21.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4599982a  // usmmla z10.s, z1.b, z25.b\n"
-      ".inst 0x45999872  // usmmla z18.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4598982e  // usmmla z14.s, z1.b, z24.b\n"
-      ".inst 0x45989876  // usmmla z22.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x4599982b  // usmmla z11.s, z1.b, z25.b\n"
-      ".inst 0x45999873  // usmmla z19.s, z3.b, z25.b\n"
-      ".inst 0x4598982f  // usmmla z15.s, z1.b, z24.b\n"
-      ".inst 0x45989877  // usmmla z23.s, z3.b, z24.b\n"
-      "32:"  // Height 3: Multiply loop: multiply skip
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 27b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z27.d, z8.d, z12.d\n"
-      "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z26.d, z9.d, z13.d\n"
-      "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z25.d, z10.d, z14.d\n"
-      "uzp2 z10.d, z10.d, z14.d\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "uzp1 z24.d, z11.d, z15.d\n"
-      "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z27.s }, p4, [x9]\n"
-      "uzp1 z16.d, z16.d, z20.d\n"
-      "uzp1 z17.d, z17.d, z21.d\n"
-      "st1w { z26.s }, p3, [x9, #1, MUL VL]\n"
-      "uzp1 z18.d, z18.d, z22.d\n"
-      "uzp1 z19.d, z19.d, z23.d\n"
-      "st1w { z25.s }, p2, [x9, #2, MUL VL]\n"
-      "st1w { z24.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
-      "st1w { z8.s }, p4, [x21]\n"
-      "st1w { z9.s }, p3, [x21, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x21, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x21, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x20]\n"
-      "st1w { z17.s }, p3, [x20, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x20, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x20, #3, MUL VL]\n"
-      "33:"  // Height 3: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
-      "bgt 24b\n"
-      "b 68f\n"
-      "34:"  // Height 4
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "35:"  // Height 4: Column loop
-      "mov x20, #0x0\n"
-      "whilelt p4.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p3.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p2.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p1.s, x20, x11\n"
-      "tbz %x[flags], #0, 36f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x22]\n"
-      "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
-      "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "ld1w { z17.s }, p4/Z, [x21]\n"
-      "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x20]\n"
-      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "zip1 z11.d, z16.d, z15.d\n"
-      "zip2 z15.d, z16.d, z15.d\n"
-      "zip1 z16.d, z17.d, z20.d\n"
-      "zip2 z20.d, z17.d, z20.d\n"
-      "zip1 z17.d, z18.d, z21.d\n"
-      "zip2 z21.d, z18.d, z21.d\n"
-      "zip1 z18.d, z19.d, z22.d\n"
-      "zip2 z22.d, z19.d, z22.d\n"
-      "zip1 z19.d, z24.d, z23.d\n"
-      "zip2 z23.d, z24.d, z23.d\n"
-      "b 37f\n"
-      "36:"  // Height 4: no accumulate
-      "mov z8.s, #0x0\n"
-      "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "mov z15.s, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "37:"  // Height 4: setup done
-      "mov x28, #0x0\n"
-      "38:"  // Height 4: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 39f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "cbnz x28, 40f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "b 40f\n"
-      "39:"  // Height 4: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "40:"  // Height 4: input setup done
-      "cmp x27, #0x10\n"
-      "ble 42f\n"
-      "41:"  // Height 4: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z31.b }, p5/Z, [x10]\n"
-      "ld1b { z30.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z29.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z28.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "trn1 z27.d, z29.d, z25.d\n"
-      "trn2 z29.d, z29.d, z25.d\n"
-      "trn1 z26.d, z28.d, z24.d\n"
-      "trn2 z28.d, z28.d, z24.d\n"
-      ".inst 0x459f9b68  // usmmla z8.s, z27.b, z31.b\n"
-      ".inst 0x459e9b6c  // usmmla z12.s, z27.b, z30.b\n"
-      ".inst 0x459f9b50  // usmmla z16.s, z26.b, z31.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x459e9b54  // usmmla z20.s, z26.b, z30.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45999b69  // usmmla z9.s, z27.b, z25.b\n"
-      ".inst 0x45999b51  // usmmla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45989b6d  // usmmla z13.s, z27.b, z24.b\n"
-      ".inst 0x45989b55  // usmmla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45999b6a  // usmmla z10.s, z27.b, z25.b\n"
-      ".inst 0x45999b52  // usmmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45989b6e  // usmmla z14.s, z27.b, z24.b\n"
-      ".inst 0x45989b56  // usmmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x45999b6b  // usmmla z11.s, z27.b, z25.b\n"
-      ".inst 0x45999b53  // usmmla z19.s, z26.b, z25.b\n"
-      ".inst 0x45989b6f  // usmmla z15.s, z27.b, z24.b\n"
-      ".inst 0x45989b57  // usmmla z23.s, z26.b, z24.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45999ba8  // usmmla z8.s, z29.b, z25.b\n"
-      ".inst 0x45999b90  // usmmla z16.s, z28.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45989bac  // usmmla z12.s, z29.b, z24.b\n"
-      ".inst 0x45989b94  // usmmla z20.s, z28.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45999ba9  // usmmla z9.s, z29.b, z25.b\n"
-      ".inst 0x45999b91  // usmmla z17.s, z28.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45989bad  // usmmla z13.s, z29.b, z24.b\n"
-      ".inst 0x45989b95  // usmmla z21.s, z28.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45999baa  // usmmla z10.s, z29.b, z25.b\n"
-      ".inst 0x45999b92  // usmmla z18.s, z28.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45989bae  // usmmla z14.s, z29.b, z24.b\n"
-      ".inst 0x45989b96  // usmmla z22.s, z28.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45999bab  // usmmla z11.s, z29.b, z25.b\n"
-      ".inst 0x45999b93  // usmmla z19.s, z28.b, z25.b\n"
-      ".inst 0x45989baf  // usmmla z15.s, z29.b, z24.b\n"
-      ".inst 0x45989b97  // usmmla z23.s, z28.b, z24.b\n"
-      "bgt 41b\n"
-      "42:"  // Height 4: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
-      "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
-      "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "trn1 z27.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      "trn1 z26.d, z3.d, z24.d\n"
-      ".inst 0x459d9b68  // usmmla z8.s, z27.b, z29.b\n"
-      ".inst 0x459c9b6c  // usmmla z12.s, z27.b, z28.b\n"
-      "trn2 z3.d, z3.d, z24.d\n"
-      ".inst 0x459d9b50  // usmmla z16.s, z26.b, z29.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x459c9b54  // usmmla z20.s, z26.b, z28.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45999b69  // usmmla z9.s, z27.b, z25.b\n"
-      ".inst 0x45999b51  // usmmla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45989b6d  // usmmla z13.s, z27.b, z24.b\n"
-      ".inst 0x45989b55  // usmmla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45999b6a  // usmmla z10.s, z27.b, z25.b\n"
-      ".inst 0x45999b52  // usmmla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45989b6e  // usmmla z14.s, z27.b, z24.b\n"
-      ".inst 0x45989b56  // usmmla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x45999b6b  // usmmla z11.s, z27.b, z25.b\n"
-      ".inst 0x45999b53  // usmmla z19.s, z26.b, z25.b\n"
-      ".inst 0x45989b6f  // usmmla z15.s, z27.b, z24.b\n"
-      ".inst 0x45989b57  // usmmla z23.s, z26.b, z24.b\n"
-      "ble 43f\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45999828  // usmmla z8.s, z1.b, z25.b\n"
-      ".inst 0x45999870  // usmmla z16.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4598982c  // usmmla z12.s, z1.b, z24.b\n"
-      ".inst 0x45989874  // usmmla z20.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45999829  // usmmla z9.s, z1.b, z25.b\n"
-      ".inst 0x45999871  // usmmla z17.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4598982d  // usmmla z13.s, z1.b, z24.b\n"
-      ".inst 0x45989875  // usmmla z21.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4599982a  // usmmla z10.s, z1.b, z25.b\n"
-      ".inst 0x45999872  // usmmla z18.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4598982e  // usmmla z14.s, z1.b, z24.b\n"
-      ".inst 0x45989876  // usmmla z22.s, z3.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x4599982b  // usmmla z11.s, z1.b, z25.b\n"
-      ".inst 0x45999873  // usmmla z19.s, z3.b, z25.b\n"
-      ".inst 0x4598982f  // usmmla z15.s, z1.b, z24.b\n"
-      ".inst 0x45989877  // usmmla z23.s, z3.b, z24.b\n"
-      "43:"  // Height 4: Multiply loop: multiply skip
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 38b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z25.d, z8.d, z12.d\n"
-      "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z24.d, z9.d, z13.d\n"
-      "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z27.d, z10.d, z14.d\n"
-      "uzp2 z10.d, z10.d, z14.d\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "uzp1 z26.d, z11.d, z15.d\n"
-      "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z25.s }, p4, [x9]\n"
-      "uzp1 z25.d, z16.d, z20.d\n"
-      "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z24.s }, p3, [x9, #1, MUL VL]\n"
-      "uzp1 z24.d, z17.d, z21.d\n"
-      "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z27.s }, p2, [x9, #2, MUL VL]\n"
-      "uzp1 z21.d, z18.d, z22.d\n"
-      "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z26.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
-      "uzp1 z20.d, z19.d, z23.d\n"
-      "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x22]\n"
-      "st1w { z9.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x22, #3, MUL VL]\n"
-      "st1w { z25.s }, p4, [x21]\n"
-      "st1w { z24.s }, p3, [x21, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x21, #2, MUL VL]\n"
-      "st1w { z20.s }, p1, [x21, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x20]\n"
-      "st1w { z17.s }, p3, [x20, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x20, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x20, #3, MUL VL]\n"
-      "44:"  // Height 4: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
-      "bgt 35b\n"
-      "b 68f\n"
-      "45:"  // Height 5
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "46:"  // Height 5: Column loop
-      "mov x20, #0x0\n"
-      "whilelt p4.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p3.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p2.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p1.s, x20, x11\n"
-      "tbz %x[flags], #0, 47f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x23]\n"
-      "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
-      "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "ld1w { z17.s }, p4/Z, [x22]\n"
-      "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x21]\n"
-      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x20]\n"
-      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "zip1 z11.d, z16.d, z15.d\n"
-      "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z16.d, z17.d, z20.d\n"
-      "zip2 z20.d, z17.d, z20.d\n"
-      "zip1 z17.d, z18.d, z21.d\n"
-      "zip2 z21.d, z18.d, z21.d\n"
-      "zip1 z18.d, z19.d, z22.d\n"
-      "zip2 z22.d, z19.d, z22.d\n"
-      "zip1 z19.d, z24.d, z23.d\n"
-      "zip2 z23.d, z24.d, z23.d\n"
-      "zip1 z24.d, z25.d, z28.d\n"
-      "zip2 z28.d, z25.d, z28.d\n"
-      "zip1 z25.d, z26.d, z29.d\n"
-      "zip2 z29.d, z26.d, z29.d\n"
-      "zip1 z26.d, z27.d, z30.d\n"
-      "zip2 z30.d, z27.d, z30.d\n"
-      "zip1 z27.d, z0.d, z31.d\n"
-      "zip2 z31.d, z0.d, z31.d\n"
-      "b 48f\n"
-      "47:"  // Height 5: no accumulate
-      "mov z8.s, #0x0\n"
-      "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "mov z15.s, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "mov z24.s, #0x0\n"
-      "mov z25.s, #0x0\n"
-      "mov z26.s, #0x0\n"
-      "mov z27.s, #0x0\n"
-      "mov z28.s, #0x0\n"
-      "mov z29.s, #0x0\n"
-      "mov z30.s, #0x0\n"
-      "mov z31.s, #0x0\n"
-      "48:"  // Height 5: setup done
-      "mov x28, #0x0\n"
-      "49:"  // Height 5: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 50f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "ldr x22, [x20, #0x20]\n"
-      "cbnz x28, 51f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "b 51f\n"
-      "50:"  // Height 5: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "51:"  // Height 5: input setup done
-      "cmp x27, #0x10\n"
-      "ble 53f\n"
-      "52:"  // Height 5: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p5/Z, [x10]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z6.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z7.b }, p0/Z, [x24]\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
-      "trn1 z3.d, z7.d, z2.d\n"
-      "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
-      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45819888  // usmmla z8.s, z4.b, z1.b\n"
-      ".inst 0x45819870  // usmmla z16.s, z3.b, z1.b\n"
-      ".inst 0x45819858  // usmmla z24.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4580988c  // usmmla z12.s, z4.b, z0.b\n"
-      ".inst 0x45809874  // usmmla z20.s, z3.b, z0.b\n"
-      ".inst 0x4580985c  // usmmla z28.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45819889  // usmmla z9.s, z4.b, z1.b\n"
-      ".inst 0x45819871  // usmmla z17.s, z3.b, z1.b\n"
-      ".inst 0x45819859  // usmmla z25.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4580988d  // usmmla z13.s, z4.b, z0.b\n"
-      ".inst 0x45809875  // usmmla z21.s, z3.b, z0.b\n"
-      ".inst 0x4580985d  // usmmla z29.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4581988a  // usmmla z10.s, z4.b, z1.b\n"
-      ".inst 0x45819872  // usmmla z18.s, z3.b, z1.b\n"
-      ".inst 0x4581985a  // usmmla z26.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4580988e  // usmmla z14.s, z4.b, z0.b\n"
-      ".inst 0x45809876  // usmmla z22.s, z3.b, z0.b\n"
-      ".inst 0x4580985e  // usmmla z30.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x4581988b  // usmmla z11.s, z4.b, z1.b\n"
-      ".inst 0x45819873  // usmmla z19.s, z3.b, z1.b\n"
-      ".inst 0x4581985b  // usmmla z27.s, z2.b, z1.b\n"
-      ".inst 0x4580988f  // usmmla z15.s, z4.b, z0.b\n"
-      ".inst 0x45809877  // usmmla z23.s, z3.b, z0.b\n"
-      ".inst 0x4580985f  // usmmla z31.s, z2.b, z0.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z0.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x458198c8  // usmmla z8.s, z6.b, z1.b\n"
-      ".inst 0x458198f0  // usmmla z16.s, z7.b, z1.b\n"
-      ".inst 0x458198b8  // usmmla z24.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x458098cc  // usmmla z12.s, z6.b, z0.b\n"
-      ".inst 0x458098f4  // usmmla z20.s, z7.b, z0.b\n"
-      ".inst 0x458098bc  // usmmla z28.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x458198c9  // usmmla z9.s, z6.b, z1.b\n"
-      ".inst 0x458198f1  // usmmla z17.s, z7.b, z1.b\n"
-      ".inst 0x458198b9  // usmmla z25.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x458098cd  // usmmla z13.s, z6.b, z0.b\n"
-      ".inst 0x458098f5  // usmmla z21.s, z7.b, z0.b\n"
-      ".inst 0x458098bd  // usmmla z29.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x458198ca  // usmmla z10.s, z6.b, z1.b\n"
-      ".inst 0x458198f2  // usmmla z18.s, z7.b, z1.b\n"
-      ".inst 0x458198ba  // usmmla z26.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x458098ce  // usmmla z14.s, z6.b, z0.b\n"
-      ".inst 0x458098f6  // usmmla z22.s, z7.b, z0.b\n"
-      ".inst 0x458098be  // usmmla z30.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x458198cb  // usmmla z11.s, z6.b, z1.b\n"
-      ".inst 0x458198f3  // usmmla z19.s, z7.b, z1.b\n"
-      ".inst 0x458198bb  // usmmla z27.s, z5.b, z1.b\n"
-      ".inst 0x458098cf  // usmmla z15.s, z6.b, z0.b\n"
-      ".inst 0x458098f7  // usmmla z23.s, z7.b, z0.b\n"
-      ".inst 0x458098bf  // usmmla z31.s, z5.b, z0.b\n"
-      "bgt 52b\n"
-      "53:"  // Height 5: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z2.b }, p5/Z, [x10]\n"
-      "subs x27, x27, #0x8\n"
-      "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z6.b }, p0/Z, [x25]\n"
-      "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z4.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
-      "trn1 z4.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
-      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x458298e8  // usmmla z8.s, z7.b, z2.b\n"
-      ".inst 0x458298d0  // usmmla z16.s, z6.b, z2.b\n"
-      ".inst 0x45829898  // usmmla z24.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x458098ec  // usmmla z12.s, z7.b, z0.b\n"
-      ".inst 0x458098d4  // usmmla z20.s, z6.b, z0.b\n"
-      ".inst 0x4580989c  // usmmla z28.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x458298e9  // usmmla z9.s, z7.b, z2.b\n"
-      ".inst 0x458298d1  // usmmla z17.s, z6.b, z2.b\n"
-      ".inst 0x45829899  // usmmla z25.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x458098ed  // usmmla z13.s, z7.b, z0.b\n"
-      ".inst 0x458098d5  // usmmla z21.s, z6.b, z0.b\n"
-      ".inst 0x4580989d  // usmmla z29.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x458298ea  // usmmla z10.s, z7.b, z2.b\n"
-      ".inst 0x458298d2  // usmmla z18.s, z6.b, z2.b\n"
-      ".inst 0x4582989a  // usmmla z26.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x458098ee  // usmmla z14.s, z7.b, z0.b\n"
-      ".inst 0x458098d6  // usmmla z22.s, z6.b, z0.b\n"
-      ".inst 0x4580989e  // usmmla z30.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x458298eb  // usmmla z11.s, z7.b, z2.b\n"
-      ".inst 0x458298d3  // usmmla z19.s, z6.b, z2.b\n"
-      ".inst 0x4582989b  // usmmla z27.s, z4.b, z2.b\n"
-      ".inst 0x458098ef  // usmmla z15.s, z7.b, z0.b\n"
-      ".inst 0x458098d7  // usmmla z23.s, z6.b, z0.b\n"
-      ".inst 0x4580989f  // usmmla z31.s, z4.b, z0.b\n"
-      "ble 54f\n"
-      "ld1b { z2.b }, p5/Z, [x10]\n"
-      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45829828  // usmmla z8.s, z1.b, z2.b\n"
-      ".inst 0x45829870  // usmmla z16.s, z3.b, z2.b\n"
-      ".inst 0x458298b8  // usmmla z24.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4580982c  // usmmla z12.s, z1.b, z0.b\n"
-      ".inst 0x45809874  // usmmla z20.s, z3.b, z0.b\n"
-      ".inst 0x458098bc  // usmmla z28.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45829829  // usmmla z9.s, z1.b, z2.b\n"
-      ".inst 0x45829871  // usmmla z17.s, z3.b, z2.b\n"
-      ".inst 0x458298b9  // usmmla z25.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4580982d  // usmmla z13.s, z1.b, z0.b\n"
-      ".inst 0x45809875  // usmmla z21.s, z3.b, z0.b\n"
-      ".inst 0x458098bd  // usmmla z29.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4582982a  // usmmla z10.s, z1.b, z2.b\n"
-      ".inst 0x45829872  // usmmla z18.s, z3.b, z2.b\n"
-      ".inst 0x458298ba  // usmmla z26.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4580982e  // usmmla z14.s, z1.b, z0.b\n"
-      ".inst 0x45809876  // usmmla z22.s, z3.b, z0.b\n"
-      ".inst 0x458098be  // usmmla z30.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x4582982b  // usmmla z11.s, z1.b, z2.b\n"
-      ".inst 0x45829873  // usmmla z19.s, z3.b, z2.b\n"
-      ".inst 0x458298bb  // usmmla z27.s, z5.b, z2.b\n"
-      ".inst 0x4580982f  // usmmla z15.s, z1.b, z0.b\n"
-      ".inst 0x45809877  // usmmla z23.s, z3.b, z0.b\n"
-      ".inst 0x458098bf  // usmmla z31.s, z5.b, z0.b\n"
-      "54:"  // Height 5: Multiply loop: multiply skip
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 49b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z1.d, z8.d, z12.d\n"
-      "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z0.d, z9.d, z13.d\n"
-      "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z3.d, z10.d, z14.d\n"
-      "uzp2 z10.d, z10.d, z14.d\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "uzp1 z2.d, z11.d, z15.d\n"
-      "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z1.s }, p4, [x9]\n"
-      "add x20, x21, x20, LSL #2\n"
-      "uzp1 z1.d, z16.d, z20.d\n"
-      "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z0.s }, p3, [x9, #1, MUL VL]\n"
-      "uzp1 z0.d, z17.d, z21.d\n"
-      "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z3.s }, p2, [x9, #2, MUL VL]\n"
-      "uzp1 z21.d, z18.d, z22.d\n"
-      "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z2.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
-      "uzp1 z20.d, z19.d, z23.d\n"
-      "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x23]\n"
-      "uzp1 z24.d, z24.d, z28.d\n"
-      "uzp1 z25.d, z25.d, z29.d\n"
-      "st1w { z9.s }, p3, [x23, #1, MUL VL]\n"
-      "uzp1 z26.d, z26.d, z30.d\n"
-      "uzp1 z27.d, z27.d, z31.d\n"
-      "st1w { z10.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z11.s }, p1, [x23, #3, MUL VL]\n"
-      "st1w { z1.s }, p4, [x22]\n"
-      "st1w { z0.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z20.s }, p1, [x22, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x21]\n"
-      "st1w { z17.s }, p3, [x21, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x21, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x21, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x20]\n"
-      "st1w { z25.s }, p3, [x20, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x20, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x20, #3, MUL VL]\n"
-      "55:"  // Height 5: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
-      "bgt 46b\n"
-      "b 68f\n"
-      "56:"  // Height 6
-      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "mov x20, #0x18\n"
-      "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
-      "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
-      "57:"  // Height 6: Column loop
-      "mov x20, #0x0\n"
-      "whilelt p4.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p3.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p2.s, x20, x11\n"
-      "incw x20\n"
-      "whilelt p1.s, x20, x11\n"
-      "tbz %x[flags], #0, 58f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x24]\n"
-      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
-      "ld1w { z17.s }, p4/Z, [x23]\n"
-      "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x22]\n"
-      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x21]\n"
-      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z11.d, z16.d, z15.d\n"
-      "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z16.d, z17.d, z20.d\n"
-      "zip2 z20.d, z17.d, z20.d\n"
-      "ld1w { z28.s }, p4/Z, [x20]\n"
-      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "zip1 z17.d, z18.d, z21.d\n"
-      "zip2 z21.d, z18.d, z21.d\n"
-      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z18.d, z19.d, z22.d\n"
-      "zip2 z22.d, z19.d, z22.d\n"
-      "zip1 z19.d, z24.d, z23.d\n"
-      "zip2 z23.d, z24.d, z23.d\n"
-      "zip1 z24.d, z25.d, z28.d\n"
-      "zip2 z28.d, z25.d, z28.d\n"
-      "zip1 z25.d, z26.d, z29.d\n"
-      "zip2 z29.d, z26.d, z29.d\n"
-      "zip1 z26.d, z27.d, z30.d\n"
-      "zip2 z30.d, z27.d, z30.d\n"
-      "zip1 z27.d, z0.d, z31.d\n"
-      "zip2 z31.d, z0.d, z31.d\n"
-      "b 59f\n"
-      "58:"  // Height 6: no accumulate
-      "mov z8.s, #0x0\n"
-      "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "mov z15.s, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "mov z21.s, #0x0\n"
-      "mov z22.s, #0x0\n"
-      "mov z23.s, #0x0\n"
-      "mov z24.s, #0x0\n"
-      "mov z25.s, #0x0\n"
-      "mov z26.s, #0x0\n"
-      "mov z27.s, #0x0\n"
-      "mov z28.s, #0x0\n"
-      "mov z29.s, #0x0\n"
-      "mov z30.s, #0x0\n"
-      "mov z31.s, #0x0\n"
-      "59:"  // Height 6: setup done
-      "mov x28, #0x0\n"
-      "60:"  // Height 6: String loop
-      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "ldr w27, [x20, x28, LSL #0x2]\n"
-      "tbz %x[flags], #3, 61f\n"
-      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
-      "add x20, x20, x21, LSL #3\n"
-      "ldr x26, [x20, #0x0]\n"
-      "ldr x25, [x20, #0x8]\n"
-      "ldr x24, [x20, #0x10]\n"
-      "ldr x23, [x20, #0x18]\n"
-      "ldr x22, [x20, #0x20]\n"
-      "ldr x21, [x20, #0x28]\n"
-      "cbnz x28, 62f\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
-      "add x26, x26, x20\n"
-      "add x25, x25, x20\n"
-      "add x24, x24, x20\n"
-      "add x23, x23, x20\n"
-      "add x22, x22, x20\n"
-      "add x21, x21, x20\n"
-      "b 62f\n"
-      "61:"  // Height 6: setup direct input
-      "mov x26, %x[input_ptr]\n"
-      "add x25, x26, x21\n"
-      "add x24, x25, x21\n"
-      "add x23, x24, x21\n"
-      "add x22, x23, x21\n"
-      "add x21, x22, x21\n"
-      "62:"  // Height 6: input setup done
-      "cmp x27, #0x10\n"
-      "ble 64f\n"
-      "63:"  // Height 6: Multiply loop: Main loop head
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p5/Z, [x10]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z6.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z7.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "ld1rqb { z0.b }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
-      "add x21, x21, #0x10\n"
-      "trn1 z3.d, z7.d, z2.d\n"
-      "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
-      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45819888  // usmmla z8.s, z4.b, z1.b\n"
-      ".inst 0x45819870  // usmmla z16.s, z3.b, z1.b\n"
-      ".inst 0x45819858  // usmmla z24.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4580988c  // usmmla z12.s, z4.b, z0.b\n"
-      ".inst 0x45809874  // usmmla z20.s, z3.b, z0.b\n"
-      ".inst 0x4580985c  // usmmla z28.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45819889  // usmmla z9.s, z4.b, z1.b\n"
-      ".inst 0x45819871  // usmmla z17.s, z3.b, z1.b\n"
-      ".inst 0x45819859  // usmmla z25.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4580988d  // usmmla z13.s, z4.b, z0.b\n"
-      ".inst 0x45809875  // usmmla z21.s, z3.b, z0.b\n"
-      ".inst 0x4580985d  // usmmla z29.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4581988a  // usmmla z10.s, z4.b, z1.b\n"
-      ".inst 0x45819872  // usmmla z18.s, z3.b, z1.b\n"
-      ".inst 0x4581985a  // usmmla z26.s, z2.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4580988e  // usmmla z14.s, z4.b, z0.b\n"
-      ".inst 0x45809876  // usmmla z22.s, z3.b, z0.b\n"
-      ".inst 0x4580985e  // usmmla z30.s, z2.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x4581988b  // usmmla z11.s, z4.b, z1.b\n"
-      ".inst 0x45819873  // usmmla z19.s, z3.b, z1.b\n"
-      ".inst 0x4581985b  // usmmla z27.s, z2.b, z1.b\n"
-      ".inst 0x4580988f  // usmmla z15.s, z4.b, z0.b\n"
-      ".inst 0x45809877  // usmmla z23.s, z3.b, z0.b\n"
-      ".inst 0x4580985f  // usmmla z31.s, z2.b, z0.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z0.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x458198c8  // usmmla z8.s, z6.b, z1.b\n"
-      ".inst 0x458198f0  // usmmla z16.s, z7.b, z1.b\n"
-      ".inst 0x458198b8  // usmmla z24.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x458098cc  // usmmla z12.s, z6.b, z0.b\n"
-      ".inst 0x458098f4  // usmmla z20.s, z7.b, z0.b\n"
-      ".inst 0x458098bc  // usmmla z28.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x458198c9  // usmmla z9.s, z6.b, z1.b\n"
-      ".inst 0x458198f1  // usmmla z17.s, z7.b, z1.b\n"
-      ".inst 0x458198b9  // usmmla z25.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x458098cd  // usmmla z13.s, z6.b, z0.b\n"
-      ".inst 0x458098f5  // usmmla z21.s, z7.b, z0.b\n"
-      ".inst 0x458098bd  // usmmla z29.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x458198ca  // usmmla z10.s, z6.b, z1.b\n"
-      ".inst 0x458198f2  // usmmla z18.s, z7.b, z1.b\n"
-      ".inst 0x458198ba  // usmmla z26.s, z5.b, z1.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x458098ce  // usmmla z14.s, z6.b, z0.b\n"
-      ".inst 0x458098f6  // usmmla z22.s, z7.b, z0.b\n"
-      ".inst 0x458098be  // usmmla z30.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x458198cb  // usmmla z11.s, z6.b, z1.b\n"
-      ".inst 0x458198f3  // usmmla z19.s, z7.b, z1.b\n"
-      ".inst 0x458198bb  // usmmla z27.s, z5.b, z1.b\n"
-      ".inst 0x458098cf  // usmmla z15.s, z6.b, z0.b\n"
-      ".inst 0x458098f7  // usmmla z23.s, z7.b, z0.b\n"
-      ".inst 0x458098bf  // usmmla z31.s, z5.b, z0.b\n"
-      "bgt 63b\n"
-      "64:"  // Height 6: Multiply loop: Single iteration only
-      "whilelt p0.b, XZR, x27\n"
-      "ld1b { z2.b }, p5/Z, [x10]\n"
-      "subs x27, x27, #0x8\n"
-      "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z6.b }, p0/Z, [x25]\n"
-      "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z4.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "ld1rqb { z0.b }, p0/Z, [x21]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
-      "trn1 z4.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
-      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x458298e8  // usmmla z8.s, z7.b, z2.b\n"
-      ".inst 0x458298d0  // usmmla z16.s, z6.b, z2.b\n"
-      ".inst 0x45829898  // usmmla z24.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x458098ec  // usmmla z12.s, z7.b, z0.b\n"
-      ".inst 0x458098d4  // usmmla z20.s, z6.b, z0.b\n"
-      ".inst 0x4580989c  // usmmla z28.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x458298e9  // usmmla z9.s, z7.b, z2.b\n"
-      ".inst 0x458298d1  // usmmla z17.s, z6.b, z2.b\n"
-      ".inst 0x45829899  // usmmla z25.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x458098ed  // usmmla z13.s, z7.b, z0.b\n"
-      ".inst 0x458098d5  // usmmla z21.s, z6.b, z0.b\n"
-      ".inst 0x4580989d  // usmmla z29.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x458298ea  // usmmla z10.s, z7.b, z2.b\n"
-      ".inst 0x458298d2  // usmmla z18.s, z6.b, z2.b\n"
-      ".inst 0x4582989a  // usmmla z26.s, z4.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x458098ee  // usmmla z14.s, z7.b, z0.b\n"
-      ".inst 0x458098d6  // usmmla z22.s, z6.b, z0.b\n"
-      ".inst 0x4580989e  // usmmla z30.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x458298eb  // usmmla z11.s, z7.b, z2.b\n"
-      ".inst 0x458298d3  // usmmla z19.s, z6.b, z2.b\n"
-      ".inst 0x4582989b  // usmmla z27.s, z4.b, z2.b\n"
-      ".inst 0x458098ef  // usmmla z15.s, z7.b, z0.b\n"
-      ".inst 0x458098d7  // usmmla z23.s, z6.b, z0.b\n"
-      ".inst 0x4580989f  // usmmla z31.s, z4.b, z0.b\n"
-      "ble 65f\n"
-      "ld1b { z2.b }, p5/Z, [x10]\n"
-      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45829828  // usmmla z8.s, z1.b, z2.b\n"
-      ".inst 0x45829870  // usmmla z16.s, z3.b, z2.b\n"
-      ".inst 0x458298b8  // usmmla z24.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x4580982c  // usmmla z12.s, z1.b, z0.b\n"
-      ".inst 0x45809874  // usmmla z20.s, z3.b, z0.b\n"
-      ".inst 0x458098bc  // usmmla z28.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45829829  // usmmla z9.s, z1.b, z2.b\n"
-      ".inst 0x45829871  // usmmla z17.s, z3.b, z2.b\n"
-      ".inst 0x458298b9  // usmmla z25.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x4580982d  // usmmla z13.s, z1.b, z0.b\n"
-      ".inst 0x45809875  // usmmla z21.s, z3.b, z0.b\n"
-      ".inst 0x458098bd  // usmmla z29.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x4582982a  // usmmla z10.s, z1.b, z2.b\n"
-      ".inst 0x45829872  // usmmla z18.s, z3.b, z2.b\n"
-      ".inst 0x458298ba  // usmmla z26.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x4580982e  // usmmla z14.s, z1.b, z0.b\n"
-      ".inst 0x45809876  // usmmla z22.s, z3.b, z0.b\n"
-      ".inst 0x458098be  // usmmla z30.s, z5.b, z0.b\n"
-      "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
-      ".inst 0x4582982b  // usmmla z11.s, z1.b, z2.b\n"
-      ".inst 0x45829873  // usmmla z19.s, z3.b, z2.b\n"
-      ".inst 0x458298bb  // usmmla z27.s, z5.b, z2.b\n"
-      ".inst 0x4580982f  // usmmla z15.s, z1.b, z0.b\n"
-      ".inst 0x45809877  // usmmla z23.s, z3.b, z0.b\n"
-      ".inst 0x458098bf  // usmmla z31.s, z5.b, z0.b\n"
-      "65:"  // Height 6: Multiply loop: multiply skip
-      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
-      "add x28, x28, #0x1\n"
-      "cmp x28, x20\n"
-      "bne 60b\n"
-      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z0.d, z8.d, z12.d\n"
-      "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z12.d, z9.d, z13.d\n"
-      "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z13.d, z10.d, z14.d\n"
-      "uzp2 z10.d, z10.d, z14.d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "uzp1 z14.d, z11.d, z15.d\n"
-      "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z0.s }, p4, [x9]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "uzp1 z15.d, z16.d, z20.d\n"
-      "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z12.s }, p3, [x9, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
-      "uzp1 z20.d, z17.d, z21.d\n"
-      "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
-      "uzp1 z21.d, z18.d, z22.d\n"
-      "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
-      "uzp1 z22.d, z19.d, z23.d\n"
-      "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x24]\n"
-      "uzp1 z23.d, z24.d, z28.d\n"
-      "uzp2 z24.d, z24.d, z28.d\n"
-      "st1w { z9.s }, p3, [x24, #1, MUL VL]\n"
-      "uzp1 z28.d, z25.d, z29.d\n"
-      "uzp2 z25.d, z25.d, z29.d\n"
-      "st1w { z10.s }, p2, [x24, #2, MUL VL]\n"
-      "uzp1 z29.d, z26.d, z30.d\n"
-      "uzp2 z26.d, z26.d, z30.d\n"
-      "st1w { z11.s }, p1, [x24, #3, MUL VL]\n"
-      "uzp1 z30.d, z27.d, z31.d\n"
-      "uzp2 z27.d, z27.d, z31.d\n"
-      "st1w { z15.s }, p4, [x23]\n"
-      "st1w { z20.s }, p3, [x23, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z22.s }, p1, [x23, #3, MUL VL]\n"
-      "st1w { z16.s }, p4, [x22]\n"
-      "st1w { z17.s }, p3, [x22, #1, MUL VL]\n"
-      "st1w { z18.s }, p2, [x22, #2, MUL VL]\n"
-      "st1w { z19.s }, p1, [x22, #3, MUL VL]\n"
-      "st1w { z23.s }, p4, [x21]\n"
-      "st1w { z28.s }, p3, [x21, #1, MUL VL]\n"
-      "st1w { z29.s }, p2, [x21, #2, MUL VL]\n"
-      "st1w { z30.s }, p1, [x21, #3, MUL VL]\n"
-      "st1w { z24.s }, p4, [x20]\n"
-      "st1w { z25.s }, p3, [x20, #1, MUL VL]\n"
-      "st1w { z26.s }, p2, [x20, #2, MUL VL]\n"
-      "st1w { z27.s }, p1, [x20, #3, MUL VL]\n"
-      "66:"  // Height 6: Writeback done
-      "decw x11, ALL, MUL #4\n"
-      "cmp x11, XZR\n"
-      "bgt 57b\n"
-      "subs %x[M], %x[M], #0x6\n"
-      "beq 68f\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "tbz %x[flags], #3, 67f\n"
-      "add x21, x21, #0x6\n"
-      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
-      "b 1b\n"
-      "67:"  // Update direct input
-      "mov x20, #0x6\n"
-      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
-      "b 1b\n"
-      "68:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
-    );
-}
-
-} // namespace arm_gemm
-#endif  // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL.hpp
index 12e99fb526..920fca738c 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL.hpp
@@ -71,7 +71,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 4, 4> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 4> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/a64fx.cpp
index 3f074fad7d..6a0aeb79b4 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/a64fx.cpp
@@ -44,18 +44,18 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -89,7 +89,7 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "beq 11f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -114,8 +114,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "5:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 6f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -134,14 +134,14 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "ble 9f\n"
       "8:"  // Height 1: Multiply loop: Main loop
       "udot z8.s, z6.b, z0.b\n"
-      "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z7.b, z0.b\n"
+      "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
       "add x26, x26, #0x4\n"
-      "subs x27, x27, #0x4\n"
       "udot z10.s, z17.b, z0.b\n"
       "udot z11.s, z16.b, z0.b\n"
+      "subs x27, x27, #0x4\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1b { z6.b }, p4/Z, [x10]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
@@ -149,14 +149,14 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "9:"  // Height 1: Multiply loop: Main loop skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "udot z8.s, z6.b, z0.b\n"
-      "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z7.b, z0.b\n"
+      "ld1b { z17.b }, p4/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n"
       "add x28, x28, #0x1\n"
-      "addvl x10, x10, #4\n"
       "cmp x28, x20\n"
       "udot z10.s, z17.b, z0.b\n"
       "udot z11.s, z16.b, z0.b\n"
+      "addvl x10, x10, #4\n"
       "bne 5b\n"
       "st1w { z8.s }, p3, [x9]\n"
       "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
@@ -171,7 +171,7 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "11:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "12:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -183,11 +183,11 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 13f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x20]\n"
       "ld1w { z13.s }, p2/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x20, #2, MUL VL]\n"
@@ -206,8 +206,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "15:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 16f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -242,8 +242,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z10.s, z17.b, z0.b\n"
       "udot z14.s, z17.b, z1.b\n"
       "udot z11.s, z16.b, z0.b\n"
-      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "udot z15.s, z16.b, z1.b\n"
+      "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1b { z6.b }, p4/Z, [x10]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
@@ -257,18 +257,18 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z13.s, z7.b, z1.b\n"
       "ld1b { z16.b }, p4/Z, [x10, #3, MUL VL]\n"
       "add x28, x28, #0x1\n"
-      "addvl x10, x10, #4\n"
       "cmp x28, x20\n"
       "udot z10.s, z17.b, z0.b\n"
       "udot z14.s, z17.b, z1.b\n"
+      "addvl x10, x10, #4\n"
       "udot z11.s, z16.b, z0.b\n"
       "udot z15.s, z16.b, z1.b\n"
       "bne 15b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "st1w { z8.s }, p3, [x9]\n"
       "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p3, [x20]\n"
@@ -283,7 +283,7 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "21:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "22:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -295,12 +295,12 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 23f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x21]\n"
       "ld1w { z13.s }, p2/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x21, #2, MUL VL]\n"
@@ -327,8 +327,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "25:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 26f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -359,8 +359,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "add x26, x26, #0x4\n"
       "subs x27, x27, #0x4\n"
       "udot z16.s, z6.b, z2.b\n"
-      "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z7.b, z0.b\n"
+      "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n"
       "add x25, x25, #0x4\n"
       "udot z13.s, z7.b, z1.b\n"
       "udot z17.s, z7.b, z2.b\n"
@@ -372,11 +372,11 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z18.s, z21.b, z2.b\n"
       "udot z11.s, z20.b, z0.b\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
+      "ld1b { z6.b }, p4/Z, [x10]\n"
       "udot z15.s, z20.b, z1.b\n"
-      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "udot z19.s, z20.b, z2.b\n"
+      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
-      "ld1b { z6.b }, p4/Z, [x10]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 28b\n"
       "29:"  // Height 3: Multiply loop: Main loop skip
@@ -385,13 +385,13 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z12.s, z6.b, z1.b\n"
       "add x28, x28, #0x1\n"
       "udot z16.s, z6.b, z2.b\n"
-      "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z7.b, z0.b\n"
+      "ld1b { z21.b }, p4/Z, [x10, #2, MUL VL]\n"
+      "cmp x28, x20\n"
       "udot z13.s, z7.b, z1.b\n"
       "udot z17.s, z7.b, z2.b\n"
       "ld1b { z20.b }, p4/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
-      "cmp x28, x20\n"
       "udot z10.s, z21.b, z0.b\n"
       "udot z14.s, z21.b, z1.b\n"
       "udot z18.s, z21.b, z2.b\n"
@@ -400,11 +400,11 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z19.s, z20.b, z2.b\n"
       "bne 25b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "st1w { z8.s }, p3, [x9]\n"
       "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p3, [x21]\n"
@@ -423,7 +423,7 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "31:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "32:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -435,13 +435,13 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 33f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x22]\n"
       "ld1w { z13.s }, p2/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p1/Z, [x22, #2, MUL VL]\n"
@@ -476,8 +476,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "35:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 36f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -527,6 +527,7 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z14.s, z25.b, z1.b\n"
       "udot z18.s, z25.b, z2.b\n"
       "udot z22.s, z25.b, z3.b\n"
+      "ld1b { z6.b }, p4/Z, [x10]\n"
       "udot z11.s, z24.b, z0.b\n"
       "udot z15.s, z24.b, z1.b\n"
       "ld1rw { z0.s }, p4/Z, [x26]\n"
@@ -535,7 +536,6 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z23.s, z24.b, z3.b\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
       "ld1rw { z3.s }, p4/Z, [x23]\n"
-      "ld1b { z6.b }, p4/Z, [x10]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 38b\n"
       "39:"  // Height 4: Multiply loop: Main loop skip
@@ -546,15 +546,15 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z16.s, z6.b, z2.b\n"
       "udot z20.s, z6.b, z3.b\n"
       "ld1b { z25.b }, p4/Z, [x10, #2, MUL VL]\n"
+      "cmp x28, x20\n"
       "udot z9.s, z7.b, z0.b\n"
       "udot z13.s, z7.b, z1.b\n"
       "udot z17.s, z7.b, z2.b\n"
       "udot z21.s, z7.b, z3.b\n"
       "ld1b { z24.b }, p4/Z, [x10, #3, MUL VL]\n"
-      "cmp x28, x20\n"
+      "addvl x10, x10, #4\n"
       "udot z10.s, z25.b, z0.b\n"
       "udot z14.s, z25.b, z1.b\n"
-      "addvl x10, x10, #4\n"
       "udot z18.s, z25.b, z2.b\n"
       "udot z22.s, z25.b, z3.b\n"
       "udot z11.s, z24.b, z0.b\n"
@@ -563,12 +563,12 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z23.s, z24.b, z3.b\n"
       "bne 35b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p3, [x9]\n"
-      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "add x22, x9, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "st1w { z8.s }, p3, [x9]\n"
       "add x20, x21, x20, LSL #2\n"
+      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
+      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p3, [x22]\n"
@@ -591,7 +591,7 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "41:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "42:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -603,16 +603,16 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 43f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x23]\n"
       "ld1w { z13.s }, p2/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p1/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p0/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p3/Z, [x22]\n"
@@ -653,8 +653,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "45:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 46f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -697,8 +697,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "add x25, x25, #0x4\n"
       "add x24, x24, #0x4\n"
       "udot z24.s, z6.b, z4.b\n"
-      "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z7.b, z0.b\n"
+      "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n"
       "add x23, x23, #0x4\n"
       "udot z13.s, z7.b, z1.b\n"
       "udot z17.s, z7.b, z2.b\n"
@@ -716,12 +716,12 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "ld1rw { z0.s }, p4/Z, [x26]\n"
       "ld1b { z6.b }, p4/Z, [x10]\n"
       "udot z15.s, z28.b, z1.b\n"
-      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "udot z19.s, z28.b, z2.b\n"
+      "ld1rw { z1.s }, p4/Z, [x25]\n"
       "ld1rw { z2.s }, p4/Z, [x24]\n"
       "udot z23.s, z28.b, z3.b\n"
-      "ld1rw { z3.s }, p4/Z, [x23]\n"
       "udot z27.s, z28.b, z4.b\n"
+      "ld1rw { z3.s }, p4/Z, [x23]\n"
       "ld1rw { z4.s }, p4/Z, [x22]\n"
       "ld1b { z7.b }, p4/Z, [x10, #1, MUL VL]\n"
       "bgt 48b\n"
@@ -732,12 +732,12 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "udot z16.s, z6.b, z2.b\n"
       "udot z20.s, z6.b, z3.b\n"
+      "cmp x28, x20\n"
       "udot z24.s, z6.b, z4.b\n"
-      "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z7.b, z0.b\n"
+      "ld1b { z29.b }, p4/Z, [x10, #2, MUL VL]\n"
       "udot z13.s, z7.b, z1.b\n"
       "udot z17.s, z7.b, z2.b\n"
-      "cmp x28, x20\n"
       "udot z21.s, z7.b, z3.b\n"
       "udot z25.s, z7.b, z4.b\n"
       "ld1b { z28.b }, p4/Z, [x10, #3, MUL VL]\n"
@@ -754,15 +754,15 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z27.s, z28.b, z4.b\n"
       "bne 45b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p3, [x9]\n"
-      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "st1w { z8.s }, p3, [x9]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
+      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z12.s }, p3, [x23]\n"
       "st1w { z13.s }, p2, [x23, #1, MUL VL]\n"
       "st1w { z14.s }, p1, [x23, #2, MUL VL]\n"
@@ -786,12 +786,11 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "b 62f\n"
       "51:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "52:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p3.s, x20, x11\n"
@@ -803,17 +802,17 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "whilelt p0.s, x20, x11\n"
       "tbz %x[flags], #0, 53f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p3/Z, [x9]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p2/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p1/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p0/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p3/Z, [x24]\n"
       "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p3/Z, [x23]\n"
@@ -862,8 +861,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "mov x28, #0x0\n"
       "55:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 56f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -951,12 +950,12 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "add x28, x28, #0x1\n"
       "udot z16.s, z6.b, z2.b\n"
       "udot z20.s, z6.b, z3.b\n"
+      "cmp x28, x20\n"
       "udot z24.s, z6.b, z4.b\n"
       "udot z28.s, z6.b, z5.b\n"
       "ld1b { z6.b }, p4/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z7.b, z0.b\n"
       "udot z13.s, z7.b, z1.b\n"
-      "cmp x28, x20\n"
       "udot z17.s, z7.b, z2.b\n"
       "udot z21.s, z7.b, z3.b\n"
       "udot z25.s, z7.b, z4.b\n"
@@ -977,17 +976,17 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "udot z31.s, z7.b, z5.b\n"
       "bne 55b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p3, [x9]\n"
-      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "st1w { z8.s }, p3, [x9]\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
+      "st1w { z9.s }, p2, [x9, #1, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
+      "st1w { z10.s }, p1, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p0, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "add x21, x22, x20, LSL #2\n"
       "st1w { z12.s }, p3, [x24]\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z13.s }, p2, [x24, #1, MUL VL]\n"
       "st1w { z14.s }, p1, [x24, #2, MUL VL]\n"
       "st1w { z15.s }, p0, [x24, #3, MUL VL]\n"
@@ -1023,8 +1022,8 @@ void sve_hybrid_u8u32_dot_6x4VL_a64fx (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "62:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp
index f2dee3c0bb..03b41dabe4 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp
@@ -44,18 +44,18 @@ void sve_hybrid_u8u32_dot_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -89,7 +89,7 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "beq 12f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -114,8 +114,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "mov x28, #0x0\n"
       "5:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 6f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -131,89 +131,89 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 9f\n"
       "8:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "udot z8.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10]\n"
+      "udot z8.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "udot z10.s, z16.b, z0.b[0]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "udot z10.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "udot z11.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "udot z8.s, z16.b, z0.b[1]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      "udot z8.s, z17.b, z0.b[1]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[1]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #6, MUL VL]\n"
+      "udot z10.s, z16.b, z0.b[1]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      "udot z10.s, z17.b, z0.b[1]\n"
       "udot z11.s, z16.b, z0.b[1]\n"
       "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
       "udot z8.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[2]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n"
       "udot z10.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "udot z11.s, z16.b, z0.b[2]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-3, MUL VL]\n"
       "udot z8.s, z17.b, z0.b[3]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[3]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
       "udot z10.s, z17.b, z0.b[3]\n"
       "udot z11.s, z16.b, z0.b[3]\n"
+      "add x26, x26, #0x10\n"
       "bgt 8b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
-      "udot z8.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10]\n"
+      "udot z8.s, z16.b, z0.b[0]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[0]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "udot z10.s, z17.b, z0.b[0]\n"
       "udot z11.s, z16.b, z0.b[0]\n"
+      "addvl x10, x10, #4\n"
       "ble 10f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z17.b, z0.b[1]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[1]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x4\n"
       "udot z10.s, z17.b, z0.b[1]\n"
       "udot z11.s, z16.b, z0.b[1]\n"
+      "addvl x10, x10, #4\n"
       "ble 10f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z17.b, z0.b[2]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[2]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x4\n"
       "udot z10.s, z17.b, z0.b[2]\n"
       "udot z11.s, z16.b, z0.b[2]\n"
+      "addvl x10, x10, #4\n"
       "ble 10f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z8.s, z17.b, z0.b[3]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[3]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "udot z10.s, z17.b, z0.b[3]\n"
       "udot z11.s, z16.b, z0.b[3]\n"
+      "addvl x10, x10, #4\n"
       "10:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -232,7 +232,7 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "12:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "13:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -244,11 +244,11 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 14f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
@@ -267,8 +267,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "mov x28, #0x0\n"
       "16:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 17f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -287,38 +287,38 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 20f\n"
       "19:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z8.s, z17.b, z1.b[0]\n"
       "udot z12.s, z17.b, z0.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z16.b, z1.b[0]\n"
       "udot z13.s, z16.b, z0.b[0]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
       "udot z10.s, z17.b, z1.b[0]\n"
       "udot z14.s, z17.b, z0.b[0]\n"
       "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "cmp x27, #0x10\n"
       "udot z11.s, z16.b, z1.b[0]\n"
       "udot z15.s, z16.b, z0.b[0]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       "udot z8.s, z17.b, z1.b[1]\n"
       "udot z12.s, z17.b, z0.b[1]\n"
       "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
+      "add x25, x25, #0x10\n"
       "udot z9.s, z16.b, z1.b[1]\n"
       "udot z13.s, z16.b, z0.b[1]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
       "udot z10.s, z17.b, z1.b[1]\n"
       "udot z14.s, z17.b, z0.b[1]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "udot z11.s, z16.b, z1.b[1]\n"
       "udot z15.s, z16.b, z0.b[1]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
       "udot z8.s, z17.b, z1.b[2]\n"
       "udot z12.s, z17.b, z0.b[2]\n"
@@ -345,50 +345,50 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "bgt 19b\n"
       "20:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z8.s, z17.b, z0.b[0]\n"
       "udot z12.s, z17.b, z1.b[0]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[0]\n"
       "udot z13.s, z16.b, z1.b[0]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "udot z10.s, z17.b, z0.b[0]\n"
       "udot z14.s, z17.b, z1.b[0]\n"
+      "addvl x10, x10, #4\n"
       "udot z11.s, z16.b, z0.b[0]\n"
       "udot z15.s, z16.b, z1.b[0]\n"
       "ble 21f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z17.b, z0.b[1]\n"
       "udot z12.s, z17.b, z1.b[1]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[1]\n"
       "udot z13.s, z16.b, z1.b[1]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x4\n"
       "udot z10.s, z17.b, z0.b[1]\n"
       "udot z14.s, z17.b, z1.b[1]\n"
+      "addvl x10, x10, #4\n"
       "udot z11.s, z16.b, z0.b[1]\n"
       "udot z15.s, z16.b, z1.b[1]\n"
       "ble 21f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z17.b, z0.b[2]\n"
       "udot z12.s, z17.b, z1.b[2]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[2]\n"
       "udot z13.s, z16.b, z1.b[2]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
+      "subs x27, x27, #0x4\n"
       "udot z10.s, z17.b, z0.b[2]\n"
       "udot z14.s, z17.b, z1.b[2]\n"
+      "addvl x10, x10, #4\n"
       "udot z11.s, z16.b, z0.b[2]\n"
       "udot z15.s, z16.b, z1.b[2]\n"
       "ble 21f\n"
@@ -396,13 +396,13 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z8.s, z17.b, z0.b[3]\n"
       "udot z12.s, z17.b, z1.b[3]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z16.b, z0.b[3]\n"
       "udot z13.s, z16.b, z1.b[3]\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "addvl x10, x10, #4\n"
       "udot z10.s, z17.b, z0.b[3]\n"
       "udot z14.s, z17.b, z1.b[3]\n"
+      "addvl x10, x10, #4\n"
       "udot z11.s, z16.b, z0.b[3]\n"
       "udot z15.s, z16.b, z1.b[3]\n"
       "21:"  // Height 2: Multiply loop: multiply skip
@@ -411,10 +411,10 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 16b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "st1w { z8.s }, p4, [x9]\n"
       "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p4, [x20]\n"
@@ -429,7 +429,7 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "23:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "24:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -441,12 +441,12 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 25f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
@@ -473,8 +473,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "mov x28, #0x0\n"
       "27:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 28f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -496,37 +496,37 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 31f\n"
       "30:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z21.b }, p5/Z, [x10]\n"
-      "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
+      "ld1b { z21.b }, p5/Z, [x10]\n"
       "udot z8.s, z21.b, z2.b[0]\n"
       "udot z12.s, z21.b, z1.b[0]\n"
-      "udot z9.s, z20.b, z2.b[0]\n"
-      "udot z13.s, z20.b, z1.b[0]\n"
+      "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z16.s, z21.b, z0.b[0]\n"
+      "udot z9.s, z20.b, z2.b[0]\n"
       "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "udot z13.s, z20.b, z1.b[0]\n"
       "udot z17.s, z20.b, z0.b[0]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
+      "cmp x27, #0x10\n"
       "udot z10.s, z21.b, z2.b[0]\n"
       "udot z14.s, z21.b, z1.b[0]\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "udot z18.s, z21.b, z0.b[0]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #4, MUL VL]\n"
       "udot z11.s, z20.b, z2.b[0]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "add x24, x24, #0x10\n"
       "udot z15.s, z20.b, z1.b[0]\n"
       "udot z19.s, z20.b, z0.b[0]\n"
       "ld1b { z20.b }, p5/Z, [x10, #5, MUL VL]\n"
       "udot z8.s, z21.b, z2.b[1]\n"
       "udot z12.s, z21.b, z1.b[1]\n"
       "udot z16.s, z21.b, z0.b[1]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #6, MUL VL]\n"
       "udot z9.s, z20.b, z2.b[1]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #6, MUL VL]\n"
       "udot z13.s, z20.b, z1.b[1]\n"
       "udot z17.s, z20.b, z0.b[1]\n"
       "ld1b { z20.b }, p5/Z, [x10, #7, MUL VL]\n"
@@ -535,31 +535,31 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z14.s, z21.b, z1.b[1]\n"
       "udot z18.s, z21.b, z0.b[1]\n"
       "udot z11.s, z20.b, z2.b[1]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "udot z15.s, z20.b, z1.b[1]\n"
       "udot z19.s, z20.b, z0.b[1]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z20.b }, p5/Z, [x10, #-7, MUL VL]\n"
       "udot z8.s, z21.b, z2.b[2]\n"
       "udot z12.s, z21.b, z1.b[2]\n"
       "udot z16.s, z21.b, z0.b[2]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "udot z9.s, z20.b, z2.b[2]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "udot z13.s, z20.b, z1.b[2]\n"
       "udot z17.s, z20.b, z0.b[2]\n"
       "ld1b { z20.b }, p5/Z, [x10, #-5, MUL VL]\n"
       "udot z10.s, z21.b, z2.b[2]\n"
       "udot z14.s, z21.b, z1.b[2]\n"
       "udot z18.s, z21.b, z0.b[2]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "udot z11.s, z20.b, z2.b[2]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "udot z15.s, z20.b, z1.b[2]\n"
       "udot z19.s, z20.b, z0.b[2]\n"
       "ld1b { z20.b }, p5/Z, [x10, #-3, MUL VL]\n"
       "udot z8.s, z21.b, z2.b[3]\n"
       "udot z12.s, z21.b, z1.b[3]\n"
       "udot z16.s, z21.b, z0.b[3]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "udot z9.s, z20.b, z2.b[3]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "udot z13.s, z20.b, z1.b[3]\n"
       "udot z17.s, z20.b, z0.b[3]\n"
       "ld1b { z20.b }, p5/Z, [x10, #-1, MUL VL]\n"
@@ -572,18 +572,18 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "bgt 30b\n"
       "31:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z21.b }, p5/Z, [x10]\n"
-      "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
+      "ld1b { z21.b }, p5/Z, [x10]\n"
       "udot z8.s, z21.b, z0.b[0]\n"
       "udot z12.s, z21.b, z1.b[0]\n"
-      "udot z9.s, z20.b, z0.b[0]\n"
-      "udot z13.s, z20.b, z1.b[0]\n"
+      "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z16.s, z21.b, z2.b[0]\n"
+      "udot z9.s, z20.b, z0.b[0]\n"
       "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "udot z13.s, z20.b, z1.b[0]\n"
       "udot z17.s, z20.b, z2.b[0]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
       "addvl x10, x10, #4\n"
@@ -596,12 +596,12 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 32f\n"
       "ld1b { z21.b }, p5/Z, [x10]\n"
       "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z21.b, z0.b[1]\n"
       "udot z12.s, z21.b, z1.b[1]\n"
       "udot z16.s, z21.b, z2.b[1]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z20.b, z0.b[1]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "udot z13.s, z20.b, z1.b[1]\n"
       "udot z17.s, z20.b, z2.b[1]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -615,12 +615,12 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 32f\n"
       "ld1b { z21.b }, p5/Z, [x10]\n"
       "ld1b { z20.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z21.b, z0.b[2]\n"
       "udot z12.s, z21.b, z1.b[2]\n"
       "udot z16.s, z21.b, z2.b[2]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z20.b, z0.b[2]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "udot z13.s, z20.b, z1.b[2]\n"
       "udot z17.s, z20.b, z2.b[2]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -637,8 +637,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z8.s, z21.b, z0.b[3]\n"
       "udot z12.s, z21.b, z1.b[3]\n"
       "udot z16.s, z21.b, z2.b[3]\n"
-      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z20.b, z0.b[3]\n"
+      "ld1b { z21.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z13.s, z20.b, z1.b[3]\n"
       "udot z17.s, z20.b, z2.b[3]\n"
       "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -655,11 +655,11 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 27b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "st1w { z8.s }, p4, [x9]\n"
       "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p4, [x21]\n"
@@ -678,7 +678,7 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "34:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "35:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -690,13 +690,13 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 36f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
@@ -731,8 +731,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "mov x28, #0x0\n"
       "38:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 39f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -757,25 +757,25 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 42f\n"
       "41:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z8.s, z25.b, z3.b[0]\n"
       "udot z12.s, z25.b, z2.b[0]\n"
-      "udot z9.s, z24.b, z3.b[0]\n"
-      "udot z13.s, z24.b, z2.b[0]\n"
       "udot z16.s, z25.b, z1.b[0]\n"
       "udot z20.s, z25.b, z0.b[0]\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      "udot z9.s, z24.b, z3.b[0]\n"
+      "udot z13.s, z24.b, z2.b[0]\n"
+      "add x24, x24, #0x10\n"
+      "add x23, x23, #0x10\n"
       "udot z17.s, z24.b, z1.b[0]\n"
       "udot z21.s, z24.b, z0.b[0]\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -804,9 +804,9 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z14.s, z25.b, z2.b[1]\n"
       "udot z18.s, z25.b, z1.b[1]\n"
       "udot z22.s, z25.b, z0.b[1]\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "udot z11.s, z24.b, z3.b[1]\n"
       "udot z15.s, z24.b, z2.b[1]\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "udot z19.s, z24.b, z1.b[1]\n"
       "udot z23.s, z24.b, z0.b[1]\n"
       "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
@@ -851,20 +851,20 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "bgt 41b\n"
       "42:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z8.s, z25.b, z0.b[0]\n"
       "udot z12.s, z25.b, z1.b[0]\n"
-      "udot z9.s, z24.b, z0.b[0]\n"
-      "udot z13.s, z24.b, z1.b[0]\n"
       "udot z16.s, z25.b, z2.b[0]\n"
       "udot z20.s, z25.b, z3.b[0]\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "udot z9.s, z24.b, z0.b[0]\n"
+      "udot z13.s, z24.b, z1.b[0]\n"
       "udot z17.s, z24.b, z2.b[0]\n"
       "udot z21.s, z24.b, z3.b[0]\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -880,12 +880,12 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 43f\n"
       "ld1b { z25.b }, p5/Z, [x10]\n"
       "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z25.b, z0.b[1]\n"
       "udot z12.s, z25.b, z1.b[1]\n"
       "udot z16.s, z25.b, z2.b[1]\n"
       "udot z20.s, z25.b, z3.b[1]\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "udot z9.s, z24.b, z0.b[1]\n"
       "udot z13.s, z24.b, z1.b[1]\n"
       "udot z17.s, z24.b, z2.b[1]\n"
@@ -903,12 +903,12 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 43f\n"
       "ld1b { z25.b }, p5/Z, [x10]\n"
       "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z25.b, z0.b[2]\n"
       "udot z12.s, z25.b, z1.b[2]\n"
       "udot z16.s, z25.b, z2.b[2]\n"
       "udot z20.s, z25.b, z3.b[2]\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x4\n"
       "udot z9.s, z24.b, z0.b[2]\n"
       "udot z13.s, z24.b, z1.b[2]\n"
       "udot z17.s, z24.b, z2.b[2]\n"
@@ -951,12 +951,12 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 38b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p4, [x9]\n"
-      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "add x22, x9, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "st1w { z8.s }, p4, [x9]\n"
       "add x20, x21, x20, LSL #2\n"
+      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z12.s }, p4, [x22]\n"
@@ -979,7 +979,7 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "45:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "46:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -991,16 +991,16 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 47f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x23, x9, x20, LSL #2\n"
+      "add x22, x23, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x23, x9, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x22]\n"
@@ -1041,8 +1041,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "mov x28, #0x0\n"
       "49:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 50f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1070,29 +1070,29 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 53f\n"
       "52:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z4.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "ld1rqb { z0.b }, p0/Z, [x22]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "ld1b { z29.b }, p5/Z, [x10]\n"
       "udot z8.s, z29.b, z4.b[0]\n"
       "udot z12.s, z29.b, z3.b[0]\n"
-      "udot z9.s, z28.b, z4.b[0]\n"
+      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z16.s, z29.b, z2.b[0]\n"
       "udot z20.s, z29.b, z1.b[0]\n"
+      "add x25, x25, #0x10\n"
       "udot z24.s, z29.b, z0.b[0]\n"
-      "udot z13.s, z28.b, z3.b[0]\n"
+      "udot z9.s, z28.b, z4.b[0]\n"
       "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      "udot z13.s, z28.b, z3.b[0]\n"
       "udot z17.s, z28.b, z2.b[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "udot z21.s, z28.b, z1.b[0]\n"
       "udot z25.s, z28.b, z0.b[0]\n"
       "ld1b { z28.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1101,8 +1101,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z18.s, z29.b, z2.b[0]\n"
       "udot z22.s, z29.b, z1.b[0]\n"
       "udot z26.s, z29.b, z0.b[0]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #4, MUL VL]\n"
       "udot z11.s, z28.b, z4.b[0]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #4, MUL VL]\n"
       "udot z15.s, z28.b, z3.b[0]\n"
       "udot z19.s, z28.b, z2.b[0]\n"
       "udot z23.s, z28.b, z1.b[0]\n"
@@ -1113,8 +1113,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z16.s, z29.b, z2.b[1]\n"
       "udot z20.s, z29.b, z1.b[1]\n"
       "udot z24.s, z29.b, z0.b[1]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #6, MUL VL]\n"
       "udot z9.s, z28.b, z4.b[1]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #6, MUL VL]\n"
       "udot z13.s, z28.b, z3.b[1]\n"
       "udot z17.s, z28.b, z2.b[1]\n"
       "udot z21.s, z28.b, z1.b[1]\n"
@@ -1127,8 +1127,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z22.s, z29.b, z1.b[1]\n"
       "udot z26.s, z29.b, z0.b[1]\n"
       "udot z11.s, z28.b, z4.b[1]\n"
-      "udot z15.s, z28.b, z3.b[1]\n"
       "ld1b { z29.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      "udot z15.s, z28.b, z3.b[1]\n"
       "udot z19.s, z28.b, z2.b[1]\n"
       "udot z23.s, z28.b, z1.b[1]\n"
       "udot z27.s, z28.b, z0.b[1]\n"
@@ -1138,8 +1138,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z16.s, z29.b, z2.b[2]\n"
       "udot z20.s, z29.b, z1.b[2]\n"
       "udot z24.s, z29.b, z0.b[2]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "udot z9.s, z28.b, z4.b[2]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #-6, MUL VL]\n"
       "udot z13.s, z28.b, z3.b[2]\n"
       "udot z17.s, z28.b, z2.b[2]\n"
       "udot z21.s, z28.b, z1.b[2]\n"
@@ -1150,8 +1150,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z18.s, z29.b, z2.b[2]\n"
       "udot z22.s, z29.b, z1.b[2]\n"
       "udot z26.s, z29.b, z0.b[2]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "udot z11.s, z28.b, z4.b[2]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #-4, MUL VL]\n"
       "udot z15.s, z28.b, z3.b[2]\n"
       "udot z19.s, z28.b, z2.b[2]\n"
       "udot z23.s, z28.b, z1.b[2]\n"
@@ -1162,8 +1162,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z16.s, z29.b, z2.b[3]\n"
       "udot z20.s, z29.b, z1.b[3]\n"
       "udot z24.s, z29.b, z0.b[3]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "udot z9.s, z28.b, z4.b[3]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #-2, MUL VL]\n"
       "udot z13.s, z28.b, z3.b[3]\n"
       "udot z17.s, z28.b, z2.b[3]\n"
       "udot z21.s, z28.b, z1.b[3]\n"
@@ -1182,23 +1182,23 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "bgt 52b\n"
       "53:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
       "ld1rqb { z4.b }, p0/Z, [x22]\n"
+      "ld1b { z29.b }, p5/Z, [x10]\n"
       "udot z8.s, z29.b, z0.b[0]\n"
       "udot z12.s, z29.b, z1.b[0]\n"
-      "udot z9.s, z28.b, z0.b[0]\n"
-      "udot z13.s, z28.b, z1.b[0]\n"
+      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z16.s, z29.b, z2.b[0]\n"
       "udot z20.s, z29.b, z3.b[0]\n"
       "udot z24.s, z29.b, z4.b[0]\n"
-      "udot z17.s, z28.b, z2.b[0]\n"
+      "udot z9.s, z28.b, z0.b[0]\n"
       "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "udot z13.s, z28.b, z1.b[0]\n"
+      "udot z17.s, z28.b, z2.b[0]\n"
       "udot z21.s, z28.b, z3.b[0]\n"
       "udot z25.s, z28.b, z4.b[0]\n"
       "ld1b { z28.b }, p5/Z, [x10, #3, MUL VL]\n"
@@ -1216,21 +1216,21 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 54f\n"
       "ld1b { z29.b }, p5/Z, [x10]\n"
       "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z29.b, z0.b[1]\n"
       "udot z12.s, z29.b, z1.b[1]\n"
       "udot z16.s, z29.b, z2.b[1]\n"
       "udot z20.s, z29.b, z3.b[1]\n"
+      "subs x27, x27, #0x4\n"
       "udot z24.s, z29.b, z4.b[1]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z28.b, z0.b[1]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z13.s, z28.b, z1.b[1]\n"
       "udot z17.s, z28.b, z2.b[1]\n"
       "udot z21.s, z28.b, z3.b[1]\n"
       "udot z25.s, z28.b, z4.b[1]\n"
       "ld1b { z28.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "udot z10.s, z29.b, z0.b[1]\n"
       "addvl x10, x10, #4\n"
+      "udot z10.s, z29.b, z0.b[1]\n"
       "udot z14.s, z29.b, z1.b[1]\n"
       "udot z18.s, z29.b, z2.b[1]\n"
       "udot z22.s, z29.b, z3.b[1]\n"
@@ -1243,21 +1243,21 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 54f\n"
       "ld1b { z29.b }, p5/Z, [x10]\n"
       "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z29.b, z0.b[2]\n"
       "udot z12.s, z29.b, z1.b[2]\n"
       "udot z16.s, z29.b, z2.b[2]\n"
       "udot z20.s, z29.b, z3.b[2]\n"
+      "subs x27, x27, #0x4\n"
       "udot z24.s, z29.b, z4.b[2]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z28.b, z0.b[2]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z13.s, z28.b, z1.b[2]\n"
       "udot z17.s, z28.b, z2.b[2]\n"
       "udot z21.s, z28.b, z3.b[2]\n"
       "udot z25.s, z28.b, z4.b[2]\n"
       "ld1b { z28.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "udot z10.s, z29.b, z0.b[2]\n"
       "addvl x10, x10, #4\n"
+      "udot z10.s, z29.b, z0.b[2]\n"
       "udot z14.s, z29.b, z1.b[2]\n"
       "udot z18.s, z29.b, z2.b[2]\n"
       "udot z22.s, z29.b, z3.b[2]\n"
@@ -1275,8 +1275,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "udot z16.s, z29.b, z2.b[3]\n"
       "udot z20.s, z29.b, z3.b[3]\n"
       "udot z24.s, z29.b, z4.b[3]\n"
-      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z9.s, z28.b, z0.b[3]\n"
+      "ld1b { z29.b }, p5/Z, [x10, #2, MUL VL]\n"
       "udot z13.s, z28.b, z1.b[3]\n"
       "udot z17.s, z28.b, z2.b[3]\n"
       "udot z21.s, z28.b, z3.b[3]\n"
@@ -1299,15 +1299,15 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 49b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p4, [x9]\n"
-      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "st1w { z8.s }, p4, [x9]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
+      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z12.s }, p4, [x23]\n"
       "st1w { z13.s }, p3, [x23, #1, MUL VL]\n"
       "st1w { z14.s }, p2, [x23, #2, MUL VL]\n"
@@ -1331,12 +1331,11 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "b 68f\n"
       "56:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "57:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1348,17 +1347,17 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 58f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "ld1w { z8.s }, p4/Z, [x9]\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "ld1w { z9.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z10.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
       "ld1w { z11.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x24]\n"
       "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z16.s }, p4/Z, [x23]\n"
@@ -1407,8 +1406,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "mov x28, #0x0\n"
       "60:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 61f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1439,29 +1438,29 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 64f\n"
       "63:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p5/Z, [x10]\n"
-      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z7.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
       "ld1rqb { z6.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "sub x27, x27, #0x10\n"
       "ld1rqb { z5.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
       "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "ld1rqb { z2.b }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
+      "add x25, x25, #0x10\n"
+      "add x24, x24, #0x10\n"
+      "ld1b { z1.b }, p5/Z, [x10]\n"
+      "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z8.s, z1.b, z7.b[0]\n"
       "udot z12.s, z1.b, z6.b[0]\n"
-      "add x21, x21, #0x10\n"
       "udot z16.s, z1.b, z5.b[0]\n"
       "udot z20.s, z1.b, z4.b[0]\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       "udot z24.s, z1.b, z3.b[0]\n"
       "udot z28.s, z1.b, z2.b[0]\n"
       "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "add x21, x21, #0x10\n"
       "udot z9.s, z0.b, z7.b[0]\n"
       "udot z13.s, z0.b, z6.b[0]\n"
       "udot z17.s, z0.b, z5.b[0]\n"
@@ -1569,24 +1568,24 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "bgt 63b\n"
       "64:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z7.b }, p5/Z, [x10]\n"
-      "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "ld1rqb { z0.b }, p0/Z, [x26]\n"
       "ld1rqb { z1.b }, p0/Z, [x25]\n"
+      "subs x27, x27, #0x4\n"
       "ld1rqb { z2.b }, p0/Z, [x24]\n"
       "ld1rqb { z3.b }, p0/Z, [x23]\n"
       "ld1rqb { z4.b }, p0/Z, [x22]\n"
       "ld1rqb { z5.b }, p0/Z, [x21]\n"
+      "ld1b { z7.b }, p5/Z, [x10]\n"
+      "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n"
       "udot z8.s, z7.b, z0.b[0]\n"
       "udot z12.s, z7.b, z1.b[0]\n"
-      "udot z9.s, z6.b, z0.b[0]\n"
-      "udot z13.s, z6.b, z1.b[0]\n"
       "udot z16.s, z7.b, z2.b[0]\n"
       "udot z20.s, z7.b, z3.b[0]\n"
       "udot z24.s, z7.b, z4.b[0]\n"
       "udot z28.s, z7.b, z5.b[0]\n"
       "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "udot z9.s, z6.b, z0.b[0]\n"
+      "udot z13.s, z6.b, z1.b[0]\n"
       "udot z17.s, z6.b, z2.b[0]\n"
       "udot z21.s, z6.b, z3.b[0]\n"
       "udot z25.s, z6.b, z4.b[0]\n"
@@ -1608,23 +1607,23 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 65f\n"
       "ld1b { z7.b }, p5/Z, [x10]\n"
       "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z7.b, z0.b[1]\n"
       "udot z12.s, z7.b, z1.b[1]\n"
       "udot z16.s, z7.b, z2.b[1]\n"
       "udot z20.s, z7.b, z3.b[1]\n"
+      "subs x27, x27, #0x4\n"
       "udot z24.s, z7.b, z4.b[1]\n"
       "udot z28.s, z7.b, z5.b[1]\n"
-      "udot z9.s, z6.b, z0.b[1]\n"
       "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "udot z9.s, z6.b, z0.b[1]\n"
       "udot z13.s, z6.b, z1.b[1]\n"
       "udot z17.s, z6.b, z2.b[1]\n"
       "udot z21.s, z6.b, z3.b[1]\n"
       "udot z25.s, z6.b, z4.b[1]\n"
       "udot z29.s, z6.b, z5.b[1]\n"
       "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "udot z10.s, z7.b, z0.b[1]\n"
       "addvl x10, x10, #4\n"
+      "udot z10.s, z7.b, z0.b[1]\n"
       "udot z14.s, z7.b, z1.b[1]\n"
       "udot z18.s, z7.b, z2.b[1]\n"
       "udot z22.s, z7.b, z3.b[1]\n"
@@ -1639,23 +1638,23 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "ble 65f\n"
       "ld1b { z7.b }, p5/Z, [x10]\n"
       "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x4\n"
       "udot z8.s, z7.b, z0.b[2]\n"
       "udot z12.s, z7.b, z1.b[2]\n"
       "udot z16.s, z7.b, z2.b[2]\n"
       "udot z20.s, z7.b, z3.b[2]\n"
+      "subs x27, x27, #0x4\n"
       "udot z24.s, z7.b, z4.b[2]\n"
       "udot z28.s, z7.b, z5.b[2]\n"
-      "udot z9.s, z6.b, z0.b[2]\n"
       "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "udot z9.s, z6.b, z0.b[2]\n"
       "udot z13.s, z6.b, z1.b[2]\n"
       "udot z17.s, z6.b, z2.b[2]\n"
       "udot z21.s, z6.b, z3.b[2]\n"
       "udot z25.s, z6.b, z4.b[2]\n"
       "udot z29.s, z6.b, z5.b[2]\n"
       "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "udot z10.s, z7.b, z0.b[2]\n"
       "addvl x10, x10, #4\n"
+      "udot z10.s, z7.b, z0.b[2]\n"
       "udot z14.s, z7.b, z1.b[2]\n"
       "udot z18.s, z7.b, z2.b[2]\n"
       "udot z22.s, z7.b, z3.b[2]\n"
@@ -1703,17 +1702,17 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "cmp x28, x20\n"
       "bne 60b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "st1w { z8.s }, p4, [x9]\n"
-      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "st1w { z8.s }, p4, [x9]\n"
       "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
+      "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
+      "add x20, x21, x20, LSL #2\n"
+      "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
-      "add x21, x22, x20, LSL #2\n"
       "st1w { z12.s }, p4, [x24]\n"
-      "add x20, x21, x20, LSL #2\n"
       "st1w { z13.s }, p3, [x24, #1, MUL VL]\n"
       "st1w { z14.s }, p2, [x24, #2, MUL VL]\n"
       "st1w { z15.s }, p1, [x24, #3, MUL VL]\n"
@@ -1749,8 +1748,8 @@ void sve_hybrid_u8u32_dot_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "68:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL.hpp
index 8437b0ea48..2c4eaaab4a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL.hpp
@@ -70,7 +70,7 @@ public:
         return true;
     }
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 6, 8, 8> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 8, 8> transforms = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp
index 898ff3a235..32cb778de4 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp
@@ -44,18 +44,18 @@ void sve_hybrid_u8u32_mmla_6x4VL (
         size_t output_offset = {};
         size_t input_initial_col = {};
         size_t input_offset = {};
-        void *output_ptr = {};
     } ka;
 
     unsigned long flags=0;
+    void *output_ptr;
     void *input_ptr;
 
     if (output_arg.is_indirect) {
-        ka.output_ptr=(void *)(output_arg.indirect.ptr);
+        output_ptr=(void *)(output_arg.indirect.ptr);
         ka.output_offset=output_arg.indirect.offset;
         flags |= 0x4;
     } else {
-        ka.output_ptr=(void *)(output_arg.direct.base);
+        output_ptr=(void *)(output_arg.direct.base);
         ka.output_offset=output_arg.direct.stride;
     }
 
@@ -89,7 +89,7 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "beq 12f\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "2:"  // Height 1: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -100,14 +100,14 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "incw x20\n"
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 3f\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z19.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "zip1 z9.d, z19.d, z13.d\n"
-      "zip2 z13.d, z19.d, z13.d\n"
+      "zip1 z9.d, z18.d, z13.d\n"
+      "zip2 z13.d, z18.d, z13.d\n"
       "zip1 z10.d, z17.d, z14.d\n"
       "zip2 z14.d, z17.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
@@ -126,8 +126,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "5:"  // Height 1: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 6f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -143,87 +143,87 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "ble 9f\n"
       "8:"  // Height 1: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z16.b }, p5/Z, [x10]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z19.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "trn1 z18.d, z19.d, z20.d\n"
-      "trn2 z19.d, z19.d, z20.d\n"
-      ".inst 0x45d09a48  // ummla z8.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45d19a4c  // ummla z12.s, z18.b, z17.b\n"
-      "ld1b { z20.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45d09a49  // ummla z9.s, z18.b, z16.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45d49a4d  // ummla z13.s, z18.b, z20.b\n"
+      "ld1rqb { z20.b }, p0/Z, [x26]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x45d19a48  // ummla z8.s, z18.b, z17.b\n"
+      ".inst 0x45d09a4c  // ummla z12.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
+      ".inst 0x45d19a49  // ummla z9.s, z18.b, z17.b\n"
+      ".inst 0x45d09a4d  // ummla z13.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45c19a4a  // ummla z10.s, z18.b, z1.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
+      ".inst 0x45d19a4a  // ummla z10.s, z18.b, z17.b\n"
       ".inst 0x45d09a4e  // ummla z14.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
+      "trn2 z20.d, z20.d, z19.d\n"
       ".inst 0x45d19a4b  // ummla z11.s, z18.b, z17.b\n"
       ".inst 0x45d09a4f  // ummla z15.s, z18.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45d09a68  // ummla z8.s, z19.b, z16.b\n"
-      "ld1b { z16.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45d19a6c  // ummla z12.s, z19.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45d09a69  // ummla z9.s, z19.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
+      ".inst 0x45d19a88  // ummla z8.s, z20.b, z17.b\n"
+      ".inst 0x45d09a8c  // ummla z12.s, z20.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n"
+      ".inst 0x45d19a89  // ummla z9.s, z20.b, z17.b\n"
+      ".inst 0x45d09a8d  // ummla z13.s, z20.b, z16.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45d19a6d  // ummla z13.s, z19.b, z17.b\n"
-      "ld1b { z3.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45d09a6a  // ummla z10.s, z19.b, z16.b\n"
+      "ld1b { z7.b }, p5/Z, [x10, #-3, MUL VL]\n"
+      ".inst 0x45d09a8a  // ummla z10.s, z20.b, z16.b\n"
+      ".inst 0x45c79a8e  // ummla z14.s, z20.b, z7.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45c39a6e  // ummla z14.s, z19.b, z3.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45d19a6b  // ummla z11.s, z19.b, z17.b\n"
-      ".inst 0x45d09a6f  // ummla z15.s, z19.b, z16.b\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45d19a8b  // ummla z11.s, z20.b, z17.b\n"
+      ".inst 0x45d09a8f  // ummla z15.s, z20.b, z16.b\n"
+      "add x26, x26, #0x10\n"
       "bgt 8b\n"
       "9:"  // Height 1: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45d19a48  // ummla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45d09a4c  // ummla z12.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45d19a49  // ummla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45d09a4d  // ummla z13.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45d19a4a  // ummla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45d09a4e  // ummla z14.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "subs x27, x27, #0x8\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45d19a4b  // ummla z11.s, z18.b, z17.b\n"
       ".inst 0x45d09a4f  // ummla z15.s, z18.b, z16.b\n"
+      "addvl x10, x10, #8\n"
       "ble 10f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45d19828  // ummla z8.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45d0982c  // ummla z12.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45d19829  // ummla z9.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45d0982d  // ummla z13.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45d1982a  // ummla z10.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45d0982e  // ummla z14.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x45d1982b  // ummla z11.s, z1.b, z17.b\n"
       ".inst 0x45d0982f  // ummla z15.s, z1.b, z16.b\n"
+      "addvl x10, x10, #8\n"
       "10:"  // Height 1: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
@@ -231,9 +231,9 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "bne 5b\n"
       "uzp1 z8.d, z8.d, z12.d\n"
       "uzp1 z9.d, z9.d, z13.d\n"
+      "st1w { z8.s }, p4, [x9]\n"
       "uzp1 z10.d, z10.d, z14.d\n"
       "uzp1 z11.d, z11.d, z15.d\n"
-      "st1w { z8.s }, p4, [x9]\n"
       "st1w { z9.s }, p3, [x9, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x9, #3, MUL VL]\n"
@@ -246,7 +246,7 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "12:"  // Height 2
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "13:"  // Height 2: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -258,19 +258,19 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 14f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x20, x9, x20, LSL #2\n"
       "ld1w { z18.s }, p4/Z, [x9]\n"
-      "ld1w { z24.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z2.s }, p3/Z, [x9, #1, MUL VL]\n"
       "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n"
       "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
-      "add x20, x9, x20, LSL #2\n"
       "ld1w { z12.s }, p4/Z, [x20]\n"
+      "zip1 z8.d, z18.d, z12.d\n"
+      "zip2 z12.d, z18.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n"
+      "zip1 z9.d, z2.d, z13.d\n"
+      "zip2 z13.d, z2.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z8.d, z18.d, z12.d\n"
-      "zip2 z12.d, z18.d, z12.d\n"
-      "zip1 z9.d, z24.d, z13.d\n"
-      "zip2 z13.d, z24.d, z13.d\n"
       "zip1 z10.d, z17.d, z14.d\n"
       "zip2 z14.d, z17.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
@@ -289,8 +289,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "16:"  // Height 2: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 17f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -309,109 +309,109 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "ble 20f\n"
       "19:"  // Height 2: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
+      "ld1rqb { z20.b }, p0/Z, [x26]\n"
+      "ld1rqb { z19.b }, p0/Z, [x25]\n"
+      "trn1 z18.d, z20.d, z19.d\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z19.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "trn1 z18.d, z19.d, z25.d\n"
-      "trn2 z19.d, z19.d, z25.d\n"
       ".inst 0x45d19a48  // ummla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45d09a4c  // ummla z12.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45d19a49  // ummla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45d09a4d  // ummla z13.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45d19a4a  // ummla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45d09a4e  // ummla z14.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
+      "trn2 z20.d, z20.d, z19.d\n"
       ".inst 0x45d19a4b  // ummla z11.s, z18.b, z17.b\n"
       ".inst 0x45d09a4f  // ummla z15.s, z18.b, z16.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45d19a68  // ummla z8.s, z19.b, z17.b\n"
+      ".inst 0x45d19a88  // ummla z8.s, z20.b, z17.b\n"
+      ".inst 0x45d09a8c  // ummla z12.s, z20.b, z16.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45d09a6c  // ummla z12.s, z19.b, z16.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45d19a69  // ummla z9.s, z19.b, z17.b\n"
+      ".inst 0x45d19a89  // ummla z9.s, z20.b, z17.b\n"
+      ".inst 0x45d09a8d  // ummla z13.s, z20.b, z16.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45d09a6d  // ummla z13.s, z19.b, z16.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45d19a6a  // ummla z10.s, z19.b, z17.b\n"
+      ".inst 0x45d19a8a  // ummla z10.s, z20.b, z17.b\n"
+      ".inst 0x45d09a8e  // ummla z14.s, z20.b, z16.b\n"
       "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45d09a6e  // ummla z14.s, z19.b, z16.b\n"
       "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45d19a6b  // ummla z11.s, z19.b, z17.b\n"
-      ".inst 0x45d09a6f  // ummla z15.s, z19.b, z16.b\n"
+      "sub x27, x27, #0x10\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45d19a8b  // ummla z11.s, z20.b, z17.b\n"
+      ".inst 0x45d09a8f  // ummla z15.s, z20.b, z16.b\n"
+      "add x26, x26, #0x10\n"
+      "add x25, x25, #0x10\n"
       "bgt 19b\n"
       "20:"  // Height 2: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z17.b }, p5/Z, [x10]\n"
-      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "ld1rqb { z19.b }, p0/Z, [x25]\n"
       "trn1 z18.d, z1.d, z19.d\n"
+      "ld1b { z17.b }, p5/Z, [x10]\n"
+      "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45d19a48  // ummla z8.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45d09a4c  // ummla z12.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
-      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45d19a49  // ummla z9.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45d09a4d  // ummla z13.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45d19a4a  // ummla z10.s, z18.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45d09a4e  // ummla z14.s, z18.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
+      "subs x27, x27, #0x8\n"
+      "trn2 z1.d, z1.d, z19.d\n"
       ".inst 0x45d19a4b  // ummla z11.s, z18.b, z17.b\n"
       ".inst 0x45d09a4f  // ummla z15.s, z18.b, z16.b\n"
+      "addvl x10, x10, #8\n"
       "ble 21f\n"
       "ld1b { z17.b }, p5/Z, [x10]\n"
       "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45d19828  // ummla z8.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45d0982c  // ummla z12.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45d19829  // ummla z9.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45d0982d  // ummla z13.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45d1982a  // ummla z10.s, z1.b, z17.b\n"
-      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45d0982e  // ummla z14.s, z1.b, z16.b\n"
+      "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n"
       "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x45d1982b  // ummla z11.s, z1.b, z17.b\n"
       ".inst 0x45d0982f  // ummla z15.s, z1.b, z16.b\n"
+      "addvl x10, x10, #8\n"
       "21:"  // Height 2: Multiply loop: multiply skip
       "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
       "add x28, x28, #0x1\n"
       "cmp x28, x20\n"
       "bne 16b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z17.d, z8.d, z12.d\n"
+      "add x20, x9, x20, LSL #2\n"
+      "uzp1 z16.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z16.d, z9.d, z13.d\n"
+      "uzp1 z17.d, z9.d, z13.d\n"
       "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z12.d, z10.d, z14.d\n"
+      "st1w { z16.s }, p4, [x9]\n"
+      "uzp1 z16.d, z10.d, z14.d\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x20, x9, x20, LSL #2\n"
-      "uzp1 z26.d, z11.d, z15.d\n"
+      "st1w { z17.s }, p3, [x9, #1, MUL VL]\n"
+      "uzp1 z2.d, z11.d, z15.d\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z17.s }, p4, [x9]\n"
-      "st1w { z16.s }, p3, [x9, #1, MUL VL]\n"
-      "st1w { z12.s }, p2, [x9, #2, MUL VL]\n"
-      "st1w { z26.s }, p1, [x9, #3, MUL VL]\n"
+      "st1w { z16.s }, p2, [x9, #2, MUL VL]\n"
+      "st1w { z2.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
       "st1w { z8.s }, p4, [x20]\n"
       "st1w { z9.s }, p3, [x20, #1, MUL VL]\n"
@@ -425,7 +425,7 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "23:"  // Height 3
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "24:"  // Height 3: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -437,28 +437,28 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 25f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z24.s }, p4/Z, [x9]\n"
-      "ld1w { z26.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x21, x9, x20, LSL #2\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x21]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x20]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "zip1 z8.d, z24.d, z12.d\n"
-      "zip2 z12.d, z24.d, z12.d\n"
-      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z9.d, z26.d, z13.d\n"
-      "zip2 z13.d, z26.d, z13.d\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
       "zip1 z17.d, z18.d, z21.d\n"
@@ -489,8 +489,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "27:"  // Height 3: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 28f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -512,92 +512,92 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "ble 31f\n"
       "30:"  // Height 3: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z27.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
+      "ld1rqb { z30.b }, p0/Z, [x26]\n"
       "ld1rqb { z24.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z26.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "trn1 z6.d, z27.d, z24.d\n"
-      "trn2 z27.d, z27.d, z24.d\n"
-      "trn1 z30.d, z26.d, z29.d\n"
-      "trn2 z26.d, z26.d, z29.d\n"
-      ".inst 0x45d998c8  // ummla z8.s, z6.b, z25.b\n"
-      ".inst 0x45dc98cc  // ummla z12.s, z6.b, z28.b\n"
-      ".inst 0x45d99bd0  // ummla z16.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45dc9bd4  // ummla z20.s, z30.b, z28.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45d998c9  // ummla z9.s, z6.b, z25.b\n"
-      ".inst 0x45d99bd1  // ummla z17.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45d898cd  // ummla z13.s, z6.b, z24.b\n"
-      ".inst 0x45d89bd5  // ummla z21.s, z30.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45d998ca  // ummla z10.s, z6.b, z25.b\n"
-      ".inst 0x45d99bd2  // ummla z18.s, z30.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45d898ce  // ummla z14.s, z6.b, z24.b\n"
-      ".inst 0x45d89bd6  // ummla z22.s, z30.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #16\n"
-      ".inst 0x45d998cb  // ummla z11.s, z6.b, z25.b\n"
-      ".inst 0x45d99bd3  // ummla z19.s, z30.b, z25.b\n"
-      ".inst 0x45d898cf  // ummla z15.s, z6.b, z24.b\n"
-      ".inst 0x45d89bd7  // ummla z23.s, z30.b, z24.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
+      "ld1rqb { z28.b }, p0/Z, [x24]\n"
+      "trn1 z27.d, z30.d, z24.d\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "trn1 z26.d, z28.d, z29.d\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45d99b68  // ummla z8.s, z27.b, z25.b\n"
       ".inst 0x45d99b50  // ummla z16.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x45d89b6c  // ummla z12.s, z27.b, z24.b\n"
       ".inst 0x45d89b54  // ummla z20.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n"
+      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45d99b69  // ummla z9.s, z27.b, z25.b\n"
       ".inst 0x45d99b51  // ummla z17.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n"
+      "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "trn2 z28.d, z28.d, z29.d\n"
       ".inst 0x45d89b6d  // ummla z13.s, z27.b, z24.b\n"
       ".inst 0x45d89b55  // ummla z21.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
+      "sub x27, x27, #0x10\n"
       ".inst 0x45d99b6a  // ummla z10.s, z27.b, z25.b\n"
       ".inst 0x45d99b52  // ummla z18.s, z26.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n"
+      "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
+      "cmp x27, #0x10\n"
       ".inst 0x45d89b6e  // ummla z14.s, z27.b, z24.b\n"
       ".inst 0x45d89b56  // ummla z22.s, z26.b, z24.b\n"
-      "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
+      "addvl x10, x10, #16\n"
       ".inst 0x45d99b6b  // ummla z11.s, z27.b, z25.b\n"
       ".inst 0x45d99b53  // ummla z19.s, z26.b, z25.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x45d89b6f  // ummla z15.s, z27.b, z24.b\n"
       ".inst 0x45d89b57  // ummla z23.s, z26.b, z24.b\n"
+      "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x45d99bc8  // ummla z8.s, z30.b, z25.b\n"
+      ".inst 0x45d99b90  // ummla z16.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45d89bcc  // ummla z12.s, z30.b, z24.b\n"
+      ".inst 0x45d89b94  // ummla z20.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n"
+      ".inst 0x45d99bc9  // ummla z9.s, z30.b, z25.b\n"
+      ".inst 0x45d99b91  // ummla z17.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n"
+      ".inst 0x45d89bcd  // ummla z13.s, z30.b, z24.b\n"
+      ".inst 0x45d89b95  // ummla z21.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n"
+      ".inst 0x45d99bca  // ummla z10.s, z30.b, z25.b\n"
+      ".inst 0x45d99b92  // ummla z18.s, z28.b, z25.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n"
+      ".inst 0x45d89bce  // ummla z14.s, z30.b, z24.b\n"
+      ".inst 0x45d89b96  // ummla z22.s, z28.b, z24.b\n"
+      "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n"
+      ".inst 0x45d99bcb  // ummla z11.s, z30.b, z25.b\n"
+      ".inst 0x45d99b93  // ummla z19.s, z28.b, z25.b\n"
+      ".inst 0x45d89bcf  // ummla z15.s, z30.b, z24.b\n"
+      ".inst 0x45d89b97  // ummla z23.s, z28.b, z24.b\n"
       "bgt 30b\n"
       "31:"  // Height 3: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z25.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
       "ld1rqb { z24.b }, p0/Z, [x25]\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
       "trn1 z27.d, z1.d, z24.d\n"
       "trn2 z1.d, z1.d, z24.d\n"
-      "trn1 z26.d, z3.d, z29.d\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "trn1 z26.d, z3.d, z28.d\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45d99b68  // ummla z8.s, z27.b, z25.b\n"
-      ".inst 0x45dc9b6c  // ummla z12.s, z27.b, z28.b\n"
-      "trn2 z3.d, z3.d, z29.d\n"
       ".inst 0x45d99b50  // ummla z16.s, z26.b, z25.b\n"
+      ".inst 0x45d89b6c  // ummla z12.s, z27.b, z24.b\n"
+      ".inst 0x45d89b54  // ummla z20.s, z26.b, z24.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45dc9b54  // ummla z20.s, z26.b, z28.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45d99b69  // ummla z9.s, z27.b, z25.b\n"
       ".inst 0x45d99b51  // ummla z17.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x45d89b6d  // ummla z13.s, z27.b, z24.b\n"
       ".inst 0x45d89b55  // ummla z21.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
+      "trn2 z3.d, z3.d, z28.d\n"
       ".inst 0x45d99b6a  // ummla z10.s, z27.b, z25.b\n"
       ".inst 0x45d99b52  // ummla z18.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
@@ -614,9 +614,9 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45d99828  // ummla z8.s, z1.b, z25.b\n"
       ".inst 0x45d99870  // ummla z16.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45d8982c  // ummla z12.s, z1.b, z24.b\n"
       ".inst 0x45d89874  // ummla z20.s, z3.b, z24.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45d99829  // ummla z9.s, z1.b, z25.b\n"
       ".inst 0x45d99871  // ummla z17.s, z3.b, z25.b\n"
@@ -641,26 +641,26 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 27b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z27.d, z8.d, z12.d\n"
+      "add x21, x9, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "uzp1 z25.d, z8.d, z12.d\n"
       "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z26.d, z9.d, z13.d\n"
+      "uzp1 z24.d, z9.d, z13.d\n"
+      "st1w { z25.s }, p4, [x9]\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z25.d, z10.d, z14.d\n"
+      "st1w { z24.s }, p3, [x9, #1, MUL VL]\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x21, x9, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
       "uzp1 z24.d, z11.d, z15.d\n"
+      "st1w { z25.s }, p2, [x9, #2, MUL VL]\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z27.s }, p4, [x9]\n"
       "uzp1 z16.d, z16.d, z20.d\n"
-      "uzp1 z17.d, z17.d, z21.d\n"
-      "st1w { z26.s }, p3, [x9, #1, MUL VL]\n"
-      "uzp1 z18.d, z18.d, z22.d\n"
-      "uzp1 z19.d, z19.d, z23.d\n"
-      "st1w { z25.s }, p2, [x9, #2, MUL VL]\n"
       "st1w { z24.s }, p1, [x9, #3, MUL VL]\n"
       "addvl x9, x9, #4\n"
+      "uzp1 z17.d, z17.d, z21.d\n"
+      "uzp1 z18.d, z18.d, z22.d\n"
       "st1w { z8.s }, p4, [x21]\n"
+      "uzp1 z19.d, z19.d, z23.d\n"
       "st1w { z9.s }, p3, [x21, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x21, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x21, #3, MUL VL]\n"
@@ -676,7 +676,7 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "34:"  // Height 4
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "35:"  // Height 4: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -688,37 +688,37 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 36f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x22, x9, x20, LSL #2\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x22]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x21]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x20]\n"
-      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
+      "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x20]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
@@ -745,8 +745,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "38:"  // Height 4: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 39f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -771,114 +771,114 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "ble 42f\n"
       "41:"  // Height 4: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z31.b }, p5/Z, [x10]\n"
-      "ld1b { z30.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z29.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqb { z30.b }, p0/Z, [x26]\n"
+      "ld1rqb { z24.b }, p0/Z, [x25]\n"
+      "trn1 z29.d, z30.d, z24.d\n"
       "ld1rqb { z28.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "add x23, x23, #0x10\n"
-      "trn1 z27.d, z29.d, z25.d\n"
-      "trn2 z29.d, z29.d, z25.d\n"
-      "trn1 z26.d, z28.d, z24.d\n"
-      "trn2 z28.d, z28.d, z24.d\n"
-      ".inst 0x45df9b68  // ummla z8.s, z27.b, z31.b\n"
-      ".inst 0x45de9b6c  // ummla z12.s, z27.b, z30.b\n"
-      ".inst 0x45df9b50  // ummla z16.s, z26.b, z31.b\n"
+      "ld1rqb { z27.b }, p0/Z, [x23]\n"
+      "trn2 z30.d, z30.d, z24.d\n"
+      "trn1 z26.d, z28.d, z27.d\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x45d99ba8  // ummla z8.s, z29.b, z25.b\n"
+      ".inst 0x45d99b50  // ummla z16.s, z26.b, z25.b\n"
+      ".inst 0x45d89bac  // ummla z12.s, z29.b, z24.b\n"
+      ".inst 0x45d89b54  // ummla z20.s, z26.b, z24.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45de9b54  // ummla z20.s, z26.b, z30.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45d99b69  // ummla z9.s, z27.b, z25.b\n"
+      ".inst 0x45d99ba9  // ummla z9.s, z29.b, z25.b\n"
       ".inst 0x45d99b51  // ummla z17.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45d89b6d  // ummla z13.s, z27.b, z24.b\n"
+      "trn2 z28.d, z28.d, z27.d\n"
+      ".inst 0x45d89bad  // ummla z13.s, z29.b, z24.b\n"
       ".inst 0x45d89b55  // ummla z21.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45d99b6a  // ummla z10.s, z27.b, z25.b\n"
+      "sub x27, x27, #0x10\n"
+      ".inst 0x45d99baa  // ummla z10.s, z29.b, z25.b\n"
       ".inst 0x45d99b52  // ummla z18.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45d89b6e  // ummla z14.s, z27.b, z24.b\n"
+      "cmp x27, #0x10\n"
+      ".inst 0x45d89bae  // ummla z14.s, z29.b, z24.b\n"
       ".inst 0x45d89b56  // ummla z22.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x45d99b6b  // ummla z11.s, z27.b, z25.b\n"
+      ".inst 0x45d99bab  // ummla z11.s, z29.b, z25.b\n"
       ".inst 0x45d99b53  // ummla z19.s, z26.b, z25.b\n"
-      ".inst 0x45d89b6f  // ummla z15.s, z27.b, z24.b\n"
-      ".inst 0x45d89b57  // ummla z23.s, z26.b, z24.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      "add x26, x26, #0x10\n"
+      ".inst 0x45d89baf  // ummla z15.s, z29.b, z24.b\n"
+      ".inst 0x45d89b57  // ummla z23.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45d99ba8  // ummla z8.s, z29.b, z25.b\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x45d99bc8  // ummla z8.s, z30.b, z25.b\n"
       ".inst 0x45d99b90  // ummla z16.s, z28.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45d89bac  // ummla z12.s, z29.b, z24.b\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45d89bcc  // ummla z12.s, z30.b, z24.b\n"
       ".inst 0x45d89b94  // ummla z20.s, z28.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45d99ba9  // ummla z9.s, z29.b, z25.b\n"
+      "add x23, x23, #0x10\n"
+      ".inst 0x45d99bc9  // ummla z9.s, z30.b, z25.b\n"
       ".inst 0x45d99b91  // ummla z17.s, z28.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45d89bad  // ummla z13.s, z29.b, z24.b\n"
+      ".inst 0x45d89bcd  // ummla z13.s, z30.b, z24.b\n"
       ".inst 0x45d89b95  // ummla z21.s, z28.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45d99baa  // ummla z10.s, z29.b, z25.b\n"
+      ".inst 0x45d99bca  // ummla z10.s, z30.b, z25.b\n"
       ".inst 0x45d99b92  // ummla z18.s, z28.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45d89bae  // ummla z14.s, z29.b, z24.b\n"
+      ".inst 0x45d89bce  // ummla z14.s, z30.b, z24.b\n"
       ".inst 0x45d89b96  // ummla z22.s, z28.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45d99bab  // ummla z11.s, z29.b, z25.b\n"
+      ".inst 0x45d99bcb  // ummla z11.s, z30.b, z25.b\n"
       ".inst 0x45d99b93  // ummla z19.s, z28.b, z25.b\n"
-      ".inst 0x45d89baf  // ummla z15.s, z29.b, z24.b\n"
+      ".inst 0x45d89bcf  // ummla z15.s, z30.b, z24.b\n"
       ".inst 0x45d89b97  // ummla z23.s, z28.b, z24.b\n"
       "bgt 41b\n"
       "42:"  // Height 4: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z29.b }, p5/Z, [x10]\n"
-      "ld1b { z28.b }, p5/Z, [x10, #1, MUL VL]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z25.b }, p0/Z, [x25]\n"
+      "ld1rqb { z24.b }, p0/Z, [x25]\n"
+      "trn1 z28.d, z1.d, z24.d\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z24.b }, p0/Z, [x23]\n"
-      "trn1 z27.d, z1.d, z25.d\n"
-      "trn2 z1.d, z1.d, z25.d\n"
-      "trn1 z26.d, z3.d, z24.d\n"
-      ".inst 0x45dd9b68  // ummla z8.s, z27.b, z29.b\n"
-      ".inst 0x45dc9b6c  // ummla z12.s, z27.b, z28.b\n"
-      "trn2 z3.d, z3.d, z24.d\n"
-      ".inst 0x45dd9b50  // ummla z16.s, z26.b, z29.b\n"
+      "ld1rqb { z27.b }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z24.d\n"
+      "trn1 z26.d, z3.d, z27.d\n"
+      "ld1b { z25.b }, p5/Z, [x10]\n"
+      "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
+      ".inst 0x45d99b88  // ummla z8.s, z28.b, z25.b\n"
+      ".inst 0x45d99b50  // ummla z16.s, z26.b, z25.b\n"
+      ".inst 0x45d89b8c  // ummla z12.s, z28.b, z24.b\n"
+      ".inst 0x45d89b54  // ummla z20.s, z26.b, z24.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45dc9b54  // ummla z20.s, z26.b, z28.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45d99b69  // ummla z9.s, z27.b, z25.b\n"
+      ".inst 0x45d99b89  // ummla z9.s, z28.b, z25.b\n"
       ".inst 0x45d99b51  // ummla z17.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45d89b6d  // ummla z13.s, z27.b, z24.b\n"
+      "subs x27, x27, #0x8\n"
+      ".inst 0x45d89b8d  // ummla z13.s, z28.b, z24.b\n"
       ".inst 0x45d89b55  // ummla z21.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45d99b6a  // ummla z10.s, z27.b, z25.b\n"
+      "trn2 z3.d, z3.d, z27.d\n"
+      ".inst 0x45d99b8a  // ummla z10.s, z28.b, z25.b\n"
       ".inst 0x45d99b52  // ummla z18.s, z26.b, z25.b\n"
       "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45d89b6e  // ummla z14.s, z27.b, z24.b\n"
+      ".inst 0x45d89b8e  // ummla z14.s, z28.b, z24.b\n"
       ".inst 0x45d89b56  // ummla z22.s, z26.b, z24.b\n"
       "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #8\n"
-      ".inst 0x45d99b6b  // ummla z11.s, z27.b, z25.b\n"
+      ".inst 0x45d99b8b  // ummla z11.s, z28.b, z25.b\n"
       ".inst 0x45d99b53  // ummla z19.s, z26.b, z25.b\n"
-      ".inst 0x45d89b6f  // ummla z15.s, z27.b, z24.b\n"
+      ".inst 0x45d89b8f  // ummla z15.s, z28.b, z24.b\n"
       ".inst 0x45d89b57  // ummla z23.s, z26.b, z24.b\n"
       "ble 43f\n"
       "ld1b { z25.b }, p5/Z, [x10]\n"
       "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45d99828  // ummla z8.s, z1.b, z25.b\n"
       ".inst 0x45d99870  // ummla z16.s, z3.b, z25.b\n"
-      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45d8982c  // ummla z12.s, z1.b, z24.b\n"
       ".inst 0x45d89874  // ummla z20.s, z3.b, z24.b\n"
+      "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n"
       "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45d99829  // ummla z9.s, z1.b, z25.b\n"
       ".inst 0x45d99871  // ummla z17.s, z3.b, z25.b\n"
@@ -903,33 +903,33 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 38b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x22, x9, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp1 z25.d, z8.d, z12.d\n"
+      "add x20, x21, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z24.d, z9.d, z13.d\n"
+      "st1w { z25.s }, p4, [x9]\n"
       "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z27.d, z10.d, z14.d\n"
+      "uzp1 z25.d, z10.d, z14.d\n"
+      "st1w { z24.s }, p3, [x9, #1, MUL VL]\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x22, x9, x20, LSL #2\n"
-      "add x21, x22, x20, LSL #2\n"
-      "add x20, x21, x20, LSL #2\n"
-      "uzp1 z26.d, z11.d, z15.d\n"
+      "uzp1 z24.d, z11.d, z15.d\n"
+      "st1w { z25.s }, p2, [x9, #2, MUL VL]\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z25.s }, p4, [x9]\n"
       "uzp1 z25.d, z16.d, z20.d\n"
+      "st1w { z24.s }, p1, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z24.s }, p3, [x9, #1, MUL VL]\n"
       "uzp1 z24.d, z17.d, z21.d\n"
+      "st1w { z8.s }, p4, [x22]\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z27.s }, p2, [x9, #2, MUL VL]\n"
       "uzp1 z21.d, z18.d, z22.d\n"
+      "st1w { z9.s }, p3, [x22, #1, MUL VL]\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z26.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
       "uzp1 z20.d, z19.d, z23.d\n"
-      "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x22]\n"
-      "st1w { z9.s }, p3, [x22, #1, MUL VL]\n"
       "st1w { z10.s }, p2, [x22, #2, MUL VL]\n"
+      "uzp2 z19.d, z19.d, z23.d\n"
       "st1w { z11.s }, p1, [x22, #3, MUL VL]\n"
       "st1w { z25.s }, p4, [x21]\n"
       "st1w { z24.s }, p3, [x21, #1, MUL VL]\n"
@@ -947,7 +947,7 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "45:"  // Height 5
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "mov x9, %x[output_ptr]\n"
       "46:"  // Height 5: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -959,46 +959,46 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 47f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "ld1w { z19.s }, p4/Z, [x9]\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "ld1w { z12.s }, p4/Z, [x23]\n"
+      "zip1 z8.d, z19.d, z12.d\n"
+      "zip2 z12.d, z19.d, z12.d\n"
       "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip1 z9.d, z17.d, z13.d\n"
+      "zip2 z13.d, z17.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x22]\n"
+      "zip1 z10.d, z18.d, z14.d\n"
+      "zip2 z14.d, z18.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
-      "ld1w { z20.s }, p4/Z, [x21]\n"
-      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x20]\n"
-      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
+      "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x21]\n"
       "zip1 z16.d, z17.d, z20.d\n"
       "zip2 z20.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
       "zip2 z21.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x20]\n"
       "zip1 z18.d, z19.d, z22.d\n"
       "zip2 z22.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip1 z19.d, z24.d, z23.d\n"
       "zip2 z23.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z24.d, z25.d, z28.d\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
@@ -1037,8 +1037,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "49:"  // Height 5: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 50f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1066,103 +1066,102 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "ble 53f\n"
       "52:"  // Height 5: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p5/Z, [x10]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
       "ld1rqb { z6.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
+      "ld1rqb { z1.b }, p0/Z, [x25]\n"
       "ld1rqb { z7.b }, p0/Z, [x24]\n"
       "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "add x24, x24, #0x10\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
+      "trn1 z5.d, z6.d, z1.d\n"
+      "trn2 z6.d, z6.d, z1.d\n"
+      "ld1rqb { z4.b }, p0/Z, [x22]\n"
       "trn1 z3.d, z7.d, z2.d\n"
       "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
+      "ld1b { z1.b }, p5/Z, [x10]\n"
+      "trn1 z2.d, z4.d, z0.d\n"
+      "trn2 z4.d, z4.d, z0.d\n"
       "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45c19888  // ummla z8.s, z4.b, z1.b\n"
+      ".inst 0x45c198a8  // ummla z8.s, z5.b, z1.b\n"
       ".inst 0x45c19870  // ummla z16.s, z3.b, z1.b\n"
       ".inst 0x45c19858  // ummla z24.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45c0988c  // ummla z12.s, z4.b, z0.b\n"
+      "sub x27, x27, #0x10\n"
+      ".inst 0x45c098ac  // ummla z12.s, z5.b, z0.b\n"
       ".inst 0x45c09874  // ummla z20.s, z3.b, z0.b\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x45c0985c  // ummla z28.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45c19889  // ummla z9.s, z4.b, z1.b\n"
+      ".inst 0x45c198a9  // ummla z9.s, z5.b, z1.b\n"
+      "add x25, x25, #0x10\n"
       ".inst 0x45c19871  // ummla z17.s, z3.b, z1.b\n"
       ".inst 0x45c19859  // ummla z25.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45c0988d  // ummla z13.s, z4.b, z0.b\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45c098ad  // ummla z13.s, z5.b, z0.b\n"
       ".inst 0x45c09875  // ummla z21.s, z3.b, z0.b\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x45c0985d  // ummla z29.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45c1988a  // ummla z10.s, z4.b, z1.b\n"
+      ".inst 0x45c198aa  // ummla z10.s, z5.b, z1.b\n"
       ".inst 0x45c19872  // ummla z18.s, z3.b, z1.b\n"
       ".inst 0x45c1985a  // ummla z26.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45c0988e  // ummla z14.s, z4.b, z0.b\n"
+      ".inst 0x45c098ae  // ummla z14.s, z5.b, z0.b\n"
       ".inst 0x45c09876  // ummla z22.s, z3.b, z0.b\n"
       ".inst 0x45c0985e  // ummla z30.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x45c1988b  // ummla z11.s, z4.b, z1.b\n"
+      ".inst 0x45c198ab  // ummla z11.s, z5.b, z1.b\n"
       ".inst 0x45c19873  // ummla z19.s, z3.b, z1.b\n"
       ".inst 0x45c1985b  // ummla z27.s, z2.b, z1.b\n"
-      ".inst 0x45c0988f  // ummla z15.s, z4.b, z0.b\n"
+      "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      ".inst 0x45c098af  // ummla z15.s, z5.b, z0.b\n"
       ".inst 0x45c09877  // ummla z23.s, z3.b, z0.b\n"
       ".inst 0x45c0985f  // ummla z31.s, z2.b, z0.b\n"
-      "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n"
       "ld1b { z0.b }, p5/Z, [x10, #-7, MUL VL]\n"
       ".inst 0x45c198c8  // ummla z8.s, z6.b, z1.b\n"
       ".inst 0x45c198f0  // ummla z16.s, z7.b, z1.b\n"
-      ".inst 0x45c198b8  // ummla z24.s, z5.b, z1.b\n"
+      ".inst 0x45c19898  // ummla z24.s, z4.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-6, MUL VL]\n"
       ".inst 0x45c098cc  // ummla z12.s, z6.b, z0.b\n"
       ".inst 0x45c098f4  // ummla z20.s, z7.b, z0.b\n"
-      ".inst 0x45c098bc  // ummla z28.s, z5.b, z0.b\n"
+      ".inst 0x45c0989c  // ummla z28.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-5, MUL VL]\n"
       ".inst 0x45c198c9  // ummla z9.s, z6.b, z1.b\n"
       ".inst 0x45c198f1  // ummla z17.s, z7.b, z1.b\n"
-      ".inst 0x45c198b9  // ummla z25.s, z5.b, z1.b\n"
+      ".inst 0x45c19899  // ummla z25.s, z4.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-4, MUL VL]\n"
       ".inst 0x45c098cd  // ummla z13.s, z6.b, z0.b\n"
       ".inst 0x45c098f5  // ummla z21.s, z7.b, z0.b\n"
-      ".inst 0x45c098bd  // ummla z29.s, z5.b, z0.b\n"
+      ".inst 0x45c0989d  // ummla z29.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-3, MUL VL]\n"
       ".inst 0x45c198ca  // ummla z10.s, z6.b, z1.b\n"
       ".inst 0x45c198f2  // ummla z18.s, z7.b, z1.b\n"
-      ".inst 0x45c198ba  // ummla z26.s, z5.b, z1.b\n"
+      ".inst 0x45c1989a  // ummla z26.s, z4.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-2, MUL VL]\n"
       ".inst 0x45c098ce  // ummla z14.s, z6.b, z0.b\n"
       ".inst 0x45c098f6  // ummla z22.s, z7.b, z0.b\n"
-      ".inst 0x45c098be  // ummla z30.s, z5.b, z0.b\n"
+      ".inst 0x45c0989e  // ummla z30.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-1, MUL VL]\n"
       ".inst 0x45c198cb  // ummla z11.s, z6.b, z1.b\n"
       ".inst 0x45c198f3  // ummla z19.s, z7.b, z1.b\n"
-      ".inst 0x45c198bb  // ummla z27.s, z5.b, z1.b\n"
+      ".inst 0x45c1989b  // ummla z27.s, z4.b, z1.b\n"
       ".inst 0x45c098cf  // ummla z15.s, z6.b, z0.b\n"
       ".inst 0x45c098f7  // ummla z23.s, z7.b, z0.b\n"
-      ".inst 0x45c098bf  // ummla z31.s, z5.b, z0.b\n"
+      ".inst 0x45c0989f  // ummla z31.s, z4.b, z0.b\n"
       "bgt 52b\n"
       "53:"  // Height 5: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z2.b }, p5/Z, [x10]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z6.b }, p0/Z, [x25]\n"
+      "ld1rqb { z4.b }, p0/Z, [x25]\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "ld1rqb { z2.b }, p0/Z, [x23]\n"
+      "trn1 z7.d, z1.d, z4.d\n"
+      "trn2 z1.d, z1.d, z4.d\n"
       "ld1rqb { z5.b }, p0/Z, [x22]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
+      "ld1b { z2.b }, p5/Z, [x10]\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
       "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
@@ -1170,6 +1169,7 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       ".inst 0x45c298d0  // ummla z16.s, z6.b, z2.b\n"
       ".inst 0x45c29898  // ummla z24.s, z4.b, z2.b\n"
       "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x45c098ec  // ummla z12.s, z7.b, z0.b\n"
       ".inst 0x45c098d4  // ummla z20.s, z6.b, z0.b\n"
       ".inst 0x45c0989c  // ummla z28.s, z4.b, z0.b\n"
@@ -1190,8 +1190,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       ".inst 0x45c098d6  // ummla z22.s, z6.b, z0.b\n"
       ".inst 0x45c0989e  // ummla z30.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x45c298eb  // ummla z11.s, z7.b, z2.b\n"
+      "addvl x10, x10, #8\n"
       ".inst 0x45c298d3  // ummla z19.s, z6.b, z2.b\n"
       ".inst 0x45c2989b  // ummla z27.s, z4.b, z2.b\n"
       ".inst 0x45c098ef  // ummla z15.s, z7.b, z0.b\n"
@@ -1203,24 +1203,24 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       ".inst 0x45c29828  // ummla z8.s, z1.b, z2.b\n"
       ".inst 0x45c29870  // ummla z16.s, z3.b, z2.b\n"
       ".inst 0x45c298b8  // ummla z24.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45c0982c  // ummla z12.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45c09874  // ummla z20.s, z3.b, z0.b\n"
       ".inst 0x45c098bc  // ummla z28.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45c29829  // ummla z9.s, z1.b, z2.b\n"
       ".inst 0x45c29871  // ummla z17.s, z3.b, z2.b\n"
       ".inst 0x45c298b9  // ummla z25.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45c0982d  // ummla z13.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45c09875  // ummla z21.s, z3.b, z0.b\n"
       ".inst 0x45c098bd  // ummla z29.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45c2982a  // ummla z10.s, z1.b, z2.b\n"
       ".inst 0x45c29872  // ummla z18.s, z3.b, z2.b\n"
       ".inst 0x45c298ba  // ummla z26.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45c0982e  // ummla z14.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45c09876  // ummla z22.s, z3.b, z0.b\n"
       ".inst 0x45c098be  // ummla z30.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
@@ -1237,39 +1237,39 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 49b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "uzp1 z1.d, z8.d, z12.d\n"
-      "uzp2 z8.d, z8.d, z12.d\n"
-      "uzp1 z0.d, z9.d, z13.d\n"
-      "uzp2 z9.d, z9.d, z13.d\n"
-      "uzp1 z3.d, z10.d, z14.d\n"
-      "uzp2 z10.d, z10.d, z14.d\n"
       "add x23, x9, x20, LSL #2\n"
       "add x22, x23, x20, LSL #2\n"
+      "uzp1 z2.d, z8.d, z12.d\n"
       "add x21, x22, x20, LSL #2\n"
+      "add x20, x21, x20, LSL #2\n"
+      "uzp2 z8.d, z8.d, z12.d\n"
+      "uzp1 z1.d, z9.d, z13.d\n"
+      "uzp2 z9.d, z9.d, z13.d\n"
+      "uzp1 z0.d, z10.d, z14.d\n"
+      "st1w { z2.s }, p4, [x9]\n"
+      "uzp2 z10.d, z10.d, z14.d\n"
       "uzp1 z2.d, z11.d, z15.d\n"
+      "st1w { z1.s }, p3, [x9, #1, MUL VL]\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z1.s }, p4, [x9]\n"
-      "add x20, x21, x20, LSL #2\n"
       "uzp1 z1.d, z16.d, z20.d\n"
+      "st1w { z0.s }, p2, [x9, #2, MUL VL]\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z0.s }, p3, [x9, #1, MUL VL]\n"
       "uzp1 z0.d, z17.d, z21.d\n"
+      "st1w { z2.s }, p1, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z3.s }, p2, [x9, #2, MUL VL]\n"
       "uzp1 z21.d, z18.d, z22.d\n"
+      "st1w { z8.s }, p4, [x23]\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z2.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
       "uzp1 z20.d, z19.d, z23.d\n"
+      "st1w { z9.s }, p3, [x23, #1, MUL VL]\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x23]\n"
       "uzp1 z24.d, z24.d, z28.d\n"
+      "st1w { z10.s }, p2, [x23, #2, MUL VL]\n"
       "uzp1 z25.d, z25.d, z29.d\n"
-      "st1w { z9.s }, p3, [x23, #1, MUL VL]\n"
       "uzp1 z26.d, z26.d, z30.d\n"
-      "uzp1 z27.d, z27.d, z31.d\n"
-      "st1w { z10.s }, p2, [x23, #2, MUL VL]\n"
       "st1w { z11.s }, p1, [x23, #3, MUL VL]\n"
+      "uzp1 z27.d, z27.d, z31.d\n"
       "st1w { z1.s }, p4, [x22]\n"
       "st1w { z0.s }, p3, [x22, #1, MUL VL]\n"
       "st1w { z21.s }, p2, [x22, #2, MUL VL]\n"
@@ -1289,12 +1289,11 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "b 68f\n"
       "56:"  // Height 6
       "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n"
       "mov x20, #0x18\n"
       "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
+      "mov x9, %x[output_ptr]\n"
       "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
-      "madd x20, x21, x20, x9\n"
-      "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n"
+      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
       "57:"  // Height 6: Column loop
       "mov x20, #0x0\n"
       "whilelt p4.s, x20, x11\n"
@@ -1306,54 +1305,54 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "whilelt p1.s, x20, x11\n"
       "tbz %x[flags], #0, 58f\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
-      "ld1w { z20.s }, p4/Z, [x9]\n"
-      "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
       "add x24, x9, x20, LSL #2\n"
       "add x23, x24, x20, LSL #2\n"
+      "ld1w { z17.s }, p4/Z, [x9]\n"
       "add x22, x23, x20, LSL #2\n"
-      "ld1w { z12.s }, p4/Z, [x24]\n"
-      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "add x21, x22, x20, LSL #2\n"
+      "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n"
+      "ld1w { z20.s }, p2/Z, [x9, #2, MUL VL]\n"
       "add x20, x21, x20, LSL #2\n"
+      "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n"
+      "ld1w { z12.s }, p4/Z, [x24]\n"
+      "zip1 z8.d, z17.d, z12.d\n"
+      "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n"
       "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n"
+      "zip2 z12.d, z17.d, z12.d\n"
+      "zip1 z9.d, z18.d, z13.d\n"
       "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n"
       "ld1w { z17.s }, p4/Z, [x23]\n"
+      "zip2 z13.d, z18.d, z13.d\n"
+      "zip1 z10.d, z20.d, z14.d\n"
       "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n"
       "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip2 z14.d, z20.d, z14.d\n"
+      "zip1 z11.d, z16.d, z15.d\n"
       "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n"
-      "zip1 z8.d, z20.d, z12.d\n"
-      "zip2 z12.d, z20.d, z12.d\n"
       "ld1w { z20.s }, p4/Z, [x22]\n"
-      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
-      "zip1 z9.d, z22.d, z13.d\n"
-      "zip2 z13.d, z22.d, z13.d\n"
-      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
-      "zip1 z10.d, z25.d, z14.d\n"
-      "zip2 z14.d, z25.d, z14.d\n"
-      "ld1w { z25.s }, p4/Z, [x21]\n"
-      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
-      "zip1 z11.d, z16.d, z15.d\n"
       "zip2 z15.d, z16.d, z15.d\n"
-      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
       "zip1 z16.d, z17.d, z20.d\n"
+      "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n"
+      "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n"
       "zip2 z20.d, z17.d, z20.d\n"
-      "ld1w { z28.s }, p4/Z, [x20]\n"
-      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
       "zip1 z17.d, z18.d, z21.d\n"
+      "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n"
+      "ld1w { z25.s }, p4/Z, [x21]\n"
       "zip2 z21.d, z18.d, z21.d\n"
-      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
-      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip1 z18.d, z19.d, z22.d\n"
+      "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n"
+      "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n"
       "zip2 z22.d, z19.d, z22.d\n"
       "zip1 z19.d, z24.d, z23.d\n"
+      "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n"
+      "ld1w { z28.s }, p4/Z, [x20]\n"
       "zip2 z23.d, z24.d, z23.d\n"
       "zip1 z24.d, z25.d, z28.d\n"
+      "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n"
+      "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n"
       "zip2 z28.d, z25.d, z28.d\n"
       "zip1 z25.d, z26.d, z29.d\n"
+      "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n"
       "zip2 z29.d, z26.d, z29.d\n"
       "zip1 z26.d, z27.d, z30.d\n"
       "zip2 z30.d, z27.d, z30.d\n"
@@ -1389,8 +1388,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "mov x28, #0x0\n"
       "60:"  // Height 6: String loop
       "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
-      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "ldr w27, [x20, x28, LSL #0x2]\n"
+      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
       "tbz %x[flags], #3, 61f\n"
       "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
       "add x20, x20, x21, LSL #3\n"
@@ -1421,113 +1420,113 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "ble 64f\n"
       "63:"  // Height 6: Multiply loop: Main loop head
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z1.b }, p5/Z, [x10]\n"
-      "sub x27, x27, #0x10\n"
-      "cmp x27, #0x10\n"
-      "ld1rqb { z6.b }, p0/Z, [x26]\n"
-      "add x26, x26, #0x10\n"
-      "ld1rqb { z3.b }, p0/Z, [x25]\n"
-      "add x25, x25, #0x10\n"
-      "ld1rqb { z7.b }, p0/Z, [x24]\n"
-      "add x24, x24, #0x10\n"
-      "ld1rqb { z2.b }, p0/Z, [x23]\n"
-      "ld1rqb { z5.b }, p0/Z, [x22]\n"
+      "ld1rqb { z7.b }, p0/Z, [x26]\n"
+      "ld1rqb { z0.b }, p0/Z, [x25]\n"
+      "trn1 z6.d, z7.d, z0.d\n"
+      "ld1rqb { z5.b }, p0/Z, [x24]\n"
+      "ld1rqb { z1.b }, p0/Z, [x23]\n"
+      "trn2 z7.d, z7.d, z0.d\n"
+      "trn1 z4.d, z5.d, z1.d\n"
+      "ld1rqb { z3.b }, p0/Z, [x22]\n"
       "ld1rqb { z0.b }, p0/Z, [x21]\n"
-      "add x23, x23, #0x10\n"
-      "add x22, x22, #0x10\n"
-      "trn1 z4.d, z6.d, z3.d\n"
-      "trn2 z6.d, z6.d, z3.d\n"
-      "add x21, x21, #0x10\n"
-      "trn1 z3.d, z7.d, z2.d\n"
-      "trn2 z7.d, z7.d, z2.d\n"
-      "trn1 z2.d, z5.d, z0.d\n"
-      "trn2 z5.d, z5.d, z0.d\n"
+      "trn2 z5.d, z5.d, z1.d\n"
+      "trn1 z2.d, z3.d, z0.d\n"
+      "trn2 z3.d, z3.d, z0.d\n"
+      "ld1b { z1.b }, p5/Z, [x10]\n"
       "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
-      ".inst 0x45c19888  // ummla z8.s, z4.b, z1.b\n"
-      ".inst 0x45c19870  // ummla z16.s, z3.b, z1.b\n"
+      ".inst 0x45c198c8  // ummla z8.s, z6.b, z1.b\n"
+      ".inst 0x45c19890  // ummla z16.s, z4.b, z1.b\n"
       ".inst 0x45c19858  // ummla z24.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n"
-      ".inst 0x45c0988c  // ummla z12.s, z4.b, z0.b\n"
-      ".inst 0x45c09874  // ummla z20.s, z3.b, z0.b\n"
+      "sub x27, x27, #0x10\n"
+      ".inst 0x45c098cc  // ummla z12.s, z6.b, z0.b\n"
+      ".inst 0x45c09894  // ummla z20.s, z4.b, z0.b\n"
+      "cmp x27, #0x10\n"
+      "add x26, x26, #0x10\n"
       ".inst 0x45c0985c  // ummla z28.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
-      ".inst 0x45c19889  // ummla z9.s, z4.b, z1.b\n"
-      ".inst 0x45c19871  // ummla z17.s, z3.b, z1.b\n"
+      ".inst 0x45c198c9  // ummla z9.s, z6.b, z1.b\n"
+      "add x25, x25, #0x10\n"
+      ".inst 0x45c19891  // ummla z17.s, z4.b, z1.b\n"
       ".inst 0x45c19859  // ummla z25.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n"
-      ".inst 0x45c0988d  // ummla z13.s, z4.b, z0.b\n"
-      ".inst 0x45c09875  // ummla z21.s, z3.b, z0.b\n"
+      "add x24, x24, #0x10\n"
+      ".inst 0x45c098cd  // ummla z13.s, z6.b, z0.b\n"
+      ".inst 0x45c09895  // ummla z21.s, z4.b, z0.b\n"
+      "add x23, x23, #0x10\n"
+      "add x22, x22, #0x10\n"
       ".inst 0x45c0985d  // ummla z29.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
-      ".inst 0x45c1988a  // ummla z10.s, z4.b, z1.b\n"
-      ".inst 0x45c19872  // ummla z18.s, z3.b, z1.b\n"
+      ".inst 0x45c198ca  // ummla z10.s, z6.b, z1.b\n"
+      "add x21, x21, #0x10\n"
+      ".inst 0x45c19892  // ummla z18.s, z4.b, z1.b\n"
       ".inst 0x45c1985a  // ummla z26.s, z2.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #6, MUL VL]\n"
-      ".inst 0x45c0988e  // ummla z14.s, z4.b, z0.b\n"
-      ".inst 0x45c09876  // ummla z22.s, z3.b, z0.b\n"
+      ".inst 0x45c098ce  // ummla z14.s, z6.b, z0.b\n"
+      ".inst 0x45c09896  // ummla z22.s, z4.b, z0.b\n"
       ".inst 0x45c0985e  // ummla z30.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
       "addvl x10, x10, #16\n"
-      ".inst 0x45c1988b  // ummla z11.s, z4.b, z1.b\n"
-      ".inst 0x45c19873  // ummla z19.s, z3.b, z1.b\n"
+      ".inst 0x45c198cb  // ummla z11.s, z6.b, z1.b\n"
+      ".inst 0x45c19893  // ummla z19.s, z4.b, z1.b\n"
       ".inst 0x45c1985b  // ummla z27.s, z2.b, z1.b\n"
-      ".inst 0x45c0988f  // ummla z15.s, z4.b, z0.b\n"
-      ".inst 0x45c09877  // ummla z23.s, z3.b, z0.b\n"
-      ".inst 0x45c0985f  // ummla z31.s, z2.b, z0.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n"
+      ".inst 0x45c098cf  // ummla z15.s, z6.b, z0.b\n"
+      ".inst 0x45c09897  // ummla z23.s, z4.b, z0.b\n"
+      ".inst 0x45c0985f  // ummla z31.s, z2.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-7, MUL VL]\n"
-      ".inst 0x45c198c8  // ummla z8.s, z6.b, z1.b\n"
-      ".inst 0x45c198f0  // ummla z16.s, z7.b, z1.b\n"
-      ".inst 0x45c198b8  // ummla z24.s, z5.b, z1.b\n"
+      ".inst 0x45c198e8  // ummla z8.s, z7.b, z1.b\n"
+      ".inst 0x45c198b0  // ummla z16.s, z5.b, z1.b\n"
+      ".inst 0x45c19878  // ummla z24.s, z3.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-6, MUL VL]\n"
-      ".inst 0x45c098cc  // ummla z12.s, z6.b, z0.b\n"
-      ".inst 0x45c098f4  // ummla z20.s, z7.b, z0.b\n"
-      ".inst 0x45c098bc  // ummla z28.s, z5.b, z0.b\n"
+      ".inst 0x45c098ec  // ummla z12.s, z7.b, z0.b\n"
+      ".inst 0x45c098b4  // ummla z20.s, z5.b, z0.b\n"
+      ".inst 0x45c0987c  // ummla z28.s, z3.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-5, MUL VL]\n"
-      ".inst 0x45c198c9  // ummla z9.s, z6.b, z1.b\n"
-      ".inst 0x45c198f1  // ummla z17.s, z7.b, z1.b\n"
-      ".inst 0x45c198b9  // ummla z25.s, z5.b, z1.b\n"
+      ".inst 0x45c198e9  // ummla z9.s, z7.b, z1.b\n"
+      ".inst 0x45c198b1  // ummla z17.s, z5.b, z1.b\n"
+      ".inst 0x45c19879  // ummla z25.s, z3.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-4, MUL VL]\n"
-      ".inst 0x45c098cd  // ummla z13.s, z6.b, z0.b\n"
-      ".inst 0x45c098f5  // ummla z21.s, z7.b, z0.b\n"
-      ".inst 0x45c098bd  // ummla z29.s, z5.b, z0.b\n"
+      ".inst 0x45c098ed  // ummla z13.s, z7.b, z0.b\n"
+      ".inst 0x45c098b5  // ummla z21.s, z5.b, z0.b\n"
+      ".inst 0x45c0987d  // ummla z29.s, z3.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-3, MUL VL]\n"
-      ".inst 0x45c198ca  // ummla z10.s, z6.b, z1.b\n"
-      ".inst 0x45c198f2  // ummla z18.s, z7.b, z1.b\n"
-      ".inst 0x45c198ba  // ummla z26.s, z5.b, z1.b\n"
+      ".inst 0x45c198ea  // ummla z10.s, z7.b, z1.b\n"
+      ".inst 0x45c198b2  // ummla z18.s, z5.b, z1.b\n"
+      ".inst 0x45c1987a  // ummla z26.s, z3.b, z1.b\n"
       "ld1b { z1.b }, p5/Z, [x10, #-2, MUL VL]\n"
-      ".inst 0x45c098ce  // ummla z14.s, z6.b, z0.b\n"
-      ".inst 0x45c098f6  // ummla z22.s, z7.b, z0.b\n"
-      ".inst 0x45c098be  // ummla z30.s, z5.b, z0.b\n"
+      ".inst 0x45c098ee  // ummla z14.s, z7.b, z0.b\n"
+      ".inst 0x45c098b6  // ummla z22.s, z5.b, z0.b\n"
+      ".inst 0x45c0987e  // ummla z30.s, z3.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #-1, MUL VL]\n"
-      ".inst 0x45c198cb  // ummla z11.s, z6.b, z1.b\n"
-      ".inst 0x45c198f3  // ummla z19.s, z7.b, z1.b\n"
-      ".inst 0x45c198bb  // ummla z27.s, z5.b, z1.b\n"
-      ".inst 0x45c098cf  // ummla z15.s, z6.b, z0.b\n"
-      ".inst 0x45c098f7  // ummla z23.s, z7.b, z0.b\n"
-      ".inst 0x45c098bf  // ummla z31.s, z5.b, z0.b\n"
+      ".inst 0x45c198eb  // ummla z11.s, z7.b, z1.b\n"
+      ".inst 0x45c198b3  // ummla z19.s, z5.b, z1.b\n"
+      ".inst 0x45c1987b  // ummla z27.s, z3.b, z1.b\n"
+      ".inst 0x45c098ef  // ummla z15.s, z7.b, z0.b\n"
+      ".inst 0x45c098b7  // ummla z23.s, z5.b, z0.b\n"
+      ".inst 0x45c0987f  // ummla z31.s, z3.b, z0.b\n"
       "bgt 63b\n"
       "64:"  // Height 6: Multiply loop: Single iteration only
       "whilelt p0.b, XZR, x27\n"
-      "ld1b { z2.b }, p5/Z, [x10]\n"
-      "subs x27, x27, #0x8\n"
       "ld1rqb { z1.b }, p0/Z, [x26]\n"
-      "ld1rqb { z6.b }, p0/Z, [x25]\n"
+      "ld1rqb { z0.b }, p0/Z, [x25]\n"
+      "trn1 z7.d, z1.d, z0.d\n"
       "ld1rqb { z3.b }, p0/Z, [x24]\n"
-      "ld1rqb { z4.b }, p0/Z, [x23]\n"
+      "ld1rqb { z2.b }, p0/Z, [x23]\n"
+      "trn2 z1.d, z1.d, z0.d\n"
+      "trn1 z6.d, z3.d, z2.d\n"
       "ld1rqb { z5.b }, p0/Z, [x22]\n"
       "ld1rqb { z0.b }, p0/Z, [x21]\n"
-      "trn1 z7.d, z1.d, z6.d\n"
-      "trn2 z1.d, z1.d, z6.d\n"
-      "trn1 z6.d, z3.d, z4.d\n"
-      "trn2 z3.d, z3.d, z4.d\n"
+      "trn2 z3.d, z3.d, z2.d\n"
       "trn1 z4.d, z5.d, z0.d\n"
       "trn2 z5.d, z5.d, z0.d\n"
+      "ld1b { z2.b }, p5/Z, [x10]\n"
       "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n"
       ".inst 0x45c298e8  // ummla z8.s, z7.b, z2.b\n"
       ".inst 0x45c298d0  // ummla z16.s, z6.b, z2.b\n"
       ".inst 0x45c29898  // ummla z24.s, z4.b, z2.b\n"
       "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
+      "subs x27, x27, #0x8\n"
       ".inst 0x45c098ec  // ummla z12.s, z7.b, z0.b\n"
       ".inst 0x45c098d4  // ummla z20.s, z6.b, z0.b\n"
       ".inst 0x45c0989c  // ummla z28.s, z4.b, z0.b\n"
@@ -1548,8 +1547,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       ".inst 0x45c098d6  // ummla z22.s, z6.b, z0.b\n"
       ".inst 0x45c0989e  // ummla z30.s, z4.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
-      "addvl x10, x10, #8\n"
       ".inst 0x45c298eb  // ummla z11.s, z7.b, z2.b\n"
+      "addvl x10, x10, #8\n"
       ".inst 0x45c298d3  // ummla z19.s, z6.b, z2.b\n"
       ".inst 0x45c2989b  // ummla z27.s, z4.b, z2.b\n"
       ".inst 0x45c098ef  // ummla z15.s, z7.b, z0.b\n"
@@ -1561,24 +1560,24 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       ".inst 0x45c29828  // ummla z8.s, z1.b, z2.b\n"
       ".inst 0x45c29870  // ummla z16.s, z3.b, z2.b\n"
       ".inst 0x45c298b8  // ummla z24.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45c0982c  // ummla z12.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n"
       ".inst 0x45c09874  // ummla z20.s, z3.b, z0.b\n"
       ".inst 0x45c098bc  // ummla z28.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n"
       ".inst 0x45c29829  // ummla z9.s, z1.b, z2.b\n"
       ".inst 0x45c29871  // ummla z17.s, z3.b, z2.b\n"
       ".inst 0x45c298b9  // ummla z25.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45c0982d  // ummla z13.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n"
       ".inst 0x45c09875  // ummla z21.s, z3.b, z0.b\n"
       ".inst 0x45c098bd  // ummla z29.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n"
       ".inst 0x45c2982a  // ummla z10.s, z1.b, z2.b\n"
       ".inst 0x45c29872  // ummla z18.s, z3.b, z2.b\n"
       ".inst 0x45c298ba  // ummla z26.s, z5.b, z2.b\n"
-      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45c0982e  // ummla z14.s, z1.b, z0.b\n"
+      "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n"
       ".inst 0x45c09876  // ummla z22.s, z3.b, z0.b\n"
       ".inst 0x45c098be  // ummla z30.s, z5.b, z0.b\n"
       "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n"
@@ -1595,46 +1594,46 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "cmp x28, x20\n"
       "bne 60b\n"
       "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+      "add x24, x9, x20, LSL #2\n"
+      "add x23, x24, x20, LSL #2\n"
       "uzp1 z0.d, z8.d, z12.d\n"
+      "add x22, x23, x20, LSL #2\n"
+      "add x21, x22, x20, LSL #2\n"
       "uzp2 z8.d, z8.d, z12.d\n"
       "uzp1 z12.d, z9.d, z13.d\n"
+      "add x20, x21, x20, LSL #2\n"
       "uzp2 z9.d, z9.d, z13.d\n"
       "uzp1 z13.d, z10.d, z14.d\n"
+      "st1w { z0.s }, p4, [x9]\n"
       "uzp2 z10.d, z10.d, z14.d\n"
-      "add x24, x9, x20, LSL #2\n"
-      "add x23, x24, x20, LSL #2\n"
-      "add x22, x23, x20, LSL #2\n"
       "uzp1 z14.d, z11.d, z15.d\n"
+      "st1w { z12.s }, p3, [x9, #1, MUL VL]\n"
       "uzp2 z11.d, z11.d, z15.d\n"
-      "st1w { z0.s }, p4, [x9]\n"
-      "add x21, x22, x20, LSL #2\n"
       "uzp1 z15.d, z16.d, z20.d\n"
+      "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "uzp2 z16.d, z16.d, z20.d\n"
-      "st1w { z12.s }, p3, [x9, #1, MUL VL]\n"
-      "add x20, x21, x20, LSL #2\n"
       "uzp1 z20.d, z17.d, z21.d\n"
+      "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
+      "addvl x9, x9, #4\n"
       "uzp2 z17.d, z17.d, z21.d\n"
-      "st1w { z13.s }, p2, [x9, #2, MUL VL]\n"
       "uzp1 z21.d, z18.d, z22.d\n"
+      "st1w { z8.s }, p4, [x24]\n"
       "uzp2 z18.d, z18.d, z22.d\n"
-      "st1w { z14.s }, p1, [x9, #3, MUL VL]\n"
-      "addvl x9, x9, #4\n"
       "uzp1 z22.d, z19.d, z23.d\n"
+      "st1w { z9.s }, p3, [x24, #1, MUL VL]\n"
       "uzp2 z19.d, z19.d, z23.d\n"
-      "st1w { z8.s }, p4, [x24]\n"
       "uzp1 z23.d, z24.d, z28.d\n"
+      "st1w { z10.s }, p2, [x24, #2, MUL VL]\n"
       "uzp2 z24.d, z24.d, z28.d\n"
-      "st1w { z9.s }, p3, [x24, #1, MUL VL]\n"
       "uzp1 z28.d, z25.d, z29.d\n"
+      "st1w { z11.s }, p1, [x24, #3, MUL VL]\n"
       "uzp2 z25.d, z25.d, z29.d\n"
-      "st1w { z10.s }, p2, [x24, #2, MUL VL]\n"
       "uzp1 z29.d, z26.d, z30.d\n"
+      "st1w { z15.s }, p4, [x23]\n"
       "uzp2 z26.d, z26.d, z30.d\n"
-      "st1w { z11.s }, p1, [x24, #3, MUL VL]\n"
       "uzp1 z30.d, z27.d, z31.d\n"
-      "uzp2 z27.d, z27.d, z31.d\n"
-      "st1w { z15.s }, p4, [x23]\n"
       "st1w { z20.s }, p3, [x23, #1, MUL VL]\n"
+      "uzp2 z27.d, z27.d, z31.d\n"
       "st1w { z21.s }, p2, [x23, #2, MUL VL]\n"
       "st1w { z22.s }, p1, [x23, #3, MUL VL]\n"
       "st1w { z16.s }, p4, [x22]\n"
@@ -1665,8 +1664,8 @@ void sve_hybrid_u8u32_mmla_6x4VL (
       "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
       "b 1b\n"
       "68:"  // Exit
-      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr)
-      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
+      : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+      : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
       : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL.hpp
index f9041edcca..ce3b070052 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL.hpp
@@ -40,8 +40,7 @@ void sve_interleaved_bf16fp32_dot_8x3VL( ARGLIST );
 class cls_sve_interleaved_bf16fp32_dot_8x3VL
 {
 public:
-    typedef bfloat16 lhs_operand_type;
-    typedef bfloat16 rhs_operand_type;
+    typedef bfloat16 operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 2, 1> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 2, 1, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 2, 1> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 2, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp
index 0646fa02eb..a6eefc1006 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_dot_8x3VL/generic.cpp
@@ -55,25 +55,25 @@ void sve_interleaved_bf16fp32_dot_8x3VL(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
+      "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
       "mov z10.b, #0x0\n"
-      "ld1h { z4.h }, p0/Z, [x22]\n"
       "mov z11.b, #0x0\n"
+      "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
       "mov z12.b, #0x0\n"
-      "ld1h { z5.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
       "mov z13.b, #0x0\n"
+      "ld1h { z4.h }, p0/Z, [x22]\n"
       "mov z14.b, #0x0\n"
       "mov z15.b, #0x0\n"
+      "ld1h { z5.h }, p0/Z, [x22, #1, MUL VL]\n"
       "mov z16.b, #0x0\n"
-      "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
       "mov z17.b, #0x0\n"
+      "ld1h { z6.h }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z18.b, #0x0\n"
-      "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
       "mov z19.b, #0x0\n"
       "mov z20.b, #0x0\n"
-      "ld1h { z6.h }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z21.b, #0x0\n"
       "mov z22.b, #0x0\n"
       "mov z23.b, #0x0\n"
@@ -151,12 +151,12 @@ void sve_interleaved_bf16fp32_dot_8x3VL(
       "ld1h { z6.h }, p0/Z, [x22, #2, MUL VL]\n"
       "bge 3b\n"
       "4:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "addvl x22, x22, #3\n"
       ".inst 0x64604088  // bfdot z8.s, z4.h, z0.h[0]\n"
       ".inst 0x6468408b  // bfdot z11.s, z4.h, z0.h[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       ".inst 0x6470408e  // bfdot z14.s, z4.h, z0.h[2]\n"
       ".inst 0x64784091  // bfdot z17.s, z4.h, z0.h[3]\n"
+      "addvl x22, x22, #3\n"
       ".inst 0x64614094  // bfdot z20.s, z4.h, z1.h[0]\n"
       ".inst 0x64694097  // bfdot z23.s, z4.h, z1.h[1]\n"
       ".inst 0x6471409a  // bfdot z26.s, z4.h, z1.h[2]\n"
@@ -183,13 +183,13 @@ void sve_interleaved_bf16fp32_dot_8x3VL(
       "add %x[Apanel], %x[Apanel], #0x20\n"
       "ld1h { z2.h }, p0/Z, [x22]\n"
       "ld1h { z1.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "ld1h { z0.h }, p0/Z, [x22, #2, MUL VL]\n"
-      "addvl x22, x22, #3\n"
       ".inst 0x64644048  // bfdot z8.s, z2.h, z4.h[0]\n"
+      "ld1h { z0.h }, p0/Z, [x22, #2, MUL VL]\n"
       ".inst 0x646c404b  // bfdot z11.s, z2.h, z4.h[1]\n"
       ".inst 0x6474404e  // bfdot z14.s, z2.h, z4.h[2]\n"
       ".inst 0x647c4051  // bfdot z17.s, z2.h, z4.h[3]\n"
       ".inst 0x64634054  // bfdot z20.s, z2.h, z3.h[0]\n"
+      "addvl x22, x22, #3\n"
       ".inst 0x646b4057  // bfdot z23.s, z2.h, z3.h[1]\n"
       ".inst 0x6473405a  // bfdot z26.s, z2.h, z3.h[2]\n"
       ".inst 0x647b405d  // bfdot z29.s, z2.h, z3.h[3]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp
index 5c0d59609b..42a7dbb37a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp
@@ -40,8 +40,7 @@ void sve_interleaved_bf16fp32_mmla_8x3VL( ARGLIST );
 class cls_sve_interleaved_bf16fp32_mmla_8x3VL
 {
 public:
-    typedef bfloat16 lhs_operand_type;
-    typedef bfloat16 rhs_operand_type;
+    typedef bfloat16 operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 4, 2> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 4, 2, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp
index 9cde63f9d7..3ffec98c16 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL/generic.cpp
@@ -55,31 +55,31 @@ void sve_interleaved_bf16fp32_mmla_8x3VL(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
-      "mov z10.b, #0x0\n"
       "ld1h { z4.h }, p0/Z, [x22]\n"
+      "mov z10.b, #0x0\n"
       "mov z11.b, #0x0\n"
+      "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
       "mov z12.b, #0x0\n"
-      "ld1h { z5.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
       "mov z13.b, #0x0\n"
+      "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
       "mov z14.b, #0x0\n"
       "mov z15.b, #0x0\n"
+      "ld1h { z5.h }, p0/Z, [x22, #1, MUL VL]\n"
       "mov z16.b, #0x0\n"
-      "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
       "mov z17.b, #0x0\n"
+      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #32]\n"
       "mov z18.b, #0x0\n"
-      "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #16]\n"
       "mov z19.b, #0x0\n"
+      "addvl x22, x22, #2\n"
       "mov z20.b, #0x0\n"
-      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #32]\n"
       "mov z21.b, #0x0\n"
+      "add %x[Apanel], %x[Apanel], #0x30\n"
       "mov z22.b, #0x0\n"
-      "addvl x22, x22, #2\n"
       "mov z23.b, #0x0\n"
       "mov z24.b, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
       "mov z25.b, #0x0\n"
       "mov z26.b, #0x0\n"
       "mov z27.b, #0x0\n"
@@ -94,77 +94,77 @@ void sve_interleaved_bf16fp32_mmla_8x3VL(
       ".inst 0x6465e40b  // bfmmla z11.s, z0.h, z5.h\n"
       ".inst 0x6464e42e  // bfmmla z14.s, z1.h, z4.h\n"
       ".inst 0x6465e431  // bfmmla z17.s, z1.h, z5.h\n"
-      "ld1h { z3.h }, p0/Z, [x22]\n"
+      "ld1h { z7.h }, p0/Z, [x22]\n"
       ".inst 0x6464e454  // bfmmla z20.s, z2.h, z4.h\n"
       ".inst 0x6465e457  // bfmmla z23.s, z2.h, z5.h\n"
-      "ld1h { z7.h }, p0/Z, [x22, #1, MUL VL]\n"
+      "ld1h { z3.h }, p0/Z, [x22, #1, MUL VL]\n"
       ".inst 0x6464e4da  // bfmmla z26.s, z6.h, z4.h\n"
       ".inst 0x6465e4dd  // bfmmla z29.s, z6.h, z5.h\n"
-      "ld1h { z4.h }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1h { z5.h }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x6463e409  // bfmmla z9.s, z0.h, z3.h\n"
+      "ld1h { z5.h }, p0/Z, [x22, #2, MUL VL]\n"
+      "ld1h { z4.h }, p0/Z, [x22, #3, MUL VL]\n"
+      ".inst 0x6467e409  // bfmmla z9.s, z0.h, z7.h\n"
+      ".inst 0x6463e40c  // bfmmla z12.s, z0.h, z3.h\n"
+      ".inst 0x6467e42f  // bfmmla z15.s, z1.h, z7.h\n"
+      ".inst 0x6463e432  // bfmmla z18.s, z1.h, z3.h\n"
       "sub x20, x20, #0x2\n"
-      ".inst 0x6467e40c  // bfmmla z12.s, z0.h, z7.h\n"
-      ".inst 0x6463e42f  // bfmmla z15.s, z1.h, z3.h\n"
+      ".inst 0x6467e455  // bfmmla z21.s, z2.h, z7.h\n"
+      ".inst 0x6463e458  // bfmmla z24.s, z2.h, z3.h\n"
       "cmp x20, #0x2\n"
-      ".inst 0x6467e432  // bfmmla z18.s, z1.h, z7.h\n"
-      ".inst 0x6463e455  // bfmmla z21.s, z2.h, z3.h\n"
-      ".inst 0x6467e458  // bfmmla z24.s, z2.h, z7.h\n"
-      ".inst 0x6463e4db  // bfmmla z27.s, z6.h, z3.h\n"
+      ".inst 0x6467e4db  // bfmmla z27.s, z6.h, z7.h\n"
+      ".inst 0x6463e4de  // bfmmla z30.s, z6.h, z3.h\n"
       "ld1h { z3.h }, p0/Z, [x22, #4, MUL VL]\n"
-      ".inst 0x6467e4de  // bfmmla z30.s, z6.h, z7.h\n"
-      ".inst 0x6464e40a  // bfmmla z10.s, z0.h, z4.h\n"
-      "ld1h { z7.h }, p0/Z, [x22, #5, MUL VL]\n"
-      ".inst 0x6465e40d  // bfmmla z13.s, z0.h, z5.h\n"
-      ".inst 0x6464e430  // bfmmla z16.s, z1.h, z4.h\n"
+      ".inst 0x6465e40a  // bfmmla z10.s, z0.h, z5.h\n"
+      ".inst 0x6464e40d  // bfmmla z13.s, z0.h, z4.h\n"
       "ld1rqh { z0.h }, p0/Z, [%x[Apanel], #16]\n"
-      ".inst 0x6465e433  // bfmmla z19.s, z1.h, z5.h\n"
-      ".inst 0x6464e456  // bfmmla z22.s, z2.h, z4.h\n"
+      ".inst 0x6465e430  // bfmmla z16.s, z1.h, z5.h\n"
+      ".inst 0x6464e433  // bfmmla z19.s, z1.h, z4.h\n"
       "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #32]\n"
-      ".inst 0x6465e459  // bfmmla z25.s, z2.h, z5.h\n"
-      ".inst 0x6464e4dc  // bfmmla z28.s, z6.h, z4.h\n"
-      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #48]\n"
-      ".inst 0x6465e4df  // bfmmla z31.s, z6.h, z5.h\n"
+      ".inst 0x6465e456  // bfmmla z22.s, z2.h, z5.h\n"
+      ".inst 0x6464e459  // bfmmla z25.s, z2.h, z4.h\n"
+      "ld1h { z7.h }, p0/Z, [x22, #5, MUL VL]\n"
+      ".inst 0x6465e4dc  // bfmmla z28.s, z6.h, z5.h\n"
+      ".inst 0x6464e4df  // bfmmla z31.s, z6.h, z4.h\n"
+      "ld1rqh { z5.h }, p0/Z, [%x[Apanel], #48]\n"
       "ld1rqh { z6.h }, p0/Z, [%x[Apanel], #64]\n"
-      "ld1h { z4.h }, p0/Z, [x22, #6, MUL VL]\n"
-      "ld1h { z5.h }, p0/Z, [x22, #7, MUL VL]\n"
-      "addvl x22, x22, #16\n"
+      "ld1h { z2.h }, p0/Z, [x22, #6, MUL VL]\n"
       ".inst 0x6463e408  // bfmmla z8.s, z0.h, z3.h\n"
+      "ld1h { z4.h }, p0/Z, [x22, #7, MUL VL]\n"
+      "addvl x22, x22, #16\n"
       ".inst 0x6467e40b  // bfmmla z11.s, z0.h, z7.h\n"
       ".inst 0x6463e42e  // bfmmla z14.s, z1.h, z3.h\n"
       ".inst 0x6467e431  // bfmmla z17.s, z1.h, z7.h\n"
-      ".inst 0x6463e454  // bfmmla z20.s, z2.h, z3.h\n"
-      ".inst 0x6467e457  // bfmmla z23.s, z2.h, z7.h\n"
+      ".inst 0x6463e4b4  // bfmmla z20.s, z5.h, z3.h\n"
+      ".inst 0x6467e4b7  // bfmmla z23.s, z5.h, z7.h\n"
       ".inst 0x6463e4da  // bfmmla z26.s, z6.h, z3.h\n"
-      "ld1h { z3.h }, p0/Z, [x22, #-8, MUL VL]\n"
       ".inst 0x6467e4dd  // bfmmla z29.s, z6.h, z7.h\n"
+      "ld1h { z3.h }, p0/Z, [x22, #-8, MUL VL]\n"
       "ld1h { z7.h }, p0/Z, [x22, #-7, MUL VL]\n"
-      ".inst 0x6464e409  // bfmmla z9.s, z0.h, z4.h\n"
-      ".inst 0x6465e40c  // bfmmla z12.s, z0.h, z5.h\n"
-      ".inst 0x6464e42f  // bfmmla z15.s, z1.h, z4.h\n"
-      ".inst 0x6465e432  // bfmmla z18.s, z1.h, z5.h\n"
-      ".inst 0x6464e455  // bfmmla z21.s, z2.h, z4.h\n"
-      ".inst 0x6465e458  // bfmmla z24.s, z2.h, z5.h\n"
-      ".inst 0x6464e4db  // bfmmla z27.s, z6.h, z4.h\n"
+      ".inst 0x6462e409  // bfmmla z9.s, z0.h, z2.h\n"
+      ".inst 0x6464e40c  // bfmmla z12.s, z0.h, z4.h\n"
+      ".inst 0x6462e42f  // bfmmla z15.s, z1.h, z2.h\n"
+      ".inst 0x6464e432  // bfmmla z18.s, z1.h, z4.h\n"
+      ".inst 0x6462e4b5  // bfmmla z21.s, z5.h, z2.h\n"
+      ".inst 0x6464e4b8  // bfmmla z24.s, z5.h, z4.h\n"
+      ".inst 0x6462e4db  // bfmmla z27.s, z6.h, z2.h\n"
+      ".inst 0x6464e4de  // bfmmla z30.s, z6.h, z4.h\n"
       "ld1h { z4.h }, p0/Z, [x22, #-6, MUL VL]\n"
-      ".inst 0x6465e4de  // bfmmla z30.s, z6.h, z5.h\n"
       ".inst 0x6463e40a  // bfmmla z10.s, z0.h, z3.h\n"
-      "ld1h { z5.h }, p0/Z, [x22, #-5, MUL VL]\n"
       ".inst 0x6467e40d  // bfmmla z13.s, z0.h, z7.h\n"
-      ".inst 0x6463e430  // bfmmla z16.s, z1.h, z3.h\n"
       "ld1rqh { z0.h }, p0/Z, [%x[Apanel], #80]\n"
+      ".inst 0x6463e430  // bfmmla z16.s, z1.h, z3.h\n"
       ".inst 0x6467e433  // bfmmla z19.s, z1.h, z7.h\n"
-      ".inst 0x6463e456  // bfmmla z22.s, z2.h, z3.h\n"
       "ld1rqh { z1.h }, p0/Z, [%x[Apanel], #96]\n"
-      ".inst 0x6467e459  // bfmmla z25.s, z2.h, z7.h\n"
+      ".inst 0x6463e4b6  // bfmmla z22.s, z5.h, z3.h\n"
+      ".inst 0x6467e4b9  // bfmmla z25.s, z5.h, z7.h\n"
+      "ld1h { z5.h }, p0/Z, [x22, #-5, MUL VL]\n"
       ".inst 0x6463e4dc  // bfmmla z28.s, z6.h, z3.h\n"
-      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #112]\n"
       ".inst 0x6467e4df  // bfmmla z31.s, z6.h, z7.h\n"
+      "ld1rqh { z2.h }, p0/Z, [%x[Apanel], #112]\n"
       "add %x[Apanel], %x[Apanel], #0x80\n"
       "addvl x22, x22, #-4\n"
       "bge 3b\n"
       "4:"  // main loop skip
-      "ld1rqh { z3.h }, p0/Z, [%x[Apanel]]\n"
+      "ld1rqh { z7.h }, p0/Z, [%x[Apanel]]\n"
       ".inst 0x6464e408  // bfmmla z8.s, z0.h, z4.h\n"
       ".inst 0x6465e40b  // bfmmla z11.s, z0.h, z5.h\n"
       ".inst 0x6464e42e  // bfmmla z14.s, z1.h, z4.h\n"
@@ -172,54 +172,54 @@ void sve_interleaved_bf16fp32_mmla_8x3VL(
       "ld1h { z6.h }, p0/Z, [x22]\n"
       ".inst 0x6464e454  // bfmmla z20.s, z2.h, z4.h\n"
       ".inst 0x6465e457  // bfmmla z23.s, z2.h, z5.h\n"
-      "ld1h { z7.h }, p0/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x6464e47a  // bfmmla z26.s, z3.h, z4.h\n"
-      ".inst 0x6465e47d  // bfmmla z29.s, z3.h, z5.h\n"
+      "ld1h { z3.h }, p0/Z, [x22, #1, MUL VL]\n"
+      ".inst 0x6464e4fa  // bfmmla z26.s, z7.h, z4.h\n"
+      ".inst 0x6465e4fd  // bfmmla z29.s, z7.h, z5.h\n"
       "ld1h { z5.h }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1h { z4.h }, p0/Z, [x22, #3, MUL VL]\n"
       ".inst 0x6466e409  // bfmmla z9.s, z0.h, z6.h\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      ".inst 0x6467e40c  // bfmmla z12.s, z0.h, z7.h\n"
+      ".inst 0x6463e40c  // bfmmla z12.s, z0.h, z3.h\n"
       ".inst 0x6466e42f  // bfmmla z15.s, z1.h, z6.h\n"
-      "addvl x22, x22, #4\n"
-      ".inst 0x6467e432  // bfmmla z18.s, z1.h, z7.h\n"
+      ".inst 0x6463e432  // bfmmla z18.s, z1.h, z3.h\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       ".inst 0x6466e455  // bfmmla z21.s, z2.h, z6.h\n"
-      ".inst 0x6467e458  // bfmmla z24.s, z2.h, z7.h\n"
-      ".inst 0x6466e47b  // bfmmla z27.s, z3.h, z6.h\n"
-      ".inst 0x6467e47e  // bfmmla z30.s, z3.h, z7.h\n"
+      ".inst 0x6463e458  // bfmmla z24.s, z2.h, z3.h\n"
+      "addvl x22, x22, #4\n"
+      ".inst 0x6466e4fb  // bfmmla z27.s, z7.h, z6.h\n"
+      ".inst 0x6463e4fe  // bfmmla z30.s, z7.h, z3.h\n"
       ".inst 0x6465e40a  // bfmmla z10.s, z0.h, z5.h\n"
       ".inst 0x6464e40d  // bfmmla z13.s, z0.h, z4.h\n"
       ".inst 0x6465e430  // bfmmla z16.s, z1.h, z5.h\n"
       ".inst 0x6464e433  // bfmmla z19.s, z1.h, z4.h\n"
       ".inst 0x6465e456  // bfmmla z22.s, z2.h, z5.h\n"
       ".inst 0x6464e459  // bfmmla z25.s, z2.h, z4.h\n"
-      ".inst 0x6465e47c  // bfmmla z28.s, z3.h, z5.h\n"
-      ".inst 0x6464e47f  // bfmmla z31.s, z3.h, z4.h\n"
+      ".inst 0x6465e4fc  // bfmmla z28.s, z7.h, z5.h\n"
+      ".inst 0x6464e4ff  // bfmmla z31.s, z7.h, z4.h\n"
       "cbz x20, 5f\n"
       "ld1h { z1.h }, p0/Z, [x22]\n"
       "ld1rqh { z7.h }, p0/Z, [%x[Apanel]]\n"
+      ".inst 0x6461e4e8  // bfmmla z8.s, z7.h, z1.h\n"
       "ld1rqh { z6.h }, p0/Z, [%x[Apanel], #16]\n"
       "ld1h { z0.h }, p0/Z, [x22, #1, MUL VL]\n"
+      ".inst 0x6460e4eb  // bfmmla z11.s, z7.h, z0.h\n"
       "ld1rqh { z5.h }, p0/Z, [%x[Apanel], #32]\n"
       "ld1rqh { z4.h }, p0/Z, [%x[Apanel], #48]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
-      ".inst 0x6461e4e8  // bfmmla z8.s, z7.h, z1.h\n"
-      "ld1h { z3.h }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1h { z2.h }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x6460e4eb  // bfmmla z11.s, z7.h, z0.h\n"
       ".inst 0x6461e4ce  // bfmmla z14.s, z6.h, z1.h\n"
       ".inst 0x6460e4d1  // bfmmla z17.s, z6.h, z0.h\n"
       ".inst 0x6461e4b4  // bfmmla z20.s, z5.h, z1.h\n"
+      "ld1h { z3.h }, p0/Z, [x22, #2, MUL VL]\n"
       ".inst 0x6460e4b7  // bfmmla z23.s, z5.h, z0.h\n"
       ".inst 0x6461e49a  // bfmmla z26.s, z4.h, z1.h\n"
-      "ld1h { z1.h }, p0/Z, [x22, #4, MUL VL]\n"
+      "ld1h { z2.h }, p0/Z, [x22, #3, MUL VL]\n"
       ".inst 0x6460e49d  // bfmmla z29.s, z4.h, z0.h\n"
+      "ld1h { z1.h }, p0/Z, [x22, #4, MUL VL]\n"
       "ld1h { z0.h }, p0/Z, [x22, #5, MUL VL]\n"
       ".inst 0x6463e4e9  // bfmmla z9.s, z7.h, z3.h\n"
       ".inst 0x6462e4ec  // bfmmla z12.s, z7.h, z2.h\n"
-      ".inst 0x6463e4cf  // bfmmla z15.s, z6.h, z3.h\n"
       "addvl x22, x22, #6\n"
+      ".inst 0x6463e4cf  // bfmmla z15.s, z6.h, z3.h\n"
       ".inst 0x6462e4d2  // bfmmla z18.s, z6.h, z2.h\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
       ".inst 0x6463e4b5  // bfmmla z21.s, z5.h, z3.h\n"
       ".inst 0x6462e4b8  // bfmmla z24.s, z5.h, z2.h\n"
       ".inst 0x6463e49b  // bfmmla z27.s, z4.h, z3.h\n"
@@ -233,53 +233,53 @@ void sve_interleaved_bf16fp32_mmla_8x3VL(
       ".inst 0x6461e49c  // bfmmla z28.s, z4.h, z1.h\n"
       ".inst 0x6460e49f  // bfmmla z31.s, z4.h, z0.h\n"
       "5:"  // multiply loop done
-      "uzp1 z2.d, z8.d, z11.d\n"
+      "uzp1 z0.d, z8.d, z11.d\n"
       "uzp2 z8.d, z8.d, z11.d\n"
-      "subs x23, x23, #0x1\n"
-      "uzp1 z1.d, z9.d, z12.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel]]\n"
+      "uzp1 z0.d, z9.d, z12.d\n"
       "uzp2 z9.d, z9.d, z12.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
       "uzp1 z0.d, z10.d, z13.d\n"
       "uzp2 z10.d, z10.d, z13.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel]]\n"
-      "uzp1 z3.d, z14.d, z17.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+      "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+      "uzp1 z0.d, z14.d, z17.d\n"
       "uzp2 z14.d, z14.d, z17.d\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "uzp1 z17.d, z15.d, z18.d\n"
+      "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+      "uzp1 z1.d, z15.d, z18.d\n"
+      "subs x23, x23, #0x1\n"
+      "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "uzp2 z15.d, z15.d, z18.d\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "uzp1 z2.d, z16.d, z19.d\n"
+      "uzp1 z17.d, z16.d, z19.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
       "uzp2 z16.d, z16.d, z19.d\n"
-      "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "uzp1 z1.d, z20.d, z23.d\n"
+      "uzp1 z0.d, z20.d, z23.d\n"
+      "st1w { z1.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
+      "addvl %x[Cpanel], %x[Cpanel], #16\n"
       "uzp2 z20.d, z20.d, z23.d\n"
-      "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
-      "uzp1 z0.d, z21.d, z24.d\n"
+      "st1w { z17.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
+      "uzp1 z23.d, z21.d, z24.d\n"
       "uzp2 z21.d, z21.d, z24.d\n"
-      "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
-      "uzp1 z23.d, z22.d, z25.d\n"
+      "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
+      "uzp1 z19.d, z22.d, z25.d\n"
       "uzp2 z22.d, z22.d, z25.d\n"
-      "st1w { z3.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
-      "uzp1 z19.d, z26.d, z29.d\n"
+      "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
+      "uzp1 z18.d, z26.d, z29.d\n"
       "uzp2 z26.d, z26.d, z29.d\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
-      "addvl %x[Cpanel], %x[Cpanel], #16\n"
-      "uzp1 z18.d, z27.d, z30.d\n"
+      "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
+      "uzp1 z17.d, z27.d, z30.d\n"
       "uzp2 z27.d, z27.d, z30.d\n"
-      "uzp1 z17.d, z28.d, z31.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
+      "uzp1 z16.d, z28.d, z31.d\n"
       "uzp2 z28.d, z28.d, z31.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
-      "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
-      "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
-      "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
-      "st1w { z23.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
+      "st1w { z23.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
+      "st1w { z19.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
       "st1w { z20.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
       "st1w { z21.s }, p0, [%x[Cpanel]]\n"
       "st1w { z22.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "st1w { z19.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "st1w { z18.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+      "st1w { z18.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+      "st1w { z17.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+      "st1w { z16.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
       "st1w { z26.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "st1w { z27.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
       "st1w { z28.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL.hpp
index 292ac1760e..e04e3d2a1a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL.hpp
@@ -40,8 +40,7 @@ void sve_interleaved_fp16_mla_8x3VL_a64fx( ARGLIST );
 class cls_sve_interleaved_fp16_mla_8x3VL
 {
 public:
-    typedef __fp16 lhs_operand_type;
-    typedef __fp16 rhs_operand_type;
+    typedef __fp16 operand_type;
     typedef __fp16 result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 1, 1> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/a64fx.cpp
index 360c61f0b4..6e19811d72 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/a64fx.cpp
@@ -54,31 +54,31 @@ void sve_interleaved_fp16_mla_8x3VL_a64fx(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
-      "mov z10.b, #0x0\n"
       "ld1h { z0.h }, p0/Z, [x22]\n"
+      "mov z10.b, #0x0\n"
       "mov z11.b, #0x0\n"
-      "mov z12.b, #0x0\n"
       "ld1h { z1.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
+      "mov z12.b, #0x0\n"
       "mov z13.b, #0x0\n"
+      "ld1h { z2.h }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z14.b, #0x0\n"
       "mov z15.b, #0x0\n"
+      "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n"
       "mov z16.b, #0x0\n"
-      "ld1h { z2.h }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z17.b, #0x0\n"
+      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n"
       "mov z18.b, #0x0\n"
-      "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n"
       "mov z19.b, #0x0\n"
+      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n"
       "mov z20.b, #0x0\n"
-      "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n"
       "mov z21.b, #0x0\n"
+      "ld1rh { z6.h }, p0/Z, [%x[Apanel], #6]\n"
       "mov z22.b, #0x0\n"
-      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n"
       "mov z23.b, #0x0\n"
       "mov z24.b, #0x0\n"
-      "ld1rh { z6.h }, p0/Z, [%x[Apanel], #6]\n"
       "mov z25.b, #0x0\n"
       "mov z26.b, #0x0\n"
       "mov z27.b, #0x0\n"
@@ -92,7 +92,7 @@ void sve_interleaved_fp16_mla_8x3VL_a64fx(
       "fmla z9.h, p0/M, z1.h, z3.h\n"
       "sub x20, x20, #0x2\n"
       "fmla z10.h, p0/M, z2.h, z3.h\n"
-      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #8]\n"
+      "ld1rh { z7.h }, p0/Z, [%x[Apanel], #8]\n"
       "fmla z11.h, p0/M, z0.h, z4.h\n"
       "fmla z12.h, p0/M, z1.h, z4.h\n"
       "fmla z13.h, p0/M, z2.h, z4.h\n"
@@ -101,63 +101,63 @@ void sve_interleaved_fp16_mla_8x3VL_a64fx(
       "fmla z15.h, p0/M, z1.h, z5.h\n"
       "cmp x20, #0x2\n"
       "fmla z16.h, p0/M, z2.h, z5.h\n"
-      "ld1rh { z7.h }, p0/Z, [%x[Apanel], #12]\n"
+      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #12]\n"
       "fmla z17.h, p0/M, z0.h, z6.h\n"
       "fmla z18.h, p0/M, z1.h, z6.h\n"
       "fmla z19.h, p0/M, z2.h, z6.h\n"
-      "ld1rh { z6.h }, p0/Z, [%x[Apanel], #14]\n"
-      "fmla z20.h, p0/M, z0.h, z3.h\n"
-      "fmla z21.h, p0/M, z1.h, z3.h\n"
-      "fmla z22.h, p0/M, z2.h, z3.h\n"
-      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #16]\n"
+      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #14]\n"
+      "fmla z20.h, p0/M, z0.h, z7.h\n"
+      "fmla z21.h, p0/M, z1.h, z7.h\n"
+      "fmla z22.h, p0/M, z2.h, z7.h\n"
+      "ld1rh { z7.h }, p0/Z, [%x[Apanel], #16]\n"
       "fmla z23.h, p0/M, z0.h, z4.h\n"
       "fmla z24.h, p0/M, z1.h, z4.h\n"
       "fmla z25.h, p0/M, z2.h, z4.h\n"
       "ld1rh { z4.h }, p0/Z, [%x[Apanel], #18]\n"
-      "fmla z26.h, p0/M, z0.h, z7.h\n"
-      "fmla z27.h, p0/M, z1.h, z7.h\n"
-      "fmla z28.h, p0/M, z2.h, z7.h\n"
+      "fmla z26.h, p0/M, z0.h, z3.h\n"
+      "fmla z27.h, p0/M, z1.h, z3.h\n"
+      "fmla z28.h, p0/M, z2.h, z3.h\n"
       "ld1rh { z3.h }, p0/Z, [%x[Apanel], #20]\n"
-      "fmla z29.h, p0/M, z0.h, z6.h\n"
-      "ld1h { z7.h }, p0/Z, [x22, #3, MUL VL]\n"
-      "fmla z30.h, p0/M, z1.h, z6.h\n"
-      "fmla z31.h, p0/M, z2.h, z6.h\n"
-      "ld1h { z6.h }, p0/Z, [x22, #4, MUL VL]\n"
-      "ld1h { z2.h }, p0/Z, [x22, #5, MUL VL]\n"
-      "addvl x22, x22, #6\n"
+      "fmla z29.h, p0/M, z0.h, z5.h\n"
+      "ld1h { z6.h }, p0/Z, [x22, #3, MUL VL]\n"
+      "fmla z30.h, p0/M, z1.h, z5.h\n"
+      "fmla z31.h, p0/M, z2.h, z5.h\n"
+      "ld1h { z2.h }, p0/Z, [x22, #4, MUL VL]\n"
+      "ld1h { z5.h }, p0/Z, [x22, #5, MUL VL]\n"
+      "fmla z8.h, p0/M, z6.h, z7.h\n"
       "ld1rh { z1.h }, p0/Z, [%x[Apanel], #22]\n"
-      "fmla z8.h, p0/M, z7.h, z5.h\n"
-      "fmla z11.h, p0/M, z7.h, z4.h\n"
-      "fmla z9.h, p0/M, z6.h, z5.h\n"
-      "fmla z12.h, p0/M, z6.h, z4.h\n"
-      "fmla z10.h, p0/M, z2.h, z5.h\n"
-      "fmla z13.h, p0/M, z2.h, z4.h\n"
-      "ld1rh { z5.h }, p0/Z, [%x[Apanel], #24]\n"
-      "fmla z14.h, p0/M, z7.h, z3.h\n"
-      "fmla z15.h, p0/M, z6.h, z3.h\n"
+      "fmla z9.h, p0/M, z2.h, z7.h\n"
+      "fmla z10.h, p0/M, z5.h, z7.h\n"
+      "fmla z11.h, p0/M, z6.h, z4.h\n"
+      "ld1rh { z7.h }, p0/Z, [%x[Apanel], #24]\n"
+      "fmla z12.h, p0/M, z2.h, z4.h\n"
+      "fmla z13.h, p0/M, z5.h, z4.h\n"
       "ld1rh { z4.h }, p0/Z, [%x[Apanel], #26]\n"
-      "fmla z16.h, p0/M, z2.h, z3.h\n"
-      "fmla z17.h, p0/M, z7.h, z1.h\n"
+      "fmla z14.h, p0/M, z6.h, z3.h\n"
+      "fmla z15.h, p0/M, z2.h, z3.h\n"
+      "addvl x22, x22, #6\n"
+      "fmla z16.h, p0/M, z5.h, z3.h\n"
       "ld1rh { z0.h }, p0/Z, [%x[Apanel], #28]\n"
-      "fmla z18.h, p0/M, z6.h, z1.h\n"
-      "fmla z19.h, p0/M, z2.h, z1.h\n"
+      "fmla z17.h, p0/M, z6.h, z1.h\n"
+      "fmla z18.h, p0/M, z2.h, z1.h\n"
+      "fmla z19.h, p0/M, z5.h, z1.h\n"
       "ld1rh { z1.h }, p0/Z, [%x[Apanel], #30]\n"
       "add %x[Apanel], %x[Apanel], #0x20\n"
-      "fmla z20.h, p0/M, z7.h, z5.h\n"
-      "fmla z21.h, p0/M, z6.h, z5.h\n"
-      "fmla z22.h, p0/M, z2.h, z5.h\n"
-      "fmla z23.h, p0/M, z7.h, z4.h\n"
+      "fmla z20.h, p0/M, z6.h, z7.h\n"
+      "fmla z21.h, p0/M, z2.h, z7.h\n"
+      "fmla z22.h, p0/M, z5.h, z7.h\n"
+      "fmla z23.h, p0/M, z6.h, z4.h\n"
       "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n"
-      "fmla z24.h, p0/M, z6.h, z4.h\n"
-      "fmla z25.h, p0/M, z2.h, z4.h\n"
+      "fmla z24.h, p0/M, z2.h, z4.h\n"
+      "fmla z25.h, p0/M, z5.h, z4.h\n"
       "ld1rh { z4.h }, p0/Z, [%x[Apanel], #2]\n"
-      "fmla z26.h, p0/M, z7.h, z0.h\n"
-      "fmla z27.h, p0/M, z6.h, z0.h\n"
-      "fmla z28.h, p0/M, z2.h, z0.h\n"
-      "fmla z29.h, p0/M, z7.h, z1.h\n"
+      "fmla z26.h, p0/M, z6.h, z0.h\n"
+      "fmla z27.h, p0/M, z2.h, z0.h\n"
+      "fmla z28.h, p0/M, z5.h, z0.h\n"
+      "fmla z29.h, p0/M, z6.h, z1.h\n"
       "ld1h { z0.h }, p0/Z, [x22]\n"
-      "fmla z30.h, p0/M, z6.h, z1.h\n"
-      "fmla z31.h, p0/M, z2.h, z1.h\n"
+      "fmla z30.h, p0/M, z2.h, z1.h\n"
+      "fmla z31.h, p0/M, z5.h, z1.h\n"
       "ld1h { z1.h }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1h { z2.h }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1rh { z5.h }, p0/Z, [%x[Apanel], #4]\n"
@@ -199,20 +199,19 @@ void sve_interleaved_fp16_mla_8x3VL_a64fx(
       "ld1h { z5.h }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1h { z4.h }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1rh { z3.h }, p0/Z, [%x[Apanel]]\n"
-      "addvl x22, x22, #3\n"
+      "fmla z8.h, p0/M, z6.h, z3.h\n"
       "ld1rh { z2.h }, p0/Z, [%x[Apanel], #2]\n"
       "ld1rh { z1.h }, p0/Z, [%x[Apanel], #4]\n"
-      "ld1rh { z0.h }, p0/Z, [%x[Apanel], #6]\n"
-      "fmla z8.h, p0/M, z6.h, z3.h\n"
       "fmla z9.h, p0/M, z5.h, z3.h\n"
+      "ld1rh { z0.h }, p0/Z, [%x[Apanel], #6]\n"
       "fmla z10.h, p0/M, z4.h, z3.h\n"
       "fmla z11.h, p0/M, z6.h, z2.h\n"
-      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #8]\n"
       "fmla z12.h, p0/M, z5.h, z2.h\n"
       "fmla z13.h, p0/M, z4.h, z2.h\n"
-      "ld1rh { z2.h }, p0/Z, [%x[Apanel], #10]\n"
+      "ld1rh { z3.h }, p0/Z, [%x[Apanel], #8]\n"
       "fmla z14.h, p0/M, z6.h, z1.h\n"
       "fmla z15.h, p0/M, z5.h, z1.h\n"
+      "ld1rh { z2.h }, p0/Z, [%x[Apanel], #10]\n"
       "fmla z16.h, p0/M, z4.h, z1.h\n"
       "fmla z17.h, p0/M, z6.h, z0.h\n"
       "ld1rh { z1.h }, p0/Z, [%x[Apanel], #12]\n"
@@ -221,9 +220,10 @@ void sve_interleaved_fp16_mla_8x3VL_a64fx(
       "ld1rh { z0.h }, p0/Z, [%x[Apanel], #14]\n"
       "fmla z20.h, p0/M, z6.h, z3.h\n"
       "fmla z21.h, p0/M, z5.h, z3.h\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
+      "addvl x22, x22, #3\n"
       "fmla z22.h, p0/M, z4.h, z3.h\n"
       "fmla z23.h, p0/M, z6.h, z2.h\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       "fmla z24.h, p0/M, z5.h, z2.h\n"
       "fmla z25.h, p0/M, z4.h, z2.h\n"
       "fmla z26.h, p0/M, z6.h, z1.h\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp
index 09180c8f36..865d011ac1 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp16_mla_8x3VL/generic.cpp
@@ -54,22 +54,22 @@ void sve_interleaved_fp16_mla_8x3VL(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
+      "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
       "mov z10.b, #0x0\n"
-      "ld1h { z2.h }, p0/Z, [x22]\n"
       "mov z11.b, #0x0\n"
+      "ld1h { z2.h }, p0/Z, [x22]\n"
       "mov z12.b, #0x0\n"
-      "ld1h { z3.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
       "mov z13.b, #0x0\n"
+      "ld1h { z3.h }, p0/Z, [x22, #1, MUL VL]\n"
       "mov z14.b, #0x0\n"
       "mov z15.b, #0x0\n"
+      "ld1h { z4.h }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z16.b, #0x0\n"
-      "ld1rqh { z0.h }, p0/Z, [%x[Apanel]]\n"
       "mov z17.b, #0x0\n"
       "mov z18.b, #0x0\n"
-      "ld1h { z4.h }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z19.b, #0x0\n"
       "mov z20.b, #0x0\n"
       "mov z21.b, #0x0\n"
@@ -147,12 +147,12 @@ void sve_interleaved_fp16_mla_8x3VL(
       "fmla z31.h, z1.h, z7.h[7]\n"
       "bge 3b\n"
       "4:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      "addvl x22, x22, #3\n"
       "fmla z8.h, z2.h, z0.h[0]\n"
       "fmla z11.h, z2.h, z0.h[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       "fmla z14.h, z2.h, z0.h[2]\n"
       "fmla z17.h, z2.h, z0.h[3]\n"
+      "addvl x22, x22, #3\n"
       "fmla z20.h, z2.h, z0.h[4]\n"
       "fmla z23.h, z2.h, z0.h[5]\n"
       "fmla z26.h, z2.h, z0.h[6]\n"
@@ -176,16 +176,16 @@ void sve_interleaved_fp16_mla_8x3VL(
       "cbz x20, 5f\n"
       "ld1rqh { z3.h }, p0/Z, [%x[Apanel]]\n"
       "ld1h { z2.h }, p0/Z, [x22]\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
+      "fmla z8.h, z2.h, z3.h[0]\n"
       "ld1h { z1.h }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1h { z0.h }, p0/Z, [x22, #2, MUL VL]\n"
-      "addvl x22, x22, #3\n"
-      "fmla z8.h, z2.h, z3.h[0]\n"
       "fmla z11.h, z2.h, z3.h[1]\n"
       "fmla z14.h, z2.h, z3.h[2]\n"
       "fmla z17.h, z2.h, z3.h[3]\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       "fmla z20.h, z2.h, z3.h[4]\n"
       "fmla z23.h, z2.h, z3.h[5]\n"
+      "addvl x22, x22, #3\n"
       "fmla z26.h, z2.h, z3.h[6]\n"
       "fmla z29.h, z2.h, z3.h[7]\n"
       "fmla z9.h, z1.h, z3.h[0]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL.hpp
index 89d65083f4..000866346f 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL.hpp
@@ -40,8 +40,7 @@ void sve_interleaved_fp32_mla_8x3VL_a64fx( ARGLIST );
 class cls_sve_interleaved_fp32_mla_8x3VL
 {
 public:
-    typedef float lhs_operand_type;
-    typedef float rhs_operand_type;
+    typedef float operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 1, 1> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/a64fx.cpp
index 0006fddb2a..ee9f58f811 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/a64fx.cpp
@@ -54,31 +54,31 @@ void sve_interleaved_fp32_mla_8x3VL_a64fx(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
-      "mov z10.b, #0x0\n"
       "ld1w { z0.s }, p0/Z, [x22]\n"
+      "mov z10.b, #0x0\n"
       "mov z11.b, #0x0\n"
-      "mov z12.b, #0x0\n"
       "ld1w { z1.s }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
+      "mov z12.b, #0x0\n"
       "mov z13.b, #0x0\n"
+      "ld1w { z2.s }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z14.b, #0x0\n"
       "mov z15.b, #0x0\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
       "mov z16.b, #0x0\n"
-      "ld1w { z2.s }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z17.b, #0x0\n"
+      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "mov z18.b, #0x0\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
       "mov z19.b, #0x0\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "mov z20.b, #0x0\n"
-      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "mov z21.b, #0x0\n"
+      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "mov z22.b, #0x0\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "mov z23.b, #0x0\n"
       "mov z24.b, #0x0\n"
-      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "mov z25.b, #0x0\n"
       "mov z26.b, #0x0\n"
       "mov z27.b, #0x0\n"
@@ -92,7 +92,7 @@ void sve_interleaved_fp32_mla_8x3VL_a64fx(
       "fmla z9.s, p0/M, z1.s, z3.s\n"
       "sub x20, x20, #0x2\n"
       "fmla z10.s, p0/M, z2.s, z3.s\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #16]\n"
       "fmla z11.s, p0/M, z0.s, z4.s\n"
       "fmla z12.s, p0/M, z1.s, z4.s\n"
       "fmla z13.s, p0/M, z2.s, z4.s\n"
@@ -101,63 +101,63 @@ void sve_interleaved_fp32_mla_8x3VL_a64fx(
       "fmla z15.s, p0/M, z1.s, z5.s\n"
       "cmp x20, #0x2\n"
       "fmla z16.s, p0/M, z2.s, z5.s\n"
-      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #24]\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #24]\n"
       "fmla z17.s, p0/M, z0.s, z6.s\n"
       "fmla z18.s, p0/M, z1.s, z6.s\n"
       "fmla z19.s, p0/M, z2.s, z6.s\n"
-      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #28]\n"
-      "fmla z20.s, p0/M, z0.s, z3.s\n"
-      "fmla z21.s, p0/M, z1.s, z3.s\n"
-      "fmla z22.s, p0/M, z2.s, z3.s\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #32]\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #28]\n"
+      "fmla z20.s, p0/M, z0.s, z7.s\n"
+      "fmla z21.s, p0/M, z1.s, z7.s\n"
+      "fmla z22.s, p0/M, z2.s, z7.s\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #32]\n"
       "fmla z23.s, p0/M, z0.s, z4.s\n"
       "fmla z24.s, p0/M, z1.s, z4.s\n"
       "fmla z25.s, p0/M, z2.s, z4.s\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #36]\n"
-      "fmla z26.s, p0/M, z0.s, z7.s\n"
-      "fmla z27.s, p0/M, z1.s, z7.s\n"
-      "fmla z28.s, p0/M, z2.s, z7.s\n"
+      "fmla z26.s, p0/M, z0.s, z3.s\n"
+      "fmla z27.s, p0/M, z1.s, z3.s\n"
+      "fmla z28.s, p0/M, z2.s, z3.s\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel], #40]\n"
-      "fmla z29.s, p0/M, z0.s, z6.s\n"
-      "ld1w { z7.s }, p0/Z, [x22, #3, MUL VL]\n"
-      "fmla z30.s, p0/M, z1.s, z6.s\n"
-      "fmla z31.s, p0/M, z2.s, z6.s\n"
-      "ld1w { z6.s }, p0/Z, [x22, #4, MUL VL]\n"
-      "ld1w { z2.s }, p0/Z, [x22, #5, MUL VL]\n"
-      "addvl x22, x22, #6\n"
+      "fmla z29.s, p0/M, z0.s, z5.s\n"
+      "ld1w { z6.s }, p0/Z, [x22, #3, MUL VL]\n"
+      "fmla z30.s, p0/M, z1.s, z5.s\n"
+      "fmla z31.s, p0/M, z2.s, z5.s\n"
+      "ld1w { z2.s }, p0/Z, [x22, #4, MUL VL]\n"
+      "ld1w { z5.s }, p0/Z, [x22, #5, MUL VL]\n"
+      "fmla z8.s, p0/M, z6.s, z7.s\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #44]\n"
-      "fmla z8.s, p0/M, z7.s, z5.s\n"
-      "fmla z11.s, p0/M, z7.s, z4.s\n"
-      "fmla z9.s, p0/M, z6.s, z5.s\n"
-      "fmla z12.s, p0/M, z6.s, z4.s\n"
-      "fmla z10.s, p0/M, z2.s, z5.s\n"
-      "fmla z13.s, p0/M, z2.s, z4.s\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #48]\n"
-      "fmla z14.s, p0/M, z7.s, z3.s\n"
-      "fmla z15.s, p0/M, z6.s, z3.s\n"
+      "fmla z9.s, p0/M, z2.s, z7.s\n"
+      "fmla z10.s, p0/M, z5.s, z7.s\n"
+      "fmla z11.s, p0/M, z6.s, z4.s\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #48]\n"
+      "fmla z12.s, p0/M, z2.s, z4.s\n"
+      "fmla z13.s, p0/M, z5.s, z4.s\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #52]\n"
-      "fmla z16.s, p0/M, z2.s, z3.s\n"
-      "fmla z17.s, p0/M, z7.s, z1.s\n"
+      "fmla z14.s, p0/M, z6.s, z3.s\n"
+      "fmla z15.s, p0/M, z2.s, z3.s\n"
+      "addvl x22, x22, #6\n"
+      "fmla z16.s, p0/M, z5.s, z3.s\n"
       "ld1rw { z0.s }, p0/Z, [%x[Apanel], #56]\n"
-      "fmla z18.s, p0/M, z6.s, z1.s\n"
-      "fmla z19.s, p0/M, z2.s, z1.s\n"
+      "fmla z17.s, p0/M, z6.s, z1.s\n"
+      "fmla z18.s, p0/M, z2.s, z1.s\n"
+      "fmla z19.s, p0/M, z5.s, z1.s\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #60]\n"
       "add %x[Apanel], %x[Apanel], #0x40\n"
-      "fmla z20.s, p0/M, z7.s, z5.s\n"
-      "fmla z21.s, p0/M, z6.s, z5.s\n"
-      "fmla z22.s, p0/M, z2.s, z5.s\n"
-      "fmla z23.s, p0/M, z7.s, z4.s\n"
+      "fmla z20.s, p0/M, z6.s, z7.s\n"
+      "fmla z21.s, p0/M, z2.s, z7.s\n"
+      "fmla z22.s, p0/M, z5.s, z7.s\n"
+      "fmla z23.s, p0/M, z6.s, z4.s\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
-      "fmla z24.s, p0/M, z6.s, z4.s\n"
-      "fmla z25.s, p0/M, z2.s, z4.s\n"
+      "fmla z24.s, p0/M, z2.s, z4.s\n"
+      "fmla z25.s, p0/M, z5.s, z4.s\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
-      "fmla z26.s, p0/M, z7.s, z0.s\n"
-      "fmla z27.s, p0/M, z6.s, z0.s\n"
-      "fmla z28.s, p0/M, z2.s, z0.s\n"
-      "fmla z29.s, p0/M, z7.s, z1.s\n"
+      "fmla z26.s, p0/M, z6.s, z0.s\n"
+      "fmla z27.s, p0/M, z2.s, z0.s\n"
+      "fmla z28.s, p0/M, z5.s, z0.s\n"
+      "fmla z29.s, p0/M, z6.s, z1.s\n"
       "ld1w { z0.s }, p0/Z, [x22]\n"
-      "fmla z30.s, p0/M, z6.s, z1.s\n"
-      "fmla z31.s, p0/M, z2.s, z1.s\n"
+      "fmla z30.s, p0/M, z2.s, z1.s\n"
+      "fmla z31.s, p0/M, z5.s, z1.s\n"
       "ld1w { z1.s }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1w { z2.s }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
@@ -199,20 +199,19 @@ void sve_interleaved_fp32_mla_8x3VL_a64fx(
       "ld1w { z5.s }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1w { z4.s }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
-      "addvl x22, x22, #3\n"
+      "fmla z8.s, p0/M, z6.s, z3.s\n"
       "ld1rw { z2.s }, p0/Z, [%x[Apanel], #4]\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #8]\n"
-      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
-      "fmla z8.s, p0/M, z6.s, z3.s\n"
       "fmla z9.s, p0/M, z5.s, z3.s\n"
+      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
       "fmla z10.s, p0/M, z4.s, z3.s\n"
       "fmla z11.s, p0/M, z6.s, z2.s\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
       "fmla z12.s, p0/M, z5.s, z2.s\n"
       "fmla z13.s, p0/M, z4.s, z2.s\n"
-      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
       "fmla z14.s, p0/M, z6.s, z1.s\n"
       "fmla z15.s, p0/M, z5.s, z1.s\n"
+      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
       "fmla z16.s, p0/M, z4.s, z1.s\n"
       "fmla z17.s, p0/M, z6.s, z0.s\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #24]\n"
@@ -221,9 +220,10 @@ void sve_interleaved_fp32_mla_8x3VL_a64fx(
       "ld1rw { z0.s }, p0/Z, [%x[Apanel], #28]\n"
       "fmla z20.s, p0/M, z6.s, z3.s\n"
       "fmla z21.s, p0/M, z5.s, z3.s\n"
-      "add %x[Apanel], %x[Apanel], #0x20\n"
+      "addvl x22, x22, #3\n"
       "fmla z22.s, p0/M, z4.s, z3.s\n"
       "fmla z23.s, p0/M, z6.s, z2.s\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "fmla z24.s, p0/M, z5.s, z2.s\n"
       "fmla z25.s, p0/M, z4.s, z2.s\n"
       "fmla z26.s, p0/M, z6.s, z1.s\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp
index 43591e9201..f459fe92c9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL/generic.cpp
@@ -54,25 +54,25 @@ void sve_interleaved_fp32_mla_8x3VL(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.b, #0x0\n"
       "mov z9.b, #0x0\n"
+      "ld1rqw { z0.s }, p0/Z, [%x[Apanel]]\n"
       "mov z10.b, #0x0\n"
-      "ld1w { z4.s }, p0/Z, [x22]\n"
       "mov z11.b, #0x0\n"
+      "ld1rqw { z1.s }, p0/Z, [%x[Apanel], #16]\n"
       "mov z12.b, #0x0\n"
-      "ld1w { z5.s }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
       "mov z13.b, #0x0\n"
+      "ld1w { z4.s }, p0/Z, [x22]\n"
       "mov z14.b, #0x0\n"
       "mov z15.b, #0x0\n"
+      "ld1w { z5.s }, p0/Z, [x22, #1, MUL VL]\n"
       "mov z16.b, #0x0\n"
-      "ld1rqw { z0.s }, p0/Z, [%x[Apanel]]\n"
       "mov z17.b, #0x0\n"
+      "ld1w { z6.s }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z18.b, #0x0\n"
-      "ld1rqw { z1.s }, p0/Z, [%x[Apanel], #16]\n"
       "mov z19.b, #0x0\n"
       "mov z20.b, #0x0\n"
-      "ld1w { z6.s }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z21.b, #0x0\n"
       "mov z22.b, #0x0\n"
       "mov z23.b, #0x0\n"
@@ -150,12 +150,12 @@ void sve_interleaved_fp32_mla_8x3VL(
       "ld1w { z6.s }, p0/Z, [x22, #2, MUL VL]\n"
       "bge 3b\n"
       "4:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "addvl x22, x22, #3\n"
       "fmla z8.s, z4.s, z0.s[0]\n"
       "fmla z11.s, z4.s, z0.s[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "fmla z14.s, z4.s, z0.s[2]\n"
       "fmla z17.s, z4.s, z0.s[3]\n"
+      "addvl x22, x22, #3\n"
       "fmla z20.s, z4.s, z1.s[0]\n"
       "fmla z23.s, z4.s, z1.s[1]\n"
       "fmla z26.s, z4.s, z1.s[2]\n"
@@ -182,13 +182,13 @@ void sve_interleaved_fp32_mla_8x3VL(
       "add %x[Apanel], %x[Apanel], #0x20\n"
       "ld1w { z2.s }, p0/Z, [x22]\n"
       "ld1w { z1.s }, p0/Z, [x22, #1, MUL VL]\n"
-      "ld1w { z0.s }, p0/Z, [x22, #2, MUL VL]\n"
-      "addvl x22, x22, #3\n"
       "fmla z8.s, z2.s, z4.s[0]\n"
+      "ld1w { z0.s }, p0/Z, [x22, #2, MUL VL]\n"
       "fmla z11.s, z2.s, z4.s[1]\n"
       "fmla z14.s, z2.s, z4.s[2]\n"
       "fmla z17.s, z2.s, z4.s[3]\n"
       "fmla z20.s, z2.s, z3.s[0]\n"
+      "addvl x22, x22, #3\n"
       "fmla z23.s, z2.s, z3.s[1]\n"
       "fmla z26.s, z2.s, z3.s[2]\n"
       "fmla z29.s, z2.s, z3.s[3]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mmla_8x3VL.hpp
index 3a0e7f4c20..71e8551b92 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mmla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mmla_8x3VL.hpp
@@ -35,8 +35,7 @@ void sve_interleaved_fp32_mmla_8x3VL(const float *, const float *, float *, int,
 
 class cls_sve_interleaved_fp32_mmla_8x3VL {
 public:
-    typedef float lhs_operand_type;
-    typedef float rhs_operand_type;
+    typedef float operand_type;
     typedef float result_type;
 
     typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
@@ -58,7 +57,7 @@ public:
     }
 
     // Use the standard fixed size transforms.
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 2, 2> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 2, 2> transforms = {};
 
     kern_type kernel=sve_interleaved_fp32_mmla_8x3VL;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL.hpp
index ac731b76ed..ed1faeccd9 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL.hpp
@@ -40,8 +40,7 @@ void sve_interleaved_s8s32_dot_8x3VL_a64fx( ARGLIST );
 class cls_sve_interleaved_s8s32_dot_8x3VL
 {
 public:
-    typedef int8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
+    typedef int8_t operand_type;
     typedef int32_t result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 4, 1> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 4, 1, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 4, 1> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 4, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/a64fx.cpp
index 1b33014f36..0159fe4923 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/a64fx.cpp
@@ -55,31 +55,31 @@ void sve_interleaved_s8s32_dot_8x3VL_a64fx(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
       "ld1b { z0.b }, p0/Z, [x22]\n"
+      "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
       "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
+      "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
+      "ld1b { z2.b }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
       "mov z16.s, #0x0\n"
-      "ld1b { z2.b }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z17.s, #0x0\n"
+      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "mov z18.s, #0x0\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
       "mov z19.s, #0x0\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "mov z20.s, #0x0\n"
-      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "mov z21.s, #0x0\n"
+      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "mov z22.s, #0x0\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
-      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "mov z25.s, #0x0\n"
       "mov z26.s, #0x0\n"
       "mov z27.s, #0x0\n"
@@ -93,7 +93,7 @@ void sve_interleaved_s8s32_dot_8x3VL_a64fx(
       "sdot z9.s, z1.b, z3.b\n"
       "sub x20, x20, #0x2\n"
       "sdot z10.s, z2.b, z3.b\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #16]\n"
       "sdot z11.s, z0.b, z4.b\n"
       "sdot z12.s, z1.b, z4.b\n"
       "sdot z13.s, z2.b, z4.b\n"
@@ -102,63 +102,63 @@ void sve_interleaved_s8s32_dot_8x3VL_a64fx(
       "sdot z15.s, z1.b, z5.b\n"
       "cmp x20, #0x2\n"
       "sdot z16.s, z2.b, z5.b\n"
-      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #24]\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #24]\n"
       "sdot z17.s, z0.b, z6.b\n"
       "sdot z18.s, z1.b, z6.b\n"
       "sdot z19.s, z2.b, z6.b\n"
-      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #28]\n"
-      "sdot z20.s, z0.b, z3.b\n"
-      "sdot z21.s, z1.b, z3.b\n"
-      "sdot z22.s, z2.b, z3.b\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #32]\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #28]\n"
+      "sdot z20.s, z0.b, z7.b\n"
+      "sdot z21.s, z1.b, z7.b\n"
+      "sdot z22.s, z2.b, z7.b\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #32]\n"
       "sdot z23.s, z0.b, z4.b\n"
       "sdot z24.s, z1.b, z4.b\n"
       "sdot z25.s, z2.b, z4.b\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #36]\n"
-      "sdot z26.s, z0.b, z7.b\n"
-      "sdot z27.s, z1.b, z7.b\n"
-      "sdot z28.s, z2.b, z7.b\n"
+      "sdot z26.s, z0.b, z3.b\n"
+      "sdot z27.s, z1.b, z3.b\n"
+      "sdot z28.s, z2.b, z3.b\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel], #40]\n"
-      "sdot z29.s, z0.b, z6.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #3, MUL VL]\n"
-      "sdot z30.s, z1.b, z6.b\n"
-      "sdot z31.s, z2.b, z6.b\n"
-      "ld1b { z6.b }, p0/Z, [x22, #4, MUL VL]\n"
-      "ld1b { z2.b }, p0/Z, [x22, #5, MUL VL]\n"
-      "addvl x22, x22, #6\n"
+      "sdot z29.s, z0.b, z5.b\n"
+      "ld1b { z6.b }, p0/Z, [x22, #3, MUL VL]\n"
+      "sdot z30.s, z1.b, z5.b\n"
+      "sdot z31.s, z2.b, z5.b\n"
+      "ld1b { z2.b }, p0/Z, [x22, #4, MUL VL]\n"
+      "ld1b { z5.b }, p0/Z, [x22, #5, MUL VL]\n"
+      "sdot z8.s, z6.b, z7.b\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #44]\n"
-      "sdot z8.s, z7.b, z5.b\n"
-      "sdot z11.s, z7.b, z4.b\n"
-      "sdot z9.s, z6.b, z5.b\n"
-      "sdot z12.s, z6.b, z4.b\n"
-      "sdot z10.s, z2.b, z5.b\n"
-      "sdot z13.s, z2.b, z4.b\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #48]\n"
-      "sdot z14.s, z7.b, z3.b\n"
-      "sdot z15.s, z6.b, z3.b\n"
+      "sdot z9.s, z2.b, z7.b\n"
+      "sdot z10.s, z5.b, z7.b\n"
+      "sdot z11.s, z6.b, z4.b\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #48]\n"
+      "sdot z12.s, z2.b, z4.b\n"
+      "sdot z13.s, z5.b, z4.b\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #52]\n"
-      "sdot z16.s, z2.b, z3.b\n"
-      "sdot z17.s, z7.b, z1.b\n"
+      "sdot z14.s, z6.b, z3.b\n"
+      "sdot z15.s, z2.b, z3.b\n"
+      "addvl x22, x22, #6\n"
+      "sdot z16.s, z5.b, z3.b\n"
       "ld1rw { z0.s }, p0/Z, [%x[Apanel], #56]\n"
-      "sdot z18.s, z6.b, z1.b\n"
-      "sdot z19.s, z2.b, z1.b\n"
+      "sdot z17.s, z6.b, z1.b\n"
+      "sdot z18.s, z2.b, z1.b\n"
+      "sdot z19.s, z5.b, z1.b\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #60]\n"
       "add %x[Apanel], %x[Apanel], #0x40\n"
-      "sdot z20.s, z7.b, z5.b\n"
-      "sdot z21.s, z6.b, z5.b\n"
-      "sdot z22.s, z2.b, z5.b\n"
-      "sdot z23.s, z7.b, z4.b\n"
+      "sdot z20.s, z6.b, z7.b\n"
+      "sdot z21.s, z2.b, z7.b\n"
+      "sdot z22.s, z5.b, z7.b\n"
+      "sdot z23.s, z6.b, z4.b\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
-      "sdot z24.s, z6.b, z4.b\n"
-      "sdot z25.s, z2.b, z4.b\n"
+      "sdot z24.s, z2.b, z4.b\n"
+      "sdot z25.s, z5.b, z4.b\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
-      "sdot z26.s, z7.b, z0.b\n"
-      "sdot z27.s, z6.b, z0.b\n"
-      "sdot z28.s, z2.b, z0.b\n"
-      "sdot z29.s, z7.b, z1.b\n"
+      "sdot z26.s, z6.b, z0.b\n"
+      "sdot z27.s, z2.b, z0.b\n"
+      "sdot z28.s, z5.b, z0.b\n"
+      "sdot z29.s, z6.b, z1.b\n"
       "ld1b { z0.b }, p0/Z, [x22]\n"
-      "sdot z30.s, z6.b, z1.b\n"
-      "sdot z31.s, z2.b, z1.b\n"
+      "sdot z30.s, z2.b, z1.b\n"
+      "sdot z31.s, z5.b, z1.b\n"
       "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1b { z2.b }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
@@ -200,20 +200,19 @@ void sve_interleaved_s8s32_dot_8x3VL_a64fx(
       "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1b { z4.b }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
-      "addvl x22, x22, #3\n"
+      "sdot z8.s, z6.b, z3.b\n"
       "ld1rw { z2.s }, p0/Z, [%x[Apanel], #4]\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #8]\n"
-      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
-      "sdot z8.s, z6.b, z3.b\n"
       "sdot z9.s, z5.b, z3.b\n"
+      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
       "sdot z10.s, z4.b, z3.b\n"
       "sdot z11.s, z6.b, z2.b\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
       "sdot z12.s, z5.b, z2.b\n"
       "sdot z13.s, z4.b, z2.b\n"
-      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
       "sdot z14.s, z6.b, z1.b\n"
       "sdot z15.s, z5.b, z1.b\n"
+      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
       "sdot z16.s, z4.b, z1.b\n"
       "sdot z17.s, z6.b, z0.b\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #24]\n"
@@ -222,9 +221,10 @@ void sve_interleaved_s8s32_dot_8x3VL_a64fx(
       "ld1rw { z0.s }, p0/Z, [%x[Apanel], #28]\n"
       "sdot z20.s, z6.b, z3.b\n"
       "sdot z21.s, z5.b, z3.b\n"
-      "add %x[Apanel], %x[Apanel], #0x20\n"
+      "addvl x22, x22, #3\n"
       "sdot z22.s, z4.b, z3.b\n"
       "sdot z23.s, z6.b, z2.b\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "sdot z24.s, z5.b, z2.b\n"
       "sdot z25.s, z4.b, z2.b\n"
       "sdot z26.s, z6.b, z1.b\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp
index 1ddf171c7e..88a086fac7 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_8x3VL/generic.cpp
@@ -55,25 +55,25 @@ void sve_interleaved_s8s32_dot_8x3VL(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
+      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
       "mov z10.s, #0x0\n"
-      "ld1b { z4.b }, p0/Z, [x22]\n"
       "mov z11.s, #0x0\n"
+      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
       "mov z12.s, #0x0\n"
-      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
       "mov z13.s, #0x0\n"
+      "ld1b { z4.b }, p0/Z, [x22]\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
+      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
       "mov z16.s, #0x0\n"
-      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
       "mov z17.s, #0x0\n"
+      "ld1b { z6.b }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z18.s, #0x0\n"
-      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
-      "ld1b { z6.b }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z21.s, #0x0\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
@@ -151,12 +151,12 @@ void sve_interleaved_s8s32_dot_8x3VL(
       "ld1b { z6.b }, p0/Z, [x22, #2, MUL VL]\n"
       "bge 3b\n"
       "4:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "addvl x22, x22, #3\n"
       "sdot z8.s, z4.b, z0.b[0]\n"
       "sdot z11.s, z4.b, z0.b[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "sdot z14.s, z4.b, z0.b[2]\n"
       "sdot z17.s, z4.b, z0.b[3]\n"
+      "addvl x22, x22, #3\n"
       "sdot z20.s, z4.b, z1.b[0]\n"
       "sdot z23.s, z4.b, z1.b[1]\n"
       "sdot z26.s, z4.b, z1.b[2]\n"
@@ -183,13 +183,13 @@ void sve_interleaved_s8s32_dot_8x3VL(
       "add %x[Apanel], %x[Apanel], #0x20\n"
       "ld1b { z2.b }, p0/Z, [x22]\n"
       "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "ld1b { z0.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "addvl x22, x22, #3\n"
       "sdot z8.s, z2.b, z4.b[0]\n"
+      "ld1b { z0.b }, p0/Z, [x22, #2, MUL VL]\n"
       "sdot z11.s, z2.b, z4.b[1]\n"
       "sdot z14.s, z2.b, z4.b[2]\n"
       "sdot z17.s, z2.b, z4.b[3]\n"
       "sdot z20.s, z2.b, z3.b[0]\n"
+      "addvl x22, x22, #3\n"
       "sdot z23.s, z2.b, z3.b[1]\n"
       "sdot z26.s, z2.b, z3.b[2]\n"
       "sdot z29.s, z2.b, z3.b[3]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL.hpp
index 5ba3e51e6f..d86943c9b2 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL.hpp
@@ -39,8 +39,7 @@ void sve_interleaved_s8s32_mmla_8x3VL( ARGLIST );
 class cls_sve_interleaved_s8s32_mmla_8x3VL
 {
 public:
-    typedef int8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
+    typedef int8_t operand_type;
     typedef int32_t result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -62,8 +61,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 8, 2> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 8, 2, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 8, 2> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 8, 2, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL/generic.cpp
index 261648eebe..afc8038956 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_mmla_8x3VL/generic.cpp
@@ -55,31 +55,31 @@ void sve_interleaved_s8s32_mmla_8x3VL(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
       "ld1b { z4.b }, p0/Z, [x22]\n"
+      "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
+      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
       "mov z12.s, #0x0\n"
-      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
       "mov z13.s, #0x0\n"
+      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
+      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
       "mov z16.s, #0x0\n"
-      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
       "mov z17.s, #0x0\n"
+      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #32]\n"
       "mov z18.s, #0x0\n"
-      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
       "mov z19.s, #0x0\n"
+      "addvl x22, x22, #2\n"
       "mov z20.s, #0x0\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #32]\n"
       "mov z21.s, #0x0\n"
+      "add %x[Apanel], %x[Apanel], #0x30\n"
       "mov z22.s, #0x0\n"
-      "addvl x22, x22, #2\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
       "mov z25.s, #0x0\n"
       "mov z26.s, #0x0\n"
       "mov z27.s, #0x0\n"
@@ -94,77 +94,77 @@ void sve_interleaved_s8s32_mmla_8x3VL(
       ".inst 0x4505980b  // smmla z11.s, z0.b, z5.b\n"
       ".inst 0x4504982e  // smmla z14.s, z1.b, z4.b\n"
       ".inst 0x45059831  // smmla z17.s, z1.b, z5.b\n"
-      "ld1b { z3.b }, p0/Z, [x22]\n"
+      "ld1b { z7.b }, p0/Z, [x22]\n"
       ".inst 0x45049854  // smmla z20.s, z2.b, z4.b\n"
       ".inst 0x45059857  // smmla z23.s, z2.b, z5.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #1, MUL VL]\n"
+      "ld1b { z3.b }, p0/Z, [x22, #1, MUL VL]\n"
       ".inst 0x450498da  // smmla z26.s, z6.b, z4.b\n"
       ".inst 0x450598dd  // smmla z29.s, z6.b, z5.b\n"
-      "ld1b { z4.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1b { z5.b }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x45039809  // smmla z9.s, z0.b, z3.b\n"
+      "ld1b { z5.b }, p0/Z, [x22, #2, MUL VL]\n"
+      "ld1b { z4.b }, p0/Z, [x22, #3, MUL VL]\n"
+      ".inst 0x45079809  // smmla z9.s, z0.b, z7.b\n"
+      ".inst 0x4503980c  // smmla z12.s, z0.b, z3.b\n"
+      ".inst 0x4507982f  // smmla z15.s, z1.b, z7.b\n"
+      ".inst 0x45039832  // smmla z18.s, z1.b, z3.b\n"
       "sub x20, x20, #0x2\n"
-      ".inst 0x4507980c  // smmla z12.s, z0.b, z7.b\n"
-      ".inst 0x4503982f  // smmla z15.s, z1.b, z3.b\n"
+      ".inst 0x45079855  // smmla z21.s, z2.b, z7.b\n"
+      ".inst 0x45039858  // smmla z24.s, z2.b, z3.b\n"
       "cmp x20, #0x2\n"
-      ".inst 0x45079832  // smmla z18.s, z1.b, z7.b\n"
-      ".inst 0x45039855  // smmla z21.s, z2.b, z3.b\n"
-      ".inst 0x45079858  // smmla z24.s, z2.b, z7.b\n"
-      ".inst 0x450398db  // smmla z27.s, z6.b, z3.b\n"
+      ".inst 0x450798db  // smmla z27.s, z6.b, z7.b\n"
+      ".inst 0x450398de  // smmla z30.s, z6.b, z3.b\n"
       "ld1b { z3.b }, p0/Z, [x22, #4, MUL VL]\n"
-      ".inst 0x450798de  // smmla z30.s, z6.b, z7.b\n"
-      ".inst 0x4504980a  // smmla z10.s, z0.b, z4.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #5, MUL VL]\n"
-      ".inst 0x4505980d  // smmla z13.s, z0.b, z5.b\n"
-      ".inst 0x45049830  // smmla z16.s, z1.b, z4.b\n"
+      ".inst 0x4505980a  // smmla z10.s, z0.b, z5.b\n"
+      ".inst 0x4504980d  // smmla z13.s, z0.b, z4.b\n"
       "ld1rqb { z0.b }, p0/Z, [%x[Apanel], #16]\n"
-      ".inst 0x45059833  // smmla z19.s, z1.b, z5.b\n"
-      ".inst 0x45049856  // smmla z22.s, z2.b, z4.b\n"
+      ".inst 0x45059830  // smmla z16.s, z1.b, z5.b\n"
+      ".inst 0x45049833  // smmla z19.s, z1.b, z4.b\n"
       "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #32]\n"
-      ".inst 0x45059859  // smmla z25.s, z2.b, z5.b\n"
-      ".inst 0x450498dc  // smmla z28.s, z6.b, z4.b\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #48]\n"
-      ".inst 0x450598df  // smmla z31.s, z6.b, z5.b\n"
+      ".inst 0x45059856  // smmla z22.s, z2.b, z5.b\n"
+      ".inst 0x45049859  // smmla z25.s, z2.b, z4.b\n"
+      "ld1b { z7.b }, p0/Z, [x22, #5, MUL VL]\n"
+      ".inst 0x450598dc  // smmla z28.s, z6.b, z5.b\n"
+      ".inst 0x450498df  // smmla z31.s, z6.b, z4.b\n"
+      "ld1rqb { z5.b }, p0/Z, [%x[Apanel], #48]\n"
       "ld1rqb { z6.b }, p0/Z, [%x[Apanel], #64]\n"
-      "ld1b { z4.b }, p0/Z, [x22, #6, MUL VL]\n"
-      "ld1b { z5.b }, p0/Z, [x22, #7, MUL VL]\n"
-      "addvl x22, x22, #16\n"
+      "ld1b { z2.b }, p0/Z, [x22, #6, MUL VL]\n"
       ".inst 0x45039808  // smmla z8.s, z0.b, z3.b\n"
+      "ld1b { z4.b }, p0/Z, [x22, #7, MUL VL]\n"
+      "addvl x22, x22, #16\n"
       ".inst 0x4507980b  // smmla z11.s, z0.b, z7.b\n"
       ".inst 0x4503982e  // smmla z14.s, z1.b, z3.b\n"
       ".inst 0x45079831  // smmla z17.s, z1.b, z7.b\n"
-      ".inst 0x45039854  // smmla z20.s, z2.b, z3.b\n"
-      ".inst 0x45079857  // smmla z23.s, z2.b, z7.b\n"
+      ".inst 0x450398b4  // smmla z20.s, z5.b, z3.b\n"
+      ".inst 0x450798b7  // smmla z23.s, z5.b, z7.b\n"
       ".inst 0x450398da  // smmla z26.s, z6.b, z3.b\n"
-      "ld1b { z3.b }, p0/Z, [x22, #-8, MUL VL]\n"
       ".inst 0x450798dd  // smmla z29.s, z6.b, z7.b\n"
+      "ld1b { z3.b }, p0/Z, [x22, #-8, MUL VL]\n"
       "ld1b { z7.b }, p0/Z, [x22, #-7, MUL VL]\n"
-      ".inst 0x45049809  // smmla z9.s, z0.b, z4.b\n"
-      ".inst 0x4505980c  // smmla z12.s, z0.b, z5.b\n"
-      ".inst 0x4504982f  // smmla z15.s, z1.b, z4.b\n"
-      ".inst 0x45059832  // smmla z18.s, z1.b, z5.b\n"
-      ".inst 0x45049855  // smmla z21.s, z2.b, z4.b\n"
-      ".inst 0x45059858  // smmla z24.s, z2.b, z5.b\n"
-      ".inst 0x450498db  // smmla z27.s, z6.b, z4.b\n"
+      ".inst 0x45029809  // smmla z9.s, z0.b, z2.b\n"
+      ".inst 0x4504980c  // smmla z12.s, z0.b, z4.b\n"
+      ".inst 0x4502982f  // smmla z15.s, z1.b, z2.b\n"
+      ".inst 0x45049832  // smmla z18.s, z1.b, z4.b\n"
+      ".inst 0x450298b5  // smmla z21.s, z5.b, z2.b\n"
+      ".inst 0x450498b8  // smmla z24.s, z5.b, z4.b\n"
+      ".inst 0x450298db  // smmla z27.s, z6.b, z2.b\n"
+      ".inst 0x450498de  // smmla z30.s, z6.b, z4.b\n"
       "ld1b { z4.b }, p0/Z, [x22, #-6, MUL VL]\n"
-      ".inst 0x450598de  // smmla z30.s, z6.b, z5.b\n"
       ".inst 0x4503980a  // smmla z10.s, z0.b, z3.b\n"
-      "ld1b { z5.b }, p0/Z, [x22, #-5, MUL VL]\n"
       ".inst 0x4507980d  // smmla z13.s, z0.b, z7.b\n"
-      ".inst 0x45039830  // smmla z16.s, z1.b, z3.b\n"
       "ld1rqb { z0.b }, p0/Z, [%x[Apanel], #80]\n"
+      ".inst 0x45039830  // smmla z16.s, z1.b, z3.b\n"
       ".inst 0x45079833  // smmla z19.s, z1.b, z7.b\n"
-      ".inst 0x45039856  // smmla z22.s, z2.b, z3.b\n"
       "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #96]\n"
-      ".inst 0x45079859  // smmla z25.s, z2.b, z7.b\n"
+      ".inst 0x450398b6  // smmla z22.s, z5.b, z3.b\n"
+      ".inst 0x450798b9  // smmla z25.s, z5.b, z7.b\n"
+      "ld1b { z5.b }, p0/Z, [x22, #-5, MUL VL]\n"
       ".inst 0x450398dc  // smmla z28.s, z6.b, z3.b\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #112]\n"
       ".inst 0x450798df  // smmla z31.s, z6.b, z7.b\n"
+      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #112]\n"
       "add %x[Apanel], %x[Apanel], #0x80\n"
       "addvl x22, x22, #-4\n"
       "bge 3b\n"
       "4:"  // main loop skip
-      "ld1rqb { z3.b }, p0/Z, [%x[Apanel]]\n"
+      "ld1rqb { z7.b }, p0/Z, [%x[Apanel]]\n"
       ".inst 0x45049808  // smmla z8.s, z0.b, z4.b\n"
       ".inst 0x4505980b  // smmla z11.s, z0.b, z5.b\n"
       ".inst 0x4504982e  // smmla z14.s, z1.b, z4.b\n"
@@ -172,54 +172,54 @@ void sve_interleaved_s8s32_mmla_8x3VL(
       "ld1b { z6.b }, p0/Z, [x22]\n"
       ".inst 0x45049854  // smmla z20.s, z2.b, z4.b\n"
       ".inst 0x45059857  // smmla z23.s, z2.b, z5.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x4504987a  // smmla z26.s, z3.b, z4.b\n"
-      ".inst 0x4505987d  // smmla z29.s, z3.b, z5.b\n"
+      "ld1b { z3.b }, p0/Z, [x22, #1, MUL VL]\n"
+      ".inst 0x450498fa  // smmla z26.s, z7.b, z4.b\n"
+      ".inst 0x450598fd  // smmla z29.s, z7.b, z5.b\n"
       "ld1b { z5.b }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1b { z4.b }, p0/Z, [x22, #3, MUL VL]\n"
       ".inst 0x45069809  // smmla z9.s, z0.b, z6.b\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      ".inst 0x4507980c  // smmla z12.s, z0.b, z7.b\n"
+      ".inst 0x4503980c  // smmla z12.s, z0.b, z3.b\n"
       ".inst 0x4506982f  // smmla z15.s, z1.b, z6.b\n"
-      "addvl x22, x22, #4\n"
-      ".inst 0x45079832  // smmla z18.s, z1.b, z7.b\n"
+      ".inst 0x45039832  // smmla z18.s, z1.b, z3.b\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       ".inst 0x45069855  // smmla z21.s, z2.b, z6.b\n"
-      ".inst 0x45079858  // smmla z24.s, z2.b, z7.b\n"
-      ".inst 0x4506987b  // smmla z27.s, z3.b, z6.b\n"
-      ".inst 0x4507987e  // smmla z30.s, z3.b, z7.b\n"
+      ".inst 0x45039858  // smmla z24.s, z2.b, z3.b\n"
+      "addvl x22, x22, #4\n"
+      ".inst 0x450698fb  // smmla z27.s, z7.b, z6.b\n"
+      ".inst 0x450398fe  // smmla z30.s, z7.b, z3.b\n"
       ".inst 0x4505980a  // smmla z10.s, z0.b, z5.b\n"
       ".inst 0x4504980d  // smmla z13.s, z0.b, z4.b\n"
       ".inst 0x45059830  // smmla z16.s, z1.b, z5.b\n"
       ".inst 0x45049833  // smmla z19.s, z1.b, z4.b\n"
       ".inst 0x45059856  // smmla z22.s, z2.b, z5.b\n"
       ".inst 0x45049859  // smmla z25.s, z2.b, z4.b\n"
-      ".inst 0x4505987c  // smmla z28.s, z3.b, z5.b\n"
-      ".inst 0x4504987f  // smmla z31.s, z3.b, z4.b\n"
+      ".inst 0x450598fc  // smmla z28.s, z7.b, z5.b\n"
+      ".inst 0x450498ff  // smmla z31.s, z7.b, z4.b\n"
       "cbz x20, 5f\n"
       "ld1b { z1.b }, p0/Z, [x22]\n"
       "ld1rqb { z7.b }, p0/Z, [%x[Apanel]]\n"
+      ".inst 0x450198e8  // smmla z8.s, z7.b, z1.b\n"
       "ld1rqb { z6.b }, p0/Z, [%x[Apanel], #16]\n"
       "ld1b { z0.b }, p0/Z, [x22, #1, MUL VL]\n"
+      ".inst 0x450098eb  // smmla z11.s, z7.b, z0.b\n"
       "ld1rqb { z5.b }, p0/Z, [%x[Apanel], #32]\n"
       "ld1rqb { z4.b }, p0/Z, [%x[Apanel], #48]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
-      ".inst 0x450198e8  // smmla z8.s, z7.b, z1.b\n"
-      "ld1b { z3.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1b { z2.b }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x450098eb  // smmla z11.s, z7.b, z0.b\n"
       ".inst 0x450198ce  // smmla z14.s, z6.b, z1.b\n"
       ".inst 0x450098d1  // smmla z17.s, z6.b, z0.b\n"
       ".inst 0x450198b4  // smmla z20.s, z5.b, z1.b\n"
+      "ld1b { z3.b }, p0/Z, [x22, #2, MUL VL]\n"
       ".inst 0x450098b7  // smmla z23.s, z5.b, z0.b\n"
       ".inst 0x4501989a  // smmla z26.s, z4.b, z1.b\n"
-      "ld1b { z1.b }, p0/Z, [x22, #4, MUL VL]\n"
+      "ld1b { z2.b }, p0/Z, [x22, #3, MUL VL]\n"
       ".inst 0x4500989d  // smmla z29.s, z4.b, z0.b\n"
+      "ld1b { z1.b }, p0/Z, [x22, #4, MUL VL]\n"
       "ld1b { z0.b }, p0/Z, [x22, #5, MUL VL]\n"
       ".inst 0x450398e9  // smmla z9.s, z7.b, z3.b\n"
       ".inst 0x450298ec  // smmla z12.s, z7.b, z2.b\n"
-      ".inst 0x450398cf  // smmla z15.s, z6.b, z3.b\n"
       "addvl x22, x22, #6\n"
+      ".inst 0x450398cf  // smmla z15.s, z6.b, z3.b\n"
       ".inst 0x450298d2  // smmla z18.s, z6.b, z2.b\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
       ".inst 0x450398b5  // smmla z21.s, z5.b, z3.b\n"
       ".inst 0x450298b8  // smmla z24.s, z5.b, z2.b\n"
       ".inst 0x4503989b  // smmla z27.s, z4.b, z3.b\n"
@@ -233,53 +233,53 @@ void sve_interleaved_s8s32_mmla_8x3VL(
       ".inst 0x4501989c  // smmla z28.s, z4.b, z1.b\n"
       ".inst 0x4500989f  // smmla z31.s, z4.b, z0.b\n"
       "5:"  // multiply loop done
-      "uzp1 z2.d, z8.d, z11.d\n"
+      "uzp1 z0.d, z8.d, z11.d\n"
       "uzp2 z8.d, z8.d, z11.d\n"
-      "subs x23, x23, #0x1\n"
-      "uzp1 z1.d, z9.d, z12.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel]]\n"
+      "uzp1 z0.d, z9.d, z12.d\n"
       "uzp2 z9.d, z9.d, z12.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
       "uzp1 z0.d, z10.d, z13.d\n"
       "uzp2 z10.d, z10.d, z13.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel]]\n"
-      "uzp1 z3.d, z14.d, z17.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+      "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+      "uzp1 z0.d, z14.d, z17.d\n"
       "uzp2 z14.d, z14.d, z17.d\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "uzp1 z17.d, z15.d, z18.d\n"
+      "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+      "uzp1 z1.d, z15.d, z18.d\n"
+      "subs x23, x23, #0x1\n"
+      "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "uzp2 z15.d, z15.d, z18.d\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "uzp1 z2.d, z16.d, z19.d\n"
+      "uzp1 z17.d, z16.d, z19.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
       "uzp2 z16.d, z16.d, z19.d\n"
-      "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "uzp1 z1.d, z20.d, z23.d\n"
+      "uzp1 z0.d, z20.d, z23.d\n"
+      "st1w { z1.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
+      "addvl %x[Cpanel], %x[Cpanel], #16\n"
       "uzp2 z20.d, z20.d, z23.d\n"
-      "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
-      "uzp1 z0.d, z21.d, z24.d\n"
+      "st1w { z17.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
+      "uzp1 z23.d, z21.d, z24.d\n"
       "uzp2 z21.d, z21.d, z24.d\n"
-      "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
-      "uzp1 z23.d, z22.d, z25.d\n"
+      "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
+      "uzp1 z19.d, z22.d, z25.d\n"
       "uzp2 z22.d, z22.d, z25.d\n"
-      "st1w { z3.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
-      "uzp1 z19.d, z26.d, z29.d\n"
+      "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
+      "uzp1 z18.d, z26.d, z29.d\n"
       "uzp2 z26.d, z26.d, z29.d\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
-      "addvl %x[Cpanel], %x[Cpanel], #16\n"
-      "uzp1 z18.d, z27.d, z30.d\n"
+      "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
+      "uzp1 z17.d, z27.d, z30.d\n"
       "uzp2 z27.d, z27.d, z30.d\n"
-      "uzp1 z17.d, z28.d, z31.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
+      "uzp1 z16.d, z28.d, z31.d\n"
       "uzp2 z28.d, z28.d, z31.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
-      "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
-      "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
-      "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
-      "st1w { z23.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
+      "st1w { z23.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
+      "st1w { z19.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
       "st1w { z20.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
       "st1w { z21.s }, p0, [%x[Cpanel]]\n"
       "st1w { z22.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "st1w { z19.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "st1w { z18.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+      "st1w { z18.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+      "st1w { z17.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+      "st1w { z16.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
       "st1w { z26.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "st1w { z27.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
       "st1w { z28.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8s8s32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8s8s32_mmla_8x3VL.hpp
deleted file mode 100644
index 072ffee1cc..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8s8s32_mmla_8x3VL.hpp
+++ /dev/null
@@ -1,107 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#pragma once
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include "../std_transforms_sve.hpp"
-#include "../performance_parameters.hpp"
-
-#define ARGLIST  \
-    const uint8_t *, const int8_t *, \
-    int32_t *, int, int, int
-
-namespace arm_gemm
-{
-// Actual kernel implementations
-void sve_interleaved_u8s8s32_mmla_8x3VL( ARGLIST );
-
-class cls_sve_interleaved_u8s8s32_mmla_8x3VL
-{
-public:
-    typedef uint8_t lhs_operand_type;
-    typedef int8_t rhs_operand_type;
-    typedef int32_t result_type;
-
-    typedef void (*kern_type)( ARGLIST );
-
-    /* Kernel blocking parameters */
-    static constexpr unsigned int out_height()
-    {
-        return 8;
-    }
-
-    static unsigned int out_width()
-    {
-        return get_vector_length<int32_t>() * 3;
-    }
-
-    static constexpr unsigned int k_unroll()
-    {
-        return 8;
-    }
-
-
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 8, 2> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 8, 2, true> transforms_quantized = {};
-    template<typename T>
-    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
-    {
-
-        if (std::is_same<T, uint32_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 61.97, 4.11, 7.93 };
-                case CPUModel::A510:
-                    return { 43.18, 3.57, 2.89 };
-                case CPUModel::V1:
-                    return { 123.47, 5.03, 11.76 };
-            }
-        }
-
-
-        if (std::is_same<T, uint8_t>::value) {
-            switch (ci->get_cpu_model()) {
-                default:
-                    return { 62.00, 4.08, 0.51 };
-                case CPUModel::A510:
-                    return { 38.02, 1.85, 0.28 };
-                case CPUModel::V1:
-                    return { 95.28, 7.99, 0.79 };
-            }
-        }
-
-        return { 1.0 };
-    }
-
-    // Default to the generic kernel
-    kern_type kernel=sve_interleaved_u8s8s32_mmla_8x3VL;
-    cls_sve_interleaved_u8s8s32_mmla_8x3VL(const CPUInfo *)
-    {
-    }
-};
-
-} // namespace arm_gemm
-
-#undef ARGLIST
-#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8s8s32_mmla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8s8s32_mmla_8x3VL/generic.cpp
deleted file mode 100644
index 50a1713b89..0000000000
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8s8s32_mmla_8x3VL/generic.cpp
+++ /dev/null
@@ -1,297 +0,0 @@
-/*
- * Copyright (c) 2024 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifdef ARM_COMPUTE_ENABLE_SVE
-
-#include <cstddef>
-#include <cstdint>
-
-namespace arm_gemm {
-
-void sve_interleaved_u8s8s32_mmla_8x3VL(
-    const uint8_t *Apanel,
-    const int8_t *Bpanel,
-    int32_t *Cpanel,
-    int ablocks,
-    int bblocks,
-    int K) {
-
-    struct KernelArgs {
-        size_t K = {};
-        const int8_t *Bpanel = {};
-        size_t bblocks = {};
-    } ka;
-
-    ka.K = (K/8) - 1;
-    ka.Bpanel = Bpanel;
-    ka.bblocks = bblocks;
-
-    __asm__ __volatile__(
-      "ptrue p0.b\n"
-      "1:"  // Height loop
-      "ldr x23, [%x[args_ptr], %[offsetof_bblocks]]\n"
-      "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n"
-      "mov x21, %x[Apanel]\n"
-      "2:"  // Width loop
-      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
-      "mov %x[Apanel], x21\n"
-      "mov z8.s, #0x0\n"
-      "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
-      "ld1b { z4.b }, p0/Z, [x22]\n"
-      "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
-      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
-      "mov z13.s, #0x0\n"
-      "mov z14.s, #0x0\n"
-      "mov z15.s, #0x0\n"
-      "mov z16.s, #0x0\n"
-      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
-      "mov z17.s, #0x0\n"
-      "mov z18.s, #0x0\n"
-      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
-      "mov z19.s, #0x0\n"
-      "mov z20.s, #0x0\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #32]\n"
-      "mov z21.s, #0x0\n"
-      "mov z22.s, #0x0\n"
-      "addvl x22, x22, #2\n"
-      "mov z23.s, #0x0\n"
-      "mov z24.s, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
-      "mov z25.s, #0x0\n"
-      "mov z26.s, #0x0\n"
-      "mov z27.s, #0x0\n"
-      "mov z28.s, #0x0\n"
-      "mov z29.s, #0x0\n"
-      "mov z30.s, #0x0\n"
-      "mov z31.s, #0x0\n"
-      "blt 4f\n"
-      "3:"  // main loop head
-      "ld1rqb { z6.b }, p0/Z, [%x[Apanel]]\n"
-      ".inst 0x45849808  // usmmla z8.s, z0.b, z4.b\n"
-      ".inst 0x4585980b  // usmmla z11.s, z0.b, z5.b\n"
-      ".inst 0x4584982e  // usmmla z14.s, z1.b, z4.b\n"
-      ".inst 0x45859831  // usmmla z17.s, z1.b, z5.b\n"
-      "ld1b { z3.b }, p0/Z, [x22]\n"
-      ".inst 0x45849854  // usmmla z20.s, z2.b, z4.b\n"
-      ".inst 0x45859857  // usmmla z23.s, z2.b, z5.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x458498da  // usmmla z26.s, z6.b, z4.b\n"
-      ".inst 0x458598dd  // usmmla z29.s, z6.b, z5.b\n"
-      "ld1b { z4.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1b { z5.b }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x45839809  // usmmla z9.s, z0.b, z3.b\n"
-      "sub x20, x20, #0x2\n"
-      ".inst 0x4587980c  // usmmla z12.s, z0.b, z7.b\n"
-      ".inst 0x4583982f  // usmmla z15.s, z1.b, z3.b\n"
-      "cmp x20, #0x2\n"
-      ".inst 0x45879832  // usmmla z18.s, z1.b, z7.b\n"
-      ".inst 0x45839855  // usmmla z21.s, z2.b, z3.b\n"
-      ".inst 0x45879858  // usmmla z24.s, z2.b, z7.b\n"
-      ".inst 0x458398db  // usmmla z27.s, z6.b, z3.b\n"
-      "ld1b { z3.b }, p0/Z, [x22, #4, MUL VL]\n"
-      ".inst 0x458798de  // usmmla z30.s, z6.b, z7.b\n"
-      ".inst 0x4584980a  // usmmla z10.s, z0.b, z4.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #5, MUL VL]\n"
-      ".inst 0x4585980d  // usmmla z13.s, z0.b, z5.b\n"
-      ".inst 0x45849830  // usmmla z16.s, z1.b, z4.b\n"
-      "ld1rqb { z0.b }, p0/Z, [%x[Apanel], #16]\n"
-      ".inst 0x45859833  // usmmla z19.s, z1.b, z5.b\n"
-      ".inst 0x45849856  // usmmla z22.s, z2.b, z4.b\n"
-      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #32]\n"
-      ".inst 0x45859859  // usmmla z25.s, z2.b, z5.b\n"
-      ".inst 0x458498dc  // usmmla z28.s, z6.b, z4.b\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #48]\n"
-      ".inst 0x458598df  // usmmla z31.s, z6.b, z5.b\n"
-      "ld1rqb { z6.b }, p0/Z, [%x[Apanel], #64]\n"
-      "ld1b { z4.b }, p0/Z, [x22, #6, MUL VL]\n"
-      "ld1b { z5.b }, p0/Z, [x22, #7, MUL VL]\n"
-      "addvl x22, x22, #16\n"
-      ".inst 0x45839808  // usmmla z8.s, z0.b, z3.b\n"
-      ".inst 0x4587980b  // usmmla z11.s, z0.b, z7.b\n"
-      ".inst 0x4583982e  // usmmla z14.s, z1.b, z3.b\n"
-      ".inst 0x45879831  // usmmla z17.s, z1.b, z7.b\n"
-      ".inst 0x45839854  // usmmla z20.s, z2.b, z3.b\n"
-      ".inst 0x45879857  // usmmla z23.s, z2.b, z7.b\n"
-      ".inst 0x458398da  // usmmla z26.s, z6.b, z3.b\n"
-      "ld1b { z3.b }, p0/Z, [x22, #-8, MUL VL]\n"
-      ".inst 0x458798dd  // usmmla z29.s, z6.b, z7.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #-7, MUL VL]\n"
-      ".inst 0x45849809  // usmmla z9.s, z0.b, z4.b\n"
-      ".inst 0x4585980c  // usmmla z12.s, z0.b, z5.b\n"
-      ".inst 0x4584982f  // usmmla z15.s, z1.b, z4.b\n"
-      ".inst 0x45859832  // usmmla z18.s, z1.b, z5.b\n"
-      ".inst 0x45849855  // usmmla z21.s, z2.b, z4.b\n"
-      ".inst 0x45859858  // usmmla z24.s, z2.b, z5.b\n"
-      ".inst 0x458498db  // usmmla z27.s, z6.b, z4.b\n"
-      "ld1b { z4.b }, p0/Z, [x22, #-6, MUL VL]\n"
-      ".inst 0x458598de  // usmmla z30.s, z6.b, z5.b\n"
-      ".inst 0x4583980a  // usmmla z10.s, z0.b, z3.b\n"
-      "ld1b { z5.b }, p0/Z, [x22, #-5, MUL VL]\n"
-      ".inst 0x4587980d  // usmmla z13.s, z0.b, z7.b\n"
-      ".inst 0x45839830  // usmmla z16.s, z1.b, z3.b\n"
-      "ld1rqb { z0.b }, p0/Z, [%x[Apanel], #80]\n"
-      ".inst 0x45879833  // usmmla z19.s, z1.b, z7.b\n"
-      ".inst 0x45839856  // usmmla z22.s, z2.b, z3.b\n"
-      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #96]\n"
-      ".inst 0x45879859  // usmmla z25.s, z2.b, z7.b\n"
-      ".inst 0x458398dc  // usmmla z28.s, z6.b, z3.b\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #112]\n"
-      ".inst 0x458798df  // usmmla z31.s, z6.b, z7.b\n"
-      "add %x[Apanel], %x[Apanel], #0x80\n"
-      "addvl x22, x22, #-4\n"
-      "bge 3b\n"
-      "4:"  // main loop skip
-      "ld1rqb { z3.b }, p0/Z, [%x[Apanel]]\n"
-      ".inst 0x45849808  // usmmla z8.s, z0.b, z4.b\n"
-      ".inst 0x4585980b  // usmmla z11.s, z0.b, z5.b\n"
-      ".inst 0x4584982e  // usmmla z14.s, z1.b, z4.b\n"
-      ".inst 0x45859831  // usmmla z17.s, z1.b, z5.b\n"
-      "ld1b { z6.b }, p0/Z, [x22]\n"
-      ".inst 0x45849854  // usmmla z20.s, z2.b, z4.b\n"
-      ".inst 0x45859857  // usmmla z23.s, z2.b, z5.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x4584987a  // usmmla z26.s, z3.b, z4.b\n"
-      ".inst 0x4585987d  // usmmla z29.s, z3.b, z5.b\n"
-      "ld1b { z5.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1b { z4.b }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x45869809  // usmmla z9.s, z0.b, z6.b\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      ".inst 0x4587980c  // usmmla z12.s, z0.b, z7.b\n"
-      ".inst 0x4586982f  // usmmla z15.s, z1.b, z6.b\n"
-      "addvl x22, x22, #4\n"
-      ".inst 0x45879832  // usmmla z18.s, z1.b, z7.b\n"
-      ".inst 0x45869855  // usmmla z21.s, z2.b, z6.b\n"
-      ".inst 0x45879858  // usmmla z24.s, z2.b, z7.b\n"
-      ".inst 0x4586987b  // usmmla z27.s, z3.b, z6.b\n"
-      ".inst 0x4587987e  // usmmla z30.s, z3.b, z7.b\n"
-      ".inst 0x4585980a  // usmmla z10.s, z0.b, z5.b\n"
-      ".inst 0x4584980d  // usmmla z13.s, z0.b, z4.b\n"
-      ".inst 0x45859830  // usmmla z16.s, z1.b, z5.b\n"
-      ".inst 0x45849833  // usmmla z19.s, z1.b, z4.b\n"
-      ".inst 0x45859856  // usmmla z22.s, z2.b, z5.b\n"
-      ".inst 0x45849859  // usmmla z25.s, z2.b, z4.b\n"
-      ".inst 0x4585987c  // usmmla z28.s, z3.b, z5.b\n"
-      ".inst 0x4584987f  // usmmla z31.s, z3.b, z4.b\n"
-      "cbz x20, 5f\n"
-      "ld1b { z1.b }, p0/Z, [x22]\n"
-      "ld1rqb { z7.b }, p0/Z, [%x[Apanel]]\n"
-      "ld1rqb { z6.b }, p0/Z, [%x[Apanel], #16]\n"
-      "ld1b { z0.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "ld1rqb { z5.b }, p0/Z, [%x[Apanel], #32]\n"
-      "ld1rqb { z4.b }, p0/Z, [%x[Apanel], #48]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
-      ".inst 0x458198e8  // usmmla z8.s, z7.b, z1.b\n"
-      "ld1b { z3.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1b { z2.b }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x458098eb  // usmmla z11.s, z7.b, z0.b\n"
-      ".inst 0x458198ce  // usmmla z14.s, z6.b, z1.b\n"
-      ".inst 0x458098d1  // usmmla z17.s, z6.b, z0.b\n"
-      ".inst 0x458198b4  // usmmla z20.s, z5.b, z1.b\n"
-      ".inst 0x458098b7  // usmmla z23.s, z5.b, z0.b\n"
-      ".inst 0x4581989a  // usmmla z26.s, z4.b, z1.b\n"
-      "ld1b { z1.b }, p0/Z, [x22, #4, MUL VL]\n"
-      ".inst 0x4580989d  // usmmla z29.s, z4.b, z0.b\n"
-      "ld1b { z0.b }, p0/Z, [x22, #5, MUL VL]\n"
-      ".inst 0x458398e9  // usmmla z9.s, z7.b, z3.b\n"
-      ".inst 0x458298ec  // usmmla z12.s, z7.b, z2.b\n"
-      ".inst 0x458398cf  // usmmla z15.s, z6.b, z3.b\n"
-      "addvl x22, x22, #6\n"
-      ".inst 0x458298d2  // usmmla z18.s, z6.b, z2.b\n"
-      ".inst 0x458398b5  // usmmla z21.s, z5.b, z3.b\n"
-      ".inst 0x458298b8  // usmmla z24.s, z5.b, z2.b\n"
-      ".inst 0x4583989b  // usmmla z27.s, z4.b, z3.b\n"
-      ".inst 0x4582989e  // usmmla z30.s, z4.b, z2.b\n"
-      ".inst 0x458198ea  // usmmla z10.s, z7.b, z1.b\n"
-      ".inst 0x458098ed  // usmmla z13.s, z7.b, z0.b\n"
-      ".inst 0x458198d0  // usmmla z16.s, z6.b, z1.b\n"
-      ".inst 0x458098d3  // usmmla z19.s, z6.b, z0.b\n"
-      ".inst 0x458198b6  // usmmla z22.s, z5.b, z1.b\n"
-      ".inst 0x458098b9  // usmmla z25.s, z5.b, z0.b\n"
-      ".inst 0x4581989c  // usmmla z28.s, z4.b, z1.b\n"
-      ".inst 0x4580989f  // usmmla z31.s, z4.b, z0.b\n"
-      "5:"  // multiply loop done
-      "uzp1 z2.d, z8.d, z11.d\n"
-      "uzp2 z8.d, z8.d, z11.d\n"
-      "subs x23, x23, #0x1\n"
-      "uzp1 z1.d, z9.d, z12.d\n"
-      "uzp2 z9.d, z9.d, z12.d\n"
-      "uzp1 z0.d, z10.d, z13.d\n"
-      "uzp2 z10.d, z10.d, z13.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel]]\n"
-      "uzp1 z3.d, z14.d, z17.d\n"
-      "uzp2 z14.d, z14.d, z17.d\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "uzp1 z17.d, z15.d, z18.d\n"
-      "uzp2 z15.d, z15.d, z18.d\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "uzp1 z2.d, z16.d, z19.d\n"
-      "uzp2 z16.d, z16.d, z19.d\n"
-      "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "uzp1 z1.d, z20.d, z23.d\n"
-      "uzp2 z20.d, z20.d, z23.d\n"
-      "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
-      "uzp1 z0.d, z21.d, z24.d\n"
-      "uzp2 z21.d, z21.d, z24.d\n"
-      "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
-      "uzp1 z23.d, z22.d, z25.d\n"
-      "uzp2 z22.d, z22.d, z25.d\n"
-      "st1w { z3.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
-      "uzp1 z19.d, z26.d, z29.d\n"
-      "uzp2 z26.d, z26.d, z29.d\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
-      "addvl %x[Cpanel], %x[Cpanel], #16\n"
-      "uzp1 z18.d, z27.d, z30.d\n"
-      "uzp2 z27.d, z27.d, z30.d\n"
-      "uzp1 z17.d, z28.d, z31.d\n"
-      "uzp2 z28.d, z28.d, z31.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
-      "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
-      "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
-      "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
-      "st1w { z23.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
-      "st1w { z20.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
-      "st1w { z21.s }, p0, [%x[Cpanel]]\n"
-      "st1w { z22.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "st1w { z19.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "st1w { z18.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
-      "st1w { z26.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
-      "st1w { z27.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
-      "st1w { z28.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
-      "addvl %x[Cpanel], %x[Cpanel], #8\n"
-      "bgt 2b\n"
-      "subs %x[ablocks], %x[ablocks], #0x1\n"
-      "bne 1b\n"
-      : [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks)
-      : [args_ptr] "r" (&ka), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_bblocks] "I" (offsetof(KernelArgs, bblocks))
-      : "cc", "memory", "p0", "x20", "x21", "x22", "x23", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
-    );
-}
-
-} // namespace arm_gemm
-#endif // ARM_COMPUTE_ENABLE_SVE
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL.hpp
index bcb3279adb..171c810c8f 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL.hpp
@@ -40,8 +40,7 @@ void sve_interleaved_u8u32_dot_8x3VL_a64fx( ARGLIST );
 class cls_sve_interleaved_u8u32_dot_8x3VL
 {
 public:
-    typedef uint8_t lhs_operand_type;
-    typedef uint8_t rhs_operand_type;
+    typedef uint8_t operand_type;
     typedef uint32_t result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -63,8 +62,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 4, 1> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 3, 4, 1, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 4, 1> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 4, 1, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp
index 171cf38fa6..e2151ef41a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp
@@ -55,31 +55,31 @@ void sve_interleaved_u8u32_dot_8x3VL_a64fx(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
       "ld1b { z0.b }, p0/Z, [x22]\n"
+      "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
-      "mov z12.s, #0x0\n"
       "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
+      "mov z12.s, #0x0\n"
       "mov z13.s, #0x0\n"
+      "ld1b { z2.b }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
       "mov z16.s, #0x0\n"
-      "ld1b { z2.b }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z17.s, #0x0\n"
+      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "mov z18.s, #0x0\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
       "mov z19.s, #0x0\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "mov z20.s, #0x0\n"
-      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
       "mov z21.s, #0x0\n"
+      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "mov z22.s, #0x0\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
-      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
       "mov z25.s, #0x0\n"
       "mov z26.s, #0x0\n"
       "mov z27.s, #0x0\n"
@@ -93,7 +93,7 @@ void sve_interleaved_u8u32_dot_8x3VL_a64fx(
       "udot z9.s, z1.b, z3.b\n"
       "sub x20, x20, #0x2\n"
       "udot z10.s, z2.b, z3.b\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #16]\n"
       "udot z11.s, z0.b, z4.b\n"
       "udot z12.s, z1.b, z4.b\n"
       "udot z13.s, z2.b, z4.b\n"
@@ -102,63 +102,63 @@ void sve_interleaved_u8u32_dot_8x3VL_a64fx(
       "udot z15.s, z1.b, z5.b\n"
       "cmp x20, #0x2\n"
       "udot z16.s, z2.b, z5.b\n"
-      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #24]\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #24]\n"
       "udot z17.s, z0.b, z6.b\n"
       "udot z18.s, z1.b, z6.b\n"
       "udot z19.s, z2.b, z6.b\n"
-      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #28]\n"
-      "udot z20.s, z0.b, z3.b\n"
-      "udot z21.s, z1.b, z3.b\n"
-      "udot z22.s, z2.b, z3.b\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #32]\n"
+      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #28]\n"
+      "udot z20.s, z0.b, z7.b\n"
+      "udot z21.s, z1.b, z7.b\n"
+      "udot z22.s, z2.b, z7.b\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #32]\n"
       "udot z23.s, z0.b, z4.b\n"
       "udot z24.s, z1.b, z4.b\n"
       "udot z25.s, z2.b, z4.b\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #36]\n"
-      "udot z26.s, z0.b, z7.b\n"
-      "udot z27.s, z1.b, z7.b\n"
-      "udot z28.s, z2.b, z7.b\n"
+      "udot z26.s, z0.b, z3.b\n"
+      "udot z27.s, z1.b, z3.b\n"
+      "udot z28.s, z2.b, z3.b\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel], #40]\n"
-      "udot z29.s, z0.b, z6.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #3, MUL VL]\n"
-      "udot z30.s, z1.b, z6.b\n"
-      "udot z31.s, z2.b, z6.b\n"
-      "ld1b { z6.b }, p0/Z, [x22, #4, MUL VL]\n"
-      "ld1b { z2.b }, p0/Z, [x22, #5, MUL VL]\n"
-      "addvl x22, x22, #6\n"
+      "udot z29.s, z0.b, z5.b\n"
+      "ld1b { z6.b }, p0/Z, [x22, #3, MUL VL]\n"
+      "udot z30.s, z1.b, z5.b\n"
+      "udot z31.s, z2.b, z5.b\n"
+      "ld1b { z2.b }, p0/Z, [x22, #4, MUL VL]\n"
+      "ld1b { z5.b }, p0/Z, [x22, #5, MUL VL]\n"
+      "udot z8.s, z6.b, z7.b\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #44]\n"
-      "udot z8.s, z7.b, z5.b\n"
-      "udot z11.s, z7.b, z4.b\n"
-      "udot z9.s, z6.b, z5.b\n"
-      "udot z12.s, z6.b, z4.b\n"
-      "udot z10.s, z2.b, z5.b\n"
-      "udot z13.s, z2.b, z4.b\n"
-      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #48]\n"
-      "udot z14.s, z7.b, z3.b\n"
-      "udot z15.s, z6.b, z3.b\n"
+      "udot z9.s, z2.b, z7.b\n"
+      "udot z10.s, z5.b, z7.b\n"
+      "udot z11.s, z6.b, z4.b\n"
+      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #48]\n"
+      "udot z12.s, z2.b, z4.b\n"
+      "udot z13.s, z5.b, z4.b\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #52]\n"
-      "udot z16.s, z2.b, z3.b\n"
-      "udot z17.s, z7.b, z1.b\n"
+      "udot z14.s, z6.b, z3.b\n"
+      "udot z15.s, z2.b, z3.b\n"
+      "addvl x22, x22, #6\n"
+      "udot z16.s, z5.b, z3.b\n"
       "ld1rw { z0.s }, p0/Z, [%x[Apanel], #56]\n"
-      "udot z18.s, z6.b, z1.b\n"
-      "udot z19.s, z2.b, z1.b\n"
+      "udot z17.s, z6.b, z1.b\n"
+      "udot z18.s, z2.b, z1.b\n"
+      "udot z19.s, z5.b, z1.b\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #60]\n"
       "add %x[Apanel], %x[Apanel], #0x40\n"
-      "udot z20.s, z7.b, z5.b\n"
-      "udot z21.s, z6.b, z5.b\n"
-      "udot z22.s, z2.b, z5.b\n"
-      "udot z23.s, z7.b, z4.b\n"
+      "udot z20.s, z6.b, z7.b\n"
+      "udot z21.s, z2.b, z7.b\n"
+      "udot z22.s, z5.b, z7.b\n"
+      "udot z23.s, z6.b, z4.b\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
-      "udot z24.s, z6.b, z4.b\n"
-      "udot z25.s, z2.b, z4.b\n"
+      "udot z24.s, z2.b, z4.b\n"
+      "udot z25.s, z5.b, z4.b\n"
       "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
-      "udot z26.s, z7.b, z0.b\n"
-      "udot z27.s, z6.b, z0.b\n"
-      "udot z28.s, z2.b, z0.b\n"
-      "udot z29.s, z7.b, z1.b\n"
+      "udot z26.s, z6.b, z0.b\n"
+      "udot z27.s, z2.b, z0.b\n"
+      "udot z28.s, z5.b, z0.b\n"
+      "udot z29.s, z6.b, z1.b\n"
       "ld1b { z0.b }, p0/Z, [x22]\n"
-      "udot z30.s, z6.b, z1.b\n"
-      "udot z31.s, z2.b, z1.b\n"
+      "udot z30.s, z2.b, z1.b\n"
+      "udot z31.s, z5.b, z1.b\n"
       "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1b { z2.b }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
@@ -200,20 +200,19 @@ void sve_interleaved_u8u32_dot_8x3VL_a64fx(
       "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
       "ld1b { z4.b }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
-      "addvl x22, x22, #3\n"
+      "udot z8.s, z6.b, z3.b\n"
       "ld1rw { z2.s }, p0/Z, [%x[Apanel], #4]\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #8]\n"
-      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
-      "udot z8.s, z6.b, z3.b\n"
       "udot z9.s, z5.b, z3.b\n"
+      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
       "udot z10.s, z4.b, z3.b\n"
       "udot z11.s, z6.b, z2.b\n"
-      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
       "udot z12.s, z5.b, z2.b\n"
       "udot z13.s, z4.b, z2.b\n"
-      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
+      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
       "udot z14.s, z6.b, z1.b\n"
       "udot z15.s, z5.b, z1.b\n"
+      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
       "udot z16.s, z4.b, z1.b\n"
       "udot z17.s, z6.b, z0.b\n"
       "ld1rw { z1.s }, p0/Z, [%x[Apanel], #24]\n"
@@ -222,9 +221,10 @@ void sve_interleaved_u8u32_dot_8x3VL_a64fx(
       "ld1rw { z0.s }, p0/Z, [%x[Apanel], #28]\n"
       "udot z20.s, z6.b, z3.b\n"
       "udot z21.s, z5.b, z3.b\n"
-      "add %x[Apanel], %x[Apanel], #0x20\n"
+      "addvl x22, x22, #3\n"
       "udot z22.s, z4.b, z3.b\n"
       "udot z23.s, z6.b, z2.b\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "udot z24.s, z5.b, z2.b\n"
       "udot z25.s, z4.b, z2.b\n"
       "udot z26.s, z6.b, z1.b\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp
index dc31a73c13..eaa3ad2428 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/generic.cpp
@@ -55,25 +55,25 @@ void sve_interleaved_u8u32_dot_8x3VL(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
+      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
       "mov z10.s, #0x0\n"
-      "ld1b { z4.b }, p0/Z, [x22]\n"
       "mov z11.s, #0x0\n"
+      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
       "mov z12.s, #0x0\n"
-      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
       "mov z13.s, #0x0\n"
+      "ld1b { z4.b }, p0/Z, [x22]\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
+      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
       "mov z16.s, #0x0\n"
-      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
       "mov z17.s, #0x0\n"
+      "ld1b { z6.b }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z18.s, #0x0\n"
-      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
       "mov z19.s, #0x0\n"
       "mov z20.s, #0x0\n"
-      "ld1b { z6.b }, p0/Z, [x22, #2, MUL VL]\n"
       "mov z21.s, #0x0\n"
       "mov z22.s, #0x0\n"
       "mov z23.s, #0x0\n"
@@ -151,12 +151,12 @@ void sve_interleaved_u8u32_dot_8x3VL(
       "ld1b { z6.b }, p0/Z, [x22, #2, MUL VL]\n"
       "bge 3b\n"
       "4:"  // main loop skip
-      "add %x[Apanel], %x[Apanel], #0x20\n"
-      "addvl x22, x22, #3\n"
       "udot z8.s, z4.b, z0.b[0]\n"
       "udot z11.s, z4.b, z0.b[1]\n"
+      "add %x[Apanel], %x[Apanel], #0x20\n"
       "udot z14.s, z4.b, z0.b[2]\n"
       "udot z17.s, z4.b, z0.b[3]\n"
+      "addvl x22, x22, #3\n"
       "udot z20.s, z4.b, z1.b[0]\n"
       "udot z23.s, z4.b, z1.b[1]\n"
       "udot z26.s, z4.b, z1.b[2]\n"
@@ -183,13 +183,13 @@ void sve_interleaved_u8u32_dot_8x3VL(
       "add %x[Apanel], %x[Apanel], #0x20\n"
       "ld1b { z2.b }, p0/Z, [x22]\n"
       "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "ld1b { z0.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "addvl x22, x22, #3\n"
       "udot z8.s, z2.b, z4.b[0]\n"
+      "ld1b { z0.b }, p0/Z, [x22, #2, MUL VL]\n"
       "udot z11.s, z2.b, z4.b[1]\n"
       "udot z14.s, z2.b, z4.b[2]\n"
       "udot z17.s, z2.b, z4.b[3]\n"
       "udot z20.s, z2.b, z3.b[0]\n"
+      "addvl x22, x22, #3\n"
       "udot z23.s, z2.b, z3.b[1]\n"
       "udot z26.s, z2.b, z3.b[2]\n"
       "udot z29.s, z2.b, z3.b[3]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp
index f6a526d879..3bbf2bbfe4 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp
@@ -39,8 +39,7 @@ void sve_interleaved_u8u32_mmla_8x3VL( ARGLIST );
 class cls_sve_interleaved_u8u32_mmla_8x3VL
 {
 public:
-    typedef uint8_t lhs_operand_type;
-    typedef uint8_t rhs_operand_type;
+    typedef uint8_t operand_type;
     typedef uint32_t result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -62,8 +61,8 @@ public:
     }
 
 
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 8, 2> transforms = {};
-    StdTransformsSVE<lhs_operand_type, rhs_operand_type, result_type, 8, 6, 8, 2, true> transforms_quantized = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 8, 2> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 8, 2, true> transforms_quantized = {};
     template<typename T>
     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
     {
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp
index e5389a771d..c66026104d 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_mmla_8x3VL/generic.cpp
@@ -55,31 +55,31 @@ void sve_interleaved_u8u32_mmla_8x3VL(
       "2:"  // Width loop
       "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
       "mov %x[Apanel], x21\n"
+      "cmp x20, #0x2\n"
       "mov z8.s, #0x0\n"
       "mov z9.s, #0x0\n"
-      "mov z10.s, #0x0\n"
       "ld1b { z4.b }, p0/Z, [x22]\n"
+      "mov z10.s, #0x0\n"
       "mov z11.s, #0x0\n"
+      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
       "mov z12.s, #0x0\n"
-      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "cmp x20, #0x2\n"
       "mov z13.s, #0x0\n"
+      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
       "mov z14.s, #0x0\n"
       "mov z15.s, #0x0\n"
+      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
       "mov z16.s, #0x0\n"
-      "ld1rqb { z0.b }, p0/Z, [%x[Apanel]]\n"
       "mov z17.s, #0x0\n"
+      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #32]\n"
       "mov z18.s, #0x0\n"
-      "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #16]\n"
       "mov z19.s, #0x0\n"
+      "addvl x22, x22, #2\n"
       "mov z20.s, #0x0\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #32]\n"
       "mov z21.s, #0x0\n"
+      "add %x[Apanel], %x[Apanel], #0x30\n"
       "mov z22.s, #0x0\n"
-      "addvl x22, x22, #2\n"
       "mov z23.s, #0x0\n"
       "mov z24.s, #0x0\n"
-      "add %x[Apanel], %x[Apanel], #0x30\n"
       "mov z25.s, #0x0\n"
       "mov z26.s, #0x0\n"
       "mov z27.s, #0x0\n"
@@ -94,77 +94,77 @@ void sve_interleaved_u8u32_mmla_8x3VL(
       ".inst 0x45c5980b  // ummla z11.s, z0.b, z5.b\n"
       ".inst 0x45c4982e  // ummla z14.s, z1.b, z4.b\n"
       ".inst 0x45c59831  // ummla z17.s, z1.b, z5.b\n"
-      "ld1b { z3.b }, p0/Z, [x22]\n"
+      "ld1b { z7.b }, p0/Z, [x22]\n"
       ".inst 0x45c49854  // ummla z20.s, z2.b, z4.b\n"
       ".inst 0x45c59857  // ummla z23.s, z2.b, z5.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #1, MUL VL]\n"
+      "ld1b { z3.b }, p0/Z, [x22, #1, MUL VL]\n"
       ".inst 0x45c498da  // ummla z26.s, z6.b, z4.b\n"
       ".inst 0x45c598dd  // ummla z29.s, z6.b, z5.b\n"
-      "ld1b { z4.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1b { z5.b }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x45c39809  // ummla z9.s, z0.b, z3.b\n"
+      "ld1b { z5.b }, p0/Z, [x22, #2, MUL VL]\n"
+      "ld1b { z4.b }, p0/Z, [x22, #3, MUL VL]\n"
+      ".inst 0x45c79809  // ummla z9.s, z0.b, z7.b\n"
+      ".inst 0x45c3980c  // ummla z12.s, z0.b, z3.b\n"
+      ".inst 0x45c7982f  // ummla z15.s, z1.b, z7.b\n"
+      ".inst 0x45c39832  // ummla z18.s, z1.b, z3.b\n"
       "sub x20, x20, #0x2\n"
-      ".inst 0x45c7980c  // ummla z12.s, z0.b, z7.b\n"
-      ".inst 0x45c3982f  // ummla z15.s, z1.b, z3.b\n"
+      ".inst 0x45c79855  // ummla z21.s, z2.b, z7.b\n"
+      ".inst 0x45c39858  // ummla z24.s, z2.b, z3.b\n"
       "cmp x20, #0x2\n"
-      ".inst 0x45c79832  // ummla z18.s, z1.b, z7.b\n"
-      ".inst 0x45c39855  // ummla z21.s, z2.b, z3.b\n"
-      ".inst 0x45c79858  // ummla z24.s, z2.b, z7.b\n"
-      ".inst 0x45c398db  // ummla z27.s, z6.b, z3.b\n"
+      ".inst 0x45c798db  // ummla z27.s, z6.b, z7.b\n"
+      ".inst 0x45c398de  // ummla z30.s, z6.b, z3.b\n"
       "ld1b { z3.b }, p0/Z, [x22, #4, MUL VL]\n"
-      ".inst 0x45c798de  // ummla z30.s, z6.b, z7.b\n"
-      ".inst 0x45c4980a  // ummla z10.s, z0.b, z4.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #5, MUL VL]\n"
-      ".inst 0x45c5980d  // ummla z13.s, z0.b, z5.b\n"
-      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
+      ".inst 0x45c5980a  // ummla z10.s, z0.b, z5.b\n"
+      ".inst 0x45c4980d  // ummla z13.s, z0.b, z4.b\n"
       "ld1rqb { z0.b }, p0/Z, [%x[Apanel], #16]\n"
-      ".inst 0x45c59833  // ummla z19.s, z1.b, z5.b\n"
-      ".inst 0x45c49856  // ummla z22.s, z2.b, z4.b\n"
+      ".inst 0x45c59830  // ummla z16.s, z1.b, z5.b\n"
+      ".inst 0x45c49833  // ummla z19.s, z1.b, z4.b\n"
       "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #32]\n"
-      ".inst 0x45c59859  // ummla z25.s, z2.b, z5.b\n"
-      ".inst 0x45c498dc  // ummla z28.s, z6.b, z4.b\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #48]\n"
-      ".inst 0x45c598df  // ummla z31.s, z6.b, z5.b\n"
+      ".inst 0x45c59856  // ummla z22.s, z2.b, z5.b\n"
+      ".inst 0x45c49859  // ummla z25.s, z2.b, z4.b\n"
+      "ld1b { z7.b }, p0/Z, [x22, #5, MUL VL]\n"
+      ".inst 0x45c598dc  // ummla z28.s, z6.b, z5.b\n"
+      ".inst 0x45c498df  // ummla z31.s, z6.b, z4.b\n"
+      "ld1rqb { z5.b }, p0/Z, [%x[Apanel], #48]\n"
       "ld1rqb { z6.b }, p0/Z, [%x[Apanel], #64]\n"
-      "ld1b { z4.b }, p0/Z, [x22, #6, MUL VL]\n"
-      "ld1b { z5.b }, p0/Z, [x22, #7, MUL VL]\n"
-      "addvl x22, x22, #16\n"
+      "ld1b { z2.b }, p0/Z, [x22, #6, MUL VL]\n"
       ".inst 0x45c39808  // ummla z8.s, z0.b, z3.b\n"
+      "ld1b { z4.b }, p0/Z, [x22, #7, MUL VL]\n"
+      "addvl x22, x22, #16\n"
       ".inst 0x45c7980b  // ummla z11.s, z0.b, z7.b\n"
       ".inst 0x45c3982e  // ummla z14.s, z1.b, z3.b\n"
       ".inst 0x45c79831  // ummla z17.s, z1.b, z7.b\n"
-      ".inst 0x45c39854  // ummla z20.s, z2.b, z3.b\n"
-      ".inst 0x45c79857  // ummla z23.s, z2.b, z7.b\n"
+      ".inst 0x45c398b4  // ummla z20.s, z5.b, z3.b\n"
+      ".inst 0x45c798b7  // ummla z23.s, z5.b, z7.b\n"
       ".inst 0x45c398da  // ummla z26.s, z6.b, z3.b\n"
-      "ld1b { z3.b }, p0/Z, [x22, #-8, MUL VL]\n"
       ".inst 0x45c798dd  // ummla z29.s, z6.b, z7.b\n"
+      "ld1b { z3.b }, p0/Z, [x22, #-8, MUL VL]\n"
       "ld1b { z7.b }, p0/Z, [x22, #-7, MUL VL]\n"
-      ".inst 0x45c49809  // ummla z9.s, z0.b, z4.b\n"
-      ".inst 0x45c5980c  // ummla z12.s, z0.b, z5.b\n"
-      ".inst 0x45c4982f  // ummla z15.s, z1.b, z4.b\n"
-      ".inst 0x45c59832  // ummla z18.s, z1.b, z5.b\n"
-      ".inst 0x45c49855  // ummla z21.s, z2.b, z4.b\n"
-      ".inst 0x45c59858  // ummla z24.s, z2.b, z5.b\n"
-      ".inst 0x45c498db  // ummla z27.s, z6.b, z4.b\n"
+      ".inst 0x45c29809  // ummla z9.s, z0.b, z2.b\n"
+      ".inst 0x45c4980c  // ummla z12.s, z0.b, z4.b\n"
+      ".inst 0x45c2982f  // ummla z15.s, z1.b, z2.b\n"
+      ".inst 0x45c49832  // ummla z18.s, z1.b, z4.b\n"
+      ".inst 0x45c298b5  // ummla z21.s, z5.b, z2.b\n"
+      ".inst 0x45c498b8  // ummla z24.s, z5.b, z4.b\n"
+      ".inst 0x45c298db  // ummla z27.s, z6.b, z2.b\n"
+      ".inst 0x45c498de  // ummla z30.s, z6.b, z4.b\n"
       "ld1b { z4.b }, p0/Z, [x22, #-6, MUL VL]\n"
-      ".inst 0x45c598de  // ummla z30.s, z6.b, z5.b\n"
       ".inst 0x45c3980a  // ummla z10.s, z0.b, z3.b\n"
-      "ld1b { z5.b }, p0/Z, [x22, #-5, MUL VL]\n"
       ".inst 0x45c7980d  // ummla z13.s, z0.b, z7.b\n"
-      ".inst 0x45c39830  // ummla z16.s, z1.b, z3.b\n"
       "ld1rqb { z0.b }, p0/Z, [%x[Apanel], #80]\n"
+      ".inst 0x45c39830  // ummla z16.s, z1.b, z3.b\n"
       ".inst 0x45c79833  // ummla z19.s, z1.b, z7.b\n"
-      ".inst 0x45c39856  // ummla z22.s, z2.b, z3.b\n"
       "ld1rqb { z1.b }, p0/Z, [%x[Apanel], #96]\n"
-      ".inst 0x45c79859  // ummla z25.s, z2.b, z7.b\n"
+      ".inst 0x45c398b6  // ummla z22.s, z5.b, z3.b\n"
+      ".inst 0x45c798b9  // ummla z25.s, z5.b, z7.b\n"
+      "ld1b { z5.b }, p0/Z, [x22, #-5, MUL VL]\n"
       ".inst 0x45c398dc  // ummla z28.s, z6.b, z3.b\n"
-      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #112]\n"
       ".inst 0x45c798df  // ummla z31.s, z6.b, z7.b\n"
+      "ld1rqb { z2.b }, p0/Z, [%x[Apanel], #112]\n"
       "add %x[Apanel], %x[Apanel], #0x80\n"
       "addvl x22, x22, #-4\n"
       "bge 3b\n"
       "4:"  // main loop skip
-      "ld1rqb { z3.b }, p0/Z, [%x[Apanel]]\n"
+      "ld1rqb { z7.b }, p0/Z, [%x[Apanel]]\n"
       ".inst 0x45c49808  // ummla z8.s, z0.b, z4.b\n"
       ".inst 0x45c5980b  // ummla z11.s, z0.b, z5.b\n"
       ".inst 0x45c4982e  // ummla z14.s, z1.b, z4.b\n"
@@ -172,54 +172,54 @@ void sve_interleaved_u8u32_mmla_8x3VL(
       "ld1b { z6.b }, p0/Z, [x22]\n"
       ".inst 0x45c49854  // ummla z20.s, z2.b, z4.b\n"
       ".inst 0x45c59857  // ummla z23.s, z2.b, z5.b\n"
-      "ld1b { z7.b }, p0/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x45c4987a  // ummla z26.s, z3.b, z4.b\n"
-      ".inst 0x45c5987d  // ummla z29.s, z3.b, z5.b\n"
+      "ld1b { z3.b }, p0/Z, [x22, #1, MUL VL]\n"
+      ".inst 0x45c498fa  // ummla z26.s, z7.b, z4.b\n"
+      ".inst 0x45c598fd  // ummla z29.s, z7.b, z5.b\n"
       "ld1b { z5.b }, p0/Z, [x22, #2, MUL VL]\n"
       "ld1b { z4.b }, p0/Z, [x22, #3, MUL VL]\n"
       ".inst 0x45c69809  // ummla z9.s, z0.b, z6.b\n"
-      "add %x[Apanel], %x[Apanel], #0x10\n"
-      ".inst 0x45c7980c  // ummla z12.s, z0.b, z7.b\n"
+      ".inst 0x45c3980c  // ummla z12.s, z0.b, z3.b\n"
       ".inst 0x45c6982f  // ummla z15.s, z1.b, z6.b\n"
-      "addvl x22, x22, #4\n"
-      ".inst 0x45c79832  // ummla z18.s, z1.b, z7.b\n"
+      ".inst 0x45c39832  // ummla z18.s, z1.b, z3.b\n"
+      "add %x[Apanel], %x[Apanel], #0x10\n"
       ".inst 0x45c69855  // ummla z21.s, z2.b, z6.b\n"
-      ".inst 0x45c79858  // ummla z24.s, z2.b, z7.b\n"
-      ".inst 0x45c6987b  // ummla z27.s, z3.b, z6.b\n"
-      ".inst 0x45c7987e  // ummla z30.s, z3.b, z7.b\n"
+      ".inst 0x45c39858  // ummla z24.s, z2.b, z3.b\n"
+      "addvl x22, x22, #4\n"
+      ".inst 0x45c698fb  // ummla z27.s, z7.b, z6.b\n"
+      ".inst 0x45c398fe  // ummla z30.s, z7.b, z3.b\n"
       ".inst 0x45c5980a  // ummla z10.s, z0.b, z5.b\n"
       ".inst 0x45c4980d  // ummla z13.s, z0.b, z4.b\n"
       ".inst 0x45c59830  // ummla z16.s, z1.b, z5.b\n"
       ".inst 0x45c49833  // ummla z19.s, z1.b, z4.b\n"
       ".inst 0x45c59856  // ummla z22.s, z2.b, z5.b\n"
       ".inst 0x45c49859  // ummla z25.s, z2.b, z4.b\n"
-      ".inst 0x45c5987c  // ummla z28.s, z3.b, z5.b\n"
-      ".inst 0x45c4987f  // ummla z31.s, z3.b, z4.b\n"
+      ".inst 0x45c598fc  // ummla z28.s, z7.b, z5.b\n"
+      ".inst 0x45c498ff  // ummla z31.s, z7.b, z4.b\n"
       "cbz x20, 5f\n"
       "ld1b { z1.b }, p0/Z, [x22]\n"
       "ld1rqb { z7.b }, p0/Z, [%x[Apanel]]\n"
+      ".inst 0x45c198e8  // ummla z8.s, z7.b, z1.b\n"
       "ld1rqb { z6.b }, p0/Z, [%x[Apanel], #16]\n"
       "ld1b { z0.b }, p0/Z, [x22, #1, MUL VL]\n"
+      ".inst 0x45c098eb  // ummla z11.s, z7.b, z0.b\n"
       "ld1rqb { z5.b }, p0/Z, [%x[Apanel], #32]\n"
       "ld1rqb { z4.b }, p0/Z, [%x[Apanel], #48]\n"
-      "add %x[Apanel], %x[Apanel], #0x40\n"
-      ".inst 0x45c198e8  // ummla z8.s, z7.b, z1.b\n"
-      "ld1b { z3.b }, p0/Z, [x22, #2, MUL VL]\n"
-      "ld1b { z2.b }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x45c098eb  // ummla z11.s, z7.b, z0.b\n"
       ".inst 0x45c198ce  // ummla z14.s, z6.b, z1.b\n"
       ".inst 0x45c098d1  // ummla z17.s, z6.b, z0.b\n"
       ".inst 0x45c198b4  // ummla z20.s, z5.b, z1.b\n"
+      "ld1b { z3.b }, p0/Z, [x22, #2, MUL VL]\n"
       ".inst 0x45c098b7  // ummla z23.s, z5.b, z0.b\n"
       ".inst 0x45c1989a  // ummla z26.s, z4.b, z1.b\n"
-      "ld1b { z1.b }, p0/Z, [x22, #4, MUL VL]\n"
+      "ld1b { z2.b }, p0/Z, [x22, #3, MUL VL]\n"
       ".inst 0x45c0989d  // ummla z29.s, z4.b, z0.b\n"
+      "ld1b { z1.b }, p0/Z, [x22, #4, MUL VL]\n"
       "ld1b { z0.b }, p0/Z, [x22, #5, MUL VL]\n"
       ".inst 0x45c398e9  // ummla z9.s, z7.b, z3.b\n"
       ".inst 0x45c298ec  // ummla z12.s, z7.b, z2.b\n"
-      ".inst 0x45c398cf  // ummla z15.s, z6.b, z3.b\n"
       "addvl x22, x22, #6\n"
+      ".inst 0x45c398cf  // ummla z15.s, z6.b, z3.b\n"
       ".inst 0x45c298d2  // ummla z18.s, z6.b, z2.b\n"
+      "add %x[Apanel], %x[Apanel], #0x40\n"
       ".inst 0x45c398b5  // ummla z21.s, z5.b, z3.b\n"
       ".inst 0x45c298b8  // ummla z24.s, z5.b, z2.b\n"
       ".inst 0x45c3989b  // ummla z27.s, z4.b, z3.b\n"
@@ -233,53 +233,53 @@ void sve_interleaved_u8u32_mmla_8x3VL(
       ".inst 0x45c1989c  // ummla z28.s, z4.b, z1.b\n"
       ".inst 0x45c0989f  // ummla z31.s, z4.b, z0.b\n"
       "5:"  // multiply loop done
-      "uzp1 z2.d, z8.d, z11.d\n"
+      "uzp1 z0.d, z8.d, z11.d\n"
       "uzp2 z8.d, z8.d, z11.d\n"
-      "subs x23, x23, #0x1\n"
-      "uzp1 z1.d, z9.d, z12.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel]]\n"
+      "uzp1 z0.d, z9.d, z12.d\n"
       "uzp2 z9.d, z9.d, z12.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
       "uzp1 z0.d, z10.d, z13.d\n"
       "uzp2 z10.d, z10.d, z13.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel]]\n"
-      "uzp1 z3.d, z14.d, z17.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+      "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+      "uzp1 z0.d, z14.d, z17.d\n"
       "uzp2 z14.d, z14.d, z17.d\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "uzp1 z17.d, z15.d, z18.d\n"
+      "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+      "uzp1 z1.d, z15.d, z18.d\n"
+      "subs x23, x23, #0x1\n"
+      "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "uzp2 z15.d, z15.d, z18.d\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "uzp1 z2.d, z16.d, z19.d\n"
+      "uzp1 z17.d, z16.d, z19.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
       "uzp2 z16.d, z16.d, z19.d\n"
-      "st1w { z8.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "uzp1 z1.d, z20.d, z23.d\n"
+      "uzp1 z0.d, z20.d, z23.d\n"
+      "st1w { z1.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
+      "addvl %x[Cpanel], %x[Cpanel], #16\n"
       "uzp2 z20.d, z20.d, z23.d\n"
-      "st1w { z9.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
-      "uzp1 z0.d, z21.d, z24.d\n"
+      "st1w { z17.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
+      "uzp1 z23.d, z21.d, z24.d\n"
       "uzp2 z21.d, z21.d, z24.d\n"
-      "st1w { z10.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
-      "uzp1 z23.d, z22.d, z25.d\n"
+      "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
+      "uzp1 z19.d, z22.d, z25.d\n"
       "uzp2 z22.d, z22.d, z25.d\n"
-      "st1w { z3.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
-      "uzp1 z19.d, z26.d, z29.d\n"
+      "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
+      "uzp1 z18.d, z26.d, z29.d\n"
       "uzp2 z26.d, z26.d, z29.d\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
-      "addvl %x[Cpanel], %x[Cpanel], #16\n"
-      "uzp1 z18.d, z27.d, z30.d\n"
+      "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
+      "uzp1 z17.d, z27.d, z30.d\n"
       "uzp2 z27.d, z27.d, z30.d\n"
-      "uzp1 z17.d, z28.d, z31.d\n"
+      "st1w { z0.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
+      "uzp1 z16.d, z28.d, z31.d\n"
       "uzp2 z28.d, z28.d, z31.d\n"
-      "st1w { z2.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
-      "st1w { z14.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
-      "st1w { z15.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
-      "st1w { z16.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
-      "st1w { z1.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
-      "st1w { z0.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
-      "st1w { z23.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
+      "st1w { z23.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
+      "st1w { z19.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
       "st1w { z20.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
       "st1w { z21.s }, p0, [%x[Cpanel]]\n"
       "st1w { z22.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
-      "st1w { z19.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
-      "st1w { z18.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
-      "st1w { z17.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
+      "st1w { z18.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
+      "st1w { z17.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
+      "st1w { z16.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
       "st1w { z26.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
       "st1w { z27.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
       "st1w { z28.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/performance_parameters.hpp b/src/core/NEON/kernels/arm_gemm/performance_parameters.hpp
index 57f779c498..ea00cccae5 100644
--- a/src/core/NEON/kernels/arm_gemm/performance_parameters.hpp
+++ b/src/core/NEON/kernels/arm_gemm/performance_parameters.hpp
@@ -26,9 +26,9 @@
 namespace arm_gemm {
 
 struct PerformanceParameters {
-    float  kernel_macs_cycle;
-    float  prepare_bytes_cycle = 0.0f;
-    float  merge_bytes_cycle   = 0.0f;
+    float	kernel_macs_cycle;
+    float	prepare_bytes_cycle = 0.0f;
+    float	merge_bytes_cycle   = 0.0f;
 
     PerformanceParameters(float k) : kernel_macs_cycle(k) { }
     PerformanceParameters(float k, float p, float m) : kernel_macs_cycle(k), prepare_bytes_cycle(p), merge_bytes_cycle(m) { }
diff --git a/src/core/NEON/kernels/arm_gemm/quantize_wrapper.hpp b/src/core/NEON/kernels/arm_gemm/quantize_wrapper.hpp
index 1a90cf7d89..d35825c428 100644
--- a/src/core/NEON/kernels/arm_gemm/quantize_wrapper.hpp
+++ b/src/core/NEON/kernels/arm_gemm/quantize_wrapper.hpp
@@ -35,9 +35,9 @@ namespace arm_gemm {
 /* Quantized wrapper - do an integer GEMM and wrap around the quantization. */
 
 template<typename To, typename Tr, typename Tgemm>
-class QuantizeWrapper : public GemmCommon<To, To, Tr> {
+class QuantizeWrapper : public GemmCommon<To, Tr> {
 private:
-    UniqueGemmCommon<To, To, Tgemm>  _subgemm = nullptr;
+    UniqueGemmCommon<To, Tgemm>  _subgemm = nullptr;
     int32_t                     *_row_sums = nullptr;
     int32_t                     *_col_sums = nullptr;
     Requantize32                 _params;
@@ -111,7 +111,7 @@ public:
 
     QuantizeWrapper(const GemmArgs &args, const Requantize32 &qp) : _params(qp), _args(args), _barrier(args._maxthreads) {
         GemmArgs newargs = GemmArgs(args._ci, args._Msize, args._Nsize, args._Ksize, args._Ksections, args._nbatches, args._nmulti, args._indirect_input, Activation(), args._maxthreads);
-        _subgemm = gemm<To, To, Tgemm>(newargs);
+        _subgemm = gemm<To, Tgemm>(newargs);
 
         if (_subgemm == nullptr) {
             return;
@@ -122,7 +122,7 @@ public:
                     const To *B, const int ldb, const int B_multi_stride,
                           Tr *C, const int ldc, const int C_batch_stride, const int C_multi_stride,
                     const Tr *bias, const int bias_multi_stride) override {
-        GemmCommon<To, To, Tr>::set_arrays(A, lda, A_batch_stride, A_multi_stride, B, ldb, B_multi_stride, C, ldc, C_batch_stride, C_multi_stride, bias, bias_multi_stride);
+        GemmCommon<To, Tr>::set_arrays(A, lda, A_batch_stride, A_multi_stride, B, ldb, B_multi_stride, C, ldc, C_batch_stride, C_multi_stride, bias, bias_multi_stride);
 
         arrays_set = true;
         set_child_arrays();
diff --git a/src/core/NEON/kernels/arm_gemm/quantized.cpp b/src/core/NEON/kernels/arm_gemm/quantized.cpp
index f24d5956e5..6da9f4be0e 100644
--- a/src/core/NEON/kernels/arm_gemm/quantized.cpp
+++ b/src/core/NEON/kernels/arm_gemm/quantized.cpp
@@ -745,10 +745,6 @@ template void requantize_block_32(const Requantize32 &qp, unsigned int width, un
                          const uint32_t *input, unsigned int in_stride, uint8_t *output, unsigned int out_stride,
                          const int32_t *row_bias, const int32_t *col_bias, unsigned int start_col);
 
-template void requantize_block_32(const Requantize32 &qp, unsigned int width, unsigned int height,
-                         const int32_t *input, unsigned int in_stride, uint8_t *output, unsigned int out_stride,
-                         const int32_t *row_bias, const int32_t *col_bias, unsigned int start_col);
-
 /*
  * Routine (and helpers) to compute row sums needed for offset correction.
  *
diff --git a/src/core/NEON/kernels/arm_gemm/std_transforms_fixed.hpp b/src/core/NEON/kernels/arm_gemm/std_transforms_fixed.hpp
index e43eb8a09d..a9cbf4ec8d 100644
--- a/src/core/NEON/kernels/arm_gemm/std_transforms_fixed.hpp
+++ b/src/core/NEON/kernels/arm_gemm/std_transforms_fixed.hpp
@@ -41,24 +41,24 @@ namespace arm_gemm {
  * The optional 'block' parameter is for kernels using dot-product type
  * instructions like UDOT and SDOT.
  */
-template<typename TInput, typename TWeight, typename TResult, unsigned int height, unsigned int width, unsigned int block=1, bool integrate_sums=false>
+template<typename TOperand, typename TResult, unsigned int height, unsigned int width, unsigned int block=1, bool integrate_sums=false>
 class StdTransformsFixed
 {
 public:
     template<typename TIn>
-    void PrepareA(TInput *out, const TIn *in, const int stride, const int y0,
+    void PrepareA(TOperand *out, const TIn *in, const int stride, const int y0,
                   const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) const {
         Interleave<height, block, VLType::None>(out, in, stride, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
 
     template<typename TIn>
-    void PrepareA_indirect(TInput *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0,
+    void PrepareA_indirect(TOperand *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0,
                            const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
         IndirectInterleave<height, block, VLType::None>(out, ptr, stringlen, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
 
     template<typename TIn>
-    void PrepareA_convolution(TInput *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen,
+    void PrepareA_convolution(TOperand *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen,
                               const int y0, const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
         ConvolutionInterleave<height, block, VLType::None>(out, ptr, stride, conv, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
@@ -68,7 +68,7 @@ public:
     }
 
     template<typename TIn>
-    void PrepareB(TWeight *out, const TIn *in, const int stride, const int x0,
+    void PrepareB(TOperand *out, const TIn *in, const int stride, const int x0,
                   const int xmax, const int k0, const int kmax, bool transposed) const {
         assert(!transposed);
         Transform<width, block,  true>(out, in, stride, x0, xmax, k0, kmax);
diff --git a/src/core/NEON/kernels/arm_gemm/std_transforms_fixed_trB.hpp b/src/core/NEON/kernels/arm_gemm/std_transforms_fixed_trB.hpp
index ec3cad0385..1db716455f 100644
--- a/src/core/NEON/kernels/arm_gemm/std_transforms_fixed_trB.hpp
+++ b/src/core/NEON/kernels/arm_gemm/std_transforms_fixed_trB.hpp
@@ -42,24 +42,24 @@ namespace arm_gemm {
  * The optional 'block' parameter is for kernels using dot-product type
  * instructions like UDOT and SDOT.
  */
-template<typename TInput, typename TWeight, typename TResult, unsigned int height, unsigned int width, unsigned int block=1, bool integrate_sums=false>
+template<typename TOperand, typename TResult, unsigned int height, unsigned int width, unsigned int block=1, bool integrate_sums=false>
 class StdTransformsFixedTRB
 {
 public:
     template<typename TIn>
-    void PrepareA(TInput *out, const TIn *in, const int stride, const int y0,
+    void PrepareA(TOperand *out, const TIn *in, const int stride, const int y0,
                   const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) const {
         Interleave<height, block, VLType::None>(out, in, stride, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
 
     template<typename TIn>
-    void PrepareA_indirect(TInput *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0,
+    void PrepareA_indirect(TOperand *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0,
                            const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
         IndirectInterleave<height, block, VLType::None>(out, ptr, stringlen, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
 
     template<typename TIn>
-    void PrepareA_convolution(TInput *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen,
+    void PrepareA_convolution(TOperand *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen,
                               const int y0, const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
         ConvolutionInterleave<height, block, VLType::None>(out, ptr, stride, conv, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
@@ -69,7 +69,7 @@ public:
     }
 
     template<typename TIn>
-    void PrepareB(TWeight *out, const TIn *in, const int stride, const int x0,
+    void PrepareB(TOperand *out, const TIn *in, const int stride, const int x0,
                   const int xmax, const int k0, const int kmax, bool transposed) const {
         if (transposed) {
             Transform<width, block, false>(out, in, stride, x0, xmax, k0, kmax);
diff --git a/src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp b/src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp
index 32d597f4af..c516bfc456 100644
--- a/src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp
+++ b/src/core/NEON/kernels/arm_gemm/std_transforms_sve.hpp
@@ -39,24 +39,24 @@ namespace arm_gemm {
  * The optional 'block' parameter is for kernels using dot-product type
  * instructions like UDOT and SDOT.
  */
-template<typename TInput, typename TWeight, typename TResult, unsigned int height, unsigned int width_vectors, unsigned int block=1, unsigned int mmla=1, bool integrate_sums=false>
+template<typename TOperand, typename TResult, unsigned int height, unsigned int width_vectors, unsigned int block=1, unsigned int mmla=1, bool integrate_sums=false>
 class StdTransformsSVE
 {
 public:
     template<typename TIn>
-    void PrepareA(TInput *out, const TIn *in, const int stride, const int y0,
+    void PrepareA(TOperand *out, const TIn *in, const int stride, const int y0,
                   const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
         Interleave<height, block, VLType::None>(out, in, stride, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
 
     template<typename TIn>
-    void PrepareA_indirect(TInput *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0,
+    void PrepareA_indirect(TOperand *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0,
                            const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
         IndirectInterleave<height, block, VLType::None>(out, ptr, stringlen, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
 
     template<typename TIn>
-    void PrepareA_convolution(TInput *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen,
+    void PrepareA_convolution(TOperand *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen,
                               const int y0, const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
         ConvolutionInterleave<height, block, VLType::None>(out, ptr, stride, conv, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
     }
@@ -66,7 +66,7 @@ public:
     }
 
     template<typename TIn>
-    void PrepareB(TWeight *out, const TIn *in, const int stride, const int x0,
+    void PrepareB(TOperand *out, const TIn *in, const int stride, const int x0,
                   const int xmax, const int k0, const int kmax, bool transposed) {
         assert (!transposed);
         Transform<width_vectors, block,  true, VLType::SVE>(out, in, stride, x0, xmax, k0, kmax);
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp
index 3690727f11..16e0822782 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a32_transpose_interleave_8way_32bit.hpp
@@ -61,8 +61,8 @@ void TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform(
 template <>
 void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x1(const uint16_t *&in0, uint16_t *out) {
   __asm volatile (
-    "VLD1.32    {d0-d3}, [%[in0]]!\n"
-    "VST1.32    {d0-d3}, [%[out]]\n"
+    "VLD1.32	{d0-d3}, [%[in0]]!\n"
+    "VST1.32	{d0-d3}, [%[out]]\n"
     ASM_PREFETCH("[%[in0], #192]")
     : [in0] "+r" (in0),
       [out] "+r" (out)
@@ -74,13 +74,13 @@ void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x1(const uint
 template <>
 void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x2(const uint16_t *&in0, const uint16_t *&in1, uint16_t *out) {
   __asm volatile (
-    "VLD1.32    {d0-d3}, [%[in0]]!\n"
-    "VST1.32    {d0-d3}, [%[out]]!\n"
+    "VLD1.32	{d0-d3}, [%[in0]]!\n"
+    "VST1.32	{d0-d3}, [%[out]]!\n"
     ASM_PREFETCH("[%[in0], #192]")
-    "VLD1.32    {d0-d3}, [%[in1]]!\n"
-    "VST1.32    {d0-d3}, [%[out]]\n"
+    "VLD1.32	{d0-d3}, [%[in1]]!\n"
+    "VST1.32	{d0-d3}, [%[out]]\n"
     ASM_PREFETCH("[%[in1], #192]")
-    "SUB    %[out], %[out], #32\n"
+    "SUB	%[out], %[out], #32\n"
     : [in0] "+r" (in0),
       [in1] "+r" (in1),
       [out] "+r" (out)
@@ -92,19 +92,19 @@ void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x2(const uint
 template <>
 void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x4(const uint16_t *&in0, const uint16_t *&in1, const uint16_t *&in2, const uint16_t *&in3, uint16_t *out) {
   __asm __volatile (
-    "VLD1.32    {d0-d3}, [%[in0]]!\n"
-    "VST1.32    {d0-d3}, [%[out]]!\n"
+    "VLD1.32	{d0-d3}, [%[in0]]!\n"
+    "VST1.32	{d0-d3}, [%[out]]!\n"
     ASM_PREFETCH("[%[in0], #192]")
-    "VLD1.32    {d0-d3}, [%[in1]]!\n"
-    "VST1.32    {d0-d3}, [%[out]]!\n"
+    "VLD1.32	{d0-d3}, [%[in1]]!\n"
+    "VST1.32	{d0-d3}, [%[out]]!\n"
     ASM_PREFETCH("[%[in1], #192]")
-    "VLD1.32    {d0-d3}, [%[in2]]!\n"
-    "VST1.32    {d0-d3}, [%[out]]!\n"
+    "VLD1.32	{d0-d3}, [%[in2]]!\n"
+    "VST1.32	{d0-d3}, [%[out]]!\n"
     ASM_PREFETCH("[%[in2], #192]")
-    "VLD1.32    {d0-d3}, [%[in3]]!\n"
-    "VST1.32    {d0-d3}, [%[out]]\n"
+    "VLD1.32	{d0-d3}, [%[in3]]!\n"
+    "VST1.32	{d0-d3}, [%[out]]\n"
     ASM_PREFETCH("[%[in3], #192]")
-    "SUB    %[out], %[out], #96\n"
+    "SUB	%[out], %[out], #96\n"
     : [in0] "+r" (in0),
       [in1] "+r" (in1),
       [in2] "+r" (in2),
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_128.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_128.hpp
index 618992c481..af5ecf5a8b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_128.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_128.hpp
@@ -34,281 +34,235 @@ void a64_transpose_interleave_128(uint32_t *out, const uint32_t *in, size_t widt
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x20\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
       "ldr q15, [x25], #0x10\n"
-      "ldr q14, [x22], #0x10\n"
+      "ldr q14, [x23], #0x10\n"
       "sub x24, x24, #0x20\n"
-      "ldr q13, [x21], #0x10\n"
-      "ldr q12, [x20], #0x10\n"
       "cmp x24, #0x20\n"
+      "ldr q13, [x22], #0x10\n"
+      "ldr q12, [x20], #0x10\n"
       "ldr q11, [x25], #0x10\n"
-      "ldr q10, [x22], #0x10\n"
-      "ldr q9, [x21], #0x10\n"
+      "ldr q10, [x23], #0x10\n"
+      "ldr q9, [x22], #0x10\n"
       "ldr q8, [x20], #0x10\n"
       "ldr q7, [x25], #0x10\n"
-      "ldr q6, [x22], #0x10\n"
-      "ldr q5, [x21], #0x10\n"
+      "ldr q6, [x23], #0x10\n"
+      "ldr q5, [x22], #0x10\n"
       "ldr q4, [x20], #0x10\n"
       "ldr q3, [x25], #0x10\n"
-      "ldr q2, [x22], #0x10\n"
-      "ldr q1, [x21], #0x10\n"
+      "ldr q2, [x23], #0x10\n"
+      "ldr q1, [x22], #0x10\n"
       "ldr q0, [x20], #0x10\n"
       "ldr q31, [x25], #0x10\n"
-      "ldr q30, [x22], #0x10\n"
-      "ldr q29, [x21], #0x10\n"
+      "ldr q30, [x23], #0x10\n"
+      "ldr q29, [x22], #0x10\n"
       "ldr q28, [x20], #0x10\n"
       "ldr q27, [x25], #0x10\n"
-      "ldr q26, [x22], #0x10\n"
-      "ldr q25, [x21], #0x10\n"
+      "ldr q26, [x23], #0x10\n"
+      "ldr q25, [x22], #0x10\n"
       "ldr q24, [x20], #0x10\n"
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "ldr q21, [x22], #0x10\n"
       "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q15, [x23, #0x0]\n"
-      "str q11, [x23, #0x10]\n"
-      "str q7, [x23, #0x20]\n"
-      "str q3, [x23, #0x30]\n"
-      "str q31, [x23, #0x40]\n"
-      "str q27, [x23, #0x50]\n"
-      "str q23, [x23, #0x60]\n"
-      "str q19, [x23, #0x70]\n"
-      "str q14, [x23, #0x80]\n"
-      "str q10, [x23, #0x90]\n"
-      "str q6, [x23, #0xa0]\n"
-      "str q2, [x23, #0xb0]\n"
-      "str q30, [x23, #0xc0]\n"
-      "str q26, [x23, #0xd0]\n"
-      "str q22, [x23, #0xe0]\n"
-      "str q18, [x23, #0xf0]\n"
-      "str q13, [x23, #0x100]\n"
-      "str q9, [x23, #0x110]\n"
-      "str q5, [x23, #0x120]\n"
-      "str q1, [x23, #0x130]\n"
-      "str q29, [x23, #0x140]\n"
-      "str q25, [x23, #0x150]\n"
-      "str q21, [x23, #0x160]\n"
-      "str q17, [x23, #0x170]\n"
-      "str q12, [x23, #0x180]\n"
-      "str q8, [x23, #0x190]\n"
-      "str q4, [x23, #0x1a0]\n"
-      "str q0, [x23, #0x1b0]\n"
-      "str q28, [x23, #0x1c0]\n"
-      "str q24, [x23, #0x1d0]\n"
-      "str q20, [x23, #0x1e0]\n"
-      "str q16, [x23, #0x1f0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "str q15, [x21, #0x0]\n"
+      "str q11, [x21, #0x10]\n"
+      "str q7, [x21, #0x20]\n"
+      "str q3, [x21, #0x30]\n"
+      "str q31, [x21, #0x40]\n"
+      "str q27, [x21, #0x50]\n"
+      "str q23, [x21, #0x60]\n"
+      "str q19, [x21, #0x70]\n"
+      "str q14, [x21, #0x80]\n"
+      "str q10, [x21, #0x90]\n"
+      "str q6, [x21, #0xa0]\n"
+      "str q2, [x21, #0xb0]\n"
+      "str q30, [x21, #0xc0]\n"
+      "str q26, [x21, #0xd0]\n"
+      "str q22, [x21, #0xe0]\n"
+      "str q18, [x21, #0xf0]\n"
+      "str q13, [x21, #0x100]\n"
+      "str q9, [x21, #0x110]\n"
+      "str q5, [x21, #0x120]\n"
+      "str q1, [x21, #0x130]\n"
+      "str q29, [x21, #0x140]\n"
+      "str q25, [x21, #0x150]\n"
+      "str q21, [x21, #0x160]\n"
+      "str q17, [x21, #0x170]\n"
+      "str q12, [x21, #0x180]\n"
+      "str q8, [x21, #0x190]\n"
+      "str q4, [x21, #0x1a0]\n"
+      "str q0, [x21, #0x1b0]\n"
+      "str q28, [x21, #0x1c0]\n"
+      "str q24, [x21, #0x1d0]\n"
+      "str q20, [x21, #0x1e0]\n"
+      "str q16, [x21, #0x1f0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x10\n"
-      "movi v16.4s, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "str q16, [x23, #0x80]\n"
-      "str q16, [x23, #0x90]\n"
-      "str q16, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "str q16, [x23, #0xc0]\n"
-      "str q16, [x23, #0xd0]\n"
-      "str q16, [x23, #0xe0]\n"
-      "str q16, [x23, #0xf0]\n"
-      "str q16, [x23, #0x100]\n"
-      "str q16, [x23, #0x110]\n"
-      "str q16, [x23, #0x120]\n"
-      "str q16, [x23, #0x130]\n"
-      "str q16, [x23, #0x140]\n"
-      "str q16, [x23, #0x150]\n"
-      "str q16, [x23, #0x160]\n"
-      "str q16, [x23, #0x170]\n"
-      "str q16, [x23, #0x180]\n"
-      "str q16, [x23, #0x190]\n"
-      "str q16, [x23, #0x1a0]\n"
-      "str q16, [x23, #0x1b0]\n"
-      "str q16, [x23, #0x1c0]\n"
-      "str q16, [x23, #0x1d0]\n"
-      "str q16, [x23, #0x1e0]\n"
-      "str q16, [x23, #0x1f0]\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 16 loop: loop
       "ldr q31, [x25], #0x10\n"
-      "ldr q30, [x22], #0x10\n"
+      "ldr q30, [x23], #0x10\n"
       "sub x24, x24, #0x10\n"
-      "ldr q29, [x21], #0x10\n"
-      "ldr q28, [x20], #0x10\n"
       "cmp x24, #0x10\n"
+      "ldr q29, [x22], #0x10\n"
+      "ldr q28, [x20], #0x10\n"
       "ldr q27, [x25], #0x10\n"
-      "ldr q26, [x22], #0x10\n"
-      "ldr q25, [x21], #0x10\n"
+      "ldr q26, [x23], #0x10\n"
+      "ldr q25, [x22], #0x10\n"
       "ldr q24, [x20], #0x10\n"
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "ldr q21, [x22], #0x10\n"
       "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q31, [x23, #0x0]\n"
-      "str q27, [x23, #0x10]\n"
-      "str q23, [x23, #0x20]\n"
-      "str q19, [x23, #0x30]\n"
-      "str q30, [x23, #0x80]\n"
-      "str q26, [x23, #0x90]\n"
-      "str q22, [x23, #0xa0]\n"
-      "str q18, [x23, #0xb0]\n"
-      "str q29, [x23, #0x100]\n"
-      "str q25, [x23, #0x110]\n"
-      "str q21, [x23, #0x120]\n"
-      "str q17, [x23, #0x130]\n"
-      "str q28, [x23, #0x180]\n"
-      "str q24, [x23, #0x190]\n"
-      "str q20, [x23, #0x1a0]\n"
-      "str q16, [x23, #0x1b0]\n"
-      "add x23, x23, #0x40\n"
+      "str q31, [x21, #0x0]\n"
+      "str q27, [x21, #0x10]\n"
+      "str q23, [x21, #0x20]\n"
+      "str q19, [x21, #0x30]\n"
+      "str q30, [x21, #0x80]\n"
+      "str q26, [x21, #0x90]\n"
+      "str q22, [x21, #0xa0]\n"
+      "str q18, [x21, #0xb0]\n"
+      "str q29, [x21, #0x100]\n"
+      "str q25, [x21, #0x110]\n"
+      "str q21, [x21, #0x120]\n"
+      "str q17, [x21, #0x130]\n"
+      "str q28, [x21, #0x180]\n"
+      "str q24, [x21, #0x190]\n"
+      "str q20, [x21, #0x1a0]\n"
+      "str q16, [x21, #0x1b0]\n"
+      "add x21, x21, #0x40\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 16 loop: skip
       "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
       "sub x24, x24, #0x4\n"
-      "ldr q17, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
       "cmp x24, #0x4\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x80]\n"
-      "str q17, [x23, #0x100]\n"
-      "str q16, [x23, #0x180]\n"
-      "add x23, x23, #0x10\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x80]\n"
+      "str q17, [x21, #0x100]\n"
+      "str q16, [x21, #0x180]\n"
+      "add x21, x21, #0x10\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr s19, [x25], #0x4\n"
-      "ldr s18, [x22], #0x4\n"
+      "ldr s18, [x23], #0x4\n"
       "sub x24, x24, #0x1\n"
-      "ldr s17, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
       "cmp x24, #0x1\n"
-      "str s19, [x23, #0x0]\n"
-      "str s18, [x23, #0x80]\n"
-      "str s17, [x23, #0x100]\n"
-      "str s16, [x23, #0x180]\n"
-      "add x23, x23, #0x4\n"
+      "ldr s17, [x22], #0x4\n"
+      "ldr s16, [x20], #0x4\n"
+      "str s19, [x21, #0x0]\n"
+      "str s18, [x21, #0x80]\n"
+      "str s17, [x21, #0x100]\n"
+      "str s16, [x21, #0x180]\n"
+      "add x21, x21, #0x4\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0x200\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x20\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Column loop
       "ldr q23, [x25], #0x10\n"
-      "sub x20, x20, #0x20\n"
       "ldr q22, [x25], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
+      "sub x20, x20, #0x20\n"
       "cmp x20, #0x20\n"
+      "ldr q21, [x25], #0x10\n"
       "ldr q20, [x25], #0x10\n"
       "ldr q19, [x25], #0x10\n"
       "ldr q18, [x25], #0x10\n"
       "ldr q17, [x25], #0x10\n"
       "ldr q16, [x25], #0x10\n"
-      "str q23, [x23, #0x0]\n"
-      "str q22, [x23, #0x10]\n"
-      "str q21, [x23, #0x20]\n"
-      "str q20, [x23, #0x30]\n"
-      "str q19, [x23, #0x40]\n"
-      "str q18, [x23, #0x50]\n"
-      "str q17, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q23, [x21, #0x0]\n"
+      "str q22, [x21, #0x10]\n"
+      "str q21, [x21, #0x20]\n"
+      "str q20, [x21, #0x30]\n"
+      "str q19, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Column loop skip
       "cmp x20, #0x10\n"
-      "movi v16.4s, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 16 loop: loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 16 loop: loop
       "ldr q19, [x25], #0x10\n"
-      "sub x20, x20, #0x10\n"
       "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
+      "sub x20, x20, #0x10\n"
       "cmp x20, #0x10\n"
+      "ldr q17, [x25], #0x10\n"
       "ldr q16, [x25], #0x10\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x10]\n"
-      "str q17, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "add x23, x23, #0x40\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 16 loop: skip
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, #0x40\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 16 loop: skip
       "cmp x20, #0x4\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr q16, [x25], #0x10\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
-      "str q16, [x23, #0x0]\n"
-      "add x23, x23, #0x10\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr s16, [x25], #0x4\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
-      "str s16, [x23, #0x0]\n"
-      "add x23, x23, #0x4\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x80\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_1x4.hpp
index 7d0460b3a0..ddd426e949 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_1x4.hpp
@@ -40,377 +40,362 @@ void a64_transpose_interleave_12_1x4(uint8_t *out, const uint8_t *in, size_t wid
 
     __asm__ __volatile__(
       "cmp %x[height], #0x8\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "cmp x28, #0x30\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x30\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q19, [x9], #0x10\n"
-      "ldr q18, [x26], #0x10\n"
-      "sub x28, x28, #0x30\n"
-      "ldr q17, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "cmp x28, #0x30\n"
-      "ldr q27, [x23], #0x10\n"
-      "ldr q26, [x22], #0x10\n"
-      "ldr q25, [x21], #0x10\n"
-      "ldr q24, [x20], #0x10\n"
-      "ldr q23, [x9], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
-      "zip1 v1.16b, v19.16b, v17.16b\n"
-      "zip1 v0.16b, v18.16b, v16.16b\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "zip2 v15.16b, v19.16b, v17.16b\n"
-      "zip2 v14.16b, v18.16b, v16.16b\n"
-      "ldr q19, [x23], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "zip1 v13.16b, v27.16b, v25.16b\n"
-      "zip1 v12.16b, v26.16b, v24.16b\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "sub x24, x24, #0x30\n"
+      "cmp x24, #0x30\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v31.16b, v21.16b, v17.16b\n"
+      "zip1 v22.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v14.16b, v21.16b, v17.16b\n"
+      "zip2 v13.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip2 v11.16b, v27.16b, v25.16b\n"
-      "zip2 v10.16b, v26.16b, v24.16b\n"
-      "ldr q9, [x9], #0x10\n"
-      "ldr q8, [x26], #0x10\n"
-      "zip1 v7.16b, v23.16b, v21.16b\n"
-      "zip1 v6.16b, v22.16b, v20.16b\n"
-      "ldr q31, [x25], #0x10\n"
-      "ldr q30, [x24], #0x10\n"
-      "zip2 v5.16b, v23.16b, v21.16b\n"
-      "zip2 v4.16b, v22.16b, v20.16b\n"
-      "ldr q29, [x23], #0x10\n"
-      "ldr q28, [x22], #0x10\n"
-      "zip1 v27.16b, v19.16b, v17.16b\n"
-      "zip1 v26.16b, v18.16b, v16.16b\n"
-      "ldr q25, [x21], #0x10\n"
-      "ldr q24, [x20], #0x10\n"
-      "zip2 v23.16b, v19.16b, v17.16b\n"
-      "zip2 v22.16b, v18.16b, v16.16b\n"
-      "zip1 v3.16b, v9.16b, v31.16b\n"
-      "zip1 v2.16b, v8.16b, v30.16b\n"
-      "zip1 v21.16b, v1.16b, v0.16b\n"
-      "zip2 v20.16b, v1.16b, v0.16b\n"
-      "zip1 v1.16b, v29.16b, v25.16b\n"
-      "zip1 v0.16b, v28.16b, v24.16b\n"
-      "zip1 v19.16b, v15.16b, v14.16b\n"
-      "zip1 v18.16b, v13.16b, v12.16b\n"
-      "zip2 v17.16b, v13.16b, v12.16b\n"
-      "zip1 v16.16b, v11.16b, v10.16b\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "zip2 v31.16b, v9.16b, v31.16b\n"
-      "zip2 v30.16b, v8.16b, v30.16b\n"
-      "str q19, [x27, #0x20]\n"
-      "zip2 v29.16b, v29.16b, v25.16b\n"
-      "zip2 v28.16b, v28.16b, v24.16b\n"
-      "str q18, [x27, #0x30]\n"
-      "zip2 v21.16b, v15.16b, v14.16b\n"
-      "zip1 v20.16b, v7.16b, v6.16b\n"
-      "str q17, [x27, #0x40]\n"
-      "zip2 v19.16b, v7.16b, v6.16b\n"
-      "zip2 v18.16b, v11.16b, v10.16b\n"
-      "str q16, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 v17.16b, v27.16b, v26.16b\n"
-      "zip2 v16.16b, v27.16b, v26.16b\n"
-      "str q21, [x27, #0x0]\n"
-      "zip1 v27.16b, v5.16b, v4.16b\n"
-      "zip2 v26.16b, v5.16b, v4.16b\n"
-      "str q20, [x27, #0x10]\n"
-      "zip1 v25.16b, v3.16b, v2.16b\n"
-      "zip1 v24.16b, v23.16b, v22.16b\n"
-      "str q19, [x27, #0x20]\n"
-      "zip2 v23.16b, v23.16b, v22.16b\n"
-      "zip1 v22.16b, v1.16b, v0.16b\n"
-      "str q18, [x27, #0x30]\n"
-      "zip2 v21.16b, v3.16b, v2.16b\n"
-      "zip1 v20.16b, v31.16b, v30.16b\n"
-      "str q17, [x27, #0x40]\n"
-      "zip2 v19.16b, v31.16b, v30.16b\n"
-      "zip2 v18.16b, v1.16b, v0.16b\n"
-      "str q16, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 v17.16b, v29.16b, v28.16b\n"
-      "zip2 v16.16b, v29.16b, v28.16b\n"
-      "str q27, [x27, #0x0]\n"
-      "str q26, [x27, #0x10]\n"
-      "str q25, [x27, #0x20]\n"
-      "str q24, [x27, #0x30]\n"
-      "str q23, [x27, #0x40]\n"
-      "str q22, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x20]\n"
-      "str q18, [x27, #0x30]\n"
-      "str q17, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 v30.16b, v19.16b, v17.16b\n"
+      "zip1 v29.16b, v18.16b, v16.16b\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v12.16b, v19.16b, v17.16b\n"
+      "zip2 v11.16b, v18.16b, v16.16b\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v10.16b, v21.16b, v17.16b\n"
+      "zip1 v9.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v8.16b, v21.16b, v17.16b\n"
+      "zip2 v7.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v6.16b, v19.16b, v17.16b\n"
+      "zip1 v5.16b, v18.16b, v16.16b\n"
+      "ldr q28, [x9], #0x10\n"
+      "ldr q27, [x28], #0x10\n"
+      "zip2 v4.16b, v19.16b, v17.16b\n"
+      "zip2 v3.16b, v18.16b, v16.16b\n"
+      "ldr q26, [x27], #0x10\n"
+      "ldr q25, [x26], #0x10\n"
+      "zip1 v2.16b, v28.16b, v26.16b\n"
+      "zip1 v1.16b, v27.16b, v25.16b\n"
+      "ldr q24, [x25], #0x10\n"
+      "ldr q23, [x23], #0x10\n"
+      "zip1 v16.16b, v31.16b, v22.16b\n"
+      "zip2 v22.16b, v31.16b, v22.16b\n"
+      "ldr q21, [x22], #0x10\n"
+      "ldr q20, [x20], #0x10\n"
+      "zip1 v0.16b, v24.16b, v21.16b\n"
+      "zip1 v31.16b, v23.16b, v20.16b\n"
+      "zip1 v19.16b, v14.16b, v13.16b\n"
+      "zip1 v18.16b, v30.16b, v29.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "zip2 v16.16b, v30.16b, v29.16b\n"
+      "zip1 v17.16b, v12.16b, v11.16b\n"
+      "str q22, [x21, #0x10]\n"
+      "str q19, [x21, #0x20]\n"
+      "zip2 v30.16b, v28.16b, v26.16b\n"
+      "zip2 v29.16b, v27.16b, v25.16b\n"
+      "str q18, [x21, #0x30]\n"
+      "zip2 v28.16b, v24.16b, v21.16b\n"
+      "zip2 v27.16b, v23.16b, v20.16b\n"
+      "str q16, [x21, #0x40]\n"
+      "zip2 v21.16b, v14.16b, v13.16b\n"
+      "zip1 v16.16b, v10.16b, v9.16b\n"
+      "str q17, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip2 v20.16b, v10.16b, v9.16b\n"
+      "zip2 v19.16b, v12.16b, v11.16b\n"
+      "zip1 v18.16b, v6.16b, v5.16b\n"
+      "zip2 v17.16b, v6.16b, v5.16b\n"
+      "str q21, [x21, #0x0]\n"
+      "str q16, [x21, #0x10]\n"
+      "zip1 v16.16b, v8.16b, v7.16b\n"
+      "zip2 v26.16b, v8.16b, v7.16b\n"
+      "str q20, [x21, #0x20]\n"
+      "zip1 v25.16b, v2.16b, v1.16b\n"
+      "zip1 v24.16b, v4.16b, v3.16b\n"
+      "str q19, [x21, #0x30]\n"
+      "zip2 v23.16b, v4.16b, v3.16b\n"
+      "zip1 v22.16b, v0.16b, v31.16b\n"
+      "str q18, [x21, #0x40]\n"
+      "zip2 v21.16b, v2.16b, v1.16b\n"
+      "zip1 v20.16b, v30.16b, v29.16b\n"
+      "str q17, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip2 v19.16b, v30.16b, v29.16b\n"
+      "zip2 v18.16b, v0.16b, v31.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v17.16b, v28.16b, v27.16b\n"
+      "zip2 v16.16b, v28.16b, v27.16b\n"
+      "str q26, [x21, #0x10]\n"
+      "str q25, [x21, #0x20]\n"
+      "str q24, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q22, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q21, [x21, #0x0]\n"
+      "str q20, [x21, #0x10]\n"
+      "str q19, [x21, #0x20]\n"
+      "str q18, [x21, #0x30]\n"
+      "str q17, [x21, #0x40]\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x28, #0xc\n"
+      "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr d19, [x9], #0x8\n"
-      "ldr d23, [x26], #0x8\n"
-      "sub x28, x28, #0xc\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
-      "cmp x28, #0xc\n"
-      "ldr d22, [x23], #0x8\n"
-      "ldr d25, [x22], #0x8\n"
-      "ldr d21, [x21], #0x8\n"
-      "ldr d16, [x20], #0x8\n"
-      "ld1 { v19.s }[2], [x9], #0x4\n"
-      "ld1 { v23.s }[2], [x26], #0x4\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v17.s }[2], [x24], #0x4\n"
-      "ld1 { v22.s }[2], [x23], #0x4\n"
-      "ld1 { v25.s }[2], [x22], #0x4\n"
-      "ld1 { v21.s }[2], [x21], #0x4\n"
-      "ld1 { v16.s }[2], [x20], #0x4\n"
-      "zip1 v24.16b, v19.16b, v18.16b\n"
-      "zip1 v20.16b, v23.16b, v17.16b\n"
-      "zip2 v19.16b, v19.16b, v18.16b\n"
-      "zip2 v18.16b, v23.16b, v17.16b\n"
-      "zip1 v23.16b, v22.16b, v21.16b\n"
-      "zip1 v17.16b, v25.16b, v16.16b\n"
-      "zip2 v22.16b, v22.16b, v21.16b\n"
-      "zip2 v16.16b, v25.16b, v16.16b\n"
-      "zip1 v21.16b, v24.16b, v20.16b\n"
-      "zip2 v20.16b, v24.16b, v20.16b\n"
+      "ldr d23, [x9], #0x8\n"
+      "ldr d22, [x28], #0x8\n"
+      "sub x24, x24, #0xc\n"
+      "cmp x24, #0xc\n"
+      "ldr d19, [x27], #0x8\n"
+      "ldr d18, [x26], #0x8\n"
+      "ldr d21, [x25], #0x8\n"
+      "ldr d25, [x23], #0x8\n"
+      "ldr d20, [x22], #0x8\n"
+      "ldr d17, [x20], #0x8\n"
+      "ld1 { v23.s }[2], [x9], #0x4\n"
+      "ld1 { v22.s }[2], [x28], #0x4\n"
+      "ld1 { v19.s }[2], [x27], #0x4\n"
+      "ld1 { v18.s }[2], [x26], #0x4\n"
+      "zip1 v24.16b, v23.16b, v19.16b\n"
+      "zip1 v16.16b, v22.16b, v18.16b\n"
+      "ld1 { v21.s }[2], [x25], #0x4\n"
+      "ld1 { v25.s }[2], [x23], #0x4\n"
+      "zip2 v19.16b, v23.16b, v19.16b\n"
+      "zip2 v18.16b, v22.16b, v18.16b\n"
+      "ld1 { v20.s }[2], [x22], #0x4\n"
+      "ld1 { v17.s }[2], [x20], #0x4\n"
+      "zip1 v23.16b, v21.16b, v20.16b\n"
+      "zip1 v22.16b, v25.16b, v17.16b\n"
+      "zip2 v21.16b, v21.16b, v20.16b\n"
+      "zip2 v20.16b, v25.16b, v17.16b\n"
+      "zip1 v17.16b, v24.16b, v16.16b\n"
+      "zip2 v16.16b, v24.16b, v16.16b\n"
+      "str q17, [x21, #0x0]\n"
       "zip1 v19.16b, v19.16b, v18.16b\n"
-      "zip1 v18.16b, v23.16b, v17.16b\n"
-      "zip2 v17.16b, v23.16b, v17.16b\n"
-      "zip1 v16.16b, v22.16b, v16.16b\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x20]\n"
-      "str q18, [x27, #0x30]\n"
-      "str q17, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 v18.16b, v23.16b, v22.16b\n"
+      "str q16, [x21, #0x10]\n"
+      "zip2 v17.16b, v23.16b, v22.16b\n"
+      "zip1 v16.16b, v21.16b, v20.16b\n"
+      "str q19, [x21, #0x20]\n"
+      "str q18, [x21, #0x30]\n"
+      "str q17, [x21, #0x40]\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x28, 10f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
+      "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr s23, [x9], #0x4\n"
-      "ldr s22, [x26], #0x4\n"
-      "sub x28, x28, #0x4\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s17, [x24], #0x4\n"
-      "cmp x28, #0x4\n"
-      "ldr s21, [x23], #0x4\n"
-      "ldr s20, [x22], #0x4\n"
-      "ldr s18, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
-      "zip1 v19.16b, v23.16b, v19.16b\n"
-      "zip1 v17.16b, v22.16b, v17.16b\n"
-      "zip1 v18.16b, v21.16b, v18.16b\n"
-      "zip1 v16.16b, v20.16b, v16.16b\n"
+      "ldr s19, [x9], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
-      "str q17, [x27, #0x0]\n"
-      "str q16, [x27, #0x30]\n"
-      "add x27, x27, #0x10\n"
+      "ldr s20, [x25], #0x4\n"
+      "ldr s19, [x23], #0x4\n"
+      "zip1 v18.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x22], #0x4\n"
+      "ldr s16, [x20], #0x4\n"
+      "zip1 v17.16b, v20.16b, v17.16b\n"
+      "zip1 v16.16b, v19.16b, v16.16b\n"
+      "str q18, [x21, #0x0]\n"
+      "zip1 v16.16b, v17.16b, v16.16b\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, #0x10\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr b23, [x9], #0x1\n"
-      "ldr b22, [x26], #0x1\n"
-      "sub x28, x28, #0x1\n"
-      "ldr b19, [x25], #0x1\n"
-      "ldr b17, [x24], #0x1\n"
-      "cmp x28, #0x1\n"
-      "ldr b21, [x23], #0x1\n"
-      "ldr b20, [x22], #0x1\n"
-      "ldr b18, [x21], #0x1\n"
-      "ldr b16, [x20], #0x1\n"
-      "zip1 v19.16b, v23.16b, v19.16b\n"
-      "zip1 v17.16b, v22.16b, v17.16b\n"
-      "zip1 v18.16b, v21.16b, v18.16b\n"
-      "zip1 v16.16b, v20.16b, v16.16b\n"
+      "ldr b19, [x9], #0x1\n"
+      "ldr b18, [x28], #0x1\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr b17, [x27], #0x1\n"
+      "ldr b16, [x26], #0x1\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
-      "str s17, [x27, #0x0]\n"
-      "str s16, [x27, #0x30]\n"
-      "add x27, x27, #0x4\n"
+      "ldr b20, [x25], #0x1\n"
+      "ldr b19, [x23], #0x1\n"
+      "zip1 v18.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x22], #0x1\n"
+      "ldr b16, [x20], #0x1\n"
+      "zip1 v17.16b, v20.16b, v17.16b\n"
+      "zip1 v16.16b, v19.16b, v16.16b\n"
+      "str s18, [x21, #0x0]\n"
+      "zip1 v16.16b, v17.16b, v16.16b\n"
+      "str s16, [x21, #0x30]\n"
+      "add x21, x21, #0x4\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x8\n"
       "add %x[out], %x[out], #0x60\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x9, %x[in]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x26, x27, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x9, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "csel x25, x25, %x[pad_row], GE\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x26, %x[in_stride]\n"
       "csel x26, x26, %x[pad_row], GT\n"
+      "csel x27, x27, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x28, x28, %x[pad_row], GT\n"
       "cmp x20, #0x30\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
-      "ldr q27, [x9], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
       "sub x20, x20, #0x30\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x30\n"
-      "ldr q26, [x9], #0x10\n"
-      "ldr q25, [x26], #0x10\n"
-      "ldr q20, [x25], #0x10\n"
-      "ldr q19, [x24], #0x10\n"
-      "ldr q24, [x9], #0x10\n"
-      "zip1 v18.16b, v27.16b, v21.16b\n"
-      "zip1 v16.16b, v22.16b, v17.16b\n"
-      "ldr q30, [x26], #0x10\n"
-      "ldr q23, [x25], #0x10\n"
-      "zip2 v29.16b, v27.16b, v21.16b\n"
-      "zip2 v28.16b, v22.16b, v17.16b\n"
-      "ldr q17, [x24], #0x10\n"
-      "zip1 v22.16b, v26.16b, v20.16b\n"
-      "zip1 v21.16b, v25.16b, v19.16b\n"
-      "zip2 v27.16b, v26.16b, v20.16b\n"
-      "zip2 v20.16b, v25.16b, v19.16b\n"
-      "zip1 v26.16b, v24.16b, v23.16b\n"
-      "zip1 v25.16b, v30.16b, v17.16b\n"
-      "zip1 v19.16b, v18.16b, v16.16b\n"
-      "zip2 v16.16b, v18.16b, v16.16b\n"
-      "zip1 v18.16b, v29.16b, v28.16b\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
-      "zip2 v23.16b, v30.16b, v17.16b\n"
-      "zip2 v17.16b, v29.16b, v28.16b\n"
-      "str q19, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "zip1 v16.16b, v22.16b, v21.16b\n"
-      "zip2 v22.16b, v22.16b, v21.16b\n"
-      "str q18, [x27, #0x20]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 v21.16b, v27.16b, v20.16b\n"
-      "zip2 v20.16b, v27.16b, v20.16b\n"
-      "str q17, [x27, #0x0]\n"
-      "zip1 v19.16b, v26.16b, v25.16b\n"
-      "zip2 v18.16b, v26.16b, v25.16b\n"
-      "str q16, [x27, #0x10]\n"
-      "zip1 v17.16b, v24.16b, v23.16b\n"
-      "zip2 v16.16b, v24.16b, v23.16b\n"
-      "str q22, [x27, #0x20]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x20]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q18, [x27, #0x0]\n"
-      "str q17, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v31.16b, v21.16b, v17.16b\n"
+      "zip1 v30.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v29.16b, v21.16b, v17.16b\n"
+      "zip2 v28.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v27.16b, v19.16b, v17.16b\n"
+      "zip1 v26.16b, v18.16b, v16.16b\n"
+      "ldr q22, [x9], #0x10\n"
+      "ldr q21, [x28], #0x10\n"
+      "zip2 v25.16b, v19.16b, v17.16b\n"
+      "zip2 v20.16b, v18.16b, v16.16b\n"
+      "ldr q19, [x27], #0x10\n"
+      "ldr q18, [x26], #0x10\n"
+      "zip1 v24.16b, v22.16b, v19.16b\n"
+      "zip1 v23.16b, v21.16b, v18.16b\n"
+      "zip1 v16.16b, v31.16b, v30.16b\n"
+      "zip2 v17.16b, v31.16b, v30.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v16.16b, v29.16b, v28.16b\n"
+      "str q17, [x21, #0x10]\n"
+      "zip2 v22.16b, v22.16b, v19.16b\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip2 v21.16b, v21.16b, v18.16b\n"
+      "zip2 v18.16b, v29.16b, v28.16b\n"
+      "zip1 v16.16b, v27.16b, v26.16b\n"
+      "zip2 v17.16b, v27.16b, v26.16b\n"
+      "str q18, [x21, #0x0]\n"
+      "str q16, [x21, #0x10]\n"
+      "zip1 v16.16b, v25.16b, v20.16b\n"
+      "zip2 v20.16b, v25.16b, v20.16b\n"
+      "str q17, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip1 v19.16b, v24.16b, v23.16b\n"
+      "zip2 v18.16b, v24.16b, v23.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v17.16b, v22.16b, v21.16b\n"
+      "zip2 v16.16b, v22.16b, v21.16b\n"
+      "str q20, [x21, #0x10]\n"
+      "str q19, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q18, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
       "ldr d19, [x9], #0x8\n"
-      "ldr d21, [x26], #0x8\n"
+      "ldr d21, [x28], #0x8\n"
       "sub x20, x20, #0xc\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
       "cmp x20, #0xc\n"
+      "ldr d18, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
       "ld1 { v19.s }[2], [x9], #0x4\n"
-      "ld1 { v21.s }[2], [x26], #0x4\n"
-      "ld1 { v18.s }[2], [x25], #0x4\n"
-      "ld1 { v16.s }[2], [x24], #0x4\n"
+      "ld1 { v21.s }[2], [x28], #0x4\n"
+      "ld1 { v18.s }[2], [x27], #0x4\n"
+      "ld1 { v16.s }[2], [x26], #0x4\n"
       "zip1 v20.16b, v19.16b, v18.16b\n"
       "zip1 v17.16b, v21.16b, v16.16b\n"
       "zip2 v19.16b, v19.16b, v18.16b\n"
-      "zip2 v16.16b, v21.16b, v16.16b\n"
-      "zip1 v18.16b, v20.16b, v17.16b\n"
+      "zip2 v18.16b, v21.16b, v16.16b\n"
+      "zip1 v16.16b, v20.16b, v17.16b\n"
       "zip2 v17.16b, v20.16b, v17.16b\n"
-      "zip1 v16.16b, v19.16b, v16.16b\n"
-      "str q18, [x27, #0x0]\n"
-      "str q17, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v16.16b, v19.16b, v18.16b\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr s19, [x9], #0x4\n"
-      "ldr s18, [x26], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
       "sub x20, x20, #0x4\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s16, [x24], #0x4\n"
       "cmp x20, #0x4\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str q16, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr b19, [x9], #0x1\n"
-      "ldr b18, [x26], #0x1\n"
+      "ldr b18, [x28], #0x1\n"
       "sub x20, x20, #0x1\n"
-      "ldr b17, [x25], #0x1\n"
-      "ldr b16, [x24], #0x1\n"
       "cmp x20, #0x1\n"
+      "ldr b17, [x27], #0x1\n"
+      "ldr b16, [x26], #0x1\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str s16, [x27, #0x0]\n"
-      "add x27, x27, #0x4\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x30\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_1x8.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_1x8.hpp
index 1eb49e290a..4e76689523 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_1x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_1x8.hpp
@@ -41,268 +41,259 @@ void a64_transpose_interleave_12_1x8(uint8_t *out, const uint8_t *in, size_t wid
     __asm__ __volatile__(
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "cmp %x[height], #0x7\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GE\n"
-      "add %x[in], x20, %x[in_stride]\n"
-      "csel x20, x20, %x[pad_row], GT\n"
-      "cmp %x[height], #0x5\n"
+      "cmp %x[height], #0x7\n"
+      "add %x[in], x22, %x[in_stride]\n"
       "csel x22, x22, %x[pad_row], GT\n"
       "csel x23, x23, %x[pad_row], GE\n"
-      "cmp %x[height], #0x3\n"
+      "cmp %x[height], #0x5\n"
+      "mov x21, %x[width]\n"
       "csel x24, x24, %x[pad_row], GT\n"
       "csel x25, x25, %x[pad_row], GE\n"
+      "cmp %x[height], #0x3\n"
+      "csel x26, x26, %x[pad_row], GT\n"
+      "csel x27, x27, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
+      "csel x28, x28, %x[pad_row], GT\n"
+      "cmp x21, #0x30\n"
+      "mov x20, %x[out]\n"
       "sub %x[height], %x[height], #0x8\n"
-      "csel x26, x26, %x[pad_row], GT\n"
-      "cmp x28, #0x30\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q1, [x9], #0x10\n"
-      "ldr q0, [x26], #0x10\n"
-      "sub x28, x28, #0x30\n"
-      "ldr q31, [x25], #0x10\n"
-      "ldr q28, [x24], #0x10\n"
-      "cmp x28, #0x30\n"
-      "ldr q27, [x23], #0x10\n"
-      "ldr q26, [x22], #0x10\n"
-      "ldr q25, [x21], #0x10\n"
-      "ldr q24, [x20], #0x10\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q25, [x28], #0x10\n"
+      "sub x21, x21, #0x30\n"
+      "cmp x21, #0x30\n"
+      "ldr q20, [x27], #0x10\n"
+      "ldr q24, [x26], #0x10\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x24], #0x10\n"
+      "zip1 v7.16b, v21.16b, v19.16b\n"
+      "zip1 v6.16b, v25.16b, v18.16b\n"
+      "ldr q17, [x23], #0x10\n"
+      "ldr q16, [x22], #0x10\n"
+      "zip1 v28.16b, v20.16b, v17.16b\n"
+      "zip1 v27.16b, v24.16b, v16.16b\n"
       "ldr q23, [x9], #0x10\n"
-      "ldr q30, [x26], #0x10\n"
-      "ldr q22, [x25], #0x10\n"
-      "ldr q21, [x24], #0x10\n"
-      "zip1 v15.16b, v1.16b, v27.16b\n"
-      "zip1 v9.16b, v0.16b, v26.16b\n"
-      "ldr q20, [x23], #0x10\n"
-      "ldr q19, [x22], #0x10\n"
-      "zip1 v18.16b, v31.16b, v25.16b\n"
-      "zip1 v29.16b, v28.16b, v24.16b\n"
-      "ldr q17, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "zip2 v14.16b, v1.16b, v27.16b\n"
-      "zip2 v13.16b, v31.16b, v25.16b\n"
-      "ldr q8, [x9], #0x10\n"
-      "ldr q7, [x26], #0x10\n"
-      "zip2 v12.16b, v0.16b, v26.16b\n"
-      "zip2 v6.16b, v28.16b, v24.16b\n"
-      "ldr q5, [x25], #0x10\n"
-      "ldr q4, [x24], #0x10\n"
-      "zip1 v3.16b, v23.16b, v20.16b\n"
-      "zip1 v11.16b, v30.16b, v19.16b\n"
-      "ldr q28, [x23], #0x10\n"
-      "ldr q27, [x22], #0x10\n"
-      "zip1 v2.16b, v22.16b, v17.16b\n"
-      "zip1 v1.16b, v21.16b, v16.16b\n"
-      "ldr q26, [x21], #0x10\n"
-      "ldr q25, [x20], #0x10\n"
-      "zip2 v24.16b, v23.16b, v20.16b\n"
-      "zip2 v23.16b, v22.16b, v17.16b\n"
-      "zip2 v22.16b, v30.16b, v19.16b\n"
-      "zip2 v21.16b, v21.16b, v16.16b\n"
-      "zip1 v0.16b, v8.16b, v28.16b\n"
-      "zip1 v10.16b, v7.16b, v27.16b\n"
-      "zip1 v31.16b, v5.16b, v26.16b\n"
-      "zip1 v30.16b, v4.16b, v25.16b\n"
-      "zip1 v20.16b, v15.16b, v18.16b\n"
-      "zip1 v19.16b, v9.16b, v29.16b\n"
-      "zip2 v18.16b, v15.16b, v18.16b\n"
-      "zip2 v16.16b, v9.16b, v29.16b\n"
-      "zip1 v29.16b, v14.16b, v13.16b\n"
-      "zip1 v17.16b, v12.16b, v6.16b\n"
-      "zip2 v9.16b, v8.16b, v28.16b\n"
-      "zip2 v28.16b, v5.16b, v26.16b\n"
-      "zip2 v8.16b, v7.16b, v27.16b\n"
-      "zip2 v27.16b, v4.16b, v25.16b\n"
-      "zip2 v7.16b, v14.16b, v13.16b\n"
-      "zip2 v6.16b, v12.16b, v6.16b\n"
-      "zip1 v5.16b, v3.16b, v2.16b\n"
-      "zip1 v4.16b, v11.16b, v1.16b\n"
-      "zip2 v3.16b, v3.16b, v2.16b\n"
-      "zip2 v2.16b, v11.16b, v1.16b\n"
-      "zip1 v26.16b, v24.16b, v23.16b\n"
-      "zip1 v25.16b, v22.16b, v21.16b\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
-      "zip2 v23.16b, v22.16b, v21.16b\n"
-      "zip1 v1.16b, v0.16b, v31.16b\n"
-      "zip1 v22.16b, v10.16b, v30.16b\n"
-      "zip1 v21.16b, v20.16b, v19.16b\n"
-      "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip1 v19.16b, v18.16b, v16.16b\n"
-      "zip2 v18.16b, v18.16b, v16.16b\n"
-      "zip1 v16.16b, v29.16b, v17.16b\n"
-      "zip2 v17.16b, v29.16b, v17.16b\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "zip2 v0.16b, v0.16b, v31.16b\n"
-      "zip2 v31.16b, v10.16b, v30.16b\n"
-      "str q19, [x27, #0x20]\n"
-      "zip1 v30.16b, v9.16b, v28.16b\n"
-      "zip1 v29.16b, v8.16b, v27.16b\n"
-      "str q18, [x27, #0x30]\n"
-      "zip2 v28.16b, v9.16b, v28.16b\n"
-      "zip2 v27.16b, v8.16b, v27.16b\n"
-      "str q16, [x27, #0x40]\n"
-      "zip1 v21.16b, v7.16b, v6.16b\n"
-      "zip2 v16.16b, v7.16b, v6.16b\n"
-      "str q17, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 v20.16b, v5.16b, v4.16b\n"
-      "zip2 v19.16b, v5.16b, v4.16b\n"
+      "ldr q22, [x28], #0x10\n"
+      "zip2 v5.16b, v21.16b, v19.16b\n"
+      "zip2 v4.16b, v20.16b, v17.16b\n"
+      "ldr q21, [x27], #0x10\n"
+      "ldr q20, [x26], #0x10\n"
+      "zip2 v3.16b, v25.16b, v18.16b\n"
+      "zip2 v2.16b, v24.16b, v16.16b\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x24], #0x10\n"
+      "zip1 v1.16b, v23.16b, v19.16b\n"
+      "zip1 v15.16b, v22.16b, v18.16b\n"
+      "ldr q17, [x23], #0x10\n"
+      "ldr q16, [x22], #0x10\n"
+      "zip1 v0.16b, v21.16b, v17.16b\n"
+      "zip1 v31.16b, v20.16b, v16.16b\n"
+      "ldr q26, [x9], #0x10\n"
+      "ldr q30, [x28], #0x10\n"
+      "zip2 v14.16b, v23.16b, v19.16b\n"
+      "zip2 v13.16b, v21.16b, v17.16b\n"
+      "ldr q25, [x27], #0x10\n"
+      "ldr q24, [x26], #0x10\n"
+      "zip2 v12.16b, v22.16b, v18.16b\n"
+      "zip2 v11.16b, v20.16b, v16.16b\n"
+      "ldr q23, [x25], #0x10\n"
+      "ldr q22, [x24], #0x10\n"
+      "zip1 v10.16b, v26.16b, v23.16b\n"
+      "zip1 v9.16b, v30.16b, v22.16b\n"
+      "ldr q21, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
+      "zip1 v29.16b, v25.16b, v21.16b\n"
+      "zip1 v8.16b, v24.16b, v17.16b\n"
+      "zip1 v19.16b, v7.16b, v28.16b\n"
+      "zip1 v16.16b, v6.16b, v27.16b\n"
+      "zip2 v28.16b, v7.16b, v28.16b\n"
+      "zip2 v18.16b, v6.16b, v27.16b\n"
+      "zip1 v27.16b, v5.16b, v4.16b\n"
+      "zip1 v20.16b, v3.16b, v2.16b\n"
+      "zip2 v7.16b, v26.16b, v23.16b\n"
+      "zip2 v26.16b, v25.16b, v21.16b\n"
+      "zip2 v6.16b, v30.16b, v22.16b\n"
+      "zip2 v25.16b, v24.16b, v17.16b\n"
+      "zip2 v5.16b, v5.16b, v4.16b\n"
+      "zip2 v4.16b, v3.16b, v2.16b\n"
+      "zip1 v3.16b, v1.16b, v0.16b\n"
+      "zip1 v2.16b, v15.16b, v31.16b\n"
+      "zip2 v1.16b, v1.16b, v0.16b\n"
+      "zip2 v0.16b, v15.16b, v31.16b\n"
+      "zip1 v31.16b, v14.16b, v13.16b\n"
+      "zip1 v30.16b, v12.16b, v11.16b\n"
+      "zip2 v24.16b, v14.16b, v13.16b\n"
+      "zip2 v23.16b, v12.16b, v11.16b\n"
+      "zip1 v22.16b, v10.16b, v29.16b\n"
+      "zip1 v21.16b, v9.16b, v8.16b\n"
+      "zip1 v17.16b, v19.16b, v16.16b\n"
+      "zip2 v16.16b, v19.16b, v16.16b\n"
+      "str q17, [x20, #0x0]\n"
+      "zip1 v19.16b, v28.16b, v18.16b\n"
+      "zip2 v18.16b, v28.16b, v18.16b\n"
+      "str q16, [x20, #0x10]\n"
+      "zip1 v17.16b, v27.16b, v20.16b\n"
+      "zip2 v16.16b, v27.16b, v20.16b\n"
+      "str q19, [x20, #0x20]\n"
+      "str q18, [x20, #0x30]\n"
+      "zip2 v29.16b, v10.16b, v29.16b\n"
+      "zip2 v20.16b, v9.16b, v8.16b\n"
+      "str q17, [x20, #0x40]\n"
+      "zip1 v28.16b, v7.16b, v26.16b\n"
+      "zip1 v27.16b, v6.16b, v25.16b\n"
+      "str q16, [x20, #0x50]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "zip2 v26.16b, v7.16b, v26.16b\n"
+      "zip2 v25.16b, v6.16b, v25.16b\n"
+      "zip1 v17.16b, v5.16b, v4.16b\n"
+      "zip2 v16.16b, v5.16b, v4.16b\n"
+      "str q17, [x20, #0x0]\n"
       "zip1 v18.16b, v3.16b, v2.16b\n"
       "zip2 v17.16b, v3.16b, v2.16b\n"
-      "str q21, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "zip1 v16.16b, v26.16b, v25.16b\n"
-      "zip2 v26.16b, v26.16b, v25.16b\n"
-      "str q20, [x27, #0x20]\n"
-      "zip1 v25.16b, v24.16b, v23.16b\n"
+      "str q16, [x20, #0x10]\n"
+      "zip1 v16.16b, v1.16b, v0.16b\n"
+      "zip2 v19.16b, v1.16b, v0.16b\n"
+      "str q18, [x20, #0x20]\n"
+      "str q17, [x20, #0x30]\n"
+      "zip1 v18.16b, v31.16b, v30.16b\n"
+      "zip2 v17.16b, v31.16b, v30.16b\n"
+      "str q16, [x20, #0x40]\n"
+      "zip1 v16.16b, v24.16b, v23.16b\n"
       "zip2 v24.16b, v24.16b, v23.16b\n"
-      "str q19, [x27, #0x30]\n"
-      "zip1 v23.16b, v1.16b, v22.16b\n"
-      "zip2 v22.16b, v1.16b, v22.16b\n"
-      "str q18, [x27, #0x40]\n"
-      "zip1 v21.16b, v0.16b, v31.16b\n"
-      "zip2 v20.16b, v0.16b, v31.16b\n"
-      "str q17, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 v19.16b, v30.16b, v29.16b\n"
-      "zip2 v18.16b, v30.16b, v29.16b\n"
-      "str q16, [x27, #0x0]\n"
-      "zip1 v17.16b, v28.16b, v27.16b\n"
-      "zip2 v16.16b, v28.16b, v27.16b\n"
-      "str q26, [x27, #0x10]\n"
-      "str q25, [x27, #0x20]\n"
-      "str q24, [x27, #0x30]\n"
-      "str q23, [x27, #0x40]\n"
-      "str q22, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x20]\n"
-      "str q18, [x27, #0x30]\n"
-      "str q17, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "str q19, [x20, #0x50]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "zip1 v23.16b, v22.16b, v21.16b\n"
+      "zip2 v22.16b, v22.16b, v21.16b\n"
+      "str q18, [x20, #0x0]\n"
+      "zip1 v21.16b, v29.16b, v20.16b\n"
+      "zip2 v20.16b, v29.16b, v20.16b\n"
+      "str q17, [x20, #0x10]\n"
+      "zip1 v19.16b, v28.16b, v27.16b\n"
+      "zip2 v18.16b, v28.16b, v27.16b\n"
+      "str q16, [x20, #0x20]\n"
+      "zip1 v17.16b, v26.16b, v25.16b\n"
+      "zip2 v16.16b, v26.16b, v25.16b\n"
+      "str q24, [x20, #0x30]\n"
+      "str q23, [x20, #0x40]\n"
+      "str q22, [x20, #0x50]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "str q21, [x20, #0x0]\n"
+      "str q20, [x20, #0x10]\n"
+      "str q19, [x20, #0x20]\n"
+      "str q18, [x20, #0x30]\n"
+      "str q17, [x20, #0x40]\n"
+      "str q16, [x20, #0x50]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x28, #0xc\n"
+      "cmp x21, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr d22, [x9], #0x8\n"
-      "ldr d23, [x26], #0x8\n"
-      "sub x28, x28, #0xc\n"
-      "ldr d21, [x25], #0x8\n"
-      "ldr d20, [x24], #0x8\n"
-      "cmp x28, #0xc\n"
-      "ldr d18, [x23], #0x8\n"
-      "ldr d17, [x22], #0x8\n"
-      "ldr d19, [x21], #0x8\n"
-      "ldr d16, [x20], #0x8\n"
-      "ld1 { v22.s }[2], [x9], #0x4\n"
-      "ld1 { v23.s }[2], [x26], #0x4\n"
-      "ld1 { v21.s }[2], [x25], #0x4\n"
-      "ld1 { v20.s }[2], [x24], #0x4\n"
-      "ld1 { v18.s }[2], [x23], #0x4\n"
-      "ld1 { v17.s }[2], [x22], #0x4\n"
-      "ld1 { v19.s }[2], [x21], #0x4\n"
-      "ld1 { v16.s }[2], [x20], #0x4\n"
-      "zip1 v25.16b, v22.16b, v18.16b\n"
-      "zip1 v24.16b, v23.16b, v17.16b\n"
-      "zip2 v22.16b, v22.16b, v18.16b\n"
-      "zip2 v23.16b, v23.16b, v17.16b\n"
-      "zip1 v18.16b, v21.16b, v19.16b\n"
-      "zip1 v17.16b, v20.16b, v16.16b\n"
-      "zip2 v21.16b, v21.16b, v19.16b\n"
-      "zip2 v16.16b, v20.16b, v16.16b\n"
-      "zip1 v20.16b, v25.16b, v18.16b\n"
-      "zip1 v19.16b, v24.16b, v17.16b\n"
-      "zip2 v18.16b, v25.16b, v18.16b\n"
-      "zip2 v17.16b, v24.16b, v17.16b\n"
-      "zip1 v22.16b, v22.16b, v21.16b\n"
-      "zip1 v16.16b, v23.16b, v16.16b\n"
-      "zip1 v21.16b, v20.16b, v19.16b\n"
-      "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip1 v19.16b, v18.16b, v17.16b\n"
-      "zip2 v18.16b, v18.16b, v17.16b\n"
-      "zip1 v17.16b, v22.16b, v16.16b\n"
-      "zip2 v16.16b, v22.16b, v16.16b\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x20]\n"
-      "str q18, [x27, #0x30]\n"
-      "str q17, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "ldr d23, [x9], #0x8\n"
+      "ldr d27, [x28], #0x8\n"
+      "sub x21, x21, #0xc\n"
+      "cmp x21, #0xc\n"
+      "ldr d21, [x27], #0x8\n"
+      "ldr d26, [x26], #0x8\n"
+      "ldr d20, [x25], #0x8\n"
+      "ldr d19, [x24], #0x8\n"
+      "ldr d17, [x23], #0x8\n"
+      "ldr d16, [x22], #0x8\n"
+      "ld1 { v23.s }[2], [x9], #0x4\n"
+      "ld1 { v27.s }[2], [x28], #0x4\n"
+      "ld1 { v21.s }[2], [x27], #0x4\n"
+      "ld1 { v26.s }[2], [x26], #0x4\n"
+      "ld1 { v20.s }[2], [x25], #0x4\n"
+      "ld1 { v19.s }[2], [x24], #0x4\n"
+      "zip1 v25.16b, v23.16b, v20.16b\n"
+      "zip1 v24.16b, v27.16b, v19.16b\n"
+      "ld1 { v17.s }[2], [x23], #0x4\n"
+      "ld1 { v16.s }[2], [x22], #0x4\n"
+      "zip1 v22.16b, v21.16b, v17.16b\n"
+      "zip1 v18.16b, v26.16b, v16.16b\n"
+      "zip2 v23.16b, v23.16b, v20.16b\n"
+      "zip2 v21.16b, v21.16b, v17.16b\n"
+      "zip2 v20.16b, v27.16b, v19.16b\n"
+      "zip2 v17.16b, v26.16b, v16.16b\n"
+      "zip1 v19.16b, v25.16b, v22.16b\n"
+      "zip1 v16.16b, v24.16b, v18.16b\n"
+      "zip2 v22.16b, v25.16b, v22.16b\n"
+      "zip2 v18.16b, v24.16b, v18.16b\n"
+      "zip1 v21.16b, v23.16b, v21.16b\n"
+      "zip1 v20.16b, v20.16b, v17.16b\n"
+      "zip1 v17.16b, v19.16b, v16.16b\n"
+      "zip2 v16.16b, v19.16b, v16.16b\n"
+      "str q17, [x20, #0x0]\n"
+      "zip1 v19.16b, v22.16b, v18.16b\n"
+      "zip2 v18.16b, v22.16b, v18.16b\n"
+      "str q16, [x20, #0x10]\n"
+      "zip1 v17.16b, v21.16b, v20.16b\n"
+      "zip2 v16.16b, v21.16b, v20.16b\n"
+      "str q19, [x20, #0x20]\n"
+      "str q18, [x20, #0x30]\n"
+      "str q17, [x20, #0x40]\n"
+      "str q16, [x20, #0x50]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x28, 10f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
+      "cmp x21, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr s23, [x9], #0x4\n"
-      "ldr s22, [x26], #0x4\n"
-      "sub x28, x28, #0x4\n"
-      "ldr s21, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
-      "cmp x28, #0x4\n"
-      "ldr s18, [x23], #0x4\n"
-      "ldr s19, [x22], #0x4\n"
-      "ldr s17, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
-      "zip1 v18.16b, v23.16b, v18.16b\n"
-      "zip1 v19.16b, v22.16b, v19.16b\n"
+      "ldr s18, [x9], #0x4\n"
+      "ldr s19, [x28], #0x4\n"
+      "sub x21, x21, #0x4\n"
+      "cmp x21, #0x4\n"
+      "ldr s21, [x27], #0x4\n"
+      "ldr s20, [x26], #0x4\n"
+      "ldr s17, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "zip1 v18.16b, v18.16b, v17.16b\n"
+      "zip1 v19.16b, v19.16b, v16.16b\n"
+      "ldr s17, [x23], #0x4\n"
+      "ldr s16, [x22], #0x4\n"
       "zip1 v17.16b, v21.16b, v17.16b\n"
       "zip1 v16.16b, v20.16b, v16.16b\n"
       "zip1 v18.16b, v18.16b, v17.16b\n"
       "zip1 v16.16b, v19.16b, v16.16b\n"
       "zip1 v17.16b, v18.16b, v16.16b\n"
       "zip2 v16.16b, v18.16b, v16.16b\n"
-      "str q17, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
+      "str q17, [x20, #0x0]\n"
+      "str q16, [x20, #0x10]\n"
+      "add x20, x20, #0x20\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x21, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr b23, [x9], #0x1\n"
-      "ldr b22, [x26], #0x1\n"
-      "sub x28, x28, #0x1\n"
-      "ldr b21, [x25], #0x1\n"
-      "ldr b20, [x24], #0x1\n"
-      "cmp x28, #0x1\n"
-      "ldr b19, [x23], #0x1\n"
-      "ldr b18, [x22], #0x1\n"
-      "ldr b17, [x21], #0x1\n"
-      "ldr b16, [x20], #0x1\n"
-      "zip1 v19.16b, v23.16b, v19.16b\n"
-      "zip1 v18.16b, v22.16b, v18.16b\n"
+      "ldr b19, [x9], #0x1\n"
+      "ldr b18, [x28], #0x1\n"
+      "sub x21, x21, #0x1\n"
+      "cmp x21, #0x1\n"
+      "ldr b21, [x27], #0x1\n"
+      "ldr b20, [x26], #0x1\n"
+      "ldr b17, [x25], #0x1\n"
+      "ldr b16, [x24], #0x1\n"
+      "zip1 v19.16b, v19.16b, v17.16b\n"
+      "zip1 v18.16b, v18.16b, v16.16b\n"
+      "ldr b17, [x23], #0x1\n"
+      "ldr b16, [x22], #0x1\n"
       "zip1 v17.16b, v21.16b, v17.16b\n"
       "zip1 v16.16b, v20.16b, v16.16b\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str d16, [x27, #0x0]\n"
-      "add x27, x27, #0x8\n"
+      "str d16, [x20, #0x0]\n"
+      "add x20, x20, #0x8\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x60\n"
       "bge 1b\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x2.hpp
index 94726570d4..eafa06ece1 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x2.hpp
@@ -40,308 +40,287 @@ void a64_transpose_interleave_12_2x2(uint16_t *out, const uint16_t *in, size_t w
 
     __asm__ __volatile__(
       "cmp %x[height], #0x8\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "cmp x28, #0x18\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q22, [x9], #0x10\n"
-      "ldr q21, [x26], #0x10\n"
-      "sub x28, x28, #0x18\n"
-      "ldr q20, [x25], #0x10\n"
-      "ldr q19, [x24], #0x10\n"
-      "cmp x28, #0x18\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q23, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "zip1 v8.8h, v22.8h, v21.8h\n"
-      "zip2 v7.8h, v22.8h, v21.8h\n"
-      "ldr q22, [x9], #0x10\n"
-      "ldr q21, [x26], #0x10\n"
-      "zip1 v6.8h, v20.8h, v19.8h\n"
-      "zip2 v5.8h, v20.8h, v19.8h\n"
-      "ldr q20, [x25], #0x10\n"
-      "ldr q19, [x24], #0x10\n"
-      "zip1 v4.8h, v18.8h, v17.8h\n"
-      "zip2 v3.8h, v18.8h, v17.8h\n"
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "sub x24, x24, #0x18\n"
+      "zip1 v10.8h, v19.8h, v18.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip2 v9.8h, v19.8h, v18.8h\n"
+      "zip1 v8.8h, v17.8h, v16.8h\n"
+      "ldr q19, [x25], #0x10\n"
       "ldr q18, [x23], #0x10\n"
+      "zip2 v7.8h, v17.8h, v16.8h\n"
+      "zip1 v6.8h, v19.8h, v18.8h\n"
       "ldr q17, [x22], #0x10\n"
-      "zip1 v2.8h, v23.8h, v16.8h\n"
-      "zip2 v1.8h, v23.8h, v16.8h\n"
-      "ldr q23, [x21], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip1 v0.8h, v22.8h, v21.8h\n"
-      "zip2 v31.8h, v22.8h, v21.8h\n"
-      "ldr q22, [x9], #0x10\n"
-      "ldr q21, [x26], #0x10\n"
-      "zip1 v30.8h, v20.8h, v19.8h\n"
-      "zip2 v29.8h, v20.8h, v19.8h\n"
+      "zip2 v5.8h, v19.8h, v18.8h\n"
+      "zip1 v4.8h, v17.8h, v16.8h\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip1 v3.8h, v21.8h, v18.8h\n"
+      "zip2 v2.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v1.8h, v17.8h, v16.8h\n"
+      "cmp x24, #0x18\n"
       "ldr q20, [x25], #0x10\n"
-      "ldr q19, [x24], #0x10\n"
-      "zip1 v28.8h, v18.8h, v17.8h\n"
-      "zip2 v27.8h, v18.8h, v17.8h\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
-      "zip1 v26.8h, v23.8h, v16.8h\n"
-      "zip2 v25.8h, v23.8h, v16.8h\n"
-      "ldr q24, [x21], #0x10\n"
+      "ldr q19, [x23], #0x10\n"
+      "zip1 v0.8h, v20.8h, v19.8h\n"
+      "zip2 v31.8h, v21.8h, v18.8h\n"
+      "ldr q30, [x22], #0x10\n"
+      "ldr q29, [x20], #0x10\n"
+      "zip1 v28.8h, v30.8h, v29.8h\n"
+      "zip2 v27.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x9], #0x10\n"
+      "ldr q16, [x28], #0x10\n"
+      "zip1 v26.8h, v17.8h, v16.8h\n"
+      "zip2 v25.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v24.8h, v17.8h, v16.8h\n"
+      "zip2 v23.8h, v17.8h, v16.8h\n"
+      "ldr q18, [x25], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
+      "zip2 v22.8h, v20.8h, v19.8h\n"
+      "zip1 v21.8h, v18.8h, v17.8h\n"
+      "ldr q20, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q8, [x27, #0x0]\n"
-      "zip1 v23.8h, v22.8h, v21.8h\n"
-      "str q7, [x27, #0x10]\n"
-      "zip2 v22.8h, v22.8h, v21.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "str q0, [x27, #0x20]\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip1 v19.8h, v18.8h, v17.8h\n"
-      "str q6, [x27, #0x30]\n"
-      "zip2 v18.8h, v18.8h, v17.8h\n"
-      "zip1 v17.8h, v24.8h, v16.8h\n"
-      "str q5, [x27, #0x40]\n"
-      "zip2 v16.8h, v24.8h, v16.8h\n"
-      "str q30, [x27, #0x50]\n"
-      "str q4, [x27, #0x60]\n"
-      "str q3, [x27, #0x70]\n"
-      "str q28, [x27, #0x80]\n"
-      "str q2, [x27, #0x90]\n"
-      "str q1, [x27, #0xa0]\n"
-      "str q26, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q31, [x27, #0x0]\n"
-      "str q23, [x27, #0x10]\n"
-      "str q22, [x27, #0x20]\n"
-      "str q29, [x27, #0x30]\n"
-      "str q21, [x27, #0x40]\n"
-      "str q20, [x27, #0x50]\n"
-      "str q27, [x27, #0x60]\n"
-      "str q19, [x27, #0x70]\n"
-      "str q18, [x27, #0x80]\n"
-      "str q25, [x27, #0x90]\n"
-      "str q17, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "str q10, [x21, #0x0]\n"
+      "zip2 v19.8h, v18.8h, v17.8h\n"
+      "str q9, [x21, #0x10]\n"
+      "zip2 v18.8h, v30.8h, v29.8h\n"
+      "zip1 v17.8h, v20.8h, v16.8h\n"
+      "str q3, [x21, #0x20]\n"
+      "zip2 v16.8h, v20.8h, v16.8h\n"
+      "str q8, [x21, #0x30]\n"
+      "str q7, [x21, #0x40]\n"
+      "str q1, [x21, #0x50]\n"
+      "str q6, [x21, #0x60]\n"
+      "str q5, [x21, #0x70]\n"
+      "str q0, [x21, #0x80]\n"
+      "str q4, [x21, #0x90]\n"
+      "str q2, [x21, #0xa0]\n"
+      "str q28, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q31, [x21, #0x0]\n"
+      "str q26, [x21, #0x10]\n"
+      "str q25, [x21, #0x20]\n"
+      "str q27, [x21, #0x30]\n"
+      "str q24, [x21, #0x40]\n"
+      "str q23, [x21, #0x50]\n"
+      "str q22, [x21, #0x60]\n"
+      "str q21, [x21, #0x70]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q17, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x28, #0xc\n"
+      "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr q23, [x9], #0x10\n"
-      "ldr q19, [x26], #0x10\n"
-      "sub x28, x28, #0xc\n"
-      "ldr q22, [x25], #0x10\n"
-      "ldr q18, [x24], #0x10\n"
-      "cmp x28, #0xc\n"
-      "ldr q21, [x23], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q20, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "zip1 v31.8h, v23.8h, v19.8h\n"
-      "zip2 v30.8h, v23.8h, v19.8h\n"
-      "ldr d29, [x9], #0x8\n"
-      "ldr d19, [x26], #0x8\n"
-      "zip1 v28.8h, v22.8h, v18.8h\n"
-      "zip2 v27.8h, v22.8h, v18.8h\n"
-      "ldr d26, [x25], #0x8\n"
-      "ldr d18, [x24], #0x8\n"
-      "zip1 v25.8h, v21.8h, v17.8h\n"
-      "zip2 v24.8h, v21.8h, v17.8h\n"
-      "ldr d23, [x23], #0x8\n"
+      "ldr q17, [x9], #0x10\n"
+      "ldr q16, [x28], #0x10\n"
+      "sub x24, x24, #0xc\n"
+      "cmp x24, #0xc\n"
+      "ldr q19, [x27], #0x10\n"
+      "ldr q18, [x26], #0x10\n"
+      "zip1 v28.8h, v17.8h, v16.8h\n"
+      "zip2 v27.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip1 v26.8h, v19.8h, v18.8h\n"
+      "zip2 v25.8h, v19.8h, v18.8h\n"
+      "ldr q19, [x22], #0x10\n"
+      "ldr q18, [x20], #0x10\n"
+      "zip1 v24.8h, v17.8h, v16.8h\n"
+      "zip2 v23.8h, v17.8h, v16.8h\n"
+      "ldr d17, [x9], #0x8\n"
+      "ldr d16, [x28], #0x8\n"
+      "zip1 v22.8h, v17.8h, v16.8h\n"
+      "zip1 v21.8h, v19.8h, v18.8h\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
+      "zip1 v20.8h, v17.8h, v16.8h\n"
+      "zip2 v19.8h, v19.8h, v18.8h\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
+      "zip1 v18.8h, v17.8h, v16.8h\n"
       "ldr d17, [x22], #0x8\n"
-      "zip1 v22.8h, v20.8h, v16.8h\n"
-      "zip2 v21.8h, v20.8h, v16.8h\n"
-      "ldr d20, [x21], #0x8\n"
       "ldr d16, [x20], #0x8\n"
-      "zip1 v19.8h, v29.8h, v19.8h\n"
-      "str q31, [x27, #0x0]\n"
-      "zip1 v18.8h, v26.8h, v18.8h\n"
-      "str q30, [x27, #0x10]\n"
-      "zip1 v17.8h, v23.8h, v17.8h\n"
-      "zip1 v16.8h, v20.8h, v16.8h\n"
-      "str q19, [x27, #0x20]\n"
-      "str q28, [x27, #0x30]\n"
-      "str q27, [x27, #0x40]\n"
-      "str q18, [x27, #0x50]\n"
-      "str q25, [x27, #0x60]\n"
-      "str q24, [x27, #0x70]\n"
-      "str q17, [x27, #0x80]\n"
-      "str q22, [x27, #0x90]\n"
-      "str q21, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 v16.8h, v17.8h, v16.8h\n"
+      "str q28, [x21, #0x0]\n"
+      "str q27, [x21, #0x10]\n"
+      "str q22, [x21, #0x20]\n"
+      "str q26, [x21, #0x30]\n"
+      "str q25, [x21, #0x40]\n"
+      "str q20, [x21, #0x50]\n"
+      "str q24, [x21, #0x60]\n"
+      "str q23, [x21, #0x70]\n"
+      "str q18, [x21, #0x80]\n"
+      "str q21, [x21, #0x90]\n"
+      "str q19, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x28, 10f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "str q16, [x27, #0x80]\n"
-      "str q16, [x27, #0x90]\n"
-      "str q16, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
+      "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr d23, [x9], #0x8\n"
-      "ldr d18, [x26], #0x8\n"
-      "sub x28, x28, #0x4\n"
-      "ldr d22, [x25], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
-      "cmp x28, #0x4\n"
-      "ldr d21, [x23], #0x8\n"
+      "ldr d19, [x9], #0x8\n"
+      "ldr d18, [x28], #0x8\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
+      "zip1 v20.8h, v19.8h, v18.8h\n"
+      "zip1 v19.8h, v17.8h, v16.8h\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
+      "zip1 v18.8h, v17.8h, v16.8h\n"
       "ldr d17, [x22], #0x8\n"
-      "ldr d20, [x21], #0x8\n"
-      "ldr d19, [x20], #0x8\n"
-      "zip1 v18.8h, v23.8h, v18.8h\n"
-      "zip1 v16.8h, v22.8h, v16.8h\n"
-      "zip1 v17.8h, v21.8h, v17.8h\n"
-      "str q18, [x27, #0x0]\n"
-      "str q16, [x27, #0x30]\n"
-      "zip1 v16.8h, v20.8h, v19.8h\n"
-      "str q17, [x27, #0x60]\n"
-      "str q16, [x27, #0x90]\n"
-      "add x27, x27, #0x10\n"
+      "ldr d16, [x20], #0x8\n"
+      "str q20, [x21, #0x0]\n"
+      "zip1 v16.8h, v17.8h, v16.8h\n"
+      "str q19, [x21, #0x30]\n"
+      "str q18, [x21, #0x60]\n"
+      "str q16, [x21, #0x90]\n"
+      "add x21, x21, #0x10\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr h23, [x9], #0x2\n"
-      "ldr h18, [x26], #0x2\n"
-      "sub x28, x28, #0x1\n"
-      "ldr h22, [x25], #0x2\n"
-      "ldr h16, [x24], #0x2\n"
-      "cmp x28, #0x1\n"
-      "ldr h21, [x23], #0x2\n"
+      "ldr h19, [x9], #0x2\n"
+      "ldr h18, [x28], #0x2\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr h17, [x27], #0x2\n"
+      "ldr h16, [x26], #0x2\n"
+      "zip1 v20.8h, v19.8h, v18.8h\n"
+      "zip1 v19.8h, v17.8h, v16.8h\n"
+      "ldr h17, [x25], #0x2\n"
+      "ldr h16, [x23], #0x2\n"
+      "zip1 v18.8h, v17.8h, v16.8h\n"
       "ldr h17, [x22], #0x2\n"
-      "ldr h20, [x21], #0x2\n"
-      "ldr h19, [x20], #0x2\n"
-      "zip1 v18.8h, v23.8h, v18.8h\n"
-      "zip1 v16.8h, v22.8h, v16.8h\n"
-      "zip1 v17.8h, v21.8h, v17.8h\n"
-      "str s18, [x27, #0x0]\n"
-      "str s16, [x27, #0x30]\n"
-      "zip1 v16.8h, v20.8h, v19.8h\n"
-      "str s17, [x27, #0x60]\n"
-      "str s16, [x27, #0x90]\n"
-      "add x27, x27, #0x4\n"
+      "ldr h16, [x20], #0x2\n"
+      "str s20, [x21, #0x0]\n"
+      "zip1 v16.8h, v17.8h, v16.8h\n"
+      "str s19, [x21, #0x30]\n"
+      "str s18, [x21, #0x60]\n"
+      "str s16, [x21, #0x90]\n"
+      "add x21, x21, #0x4\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x8\n"
       "add %x[out], %x[out], #0xc0\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x9, %x[in]\n"
       "mov x20, %x[width]\n"
+      "add x28, x9, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x2\n"
-      "add x26, x9, %x[in_stride]\n"
-      "add %x[in], x26, %x[in_stride]\n"
-      "csel x26, x26, %x[pad_row], GT\n"
+      "add %x[in], x28, %x[in_stride]\n"
+      "csel x28, x28, %x[pad_row], GT\n"
       "cmp x20, #0x18\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
-      "ldr q19, [x9], #0x10\n"
-      "ldr q16, [x26], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x2\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
+      "ldr q17, [x9], #0x10\n"
+      "ldr q16, [x28], #0x10\n"
       "sub x20, x20, #0x18\n"
-      "ldr q22, [x9], #0x10\n"
-      "cmp x20, #0x18\n"
-      "ldr q18, [x26], #0x10\n"
+      "zip1 v22.8h, v17.8h, v16.8h\n"
       "ldr q21, [x9], #0x10\n"
-      "ldr q20, [x26], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v17.8h, v17.8h, v16.8h\n"
+      "zip1 v20.8h, v21.8h, v18.8h\n"
+      "ldr q19, [x9], #0x10\n"
+      "ldr q16, [x28], #0x10\n"
+      "str q22, [x21, #0x0]\n"
+      "cmp x20, #0x18\n"
+      "str q17, [x21, #0x10]\n"
+      "zip2 v18.8h, v21.8h, v18.8h\n"
       "zip1 v17.8h, v19.8h, v16.8h\n"
+      "str q20, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
       "zip2 v16.8h, v19.8h, v16.8h\n"
-      "zip1 v19.8h, v22.8h, v18.8h\n"
-      "zip2 v18.8h, v22.8h, v18.8h\n"
-      "str q17, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "zip1 v17.8h, v21.8h, v20.8h\n"
-      "zip2 v16.8h, v21.8h, v20.8h\n"
-      "str q19, [x27, #0x20]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q18, [x27, #0x0]\n"
-      "str q17, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "str q18, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
       "ldr q20, [x9], #0x10\n"
-      "ldr q17, [x26], #0x10\n"
+      "ldr q17, [x28], #0x10\n"
       "sub x20, x20, #0xc\n"
-      "ldr d19, [x9], #0x8\n"
       "cmp x20, #0xc\n"
-      "ldr d16, [x26], #0x8\n"
+      "ldr d19, [x9], #0x8\n"
+      "ldr d16, [x28], #0x8\n"
       "zip1 v18.8h, v20.8h, v17.8h\n"
       "zip2 v17.8h, v20.8h, v17.8h\n"
       "zip1 v16.8h, v19.8h, v16.8h\n"
-      "str q18, [x27, #0x0]\n"
-      "str q17, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q18, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr d17, [x9], #0x8\n"
-      "ldr d16, [x26], #0x8\n"
+      "ldr d16, [x28], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
       "zip1 v16.8h, v17.8h, v16.8h\n"
-      "str q16, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr h17, [x9], #0x2\n"
-      "ldr h16, [x26], #0x2\n"
+      "ldr h16, [x28], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
       "zip1 v16.8h, v17.8h, v16.8h\n"
-      "str s16, [x27, #0x0]\n"
-      "add x27, x27, #0x4\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x30\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x4.hpp
index b9dca66e7b..67493393a0 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x4.hpp
@@ -40,412 +40,388 @@ void a64_transpose_interleave_12_2x4(uint16_t *out, const uint16_t *in, size_t w
 
     __asm__ __volatile__(
       "cmp %x[height], #0x8\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "cmp x28, #0x18\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q27, [x9], #0x10\n"
-      "ldr q26, [x26], #0x10\n"
-      "sub x28, x28, #0x18\n"
-      "ldr q25, [x25], #0x10\n"
-      "ldr q24, [x24], #0x10\n"
-      "cmp x28, #0x18\n"
-      "ldr q23, [x23], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
-      "ldr q20, [x20], #0x10\n"
-      "ldr q19, [x9], #0x10\n"
-      "ldr q18, [x26], #0x10\n"
-      "zip1 v15.8h, v27.8h, v25.8h\n"
-      "zip1 v14.8h, v26.8h, v24.8h\n"
-      "ldr q17, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "zip2 v13.8h, v27.8h, v25.8h\n"
-      "zip2 v12.8h, v26.8h, v24.8h\n"
-      "ldr q1, [x23], #0x10\n"
-      "ldr q0, [x22], #0x10\n"
-      "zip1 v11.8h, v23.8h, v21.8h\n"
-      "zip1 v10.8h, v22.8h, v20.8h\n"
-      "ldr q31, [x21], #0x10\n"
-      "ldr q24, [x20], #0x10\n"
-      "zip2 v30.8h, v23.8h, v21.8h\n"
-      "zip2 v29.8h, v22.8h, v20.8h\n"
-      "ldr q23, [x9], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
-      "zip1 v28.8h, v19.8h, v17.8h\n"
-      "zip1 v27.8h, v18.8h, v16.8h\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "zip2 v9.8h, v19.8h, v17.8h\n"
-      "zip2 v8.8h, v18.8h, v16.8h\n"
-      "ldr q19, [x23], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "zip1 v26.8h, v1.8h, v31.8h\n"
-      "zip1 v25.8h, v0.8h, v24.8h\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "sub x24, x24, #0x18\n"
+      "cmp x24, #0x18\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v13.8h, v21.8h, v17.8h\n"
+      "zip1 v12.8h, v20.8h, v16.8h\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v11.8h, v21.8h, v17.8h\n"
+      "zip2 v10.8h, v20.8h, v16.8h\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v9.8h, v19.8h, v17.8h\n"
+      "zip1 v8.8h, v18.8h, v16.8h\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v7.8h, v19.8h, v17.8h\n"
+      "zip2 v6.8h, v18.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v27.8h, v21.8h, v17.8h\n"
+      "zip1 v22.8h, v20.8h, v16.8h\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v5.8h, v21.8h, v17.8h\n"
+      "zip2 v4.8h, v20.8h, v16.8h\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v26.8h, v19.8h, v17.8h\n"
+      "zip1 v25.8h, v18.8h, v16.8h\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v3.8h, v19.8h, v17.8h\n"
+      "zip2 v2.8h, v18.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v24.8h, v21.8h, v17.8h\n"
+      "zip1 v23.8h, v20.8h, v16.8h\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v1.8h, v21.8h, v17.8h\n"
+      "zip2 v0.8h, v20.8h, v16.8h\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip2 v7.8h, v1.8h, v31.8h\n"
-      "zip2 v6.8h, v0.8h, v24.8h\n"
-      "zip1 v5.8h, v23.8h, v21.8h\n"
-      "zip1 v24.8h, v22.8h, v20.8h\n"
-      "zip2 v4.8h, v23.8h, v21.8h\n"
-      "zip2 v3.8h, v22.8h, v20.8h\n"
-      "zip1 v2.8h, v19.8h, v17.8h\n"
-      "zip1 v1.8h, v18.8h, v16.8h\n"
-      "zip2 v0.8h, v19.8h, v17.8h\n"
-      "zip2 v31.8h, v18.8h, v16.8h\n"
-      "zip1 v16.8h, v15.8h, v14.8h\n"
-      "zip2 v23.8h, v15.8h, v14.8h\n"
-      "zip1 v22.8h, v13.8h, v12.8h\n"
-      "zip2 v21.8h, v13.8h, v12.8h\n"
-      "zip1 v20.8h, v28.8h, v27.8h\n"
-      "zip2 v19.8h, v28.8h, v27.8h\n"
+      "zip1 v31.8h, v19.8h, v17.8h\n"
+      "zip1 v30.8h, v18.8h, v16.8h\n"
+      "zip2 v29.8h, v19.8h, v17.8h\n"
+      "zip2 v28.8h, v18.8h, v16.8h\n"
+      "zip1 v17.8h, v13.8h, v12.8h\n"
+      "zip2 v16.8h, v13.8h, v12.8h\n"
+      "str q17, [x21, #0x0]\n"
       "zip1 v18.8h, v11.8h, v10.8h\n"
       "zip2 v17.8h, v11.8h, v10.8h\n"
-      "str q16, [x27, #0x0]\n"
-      "zip1 v16.8h, v30.8h, v29.8h\n"
-      "zip2 v30.8h, v30.8h, v29.8h\n"
-      "str q23, [x27, #0x10]\n"
-      "zip1 v29.8h, v26.8h, v25.8h\n"
-      "zip2 v28.8h, v26.8h, v25.8h\n"
-      "str q22, [x27, #0x20]\n"
-      "str q21, [x27, #0x30]\n"
-      "zip1 v27.8h, v9.8h, v8.8h\n"
-      "zip2 v26.8h, v9.8h, v8.8h\n"
-      "str q20, [x27, #0x40]\n"
-      "zip1 v25.8h, v5.8h, v24.8h\n"
-      "zip2 v24.8h, v5.8h, v24.8h\n"
-      "str q19, [x27, #0x50]\n"
-      "zip1 v23.8h, v4.8h, v3.8h\n"
-      "zip2 v22.8h, v4.8h, v3.8h\n"
-      "str q18, [x27, #0x60]\n"
-      "zip1 v21.8h, v7.8h, v6.8h\n"
-      "zip2 v20.8h, v7.8h, v6.8h\n"
-      "str q17, [x27, #0x70]\n"
-      "zip1 v19.8h, v2.8h, v1.8h\n"
-      "zip2 v18.8h, v2.8h, v1.8h\n"
-      "str q16, [x27, #0x80]\n"
-      "zip1 v17.8h, v0.8h, v31.8h\n"
-      "zip2 v16.8h, v0.8h, v31.8h\n"
-      "str q30, [x27, #0x90]\n"
-      "str q29, [x27, #0xa0]\n"
-      "str q28, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q27, [x27, #0x0]\n"
-      "str q26, [x27, #0x10]\n"
-      "str q25, [x27, #0x20]\n"
-      "str q24, [x27, #0x30]\n"
-      "str q23, [x27, #0x40]\n"
-      "str q22, [x27, #0x50]\n"
-      "str q21, [x27, #0x60]\n"
-      "str q20, [x27, #0x70]\n"
-      "str q19, [x27, #0x80]\n"
-      "str q18, [x27, #0x90]\n"
-      "str q17, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "str q16, [x21, #0x10]\n"
+      "zip1 v16.8h, v27.8h, v22.8h\n"
+      "zip2 v22.8h, v27.8h, v22.8h\n"
+      "str q18, [x21, #0x20]\n"
+      "zip1 v21.8h, v9.8h, v8.8h\n"
+      "zip2 v20.8h, v9.8h, v8.8h\n"
+      "str q17, [x21, #0x30]\n"
+      "zip1 v19.8h, v7.8h, v6.8h\n"
+      "zip2 v18.8h, v7.8h, v6.8h\n"
+      "str q16, [x21, #0x40]\n"
+      "zip1 v17.8h, v26.8h, v25.8h\n"
+      "zip2 v16.8h, v26.8h, v25.8h\n"
+      "str q22, [x21, #0x50]\n"
+      "str q21, [x21, #0x60]\n"
+      "zip1 v27.8h, v5.8h, v4.8h\n"
+      "zip2 v26.8h, v5.8h, v4.8h\n"
+      "str q20, [x21, #0x70]\n"
+      "zip1 v25.8h, v24.8h, v23.8h\n"
+      "zip2 v24.8h, v24.8h, v23.8h\n"
+      "str q19, [x21, #0x80]\n"
+      "zip1 v23.8h, v1.8h, v0.8h\n"
+      "zip2 v22.8h, v1.8h, v0.8h\n"
+      "str q18, [x21, #0x90]\n"
+      "zip1 v21.8h, v3.8h, v2.8h\n"
+      "zip2 v20.8h, v3.8h, v2.8h\n"
+      "str q17, [x21, #0xa0]\n"
+      "zip1 v19.8h, v31.8h, v30.8h\n"
+      "zip2 v18.8h, v31.8h, v30.8h\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip1 v17.8h, v29.8h, v28.8h\n"
+      "zip2 v16.8h, v29.8h, v28.8h\n"
+      "str q27, [x21, #0x0]\n"
+      "str q26, [x21, #0x10]\n"
+      "str q25, [x21, #0x20]\n"
+      "str q24, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q22, [x21, #0x50]\n"
+      "str q21, [x21, #0x60]\n"
+      "str q20, [x21, #0x70]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q17, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x28, #0xc\n"
+      "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr q23, [x9], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
-      "sub x28, x28, #0xc\n"
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "sub x24, x24, #0xc\n"
+      "cmp x24, #0xc\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v25.8h, v19.8h, v17.8h\n"
+      "zip1 v24.8h, v18.8h, v16.8h\n"
       "ldr q21, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "cmp x28, #0xc\n"
-      "ldr q1, [x23], #0x10\n"
-      "ldr q0, [x22], #0x10\n"
-      "ldr q28, [x21], #0x10\n"
-      "ldr q27, [x20], #0x10\n"
-      "ldr d20, [x9], #0x8\n"
-      "ldr d19, [x26], #0x8\n"
-      "zip1 v26.8h, v23.8h, v21.8h\n"
-      "zip1 v25.8h, v22.8h, v16.8h\n"
-      "ldr d18, [x25], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
-      "zip2 v31.8h, v23.8h, v21.8h\n"
-      "zip2 v24.8h, v22.8h, v16.8h\n"
-      "ldr d23, [x23], #0x8\n"
-      "ldr d22, [x22], #0x8\n"
-      "zip1 v30.8h, v1.8h, v28.8h\n"
-      "zip1 v29.8h, v0.8h, v27.8h\n"
-      "ldr d21, [x21], #0x8\n"
+      "ldr q20, [x23], #0x10\n"
+      "zip2 v31.8h, v19.8h, v17.8h\n"
+      "zip2 v23.8h, v18.8h, v16.8h\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v30.8h, v21.8h, v17.8h\n"
+      "zip1 v29.8h, v20.8h, v16.8h\n"
+      "ldr d19, [x9], #0x8\n"
+      "ldr d18, [x28], #0x8\n"
+      "zip2 v28.8h, v21.8h, v17.8h\n"
+      "zip2 v27.8h, v20.8h, v16.8h\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
+      "zip1 v26.8h, v19.8h, v17.8h\n"
+      "zip1 v22.8h, v18.8h, v16.8h\n"
+      "ldr d21, [x25], #0x8\n"
+      "ldr d20, [x23], #0x8\n"
+      "zip1 v19.8h, v25.8h, v24.8h\n"
+      "zip2 v18.8h, v25.8h, v24.8h\n"
+      "ldr d17, [x22], #0x8\n"
       "ldr d16, [x20], #0x8\n"
-      "zip2 v28.8h, v1.8h, v28.8h\n"
-      "zip2 v27.8h, v0.8h, v27.8h\n"
-      "zip1 v20.8h, v20.8h, v18.8h\n"
-      "zip1 v19.8h, v19.8h, v17.8h\n"
-      "zip1 v18.8h, v26.8h, v25.8h\n"
-      "zip2 v17.8h, v26.8h, v25.8h\n"
-      "zip1 v26.8h, v23.8h, v21.8h\n"
-      "zip1 v25.8h, v22.8h, v16.8h\n"
-      "zip1 v16.8h, v31.8h, v24.8h\n"
-      "zip2 v24.8h, v31.8h, v24.8h\n"
-      "zip1 v23.8h, v20.8h, v19.8h\n"
-      "zip2 v22.8h, v20.8h, v19.8h\n"
-      "str q18, [x27, #0x0]\n"
+      "zip1 v25.8h, v21.8h, v17.8h\n"
+      "zip1 v24.8h, v20.8h, v16.8h\n"
+      "zip1 v17.8h, v31.8h, v23.8h\n"
+      "zip2 v16.8h, v31.8h, v23.8h\n"
+      "str q19, [x21, #0x0]\n"
+      "zip1 v23.8h, v26.8h, v22.8h\n"
+      "zip2 v22.8h, v26.8h, v22.8h\n"
+      "str q18, [x21, #0x10]\n"
       "zip1 v21.8h, v30.8h, v29.8h\n"
       "zip2 v20.8h, v30.8h, v29.8h\n"
-      "str q17, [x27, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
       "zip1 v19.8h, v28.8h, v27.8h\n"
       "zip2 v18.8h, v28.8h, v27.8h\n"
-      "str q16, [x27, #0x20]\n"
-      "zip1 v17.8h, v26.8h, v25.8h\n"
-      "zip2 v16.8h, v26.8h, v25.8h\n"
-      "str q24, [x27, #0x30]\n"
-      "str q23, [x27, #0x40]\n"
-      "str q22, [x27, #0x50]\n"
-      "str q21, [x27, #0x60]\n"
-      "str q20, [x27, #0x70]\n"
-      "str q19, [x27, #0x80]\n"
-      "str q18, [x27, #0x90]\n"
-      "str q17, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "str q16, [x21, #0x30]\n"
+      "zip1 v17.8h, v25.8h, v24.8h\n"
+      "zip2 v16.8h, v25.8h, v24.8h\n"
+      "str q23, [x21, #0x40]\n"
+      "str q22, [x21, #0x50]\n"
+      "str q21, [x21, #0x60]\n"
+      "str q20, [x21, #0x70]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q17, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x28, 10f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "str q16, [x27, #0x80]\n"
-      "str q16, [x27, #0x90]\n"
-      "str q16, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
+      "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr d23, [x9], #0x8\n"
-      "ldr d22, [x26], #0x8\n"
-      "sub x28, x28, #0x4\n"
+      "ldr d19, [x9], #0x8\n"
+      "ldr d18, [x28], #0x8\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
+      "zip1 v17.8h, v19.8h, v17.8h\n"
+      "zip1 v16.8h, v18.8h, v16.8h\n"
       "ldr d18, [x25], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
-      "cmp x28, #0x4\n"
-      "ldr d20, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d19, [x21], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "zip1 v20.8h, v17.8h, v16.8h\n"
+      "zip2 v19.8h, v17.8h, v16.8h\n"
+      "ldr d17, [x22], #0x8\n"
       "ldr d16, [x20], #0x8\n"
-      "zip1 v18.8h, v23.8h, v18.8h\n"
-      "zip1 v17.8h, v22.8h, v17.8h\n"
-      "zip1 v20.8h, v20.8h, v19.8h\n"
+      "zip1 v18.8h, v18.8h, v17.8h\n"
       "zip1 v16.8h, v21.8h, v16.8h\n"
-      "zip1 v19.8h, v18.8h, v17.8h\n"
-      "zip2 v18.8h, v18.8h, v17.8h\n"
-      "zip1 v17.8h, v20.8h, v16.8h\n"
-      "zip2 v16.8h, v20.8h, v16.8h\n"
-      "str q19, [x27, #0x0]\n"
-      "str q18, [x27, #0x10]\n"
-      "str q17, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "add x27, x27, #0x20\n"
+      "str q20, [x21, #0x0]\n"
+      "zip1 v17.8h, v18.8h, v16.8h\n"
+      "zip2 v16.8h, v18.8h, v16.8h\n"
+      "str q19, [x21, #0x10]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x70]\n"
+      "add x21, x21, #0x20\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr h23, [x9], #0x2\n"
-      "ldr h22, [x26], #0x2\n"
-      "sub x28, x28, #0x1\n"
-      "ldr h19, [x25], #0x2\n"
-      "ldr h17, [x24], #0x2\n"
-      "cmp x28, #0x1\n"
-      "ldr h21, [x23], #0x2\n"
-      "ldr h20, [x22], #0x2\n"
-      "ldr h18, [x21], #0x2\n"
-      "ldr h16, [x20], #0x2\n"
-      "zip1 v19.8h, v23.8h, v19.8h\n"
-      "zip1 v17.8h, v22.8h, v17.8h\n"
-      "zip1 v18.8h, v21.8h, v18.8h\n"
-      "zip1 v16.8h, v20.8h, v16.8h\n"
+      "ldr h19, [x9], #0x2\n"
+      "ldr h18, [x28], #0x2\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr h17, [x27], #0x2\n"
+      "ldr h16, [x26], #0x2\n"
       "zip1 v17.8h, v19.8h, v17.8h\n"
       "zip1 v16.8h, v18.8h, v16.8h\n"
-      "str d17, [x27, #0x0]\n"
-      "str d16, [x27, #0x60]\n"
-      "add x27, x27, #0x8\n"
+      "ldr h20, [x25], #0x2\n"
+      "ldr h19, [x23], #0x2\n"
+      "zip1 v18.8h, v17.8h, v16.8h\n"
+      "ldr h17, [x22], #0x2\n"
+      "ldr h16, [x20], #0x2\n"
+      "zip1 v17.8h, v20.8h, v17.8h\n"
+      "zip1 v16.8h, v19.8h, v16.8h\n"
+      "str d18, [x21, #0x0]\n"
+      "zip1 v16.8h, v17.8h, v16.8h\n"
+      "str d16, [x21, #0x60]\n"
+      "add x21, x21, #0x8\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x8\n"
       "add %x[out], %x[out], #0xc0\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x9, %x[in]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x26, x27, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x9, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "csel x25, x25, %x[pad_row], GE\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x26, %x[in_stride]\n"
       "csel x26, x26, %x[pad_row], GT\n"
+      "csel x27, x27, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x28, x28, %x[pad_row], GT\n"
       "cmp x20, #0x18\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
-      "ldr q23, [x9], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
       "sub x20, x20, #0x18\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
       "cmp x20, #0x18\n"
-      "ldr q26, [x9], #0x10\n"
-      "ldr q25, [x26], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v31.8h, v19.8h, v17.8h\n"
+      "zip1 v30.8h, v18.8h, v16.8h\n"
+      "ldr q22, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v29.8h, v19.8h, v17.8h\n"
+      "zip2 v28.8h, v18.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v27.8h, v22.8h, v17.8h\n"
+      "zip1 v21.8h, v20.8h, v16.8h\n"
       "ldr q19, [x9], #0x10\n"
-      "zip1 v31.8h, v23.8h, v21.8h\n"
-      "zip1 v30.8h, v22.8h, v16.8h\n"
-      "ldr q18, [x26], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
-      "zip2 v29.8h, v23.8h, v21.8h\n"
-      "zip2 v23.8h, v22.8h, v16.8h\n"
-      "ldr q16, [x24], #0x10\n"
-      "zip1 v22.8h, v26.8h, v24.8h\n"
-      "zip1 v21.8h, v25.8h, v20.8h\n"
-      "zip2 v28.8h, v26.8h, v24.8h\n"
-      "zip2 v20.8h, v25.8h, v20.8h\n"
-      "zip1 v27.8h, v19.8h, v17.8h\n"
-      "zip1 v26.8h, v18.8h, v16.8h\n"
-      "zip2 v25.8h, v19.8h, v17.8h\n"
-      "zip2 v24.8h, v18.8h, v16.8h\n"
-      "zip1 v19.8h, v31.8h, v30.8h\n"
-      "zip2 v18.8h, v31.8h, v30.8h\n"
-      "zip1 v17.8h, v29.8h, v23.8h\n"
-      "zip2 v16.8h, v29.8h, v23.8h\n"
-      "zip1 v23.8h, v22.8h, v21.8h\n"
-      "zip2 v22.8h, v22.8h, v21.8h\n"
-      "str q19, [x27, #0x0]\n"
-      "str q18, [x27, #0x10]\n"
-      "zip1 v21.8h, v28.8h, v20.8h\n"
-      "zip2 v20.8h, v28.8h, v20.8h\n"
-      "str q17, [x27, #0x20]\n"
-      "zip1 v19.8h, v27.8h, v26.8h\n"
-      "zip2 v18.8h, v27.8h, v26.8h\n"
-      "str q16, [x27, #0x30]\n"
-      "zip1 v17.8h, v25.8h, v24.8h\n"
-      "zip2 v16.8h, v25.8h, v24.8h\n"
-      "str q23, [x27, #0x40]\n"
-      "str q22, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x20]\n"
-      "str q18, [x27, #0x30]\n"
-      "str q17, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v26.8h, v22.8h, v17.8h\n"
+      "zip2 v20.8h, v20.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v25.8h, v19.8h, v17.8h\n"
+      "zip1 v24.8h, v18.8h, v16.8h\n"
+      "zip2 v23.8h, v19.8h, v17.8h\n"
+      "zip2 v22.8h, v18.8h, v16.8h\n"
+      "zip1 v17.8h, v31.8h, v30.8h\n"
+      "zip2 v16.8h, v31.8h, v30.8h\n"
+      "str q17, [x21, #0x0]\n"
+      "zip1 v19.8h, v29.8h, v28.8h\n"
+      "zip2 v18.8h, v29.8h, v28.8h\n"
+      "str q16, [x21, #0x10]\n"
+      "zip1 v17.8h, v27.8h, v21.8h\n"
+      "zip2 v16.8h, v27.8h, v21.8h\n"
+      "str q19, [x21, #0x20]\n"
+      "str q18, [x21, #0x30]\n"
+      "zip1 v21.8h, v26.8h, v20.8h\n"
+      "zip2 v20.8h, v26.8h, v20.8h\n"
+      "str q17, [x21, #0x40]\n"
+      "zip1 v19.8h, v25.8h, v24.8h\n"
+      "zip2 v18.8h, v25.8h, v24.8h\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip1 v17.8h, v23.8h, v22.8h\n"
+      "zip2 v16.8h, v23.8h, v22.8h\n"
+      "str q21, [x21, #0x0]\n"
+      "str q20, [x21, #0x10]\n"
+      "str q19, [x21, #0x20]\n"
+      "str q18, [x21, #0x30]\n"
+      "str q17, [x21, #0x40]\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
-      "ldr q25, [x9], #0x10\n"
-      "ldr q24, [x26], #0x10\n"
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
+      "ldr q21, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
       "sub x20, x20, #0xc\n"
-      "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0xc\n"
-      "ldr d22, [x9], #0x8\n"
-      "ldr d23, [x26], #0x8\n"
-      "ldr d21, [x25], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
-      "zip1 v20.8h, v25.8h, v18.8h\n"
-      "zip1 v19.8h, v24.8h, v17.8h\n"
-      "zip2 v18.8h, v25.8h, v18.8h\n"
-      "zip2 v17.8h, v24.8h, v17.8h\n"
-      "zip1 v22.8h, v22.8h, v21.8h\n"
-      "zip1 v16.8h, v23.8h, v16.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip1 v19.8h, v18.8h, v17.8h\n"
-      "zip2 v18.8h, v18.8h, v17.8h\n"
-      "zip1 v17.8h, v22.8h, v16.8h\n"
-      "zip2 v16.8h, v22.8h, v16.8h\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x20]\n"
-      "str q18, [x27, #0x30]\n"
-      "str q17, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v24.8h, v21.8h, v17.8h\n"
+      "zip1 v23.8h, v18.8h, v16.8h\n"
+      "ldr d20, [x9], #0x8\n"
+      "ldr d19, [x28], #0x8\n"
+      "zip2 v22.8h, v21.8h, v17.8h\n"
+      "zip2 v18.8h, v18.8h, v16.8h\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
+      "zip1 v21.8h, v20.8h, v17.8h\n"
+      "zip1 v20.8h, v19.8h, v16.8h\n"
+      "zip1 v17.8h, v24.8h, v23.8h\n"
+      "zip2 v16.8h, v24.8h, v23.8h\n"
+      "str q17, [x21, #0x0]\n"
+      "zip1 v19.8h, v22.8h, v18.8h\n"
+      "zip2 v18.8h, v22.8h, v18.8h\n"
+      "str q16, [x21, #0x10]\n"
+      "zip1 v17.8h, v21.8h, v20.8h\n"
+      "zip2 v16.8h, v21.8h, v20.8h\n"
+      "str q19, [x21, #0x20]\n"
+      "str q18, [x21, #0x30]\n"
+      "str q17, [x21, #0x40]\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr d18, [x9], #0x8\n"
-      "ldr d19, [x26], #0x8\n"
+      "ldr d19, [x28], #0x8\n"
       "sub x20, x20, #0x4\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
       "cmp x20, #0x4\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
       "zip1 v18.8h, v18.8h, v17.8h\n"
       "zip1 v16.8h, v19.8h, v16.8h\n"
       "zip1 v17.8h, v18.8h, v16.8h\n"
       "zip2 v16.8h, v18.8h, v16.8h\n"
-      "str q17, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q17, [x21, #0x0]\n"
+      "str q16, [x21, #0x10]\n"
+      "add x21, x21, #0x20\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr h19, [x9], #0x2\n"
-      "ldr h18, [x26], #0x2\n"
+      "ldr h18, [x28], #0x2\n"
       "sub x20, x20, #0x1\n"
-      "ldr h17, [x25], #0x2\n"
-      "ldr h16, [x24], #0x2\n"
       "cmp x20, #0x1\n"
+      "ldr h17, [x27], #0x2\n"
+      "ldr h16, [x26], #0x2\n"
       "zip1 v17.8h, v19.8h, v17.8h\n"
       "zip1 v16.8h, v18.8h, v16.8h\n"
       "zip1 v16.8h, v17.8h, v16.8h\n"
-      "str d16, [x27, #0x0]\n"
-      "add x27, x27, #0x8\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x60\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x4_fp32bf16.hpp
index 07326e7c98..fe554a65f9 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x4_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_2x4_fp32bf16.hpp
@@ -40,607 +40,592 @@ void a64_transpose_interleave_12_2x4_fp32bf16(bfloat16 *out, const float *in, si
 
     __asm__ __volatile__(
       "cmp %x[height], #0x8\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "cmp x28, #0x18\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q18, [x9], #0x10\n"
-      "ldr q24, [x26], #0x10\n"
-      "sub x28, x28, #0x18\n"
-      "ldr q23, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "cmp x28, #0x18\n"
-      "ldr q29, [x23], #0x10\n"
-      "ldr q7, [x22], #0x10\n"
-      "ldr q5, [x21], #0x10\n"
-      "ldr q12, [x20], #0x10\n"
-      "ldr q1, [x9], #0x10\n"
-      "ldr q13, [x26], #0x10\n"
-      "zip1 v6.4s, v18.4s, v23.4s\n"
-      "zip1 v26.4s, v24.4s, v20.4s\n"
-      "ldr q9, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
-      "zip2 v23.4s, v18.4s, v23.4s\n"
-      "zip2 v19.4s, v24.4s, v20.4s\n"
-      "ldr q22, [x23], #0x10\n"
-      "ldr q27, [x22], #0x10\n"
-      "zip1 v15.4s, v29.4s, v5.4s\n"
-      "zip1 v2.4s, v7.4s, v12.4s\n"
-      "ldr q25, [x21], #0x10\n"
-      "ldr q28, [x20], #0x10\n"
-      "zip2 v21.4s, v29.4s, v5.4s\n"
-      "zip2 v11.4s, v7.4s, v12.4s\n"
-      "ldr q3, [x9], #0x10\n"
-      "ldr q31, [x26], #0x10\n"
-      "zip1 v12.4s, v1.4s, v9.4s\n"
-      "zip1 v14.4s, v13.4s, v17.4s\n"
-      "ldr q30, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "zip2 v1.4s, v1.4s, v9.4s\n"
-      "zip2 v0.4s, v13.4s, v17.4s\n"
-      "ldr q13, [x23], #0x10\n"
-      "ldr q8, [x22], #0x10\n"
-      "zip1 v9.4s, v22.4s, v25.4s\n"
-      "zip1 v29.4s, v27.4s, v28.4s\n"
-      "ldr q5, [x21], #0x10\n"
-      "ldr q17, [x20], #0x10\n"
-      "zip2 v20.4s, v22.4s, v25.4s\n"
-      "zip2 v28.4s, v27.4s, v28.4s\n"
-      "ldr q7, [x9], #0x10\n"
-      "ldr q24, [x26], #0x10\n"
-      "zip1 v25.4s, v3.4s, v30.4s\n"
-      "zip1 v10.4s, v31.4s, v16.4s\n"
-      "ldr q4, [x25], #0x10\n"
-      "ldr q22, [x24], #0x10\n"
-      "zip2 v27.4s, v3.4s, v30.4s\n"
-      "zip2 v3.4s, v31.4s, v16.4s\n"
-      "ldr q31, [x23], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "zip1 v16.4s, v13.4s, v5.4s\n"
-      "zip1 v30.4s, v8.4s, v17.4s\n"
-      "zip2 v13.4s, v13.4s, v5.4s\n"
-      "ldr q5, [x21], #0x10\n"
-      "zip2 v17.4s, v8.4s, v17.4s\n"
-      "zip1 v8.4s, v7.4s, v4.4s\n"
-      "zip2 v7.4s, v7.4s, v4.4s\n"
-      "zip1 v4.4s, v24.4s, v22.4s\n"
-      "zip2 v22.4s, v24.4s, v22.4s\n"
-      "zip1 v24.4s, v31.4s, v5.4s\n"
-      "zip2 v31.4s, v31.4s, v5.4s\n"
-      "zip1 v5.4s, v6.4s, v26.4s\n"
-      "zip2 v6.4s, v6.4s, v26.4s\n"
+      "ldr q15, [x9], #0x10\n"
+      "ldr q17, [x28], #0x10\n"
+      "sub x24, x24, #0x18\n"
+      "cmp x24, #0x18\n"
+      "ldr q16, [x27], #0x10\n"
+      "ldr q20, [x26], #0x10\n"
+      "zip1 v6.4s, v15.4s, v16.4s\n"
+      "zip1 v11.4s, v17.4s, v20.4s\n"
+      "ldr q2, [x25], #0x10\n"
+      "ldr q4, [x23], #0x10\n"
+      "zip2 v22.4s, v15.4s, v16.4s\n"
+      "zip2 v18.4s, v17.4s, v20.4s\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q26, [x20], #0x10\n"
-      ".inst 0x0ea168a5  // bfcvtn v5.4h, v5.4s\n"
-      ".inst 0x4ea168c5  // bfcvtn2 v5.8h, v6.4s\n"
-      "zip1 v6.4s, v23.4s, v19.4s\n"
-      "zip2 v19.4s, v23.4s, v19.4s\n"
-      "ldr q23, [x9], #0x10\n"
-      ".inst 0x0ea168c6  // bfcvtn v6.4h, v6.4s\n"
-      ".inst 0x4ea16a66  // bfcvtn2 v6.8h, v19.4s\n"
-      "zip1 v19.4s, v18.4s, v26.4s\n"
-      "zip2 v26.4s, v18.4s, v26.4s\n"
-      "zip1 v18.4s, v12.4s, v14.4s\n"
-      "zip2 v12.4s, v12.4s, v14.4s\n"
-      "ldr q14, [x26], #0x10\n"
-      ".inst 0x0ea16a52  // bfcvtn v18.4h, v18.4s\n"
-      ".inst 0x4ea16992  // bfcvtn2 v18.8h, v12.4s\n"
-      "zip1 v12.4s, v1.4s, v0.4s\n"
-      "zip2 v1.4s, v1.4s, v0.4s\n"
-      "ldr q0, [x25], #0x10\n"
-      ".inst 0x0ea1698c  // bfcvtn v12.4h, v12.4s\n"
-      ".inst 0x4ea1682c  // bfcvtn2 v12.8h, v1.4s\n"
-      "zip1 v1.4s, v25.4s, v10.4s\n"
-      "zip2 v10.4s, v25.4s, v10.4s\n"
-      "ldr q25, [x24], #0x10\n"
+      "zip1 v9.4s, v2.4s, v17.4s\n"
+      "zip1 v10.4s, v4.4s, v26.4s\n"
+      "ldr q16, [x9], #0x10\n"
+      "ldr q27, [x28], #0x10\n"
+      "zip2 v3.4s, v2.4s, v17.4s\n"
+      "zip2 v30.4s, v4.4s, v26.4s\n"
+      "ldr q13, [x27], #0x10\n"
+      "ldr q1, [x26], #0x10\n"
+      "zip1 v23.4s, v16.4s, v13.4s\n"
+      "zip1 v5.4s, v27.4s, v1.4s\n"
+      "ldr q26, [x25], #0x10\n"
+      "ldr q14, [x23], #0x10\n"
+      "zip2 v0.4s, v16.4s, v13.4s\n"
+      "zip2 v2.4s, v27.4s, v1.4s\n"
+      "ldr q15, [x22], #0x10\n"
+      "ldr q8, [x20], #0x10\n"
+      "zip1 v31.4s, v26.4s, v15.4s\n"
+      "zip1 v4.4s, v14.4s, v8.4s\n"
+      "ldr q28, [x9], #0x10\n"
+      "ldr q19, [x28], #0x10\n"
+      "zip2 v21.4s, v26.4s, v15.4s\n"
+      "zip2 v16.4s, v14.4s, v8.4s\n"
+      "ldr q15, [x27], #0x10\n"
+      "ldr q1, [x26], #0x10\n"
+      "zip1 v17.4s, v28.4s, v15.4s\n"
+      "zip1 v8.4s, v19.4s, v1.4s\n"
+      "ldr q27, [x25], #0x10\n"
+      "ldr q20, [x23], #0x10\n"
+      "zip2 v7.4s, v28.4s, v15.4s\n"
+      "zip2 v15.4s, v19.4s, v1.4s\n"
+      "ldr q12, [x22], #0x10\n"
+      "ldr q25, [x20], #0x10\n"
+      "zip1 v14.4s, v27.4s, v12.4s\n"
+      "zip1 v26.4s, v20.4s, v25.4s\n"
+      "ldr q13, [x9], #0x10\n"
+      "ldr q29, [x28], #0x10\n"
+      "zip2 v28.4s, v27.4s, v12.4s\n"
+      "zip2 v12.4s, v20.4s, v25.4s\n"
+      "ldr q27, [x27], #0x10\n"
+      "ldr q20, [x26], #0x10\n"
+      "zip1 v19.4s, v13.4s, v27.4s\n"
+      "zip1 v25.4s, v29.4s, v20.4s\n"
+      "ldr q24, [x25], #0x10\n"
+      "ldr q1, [x23], #0x10\n"
+      "zip2 v27.4s, v13.4s, v27.4s\n"
+      "zip2 v13.4s, v29.4s, v20.4s\n"
+      "ldr q20, [x22], #0x10\n"
+      "zip1 v29.4s, v24.4s, v20.4s\n"
+      "zip2 v20.4s, v24.4s, v20.4s\n"
+      "zip1 v24.4s, v6.4s, v11.4s\n"
+      ".inst 0x0ea16b18  // bfcvtn v24.4h, v24.4s\n"
+      "zip2 v11.4s, v6.4s, v11.4s\n"
+      "ldr q6, [x20], #0x10\n"
+      ".inst 0x4ea16978  // bfcvtn2 v24.8h, v11.4s\n"
+      "zip1 v11.4s, v1.4s, v6.4s\n"
+      "zip2 v6.4s, v1.4s, v6.4s\n"
+      "zip1 v1.4s, v22.4s, v18.4s\n"
       ".inst 0x0ea16821  // bfcvtn v1.4h, v1.4s\n"
-      ".inst 0x4ea16941  // bfcvtn2 v1.8h, v10.4s\n"
-      "zip1 v10.4s, v23.4s, v0.4s\n"
-      "zip2 v23.4s, v23.4s, v0.4s\n"
-      "zip1 v0.4s, v14.4s, v25.4s\n"
-      "zip2 v25.4s, v14.4s, v25.4s\n"
-      "zip1 v14.4s, v27.4s, v3.4s\n"
-      "zip2 v3.4s, v27.4s, v3.4s\n"
-      "ldr q27, [x23], #0x10\n"
-      ".inst 0x0ea169ce  // bfcvtn v14.4h, v14.4s\n"
-      ".inst 0x4ea1686e  // bfcvtn2 v14.8h, v3.4s\n"
-      "zip1 v3.4s, v15.4s, v2.4s\n"
-      "zip2 v15.4s, v15.4s, v2.4s\n"
-      "ldr q2, [x22], #0x10\n"
+      "zip2 v18.4s, v22.4s, v18.4s\n"
+      "ldr q22, [x9], #0x10\n"
+      ".inst 0x4ea16a41  // bfcvtn2 v1.8h, v18.4s\n"
+      "zip1 v18.4s, v23.4s, v5.4s\n"
+      ".inst 0x0ea16a52  // bfcvtn v18.4h, v18.4s\n"
+      "zip2 v5.4s, v23.4s, v5.4s\n"
+      "ldr q23, [x28], #0x10\n"
+      ".inst 0x4ea168b2  // bfcvtn2 v18.8h, v5.4s\n"
+      "zip1 v5.4s, v0.4s, v2.4s\n"
+      ".inst 0x0ea168a5  // bfcvtn v5.4h, v5.4s\n"
+      "zip2 v0.4s, v0.4s, v2.4s\n"
+      "ldr q2, [x27], #0x10\n"
+      ".inst 0x4ea16805  // bfcvtn2 v5.8h, v0.4s\n"
+      "zip1 v0.4s, v22.4s, v2.4s\n"
+      "zip2 v2.4s, v22.4s, v2.4s\n"
+      "zip1 v22.4s, v17.4s, v8.4s\n"
+      ".inst 0x0ea16ad6  // bfcvtn v22.4h, v22.4s\n"
+      "zip2 v8.4s, v17.4s, v8.4s\n"
+      "ldr q17, [x26], #0x10\n"
+      ".inst 0x4ea16916  // bfcvtn2 v22.8h, v8.4s\n"
+      "zip1 v8.4s, v23.4s, v17.4s\n"
+      "zip2 v23.4s, v23.4s, v17.4s\n"
+      "zip1 v17.4s, v7.4s, v15.4s\n"
+      ".inst 0x0ea16a31  // bfcvtn v17.4h, v17.4s\n"
+      "zip2 v7.4s, v7.4s, v15.4s\n"
+      "ldr q15, [x25], #0x10\n"
+      ".inst 0x4ea168f1  // bfcvtn2 v17.8h, v7.4s\n"
+      "zip1 v7.4s, v9.4s, v10.4s\n"
+      ".inst 0x0ea168e7  // bfcvtn v7.4h, v7.4s\n"
+      "zip2 v10.4s, v9.4s, v10.4s\n"
+      "ldr q9, [x23], #0x10\n"
+      ".inst 0x4ea16947  // bfcvtn2 v7.8h, v10.4s\n"
+      "zip1 v10.4s, v3.4s, v30.4s\n"
+      ".inst 0x0ea1694a  // bfcvtn v10.4h, v10.4s\n"
+      "zip2 v30.4s, v3.4s, v30.4s\n"
+      "ldr q3, [x22], #0x10\n"
+      ".inst 0x4ea16bca  // bfcvtn2 v10.8h, v30.4s\n"
+      "zip1 v30.4s, v15.4s, v3.4s\n"
+      "zip2 v15.4s, v15.4s, v3.4s\n"
+      "zip1 v3.4s, v31.4s, v4.4s\n"
       ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
-      ".inst 0x4ea169e3  // bfcvtn2 v3.8h, v15.4s\n"
-      "zip1 v15.4s, v21.4s, v11.4s\n"
-      "zip2 v11.4s, v21.4s, v11.4s\n"
-      "ldr q21, [x21], #0x10\n"
-      ".inst 0x0ea169ef  // bfcvtn v15.4h, v15.4s\n"
-      ".inst 0x4ea1696f  // bfcvtn2 v15.8h, v11.4s\n"
-      "zip1 v11.4s, v9.4s, v29.4s\n"
-      "zip2 v29.4s, v9.4s, v29.4s\n"
-      "ldr q9, [x20], #0x10\n"
-      ".inst 0x0ea1696b  // bfcvtn v11.4h, v11.4s\n"
-      ".inst 0x4ea16bab  // bfcvtn2 v11.8h, v29.4s\n"
-      "zip1 v29.4s, v27.4s, v21.4s\n"
-      "zip2 v21.4s, v27.4s, v21.4s\n"
-      "zip1 v27.4s, v2.4s, v9.4s\n"
-      "zip2 v2.4s, v2.4s, v9.4s\n"
-      "zip1 v9.4s, v20.4s, v28.4s\n"
-      "zip2 v28.4s, v20.4s, v28.4s\n"
-      "ldr q20, [x9], #0x10\n"
+      "zip2 v31.4s, v31.4s, v4.4s\n"
+      "ldr q4, [x20], #0x10\n"
+      ".inst 0x4ea16be3  // bfcvtn2 v3.8h, v31.4s\n"
+      "zip1 v31.4s, v9.4s, v4.4s\n"
+      "zip2 v4.4s, v9.4s, v4.4s\n"
+      "zip1 v9.4s, v21.4s, v16.4s\n"
       ".inst 0x0ea16929  // bfcvtn v9.4h, v9.4s\n"
-      ".inst 0x4ea16b89  // bfcvtn2 v9.8h, v28.4s\n"
-      "zip1 v28.4s, v16.4s, v30.4s\n"
-      "zip2 v30.4s, v16.4s, v30.4s\n"
-      "ldr q16, [x26], #0x10\n"
-      ".inst 0x0ea16b9c  // bfcvtn v28.4h, v28.4s\n"
-      ".inst 0x4ea16bdc  // bfcvtn2 v28.8h, v30.4s\n"
-      "zip1 v30.4s, v13.4s, v17.4s\n"
-      "zip2 v13.4s, v13.4s, v17.4s\n"
-      "ldr q17, [x25], #0x10\n"
-      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
-      ".inst 0x4ea169be  // bfcvtn2 v30.8h, v13.4s\n"
-      "zip1 v13.4s, v8.4s, v4.4s\n"
-      "zip2 v4.4s, v8.4s, v4.4s\n"
-      "ldr q8, [x24], #0x10\n"
-      ".inst 0x0ea169ad  // bfcvtn v13.4h, v13.4s\n"
-      ".inst 0x4ea1688d  // bfcvtn2 v13.8h, v4.4s\n"
-      "zip1 v4.4s, v20.4s, v17.4s\n"
-      "zip2 v20.4s, v20.4s, v17.4s\n"
-      "zip1 v17.4s, v16.4s, v8.4s\n"
-      "zip2 v8.4s, v16.4s, v8.4s\n"
-      "zip1 v16.4s, v7.4s, v22.4s\n"
-      "zip2 v22.4s, v7.4s, v22.4s\n"
-      "ldr q7, [x23], #0x10\n"
+      "zip2 v16.4s, v21.4s, v16.4s\n"
+      "ldr q21, [x9], #0x10\n"
+      ".inst 0x4ea16a09  // bfcvtn2 v9.8h, v16.4s\n"
+      "zip1 v16.4s, v14.4s, v26.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      ".inst 0x4ea16ad0  // bfcvtn2 v16.8h, v22.4s\n"
-      "zip1 v22.4s, v10.4s, v0.4s\n"
-      "zip2 v10.4s, v10.4s, v0.4s\n"
-      "ldr q0, [x22], #0x10\n"
-      ".inst 0x0ea16ad6  // bfcvtn v22.4h, v22.4s\n"
-      ".inst 0x4ea16956  // bfcvtn2 v22.8h, v10.4s\n"
-      "zip1 v10.4s, v23.4s, v25.4s\n"
-      "zip2 v23.4s, v23.4s, v25.4s\n"
-      "ldr q25, [x21], #0x10\n"
-      ".inst 0x0ea1694a  // bfcvtn v10.4h, v10.4s\n"
-      ".inst 0x4ea16aea  // bfcvtn2 v10.8h, v23.4s\n"
-      "ldr q23, [x20], #0x10\n"
-      "str q5, [x27, #0x0]\n"
-      "zip1 v5.4s, v4.4s, v17.4s\n"
-      "zip2 v4.4s, v4.4s, v17.4s\n"
-      "str q6, [x27, #0x10]\n"
-      "zip1 v17.4s, v7.4s, v25.4s\n"
-      "zip2 v7.4s, v7.4s, v25.4s\n"
-      "str q18, [x27, #0x20]\n"
-      "zip1 v18.4s, v0.4s, v23.4s\n"
-      "zip2 v0.4s, v0.4s, v23.4s\n"
-      "str q12, [x27, #0x30]\n"
-      "zip1 v6.4s, v20.4s, v8.4s\n"
-      "zip1 v23.4s, v24.4s, v19.4s\n"
-      "str q1, [x27, #0x40]\n"
-      "zip1 v25.4s, v31.4s, v26.4s\n"
-      "zip1 v12.4s, v29.4s, v27.4s\n"
-      "str q14, [x27, #0x50]\n"
-      "zip1 v14.4s, v21.4s, v2.4s\n"
-      "zip1 v1.4s, v17.4s, v18.4s\n"
-      "str q3, [x27, #0x60]\n"
-      "zip1 v3.4s, v7.4s, v0.4s\n"
-      ".inst 0x0ea168a5  // bfcvtn v5.4h, v5.4s\n"
-      "str q15, [x27, #0x70]\n"
-      ".inst 0x0ea168c6  // bfcvtn v6.4h, v6.4s\n"
-      "zip2 v20.4s, v20.4s, v8.4s\n"
-      "str q11, [x27, #0x80]\n"
-      ".inst 0x0ea16af7  // bfcvtn v23.4h, v23.4s\n"
-      "zip2 v24.4s, v24.4s, v19.4s\n"
-      "str q9, [x27, #0x90]\n"
-      ".inst 0x0ea16b28  // bfcvtn v8.4h, v25.4s\n"
-      "zip2 v11.4s, v31.4s, v26.4s\n"
-      "str q28, [x27, #0xa0]\n"
-      ".inst 0x0ea16993  // bfcvtn v19.4h, v12.4s\n"
-      "zip2 v27.4s, v29.4s, v27.4s\n"
-      "str q30, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip2 v14.4s, v14.4s, v26.4s\n"
+      "ldr q26, [x28], #0x10\n"
+      ".inst 0x4ea169d0  // bfcvtn2 v16.8h, v14.4s\n"
+      "zip1 v14.4s, v28.4s, v12.4s\n"
       ".inst 0x0ea169ce  // bfcvtn v14.4h, v14.4s\n"
-      "zip2 v31.4s, v21.4s, v2.4s\n"
-      "str q13, [x27, #0x0]\n"
-      ".inst 0x0ea1683e  // bfcvtn v30.4h, v1.4s\n"
-      "zip2 v15.4s, v17.4s, v18.4s\n"
-      "str q16, [x27, #0x10]\n"
-      ".inst 0x0ea1686c  // bfcvtn v12.4h, v3.4s\n"
-      "zip2 v9.4s, v7.4s, v0.4s\n"
-      "str q22, [x27, #0x20]\n"
+      "zip2 v12.4s, v28.4s, v12.4s\n"
+      "ldr q28, [x27], #0x10\n"
+      ".inst 0x4ea1698e  // bfcvtn2 v14.8h, v12.4s\n"
+      "zip1 v12.4s, v21.4s, v28.4s\n"
+      "zip2 v28.4s, v21.4s, v28.4s\n"
+      "zip1 v21.4s, v19.4s, v25.4s\n"
+      ".inst 0x0ea16ab5  // bfcvtn v21.4h, v21.4s\n"
+      "zip2 v19.4s, v19.4s, v25.4s\n"
+      "ldr q25, [x26], #0x10\n"
+      ".inst 0x4ea16a75  // bfcvtn2 v21.8h, v19.4s\n"
+      "zip1 v19.4s, v26.4s, v25.4s\n"
+      "zip2 v25.4s, v26.4s, v25.4s\n"
+      "zip1 v26.4s, v27.4s, v13.4s\n"
+      ".inst 0x0ea16b5a  // bfcvtn v26.4h, v26.4s\n"
+      "zip2 v13.4s, v27.4s, v13.4s\n"
+      "ldr q27, [x25], #0x10\n"
+      ".inst 0x4ea169ba  // bfcvtn2 v26.8h, v13.4s\n"
+      "zip1 v13.4s, v0.4s, v8.4s\n"
+      ".inst 0x0ea169ad  // bfcvtn v13.4h, v13.4s\n"
+      "zip2 v8.4s, v0.4s, v8.4s\n"
+      "ldr q0, [x23], #0x10\n"
+      ".inst 0x4ea1690d  // bfcvtn2 v13.8h, v8.4s\n"
+      "zip1 v8.4s, v2.4s, v23.4s\n"
+      ".inst 0x0ea16908  // bfcvtn v8.4h, v8.4s\n"
+      "zip2 v23.4s, v2.4s, v23.4s\n"
+      "ldr q2, [x22], #0x10\n"
+      ".inst 0x4ea16ae8  // bfcvtn2 v8.8h, v23.4s\n"
+      "ldr q23, [x20], #0x10\n"
+      "str q24, [x21, #0x0]\n"
+      "zip1 v24.4s, v27.4s, v2.4s\n"
+      "zip2 v27.4s, v27.4s, v2.4s\n"
+      "zip1 v2.4s, v0.4s, v23.4s\n"
+      "zip2 v23.4s, v0.4s, v23.4s\n"
+      "str q1, [x21, #0x10]\n"
+      "zip1 v0.4s, v12.4s, v19.4s\n"
+      "zip1 v1.4s, v28.4s, v25.4s\n"
+      "str q18, [x21, #0x20]\n"
+      "zip1 v18.4s, v29.4s, v11.4s\n"
+      ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
+      "str q5, [x21, #0x30]\n"
+      "zip1 v5.4s, v20.4s, v6.4s\n"
+      "zip2 v19.4s, v12.4s, v19.4s\n"
+      "str q22, [x21, #0x40]\n"
+      "zip1 v12.4s, v30.4s, v31.4s\n"
+      "zip1 v22.4s, v15.4s, v4.4s\n"
+      "str q17, [x21, #0x50]\n"
+      "zip1 v17.4s, v24.4s, v2.4s\n"
+      ".inst 0x0ea16821  // bfcvtn v1.4h, v1.4s\n"
+      "str q7, [x21, #0x60]\n"
+      "zip1 v7.4s, v27.4s, v23.4s\n"
+      "zip2 v25.4s, v28.4s, v25.4s\n"
+      "str q10, [x21, #0x70]\n"
+      ".inst 0x0ea16a52  // bfcvtn v18.4h, v18.4s\n"
+      "zip2 v29.4s, v29.4s, v11.4s\n"
+      "str q3, [x21, #0x80]\n"
+      ".inst 0x0ea168ab  // bfcvtn v11.4h, v5.4s\n"
+      "zip2 v10.4s, v20.4s, v6.4s\n"
+      "str q9, [x21, #0x90]\n"
+      ".inst 0x0ea16986  // bfcvtn v6.4h, v12.4s\n"
+      "zip2 v12.4s, v30.4s, v31.4s\n"
+      "str q16, [x21, #0xa0]\n"
+      ".inst 0x0ea16ac5  // bfcvtn v5.4h, v22.4s\n"
+      "zip2 v4.4s, v15.4s, v4.4s\n"
+      "str q14, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      ".inst 0x0ea16a2f  // bfcvtn v15.4h, v17.4s\n"
+      "zip2 v20.4s, v24.4s, v2.4s\n"
+      "str q21, [x21, #0x0]\n"
+      ".inst 0x0ea168fc  // bfcvtn v28.4h, v7.4s\n"
+      "zip2 v30.4s, v27.4s, v23.4s\n"
+      "str q26, [x21, #0x10]\n"
+      ".inst 0x4ea16a60  // bfcvtn2 v0.8h, v19.4s\n"
+      ".inst 0x4ea16b21  // bfcvtn2 v1.8h, v25.4s\n"
+      "str q13, [x21, #0x20]\n"
+      ".inst 0x4ea16bb2  // bfcvtn2 v18.8h, v29.4s\n"
+      ".inst 0x4ea1694b  // bfcvtn2 v11.8h, v10.4s\n"
+      "str q8, [x21, #0x30]\n"
+      ".inst 0x4ea16986  // bfcvtn2 v6.8h, v12.4s\n"
       ".inst 0x4ea16885  // bfcvtn2 v5.8h, v4.4s\n"
-      ".inst 0x4ea16a86  // bfcvtn2 v6.8h, v20.4s\n"
-      "str q10, [x27, #0x30]\n"
-      ".inst 0x4ea16b17  // bfcvtn2 v23.8h, v24.4s\n"
-      ".inst 0x4ea16968  // bfcvtn2 v8.8h, v11.4s\n"
-      ".inst 0x4ea16b73  // bfcvtn2 v19.8h, v27.4s\n"
-      ".inst 0x4ea16bee  // bfcvtn2 v14.8h, v31.4s\n"
-      ".inst 0x4ea169fe  // bfcvtn2 v30.8h, v15.4s\n"
-      ".inst 0x4ea1692c  // bfcvtn2 v12.8h, v9.4s\n"
-      "str q5, [x27, #0x40]\n"
-      "str q6, [x27, #0x50]\n"
-      "str q23, [x27, #0x60]\n"
-      "str q8, [x27, #0x70]\n"
-      "str q19, [x27, #0x80]\n"
-      "str q14, [x27, #0x90]\n"
-      "str q30, [x27, #0xa0]\n"
-      "str q12, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "str q0, [x21, #0x40]\n"
+      ".inst 0x4ea16a8f  // bfcvtn2 v15.8h, v20.4s\n"
+      ".inst 0x4ea16bdc  // bfcvtn2 v28.8h, v30.4s\n"
+      "str q1, [x21, #0x50]\n"
+      "str q18, [x21, #0x60]\n"
+      "str q11, [x21, #0x70]\n"
+      "str q6, [x21, #0x80]\n"
+      "str q5, [x21, #0x90]\n"
+      "str q15, [x21, #0xa0]\n"
+      "str q28, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x28, #0xc\n"
+      "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr q5, [x9], #0x10\n"
-      "ldr q9, [x26], #0x10\n"
-      "sub x28, x28, #0xc\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q12, [x24], #0x10\n"
-      "cmp x28, #0xc\n"
-      "ldr q0, [x23], #0x10\n"
-      "ldr q29, [x22], #0x10\n"
-      "ldr q23, [x21], #0x10\n"
-      "ldr q27, [x20], #0x10\n"
-      "ldr q17, [x9], #0x10\n"
-      "ldr q31, [x26], #0x10\n"
-      "zip1 v16.4s, v5.4s, v19.4s\n"
-      "zip1 v7.4s, v9.4s, v12.4s\n"
-      "ldr q10, [x25], #0x10\n"
-      "ldr q11, [x24], #0x10\n"
-      "zip2 v20.4s, v5.4s, v19.4s\n"
-      "zip2 v28.4s, v9.4s, v12.4s\n"
-      "ldr q21, [x23], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
-      "zip1 v24.4s, v0.4s, v23.4s\n"
-      "zip1 v9.4s, v29.4s, v27.4s\n"
-      "ldr q18, [x21], #0x10\n"
-      "ldr q25, [x20], #0x10\n"
-      "zip2 v8.4s, v0.4s, v23.4s\n"
-      "zip2 v4.4s, v29.4s, v27.4s\n"
-      "ldr q27, [x9], #0x10\n"
+      "ldr q20, [x9], #0x10\n"
+      "ldr q9, [x28], #0x10\n"
+      "sub x24, x24, #0xc\n"
+      "cmp x24, #0xc\n"
+      "ldr q8, [x27], #0x10\n"
+      "ldr q1, [x26], #0x10\n"
+      "zip1 v7.4s, v20.4s, v8.4s\n"
+      "zip1 v19.4s, v9.4s, v1.4s\n"
+      "ldr q6, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip2 v5.4s, v20.4s, v8.4s\n"
+      "zip2 v18.4s, v9.4s, v1.4s\n"
+      "ldr q27, [x22], #0x10\n"
+      "ldr q14, [x20], #0x10\n"
+      "zip1 v26.4s, v6.4s, v27.4s\n"
+      "zip1 v15.4s, v16.4s, v14.4s\n"
+      "ldr q1, [x9], #0x10\n"
+      "ldr q30, [x28], #0x10\n"
+      "zip2 v24.4s, v6.4s, v27.4s\n"
+      "zip2 v25.4s, v16.4s, v14.4s\n"
+      "ldr q13, [x27], #0x10\n"
+      "ldr q17, [x26], #0x10\n"
+      "zip1 v10.4s, v1.4s, v13.4s\n"
+      "zip1 v16.4s, v30.4s, v17.4s\n"
+      "ldr q4, [x25], #0x10\n"
+      "ldr q11, [x23], #0x10\n"
+      "zip2 v0.4s, v1.4s, v13.4s\n"
+      "zip2 v27.4s, v30.4s, v17.4s\n"
+      "ldr q28, [x22], #0x10\n"
+      "ldr q12, [x20], #0x10\n"
+      "zip1 v22.4s, v4.4s, v28.4s\n"
+      "zip1 v13.4s, v11.4s, v12.4s\n"
+      "ldr q31, [x9], #0x10\n"
+      "ldr q17, [x28], #0x10\n"
+      "zip2 v14.4s, v4.4s, v28.4s\n"
+      "zip2 v12.4s, v11.4s, v12.4s\n"
+      "ldr q2, [x27], #0x10\n"
       "ldr q3, [x26], #0x10\n"
-      "zip1 v2.4s, v17.4s, v10.4s\n"
-      "zip1 v6.4s, v31.4s, v11.4s\n"
-      "ldr q1, [x25], #0x10\n"
-      "ldr q29, [x24], #0x10\n"
-      "zip2 v15.4s, v17.4s, v10.4s\n"
-      "zip2 v5.4s, v31.4s, v11.4s\n"
-      "ldr q31, [x23], #0x10\n"
-      "ldr q12, [x22], #0x10\n"
-      "zip1 v13.4s, v21.4s, v18.4s\n"
-      "zip1 v10.4s, v22.4s, v25.4s\n"
-      "ldr q19, [x21], #0x10\n"
-      "ldr q23, [x20], #0x10\n"
-      "zip2 v14.4s, v21.4s, v18.4s\n"
-      "zip2 v30.4s, v22.4s, v25.4s\n"
-      "zip1 v0.4s, v27.4s, v1.4s\n"
-      "zip1 v26.4s, v3.4s, v29.4s\n"
-      "zip2 v1.4s, v27.4s, v1.4s\n"
-      "zip2 v11.4s, v3.4s, v29.4s\n"
-      "zip1 v18.4s, v31.4s, v19.4s\n"
-      "zip1 v17.4s, v12.4s, v23.4s\n"
-      "zip2 v21.4s, v31.4s, v19.4s\n"
-      "zip2 v27.4s, v12.4s, v23.4s\n"
-      "zip1 v3.4s, v16.4s, v7.4s\n"
-      "zip1 v23.4s, v20.4s, v28.4s\n"
-      "zip1 v25.4s, v2.4s, v6.4s\n"
-      "zip1 v12.4s, v15.4s, v5.4s\n"
-      "zip1 v31.4s, v0.4s, v26.4s\n"
-      "zip1 v19.4s, v1.4s, v11.4s\n"
-      "zip1 v29.4s, v24.4s, v9.4s\n"
-      "zip1 v22.4s, v8.4s, v4.4s\n"
-      ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
-      "zip2 v16.4s, v16.4s, v7.4s\n"
-      "zip1 v7.4s, v13.4s, v10.4s\n"
-      ".inst 0x0ea16af7  // bfcvtn v23.4h, v23.4s\n"
-      "zip2 v20.4s, v20.4s, v28.4s\n"
-      "zip1 v28.4s, v14.4s, v30.4s\n"
-      ".inst 0x0ea16b39  // bfcvtn v25.4h, v25.4s\n"
-      "zip2 v6.4s, v2.4s, v6.4s\n"
-      "zip1 v2.4s, v18.4s, v17.4s\n"
-      ".inst 0x0ea1698c  // bfcvtn v12.4h, v12.4s\n"
-      "zip2 v5.4s, v15.4s, v5.4s\n"
-      "zip1 v15.4s, v21.4s, v27.4s\n"
+      "zip1 v8.4s, v31.4s, v2.4s\n"
+      "zip1 v4.4s, v17.4s, v3.4s\n"
+      "ldr q23, [x25], #0x10\n"
+      "ldr q1, [x23], #0x10\n"
+      "zip2 v28.4s, v31.4s, v2.4s\n"
+      "zip2 v29.4s, v17.4s, v3.4s\n"
+      "ldr q11, [x22], #0x10\n"
+      "ldr q17, [x20], #0x10\n"
+      "zip1 v9.4s, v23.4s, v11.4s\n"
+      "zip1 v21.4s, v1.4s, v17.4s\n"
+      "zip2 v11.4s, v23.4s, v11.4s\n"
+      "zip2 v17.4s, v1.4s, v17.4s\n"
+      "zip1 v2.4s, v7.4s, v19.4s\n"
+      "zip1 v31.4s, v5.4s, v18.4s\n"
+      "zip1 v3.4s, v10.4s, v16.4s\n"
+      "zip1 v6.4s, v0.4s, v27.4s\n"
+      "zip1 v1.4s, v8.4s, v4.4s\n"
+      "zip1 v30.4s, v28.4s, v29.4s\n"
+      "zip1 v20.4s, v26.4s, v15.4s\n"
+      "zip1 v23.4s, v24.4s, v25.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "zip2 v7.4s, v7.4s, v19.4s\n"
+      "zip1 v19.4s, v22.4s, v13.4s\n"
       ".inst 0x0ea16bff  // bfcvtn v31.4h, v31.4s\n"
-      "zip2 v26.4s, v0.4s, v26.4s\n"
-      ".inst 0x0ea16a60  // bfcvtn v0.4h, v19.4s\n"
-      "zip2 v1.4s, v1.4s, v11.4s\n"
-      ".inst 0x0ea16bab  // bfcvtn v11.4h, v29.4s\n"
-      "zip2 v9.4s, v24.4s, v9.4s\n"
-      ".inst 0x0ea16ad3  // bfcvtn v19.4h, v22.4s\n"
-      "zip2 v24.4s, v8.4s, v4.4s\n"
-      ".inst 0x0ea168e8  // bfcvtn v8.4h, v7.4s\n"
-      "zip2 v4.4s, v13.4s, v10.4s\n"
-      ".inst 0x0ea16b96  // bfcvtn v22.4h, v28.4s\n"
-      "zip2 v29.4s, v14.4s, v30.4s\n"
-      ".inst 0x0ea1685c  // bfcvtn v28.4h, v2.4s\n"
-      "zip2 v14.4s, v18.4s, v17.4s\n"
-      ".inst 0x0ea169f2  // bfcvtn v18.4h, v15.4s\n"
-      "zip2 v17.4s, v21.4s, v27.4s\n"
+      "zip2 v18.4s, v5.4s, v18.4s\n"
+      "zip1 v5.4s, v14.4s, v12.4s\n"
+      ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
+      "zip2 v16.4s, v10.4s, v16.4s\n"
+      "zip1 v10.4s, v9.4s, v21.4s\n"
+      ".inst 0x0ea168c6  // bfcvtn v6.4h, v6.4s\n"
+      "zip2 v0.4s, v0.4s, v27.4s\n"
+      "zip1 v27.4s, v11.4s, v17.4s\n"
+      ".inst 0x0ea16821  // bfcvtn v1.4h, v1.4s\n"
+      "zip2 v4.4s, v8.4s, v4.4s\n"
+      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
+      "zip2 v29.4s, v28.4s, v29.4s\n"
+      ".inst 0x0ea16a9c  // bfcvtn v28.4h, v20.4s\n"
+      "zip2 v15.4s, v26.4s, v15.4s\n"
+      ".inst 0x0ea16ae8  // bfcvtn v8.4h, v23.4s\n"
+      "zip2 v26.4s, v24.4s, v25.4s\n"
+      ".inst 0x0ea16a79  // bfcvtn v25.4h, v19.4s\n"
+      "zip2 v24.4s, v22.4s, v13.4s\n"
+      ".inst 0x0ea168b7  // bfcvtn v23.4h, v5.4s\n"
+      "zip2 v22.4s, v14.4s, v12.4s\n"
+      ".inst 0x0ea16945  // bfcvtn v5.4h, v10.4s\n"
+      "zip2 v20.4s, v9.4s, v21.4s\n"
+      ".inst 0x0ea16b73  // bfcvtn v19.4h, v27.4s\n"
+      "zip2 v17.4s, v11.4s, v17.4s\n"
+      ".inst 0x4ea168e2  // bfcvtn2 v2.8h, v7.4s\n"
+      ".inst 0x4ea16a5f  // bfcvtn2 v31.8h, v18.4s\n"
+      "str q2, [x21, #0x0]\n"
       ".inst 0x4ea16a03  // bfcvtn2 v3.8h, v16.4s\n"
-      ".inst 0x4ea16a97  // bfcvtn2 v23.8h, v20.4s\n"
-      ".inst 0x4ea168d9  // bfcvtn2 v25.8h, v6.4s\n"
-      ".inst 0x4ea168ac  // bfcvtn2 v12.8h, v5.4s\n"
-      ".inst 0x4ea16b5f  // bfcvtn2 v31.8h, v26.4s\n"
-      ".inst 0x4ea16820  // bfcvtn2 v0.8h, v1.4s\n"
-      ".inst 0x4ea1692b  // bfcvtn2 v11.8h, v9.4s\n"
-      ".inst 0x4ea16b13  // bfcvtn2 v19.8h, v24.4s\n"
-      "str q3, [x27, #0x0]\n"
-      ".inst 0x4ea16888  // bfcvtn2 v8.8h, v4.4s\n"
-      ".inst 0x4ea16bb6  // bfcvtn2 v22.8h, v29.4s\n"
-      "str q23, [x27, #0x10]\n"
-      ".inst 0x4ea169dc  // bfcvtn2 v28.8h, v14.4s\n"
-      ".inst 0x4ea16a32  // bfcvtn2 v18.8h, v17.4s\n"
-      "str q25, [x27, #0x20]\n"
-      "str q12, [x27, #0x30]\n"
-      "str q31, [x27, #0x40]\n"
-      "str q0, [x27, #0x50]\n"
-      "str q11, [x27, #0x60]\n"
-      "str q19, [x27, #0x70]\n"
-      "str q8, [x27, #0x80]\n"
-      "str q22, [x27, #0x90]\n"
-      "str q28, [x27, #0xa0]\n"
-      "str q18, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      ".inst 0x4ea16806  // bfcvtn2 v6.8h, v0.4s\n"
+      "str q31, [x21, #0x10]\n"
+      ".inst 0x4ea16881  // bfcvtn2 v1.8h, v4.4s\n"
+      ".inst 0x4ea16bbe  // bfcvtn2 v30.8h, v29.4s\n"
+      "str q3, [x21, #0x20]\n"
+      ".inst 0x4ea169fc  // bfcvtn2 v28.8h, v15.4s\n"
+      ".inst 0x4ea16b48  // bfcvtn2 v8.8h, v26.4s\n"
+      "str q6, [x21, #0x30]\n"
+      ".inst 0x4ea16b19  // bfcvtn2 v25.8h, v24.4s\n"
+      ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
+      "str q1, [x21, #0x40]\n"
+      ".inst 0x4ea16a85  // bfcvtn2 v5.8h, v20.4s\n"
+      ".inst 0x4ea16a33  // bfcvtn2 v19.8h, v17.4s\n"
+      "str q30, [x21, #0x50]\n"
+      "str q28, [x21, #0x60]\n"
+      "str q8, [x21, #0x70]\n"
+      "str q25, [x21, #0x80]\n"
+      "str q23, [x21, #0x90]\n"
+      "str q5, [x21, #0xa0]\n"
+      "str q19, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x28, 10f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "str q16, [x27, #0x80]\n"
-      "str q16, [x27, #0x90]\n"
-      "str q16, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
+      "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr q25, [x9], #0x10\n"
-      "ldr q24, [x26], #0x10\n"
-      "sub x28, x28, #0x4\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "cmp x28, #0x4\n"
-      "ldr q23, [x23], #0x10\n"
-      "ldr q19, [x22], #0x10\n"
-      "ldr q18, [x21], #0x10\n"
-      "ldr q17, [x20], #0x10\n"
-      "zip1 v22.4s, v25.4s, v21.4s\n"
-      "zip1 v16.4s, v24.4s, v20.4s\n"
-      "zip2 v21.4s, v25.4s, v21.4s\n"
-      "zip2 v20.4s, v24.4s, v20.4s\n"
-      "zip1 v27.4s, v23.4s, v18.4s\n"
-      "zip1 v26.4s, v19.4s, v17.4s\n"
-      "zip2 v25.4s, v23.4s, v18.4s\n"
-      "zip2 v24.4s, v19.4s, v17.4s\n"
-      "zip1 v19.4s, v22.4s, v16.4s\n"
-      "zip1 v18.4s, v21.4s, v20.4s\n"
+      "ldr q23, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.4s, v23.4s, v17.4s\n"
+      "zip1 v21.4s, v20.4s, v16.4s\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v28.4s, v23.4s, v17.4s\n"
+      "zip2 v20.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v27.4s, v19.4s, v17.4s\n"
+      "zip1 v26.4s, v18.4s, v16.4s\n"
+      "zip2 v25.4s, v19.4s, v17.4s\n"
+      "zip2 v24.4s, v18.4s, v16.4s\n"
+      "zip1 v19.4s, v22.4s, v21.4s\n"
+      "zip1 v18.4s, v28.4s, v20.4s\n"
       "zip1 v17.4s, v27.4s, v26.4s\n"
-      "zip2 v23.4s, v22.4s, v16.4s\n"
       "zip1 v16.4s, v25.4s, v24.4s\n"
-      "zip2 v22.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a75  // bfcvtn v21.4h, v19.4s\n"
-      ".inst 0x0ea16a54  // bfcvtn v20.4h, v18.4s\n"
+      ".inst 0x0ea16a77  // bfcvtn v23.4h, v19.4s\n"
+      "zip2 v22.4s, v22.4s, v21.4s\n"
+      ".inst 0x0ea16a55  // bfcvtn v21.4h, v18.4s\n"
+      "zip2 v20.4s, v28.4s, v20.4s\n"
       ".inst 0x0ea16a33  // bfcvtn v19.4h, v17.4s\n"
       "zip2 v18.4s, v27.4s, v26.4s\n"
       ".inst 0x0ea16a11  // bfcvtn v17.4h, v16.4s\n"
       "zip2 v16.4s, v25.4s, v24.4s\n"
-      ".inst 0x4ea16af5  // bfcvtn2 v21.8h, v23.4s\n"
-      ".inst 0x4ea16ad4  // bfcvtn2 v20.8h, v22.4s\n"
+      ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
+      ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
+      "str q23, [x21, #0x0]\n"
       ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
       ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x60]\n"
-      "str q17, [x27, #0x70]\n"
-      "add x27, x27, #0x20\n"
+      "str q21, [x21, #0x10]\n"
+      "str q19, [x21, #0x60]\n"
+      "str q17, [x21, #0x70]\n"
+      "add x21, x21, #0x20\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr s23, [x9], #0x4\n"
-      "ldr s22, [x26], #0x4\n"
-      "sub x28, x28, #0x1\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s17, [x24], #0x4\n"
-      "cmp x28, #0x1\n"
-      "ldr s21, [x23], #0x4\n"
-      "ldr s20, [x22], #0x4\n"
-      "ldr s18, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
-      "zip1 v19.4s, v23.4s, v19.4s\n"
-      "zip1 v17.4s, v22.4s, v17.4s\n"
-      "zip1 v18.4s, v21.4s, v18.4s\n"
-      "zip1 v16.4s, v20.4s, v16.4s\n"
+      "ldr s19, [x9], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
       "zip1 v17.4s, v19.4s, v17.4s\n"
       "zip1 v16.4s, v18.4s, v16.4s\n"
-      ".inst 0x0ea16a31  // bfcvtn v17.4h, v17.4s\n"
+      "ldr s20, [x25], #0x4\n"
+      "ldr s19, [x23], #0x4\n"
+      "zip1 v16.4s, v17.4s, v16.4s\n"
+      ".inst 0x0ea16a12  // bfcvtn v18.4h, v16.4s\n"
+      "ldr s17, [x22], #0x4\n"
+      "ldr s16, [x20], #0x4\n"
+      "zip1 v17.4s, v20.4s, v17.4s\n"
+      "zip1 v16.4s, v19.4s, v16.4s\n"
+      "zip1 v16.4s, v17.4s, v16.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      "str d17, [x27, #0x0]\n"
-      "str d16, [x27, #0x60]\n"
-      "add x27, x27, #0x8\n"
+      "str d18, [x21, #0x0]\n"
+      "str d16, [x21, #0x60]\n"
+      "add x21, x21, #0x8\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x8\n"
       "add %x[out], %x[out], #0xc0\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x9, %x[in]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x26, x27, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x9, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "csel x25, x25, %x[pad_row], GE\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x26, %x[in_stride]\n"
       "csel x26, x26, %x[pad_row], GT\n"
+      "csel x27, x27, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x28, x28, %x[pad_row], GT\n"
       "cmp x20, #0x18\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
-      "ldr q24, [x9], #0x10\n"
-      "ldr q20, [x26], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
+      "ldr q22, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
       "sub x20, x20, #0x18\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x18\n"
-      "ldr q25, [x9], #0x10\n"
-      "ldr q23, [x26], #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "ldr q27, [x9], #0x10\n"
-      "zip1 v21.4s, v24.4s, v19.4s\n"
-      "zip1 v6.4s, v20.4s, v17.4s\n"
-      "ldr q26, [x26], #0x10\n"
-      "ldr q22, [x25], #0x10\n"
-      "zip2 v19.4s, v24.4s, v19.4s\n"
-      "zip2 v0.4s, v20.4s, v17.4s\n"
-      "ldr q20, [x24], #0x10\n"
-      "ldr q3, [x9], #0x10\n"
-      "zip1 v17.4s, v25.4s, v18.4s\n"
-      "zip1 v4.4s, v23.4s, v16.4s\n"
-      "ldr q1, [x26], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
-      "zip2 v5.4s, v25.4s, v18.4s\n"
-      "zip2 v16.4s, v23.4s, v16.4s\n"
-      "ldr q23, [x24], #0x10\n"
-      "ldr q30, [x9], #0x10\n"
-      "zip1 v31.4s, v27.4s, v22.4s\n"
-      "zip1 v2.4s, v26.4s, v20.4s\n"
-      "ldr q25, [x26], #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "zip2 v29.4s, v27.4s, v22.4s\n"
-      "zip2 v28.4s, v26.4s, v20.4s\n"
-      "ldr q20, [x24], #0x10\n"
-      "ldr q27, [x9], #0x10\n"
-      "zip1 v22.4s, v3.4s, v24.4s\n"
-      "zip1 v26.4s, v1.4s, v23.4s\n"
-      "ldr q7, [x26], #0x10\n"
-      "ldr q11, [x25], #0x10\n"
-      "zip2 v24.4s, v3.4s, v24.4s\n"
-      "zip2 v14.4s, v1.4s, v23.4s\n"
-      "ldr q3, [x24], #0x10\n"
-      "zip1 v23.4s, v30.4s, v18.4s\n"
-      "zip1 v13.4s, v25.4s, v20.4s\n"
-      "zip2 v18.4s, v30.4s, v18.4s\n"
-      "zip2 v12.4s, v25.4s, v20.4s\n"
-      "zip1 v15.4s, v27.4s, v11.4s\n"
-      "zip1 v9.4s, v7.4s, v3.4s\n"
-      "zip2 v8.4s, v27.4s, v11.4s\n"
-      "zip2 v10.4s, v7.4s, v3.4s\n"
-      "zip1 v11.4s, v21.4s, v6.4s\n"
-      "zip1 v7.4s, v19.4s, v0.4s\n"
-      "zip1 v1.4s, v17.4s, v4.4s\n"
-      "zip1 v3.4s, v5.4s, v16.4s\n"
-      "zip1 v30.4s, v31.4s, v2.4s\n"
-      "zip1 v25.4s, v29.4s, v28.4s\n"
-      "zip1 v27.4s, v22.4s, v26.4s\n"
-      "zip1 v20.4s, v24.4s, v14.4s\n"
-      ".inst 0x0ea1696b  // bfcvtn v11.4h, v11.4s\n"
-      "zip2 v6.4s, v21.4s, v6.4s\n"
-      "zip1 v21.4s, v23.4s, v13.4s\n"
-      ".inst 0x0ea168e7  // bfcvtn v7.4h, v7.4s\n"
-      "zip2 v0.4s, v19.4s, v0.4s\n"
-      "zip1 v19.4s, v18.4s, v12.4s\n"
-      ".inst 0x0ea16821  // bfcvtn v1.4h, v1.4s\n"
-      "zip2 v4.4s, v17.4s, v4.4s\n"
-      "zip1 v17.4s, v15.4s, v9.4s\n"
-      ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
-      "zip2 v5.4s, v5.4s, v16.4s\n"
-      "zip1 v16.4s, v8.4s, v10.4s\n"
-      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
-      "zip2 v2.4s, v31.4s, v2.4s\n"
-      ".inst 0x0ea16b3f  // bfcvtn v31.4h, v25.4s\n"
-      "zip2 v29.4s, v29.4s, v28.4s\n"
-      ".inst 0x0ea16b7c  // bfcvtn v28.4h, v27.4s\n"
-      "zip2 v27.4s, v22.4s, v26.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v19.4s, v22.4s, v17.4s\n"
+      "zip1 v21.4s, v18.4s, v16.4s\n"
+      "ldr q24, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v10.4s, v22.4s, v17.4s\n"
+      "zip2 v2.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.4s, v24.4s, v17.4s\n"
+      "zip1 v4.4s, v20.4s, v16.4s\n"
+      "ldr q23, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v29.4s, v24.4s, v17.4s\n"
+      "zip2 v1.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v30.4s, v23.4s, v17.4s\n"
+      "zip1 v31.4s, v18.4s, v16.4s\n"
+      "ldr q24, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v23.4s, v23.4s, v17.4s\n"
+      "zip2 v28.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v25.4s, v24.4s, v17.4s\n"
+      "zip1 v26.4s, v20.4s, v16.4s\n"
+      "ldr q14, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v24.4s, v24.4s, v17.4s\n"
+      "zip2 v15.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v12.4s, v14.4s, v17.4s\n"
+      "zip1 v13.4s, v18.4s, v16.4s\n"
+      "ldr q7, [x9], #0x10\n"
+      "ldr q3, [x28], #0x10\n"
+      "zip2 v0.4s, v14.4s, v17.4s\n"
+      "zip2 v9.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v14.4s, v7.4s, v17.4s\n"
+      "zip1 v8.4s, v3.4s, v16.4s\n"
+      "zip2 v7.4s, v7.4s, v17.4s\n"
+      "zip2 v11.4s, v3.4s, v16.4s\n"
+      "zip1 v18.4s, v19.4s, v21.4s\n"
+      "zip1 v6.4s, v10.4s, v2.4s\n"
+      "zip1 v5.4s, v22.4s, v4.4s\n"
+      "zip1 v16.4s, v29.4s, v1.4s\n"
+      "zip1 v27.4s, v30.4s, v31.4s\n"
+      "zip1 v3.4s, v23.4s, v28.4s\n"
+      "zip1 v17.4s, v25.4s, v26.4s\n"
+      "zip1 v20.4s, v24.4s, v15.4s\n"
+      ".inst 0x0ea16a52  // bfcvtn v18.4h, v18.4s\n"
+      "zip2 v19.4s, v19.4s, v21.4s\n"
+      "zip1 v21.4s, v12.4s, v13.4s\n"
+      ".inst 0x0ea168c6  // bfcvtn v6.4h, v6.4s\n"
+      "zip2 v10.4s, v10.4s, v2.4s\n"
+      "zip1 v2.4s, v0.4s, v9.4s\n"
+      ".inst 0x0ea168a5  // bfcvtn v5.4h, v5.4s\n"
+      "zip2 v4.4s, v22.4s, v4.4s\n"
+      "zip1 v22.4s, v14.4s, v8.4s\n"
+      ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
+      "zip2 v1.4s, v29.4s, v1.4s\n"
+      "zip1 v29.4s, v7.4s, v11.4s\n"
+      ".inst 0x0ea16b7b  // bfcvtn v27.4h, v27.4s\n"
+      "zip2 v30.4s, v30.4s, v31.4s\n"
+      ".inst 0x0ea1687f  // bfcvtn v31.4h, v3.4s\n"
+      "zip2 v23.4s, v23.4s, v28.4s\n"
+      ".inst 0x0ea16a23  // bfcvtn v3.4h, v17.4s\n"
+      "zip2 v28.4s, v25.4s, v26.4s\n"
       ".inst 0x0ea16a9a  // bfcvtn v26.4h, v20.4s\n"
-      "zip2 v25.4s, v24.4s, v14.4s\n"
+      "zip2 v25.4s, v24.4s, v15.4s\n"
       ".inst 0x0ea16ab8  // bfcvtn v24.4h, v21.4s\n"
-      "zip2 v22.4s, v23.4s, v13.4s\n"
-      ".inst 0x0ea16a75  // bfcvtn v21.4h, v19.4s\n"
-      "zip2 v20.4s, v18.4s, v12.4s\n"
-      ".inst 0x0ea16a33  // bfcvtn v19.4h, v17.4s\n"
-      "zip2 v18.4s, v15.4s, v9.4s\n"
-      ".inst 0x0ea16a11  // bfcvtn v17.4h, v16.4s\n"
-      "zip2 v16.4s, v8.4s, v10.4s\n"
-      ".inst 0x4ea168cb  // bfcvtn2 v11.8h, v6.4s\n"
-      ".inst 0x4ea16807  // bfcvtn2 v7.8h, v0.4s\n"
-      ".inst 0x4ea16881  // bfcvtn2 v1.8h, v4.4s\n"
-      ".inst 0x4ea168a3  // bfcvtn2 v3.8h, v5.4s\n"
-      ".inst 0x4ea1685e  // bfcvtn2 v30.8h, v2.4s\n"
-      ".inst 0x4ea16bbf  // bfcvtn2 v31.8h, v29.4s\n"
-      "str q11, [x27, #0x0]\n"
-      "str q7, [x27, #0x10]\n"
-      ".inst 0x4ea16b7c  // bfcvtn2 v28.8h, v27.4s\n"
+      "zip2 v12.4s, v12.4s, v13.4s\n"
+      ".inst 0x0ea16855  // bfcvtn v21.4h, v2.4s\n"
+      "zip2 v13.4s, v0.4s, v9.4s\n"
+      ".inst 0x0ea16ac2  // bfcvtn v2.4h, v22.4s\n"
+      "zip2 v0.4s, v14.4s, v8.4s\n"
+      ".inst 0x0ea16ba9  // bfcvtn v9.4h, v29.4s\n"
+      "zip2 v17.4s, v7.4s, v11.4s\n"
+      ".inst 0x4ea16a72  // bfcvtn2 v18.8h, v19.4s\n"
+      ".inst 0x4ea16946  // bfcvtn2 v6.8h, v10.4s\n"
+      "str q18, [x21, #0x0]\n"
+      ".inst 0x4ea16885  // bfcvtn2 v5.8h, v4.4s\n"
+      ".inst 0x4ea16830  // bfcvtn2 v16.8h, v1.4s\n"
+      "str q6, [x21, #0x10]\n"
+      ".inst 0x4ea16bdb  // bfcvtn2 v27.8h, v30.4s\n"
+      ".inst 0x4ea16aff  // bfcvtn2 v31.8h, v23.4s\n"
+      "str q5, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      ".inst 0x4ea16b83  // bfcvtn2 v3.8h, v28.4s\n"
       ".inst 0x4ea16b3a  // bfcvtn2 v26.8h, v25.4s\n"
-      "str q1, [x27, #0x20]\n"
-      ".inst 0x4ea16ad8  // bfcvtn2 v24.8h, v22.4s\n"
-      ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
-      "str q3, [x27, #0x30]\n"
-      ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
-      ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
-      "str q30, [x27, #0x40]\n"
-      "str q31, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q28, [x27, #0x0]\n"
-      "str q26, [x27, #0x10]\n"
-      "str q24, [x27, #0x20]\n"
-      "str q21, [x27, #0x30]\n"
-      "str q19, [x27, #0x40]\n"
-      "str q17, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "str q27, [x21, #0x40]\n"
+      ".inst 0x4ea16998  // bfcvtn2 v24.8h, v12.4s\n"
+      ".inst 0x4ea169b5  // bfcvtn2 v21.8h, v13.4s\n"
+      "str q31, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      ".inst 0x4ea16802  // bfcvtn2 v2.8h, v0.4s\n"
+      ".inst 0x4ea16a29  // bfcvtn2 v9.8h, v17.4s\n"
+      "str q3, [x21, #0x0]\n"
+      "str q26, [x21, #0x10]\n"
+      "str q24, [x21, #0x20]\n"
+      "str q21, [x21, #0x30]\n"
+      "str q2, [x21, #0x40]\n"
+      "str q9, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
-      "ldr q24, [x9], #0x10\n"
-      "ldr q23, [x26], #0x10\n"
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
       "sub x20, x20, #0xc\n"
-      "ldr q22, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
       "cmp x20, #0xc\n"
-      "ldr q28, [x9], #0x10\n"
-      "ldr q27, [x26], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v26.4s, v19.4s, v17.4s\n"
+      "zip1 v25.4s, v18.4s, v16.4s\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v24.4s, v19.4s, v17.4s\n"
+      "zip2 v23.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v2.4s, v21.4s, v17.4s\n"
+      "zip1 v22.4s, v20.4s, v16.4s\n"
       "ldr q19, [x9], #0x10\n"
-      "zip1 v26.4s, v24.4s, v22.4s\n"
-      "zip1 v25.4s, v23.4s, v16.4s\n"
-      "ldr q18, [x26], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
-      "zip2 v24.4s, v24.4s, v22.4s\n"
-      "zip2 v23.4s, v23.4s, v16.4s\n"
-      "ldr q16, [x24], #0x10\n"
-      "zip1 v2.4s, v28.4s, v21.4s\n"
-      "zip1 v22.4s, v27.4s, v20.4s\n"
-      "zip2 v1.4s, v28.4s, v21.4s\n"
-      "zip2 v0.4s, v27.4s, v20.4s\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v1.4s, v21.4s, v17.4s\n"
+      "zip2 v0.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
       "zip1 v31.4s, v19.4s, v17.4s\n"
       "zip1 v30.4s, v18.4s, v16.4s\n"
       "zip2 v29.4s, v19.4s, v17.4s\n"
@@ -665,75 +650,66 @@ void a64_transpose_interleave_12_2x4_fp32bf16(bfloat16 *out, const float *in, si
       "zip2 v16.4s, v29.4s, v28.4s\n"
       ".inst 0x4ea16b5b  // bfcvtn2 v27.8h, v26.4s\n"
       ".inst 0x4ea16b19  // bfcvtn2 v25.8h, v24.4s\n"
+      "str q27, [x21, #0x0]\n"
       ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
       ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
+      "str q25, [x21, #0x10]\n"
       ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
       ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
-      "str q27, [x27, #0x0]\n"
-      "str q25, [x27, #0x10]\n"
-      "str q23, [x27, #0x20]\n"
-      "str q21, [x27, #0x30]\n"
-      "str q19, [x27, #0x40]\n"
-      "str q17, [x27, #0x50]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q23, [x21, #0x20]\n"
+      "str q21, [x21, #0x30]\n"
+      "str q19, [x21, #0x40]\n"
+      "str q17, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
-      "ldr q21, [x9], #0x10\n"
-      "ldr q20, [x26], #0x10\n"
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
+      "ldr q20, [x9], #0x10\n"
+      "ldr q19, [x28], #0x10\n"
       "sub x20, x20, #0x4\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x4\n"
-      "zip1 v18.4s, v21.4s, v19.4s\n"
-      "zip1 v16.4s, v20.4s, v17.4s\n"
-      "zip2 v21.4s, v21.4s, v19.4s\n"
-      "zip2 v20.4s, v20.4s, v17.4s\n"
-      "zip1 v17.4s, v18.4s, v16.4s\n"
-      "zip2 v19.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.4s, v20.4s, v17.4s\n"
+      "zip1 v18.4s, v19.4s, v16.4s\n"
+      "zip2 v21.4s, v20.4s, v17.4s\n"
+      "zip2 v20.4s, v19.4s, v16.4s\n"
+      "zip1 v17.4s, v22.4s, v18.4s\n"
       "zip1 v16.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a32  // bfcvtn v18.4h, v17.4s\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      ".inst 0x4ea16a72  // bfcvtn2 v18.8h, v19.4s\n"
-      ".inst 0x4ea16a30  // bfcvtn2 v16.8h, v17.4s\n"
-      "str q18, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      ".inst 0x0ea16a33  // bfcvtn v19.4h, v17.4s\n"
+      "zip2 v18.4s, v22.4s, v18.4s\n"
+      ".inst 0x0ea16a11  // bfcvtn v17.4h, v16.4s\n"
+      "zip2 v16.4s, v21.4s, v20.4s\n"
+      ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
+      ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
+      "str q19, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "add x21, x21, #0x20\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr s19, [x9], #0x4\n"
-      "ldr s18, [x26], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
       "sub x20, x20, #0x1\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s16, [x24], #0x4\n"
       "cmp x20, #0x1\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
       "zip1 v17.4s, v19.4s, v17.4s\n"
       "zip1 v16.4s, v18.4s, v16.4s\n"
       "zip1 v16.4s, v17.4s, v16.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      "str d16, [x27, #0x0]\n"
-      "add x27, x27, #0x8\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x60\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_s8s16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_s8s16.hpp
index ee37b4378f..0d6f8b1cd4 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_s8s16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_s8s16.hpp
@@ -34,236 +34,222 @@ void a64_transpose_interleave_12_s8s16(int16_t *out, const int8_t *in, size_t wi
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
       "ldr q3, [x25], #0x10\n"
-      "ldr q21, [x22], #0x10\n"
+      "ldr q21, [x23], #0x10\n"
+      "sshll2 v20.8h, v3.16b, #0x0\n"
+      "sshll v2.8h, v21.8b, #0x0\n"
+      "ldr q1, [x22], #0x10\n"
+      "ldr q19, [x20], #0x10\n"
+      "sshll2 v18.8h, v1.16b, #0x0\n"
+      "sshll v0.8h, v19.8b, #0x0\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d16, [x22], #0x8\n"
+      "sshll v31.8h, v17.8b, #0x0\n"
+      "sshll v30.8h, v16.8b, #0x0\n"
+      "ldr d29, [x23], #0x8\n"
+      "ldr d28, [x20], #0x8\n"
+      "sshll2 v27.8h, v21.16b, #0x0\n"
+      "sshll2 v26.8h, v19.16b, #0x0\n"
+      "dup v25.2d, v20.d[0]\n"
+      "dup v24.2d, v2.d[1]\n"
       "sub x24, x24, #0x18\n"
-      "ldr q2, [x21], #0x10\n"
-      "ldr q20, [x20], #0x10\n"
       "cmp x24, #0x18\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d17, [x21], #0x8\n"
-      "ldr d1, [x22], #0x8\n"
-      "sshll2 v18.8h, v3.16b, #0x0\n"
-      "sshll v0.8h, v21.8b, #0x0\n"
-      "ldr d31, [x20], #0x8\n"
-      "sshll2 v16.8h, v2.16b, #0x0\n"
-      "sshll v30.8h, v20.8b, #0x0\n"
-      "sshll v29.8h, v19.8b, #0x0\n"
-      "sshll v28.8h, v17.8b, #0x0\n"
-      "sshll2 v27.8h, v21.16b, #0x0\n"
-      "sshll2 v26.8h, v20.16b, #0x0\n"
-      "dup v17.2d, v18.d[0]\n"
-      "dup v25.2d, v0.d[1]\n"
-      "dup v24.2d, v16.d[0]\n"
-      "dup v23.2d, v30.d[1]\n"
-      "dup v22.2d, v18.d[1]\n"
-      "dup v21.2d, v29.d[1]\n"
-      "dup v20.2d, v16.d[1]\n"
-      "dup v19.2d, v28.d[1]\n"
-      "sshll v16.8h, v3.8b, #0x0\n"
-      "sshll v18.8h, v2.8b, #0x0\n"
-      "mov v17.d[1], v0.d[0]\n"
-      "mov v25.d[1], v27.d[0]\n"
-      "mov v24.d[1], v30.d[0]\n"
-      "mov v23.d[1], v26.d[0]\n"
-      "str q16, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "sshll v17.8h, v1.8b, #0x0\n"
-      "sshll v16.8h, v31.8b, #0x0\n"
-      "str q25, [x23, #0x20]\n"
-      "mov v22.d[1], v29.d[0]\n"
-      "mov v21.d[1], v27.d[1]\n"
-      "str q18, [x23, #0x30]\n"
-      "mov v20.d[1], v28.d[0]\n"
-      "mov v19.d[1], v26.d[1]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q23, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q22, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q17, [x23, #0x20]\n"
-      "str q20, [x23, #0x30]\n"
-      "str q19, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "dup v23.2d, v18.d[0]\n"
+      "dup v22.2d, v0.d[1]\n"
+      "dup v21.2d, v20.d[1]\n"
+      "dup v20.2d, v31.d[1]\n"
+      "dup v19.2d, v18.d[1]\n"
+      "dup v18.2d, v30.d[1]\n"
+      "sshll v17.8h, v3.8b, #0x0\n"
+      "sshll v16.8h, v1.8b, #0x0\n"
+      "str q17, [x21, #0x0]\n"
+      "mov v25.d[1], v2.d[0]\n"
+      "mov v24.d[1], v27.d[0]\n"
+      "str q25, [x21, #0x10]\n"
+      "mov v23.d[1], v0.d[0]\n"
+      "mov v22.d[1], v26.d[0]\n"
+      "str q24, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "sshll v17.8h, v29.8b, #0x0\n"
+      "sshll v16.8h, v28.8b, #0x0\n"
+      "str q23, [x21, #0x40]\n"
+      "mov v21.d[1], v31.d[0]\n"
+      "mov v20.d[1], v27.d[1]\n"
+      "str q22, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "mov v19.d[1], v30.d[0]\n"
+      "mov v18.d[1], v26.d[1]\n"
+      "str q21, [x21, #0x0]\n"
+      "str q20, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q19, [x21, #0x30]\n"
+      "str q18, [x21, #0x40]\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr d19, [x22], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "ldr d18, [x20], #0x8\n"
       "sub x24, x24, #0xc\n"
-      "ld1 { v19.s }[2], [x22], #0x4\n"
-      "ldr d17, [x25], #0x8\n"
       "cmp x24, #0xc\n"
+      "ld1 { v19.s }[2], [x23], #0x4\n"
       "ld1 { v18.s }[2], [x20], #0x4\n"
-      "ldr d16, [x21], #0x8\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v16.s }[2], [x21], #0x4\n"
       "sshll v25.8h, v19.8b, #0x0\n"
-      "sshll2 v24.8h, v19.16b, #0x0\n"
-      "sshll v23.8h, v18.8b, #0x0\n"
+      "sshll v24.8h, v18.8b, #0x0\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d16, [x22], #0x8\n"
+      "sshll2 v23.8h, v19.16b, #0x0\n"
       "sshll2 v22.8h, v18.16b, #0x0\n"
+      "ld1 { v17.s }[2], [x25], #0x4\n"
+      "ld1 { v16.s }[2], [x22], #0x4\n"
       "sshll2 v21.8h, v17.16b, #0x0\n"
       "sshll2 v20.8h, v16.16b, #0x0\n"
       "dup v19.2d, v25.d[1]\n"
-      "sshll v18.8h, v17.8b, #0x0\n"
-      "dup v17.2d, v23.d[1]\n"
+      "dup v18.2d, v24.d[1]\n"
+      "sshll v17.8h, v17.8b, #0x0\n"
       "sshll v16.8h, v16.8b, #0x0\n"
+      "str q17, [x21, #0x0]\n"
       "mov v21.d[1], v25.d[0]\n"
-      "mov v19.d[1], v24.d[0]\n"
-      "mov v20.d[1], v23.d[0]\n"
-      "mov v17.d[1], v22.d[0]\n"
-      "str q18, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q19, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q20, [x23, #0x40]\n"
-      "str q17, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "mov v19.d[1], v23.d[0]\n"
+      "str q21, [x21, #0x10]\n"
+      "mov v20.d[1], v24.d[0]\n"
+      "mov v18.d[1], v22.d[0]\n"
+      "str q19, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "str q20, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr s19, [x25], #0x4\n"
-      "ldr s18, [x22], #0x4\n"
+      "ldr s18, [x23], #0x4\n"
       "sub x24, x24, #0x4\n"
-      "ldr s17, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
       "cmp x24, #0x4\n"
+      "ldr s17, [x22], #0x4\n"
+      "ldr s16, [x20], #0x4\n"
       "sshll v19.8h, v19.8b, #0x0\n"
       "sshll v18.8h, v18.8b, #0x0\n"
       "sshll v17.8h, v17.8b, #0x0\n"
       "sshll v16.8h, v16.8b, #0x0\n"
-      "str d19, [x23, #0x0]\n"
-      "str d18, [x23, #0x18]\n"
-      "str d17, [x23, #0x30]\n"
-      "str d16, [x23, #0x48]\n"
-      "add x23, x23, #0x8\n"
+      "str d19, [x21, #0x0]\n"
+      "str d18, [x21, #0x18]\n"
+      "str d17, [x21, #0x30]\n"
+      "str d16, [x21, #0x48]\n"
+      "add x21, x21, #0x8\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr b19, [x25], #0x1\n"
-      "ldr b18, [x22], #0x1\n"
+      "ldr b18, [x23], #0x1\n"
       "sub x24, x24, #0x1\n"
-      "ldr b17, [x21], #0x1\n"
-      "ldr b16, [x20], #0x1\n"
       "cmp x24, #0x1\n"
+      "ldr b17, [x22], #0x1\n"
+      "ldr b16, [x20], #0x1\n"
       "sshll v19.8h, v19.8b, #0x0\n"
       "sshll v18.8h, v18.8b, #0x0\n"
       "sshll v17.8h, v17.8b, #0x0\n"
       "sshll v16.8h, v16.8b, #0x0\n"
-      "str h19, [x23, #0x0]\n"
-      "str h18, [x23, #0x18]\n"
-      "str h17, [x23, #0x30]\n"
-      "str h16, [x23, #0x48]\n"
-      "add x23, x23, #0x2\n"
+      "str h19, [x21, #0x0]\n"
+      "str h18, [x21, #0x18]\n"
+      "str h17, [x21, #0x30]\n"
+      "str h16, [x21, #0x48]\n"
+      "add x21, x21, #0x2\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0x60\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x18\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
-      "ldr q18, [x25], #0x10\n"
-      "sub x20, x20, #0x18\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
+      "ldr q20, [x25], #0x10\n"
       "ldr d16, [x25], #0x8\n"
+      "sshll2 v19.8h, v20.16b, #0x0\n"
+      "sshll v18.8h, v16.8b, #0x0\n"
+      "dup v17.2d, v19.d[1]\n"
+      "sub x20, x20, #0x18\n"
+      "sshll v16.8h, v20.8b, #0x0\n"
+      "str q16, [x21, #0x0]\n"
+      "dup v16.2d, v19.d[0]\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
       "cmp x20, #0x18\n"
-      "sshll2 v17.8h, v18.16b, #0x0\n"
-      "sshll v16.8h, v16.8b, #0x0\n"
-      "sshll v19.8h, v18.8b, #0x0\n"
-      "dup v18.2d, v17.d[1]\n"
-      "dup v17.2d, v17.d[0]\n"
-      "mov v18.d[1], v16.d[0]\n"
-      "dup v16.2d, v16.d[1]\n"
-      "str q19, [x23, #0x0]\n"
-      "str d17, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q18, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "mov v17.d[1], v18.d[0]\n"
+      "dup v16.2d, v18.d[1]\n"
+      "str q17, [x21, #0x0]\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
       "ldr d16, [x25], #0x8\n"
-      "sub x20, x20, #0xc\n"
       "ld1 { v16.s }[2], [x25], #0x4\n"
+      "sub x20, x20, #0xc\n"
       "cmp x20, #0xc\n"
       "sshll v17.8h, v16.8b, #0x0\n"
       "sshll2 v16.8h, v16.16b, #0x0\n"
-      "str q17, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q17, [x21, #0x0]\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr s16, [x25], #0x4\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
       "sshll v16.8h, v16.8b, #0x0\n"
-      "str d16, [x23, #0x0]\n"
-      "add x23, x23, #0x8\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr b16, [x25], #0x1\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
       "sshll v16.8h, v16.8b, #0x0\n"
-      "str h16, [x23, #0x0]\n"
-      "add x23, x23, #0x2\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str h16, [x21, #0x0]\n"
+      "add x21, x21, #0x2\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x18\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_u8u16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_u8u16.hpp
index 3bfa4a0e42..b263e6c41d 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_u8u16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_12_u8u16.hpp
@@ -34,236 +34,222 @@ void a64_transpose_interleave_12_u8u16(uint16_t *out, const uint8_t *in, size_t
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
       "ldr q3, [x25], #0x10\n"
-      "ldr q21, [x22], #0x10\n"
+      "ldr q21, [x23], #0x10\n"
+      "ushll2 v20.8h, v3.16b, #0x0\n"
+      "ushll v2.8h, v21.8b, #0x0\n"
+      "ldr q1, [x22], #0x10\n"
+      "ldr q19, [x20], #0x10\n"
+      "ushll2 v18.8h, v1.16b, #0x0\n"
+      "ushll v0.8h, v19.8b, #0x0\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d16, [x22], #0x8\n"
+      "ushll v31.8h, v17.8b, #0x0\n"
+      "ushll v30.8h, v16.8b, #0x0\n"
+      "ldr d29, [x23], #0x8\n"
+      "ldr d28, [x20], #0x8\n"
+      "ushll2 v27.8h, v21.16b, #0x0\n"
+      "ushll2 v26.8h, v19.16b, #0x0\n"
+      "dup v25.2d, v20.d[0]\n"
+      "dup v24.2d, v2.d[1]\n"
       "sub x24, x24, #0x18\n"
-      "ldr q2, [x21], #0x10\n"
-      "ldr q20, [x20], #0x10\n"
       "cmp x24, #0x18\n"
-      "ldr d19, [x25], #0x8\n"
-      "ldr d17, [x21], #0x8\n"
-      "ldr d1, [x22], #0x8\n"
-      "ushll2 v18.8h, v3.16b, #0x0\n"
-      "ushll v0.8h, v21.8b, #0x0\n"
-      "ldr d31, [x20], #0x8\n"
-      "ushll2 v16.8h, v2.16b, #0x0\n"
-      "ushll v30.8h, v20.8b, #0x0\n"
-      "ushll v29.8h, v19.8b, #0x0\n"
-      "ushll v28.8h, v17.8b, #0x0\n"
-      "ushll2 v27.8h, v21.16b, #0x0\n"
-      "ushll2 v26.8h, v20.16b, #0x0\n"
-      "dup v17.2d, v18.d[0]\n"
-      "dup v25.2d, v0.d[1]\n"
-      "dup v24.2d, v16.d[0]\n"
-      "dup v23.2d, v30.d[1]\n"
-      "dup v22.2d, v18.d[1]\n"
-      "dup v21.2d, v29.d[1]\n"
-      "dup v20.2d, v16.d[1]\n"
-      "dup v19.2d, v28.d[1]\n"
-      "ushll v16.8h, v3.8b, #0x0\n"
-      "ushll v18.8h, v2.8b, #0x0\n"
-      "mov v17.d[1], v0.d[0]\n"
-      "mov v25.d[1], v27.d[0]\n"
-      "mov v24.d[1], v30.d[0]\n"
-      "mov v23.d[1], v26.d[0]\n"
-      "str q16, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "ushll v17.8h, v1.8b, #0x0\n"
-      "ushll v16.8h, v31.8b, #0x0\n"
-      "str q25, [x23, #0x20]\n"
-      "mov v22.d[1], v29.d[0]\n"
-      "mov v21.d[1], v27.d[1]\n"
-      "str q18, [x23, #0x30]\n"
-      "mov v20.d[1], v28.d[0]\n"
-      "mov v19.d[1], v26.d[1]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q23, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q22, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q17, [x23, #0x20]\n"
-      "str q20, [x23, #0x30]\n"
-      "str q19, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "dup v23.2d, v18.d[0]\n"
+      "dup v22.2d, v0.d[1]\n"
+      "dup v21.2d, v20.d[1]\n"
+      "dup v20.2d, v31.d[1]\n"
+      "dup v19.2d, v18.d[1]\n"
+      "dup v18.2d, v30.d[1]\n"
+      "ushll v17.8h, v3.8b, #0x0\n"
+      "ushll v16.8h, v1.8b, #0x0\n"
+      "str q17, [x21, #0x0]\n"
+      "mov v25.d[1], v2.d[0]\n"
+      "mov v24.d[1], v27.d[0]\n"
+      "str q25, [x21, #0x10]\n"
+      "mov v23.d[1], v0.d[0]\n"
+      "mov v22.d[1], v26.d[0]\n"
+      "str q24, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "ushll v17.8h, v29.8b, #0x0\n"
+      "ushll v16.8h, v28.8b, #0x0\n"
+      "str q23, [x21, #0x40]\n"
+      "mov v21.d[1], v31.d[0]\n"
+      "mov v20.d[1], v27.d[1]\n"
+      "str q22, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "mov v19.d[1], v30.d[0]\n"
+      "mov v18.d[1], v26.d[1]\n"
+      "str q21, [x21, #0x0]\n"
+      "str q20, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q19, [x21, #0x30]\n"
+      "str q18, [x21, #0x40]\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr d19, [x22], #0x8\n"
+      "ldr d19, [x23], #0x8\n"
       "ldr d18, [x20], #0x8\n"
       "sub x24, x24, #0xc\n"
-      "ld1 { v19.s }[2], [x22], #0x4\n"
-      "ldr d17, [x25], #0x8\n"
       "cmp x24, #0xc\n"
+      "ld1 { v19.s }[2], [x23], #0x4\n"
       "ld1 { v18.s }[2], [x20], #0x4\n"
-      "ldr d16, [x21], #0x8\n"
-      "ld1 { v17.s }[2], [x25], #0x4\n"
-      "ld1 { v16.s }[2], [x21], #0x4\n"
       "ushll v25.8h, v19.8b, #0x0\n"
-      "ushll2 v24.8h, v19.16b, #0x0\n"
-      "ushll v23.8h, v18.8b, #0x0\n"
+      "ushll v24.8h, v18.8b, #0x0\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d16, [x22], #0x8\n"
+      "ushll2 v23.8h, v19.16b, #0x0\n"
       "ushll2 v22.8h, v18.16b, #0x0\n"
+      "ld1 { v17.s }[2], [x25], #0x4\n"
+      "ld1 { v16.s }[2], [x22], #0x4\n"
       "ushll2 v21.8h, v17.16b, #0x0\n"
       "ushll2 v20.8h, v16.16b, #0x0\n"
       "dup v19.2d, v25.d[1]\n"
-      "ushll v18.8h, v17.8b, #0x0\n"
-      "dup v17.2d, v23.d[1]\n"
+      "dup v18.2d, v24.d[1]\n"
+      "ushll v17.8h, v17.8b, #0x0\n"
       "ushll v16.8h, v16.8b, #0x0\n"
+      "str q17, [x21, #0x0]\n"
       "mov v21.d[1], v25.d[0]\n"
-      "mov v19.d[1], v24.d[0]\n"
-      "mov v20.d[1], v23.d[0]\n"
-      "mov v17.d[1], v22.d[0]\n"
-      "str q18, [x23, #0x0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q19, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q20, [x23, #0x40]\n"
-      "str q17, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "mov v19.d[1], v23.d[0]\n"
+      "str q21, [x21, #0x10]\n"
+      "mov v20.d[1], v24.d[0]\n"
+      "mov v18.d[1], v22.d[0]\n"
+      "str q19, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "str q20, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr s19, [x25], #0x4\n"
-      "ldr s18, [x22], #0x4\n"
+      "ldr s18, [x23], #0x4\n"
       "sub x24, x24, #0x4\n"
-      "ldr s17, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
       "cmp x24, #0x4\n"
+      "ldr s17, [x22], #0x4\n"
+      "ldr s16, [x20], #0x4\n"
       "ushll v19.8h, v19.8b, #0x0\n"
       "ushll v18.8h, v18.8b, #0x0\n"
       "ushll v17.8h, v17.8b, #0x0\n"
       "ushll v16.8h, v16.8b, #0x0\n"
-      "str d19, [x23, #0x0]\n"
-      "str d18, [x23, #0x18]\n"
-      "str d17, [x23, #0x30]\n"
-      "str d16, [x23, #0x48]\n"
-      "add x23, x23, #0x8\n"
+      "str d19, [x21, #0x0]\n"
+      "str d18, [x21, #0x18]\n"
+      "str d17, [x21, #0x30]\n"
+      "str d16, [x21, #0x48]\n"
+      "add x21, x21, #0x8\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr b19, [x25], #0x1\n"
-      "ldr b18, [x22], #0x1\n"
+      "ldr b18, [x23], #0x1\n"
       "sub x24, x24, #0x1\n"
-      "ldr b17, [x21], #0x1\n"
-      "ldr b16, [x20], #0x1\n"
       "cmp x24, #0x1\n"
+      "ldr b17, [x22], #0x1\n"
+      "ldr b16, [x20], #0x1\n"
       "ushll v19.8h, v19.8b, #0x0\n"
       "ushll v18.8h, v18.8b, #0x0\n"
       "ushll v17.8h, v17.8b, #0x0\n"
       "ushll v16.8h, v16.8b, #0x0\n"
-      "str h19, [x23, #0x0]\n"
-      "str h18, [x23, #0x18]\n"
-      "str h17, [x23, #0x30]\n"
-      "str h16, [x23, #0x48]\n"
-      "add x23, x23, #0x2\n"
+      "str h19, [x21, #0x0]\n"
+      "str h18, [x21, #0x18]\n"
+      "str h17, [x21, #0x30]\n"
+      "str h16, [x21, #0x48]\n"
+      "add x21, x21, #0x2\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0x60\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x18\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
-      "ldr q18, [x25], #0x10\n"
-      "sub x20, x20, #0x18\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
+      "ldr q20, [x25], #0x10\n"
       "ldr d16, [x25], #0x8\n"
+      "ushll2 v19.8h, v20.16b, #0x0\n"
+      "ushll v18.8h, v16.8b, #0x0\n"
+      "dup v17.2d, v19.d[1]\n"
+      "sub x20, x20, #0x18\n"
+      "ushll v16.8h, v20.8b, #0x0\n"
+      "str q16, [x21, #0x0]\n"
+      "dup v16.2d, v19.d[0]\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
       "cmp x20, #0x18\n"
-      "ushll2 v17.8h, v18.16b, #0x0\n"
-      "ushll v16.8h, v16.8b, #0x0\n"
-      "ushll v19.8h, v18.8b, #0x0\n"
-      "dup v18.2d, v17.d[1]\n"
-      "dup v17.2d, v17.d[0]\n"
-      "mov v18.d[1], v16.d[0]\n"
-      "dup v16.2d, v16.d[1]\n"
-      "str q19, [x23, #0x0]\n"
-      "str d17, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q18, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "mov v17.d[1], v18.d[0]\n"
+      "dup v16.2d, v18.d[1]\n"
+      "str q17, [x21, #0x0]\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
       "ldr d16, [x25], #0x8\n"
-      "sub x20, x20, #0xc\n"
       "ld1 { v16.s }[2], [x25], #0x4\n"
+      "sub x20, x20, #0xc\n"
       "cmp x20, #0xc\n"
       "ushll v17.8h, v16.8b, #0x0\n"
       "ushll2 v16.8h, v16.16b, #0x0\n"
-      "str q17, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q17, [x21, #0x0]\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr s16, [x25], #0x4\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
       "ushll v16.8h, v16.8b, #0x0\n"
-      "str d16, [x23, #0x0]\n"
-      "add x23, x23, #0x8\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr b16, [x25], #0x1\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
       "ushll v16.8h, v16.8b, #0x0\n"
-      "str h16, [x23, #0x0]\n"
-      "add x23, x23, #0x2\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str h16, [x21, #0x0]\n"
+      "add x21, x21, #0x2\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x18\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x4.hpp
index 7c79c5f7f0..087dc923e8 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x4.hpp
@@ -40,16 +40,14 @@ void a64_transpose_interleave_16_1x4(uint8_t *out, const uint8_t *in, size_t wid
 
     __asm__ __volatile__(
       "cmp %x[height], #0x10\n"
-      "blt 9f\n"
+      "blt 8f\n"
       "1:"  // Main row loop: Head
       "mov x17, %x[in]\n"
-      "mov x16, %x[width]\n"
-      "mov x15, %x[out]\n"
-      "sub %x[height], %x[height], #0x10\n"
-      "add x14, x17, %x[in_stride]\n"
+      "add x16, x17, %x[in_stride]\n"
+      "add x15, x16, %x[in_stride]\n"
+      "add x14, x15, %x[in_stride]\n"
       "add x13, x14, %x[in_stride]\n"
       "add x12, x13, %x[in_stride]\n"
-      "cmp x16, #0x10\n"
       "add x11, x12, %x[in_stride]\n"
       "add x10, x11, %x[in_stride]\n"
       "add x9, x10, %x[in_stride]\n"
@@ -57,268 +55,244 @@ void a64_transpose_interleave_16_1x4(uint8_t *out, const uint8_t *in, size_t wid
       "add x27, x28, %x[in_stride]\n"
       "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x10\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x10\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
-      "ldr q19, [x17], #0x10\n"
-      "ldr q18, [x14], #0x10\n"
-      "sub x16, x16, #0x10\n"
-      "ldr q17, [x13], #0x10\n"
-      "ldr q16, [x12], #0x10\n"
-      "cmp x16, #0x10\n"
-      "ldr q24, [x11], #0x10\n"
-      "ldr q23, [x10], #0x10\n"
-      "ldr q22, [x9], #0x10\n"
-      "ldr q21, [x28], #0x10\n"
-      "ldr q30, [x27], #0x10\n"
-      "ldr q29, [x26], #0x10\n"
-      "zip1 v3.16b, v19.16b, v17.16b\n"
-      "zip1 v2.16b, v18.16b, v16.16b\n"
-      "ldr q28, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "zip2 v1.16b, v19.16b, v17.16b\n"
-      "zip2 v27.16b, v18.16b, v16.16b\n"
-      "ldr q19, [x23], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "zip1 v26.16b, v24.16b, v22.16b\n"
-      "zip1 v25.16b, v23.16b, v21.16b\n"
-      "ldr q17, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "zip2 v24.16b, v24.16b, v22.16b\n"
-      "zip2 v23.16b, v23.16b, v21.16b\n"
-      "zip1 v22.16b, v30.16b, v28.16b\n"
-      "zip1 v21.16b, v29.16b, v20.16b\n"
-      "zip2 v0.16b, v30.16b, v28.16b\n"
-      "zip2 v20.16b, v29.16b, v20.16b\n"
+      "ldr q21, [x17], #0x10\n"
+      "ldr q20, [x16], #0x10\n"
+      "sub x24, x24, #0x10\n"
+      "cmp x24, #0x10\n"
+      "ldr q17, [x15], #0x10\n"
+      "ldr q16, [x14], #0x10\n"
+      "zip1 v3.16b, v21.16b, v17.16b\n"
+      "zip1 v2.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x13], #0x10\n"
+      "ldr q18, [x12], #0x10\n"
+      "zip2 v1.16b, v21.16b, v17.16b\n"
+      "zip2 v0.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x11], #0x10\n"
+      "ldr q16, [x10], #0x10\n"
       "zip1 v31.16b, v19.16b, v17.16b\n"
       "zip1 v30.16b, v18.16b, v16.16b\n"
-      "zip2 v29.16b, v19.16b, v17.16b\n"
-      "zip2 v28.16b, v18.16b, v16.16b\n"
-      "zip1 v19.16b, v3.16b, v2.16b\n"
-      "zip2 v18.16b, v3.16b, v2.16b\n"
-      "zip1 v17.16b, v1.16b, v27.16b\n"
-      "zip2 v16.16b, v1.16b, v27.16b\n"
-      "zip1 v27.16b, v26.16b, v25.16b\n"
-      "zip2 v26.16b, v26.16b, v25.16b\n"
-      "zip1 v25.16b, v24.16b, v23.16b\n"
+      "ldr q25, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v24.16b, v19.16b, v17.16b\n"
+      "zip2 v23.16b, v18.16b, v16.16b\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.16b, v25.16b, v17.16b\n"
+      "zip1 v21.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v29.16b, v25.16b, v17.16b\n"
+      "zip2 v20.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v28.16b, v19.16b, v17.16b\n"
+      "zip1 v27.16b, v18.16b, v16.16b\n"
+      "zip2 v26.16b, v19.16b, v17.16b\n"
+      "zip2 v25.16b, v18.16b, v16.16b\n"
+      "zip1 v16.16b, v3.16b, v2.16b\n"
+      "zip2 v17.16b, v3.16b, v2.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v16.16b, v1.16b, v0.16b\n"
+      "zip2 v19.16b, v1.16b, v0.16b\n"
+      "str q17, [x21, #0x10]\n"
+      "zip1 v18.16b, v31.16b, v30.16b\n"
+      "zip2 v17.16b, v31.16b, v30.16b\n"
+      "str q16, [x21, #0x20]\n"
+      "zip1 v16.16b, v24.16b, v23.16b\n"
       "zip2 v24.16b, v24.16b, v23.16b\n"
-      "str q19, [x15, #0x0]\n"
+      "str q19, [x21, #0x30]\n"
       "zip1 v23.16b, v22.16b, v21.16b\n"
       "zip2 v22.16b, v22.16b, v21.16b\n"
-      "str q18, [x15, #0x10]\n"
-      "zip1 v21.16b, v0.16b, v20.16b\n"
-      "zip2 v20.16b, v0.16b, v20.16b\n"
-      "str q17, [x15, #0x20]\n"
-      "zip1 v19.16b, v31.16b, v30.16b\n"
-      "zip2 v18.16b, v31.16b, v30.16b\n"
-      "str q16, [x15, #0x30]\n"
-      "zip1 v17.16b, v29.16b, v28.16b\n"
-      "zip2 v16.16b, v29.16b, v28.16b\n"
-      "str q27, [x15, #0x40]\n"
-      "str q26, [x15, #0x50]\n"
-      "str q25, [x15, #0x60]\n"
-      "str q24, [x15, #0x70]\n"
-      "str q23, [x15, #0x80]\n"
-      "str q22, [x15, #0x90]\n"
-      "str q21, [x15, #0xa0]\n"
-      "str q20, [x15, #0xb0]\n"
-      "str q19, [x15, #0xc0]\n"
-      "str q18, [x15, #0xd0]\n"
-      "str q17, [x15, #0xe0]\n"
-      "str q16, [x15, #0xf0]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "str q18, [x21, #0x40]\n"
+      "zip1 v21.16b, v29.16b, v20.16b\n"
+      "zip2 v20.16b, v29.16b, v20.16b\n"
+      "str q17, [x21, #0x50]\n"
+      "zip1 v19.16b, v28.16b, v27.16b\n"
+      "zip2 v18.16b, v28.16b, v27.16b\n"
+      "str q16, [x21, #0x60]\n"
+      "zip1 v17.16b, v26.16b, v25.16b\n"
+      "zip2 v16.16b, v26.16b, v25.16b\n"
+      "str q24, [x21, #0x70]\n"
+      "str q23, [x21, #0x80]\n"
+      "str q22, [x21, #0x90]\n"
+      "str q21, [x21, #0xa0]\n"
+      "str q20, [x21, #0xb0]\n"
+      "str q19, [x21, #0xc0]\n"
+      "str q18, [x21, #0xd0]\n"
+      "str q17, [x21, #0xe0]\n"
+      "str q16, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x16, 8f\n"
-      "cmp x16, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x15, #0x0]\n"
-      "str q16, [x15, #0x10]\n"
-      "str q16, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "str q16, [x15, #0x40]\n"
-      "str q16, [x15, #0x50]\n"
-      "str q16, [x15, #0x60]\n"
-      "str q16, [x15, #0x70]\n"
-      "str q16, [x15, #0x80]\n"
-      "str q16, [x15, #0x90]\n"
-      "str q16, [x15, #0xa0]\n"
-      "str q16, [x15, #0xb0]\n"
-      "str q16, [x15, #0xc0]\n"
-      "str q16, [x15, #0xd0]\n"
-      "str q16, [x15, #0xe0]\n"
-      "str q16, [x15, #0xf0]\n"
+      "cmp x24, #0x4\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 4 loop: loop
-      "ldr s23, [x17], #0x4\n"
-      "ldr s21, [x14], #0x4\n"
-      "sub x16, x16, #0x4\n"
-      "ldr s20, [x13], #0x4\n"
-      "ldr s19, [x12], #0x4\n"
-      "cmp x16, #0x4\n"
-      "ldr s22, [x11], #0x4\n"
-      "ldr s18, [x10], #0x4\n"
-      "ldr s17, [x9], #0x4\n"
-      "ldr s16, [x28], #0x4\n"
-      "ldr s27, [x27], #0x4\n"
-      "ldr s26, [x26], #0x4\n"
-      "zip1 v25.16b, v23.16b, v20.16b\n"
-      "zip1 v21.16b, v21.16b, v19.16b\n"
+      "ldr s19, [x17], #0x4\n"
+      "ldr s18, [x16], #0x4\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr s17, [x15], #0x4\n"
+      "ldr s16, [x14], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr s19, [x13], #0x4\n"
+      "ldr s18, [x12], #0x4\n"
+      "zip1 v22.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x11], #0x4\n"
+      "ldr s16, [x10], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr s19, [x9], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
+      "zip1 v21.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
       "ldr s20, [x25], #0x4\n"
-      "ldr s19, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s23, [x22], #0x4\n"
-      "zip1 v22.16b, v22.16b, v17.16b\n"
-      "zip1 v17.16b, v18.16b, v16.16b\n"
-      "ldr s18, [x21], #0x4\n"
+      "ldr s19, [x23], #0x4\n"
+      "zip1 v18.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x22], #0x4\n"
       "ldr s16, [x20], #0x4\n"
-      "zip1 v21.16b, v25.16b, v21.16b\n"
-      "zip1 v20.16b, v27.16b, v20.16b\n"
-      "zip1 v19.16b, v26.16b, v19.16b\n"
-      "zip1 v17.16b, v22.16b, v17.16b\n"
-      "zip1 v18.16b, v24.16b, v18.16b\n"
-      "zip1 v16.16b, v23.16b, v16.16b\n"
-      "str q21, [x15, #0x0]\n"
-      "str q17, [x15, #0x40]\n"
-      "zip1 v17.16b, v20.16b, v19.16b\n"
-      "zip1 v16.16b, v18.16b, v16.16b\n"
-      "str q17, [x15, #0x80]\n"
-      "str q16, [x15, #0xc0]\n"
-      "add x15, x15, #0x10\n"
+      "zip1 v17.16b, v20.16b, v17.16b\n"
+      "zip1 v16.16b, v19.16b, v16.16b\n"
+      "str q22, [x21, #0x0]\n"
+      "zip1 v16.16b, v17.16b, v16.16b\n"
+      "str q21, [x21, #0x40]\n"
+      "str q18, [x21, #0x80]\n"
+      "str q16, [x21, #0xc0]\n"
+      "add x21, x21, #0x10\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 4 loop: skip
-      "cmp x16, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 1 loop: loop
-      "ldr b23, [x17], #0x1\n"
-      "ldr b21, [x14], #0x1\n"
-      "sub x16, x16, #0x1\n"
-      "ldr b20, [x13], #0x1\n"
-      "ldr b19, [x12], #0x1\n"
-      "cmp x16, #0x1\n"
-      "ldr b22, [x11], #0x1\n"
-      "ldr b18, [x10], #0x1\n"
-      "ldr b17, [x9], #0x1\n"
-      "ldr b16, [x28], #0x1\n"
-      "ldr b27, [x27], #0x1\n"
-      "ldr b26, [x26], #0x1\n"
-      "zip1 v25.16b, v23.16b, v20.16b\n"
-      "zip1 v21.16b, v21.16b, v19.16b\n"
+      "ldr b19, [x17], #0x1\n"
+      "ldr b18, [x16], #0x1\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr b17, [x15], #0x1\n"
+      "ldr b16, [x14], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr b19, [x13], #0x1\n"
+      "ldr b18, [x12], #0x1\n"
+      "zip1 v22.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x11], #0x1\n"
+      "ldr b16, [x10], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr b19, [x9], #0x1\n"
+      "ldr b18, [x28], #0x1\n"
+      "zip1 v21.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x27], #0x1\n"
+      "ldr b16, [x26], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
       "ldr b20, [x25], #0x1\n"
-      "ldr b19, [x24], #0x1\n"
-      "ldr b24, [x23], #0x1\n"
-      "ldr b23, [x22], #0x1\n"
-      "zip1 v22.16b, v22.16b, v17.16b\n"
-      "zip1 v17.16b, v18.16b, v16.16b\n"
-      "ldr b18, [x21], #0x1\n"
+      "ldr b19, [x23], #0x1\n"
+      "zip1 v18.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x22], #0x1\n"
       "ldr b16, [x20], #0x1\n"
-      "zip1 v21.16b, v25.16b, v21.16b\n"
-      "zip1 v20.16b, v27.16b, v20.16b\n"
-      "zip1 v19.16b, v26.16b, v19.16b\n"
-      "zip1 v17.16b, v22.16b, v17.16b\n"
-      "zip1 v18.16b, v24.16b, v18.16b\n"
-      "zip1 v16.16b, v23.16b, v16.16b\n"
-      "str s21, [x15, #0x0]\n"
-      "str s17, [x15, #0x40]\n"
-      "zip1 v17.16b, v20.16b, v19.16b\n"
-      "zip1 v16.16b, v18.16b, v16.16b\n"
-      "str s17, [x15, #0x80]\n"
-      "str s16, [x15, #0xc0]\n"
-      "add x15, x15, #0x4\n"
+      "zip1 v17.16b, v20.16b, v17.16b\n"
+      "zip1 v16.16b, v19.16b, v16.16b\n"
+      "str s22, [x21, #0x0]\n"
+      "zip1 v16.16b, v17.16b, v16.16b\n"
+      "str s21, [x21, #0x40]\n"
+      "str s18, [x21, #0x80]\n"
+      "str s16, [x21, #0xc0]\n"
+      "add x21, x21, #0x4\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 1 loop: skip
-      "8:"  // Main row loop: odd col skip
       "cmp %x[height], #0x10\n"
       "add %x[out], %x[out], #0x100\n"
       "bge 1b\n"
-      "cbz %x[height], 18f\n"
-      "9:"  // Main loop skip
-      "10:"  // Tail row loop: Head
+      "cbz %x[height], 16f\n"
+      "8:"  // Main loop skip
+      "9:"  // Tail row loop: Head
       "mov x17, %x[in]\n"
+      "add x16, x17, %x[in_stride]\n"
+      "add x15, x16, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x14, x15, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x15, %x[out]\n"
-      "add x14, x17, %x[in_stride]\n"
-      "add x13, x14, %x[in_stride]\n"
-      "add x12, x13, %x[in_stride]\n"
-      "csel x13, x13, %x[pad_row], GE\n"
-      "add %x[in], x12, %x[in_stride]\n"
-      "csel x12, x12, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x14, %x[in_stride]\n"
       "csel x14, x14, %x[pad_row], GT\n"
+      "csel x15, x15, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x16, x16, %x[pad_row], GT\n"
       "cmp x20, #0x10\n"
-      "blt 12f\n"
-      "11:"  // Tail row loop: Column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "blt 11f\n"
+      "10:"  // Tail row loop: Column loop
       "ldr q20, [x17], #0x10\n"
-      "ldr q21, [x14], #0x10\n"
+      "ldr q21, [x16], #0x10\n"
       "sub x20, x20, #0x10\n"
-      "ldr q19, [x13], #0x10\n"
-      "ldr q16, [x12], #0x10\n"
       "cmp x20, #0x10\n"
+      "ldr q19, [x15], #0x10\n"
+      "ldr q16, [x14], #0x10\n"
       "zip1 v18.16b, v20.16b, v19.16b\n"
       "zip1 v17.16b, v21.16b, v16.16b\n"
       "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip2 v16.16b, v21.16b, v16.16b\n"
-      "zip1 v19.16b, v18.16b, v17.16b\n"
+      "zip2 v19.16b, v21.16b, v16.16b\n"
+      "zip1 v16.16b, v18.16b, v17.16b\n"
       "zip2 v18.16b, v18.16b, v17.16b\n"
-      "zip1 v17.16b, v20.16b, v16.16b\n"
-      "zip2 v16.16b, v20.16b, v16.16b\n"
-      "str q19, [x15, #0x0]\n"
-      "str q18, [x15, #0x10]\n"
-      "str q17, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "bge 11b\n"
-      "12:"  // Tail row loop: Column loop skip
-      "cbz x20, 17f\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v17.16b, v20.16b, v19.16b\n"
+      "zip2 v16.16b, v20.16b, v19.16b\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 10b\n"
+      "11:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x15, #0x0]\n"
-      "str q16, [x15, #0x10]\n"
-      "str q16, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: width 4 loop: loop
+      "blt 13f\n"
+      "12:"  // Tail row loop: width 4 loop: loop
       "ldr s19, [x17], #0x4\n"
-      "ldr s18, [x14], #0x4\n"
+      "ldr s18, [x16], #0x4\n"
       "sub x20, x20, #0x4\n"
-      "ldr s17, [x13], #0x4\n"
-      "ldr s16, [x12], #0x4\n"
       "cmp x20, #0x4\n"
+      "ldr s17, [x15], #0x4\n"
+      "ldr s16, [x14], #0x4\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str q16, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 1 loop: loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 1 loop: loop
       "ldr b19, [x17], #0x1\n"
-      "ldr b18, [x14], #0x1\n"
+      "ldr b18, [x16], #0x1\n"
       "sub x20, x20, #0x1\n"
-      "ldr b17, [x13], #0x1\n"
-      "ldr b16, [x12], #0x1\n"
       "cmp x20, #0x1\n"
+      "ldr b17, [x15], #0x1\n"
+      "ldr b16, [x14], #0x1\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str s16, [x15, #0x0]\n"
-      "add x15, x15, #0x4\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 1 loop: skip
-      "17:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x40\n"
-      "bge 10b\n"
-      "18:"  // Done
+      "bge 9b\n"
+      "16:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp
index 586696fcc5..93c95048a8 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp
@@ -41,232 +41,221 @@ void a64_transpose_interleave_16_1x8(uint8_t *out, const uint8_t *in, size_t wid
     __asm__ __volatile__(
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "cmp %x[height], #0x7\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GE\n"
-      "add %x[in], x20, %x[in_stride]\n"
-      "csel x20, x20, %x[pad_row], GT\n"
-      "cmp %x[height], #0x5\n"
+      "cmp %x[height], #0x7\n"
+      "add %x[in], x22, %x[in_stride]\n"
       "csel x22, x22, %x[pad_row], GT\n"
       "csel x23, x23, %x[pad_row], GE\n"
-      "cmp %x[height], #0x3\n"
+      "cmp %x[height], #0x5\n"
+      "mov x21, %x[width]\n"
       "csel x24, x24, %x[pad_row], GT\n"
       "csel x25, x25, %x[pad_row], GE\n"
+      "cmp %x[height], #0x3\n"
+      "csel x26, x26, %x[pad_row], GT\n"
+      "csel x27, x27, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
+      "csel x28, x28, %x[pad_row], GT\n"
+      "cmp x21, #0x20\n"
+      "mov x20, %x[out]\n"
       "sub %x[height], %x[height], #0x8\n"
-      "csel x26, x26, %x[pad_row], GT\n"
-      "cmp x28, #0x20\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q24, [x9], #0x10\n"
-      "ldr q0, [x26], #0x10\n"
-      "sub x28, x28, #0x20\n"
-      "ldr q31, [x25], #0x10\n"
-      "ldr q30, [x24], #0x10\n"
-      "cmp x28, #0x20\n"
-      "ldr q23, [x23], #0x10\n"
-      "ldr q29, [x22], #0x10\n"
-      "ldr q22, [x21], #0x10\n"
-      "ldr q21, [x20], #0x10\n"
-      "ldr q28, [x9], #0x10\n"
-      "ldr q4, [x26], #0x10\n"
-      "ldr q27, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "zip1 v3.16b, v24.16b, v23.16b\n"
-      "zip1 v2.16b, v0.16b, v29.16b\n"
-      "ldr q19, [x23], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "zip1 v26.16b, v31.16b, v22.16b\n"
-      "zip1 v25.16b, v30.16b, v21.16b\n"
-      "ldr q17, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
-      "zip2 v23.16b, v31.16b, v22.16b\n"
-      "zip2 v22.16b, v0.16b, v29.16b\n"
-      "zip2 v21.16b, v30.16b, v21.16b\n"
-      "zip1 v0.16b, v28.16b, v19.16b\n"
-      "zip1 v31.16b, v4.16b, v18.16b\n"
-      "zip1 v30.16b, v27.16b, v17.16b\n"
-      "zip1 v29.16b, v20.16b, v16.16b\n"
-      "zip2 v1.16b, v28.16b, v19.16b\n"
-      "zip2 v28.16b, v27.16b, v17.16b\n"
-      "zip2 v27.16b, v4.16b, v18.16b\n"
-      "zip2 v20.16b, v20.16b, v16.16b\n"
-      "zip1 v19.16b, v3.16b, v26.16b\n"
-      "zip1 v18.16b, v2.16b, v25.16b\n"
-      "zip2 v17.16b, v3.16b, v26.16b\n"
-      "zip2 v16.16b, v2.16b, v25.16b\n"
-      "zip1 v26.16b, v24.16b, v23.16b\n"
-      "zip1 v25.16b, v22.16b, v21.16b\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
+      "ldr q23, [x9], #0x10\n"
+      "ldr q22, [x28], #0x10\n"
+      "sub x21, x21, #0x20\n"
+      "cmp x21, #0x20\n"
+      "ldr q20, [x27], #0x10\n"
+      "ldr q21, [x26], #0x10\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x24], #0x10\n"
+      "zip1 v5.16b, v23.16b, v19.16b\n"
+      "zip1 v4.16b, v22.16b, v18.16b\n"
+      "ldr q17, [x23], #0x10\n"
+      "ldr q16, [x22], #0x10\n"
+      "zip1 v3.16b, v20.16b, v17.16b\n"
+      "zip1 v31.16b, v21.16b, v16.16b\n"
+      "ldr q25, [x9], #0x10\n"
+      "ldr q24, [x28], #0x10\n"
+      "zip2 v2.16b, v23.16b, v19.16b\n"
+      "zip2 v30.16b, v20.16b, v17.16b\n"
+      "ldr q23, [x27], #0x10\n"
+      "ldr q20, [x26], #0x10\n"
+      "zip2 v22.16b, v22.16b, v18.16b\n"
+      "zip2 v21.16b, v21.16b, v16.16b\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x24], #0x10\n"
+      "zip1 v29.16b, v25.16b, v19.16b\n"
+      "zip1 v28.16b, v24.16b, v18.16b\n"
+      "ldr q17, [x23], #0x10\n"
+      "ldr q16, [x22], #0x10\n"
+      "zip1 v27.16b, v23.16b, v17.16b\n"
+      "zip1 v26.16b, v20.16b, v16.16b\n"
+      "zip2 v1.16b, v25.16b, v19.16b\n"
+      "zip2 v25.16b, v23.16b, v17.16b\n"
+      "zip2 v24.16b, v24.16b, v18.16b\n"
+      "zip2 v16.16b, v20.16b, v16.16b\n"
+      "zip1 v0.16b, v5.16b, v3.16b\n"
+      "zip1 v17.16b, v4.16b, v31.16b\n"
+      "zip2 v20.16b, v5.16b, v3.16b\n"
+      "zip2 v19.16b, v4.16b, v31.16b\n"
+      "zip1 v31.16b, v2.16b, v30.16b\n"
+      "zip1 v18.16b, v22.16b, v21.16b\n"
+      "zip2 v30.16b, v2.16b, v30.16b\n"
       "zip2 v23.16b, v22.16b, v21.16b\n"
-      "zip1 v22.16b, v0.16b, v30.16b\n"
-      "zip1 v21.16b, v31.16b, v29.16b\n"
-      "zip2 v0.16b, v0.16b, v30.16b\n"
-      "zip2 v31.16b, v31.16b, v29.16b\n"
-      "zip1 v30.16b, v1.16b, v28.16b\n"
-      "zip1 v29.16b, v27.16b, v20.16b\n"
-      "zip2 v28.16b, v1.16b, v28.16b\n"
-      "zip2 v27.16b, v27.16b, v20.16b\n"
-      "zip1 v20.16b, v19.16b, v18.16b\n"
-      "zip2 v19.16b, v19.16b, v18.16b\n"
-      "zip1 v18.16b, v17.16b, v16.16b\n"
-      "zip2 v17.16b, v17.16b, v16.16b\n"
-      "zip1 v16.16b, v26.16b, v25.16b\n"
-      "zip2 v26.16b, v26.16b, v25.16b\n"
-      "zip1 v25.16b, v24.16b, v23.16b\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
-      "str q20, [x27, #0x0]\n"
-      "str q19, [x27, #0x10]\n"
+      "zip1 v22.16b, v29.16b, v27.16b\n"
+      "zip1 v21.16b, v28.16b, v26.16b\n"
+      "zip2 v29.16b, v29.16b, v27.16b\n"
+      "zip2 v28.16b, v28.16b, v26.16b\n"
+      "zip1 v27.16b, v1.16b, v25.16b\n"
+      "zip1 v26.16b, v24.16b, v16.16b\n"
+      "zip2 v25.16b, v1.16b, v25.16b\n"
+      "zip2 v24.16b, v24.16b, v16.16b\n"
+      "zip1 v16.16b, v0.16b, v17.16b\n"
+      "zip2 v17.16b, v0.16b, v17.16b\n"
+      "str q16, [x20, #0x0]\n"
+      "zip1 v16.16b, v20.16b, v19.16b\n"
+      "zip2 v20.16b, v20.16b, v19.16b\n"
+      "str q17, [x20, #0x10]\n"
+      "zip1 v19.16b, v31.16b, v18.16b\n"
+      "zip2 v18.16b, v31.16b, v18.16b\n"
+      "str q16, [x20, #0x20]\n"
+      "zip1 v17.16b, v30.16b, v23.16b\n"
+      "zip2 v16.16b, v30.16b, v23.16b\n"
+      "str q20, [x20, #0x30]\n"
+      "str q19, [x20, #0x40]\n"
       "zip1 v23.16b, v22.16b, v21.16b\n"
       "zip2 v22.16b, v22.16b, v21.16b\n"
-      "str q18, [x27, #0x20]\n"
-      "zip1 v21.16b, v0.16b, v31.16b\n"
-      "zip2 v20.16b, v0.16b, v31.16b\n"
-      "str q17, [x27, #0x30]\n"
-      "zip1 v19.16b, v30.16b, v29.16b\n"
-      "zip2 v18.16b, v30.16b, v29.16b\n"
-      "str q16, [x27, #0x40]\n"
-      "zip1 v17.16b, v28.16b, v27.16b\n"
-      "zip2 v16.16b, v28.16b, v27.16b\n"
-      "str q26, [x27, #0x50]\n"
-      "str q25, [x27, #0x60]\n"
-      "str q24, [x27, #0x70]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q23, [x27, #0x0]\n"
-      "str q22, [x27, #0x10]\n"
-      "str q21, [x27, #0x20]\n"
-      "str q20, [x27, #0x30]\n"
-      "str q19, [x27, #0x40]\n"
-      "str q18, [x27, #0x50]\n"
-      "str q17, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "str q18, [x20, #0x50]\n"
+      "zip1 v21.16b, v29.16b, v28.16b\n"
+      "zip2 v20.16b, v29.16b, v28.16b\n"
+      "str q17, [x20, #0x60]\n"
+      "zip1 v19.16b, v27.16b, v26.16b\n"
+      "zip2 v18.16b, v27.16b, v26.16b\n"
+      "str q16, [x20, #0x70]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "zip1 v17.16b, v25.16b, v24.16b\n"
+      "zip2 v16.16b, v25.16b, v24.16b\n"
+      "str q23, [x20, #0x0]\n"
+      "str q22, [x20, #0x10]\n"
+      "str q21, [x20, #0x20]\n"
+      "str q20, [x20, #0x30]\n"
+      "str q19, [x20, #0x40]\n"
+      "str q18, [x20, #0x50]\n"
+      "str q17, [x20, #0x60]\n"
+      "str q16, [x20, #0x70]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x28, #0x10\n"
+      "cmp x21, #0x10\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr q24, [x9], #0x10\n"
-      "ldr q25, [x26], #0x10\n"
-      "sub x28, x28, #0x10\n"
-      "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x24], #0x10\n"
-      "cmp x28, #0x10\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "zip1 v20.16b, v24.16b, v18.16b\n"
-      "zip1 v19.16b, v25.16b, v17.16b\n"
-      "zip2 v24.16b, v24.16b, v18.16b\n"
-      "zip2 v25.16b, v25.16b, v17.16b\n"
-      "zip1 v18.16b, v23.16b, v21.16b\n"
-      "zip1 v17.16b, v22.16b, v16.16b\n"
-      "zip2 v23.16b, v23.16b, v21.16b\n"
-      "zip2 v16.16b, v22.16b, v16.16b\n"
-      "zip1 v22.16b, v20.16b, v18.16b\n"
-      "zip1 v21.16b, v19.16b, v17.16b\n"
-      "zip2 v20.16b, v20.16b, v18.16b\n"
-      "zip2 v19.16b, v19.16b, v17.16b\n"
-      "zip1 v18.16b, v24.16b, v23.16b\n"
-      "zip1 v17.16b, v25.16b, v16.16b\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
-      "zip2 v16.16b, v25.16b, v16.16b\n"
-      "zip1 v23.16b, v22.16b, v21.16b\n"
-      "zip2 v22.16b, v22.16b, v21.16b\n"
-      "zip1 v21.16b, v20.16b, v19.16b\n"
-      "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip1 v19.16b, v18.16b, v17.16b\n"
-      "zip2 v18.16b, v18.16b, v17.16b\n"
-      "zip1 v17.16b, v24.16b, v16.16b\n"
+      "ldr q25, [x9], #0x10\n"
+      "ldr q27, [x28], #0x10\n"
+      "sub x21, x21, #0x10\n"
+      "cmp x21, #0x10\n"
+      "ldr q26, [x27], #0x10\n"
+      "ldr q24, [x26], #0x10\n"
+      "ldr q22, [x25], #0x10\n"
+      "ldr q21, [x24], #0x10\n"
+      "zip1 v20.16b, v25.16b, v22.16b\n"
+      "zip1 v23.16b, v27.16b, v21.16b\n"
+      "ldr q17, [x23], #0x10\n"
+      "ldr q16, [x22], #0x10\n"
+      "zip1 v19.16b, v26.16b, v17.16b\n"
+      "zip1 v18.16b, v24.16b, v16.16b\n"
+      "zip2 v25.16b, v25.16b, v22.16b\n"
+      "zip2 v22.16b, v26.16b, v17.16b\n"
+      "zip2 v21.16b, v27.16b, v21.16b\n"
       "zip2 v16.16b, v24.16b, v16.16b\n"
-      "str q23, [x27, #0x0]\n"
-      "str q22, [x27, #0x10]\n"
-      "str q21, [x27, #0x20]\n"
-      "str q20, [x27, #0x30]\n"
-      "str q19, [x27, #0x40]\n"
-      "str q18, [x27, #0x50]\n"
-      "str q17, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 v24.16b, v20.16b, v19.16b\n"
+      "zip1 v17.16b, v23.16b, v18.16b\n"
+      "zip2 v20.16b, v20.16b, v19.16b\n"
+      "zip2 v19.16b, v23.16b, v18.16b\n"
+      "zip1 v23.16b, v25.16b, v22.16b\n"
+      "zip1 v18.16b, v21.16b, v16.16b\n"
+      "zip2 v22.16b, v25.16b, v22.16b\n"
+      "zip2 v21.16b, v21.16b, v16.16b\n"
+      "zip1 v16.16b, v24.16b, v17.16b\n"
+      "zip2 v17.16b, v24.16b, v17.16b\n"
+      "str q16, [x20, #0x0]\n"
+      "zip1 v16.16b, v20.16b, v19.16b\n"
+      "zip2 v20.16b, v20.16b, v19.16b\n"
+      "str q17, [x20, #0x10]\n"
+      "zip1 v19.16b, v23.16b, v18.16b\n"
+      "zip2 v18.16b, v23.16b, v18.16b\n"
+      "str q16, [x20, #0x20]\n"
+      "zip1 v17.16b, v22.16b, v21.16b\n"
+      "zip2 v16.16b, v22.16b, v21.16b\n"
+      "str q20, [x20, #0x30]\n"
+      "str q19, [x20, #0x40]\n"
+      "str q18, [x20, #0x50]\n"
+      "str q17, [x20, #0x60]\n"
+      "str q16, [x20, #0x70]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x28, 10f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
+      "cmp x21, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr s23, [x9], #0x4\n"
-      "ldr s22, [x26], #0x4\n"
-      "sub x28, x28, #0x4\n"
-      "ldr s21, [x25], #0x4\n"
-      "ldr s20, [x24], #0x4\n"
-      "cmp x28, #0x4\n"
-      "ldr s18, [x23], #0x4\n"
-      "ldr s19, [x22], #0x4\n"
-      "ldr s17, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
-      "zip1 v18.16b, v23.16b, v18.16b\n"
-      "zip1 v19.16b, v22.16b, v19.16b\n"
+      "ldr s18, [x9], #0x4\n"
+      "ldr s19, [x28], #0x4\n"
+      "sub x21, x21, #0x4\n"
+      "cmp x21, #0x4\n"
+      "ldr s21, [x27], #0x4\n"
+      "ldr s20, [x26], #0x4\n"
+      "ldr s17, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "zip1 v18.16b, v18.16b, v17.16b\n"
+      "zip1 v19.16b, v19.16b, v16.16b\n"
+      "ldr s17, [x23], #0x4\n"
+      "ldr s16, [x22], #0x4\n"
       "zip1 v17.16b, v21.16b, v17.16b\n"
       "zip1 v16.16b, v20.16b, v16.16b\n"
       "zip1 v18.16b, v18.16b, v17.16b\n"
       "zip1 v16.16b, v19.16b, v16.16b\n"
       "zip1 v17.16b, v18.16b, v16.16b\n"
       "zip2 v16.16b, v18.16b, v16.16b\n"
-      "str q17, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
+      "str q17, [x20, #0x0]\n"
+      "str q16, [x20, #0x10]\n"
+      "add x20, x20, #0x20\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x21, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr b23, [x9], #0x1\n"
-      "ldr b22, [x26], #0x1\n"
-      "sub x28, x28, #0x1\n"
-      "ldr b21, [x25], #0x1\n"
-      "ldr b20, [x24], #0x1\n"
-      "cmp x28, #0x1\n"
-      "ldr b19, [x23], #0x1\n"
-      "ldr b18, [x22], #0x1\n"
-      "ldr b17, [x21], #0x1\n"
-      "ldr b16, [x20], #0x1\n"
-      "zip1 v19.16b, v23.16b, v19.16b\n"
-      "zip1 v18.16b, v22.16b, v18.16b\n"
+      "ldr b19, [x9], #0x1\n"
+      "ldr b18, [x28], #0x1\n"
+      "sub x21, x21, #0x1\n"
+      "cmp x21, #0x1\n"
+      "ldr b21, [x27], #0x1\n"
+      "ldr b20, [x26], #0x1\n"
+      "ldr b17, [x25], #0x1\n"
+      "ldr b16, [x24], #0x1\n"
+      "zip1 v19.16b, v19.16b, v17.16b\n"
+      "zip1 v18.16b, v18.16b, v16.16b\n"
+      "ldr b17, [x23], #0x1\n"
+      "ldr b16, [x22], #0x1\n"
       "zip1 v17.16b, v21.16b, v17.16b\n"
       "zip1 v16.16b, v20.16b, v16.16b\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str d16, [x27, #0x0]\n"
-      "add x27, x27, #0x8\n"
+      "str d16, [x20, #0x0]\n"
+      "add x20, x20, #0x8\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x80\n"
       "bge 1b\n"
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x2.hpp
index 8186b1f475..b1efe81b35 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x2.hpp
@@ -40,215 +40,189 @@ void a64_transpose_interleave_16_2x2(uint16_t *out, const uint16_t *in, size_t w
 
     __asm__ __volatile__(
       "cmp %x[height], #0x8\n"
-      "blt 9f\n"
+      "blt 8f\n"
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "cmp x28, #0x10\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x10\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
-      "ldr q22, [x9], #0x10\n"
-      "ldr q21, [x26], #0x10\n"
-      "sub x28, x28, #0x10\n"
-      "ldr q20, [x25], #0x10\n"
-      "ldr q19, [x24], #0x10\n"
-      "cmp x28, #0x10\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q23, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "zip1 v0.8h, v22.8h, v21.8h\n"
-      "zip2 v31.8h, v22.8h, v21.8h\n"
-      "ldr q22, [x9], #0x10\n"
-      "ldr q21, [x26], #0x10\n"
-      "zip1 v30.8h, v20.8h, v19.8h\n"
-      "zip2 v29.8h, v20.8h, v19.8h\n"
-      "ldr q20, [x25], #0x10\n"
-      "ldr q19, [x24], #0x10\n"
-      "zip1 v28.8h, v18.8h, v17.8h\n"
-      "zip2 v27.8h, v18.8h, v17.8h\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
-      "zip1 v26.8h, v23.8h, v16.8h\n"
-      "zip2 v25.8h, v23.8h, v16.8h\n"
-      "ldr q24, [x21], #0x10\n"
+      "ldr q17, [x9], #0x10\n"
+      "ldr q16, [x28], #0x10\n"
+      "sub x24, x24, #0x10\n"
+      "cmp x24, #0x10\n"
+      "ldr q19, [x27], #0x10\n"
+      "ldr q18, [x26], #0x10\n"
+      "zip1 v1.8h, v17.8h, v16.8h\n"
+      "zip2 v0.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip1 v31.8h, v19.8h, v18.8h\n"
+      "zip2 v30.8h, v19.8h, v18.8h\n"
+      "ldr q29, [x22], #0x10\n"
+      "ldr q18, [x20], #0x10\n"
+      "zip1 v28.8h, v17.8h, v16.8h\n"
+      "zip2 v27.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x9], #0x10\n"
+      "ldr q16, [x28], #0x10\n"
+      "zip1 v26.8h, v17.8h, v16.8h\n"
+      "zip2 v25.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v24.8h, v17.8h, v16.8h\n"
+      "zip2 v23.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip1 v22.8h, v17.8h, v16.8h\n"
+      "zip2 v21.8h, v17.8h, v16.8h\n"
+      "ldr q20, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip1 v23.8h, v22.8h, v21.8h\n"
-      "zip2 v22.8h, v22.8h, v21.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "str q0, [x27, #0x0]\n"
-      "zip1 v19.8h, v18.8h, v17.8h\n"
-      "zip2 v18.8h, v18.8h, v17.8h\n"
-      "str q31, [x27, #0x10]\n"
-      "zip1 v17.8h, v24.8h, v16.8h\n"
-      "zip2 v16.8h, v24.8h, v16.8h\n"
-      "str q23, [x27, #0x20]\n"
-      "str q22, [x27, #0x30]\n"
-      "str q30, [x27, #0x40]\n"
-      "str q29, [x27, #0x50]\n"
-      "str q21, [x27, #0x60]\n"
-      "str q20, [x27, #0x70]\n"
-      "str q28, [x27, #0x80]\n"
-      "str q27, [x27, #0x90]\n"
-      "str q19, [x27, #0xa0]\n"
-      "str q18, [x27, #0xb0]\n"
-      "str q26, [x27, #0xc0]\n"
-      "str q25, [x27, #0xd0]\n"
-      "str q17, [x27, #0xe0]\n"
-      "str q16, [x27, #0xf0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 v19.8h, v29.8h, v18.8h\n"
+      "zip2 v18.8h, v29.8h, v18.8h\n"
+      "zip1 v17.8h, v20.8h, v16.8h\n"
+      "zip2 v16.8h, v20.8h, v16.8h\n"
+      "str q1, [x21, #0x0]\n"
+      "str q0, [x21, #0x10]\n"
+      "str q26, [x21, #0x20]\n"
+      "str q25, [x21, #0x30]\n"
+      "str q31, [x21, #0x40]\n"
+      "str q30, [x21, #0x50]\n"
+      "str q24, [x21, #0x60]\n"
+      "str q23, [x21, #0x70]\n"
+      "str q28, [x21, #0x80]\n"
+      "str q27, [x21, #0x90]\n"
+      "str q22, [x21, #0xa0]\n"
+      "str q21, [x21, #0xb0]\n"
+      "str q19, [x21, #0xc0]\n"
+      "str q18, [x21, #0xd0]\n"
+      "str q17, [x21, #0xe0]\n"
+      "str q16, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x28, 8f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "str q16, [x27, #0x80]\n"
-      "str q16, [x27, #0x90]\n"
-      "str q16, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
-      "str q16, [x27, #0xc0]\n"
-      "str q16, [x27, #0xd0]\n"
-      "str q16, [x27, #0xe0]\n"
-      "str q16, [x27, #0xf0]\n"
+      "cmp x24, #0x4\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 4 loop: loop
-      "ldr d23, [x9], #0x8\n"
-      "ldr d18, [x26], #0x8\n"
-      "sub x28, x28, #0x4\n"
-      "ldr d22, [x25], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
-      "cmp x28, #0x4\n"
-      "ldr d21, [x23], #0x8\n"
+      "ldr d19, [x9], #0x8\n"
+      "ldr d18, [x28], #0x8\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
+      "zip1 v20.8h, v19.8h, v18.8h\n"
+      "zip1 v19.8h, v17.8h, v16.8h\n"
+      "ldr d17, [x25], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
+      "zip1 v18.8h, v17.8h, v16.8h\n"
       "ldr d17, [x22], #0x8\n"
-      "ldr d20, [x21], #0x8\n"
-      "ldr d19, [x20], #0x8\n"
-      "zip1 v18.8h, v23.8h, v18.8h\n"
-      "zip1 v16.8h, v22.8h, v16.8h\n"
-      "zip1 v17.8h, v21.8h, v17.8h\n"
-      "str q18, [x27, #0x0]\n"
-      "str q16, [x27, #0x40]\n"
-      "zip1 v16.8h, v20.8h, v19.8h\n"
-      "str q17, [x27, #0x80]\n"
-      "str q16, [x27, #0xc0]\n"
-      "add x27, x27, #0x10\n"
+      "ldr d16, [x20], #0x8\n"
+      "str q20, [x21, #0x0]\n"
+      "zip1 v16.8h, v17.8h, v16.8h\n"
+      "str q19, [x21, #0x40]\n"
+      "str q18, [x21, #0x80]\n"
+      "str q16, [x21, #0xc0]\n"
+      "add x21, x21, #0x10\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 1 loop: loop
-      "ldr h23, [x9], #0x2\n"
-      "ldr h18, [x26], #0x2\n"
-      "sub x28, x28, #0x1\n"
-      "ldr h22, [x25], #0x2\n"
-      "ldr h16, [x24], #0x2\n"
-      "cmp x28, #0x1\n"
-      "ldr h21, [x23], #0x2\n"
+      "ldr h19, [x9], #0x2\n"
+      "ldr h18, [x28], #0x2\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr h17, [x27], #0x2\n"
+      "ldr h16, [x26], #0x2\n"
+      "zip1 v20.8h, v19.8h, v18.8h\n"
+      "zip1 v19.8h, v17.8h, v16.8h\n"
+      "ldr h17, [x25], #0x2\n"
+      "ldr h16, [x23], #0x2\n"
+      "zip1 v18.8h, v17.8h, v16.8h\n"
       "ldr h17, [x22], #0x2\n"
-      "ldr h20, [x21], #0x2\n"
-      "ldr h19, [x20], #0x2\n"
-      "zip1 v18.8h, v23.8h, v18.8h\n"
-      "zip1 v16.8h, v22.8h, v16.8h\n"
-      "zip1 v17.8h, v21.8h, v17.8h\n"
-      "str s18, [x27, #0x0]\n"
-      "str s16, [x27, #0x40]\n"
-      "zip1 v16.8h, v20.8h, v19.8h\n"
-      "str s17, [x27, #0x80]\n"
-      "str s16, [x27, #0xc0]\n"
-      "add x27, x27, #0x4\n"
+      "ldr h16, [x20], #0x2\n"
+      "str s20, [x21, #0x0]\n"
+      "zip1 v16.8h, v17.8h, v16.8h\n"
+      "str s19, [x21, #0x40]\n"
+      "str s18, [x21, #0x80]\n"
+      "str s16, [x21, #0xc0]\n"
+      "add x21, x21, #0x4\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 1 loop: skip
-      "8:"  // Main row loop: odd col skip
       "cmp %x[height], #0x8\n"
       "add %x[out], %x[out], #0x100\n"
       "bge 1b\n"
-      "cbz %x[height], 18f\n"
-      "9:"  // Main loop skip
-      "10:"  // Tail row loop: Head
+      "cbz %x[height], 16f\n"
+      "8:"  // Main loop skip
+      "9:"  // Tail row loop: Head
       "mov x9, %x[in]\n"
       "mov x20, %x[width]\n"
+      "add x28, x9, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x2\n"
-      "add x26, x9, %x[in_stride]\n"
-      "add %x[in], x26, %x[in_stride]\n"
-      "csel x26, x26, %x[pad_row], GT\n"
+      "add %x[in], x28, %x[in_stride]\n"
+      "csel x28, x28, %x[pad_row], GT\n"
       "cmp x20, #0x10\n"
-      "blt 12f\n"
-      "11:"  // Tail row loop: Column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x2\n"
+      "blt 11f\n"
+      "10:"  // Tail row loop: Column loop
       "ldr q18, [x9], #0x10\n"
-      "ldr q17, [x26], #0x10\n"
+      "ldr q17, [x28], #0x10\n"
       "sub x20, x20, #0x10\n"
-      "ldr q20, [x9], #0x10\n"
       "cmp x20, #0x10\n"
-      "ldr q16, [x26], #0x10\n"
+      "ldr q20, [x9], #0x10\n"
+      "ldr q16, [x28], #0x10\n"
       "zip1 v19.8h, v18.8h, v17.8h\n"
       "zip2 v18.8h, v18.8h, v17.8h\n"
       "zip1 v17.8h, v20.8h, v16.8h\n"
       "zip2 v16.8h, v20.8h, v16.8h\n"
-      "str q19, [x27, #0x0]\n"
-      "str q18, [x27, #0x10]\n"
-      "str q17, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 11b\n"
-      "12:"  // Tail row loop: Column loop skip
-      "cbz x20, 17f\n"
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 10b\n"
+      "11:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: width 4 loop: loop
+      "blt 13f\n"
+      "12:"  // Tail row loop: width 4 loop: loop
       "ldr d17, [x9], #0x8\n"
-      "ldr d16, [x26], #0x8\n"
+      "ldr d16, [x28], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
       "zip1 v16.8h, v17.8h, v16.8h\n"
-      "str q16, [x27, #0x0]\n"
-      "add x27, x27, #0x10\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 1 loop: loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 1 loop: loop
       "ldr h17, [x9], #0x2\n"
-      "ldr h16, [x26], #0x2\n"
+      "ldr h16, [x28], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
       "zip1 v16.8h, v17.8h, v16.8h\n"
-      "str s16, [x27, #0x0]\n"
-      "add x27, x27, #0x4\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 1 loop: skip
-      "17:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x40\n"
-      "bge 10b\n"
-      "18:"  // Done
+      "bge 9b\n"
+      "16:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "v0", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v0", "v1", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp
index 1ed8708f4f..02ae1ade30 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp
@@ -40,481 +40,451 @@ void a64_transpose_interleave_16_2x4(uint16_t *out, const uint16_t *in, size_t w
 
     __asm__ __volatile__(
       "cmp %x[height], #0x8\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "cmp x28, #0x20\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x20\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q0, [x9], #0x10\n"
-      "ldr q18, [x26], #0x10\n"
-      "sub x28, x28, #0x20\n"
-      "ldr q22, [x25], #0x10\n"
-      "ldr q27, [x24], #0x10\n"
-      "cmp x28, #0x20\n"
-      "ldr q29, [x23], #0x10\n"
-      "ldr q12, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
-      "ldr q5, [x20], #0x10\n"
-      "ldr q7, [x9], #0x10\n"
-      "ldr q16, [x26], #0x10\n"
-      "zip1 v26.8h, v0.8h, v22.8h\n"
-      "zip1 v3.8h, v18.8h, v27.8h\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q28, [x24], #0x10\n"
-      "zip2 v25.8h, v0.8h, v22.8h\n"
-      "zip2 v10.8h, v18.8h, v27.8h\n"
-      "ldr q1, [x23], #0x10\n"
-      "ldr q20, [x22], #0x10\n"
-      "zip1 v14.8h, v29.8h, v17.8h\n"
-      "zip1 v19.8h, v12.8h, v5.8h\n"
-      "ldr q0, [x21], #0x10\n"
-      "ldr q13, [x20], #0x10\n"
-      "zip2 v22.8h, v29.8h, v17.8h\n"
-      "zip2 v15.8h, v12.8h, v5.8h\n"
+      "ldr q23, [x9], #0x10\n"
+      "ldr q29, [x28], #0x10\n"
+      "sub x24, x24, #0x20\n"
+      "cmp x24, #0x20\n"
+      "ldr q13, [x27], #0x10\n"
+      "ldr q12, [x26], #0x10\n"
+      "zip1 v20.8h, v23.8h, v13.8h\n"
+      "zip1 v28.8h, v29.8h, v12.8h\n"
+      "ldr q18, [x25], #0x10\n"
+      "ldr q9, [x23], #0x10\n"
+      "zip2 v22.8h, v23.8h, v13.8h\n"
+      "zip2 v1.8h, v29.8h, v12.8h\n"
+      "ldr q27, [x22], #0x10\n"
+      "ldr q3, [x20], #0x10\n"
+      "zip1 v4.8h, v18.8h, v27.8h\n"
+      "zip1 v26.8h, v9.8h, v3.8h\n"
       "ldr q17, [x9], #0x10\n"
-      "ldr q23, [x26], #0x10\n"
-      "zip1 v18.8h, v7.8h, v21.8h\n"
-      "zip1 v5.8h, v16.8h, v28.8h\n"
-      "ldr q31, [x25], #0x10\n"
-      "ldr q12, [x24], #0x10\n"
-      "zip2 v24.8h, v7.8h, v21.8h\n"
-      "zip2 v28.8h, v16.8h, v28.8h\n"
-      "ldr q4, [x23], #0x10\n"
-      "ldr q6, [x22], #0x10\n"
-      "zip1 v30.8h, v1.8h, v0.8h\n"
-      "zip1 v21.8h, v20.8h, v13.8h\n"
-      "ldr q7, [x21], #0x10\n"
-      "ldr q9, [x20], #0x10\n"
-      "zip2 v27.8h, v1.8h, v0.8h\n"
-      "zip2 v0.8h, v20.8h, v13.8h\n"
-      "ldr q2, [x9], #0x10\n"
-      "ldr q8, [x26], #0x10\n"
-      "zip1 v29.8h, v17.8h, v31.8h\n"
-      "zip1 v16.8h, v23.8h, v12.8h\n"
-      "ldr q13, [x25], #0x10\n"
-      "ldr q1, [x24], #0x10\n"
-      "zip2 v20.8h, v17.8h, v31.8h\n"
-      "zip2 v23.8h, v23.8h, v12.8h\n"
-      "ldr q17, [x23], #0x10\n"
-      "ldr q31, [x22], #0x10\n"
-      "zip1 v12.8h, v4.8h, v7.8h\n"
-      "zip1 v11.8h, v6.8h, v9.8h\n"
-      "zip2 v7.8h, v4.8h, v7.8h\n"
-      "ldr q4, [x21], #0x10\n"
-      "zip2 v6.8h, v6.8h, v9.8h\n"
-      "zip1 v9.8h, v2.8h, v13.8h\n"
-      "zip2 v13.8h, v2.8h, v13.8h\n"
-      "zip1 v2.8h, v8.8h, v1.8h\n"
-      "zip2 v1.8h, v8.8h, v1.8h\n"
-      "zip1 v8.8h, v17.8h, v4.8h\n"
-      "zip2 v4.8h, v17.8h, v4.8h\n"
-      "zip1 v17.8h, v26.8h, v3.8h\n"
-      "zip2 v26.8h, v26.8h, v3.8h\n"
+      "ldr q2, [x28], #0x10\n"
+      "zip2 v15.8h, v18.8h, v27.8h\n"
+      "zip2 v12.8h, v9.8h, v3.8h\n"
+      "ldr q23, [x27], #0x10\n"
+      "ldr q14, [x26], #0x10\n"
+      "zip1 v19.8h, v17.8h, v23.8h\n"
+      "zip1 v21.8h, v2.8h, v14.8h\n"
+      "ldr q6, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v27.8h, v17.8h, v23.8h\n"
+      "zip2 v17.8h, v2.8h, v14.8h\n"
+      "ldr q0, [x22], #0x10\n"
       "ldr q3, [x20], #0x10\n"
-      "str q17, [x27, #0x0]\n"
-      "zip1 v17.8h, v25.8h, v10.8h\n"
-      "zip2 v25.8h, v25.8h, v10.8h\n"
-      "zip1 v10.8h, v18.8h, v5.8h\n"
-      "zip2 v5.8h, v18.8h, v5.8h\n"
-      "zip1 v18.8h, v31.8h, v3.8h\n"
-      "zip2 v3.8h, v31.8h, v3.8h\n"
-      "str q26, [x27, #0x10]\n"
-      "zip1 v31.8h, v24.8h, v28.8h\n"
-      "zip2 v26.8h, v24.8h, v28.8h\n"
-      "str q17, [x27, #0x20]\n"
-      "zip1 v28.8h, v14.8h, v19.8h\n"
-      "zip2 v19.8h, v14.8h, v19.8h\n"
-      "str q25, [x27, #0x30]\n"
-      "zip1 v24.8h, v22.8h, v15.8h\n"
-      "zip2 v15.8h, v22.8h, v15.8h\n"
-      "str q10, [x27, #0x40]\n"
-      "zip1 v25.8h, v30.8h, v21.8h\n"
-      "zip2 v14.8h, v30.8h, v21.8h\n"
-      "str q5, [x27, #0x50]\n"
-      "zip1 v5.8h, v27.8h, v0.8h\n"
-      "zip2 v0.8h, v27.8h, v0.8h\n"
-      "str q31, [x27, #0x60]\n"
-      "str q26, [x27, #0x70]\n"
-      "zip1 v31.8h, v29.8h, v16.8h\n"
-      "zip2 v22.8h, v29.8h, v16.8h\n"
-      "str q28, [x27, #0x80]\n"
-      "zip1 v10.8h, v20.8h, v23.8h\n"
-      "zip2 v28.8h, v20.8h, v23.8h\n"
-      "str q19, [x27, #0x90]\n"
-      "zip1 v27.8h, v9.8h, v2.8h\n"
-      "zip2 v20.8h, v9.8h, v2.8h\n"
-      "str q24, [x27, #0xa0]\n"
-      "zip1 v2.8h, v13.8h, v1.8h\n"
-      "zip2 v24.8h, v13.8h, v1.8h\n"
-      "str q15, [x27, #0xb0]\n"
-      "zip1 v23.8h, v12.8h, v11.8h\n"
-      "zip2 v26.8h, v12.8h, v11.8h\n"
-      "str q25, [x27, #0xc0]\n"
-      "zip1 v21.8h, v7.8h, v6.8h\n"
-      "zip2 v29.8h, v7.8h, v6.8h\n"
-      "str q14, [x27, #0xd0]\n"
-      "zip1 v19.8h, v8.8h, v18.8h\n"
-      "zip2 v18.8h, v8.8h, v18.8h\n"
-      "str q5, [x27, #0xe0]\n"
-      "zip1 v17.8h, v4.8h, v3.8h\n"
-      "zip2 v16.8h, v4.8h, v3.8h\n"
-      "str q0, [x27, #0xf0]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q31, [x27, #0x0]\n"
-      "str q22, [x27, #0x10]\n"
-      "str q10, [x27, #0x20]\n"
-      "str q28, [x27, #0x30]\n"
-      "str q27, [x27, #0x40]\n"
-      "str q20, [x27, #0x50]\n"
-      "str q2, [x27, #0x60]\n"
-      "str q24, [x27, #0x70]\n"
-      "str q23, [x27, #0x80]\n"
-      "str q26, [x27, #0x90]\n"
-      "str q21, [x27, #0xa0]\n"
-      "str q29, [x27, #0xb0]\n"
-      "str q19, [x27, #0xc0]\n"
-      "str q18, [x27, #0xd0]\n"
-      "str q17, [x27, #0xe0]\n"
-      "str q16, [x27, #0xf0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 v16.8h, v6.8h, v0.8h\n"
+      "zip1 v30.8h, v18.8h, v3.8h\n"
+      "ldr q2, [x9], #0x10\n"
+      "ldr q13, [x28], #0x10\n"
+      "zip2 v31.8h, v6.8h, v0.8h\n"
+      "zip2 v8.8h, v18.8h, v3.8h\n"
+      "ldr q14, [x27], #0x10\n"
+      "ldr q3, [x26], #0x10\n"
+      "zip1 v11.8h, v2.8h, v14.8h\n"
+      "zip1 v29.8h, v13.8h, v3.8h\n"
+      "ldr q25, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v23.8h, v2.8h, v14.8h\n"
+      "zip2 v10.8h, v13.8h, v3.8h\n"
+      "ldr q7, [x22], #0x10\n"
+      "ldr q6, [x20], #0x10\n"
+      "zip1 v14.8h, v25.8h, v7.8h\n"
+      "zip1 v13.8h, v18.8h, v6.8h\n"
+      "ldr q2, [x9], #0x10\n"
+      "ldr q5, [x28], #0x10\n"
+      "zip2 v9.8h, v25.8h, v7.8h\n"
+      "zip2 v7.8h, v18.8h, v6.8h\n"
+      "ldr q6, [x27], #0x10\n"
+      "ldr q24, [x26], #0x10\n"
+      "zip1 v25.8h, v2.8h, v6.8h\n"
+      "zip1 v3.8h, v5.8h, v24.8h\n"
+      "ldr q0, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v2.8h, v2.8h, v6.8h\n"
+      "zip2 v24.8h, v5.8h, v24.8h\n"
+      "ldr q5, [x22], #0x10\n"
+      "zip1 v6.8h, v0.8h, v5.8h\n"
+      "zip2 v5.8h, v0.8h, v5.8h\n"
+      "zip1 v0.8h, v20.8h, v28.8h\n"
+      "zip2 v28.8h, v20.8h, v28.8h\n"
+      "ldr q20, [x20], #0x10\n"
+      "str q0, [x21, #0x0]\n"
+      "zip1 v0.8h, v18.8h, v20.8h\n"
+      "zip2 v20.8h, v18.8h, v20.8h\n"
+      "str q28, [x21, #0x10]\n"
+      "zip1 v18.8h, v22.8h, v1.8h\n"
+      "zip2 v28.8h, v22.8h, v1.8h\n"
+      "str q18, [x21, #0x20]\n"
+      "zip1 v22.8h, v19.8h, v21.8h\n"
+      "zip2 v19.8h, v19.8h, v21.8h\n"
+      "str q28, [x21, #0x30]\n"
+      "zip1 v18.8h, v27.8h, v17.8h\n"
+      "zip2 v17.8h, v27.8h, v17.8h\n"
+      "str q22, [x21, #0x40]\n"
+      "zip1 v27.8h, v4.8h, v26.8h\n"
+      "zip2 v26.8h, v4.8h, v26.8h\n"
+      "str q19, [x21, #0x50]\n"
+      "zip1 v22.8h, v15.8h, v12.8h\n"
+      "zip2 v21.8h, v15.8h, v12.8h\n"
+      "str q18, [x21, #0x60]\n"
+      "zip1 v19.8h, v16.8h, v30.8h\n"
+      "zip2 v18.8h, v16.8h, v30.8h\n"
+      "str q17, [x21, #0x70]\n"
+      "zip1 v17.8h, v31.8h, v8.8h\n"
+      "zip2 v16.8h, v31.8h, v8.8h\n"
+      "str q27, [x21, #0x80]\n"
+      "str q26, [x21, #0x90]\n"
+      "zip1 v31.8h, v11.8h, v29.8h\n"
+      "zip2 v30.8h, v11.8h, v29.8h\n"
+      "str q22, [x21, #0xa0]\n"
+      "zip1 v29.8h, v23.8h, v10.8h\n"
+      "zip2 v28.8h, v23.8h, v10.8h\n"
+      "str q21, [x21, #0xb0]\n"
+      "zip1 v27.8h, v25.8h, v3.8h\n"
+      "zip2 v26.8h, v25.8h, v3.8h\n"
+      "str q19, [x21, #0xc0]\n"
+      "zip1 v25.8h, v2.8h, v24.8h\n"
+      "zip2 v24.8h, v2.8h, v24.8h\n"
+      "str q18, [x21, #0xd0]\n"
+      "zip1 v23.8h, v14.8h, v13.8h\n"
+      "zip2 v22.8h, v14.8h, v13.8h\n"
+      "str q17, [x21, #0xe0]\n"
+      "zip1 v21.8h, v9.8h, v7.8h\n"
+      "zip2 v19.8h, v9.8h, v7.8h\n"
+      "str q16, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip1 v2.8h, v6.8h, v0.8h\n"
+      "zip2 v18.8h, v6.8h, v0.8h\n"
+      "zip1 v17.8h, v5.8h, v20.8h\n"
+      "zip2 v16.8h, v5.8h, v20.8h\n"
+      "str q31, [x21, #0x0]\n"
+      "str q30, [x21, #0x10]\n"
+      "str q29, [x21, #0x20]\n"
+      "str q28, [x21, #0x30]\n"
+      "str q27, [x21, #0x40]\n"
+      "str q26, [x21, #0x50]\n"
+      "str q25, [x21, #0x60]\n"
+      "str q24, [x21, #0x70]\n"
+      "str q23, [x21, #0x80]\n"
+      "str q22, [x21, #0x90]\n"
+      "str q21, [x21, #0xa0]\n"
+      "str q19, [x21, #0xb0]\n"
+      "str q2, [x21, #0xc0]\n"
+      "str q18, [x21, #0xd0]\n"
+      "str q17, [x21, #0xe0]\n"
+      "str q16, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x28, #0x10\n"
+      "cmp x24, #0x10\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr q19, [x9], #0x10\n"
-      "ldr q18, [x26], #0x10\n"
-      "sub x28, x28, #0x10\n"
-      "ldr q17, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "cmp x28, #0x10\n"
-      "ldr q31, [x23], #0x10\n"
-      "ldr q30, [x22], #0x10\n"
-      "ldr q23, [x21], #0x10\n"
-      "ldr q22, [x20], #0x10\n"
-      "ldr q29, [x9], #0x10\n"
-      "ldr q28, [x26], #0x10\n"
-      "zip1 v27.8h, v19.8h, v17.8h\n"
-      "zip1 v26.8h, v18.8h, v16.8h\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "zip2 v25.8h, v19.8h, v17.8h\n"
-      "zip2 v24.8h, v18.8h, v16.8h\n"
-      "ldr q19, [x23], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "zip1 v3.8h, v31.8h, v23.8h\n"
-      "zip1 v2.8h, v30.8h, v22.8h\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "sub x24, x24, #0x10\n"
+      "cmp x24, #0x10\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v3.8h, v21.8h, v17.8h\n"
+      "zip1 v2.8h, v20.8h, v16.8h\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v1.8h, v21.8h, v17.8h\n"
+      "zip2 v24.8h, v20.8h, v16.8h\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip2 v1.8h, v31.8h, v23.8h\n"
-      "zip2 v0.8h, v30.8h, v22.8h\n"
-      "zip1 v23.8h, v29.8h, v21.8h\n"
-      "zip1 v22.8h, v28.8h, v20.8h\n"
-      "zip2 v21.8h, v29.8h, v21.8h\n"
-      "zip2 v20.8h, v28.8h, v20.8h\n"
-      "zip1 v31.8h, v19.8h, v17.8h\n"
-      "zip1 v30.8h, v18.8h, v16.8h\n"
-      "zip2 v29.8h, v19.8h, v17.8h\n"
-      "zip2 v28.8h, v18.8h, v16.8h\n"
-      "zip1 v19.8h, v27.8h, v26.8h\n"
-      "zip2 v18.8h, v27.8h, v26.8h\n"
-      "zip1 v17.8h, v25.8h, v24.8h\n"
-      "zip2 v16.8h, v25.8h, v24.8h\n"
-      "zip1 v27.8h, v23.8h, v22.8h\n"
-      "zip2 v26.8h, v23.8h, v22.8h\n"
-      "zip1 v25.8h, v21.8h, v20.8h\n"
+      "zip1 v0.8h, v19.8h, v17.8h\n"
+      "zip1 v31.8h, v18.8h, v16.8h\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v30.8h, v19.8h, v17.8h\n"
+      "zip2 v29.8h, v18.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v23.8h, v21.8h, v17.8h\n"
+      "zip1 v22.8h, v20.8h, v16.8h\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v21.8h, v21.8h, v17.8h\n"
+      "zip2 v20.8h, v20.8h, v16.8h\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v28.8h, v19.8h, v17.8h\n"
+      "zip1 v27.8h, v18.8h, v16.8h\n"
+      "zip2 v26.8h, v19.8h, v17.8h\n"
+      "zip2 v25.8h, v18.8h, v16.8h\n"
+      "zip1 v16.8h, v3.8h, v2.8h\n"
+      "zip2 v17.8h, v3.8h, v2.8h\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v16.8h, v1.8h, v24.8h\n"
+      "zip2 v19.8h, v1.8h, v24.8h\n"
+      "str q17, [x21, #0x10]\n"
+      "zip1 v18.8h, v23.8h, v22.8h\n"
+      "zip2 v17.8h, v23.8h, v22.8h\n"
+      "str q16, [x21, #0x20]\n"
+      "zip1 v16.8h, v21.8h, v20.8h\n"
       "zip2 v24.8h, v21.8h, v20.8h\n"
-      "str q19, [x27, #0x0]\n"
-      "zip1 v23.8h, v3.8h, v2.8h\n"
-      "zip2 v22.8h, v3.8h, v2.8h\n"
-      "str q18, [x27, #0x10]\n"
-      "zip1 v21.8h, v1.8h, v0.8h\n"
-      "zip2 v20.8h, v1.8h, v0.8h\n"
-      "str q17, [x27, #0x20]\n"
-      "zip1 v19.8h, v31.8h, v30.8h\n"
-      "zip2 v18.8h, v31.8h, v30.8h\n"
-      "str q16, [x27, #0x30]\n"
-      "zip1 v17.8h, v29.8h, v28.8h\n"
-      "zip2 v16.8h, v29.8h, v28.8h\n"
-      "str q27, [x27, #0x40]\n"
-      "str q26, [x27, #0x50]\n"
-      "str q25, [x27, #0x60]\n"
-      "str q24, [x27, #0x70]\n"
-      "str q23, [x27, #0x80]\n"
-      "str q22, [x27, #0x90]\n"
-      "str q21, [x27, #0xa0]\n"
-      "str q20, [x27, #0xb0]\n"
-      "str q19, [x27, #0xc0]\n"
-      "str q18, [x27, #0xd0]\n"
-      "str q17, [x27, #0xe0]\n"
-      "str q16, [x27, #0xf0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "str q19, [x21, #0x30]\n"
+      "zip1 v23.8h, v0.8h, v31.8h\n"
+      "zip2 v22.8h, v0.8h, v31.8h\n"
+      "str q18, [x21, #0x40]\n"
+      "zip1 v21.8h, v30.8h, v29.8h\n"
+      "zip2 v20.8h, v30.8h, v29.8h\n"
+      "str q17, [x21, #0x50]\n"
+      "zip1 v19.8h, v28.8h, v27.8h\n"
+      "zip2 v18.8h, v28.8h, v27.8h\n"
+      "str q16, [x21, #0x60]\n"
+      "zip1 v17.8h, v26.8h, v25.8h\n"
+      "zip2 v16.8h, v26.8h, v25.8h\n"
+      "str q24, [x21, #0x70]\n"
+      "str q23, [x21, #0x80]\n"
+      "str q22, [x21, #0x90]\n"
+      "str q21, [x21, #0xa0]\n"
+      "str q20, [x21, #0xb0]\n"
+      "str q19, [x21, #0xc0]\n"
+      "str q18, [x21, #0xd0]\n"
+      "str q17, [x21, #0xe0]\n"
+      "str q16, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x28, 10f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "str q16, [x27, #0x80]\n"
-      "str q16, [x27, #0x90]\n"
-      "str q16, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
-      "str q16, [x27, #0xc0]\n"
-      "str q16, [x27, #0xd0]\n"
-      "str q16, [x27, #0xe0]\n"
-      "str q16, [x27, #0xf0]\n"
+      "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr d23, [x9], #0x8\n"
-      "ldr d22, [x26], #0x8\n"
-      "sub x28, x28, #0x4\n"
+      "ldr d19, [x9], #0x8\n"
+      "ldr d18, [x28], #0x8\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
+      "zip1 v17.8h, v19.8h, v17.8h\n"
+      "zip1 v16.8h, v18.8h, v16.8h\n"
       "ldr d18, [x25], #0x8\n"
-      "ldr d17, [x24], #0x8\n"
-      "cmp x28, #0x4\n"
-      "ldr d20, [x23], #0x8\n"
-      "ldr d21, [x22], #0x8\n"
-      "ldr d19, [x21], #0x8\n"
+      "ldr d21, [x23], #0x8\n"
+      "zip1 v20.8h, v17.8h, v16.8h\n"
+      "zip2 v19.8h, v17.8h, v16.8h\n"
+      "ldr d17, [x22], #0x8\n"
       "ldr d16, [x20], #0x8\n"
-      "zip1 v18.8h, v23.8h, v18.8h\n"
-      "zip1 v17.8h, v22.8h, v17.8h\n"
-      "zip1 v20.8h, v20.8h, v19.8h\n"
+      "zip1 v18.8h, v18.8h, v17.8h\n"
       "zip1 v16.8h, v21.8h, v16.8h\n"
-      "zip1 v19.8h, v18.8h, v17.8h\n"
-      "zip2 v18.8h, v18.8h, v17.8h\n"
-      "zip1 v17.8h, v20.8h, v16.8h\n"
-      "zip2 v16.8h, v20.8h, v16.8h\n"
-      "str q19, [x27, #0x0]\n"
-      "str q18, [x27, #0x10]\n"
-      "str q17, [x27, #0x80]\n"
-      "str q16, [x27, #0x90]\n"
-      "add x27, x27, #0x20\n"
+      "str q20, [x21, #0x0]\n"
+      "zip1 v17.8h, v18.8h, v16.8h\n"
+      "zip2 v16.8h, v18.8h, v16.8h\n"
+      "str q19, [x21, #0x10]\n"
+      "str q17, [x21, #0x80]\n"
+      "str q16, [x21, #0x90]\n"
+      "add x21, x21, #0x20\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr h23, [x9], #0x2\n"
-      "ldr h22, [x26], #0x2\n"
-      "sub x28, x28, #0x1\n"
-      "ldr h19, [x25], #0x2\n"
-      "ldr h17, [x24], #0x2\n"
-      "cmp x28, #0x1\n"
-      "ldr h21, [x23], #0x2\n"
-      "ldr h20, [x22], #0x2\n"
-      "ldr h18, [x21], #0x2\n"
-      "ldr h16, [x20], #0x2\n"
-      "zip1 v19.8h, v23.8h, v19.8h\n"
-      "zip1 v17.8h, v22.8h, v17.8h\n"
-      "zip1 v18.8h, v21.8h, v18.8h\n"
-      "zip1 v16.8h, v20.8h, v16.8h\n"
+      "ldr h19, [x9], #0x2\n"
+      "ldr h18, [x28], #0x2\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr h17, [x27], #0x2\n"
+      "ldr h16, [x26], #0x2\n"
       "zip1 v17.8h, v19.8h, v17.8h\n"
       "zip1 v16.8h, v18.8h, v16.8h\n"
-      "str d17, [x27, #0x0]\n"
-      "str d16, [x27, #0x80]\n"
-      "add x27, x27, #0x8\n"
+      "ldr h20, [x25], #0x2\n"
+      "ldr h19, [x23], #0x2\n"
+      "zip1 v18.8h, v17.8h, v16.8h\n"
+      "ldr h17, [x22], #0x2\n"
+      "ldr h16, [x20], #0x2\n"
+      "zip1 v17.8h, v20.8h, v17.8h\n"
+      "zip1 v16.8h, v19.8h, v16.8h\n"
+      "str d18, [x21, #0x0]\n"
+      "zip1 v16.8h, v17.8h, v16.8h\n"
+      "str d16, [x21, #0x80]\n"
+      "add x21, x21, #0x8\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x8\n"
       "add %x[out], %x[out], #0x100\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x9, %x[in]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x26, x27, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x9, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "csel x25, x25, %x[pad_row], GE\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x26, %x[in_stride]\n"
       "csel x26, x26, %x[pad_row], GT\n"
+      "csel x27, x27, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x28, x28, %x[pad_row], GT\n"
       "cmp x20, #0x20\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
-      "ldr q20, [x9], #0x10\n"
-      "ldr q19, [x26], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
       "sub x20, x20, #0x20\n"
-      "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x20\n"
-      "ldr q23, [x9], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "ldr q29, [x9], #0x10\n"
-      "zip1 v4.8h, v20.8h, v18.8h\n"
-      "zip1 v3.8h, v19.8h, v17.8h\n"
-      "ldr q28, [x26], #0x10\n"
-      "ldr q27, [x25], #0x10\n"
-      "zip2 v2.8h, v20.8h, v18.8h\n"
-      "zip2 v1.8h, v19.8h, v17.8h\n"
-      "ldr q20, [x24], #0x10\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v4.8h, v21.8h, v17.8h\n"
+      "zip1 v3.8h, v20.8h, v16.8h\n"
       "ldr q19, [x9], #0x10\n"
-      "zip1 v26.8h, v23.8h, v21.8h\n"
-      "zip1 v25.8h, v22.8h, v16.8h\n"
-      "ldr q18, [x26], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
-      "zip2 v24.8h, v23.8h, v21.8h\n"
-      "zip2 v23.8h, v22.8h, v16.8h\n"
-      "ldr q16, [x24], #0x10\n"
-      "zip1 v22.8h, v29.8h, v27.8h\n"
-      "zip1 v21.8h, v28.8h, v20.8h\n"
-      "zip2 v0.8h, v29.8h, v27.8h\n"
-      "zip2 v31.8h, v28.8h, v20.8h\n"
-      "zip1 v30.8h, v19.8h, v17.8h\n"
-      "zip1 v29.8h, v18.8h, v16.8h\n"
-      "zip2 v28.8h, v19.8h, v17.8h\n"
-      "zip2 v27.8h, v18.8h, v16.8h\n"
-      "zip1 v20.8h, v4.8h, v3.8h\n"
-      "zip2 v19.8h, v4.8h, v3.8h\n"
-      "zip1 v18.8h, v2.8h, v1.8h\n"
-      "zip2 v17.8h, v2.8h, v1.8h\n"
-      "zip1 v16.8h, v26.8h, v25.8h\n"
-      "zip2 v26.8h, v26.8h, v25.8h\n"
-      "zip1 v25.8h, v24.8h, v23.8h\n"
-      "zip2 v24.8h, v24.8h, v23.8h\n"
-      "str q20, [x27, #0x0]\n"
-      "str q19, [x27, #0x10]\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v2.8h, v21.8h, v17.8h\n"
+      "zip2 v1.8h, v20.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v0.8h, v19.8h, v17.8h\n"
+      "zip1 v31.8h, v18.8h, v16.8h\n"
+      "ldr q24, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v30.8h, v19.8h, v17.8h\n"
+      "zip2 v23.8h, v18.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.8h, v24.8h, v17.8h\n"
+      "zip1 v21.8h, v20.8h, v16.8h\n"
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v29.8h, v24.8h, v17.8h\n"
+      "zip2 v28.8h, v20.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v27.8h, v19.8h, v17.8h\n"
+      "zip1 v26.8h, v18.8h, v16.8h\n"
+      "zip2 v25.8h, v19.8h, v17.8h\n"
+      "zip2 v24.8h, v18.8h, v16.8h\n"
+      "zip1 v16.8h, v4.8h, v3.8h\n"
+      "zip2 v17.8h, v4.8h, v3.8h\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v16.8h, v2.8h, v1.8h\n"
+      "zip2 v20.8h, v2.8h, v1.8h\n"
+      "str q17, [x21, #0x10]\n"
+      "zip1 v19.8h, v0.8h, v31.8h\n"
+      "zip2 v18.8h, v0.8h, v31.8h\n"
+      "str q16, [x21, #0x20]\n"
+      "zip1 v17.8h, v30.8h, v23.8h\n"
+      "zip2 v16.8h, v30.8h, v23.8h\n"
+      "str q20, [x21, #0x30]\n"
+      "str q19, [x21, #0x40]\n"
       "zip1 v23.8h, v22.8h, v21.8h\n"
       "zip2 v22.8h, v22.8h, v21.8h\n"
-      "str q18, [x27, #0x20]\n"
-      "zip1 v21.8h, v0.8h, v31.8h\n"
-      "zip2 v20.8h, v0.8h, v31.8h\n"
-      "str q17, [x27, #0x30]\n"
-      "zip1 v19.8h, v30.8h, v29.8h\n"
-      "zip2 v18.8h, v30.8h, v29.8h\n"
-      "str q16, [x27, #0x40]\n"
-      "zip1 v17.8h, v28.8h, v27.8h\n"
-      "zip2 v16.8h, v28.8h, v27.8h\n"
-      "str q26, [x27, #0x50]\n"
-      "str q25, [x27, #0x60]\n"
-      "str q24, [x27, #0x70]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "str q23, [x27, #0x0]\n"
-      "str q22, [x27, #0x10]\n"
-      "str q21, [x27, #0x20]\n"
-      "str q20, [x27, #0x30]\n"
-      "str q19, [x27, #0x40]\n"
-      "str q18, [x27, #0x50]\n"
-      "str q17, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "str q18, [x21, #0x50]\n"
+      "zip1 v21.8h, v29.8h, v28.8h\n"
+      "zip2 v20.8h, v29.8h, v28.8h\n"
+      "str q17, [x21, #0x60]\n"
+      "zip1 v19.8h, v27.8h, v26.8h\n"
+      "zip2 v18.8h, v27.8h, v26.8h\n"
+      "str q16, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip1 v17.8h, v25.8h, v24.8h\n"
+      "zip2 v16.8h, v25.8h, v24.8h\n"
+      "str q23, [x21, #0x0]\n"
+      "str q22, [x21, #0x10]\n"
+      "str q21, [x21, #0x20]\n"
+      "str q20, [x21, #0x30]\n"
+      "str q19, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0x10\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
-      "ldr q20, [x9], #0x10\n"
-      "ldr q19, [x26], #0x10\n"
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
       "sub x20, x20, #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x10\n"
-      "ldr q24, [x9], #0x10\n"
-      "ldr q25, [x26], #0x10\n"
-      "ldr q23, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "zip1 v22.8h, v20.8h, v18.8h\n"
-      "zip1 v21.8h, v19.8h, v17.8h\n"
-      "zip2 v20.8h, v20.8h, v18.8h\n"
-      "zip2 v19.8h, v19.8h, v17.8h\n"
-      "zip1 v18.8h, v24.8h, v23.8h\n"
-      "zip1 v17.8h, v25.8h, v16.8h\n"
-      "zip2 v24.8h, v24.8h, v23.8h\n"
-      "zip2 v16.8h, v25.8h, v16.8h\n"
-      "zip1 v23.8h, v22.8h, v21.8h\n"
-      "zip2 v22.8h, v22.8h, v21.8h\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v25.8h, v19.8h, v17.8h\n"
+      "zip1 v24.8h, v18.8h, v16.8h\n"
+      "ldr q22, [x9], #0x10\n"
+      "ldr q21, [x28], #0x10\n"
+      "zip2 v20.8h, v19.8h, v17.8h\n"
+      "zip2 v19.8h, v18.8h, v16.8h\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v23.8h, v22.8h, v17.8h\n"
+      "zip1 v18.8h, v21.8h, v16.8h\n"
+      "zip2 v22.8h, v22.8h, v17.8h\n"
+      "zip2 v21.8h, v21.8h, v16.8h\n"
+      "zip1 v16.8h, v25.8h, v24.8h\n"
+      "zip2 v17.8h, v25.8h, v24.8h\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v16.8h, v20.8h, v19.8h\n"
       "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip1 v19.8h, v18.8h, v17.8h\n"
-      "zip2 v18.8h, v18.8h, v17.8h\n"
-      "zip1 v17.8h, v24.8h, v16.8h\n"
-      "zip2 v16.8h, v24.8h, v16.8h\n"
-      "str q23, [x27, #0x0]\n"
-      "str q22, [x27, #0x10]\n"
-      "str q21, [x27, #0x20]\n"
-      "str q20, [x27, #0x30]\n"
-      "str q19, [x27, #0x40]\n"
-      "str q18, [x27, #0x50]\n"
-      "str q17, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q17, [x21, #0x10]\n"
+      "zip1 v19.8h, v23.8h, v18.8h\n"
+      "zip2 v18.8h, v23.8h, v18.8h\n"
+      "str q16, [x21, #0x20]\n"
+      "zip1 v17.8h, v22.8h, v21.8h\n"
+      "zip2 v16.8h, v22.8h, v21.8h\n"
+      "str q20, [x21, #0x30]\n"
+      "str q19, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr d18, [x9], #0x8\n"
-      "ldr d19, [x26], #0x8\n"
+      "ldr d19, [x28], #0x8\n"
       "sub x20, x20, #0x4\n"
-      "ldr d17, [x25], #0x8\n"
-      "ldr d16, [x24], #0x8\n"
       "cmp x20, #0x4\n"
+      "ldr d17, [x27], #0x8\n"
+      "ldr d16, [x26], #0x8\n"
       "zip1 v18.8h, v18.8h, v17.8h\n"
       "zip1 v16.8h, v19.8h, v16.8h\n"
       "zip1 v17.8h, v18.8h, v16.8h\n"
       "zip2 v16.8h, v18.8h, v16.8h\n"
-      "str q17, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q17, [x21, #0x0]\n"
+      "str q16, [x21, #0x10]\n"
+      "add x21, x21, #0x20\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr h19, [x9], #0x2\n"
-      "ldr h18, [x26], #0x2\n"
+      "ldr h18, [x28], #0x2\n"
       "sub x20, x20, #0x1\n"
-      "ldr h17, [x25], #0x2\n"
-      "ldr h16, [x24], #0x2\n"
       "cmp x20, #0x1\n"
+      "ldr h17, [x27], #0x2\n"
+      "ldr h16, [x26], #0x2\n"
       "zip1 v17.8h, v19.8h, v17.8h\n"
       "zip1 v16.8h, v18.8h, v16.8h\n"
       "zip1 v16.8h, v17.8h, v16.8h\n"
-      "str d16, [x27, #0x0]\n"
-      "add x27, x27, #0x8\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x80\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4_fp32bf16.hpp
index a2f64768da..435398da0b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4_fp32bf16.hpp
@@ -40,313 +40,294 @@ void a64_transpose_interleave_16_2x4_fp32bf16(bfloat16 *out, const float *in, si
 
     __asm__ __volatile__(
       "cmp %x[height], #0x8\n"
-      "blt 9f\n"
+      "blt 8f\n"
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "cmp x28, #0x10\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x10\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
-      "ldr q2, [x9], #0x10\n"
-      "ldr q3, [x26], #0x10\n"
-      "sub x28, x28, #0x10\n"
-      "ldr q17, [x25], #0x10\n"
-      "ldr q4, [x24], #0x10\n"
-      "cmp x28, #0x10\n"
-      "ldr q16, [x23], #0x10\n"
+      "ldr q13, [x9], #0x10\n"
+      "ldr q12, [x28], #0x10\n"
+      "sub x24, x24, #0x10\n"
+      "cmp x24, #0x10\n"
+      "ldr q1, [x27], #0x10\n"
+      "ldr q9, [x26], #0x10\n"
+      "zip1 v19.4s, v13.4s, v1.4s\n"
+      "zip1 v14.4s, v12.4s, v9.4s\n"
+      "ldr q15, [x25], #0x10\n"
+      "ldr q4, [x23], #0x10\n"
+      "zip2 v8.4s, v13.4s, v1.4s\n"
+      "zip2 v28.4s, v12.4s, v9.4s\n"
+      "ldr q0, [x22], #0x10\n"
+      "ldr q1, [x20], #0x10\n"
+      "zip1 v16.4s, v15.4s, v0.4s\n"
+      "zip1 v5.4s, v4.4s, v1.4s\n"
+      "ldr q25, [x9], #0x10\n"
+      "ldr q24, [x28], #0x10\n"
+      "zip2 v3.4s, v15.4s, v0.4s\n"
+      "zip2 v2.4s, v4.4s, v1.4s\n"
+      "ldr q21, [x27], #0x10\n"
+      "ldr q30, [x26], #0x10\n"
+      "zip1 v18.4s, v25.4s, v21.4s\n"
+      "zip1 v27.4s, v24.4s, v30.4s\n"
+      "ldr q22, [x25], #0x10\n"
+      "ldr q20, [x23], #0x10\n"
+      "zip2 v9.4s, v25.4s, v21.4s\n"
+      "zip2 v10.4s, v24.4s, v30.4s\n"
       "ldr q1, [x22], #0x10\n"
-      "ldr q5, [x21], #0x10\n"
-      "ldr q15, [x20], #0x10\n"
-      "ldr q19, [x9], #0x10\n"
-      "ldr q11, [x26], #0x10\n"
-      "zip1 v14.4s, v2.4s, v17.4s\n"
-      "zip1 v27.4s, v3.4s, v4.4s\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q18, [x24], #0x10\n"
-      "zip2 v12.4s, v2.4s, v17.4s\n"
-      "zip2 v7.4s, v3.4s, v4.4s\n"
-      "ldr q22, [x23], #0x10\n"
-      "ldr q9, [x22], #0x10\n"
-      "zip1 v17.4s, v16.4s, v5.4s\n"
-      "zip1 v29.4s, v1.4s, v15.4s\n"
-      "ldr q30, [x21], #0x10\n"
-      "ldr q8, [x20], #0x10\n"
-      "zip2 v28.4s, v16.4s, v5.4s\n"
-      "zip2 v26.4s, v1.4s, v15.4s\n"
+      "ldr q21, [x20], #0x10\n"
+      "zip1 v25.4s, v22.4s, v1.4s\n"
+      "zip1 v7.4s, v20.4s, v21.4s\n"
       "ldr q31, [x9], #0x10\n"
-      "ldr q13, [x26], #0x10\n"
-      "zip1 v20.4s, v19.4s, v21.4s\n"
-      "zip1 v10.4s, v11.4s, v18.4s\n"
-      "ldr q2, [x25], #0x10\n"
-      "ldr q23, [x24], #0x10\n"
-      "zip2 v4.4s, v19.4s, v21.4s\n"
-      "zip2 v6.4s, v11.4s, v18.4s\n"
-      "ldr q21, [x23], #0x10\n"
+      "ldr q17, [x28], #0x10\n"
+      "zip2 v30.4s, v22.4s, v1.4s\n"
+      "zip2 v20.4s, v20.4s, v21.4s\n"
+      "ldr q15, [x27], #0x10\n"
+      "ldr q24, [x26], #0x10\n"
+      "zip1 v6.4s, v31.4s, v15.4s\n"
+      "zip1 v4.4s, v17.4s, v24.4s\n"
+      "ldr q12, [x25], #0x10\n"
+      "ldr q29, [x23], #0x10\n"
+      "zip2 v22.4s, v31.4s, v15.4s\n"
+      "zip2 v26.4s, v17.4s, v24.4s\n"
+      "ldr q0, [x22], #0x10\n"
+      "ldr q24, [x20], #0x10\n"
+      "zip1 v17.4s, v12.4s, v0.4s\n"
+      "zip1 v31.4s, v29.4s, v24.4s\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q1, [x28], #0x10\n"
+      "zip2 v23.4s, v12.4s, v0.4s\n"
+      "zip2 v24.4s, v29.4s, v24.4s\n"
+      "ldr q11, [x27], #0x10\n"
+      "ldr q29, [x26], #0x10\n"
+      "zip1 v0.4s, v21.4s, v11.4s\n"
+      "zip1 v13.4s, v1.4s, v29.4s\n"
+      "ldr q15, [x25], #0x10\n"
+      "ldr q12, [x23], #0x10\n"
+      "zip2 v21.4s, v21.4s, v11.4s\n"
+      "zip2 v29.4s, v1.4s, v29.4s\n"
       "ldr q1, [x22], #0x10\n"
-      "zip1 v24.4s, v22.4s, v30.4s\n"
-      "zip1 v16.4s, v9.4s, v8.4s\n"
-      "ldr q18, [x21], #0x10\n"
-      "ldr q15, [x20], #0x10\n"
-      "zip2 v22.4s, v22.4s, v30.4s\n"
-      "zip2 v25.4s, v9.4s, v8.4s\n"
-      "ldr q3, [x9], #0x10\n"
-      "ldr q5, [x26], #0x10\n"
-      "zip1 v8.4s, v31.4s, v2.4s\n"
-      "zip1 v30.4s, v13.4s, v23.4s\n"
-      "ldr q0, [x25], #0x10\n"
-      "ldr q19, [x24], #0x10\n"
-      "zip2 v31.4s, v31.4s, v2.4s\n"
-      "zip2 v2.4s, v13.4s, v23.4s\n"
-      "ldr q11, [x23], #0x10\n"
-      "ldr q9, [x22], #0x10\n"
-      "zip1 v13.4s, v21.4s, v18.4s\n"
-      "zip1 v23.4s, v1.4s, v15.4s\n"
-      "zip2 v21.4s, v21.4s, v18.4s\n"
-      "ldr q18, [x21], #0x10\n"
-      "zip2 v15.4s, v1.4s, v15.4s\n"
-      "zip1 v1.4s, v3.4s, v0.4s\n"
-      "zip2 v0.4s, v3.4s, v0.4s\n"
-      "zip1 v3.4s, v5.4s, v19.4s\n"
-      "zip2 v5.4s, v5.4s, v19.4s\n"
-      "zip1 v19.4s, v11.4s, v18.4s\n"
-      "zip2 v18.4s, v11.4s, v18.4s\n"
-      "zip1 v11.4s, v14.4s, v27.4s\n"
-      "zip2 v14.4s, v14.4s, v27.4s\n"
-      "ldr q27, [x20], #0x10\n"
-      ".inst 0x0ea1696b  // bfcvtn v11.4h, v11.4s\n"
-      ".inst 0x4ea169cb  // bfcvtn2 v11.8h, v14.4s\n"
-      "zip1 v14.4s, v12.4s, v7.4s\n"
-      "zip2 v7.4s, v12.4s, v7.4s\n"
-      "zip1 v12.4s, v9.4s, v27.4s\n"
-      "zip2 v27.4s, v9.4s, v27.4s\n"
-      "zip1 v9.4s, v20.4s, v10.4s\n"
-      "zip2 v20.4s, v20.4s, v10.4s\n"
-      "str q11, [x27, #0x0]\n"
-      "zip1 v10.4s, v4.4s, v6.4s\n"
-      "zip1 v11.4s, v8.4s, v30.4s\n"
-      ".inst 0x0ea169ce  // bfcvtn v14.4h, v14.4s\n"
-      "zip2 v4.4s, v4.4s, v6.4s\n"
-      "zip1 v6.4s, v31.4s, v2.4s\n"
-      ".inst 0x0ea16929  // bfcvtn v9.4h, v9.4s\n"
-      ".inst 0x0ea1694a  // bfcvtn v10.4h, v10.4s\n"
-      ".inst 0x0ea1696b  // bfcvtn v11.4h, v11.4s\n"
-      "zip2 v30.4s, v8.4s, v30.4s\n"
-      "zip1 v8.4s, v1.4s, v3.4s\n"
-      ".inst 0x0ea168c6  // bfcvtn v6.4h, v6.4s\n"
-      "zip2 v31.4s, v31.4s, v2.4s\n"
-      "zip1 v2.4s, v0.4s, v5.4s\n"
-      "zip2 v3.4s, v1.4s, v3.4s\n"
-      "zip1 v1.4s, v17.4s, v29.4s\n"
+      "zip1 v11.4s, v15.4s, v1.4s\n"
+      "zip2 v1.4s, v15.4s, v1.4s\n"
+      "zip1 v15.4s, v19.4s, v14.4s\n"
+      ".inst 0x0ea169ef  // bfcvtn v15.4h, v15.4s\n"
+      "zip2 v14.4s, v19.4s, v14.4s\n"
+      "ldr q19, [x20], #0x10\n"
+      ".inst 0x4ea169cf  // bfcvtn2 v15.8h, v14.4s\n"
+      "str q15, [x21, #0x0]\n"
+      "zip1 v14.4s, v12.4s, v19.4s\n"
+      "zip2 v15.4s, v12.4s, v19.4s\n"
+      "zip1 v12.4s, v8.4s, v28.4s\n"
+      "zip1 v19.4s, v18.4s, v27.4s\n"
+      ".inst 0x0ea1698c  // bfcvtn v12.4h, v12.4s\n"
+      "zip2 v28.4s, v8.4s, v28.4s\n"
+      "zip1 v8.4s, v9.4s, v10.4s\n"
+      ".inst 0x0ea16a73  // bfcvtn v19.4h, v19.4s\n"
+      "zip2 v18.4s, v18.4s, v27.4s\n"
+      "zip1 v27.4s, v6.4s, v4.4s\n"
       ".inst 0x0ea16908  // bfcvtn v8.4h, v8.4s\n"
-      "zip2 v5.4s, v0.4s, v5.4s\n"
-      "zip1 v0.4s, v28.4s, v26.4s\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      "zip2 v17.4s, v17.4s, v29.4s\n"
-      "zip1 v29.4s, v24.4s, v16.4s\n"
-      ".inst 0x0ea16821  // bfcvtn v1.4h, v1.4s\n"
+      "zip2 v10.4s, v9.4s, v10.4s\n"
+      "zip1 v9.4s, v22.4s, v26.4s\n"
+      ".inst 0x0ea16b7b  // bfcvtn v27.4h, v27.4s\n"
+      "zip2 v6.4s, v6.4s, v4.4s\n"
+      "zip1 v4.4s, v0.4s, v13.4s\n"
+      ".inst 0x0ea16929  // bfcvtn v9.4h, v9.4s\n"
+      "zip2 v22.4s, v22.4s, v26.4s\n"
+      "zip1 v26.4s, v21.4s, v29.4s\n"
+      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
+      "zip2 v13.4s, v0.4s, v13.4s\n"
+      "zip1 v0.4s, v16.4s, v5.4s\n"
+      ".inst 0x0ea16b5a  // bfcvtn v26.4h, v26.4s\n"
+      "zip2 v21.4s, v21.4s, v29.4s\n"
+      "zip1 v29.4s, v3.4s, v2.4s\n"
       ".inst 0x0ea16800  // bfcvtn v0.4h, v0.4s\n"
-      "zip2 v28.4s, v28.4s, v26.4s\n"
-      "zip1 v26.4s, v22.4s, v25.4s\n"
-      "zip2 v24.4s, v24.4s, v16.4s\n"
-      "zip1 v16.4s, v13.4s, v23.4s\n"
+      "zip2 v5.4s, v16.4s, v5.4s\n"
+      "zip1 v16.4s, v25.4s, v7.4s\n"
       ".inst 0x0ea16bbd  // bfcvtn v29.4h, v29.4s\n"
-      "zip2 v25.4s, v22.4s, v25.4s\n"
-      "zip1 v22.4s, v21.4s, v15.4s\n"
-      ".inst 0x0ea16b5a  // bfcvtn v26.4h, v26.4s\n"
-      "zip2 v13.4s, v13.4s, v23.4s\n"
-      "zip1 v23.4s, v19.4s, v12.4s\n"
+      "zip2 v2.4s, v3.4s, v2.4s\n"
+      "zip1 v3.4s, v30.4s, v20.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      ".inst 0x0ea16ad6  // bfcvtn v22.4h, v22.4s\n"
-      "zip2 v21.4s, v21.4s, v15.4s\n"
-      "zip1 v15.4s, v18.4s, v27.4s\n"
-      "zip2 v19.4s, v19.4s, v12.4s\n"
+      "zip2 v7.4s, v25.4s, v7.4s\n"
+      "zip1 v25.4s, v17.4s, v31.4s\n"
+      ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
+      "zip2 v30.4s, v30.4s, v20.4s\n"
+      "zip1 v20.4s, v23.4s, v24.4s\n"
+      ".inst 0x0ea16b39  // bfcvtn v25.4h, v25.4s\n"
+      "zip2 v17.4s, v17.4s, v31.4s\n"
+      "zip1 v31.4s, v11.4s, v14.4s\n"
+      ".inst 0x0ea16a94  // bfcvtn v20.4h, v20.4s\n"
+      "zip2 v24.4s, v23.4s, v24.4s\n"
+      "zip1 v23.4s, v1.4s, v15.4s\n"
+      ".inst 0x0ea16bff  // bfcvtn v31.4h, v31.4s\n"
+      "zip2 v14.4s, v11.4s, v14.4s\n"
       ".inst 0x0ea16af7  // bfcvtn v23.4h, v23.4s\n"
-      "zip2 v12.4s, v18.4s, v27.4s\n"
-      ".inst 0x4ea168ee  // bfcvtn2 v14.8h, v7.4s\n"
-      ".inst 0x4ea16a89  // bfcvtn2 v9.8h, v20.4s\n"
-      ".inst 0x0ea169f2  // bfcvtn v18.4h, v15.4s\n"
-      ".inst 0x4ea1688a  // bfcvtn2 v10.8h, v4.4s\n"
-      ".inst 0x4ea16bcb  // bfcvtn2 v11.8h, v30.4s\n"
-      ".inst 0x4ea16be6  // bfcvtn2 v6.8h, v31.4s\n"
-      ".inst 0x4ea16868  // bfcvtn2 v8.8h, v3.4s\n"
-      ".inst 0x4ea168a2  // bfcvtn2 v2.8h, v5.4s\n"
-      "str q14, [x27, #0x10]\n"
-      ".inst 0x4ea16a21  // bfcvtn2 v1.8h, v17.4s\n"
-      ".inst 0x4ea16b80  // bfcvtn2 v0.8h, v28.4s\n"
-      "str q9, [x27, #0x20]\n"
-      ".inst 0x4ea16b1d  // bfcvtn2 v29.8h, v24.4s\n"
-      ".inst 0x4ea16b3a  // bfcvtn2 v26.8h, v25.4s\n"
-      "str q10, [x27, #0x30]\n"
-      ".inst 0x4ea169b0  // bfcvtn2 v16.8h, v13.4s\n"
-      ".inst 0x4ea16ab6  // bfcvtn2 v22.8h, v21.4s\n"
-      "str q11, [x27, #0x40]\n"
-      ".inst 0x4ea16a77  // bfcvtn2 v23.8h, v19.4s\n"
-      ".inst 0x4ea16992  // bfcvtn2 v18.8h, v12.4s\n"
-      "str q6, [x27, #0x50]\n"
-      "str q8, [x27, #0x60]\n"
-      "str q2, [x27, #0x70]\n"
-      "str q1, [x27, #0x80]\n"
-      "str q0, [x27, #0x90]\n"
-      "str q29, [x27, #0xa0]\n"
-      "str q26, [x27, #0xb0]\n"
-      "str q16, [x27, #0xc0]\n"
-      "str q22, [x27, #0xd0]\n"
-      "str q23, [x27, #0xe0]\n"
-      "str q18, [x27, #0xf0]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip2 v1.4s, v1.4s, v15.4s\n"
+      ".inst 0x4ea16b8c  // bfcvtn2 v12.8h, v28.4s\n"
+      "str q12, [x21, #0x10]\n"
+      ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
+      ".inst 0x4ea16948  // bfcvtn2 v8.8h, v10.4s\n"
+      "str q19, [x21, #0x20]\n"
+      ".inst 0x4ea168db  // bfcvtn2 v27.8h, v6.4s\n"
+      ".inst 0x4ea16ac9  // bfcvtn2 v9.8h, v22.4s\n"
+      "str q8, [x21, #0x30]\n"
+      ".inst 0x4ea169a4  // bfcvtn2 v4.8h, v13.4s\n"
+      ".inst 0x4ea16aba  // bfcvtn2 v26.8h, v21.4s\n"
+      "str q27, [x21, #0x40]\n"
+      ".inst 0x4ea168a0  // bfcvtn2 v0.8h, v5.4s\n"
+      ".inst 0x4ea1685d  // bfcvtn2 v29.8h, v2.4s\n"
+      "str q9, [x21, #0x50]\n"
+      ".inst 0x4ea168f0  // bfcvtn2 v16.8h, v7.4s\n"
+      ".inst 0x4ea16bc3  // bfcvtn2 v3.8h, v30.4s\n"
+      "str q4, [x21, #0x60]\n"
+      ".inst 0x4ea16a39  // bfcvtn2 v25.8h, v17.4s\n"
+      ".inst 0x4ea16b14  // bfcvtn2 v20.8h, v24.4s\n"
+      "str q26, [x21, #0x70]\n"
+      ".inst 0x4ea169df  // bfcvtn2 v31.8h, v14.4s\n"
+      ".inst 0x4ea16837  // bfcvtn2 v23.8h, v1.4s\n"
+      "str q0, [x21, #0x80]\n"
+      "str q29, [x21, #0x90]\n"
+      "str q16, [x21, #0xa0]\n"
+      "str q3, [x21, #0xb0]\n"
+      "str q25, [x21, #0xc0]\n"
+      "str q20, [x21, #0xd0]\n"
+      "str q31, [x21, #0xe0]\n"
+      "str q23, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x28, 8f\n"
-      "cmp x28, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "str q16, [x27, #0x80]\n"
-      "str q16, [x27, #0x90]\n"
-      "str q16, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
-      "str q16, [x27, #0xc0]\n"
-      "str q16, [x27, #0xd0]\n"
-      "str q16, [x27, #0xe0]\n"
-      "str q16, [x27, #0xf0]\n"
+      "cmp x24, #0x4\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 4 loop: loop
-      "ldr q25, [x9], #0x10\n"
-      "ldr q24, [x26], #0x10\n"
-      "sub x28, x28, #0x4\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "cmp x28, #0x4\n"
-      "ldr q23, [x23], #0x10\n"
-      "ldr q19, [x22], #0x10\n"
-      "ldr q18, [x21], #0x10\n"
-      "ldr q17, [x20], #0x10\n"
-      "zip1 v22.4s, v25.4s, v21.4s\n"
-      "zip1 v16.4s, v24.4s, v20.4s\n"
-      "zip2 v21.4s, v25.4s, v21.4s\n"
-      "zip2 v20.4s, v24.4s, v20.4s\n"
-      "zip1 v27.4s, v23.4s, v18.4s\n"
-      "zip1 v26.4s, v19.4s, v17.4s\n"
-      "zip2 v25.4s, v23.4s, v18.4s\n"
-      "zip2 v24.4s, v19.4s, v17.4s\n"
-      "zip1 v19.4s, v22.4s, v16.4s\n"
-      "zip1 v18.4s, v21.4s, v20.4s\n"
+      "ldr q23, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.4s, v23.4s, v17.4s\n"
+      "zip1 v21.4s, v20.4s, v16.4s\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v28.4s, v23.4s, v17.4s\n"
+      "zip2 v20.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v27.4s, v19.4s, v17.4s\n"
+      "zip1 v26.4s, v18.4s, v16.4s\n"
+      "zip2 v25.4s, v19.4s, v17.4s\n"
+      "zip2 v24.4s, v18.4s, v16.4s\n"
+      "zip1 v19.4s, v22.4s, v21.4s\n"
+      "zip1 v18.4s, v28.4s, v20.4s\n"
       "zip1 v17.4s, v27.4s, v26.4s\n"
-      "zip2 v23.4s, v22.4s, v16.4s\n"
       "zip1 v16.4s, v25.4s, v24.4s\n"
-      "zip2 v22.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a75  // bfcvtn v21.4h, v19.4s\n"
-      ".inst 0x0ea16a54  // bfcvtn v20.4h, v18.4s\n"
+      ".inst 0x0ea16a77  // bfcvtn v23.4h, v19.4s\n"
+      "zip2 v22.4s, v22.4s, v21.4s\n"
+      ".inst 0x0ea16a55  // bfcvtn v21.4h, v18.4s\n"
+      "zip2 v20.4s, v28.4s, v20.4s\n"
       ".inst 0x0ea16a33  // bfcvtn v19.4h, v17.4s\n"
       "zip2 v18.4s, v27.4s, v26.4s\n"
       ".inst 0x0ea16a11  // bfcvtn v17.4h, v16.4s\n"
       "zip2 v16.4s, v25.4s, v24.4s\n"
-      ".inst 0x4ea16af5  // bfcvtn2 v21.8h, v23.4s\n"
-      ".inst 0x4ea16ad4  // bfcvtn2 v20.8h, v22.4s\n"
+      ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
+      ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
+      "str q23, [x21, #0x0]\n"
       ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
       ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0x80]\n"
-      "str q17, [x27, #0x90]\n"
-      "add x27, x27, #0x20\n"
+      "str q21, [x21, #0x10]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q17, [x21, #0x90]\n"
+      "add x21, x21, #0x20\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 1 loop: loop
-      "ldr s23, [x9], #0x4\n"
-      "ldr s22, [x26], #0x4\n"
-      "sub x28, x28, #0x1\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s17, [x24], #0x4\n"
-      "cmp x28, #0x1\n"
-      "ldr s21, [x23], #0x4\n"
-      "ldr s20, [x22], #0x4\n"
-      "ldr s18, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
-      "zip1 v19.4s, v23.4s, v19.4s\n"
-      "zip1 v17.4s, v22.4s, v17.4s\n"
-      "zip1 v18.4s, v21.4s, v18.4s\n"
-      "zip1 v16.4s, v20.4s, v16.4s\n"
+      "ldr s19, [x9], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
       "zip1 v17.4s, v19.4s, v17.4s\n"
       "zip1 v16.4s, v18.4s, v16.4s\n"
-      ".inst 0x0ea16a31  // bfcvtn v17.4h, v17.4s\n"
+      "ldr s20, [x25], #0x4\n"
+      "ldr s19, [x23], #0x4\n"
+      "zip1 v16.4s, v17.4s, v16.4s\n"
+      ".inst 0x0ea16a12  // bfcvtn v18.4h, v16.4s\n"
+      "ldr s17, [x22], #0x4\n"
+      "ldr s16, [x20], #0x4\n"
+      "zip1 v17.4s, v20.4s, v17.4s\n"
+      "zip1 v16.4s, v19.4s, v16.4s\n"
+      "zip1 v16.4s, v17.4s, v16.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      "str d17, [x27, #0x0]\n"
-      "str d16, [x27, #0x80]\n"
-      "add x27, x27, #0x8\n"
+      "str d18, [x21, #0x0]\n"
+      "str d16, [x21, #0x80]\n"
+      "add x21, x21, #0x8\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 1 loop: skip
-      "8:"  // Main row loop: odd col skip
       "cmp %x[height], #0x8\n"
       "add %x[out], %x[out], #0x100\n"
       "bge 1b\n"
-      "cbz %x[height], 18f\n"
-      "9:"  // Main loop skip
-      "10:"  // Tail row loop: Head
+      "cbz %x[height], 16f\n"
+      "8:"  // Main loop skip
+      "9:"  // Tail row loop: Head
       "mov x9, %x[in]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x26, x27, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x9, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "csel x25, x25, %x[pad_row], GE\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x26, %x[in_stride]\n"
       "csel x26, x26, %x[pad_row], GT\n"
+      "csel x27, x27, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x28, x28, %x[pad_row], GT\n"
       "cmp x20, #0x10\n"
-      "blt 12f\n"
-      "11:"  // Tail row loop: Column loop
-      "ldr q20, [x9], #0x10\n"
-      "ldr q19, [x26], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "blt 11f\n"
+      "10:"  // Tail row loop: Column loop
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
       "sub x20, x20, #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x10\n"
-      "ldr q0, [x9], #0x10\n"
-      "ldr q31, [x26], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "ldr q23, [x9], #0x10\n"
-      "zip1 v30.4s, v20.4s, v18.4s\n"
-      "zip1 v29.4s, v19.4s, v17.4s\n"
-      "ldr q22, [x26], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
-      "zip2 v28.4s, v20.4s, v18.4s\n"
-      "zip2 v27.4s, v19.4s, v17.4s\n"
-      "ldr q20, [x24], #0x10\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v30.4s, v21.4s, v17.4s\n"
+      "zip1 v29.4s, v20.4s, v16.4s\n"
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v28.4s, v21.4s, v17.4s\n"
+      "zip2 v27.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v26.4s, v19.4s, v17.4s\n"
+      "zip1 v25.4s, v18.4s, v16.4s\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v8.4s, v19.4s, v17.4s\n"
+      "zip2 v24.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v7.4s, v21.4s, v17.4s\n"
+      "zip1 v6.4s, v20.4s, v16.4s\n"
       "ldr q19, [x9], #0x10\n"
-      "zip1 v26.4s, v0.4s, v24.4s\n"
-      "zip1 v25.4s, v31.4s, v16.4s\n"
-      "ldr q18, [x26], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
-      "zip2 v8.4s, v0.4s, v24.4s\n"
-      "zip2 v24.4s, v31.4s, v16.4s\n"
-      "ldr q16, [x24], #0x10\n"
-      "zip1 v7.4s, v23.4s, v21.4s\n"
-      "zip1 v6.4s, v22.4s, v20.4s\n"
-      "zip2 v5.4s, v23.4s, v21.4s\n"
-      "zip2 v4.4s, v22.4s, v20.4s\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v5.4s, v21.4s, v17.4s\n"
+      "zip2 v4.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
       "zip1 v3.4s, v19.4s, v17.4s\n"
       "zip1 v2.4s, v18.4s, v16.4s\n"
       "zip2 v1.4s, v19.4s, v17.4s\n"
@@ -377,81 +358,70 @@ void a64_transpose_interleave_16_2x4_fp32bf16(bfloat16 *out, const float *in, si
       "zip2 v16.4s, v1.4s, v0.4s\n"
       ".inst 0x4ea16bdf  // bfcvtn2 v31.8h, v30.4s\n"
       ".inst 0x4ea16b9d  // bfcvtn2 v29.8h, v28.4s\n"
+      "str q31, [x21, #0x0]\n"
       ".inst 0x4ea16b5b  // bfcvtn2 v27.8h, v26.4s\n"
       ".inst 0x4ea16b19  // bfcvtn2 v25.8h, v24.4s\n"
+      "str q29, [x21, #0x10]\n"
       ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
       ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
+      "str q27, [x21, #0x20]\n"
       ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
       ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
-      "str q31, [x27, #0x0]\n"
-      "str q29, [x27, #0x10]\n"
-      "str q27, [x27, #0x20]\n"
-      "str q25, [x27, #0x30]\n"
-      "str q23, [x27, #0x40]\n"
-      "str q21, [x27, #0x50]\n"
-      "str q19, [x27, #0x60]\n"
-      "str q17, [x27, #0x70]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 11b\n"
-      "12:"  // Tail row loop: Column loop skip
-      "cbz x20, 17f\n"
+      "str q25, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q21, [x21, #0x50]\n"
+      "str q19, [x21, #0x60]\n"
+      "str q17, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 10b\n"
+      "11:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: width 4 loop: loop
-      "ldr q21, [x9], #0x10\n"
-      "ldr q20, [x26], #0x10\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: width 4 loop: loop
+      "ldr q20, [x9], #0x10\n"
+      "ldr q19, [x28], #0x10\n"
       "sub x20, x20, #0x4\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x4\n"
-      "zip1 v18.4s, v21.4s, v19.4s\n"
-      "zip1 v16.4s, v20.4s, v17.4s\n"
-      "zip2 v21.4s, v21.4s, v19.4s\n"
-      "zip2 v20.4s, v20.4s, v17.4s\n"
-      "zip1 v17.4s, v18.4s, v16.4s\n"
-      "zip2 v19.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.4s, v20.4s, v17.4s\n"
+      "zip1 v18.4s, v19.4s, v16.4s\n"
+      "zip2 v21.4s, v20.4s, v17.4s\n"
+      "zip2 v20.4s, v19.4s, v16.4s\n"
+      "zip1 v17.4s, v22.4s, v18.4s\n"
       "zip1 v16.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a32  // bfcvtn v18.4h, v17.4s\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      ".inst 0x4ea16a72  // bfcvtn2 v18.8h, v19.4s\n"
-      ".inst 0x4ea16a30  // bfcvtn2 v16.8h, v17.4s\n"
-      "str q18, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: width 4 loop: skip
+      ".inst 0x0ea16a33  // bfcvtn v19.4h, v17.4s\n"
+      "zip2 v18.4s, v22.4s, v18.4s\n"
+      ".inst 0x0ea16a11  // bfcvtn v17.4h, v16.4s\n"
+      "zip2 v16.4s, v21.4s, v20.4s\n"
+      ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
+      ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
+      "str q19, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "add x21, x21, #0x20\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 1 loop: loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 1 loop: loop
       "ldr s19, [x9], #0x4\n"
-      "ldr s18, [x26], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
       "sub x20, x20, #0x1\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s16, [x24], #0x4\n"
       "cmp x20, #0x1\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
       "zip1 v17.4s, v19.4s, v17.4s\n"
       "zip1 v16.4s, v18.4s, v16.4s\n"
       "zip1 v16.4s, v17.4s, v16.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      "str d16, [x27, #0x0]\n"
-      "add x27, x27, #0x8\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 1 loop: skip
-      "17:"  // Tail row loop: odd col skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x80\n"
-      "bge 10b\n"
-      "18:"  // Done
+      "bge 9b\n"
+      "16:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24.hpp
index 0f00300c54..a49acf1449 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24.hpp
@@ -34,206 +34,192 @@ void a64_transpose_interleave_24(uint16_t *out, const uint16_t *in, size_t width
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
       "ldr q1, [x25], #0x10\n"
-      "ldr q0, [x21], #0x10\n"
+      "ldr q0, [x22], #0x10\n"
       "sub x24, x24, #0x18\n"
-      "ldr q17, [x25], #0x10\n"
-      "ldr q31, [x22], #0x10\n"
       "cmp x24, #0x18\n"
-      "ldr q16, [x21], #0x10\n"
-      "ldr q30, [x20], #0x10\n"
-      "ldr q29, [x25], #0x10\n"
-      "ldr q28, [x21], #0x10\n"
-      "ldr q27, [x22], #0x10\n"
-      "dup v26.2d, v17.d[0]\n"
-      "dup v25.2d, v31.d[1]\n"
-      "ldr q24, [x20], #0x10\n"
-      "ldr q23, [x22], #0x10\n"
-      "dup v22.2d, v16.d[0]\n"
-      "dup v21.2d, v30.d[1]\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q31, [x23], #0x10\n"
+      "dup v30.2d, v17.d[0]\n"
+      "dup v29.2d, v31.d[1]\n"
+      "ldr q16, [x22], #0x10\n"
+      "ldr q28, [x20], #0x10\n"
+      "dup v27.2d, v16.d[0]\n"
+      "dup v26.2d, v28.d[1]\n"
+      "ldr q25, [x25], #0x10\n"
+      "ldr q24, [x22], #0x10\n"
+      "dup v23.2d, v17.d[1]\n"
+      "dup v22.2d, v25.d[1]\n"
+      "ldr q21, [x23], #0x10\n"
       "ldr q20, [x20], #0x10\n"
-      "dup v19.2d, v17.d[1]\n"
-      "dup v18.2d, v29.d[1]\n"
-      "str q1, [x23, #0x0]\n"
-      "dup v17.2d, v16.d[1]\n"
-      "dup v16.2d, v28.d[1]\n"
-      "mov v26.d[1], v31.d[0]\n"
-      "mov v25.d[1], v27.d[0]\n"
-      "mov v22.d[1], v30.d[0]\n"
-      "mov v21.d[1], v24.d[0]\n"
-      "str q26, [x23, #0x10]\n"
-      "str q25, [x23, #0x20]\n"
-      "mov v19.d[1], v29.d[0]\n"
-      "mov v18.d[1], v27.d[1]\n"
-      "str q0, [x23, #0x30]\n"
-      "mov v17.d[1], v28.d[0]\n"
-      "mov v16.d[1], v24.d[1]\n"
-      "str q22, [x23, #0x40]\n"
-      "str q21, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x10]\n"
-      "str q23, [x23, #0x20]\n"
-      "str q17, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q20, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "dup v19.2d, v16.d[1]\n"
+      "dup v18.2d, v24.d[1]\n"
+      "ldr q17, [x23], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "mov v30.d[1], v31.d[0]\n"
+      "mov v29.d[1], v21.d[0]\n"
+      "mov v27.d[1], v28.d[0]\n"
+      "mov v26.d[1], v20.d[0]\n"
+      "str q1, [x21, #0x0]\n"
+      "str q30, [x21, #0x10]\n"
+      "mov v23.d[1], v25.d[0]\n"
+      "mov v22.d[1], v21.d[1]\n"
+      "str q29, [x21, #0x20]\n"
+      "mov v19.d[1], v24.d[0]\n"
+      "mov v18.d[1], v20.d[1]\n"
+      "str q0, [x21, #0x30]\n"
+      "str q27, [x21, #0x40]\n"
+      "str q26, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q23, [x21, #0x0]\n"
+      "str q22, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q19, [x21, #0x30]\n"
+      "str q18, [x21, #0x40]\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr q25, [x22], #0x10\n"
-      "ldr q24, [x20], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
+      "ldr q23, [x20], #0x10\n"
+      "dup v22.2d, v17.d[1]\n"
+      "dup v21.2d, v23.d[1]\n"
+      "ldr q20, [x25], #0x10\n"
+      "ldr q19, [x22], #0x10\n"
       "sub x24, x24, #0xc\n"
-      "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x21], #0x10\n"
       "cmp x24, #0xc\n"
-      "ldr d21, [x25], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "ldr d19, [x21], #0x8\n"
-      "ldr d18, [x20], #0x8\n"
-      "dup v17.2d, v25.d[1]\n"
-      "dup v16.2d, v24.d[1]\n"
-      "str q23, [x23, #0x0]\n"
-      "mov v21.d[1], v25.d[0]\n"
-      "mov v17.d[1], v20.d[0]\n"
-      "mov v19.d[1], v24.d[0]\n"
-      "mov v16.d[1], v18.d[0]\n"
-      "str q21, [x23, #0x10]\n"
-      "str q17, [x23, #0x20]\n"
-      "str q22, [x23, #0x30]\n"
-      "str q19, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "ldr d18, [x25], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
+      "mov v18.d[1], v17.d[0]\n"
+      "mov v22.d[1], v16.d[0]\n"
+      "ldr d17, [x22], #0x8\n"
+      "ldr d16, [x20], #0x8\n"
+      "mov v17.d[1], v23.d[0]\n"
+      "mov v21.d[1], v16.d[0]\n"
+      "str q20, [x21, #0x0]\n"
+      "str q18, [x21, #0x10]\n"
+      "str q22, [x21, #0x20]\n"
+      "str q19, [x21, #0x30]\n"
+      "str q17, [x21, #0x40]\n"
+      "str q21, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr d19, [x25], #0x8\n"
-      "ldr d18, [x22], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "sub x24, x24, #0x4\n"
-      "ldr d17, [x21], #0x8\n"
-      "ldr d16, [x20], #0x8\n"
       "cmp x24, #0x4\n"
-      "str d19, [x23, #0x0]\n"
-      "str d18, [x23, #0x18]\n"
-      "str d17, [x23, #0x30]\n"
-      "str d16, [x23, #0x48]\n"
-      "add x23, x23, #0x8\n"
+      "ldr d17, [x22], #0x8\n"
+      "ldr d16, [x20], #0x8\n"
+      "str d19, [x21, #0x0]\n"
+      "str d18, [x21, #0x18]\n"
+      "str d17, [x21, #0x30]\n"
+      "str d16, [x21, #0x48]\n"
+      "add x21, x21, #0x8\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr h19, [x25], #0x2\n"
-      "ldr h18, [x22], #0x2\n"
+      "ldr h18, [x23], #0x2\n"
       "sub x24, x24, #0x1\n"
-      "ldr h17, [x21], #0x2\n"
-      "ldr h16, [x20], #0x2\n"
       "cmp x24, #0x1\n"
-      "str h19, [x23, #0x0]\n"
-      "str h18, [x23, #0x18]\n"
-      "str h17, [x23, #0x30]\n"
-      "str h16, [x23, #0x48]\n"
-      "add x23, x23, #0x2\n"
+      "ldr h17, [x22], #0x2\n"
+      "ldr h16, [x20], #0x2\n"
+      "str h19, [x21, #0x0]\n"
+      "str h18, [x21, #0x18]\n"
+      "str h17, [x21, #0x30]\n"
+      "str h16, [x21, #0x48]\n"
+      "add x21, x21, #0x2\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0x60\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x18\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
       "ldr q19, [x25], #0x10\n"
-      "sub x20, x20, #0x18\n"
       "ldr q16, [x25], #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "cmp x20, #0x18\n"
-      "dup v17.2d, v16.d[1]\n"
+      "dup v18.2d, v16.d[1]\n"
+      "sub x20, x20, #0x18\n"
+      "ldr q17, [x25], #0x10\n"
       "dup v16.2d, v16.d[0]\n"
-      "str q19, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "mov v17.d[1], v18.d[0]\n"
-      "dup v16.2d, v18.d[1]\n"
-      "str q17, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "str q19, [x21, #0x0]\n"
+      "cmp x20, #0x18\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "mov v18.d[1], v17.d[0]\n"
+      "dup v16.2d, v17.d[1]\n"
+      "str q18, [x21, #0x0]\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
       "ldr q17, [x25], #0x10\n"
-      "sub x20, x20, #0xc\n"
       "ldr d16, [x25], #0x8\n"
+      "sub x20, x20, #0xc\n"
       "cmp x20, #0xc\n"
-      "str q17, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q17, [x21, #0x0]\n"
+      "str d16, [x21, #0x10]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str d16, [x23, #0x10]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr d16, [x25], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
-      "str d16, [x23, #0x0]\n"
-      "add x23, x23, #0x8\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr h16, [x25], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
-      "str h16, [x23, #0x0]\n"
-      "add x23, x23, #0x2\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str h16, [x21, #0x0]\n"
+      "add x21, x21, #0x2\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x18\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp
index 3e0ab6d955..d8edd806eb 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp
@@ -40,675 +40,634 @@ void a64_transpose_interleave_24_2x4_fp32bf16(bfloat16 *out, const float *in, si
 
     __asm__ __volatile__(
       "cmp %x[height], #0x8\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x9, %x[in]\n"
-      "mov x28, %x[width]\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x9, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "cmp x28, #0x18\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
-      "ldr q28, [x9], #0x10\n"
-      "ldr q17, [x26], #0x10\n"
-      "sub x28, x28, #0x18\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q15, [x24], #0x10\n"
-      "cmp x28, #0x18\n"
-      "ldr q20, [x23], #0x10\n"
-      "ldr q30, [x22], #0x10\n"
-      "ldr q9, [x21], #0x10\n"
-      "ldr q14, [x20], #0x10\n"
+      "ldr q15, [x9], #0x10\n"
+      "ldr q1, [x28], #0x10\n"
+      "sub x24, x24, #0x18\n"
+      "cmp x24, #0x18\n"
+      "ldr q0, [x27], #0x10\n"
+      "ldr q27, [x26], #0x10\n"
+      "zip1 v18.4s, v15.4s, v0.4s\n"
+      "zip1 v20.4s, v1.4s, v27.4s\n"
+      "ldr q13, [x25], #0x10\n"
+      "ldr q14, [x23], #0x10\n"
+      "zip2 v16.4s, v15.4s, v0.4s\n"
+      "zip2 v3.4s, v1.4s, v27.4s\n"
+      "ldr q12, [x22], #0x10\n"
+      "ldr q11, [x20], #0x10\n"
+      "zip1 v4.4s, v13.4s, v12.4s\n"
+      "zip1 v28.4s, v14.4s, v11.4s\n"
+      "ldr q5, [x9], #0x10\n"
+      "ldr q30, [x28], #0x10\n"
+      "zip2 v23.4s, v13.4s, v12.4s\n"
+      "zip2 v19.4s, v14.4s, v11.4s\n"
+      "ldr q25, [x27], #0x10\n"
+      "ldr q11, [x26], #0x10\n"
+      "zip1 v21.4s, v5.4s, v25.4s\n"
+      "zip1 v14.4s, v30.4s, v11.4s\n"
+      "ldr q6, [x25], #0x10\n"
+      "ldr q27, [x23], #0x10\n"
+      "zip2 v29.4s, v5.4s, v25.4s\n"
+      "zip2 v17.4s, v30.4s, v11.4s\n"
+      "ldr q2, [x22], #0x10\n"
+      "ldr q10, [x20], #0x10\n"
+      "zip1 v11.4s, v6.4s, v2.4s\n"
+      "zip1 v1.4s, v27.4s, v10.4s\n"
+      "ldr q8, [x9], #0x10\n"
+      "ldr q5, [x28], #0x10\n"
+      "zip2 v24.4s, v6.4s, v2.4s\n"
+      "zip2 v0.4s, v27.4s, v10.4s\n"
+      "ldr q6, [x27], #0x10\n"
+      "ldr q31, [x26], #0x10\n"
+      "zip1 v12.4s, v8.4s, v6.4s\n"
+      "zip1 v10.4s, v5.4s, v31.4s\n"
+      "ldr q30, [x25], #0x10\n"
+      "ldr q2, [x23], #0x10\n"
+      "zip2 v9.4s, v8.4s, v6.4s\n"
+      "zip2 v13.4s, v5.4s, v31.4s\n"
+      "ldr q7, [x22], #0x10\n"
+      "ldr q8, [x20], #0x10\n"
+      "zip1 v27.4s, v30.4s, v7.4s\n"
+      "zip1 v31.4s, v2.4s, v8.4s\n"
+      "ldr q5, [x9], #0x10\n"
+      "ldr q26, [x28], #0x10\n"
+      "zip2 v22.4s, v30.4s, v7.4s\n"
+      "zip2 v8.4s, v2.4s, v8.4s\n"
+      "ldr q2, [x27], #0x10\n"
+      "ldr q6, [x26], #0x10\n"
+      "zip1 v25.4s, v5.4s, v2.4s\n"
+      "zip1 v15.4s, v26.4s, v6.4s\n"
+      "ldr q7, [x25], #0x10\n"
+      "ldr q30, [x23], #0x10\n"
+      "zip2 v5.4s, v5.4s, v2.4s\n"
+      "zip2 v26.4s, v26.4s, v6.4s\n"
+      "ldr q2, [x22], #0x10\n"
+      "zip1 v6.4s, v7.4s, v2.4s\n"
+      "zip2 v7.4s, v7.4s, v2.4s\n"
+      "zip1 v2.4s, v18.4s, v20.4s\n"
+      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
+      "zip2 v20.4s, v18.4s, v20.4s\n"
+      "ldr q18, [x20], #0x10\n"
+      ".inst 0x4ea16a82  // bfcvtn2 v2.8h, v20.4s\n"
+      "zip1 v20.4s, v30.4s, v18.4s\n"
+      "zip2 v18.4s, v30.4s, v18.4s\n"
+      "zip1 v30.4s, v16.4s, v3.4s\n"
+      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
+      "zip2 v3.4s, v16.4s, v3.4s\n"
       "ldr q16, [x9], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
-      "zip1 v27.4s, v28.4s, v19.4s\n"
-      "zip1 v7.4s, v17.4s, v15.4s\n"
-      "ldr q12, [x25], #0x10\n"
-      "ldr q24, [x24], #0x10\n"
-      "zip2 v6.4s, v28.4s, v19.4s\n"
-      "zip2 v26.4s, v17.4s, v15.4s\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q23, [x22], #0x10\n"
-      "zip1 v25.4s, v20.4s, v9.4s\n"
-      "zip1 v11.4s, v30.4s, v14.4s\n"
-      "ldr q2, [x21], #0x10\n"
-      "ldr q0, [x20], #0x10\n"
-      "zip2 v28.4s, v20.4s, v9.4s\n"
-      "zip2 v31.4s, v30.4s, v14.4s\n"
-      "ldr q13, [x9], #0x10\n"
-      "ldr q10, [x26], #0x10\n"
-      "zip1 v14.4s, v16.4s, v12.4s\n"
-      "zip1 v15.4s, v22.4s, v24.4s\n"
-      "ldr q1, [x25], #0x10\n"
-      "ldr q30, [x24], #0x10\n"
-      "zip2 v3.4s, v16.4s, v12.4s\n"
-      "zip2 v20.4s, v22.4s, v24.4s\n"
-      "ldr q17, [x23], #0x10\n"
-      "ldr q9, [x22], #0x10\n"
-      "zip1 v21.4s, v18.4s, v2.4s\n"
-      "zip1 v22.4s, v23.4s, v0.4s\n"
-      "ldr q5, [x21], #0x10\n"
-      "ldr q4, [x20], #0x10\n"
-      "zip2 v16.4s, v18.4s, v2.4s\n"
-      "zip2 v12.4s, v23.4s, v0.4s\n"
-      "ldr q23, [x9], #0x10\n"
-      "ldr q24, [x26], #0x10\n"
-      "zip1 v2.4s, v13.4s, v1.4s\n"
-      "zip1 v18.4s, v10.4s, v30.4s\n"
-      "ldr q29, [x25], #0x10\n"
-      "ldr q0, [x24], #0x10\n"
-      "zip2 v1.4s, v13.4s, v1.4s\n"
-      "zip2 v30.4s, v10.4s, v30.4s\n"
-      "ldr q13, [x23], #0x10\n"
-      "ldr q10, [x22], #0x10\n"
-      "zip1 v8.4s, v17.4s, v5.4s\n"
-      "zip1 v19.4s, v9.4s, v4.4s\n"
-      "zip2 v5.4s, v17.4s, v5.4s\n"
-      "ldr q17, [x21], #0x10\n"
-      "zip2 v9.4s, v9.4s, v4.4s\n"
-      "zip1 v4.4s, v23.4s, v29.4s\n"
-      "zip2 v23.4s, v23.4s, v29.4s\n"
-      "zip1 v29.4s, v24.4s, v0.4s\n"
-      "zip2 v24.4s, v24.4s, v0.4s\n"
-      "zip1 v0.4s, v13.4s, v17.4s\n"
-      "zip2 v17.4s, v13.4s, v17.4s\n"
-      "zip1 v13.4s, v27.4s, v7.4s\n"
-      "zip2 v7.4s, v27.4s, v7.4s\n"
-      "ldr q27, [x20], #0x10\n"
+      ".inst 0x4ea1687e  // bfcvtn2 v30.8h, v3.4s\n"
+      "zip1 v3.4s, v21.4s, v14.4s\n"
+      ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
+      "zip2 v21.4s, v21.4s, v14.4s\n"
+      "ldr q14, [x28], #0x10\n"
+      ".inst 0x4ea16aa3  // bfcvtn2 v3.8h, v21.4s\n"
+      "zip1 v21.4s, v29.4s, v17.4s\n"
+      ".inst 0x0ea16ab5  // bfcvtn v21.4h, v21.4s\n"
+      "zip2 v29.4s, v29.4s, v17.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      ".inst 0x4ea16bb5  // bfcvtn2 v21.8h, v29.4s\n"
+      "zip1 v29.4s, v16.4s, v17.4s\n"
+      "zip2 v16.4s, v16.4s, v17.4s\n"
+      "zip1 v17.4s, v12.4s, v10.4s\n"
+      ".inst 0x0ea16a31  // bfcvtn v17.4h, v17.4s\n"
+      "zip2 v10.4s, v12.4s, v10.4s\n"
+      "ldr q12, [x26], #0x10\n"
+      ".inst 0x4ea16951  // bfcvtn2 v17.8h, v10.4s\n"
+      "zip1 v10.4s, v14.4s, v12.4s\n"
+      "zip2 v14.4s, v14.4s, v12.4s\n"
+      "zip1 v12.4s, v9.4s, v13.4s\n"
+      ".inst 0x0ea1698c  // bfcvtn v12.4h, v12.4s\n"
+      "zip2 v13.4s, v9.4s, v13.4s\n"
+      "ldr q9, [x25], #0x10\n"
+      ".inst 0x4ea169ac  // bfcvtn2 v12.8h, v13.4s\n"
+      "zip1 v13.4s, v25.4s, v15.4s\n"
       ".inst 0x0ea169ad  // bfcvtn v13.4h, v13.4s\n"
-      ".inst 0x4ea168ed  // bfcvtn2 v13.8h, v7.4s\n"
-      "zip1 v7.4s, v6.4s, v26.4s\n"
-      "zip2 v26.4s, v6.4s, v26.4s\n"
-      "ldr q6, [x9], #0x10\n"
-      ".inst 0x0ea168e7  // bfcvtn v7.4h, v7.4s\n"
-      ".inst 0x4ea16b47  // bfcvtn2 v7.8h, v26.4s\n"
-      "zip1 v26.4s, v10.4s, v27.4s\n"
-      "zip2 v27.4s, v10.4s, v27.4s\n"
-      "zip1 v10.4s, v14.4s, v15.4s\n"
-      "zip2 v14.4s, v14.4s, v15.4s\n"
-      "ldr q15, [x26], #0x10\n"
-      ".inst 0x0ea1694a  // bfcvtn v10.4h, v10.4s\n"
-      ".inst 0x4ea169ca  // bfcvtn2 v10.8h, v14.4s\n"
-      "zip1 v14.4s, v3.4s, v20.4s\n"
-      "zip2 v20.4s, v3.4s, v20.4s\n"
-      "ldr q3, [x25], #0x10\n"
+      "zip2 v25.4s, v25.4s, v15.4s\n"
+      "ldr q15, [x23], #0x10\n"
+      ".inst 0x4ea16b2d  // bfcvtn2 v13.8h, v25.4s\n"
+      "zip1 v25.4s, v5.4s, v26.4s\n"
+      ".inst 0x0ea16b39  // bfcvtn v25.4h, v25.4s\n"
+      "zip2 v5.4s, v5.4s, v26.4s\n"
+      "ldr q26, [x22], #0x10\n"
+      ".inst 0x4ea168b9  // bfcvtn2 v25.8h, v5.4s\n"
+      "zip1 v5.4s, v9.4s, v26.4s\n"
+      "zip2 v9.4s, v9.4s, v26.4s\n"
+      "zip1 v26.4s, v29.4s, v10.4s\n"
+      ".inst 0x0ea16b5a  // bfcvtn v26.4h, v26.4s\n"
+      "zip2 v10.4s, v29.4s, v10.4s\n"
+      "ldr q29, [x20], #0x10\n"
+      ".inst 0x4ea1695a  // bfcvtn2 v26.8h, v10.4s\n"
+      "zip1 v10.4s, v15.4s, v29.4s\n"
+      "zip2 v15.4s, v15.4s, v29.4s\n"
+      "zip1 v29.4s, v16.4s, v14.4s\n"
+      ".inst 0x0ea16bbd  // bfcvtn v29.4h, v29.4s\n"
+      "zip2 v14.4s, v16.4s, v14.4s\n"
+      "ldr q16, [x9], #0x10\n"
+      ".inst 0x4ea169dd  // bfcvtn2 v29.8h, v14.4s\n"
+      "zip1 v14.4s, v4.4s, v28.4s\n"
       ".inst 0x0ea169ce  // bfcvtn v14.4h, v14.4s\n"
-      ".inst 0x4ea16a8e  // bfcvtn2 v14.8h, v20.4s\n"
-      "zip1 v20.4s, v2.4s, v18.4s\n"
-      "zip2 v18.4s, v2.4s, v18.4s\n"
-      "ldr q2, [x24], #0x10\n"
-      ".inst 0x0ea16a94  // bfcvtn v20.4h, v20.4s\n"
-      ".inst 0x4ea16a54  // bfcvtn2 v20.8h, v18.4s\n"
-      "zip1 v18.4s, v6.4s, v3.4s\n"
-      "zip2 v3.4s, v6.4s, v3.4s\n"
-      "zip1 v6.4s, v15.4s, v2.4s\n"
-      "zip2 v15.4s, v15.4s, v2.4s\n"
-      "zip1 v2.4s, v1.4s, v30.4s\n"
-      "zip2 v30.4s, v1.4s, v30.4s\n"
-      "ldr q1, [x23], #0x10\n"
-      ".inst 0x0ea16842  // bfcvtn v2.4h, v2.4s\n"
-      ".inst 0x4ea16bc2  // bfcvtn2 v2.8h, v30.4s\n"
-      "zip1 v30.4s, v4.4s, v29.4s\n"
-      "zip2 v4.4s, v4.4s, v29.4s\n"
-      "ldr q29, [x22], #0x10\n"
-      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
-      ".inst 0x4ea1689e  // bfcvtn2 v30.8h, v4.4s\n"
-      "zip1 v4.4s, v23.4s, v24.4s\n"
-      "zip2 v24.4s, v23.4s, v24.4s\n"
-      "ldr q23, [x21], #0x10\n"
+      "zip2 v4.4s, v4.4s, v28.4s\n"
+      "ldr q28, [x28], #0x10\n"
+      ".inst 0x4ea1688e  // bfcvtn2 v14.8h, v4.4s\n"
+      "zip1 v4.4s, v23.4s, v19.4s\n"
       ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
-      ".inst 0x4ea16b04  // bfcvtn2 v4.8h, v24.4s\n"
-      "zip1 v24.4s, v18.4s, v6.4s\n"
-      "zip2 v18.4s, v18.4s, v6.4s\n"
-      "ldr q6, [x20], #0x10\n"
-      ".inst 0x0ea16b18  // bfcvtn v24.4h, v24.4s\n"
-      ".inst 0x4ea16a58  // bfcvtn2 v24.8h, v18.4s\n"
-      "zip1 v18.4s, v1.4s, v23.4s\n"
-      "zip2 v1.4s, v1.4s, v23.4s\n"
-      "zip1 v23.4s, v29.4s, v6.4s\n"
-      "zip2 v29.4s, v29.4s, v6.4s\n"
-      "zip1 v6.4s, v3.4s, v15.4s\n"
-      "zip2 v3.4s, v3.4s, v15.4s\n"
-      "ldr q15, [x9], #0x10\n"
-      ".inst 0x0ea168c6  // bfcvtn v6.4h, v6.4s\n"
-      ".inst 0x4ea16866  // bfcvtn2 v6.8h, v3.4s\n"
-      "zip1 v3.4s, v25.4s, v11.4s\n"
-      "zip2 v11.4s, v25.4s, v11.4s\n"
-      "ldr q25, [x26], #0x10\n"
-      ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
-      ".inst 0x4ea16963  // bfcvtn2 v3.8h, v11.4s\n"
-      "zip1 v11.4s, v28.4s, v31.4s\n"
-      "zip2 v28.4s, v28.4s, v31.4s\n"
-      "ldr q31, [x25], #0x10\n"
+      "zip2 v19.4s, v23.4s, v19.4s\n"
+      "ldr q23, [x27], #0x10\n"
+      ".inst 0x4ea16a64  // bfcvtn2 v4.8h, v19.4s\n"
+      "zip1 v19.4s, v16.4s, v23.4s\n"
+      "zip2 v16.4s, v16.4s, v23.4s\n"
+      "zip1 v23.4s, v11.4s, v1.4s\n"
+      ".inst 0x0ea16af7  // bfcvtn v23.4h, v23.4s\n"
+      "zip2 v1.4s, v11.4s, v1.4s\n"
+      "ldr q11, [x26], #0x10\n"
+      ".inst 0x4ea16837  // bfcvtn2 v23.8h, v1.4s\n"
+      "zip1 v1.4s, v28.4s, v11.4s\n"
+      "zip2 v28.4s, v28.4s, v11.4s\n"
+      "zip1 v11.4s, v19.4s, v1.4s\n"
       ".inst 0x0ea1696b  // bfcvtn v11.4h, v11.4s\n"
-      ".inst 0x4ea16b8b  // bfcvtn2 v11.8h, v28.4s\n"
-      "zip1 v28.4s, v21.4s, v22.4s\n"
-      "zip2 v21.4s, v21.4s, v22.4s\n"
-      "ldr q22, [x24], #0x10\n"
-      ".inst 0x0ea16b9c  // bfcvtn v28.4h, v28.4s\n"
-      ".inst 0x4ea16abc  // bfcvtn2 v28.8h, v21.4s\n"
-      "zip1 v21.4s, v15.4s, v31.4s\n"
-      "zip2 v31.4s, v15.4s, v31.4s\n"
-      "zip1 v15.4s, v25.4s, v22.4s\n"
-      "zip2 v22.4s, v25.4s, v22.4s\n"
-      "zip1 v25.4s, v16.4s, v12.4s\n"
-      "zip2 v16.4s, v16.4s, v12.4s\n"
-      "ldr q12, [x23], #0x10\n"
-      ".inst 0x0ea16b39  // bfcvtn v25.4h, v25.4s\n"
-      ".inst 0x4ea16a19  // bfcvtn2 v25.8h, v16.4s\n"
-      "zip1 v16.4s, v21.4s, v15.4s\n"
-      "zip2 v21.4s, v21.4s, v15.4s\n"
-      "ldr q15, [x22], #0x10\n"
+      "zip2 v19.4s, v19.4s, v1.4s\n"
+      "ldr q1, [x25], #0x10\n"
+      ".inst 0x4ea16a6b  // bfcvtn2 v11.8h, v19.4s\n"
+      "zip1 v19.4s, v16.4s, v28.4s\n"
+      ".inst 0x0ea16a73  // bfcvtn v19.4h, v19.4s\n"
+      "zip2 v16.4s, v16.4s, v28.4s\n"
+      "ldr q28, [x23], #0x10\n"
+      ".inst 0x4ea16a13  // bfcvtn2 v19.8h, v16.4s\n"
+      "zip1 v16.4s, v24.4s, v0.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      ".inst 0x4ea16ab0  // bfcvtn2 v16.8h, v21.4s\n"
-      "zip1 v21.4s, v31.4s, v22.4s\n"
-      "zip2 v22.4s, v31.4s, v22.4s\n"
-      "ldr q31, [x21], #0x10\n"
-      ".inst 0x0ea16ab5  // bfcvtn v21.4h, v21.4s\n"
-      ".inst 0x4ea16ad5  // bfcvtn2 v21.8h, v22.4s\n"
-      "ldr q22, [x20], #0x10\n"
-      "str q13, [x27, #0x0]\n"
-      "zip1 v13.4s, v8.4s, v19.4s\n"
-      "zip2 v19.4s, v8.4s, v19.4s\n"
-      "str q7, [x27, #0x10]\n"
-      "zip1 v8.4s, v12.4s, v31.4s\n"
-      "zip2 v12.4s, v12.4s, v31.4s\n"
-      "str q10, [x27, #0x20]\n"
-      "zip1 v7.4s, v15.4s, v22.4s\n"
-      "zip2 v15.4s, v15.4s, v22.4s\n"
-      "str q14, [x27, #0x30]\n"
-      "zip1 v14.4s, v5.4s, v9.4s\n"
-      "zip1 v22.4s, v0.4s, v26.4s\n"
-      "str q20, [x27, #0x40]\n"
-      "zip1 v31.4s, v17.4s, v27.4s\n"
-      "zip1 v10.4s, v18.4s, v23.4s\n"
-      "str q2, [x27, #0x50]\n"
-      "zip1 v20.4s, v1.4s, v29.4s\n"
-      "zip1 v2.4s, v8.4s, v7.4s\n"
-      "str q30, [x27, #0x60]\n"
-      "zip1 v30.4s, v12.4s, v15.4s\n"
-      ".inst 0x0ea169ad  // bfcvtn v13.4h, v13.4s\n"
-      "str q4, [x27, #0x70]\n"
-      ".inst 0x0ea169c4  // bfcvtn v4.4h, v14.4s\n"
-      "zip2 v9.4s, v5.4s, v9.4s\n"
-      "str q24, [x27, #0x80]\n"
-      ".inst 0x0ea16ad8  // bfcvtn v24.4h, v22.4s\n"
-      "zip2 v26.4s, v0.4s, v26.4s\n"
-      "str q6, [x27, #0x90]\n"
-      ".inst 0x0ea16bee  // bfcvtn v14.4h, v31.4s\n"
-      "zip2 v5.4s, v17.4s, v27.4s\n"
-      "str q16, [x27, #0xa0]\n"
-      ".inst 0x0ea1695b  // bfcvtn v27.4h, v10.4s\n"
-      "zip2 v17.4s, v18.4s, v23.4s\n"
-      "str q21, [x27, #0xb0]\n"
-      ".inst 0x0ea16a9f  // bfcvtn v31.4h, v20.4s\n"
-      "zip2 v23.4s, v1.4s, v29.4s\n"
-      "str q3, [x27, #0xc0]\n"
-      ".inst 0x0ea16852  // bfcvtn v18.4h, v2.4s\n"
-      "zip2 v20.4s, v8.4s, v7.4s\n"
-      "str q11, [x27, #0xd0]\n"
-      ".inst 0x0ea16bd0  // bfcvtn v16.4h, v30.4s\n"
-      "zip2 v15.4s, v12.4s, v15.4s\n"
-      "str q28, [x27, #0xe0]\n"
-      ".inst 0x4ea16a6d  // bfcvtn2 v13.8h, v19.4s\n"
-      ".inst 0x4ea16924  // bfcvtn2 v4.8h, v9.4s\n"
-      "str q25, [x27, #0xf0]\n"
-      ".inst 0x4ea16b58  // bfcvtn2 v24.8h, v26.4s\n"
-      ".inst 0x4ea168ae  // bfcvtn2 v14.8h, v5.4s\n"
-      ".inst 0x4ea16a3b  // bfcvtn2 v27.8h, v17.4s\n"
-      ".inst 0x4ea16aff  // bfcvtn2 v31.8h, v23.4s\n"
-      ".inst 0x4ea16a92  // bfcvtn2 v18.8h, v20.4s\n"
-      ".inst 0x4ea169f0  // bfcvtn2 v16.8h, v15.4s\n"
-      "str q13, [x27, #0x100]\n"
-      "str q4, [x27, #0x110]\n"
-      "str q24, [x27, #0x120]\n"
-      "str q14, [x27, #0x130]\n"
-      "str q27, [x27, #0x140]\n"
-      "str q31, [x27, #0x150]\n"
-      "str q18, [x27, #0x160]\n"
-      "str q16, [x27, #0x170]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip2 v24.4s, v24.4s, v0.4s\n"
+      "ldr q0, [x22], #0x10\n"
+      ".inst 0x4ea16b10  // bfcvtn2 v16.8h, v24.4s\n"
+      "ldr q24, [x20], #0x10\n"
+      "str q2, [x21, #0x0]\n"
+      "zip1 v2.4s, v1.4s, v0.4s\n"
+      "zip2 v0.4s, v1.4s, v0.4s\n"
+      "zip1 v1.4s, v28.4s, v24.4s\n"
+      "zip2 v28.4s, v28.4s, v24.4s\n"
+      "str q30, [x21, #0x10]\n"
+      "zip1 v24.4s, v27.4s, v31.4s\n"
+      "zip1 v30.4s, v22.4s, v8.4s\n"
+      "str q3, [x21, #0x20]\n"
+      "zip1 v3.4s, v6.4s, v20.4s\n"
+      ".inst 0x0ea16b18  // bfcvtn v24.4h, v24.4s\n"
+      "str q21, [x21, #0x30]\n"
+      "zip1 v21.4s, v7.4s, v18.4s\n"
+      "zip2 v31.4s, v27.4s, v31.4s\n"
+      "str q17, [x21, #0x40]\n"
+      "zip1 v17.4s, v5.4s, v10.4s\n"
+      "zip1 v27.4s, v9.4s, v15.4s\n"
+      "str q12, [x21, #0x50]\n"
+      "zip1 v12.4s, v2.4s, v1.4s\n"
+      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
+      "str q13, [x21, #0x60]\n"
+      "zip1 v13.4s, v0.4s, v28.4s\n"
+      "zip2 v22.4s, v22.4s, v8.4s\n"
+      "str q25, [x21, #0x70]\n"
+      ".inst 0x0ea16879  // bfcvtn v25.4h, v3.4s\n"
+      "zip2 v8.4s, v6.4s, v20.4s\n"
+      "str q26, [x21, #0x80]\n"
+      ".inst 0x0ea16aa3  // bfcvtn v3.4h, v21.4s\n"
+      "zip2 v18.4s, v7.4s, v18.4s\n"
+      "str q29, [x21, #0x90]\n"
+      ".inst 0x0ea16a27  // bfcvtn v7.4h, v17.4s\n"
+      "zip2 v21.4s, v5.4s, v10.4s\n"
+      "str q11, [x21, #0xa0]\n"
+      ".inst 0x0ea16b65  // bfcvtn v5.4h, v27.4s\n"
+      "zip2 v15.4s, v9.4s, v15.4s\n"
+      "str q19, [x21, #0xb0]\n"
+      ".inst 0x0ea16991  // bfcvtn v17.4h, v12.4s\n"
+      "zip2 v20.4s, v2.4s, v1.4s\n"
+      "str q14, [x21, #0xc0]\n"
+      ".inst 0x0ea169bb  // bfcvtn v27.4h, v13.4s\n"
+      "zip2 v29.4s, v0.4s, v28.4s\n"
+      "str q4, [x21, #0xd0]\n"
+      ".inst 0x4ea16bf8  // bfcvtn2 v24.8h, v31.4s\n"
+      ".inst 0x4ea16ade  // bfcvtn2 v30.8h, v22.4s\n"
+      "str q23, [x21, #0xe0]\n"
+      ".inst 0x4ea16919  // bfcvtn2 v25.8h, v8.4s\n"
+      ".inst 0x4ea16a43  // bfcvtn2 v3.8h, v18.4s\n"
+      "str q16, [x21, #0xf0]\n"
+      ".inst 0x4ea16aa7  // bfcvtn2 v7.8h, v21.4s\n"
+      ".inst 0x4ea169e5  // bfcvtn2 v5.8h, v15.4s\n"
+      "str q24, [x21, #0x100]\n"
+      ".inst 0x4ea16a91  // bfcvtn2 v17.8h, v20.4s\n"
+      ".inst 0x4ea16bbb  // bfcvtn2 v27.8h, v29.4s\n"
+      "str q30, [x21, #0x110]\n"
+      "str q25, [x21, #0x120]\n"
+      "str q3, [x21, #0x130]\n"
+      "str q7, [x21, #0x140]\n"
+      "str q5, [x21, #0x150]\n"
+      "str q17, [x21, #0x160]\n"
+      "str q27, [x21, #0x170]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x28, 10f\n"
-      "cmp x28, #0x10\n"
-      "movi v6.16b, #0x0\n"
-      "str q6, [x27, #0x0]\n"
-      "str q6, [x27, #0x10]\n"
-      "str q6, [x27, #0x20]\n"
-      "str q6, [x27, #0x30]\n"
-      "str q6, [x27, #0x40]\n"
-      "str q6, [x27, #0x50]\n"
-      "str q6, [x27, #0x60]\n"
-      "str q6, [x27, #0x70]\n"
-      "str q6, [x27, #0x80]\n"
-      "str q6, [x27, #0x90]\n"
-      "str q6, [x27, #0xa0]\n"
-      "str q6, [x27, #0xb0]\n"
-      "str q6, [x27, #0xc0]\n"
-      "str q6, [x27, #0xd0]\n"
-      "str q6, [x27, #0xe0]\n"
-      "str q6, [x27, #0xf0]\n"
-      "str q6, [x27, #0x100]\n"
-      "str q6, [x27, #0x110]\n"
-      "str q6, [x27, #0x120]\n"
-      "str q6, [x27, #0x130]\n"
-      "str q6, [x27, #0x140]\n"
-      "str q6, [x27, #0x150]\n"
-      "str q6, [x27, #0x160]\n"
-      "str q6, [x27, #0x170]\n"
+      "cmp x24, #0x10\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 16 loop: loop
-      "ldr q19, [x9], #0x10\n"
-      "ldr q24, [x26], #0x10\n"
-      "sub x28, x28, #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "ldr q21, [x24], #0x10\n"
-      "cmp x28, #0x10\n"
-      "ldr q15, [x23], #0x10\n"
-      "ldr q20, [x22], #0x10\n"
-      "ldr q30, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "ldr q6, [x9], #0x10\n"
-      "ldr q10, [x26], #0x10\n"
-      "zip1 v4.4s, v19.4s, v18.4s\n"
-      "zip1 v17.4s, v24.4s, v21.4s\n"
-      "ldr q13, [x25], #0x10\n"
-      "ldr q31, [x24], #0x10\n"
-      "zip2 v29.4s, v19.4s, v18.4s\n"
-      "zip2 v26.4s, v24.4s, v21.4s\n"
-      "ldr q21, [x23], #0x10\n"
-      "ldr q25, [x22], #0x10\n"
-      "zip1 v18.4s, v15.4s, v30.4s\n"
-      "zip1 v7.4s, v20.4s, v16.4s\n"
-      "ldr q19, [x21], #0x10\n"
+      "ldr q9, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "sub x24, x24, #0x10\n"
+      "cmp x24, #0x10\n"
+      "ldr q15, [x27], #0x10\n"
+      "ldr q17, [x26], #0x10\n"
+      "zip1 v14.4s, v9.4s, v15.4s\n"
+      "zip1 v11.4s, v18.4s, v17.4s\n"
+      "ldr q7, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip2 v12.4s, v9.4s, v15.4s\n"
+      "zip2 v6.4s, v18.4s, v17.4s\n"
+      "ldr q15, [x22], #0x10\n"
       "ldr q3, [x20], #0x10\n"
-      "zip2 v24.4s, v15.4s, v30.4s\n"
-      "zip2 v15.4s, v20.4s, v16.4s\n"
-      "ldr q27, [x9], #0x10\n"
-      "ldr q9, [x26], #0x10\n"
-      "zip1 v2.4s, v6.4s, v13.4s\n"
-      "zip1 v30.4s, v10.4s, v31.4s\n"
-      "ldr q0, [x25], #0x10\n"
-      "ldr q11, [x24], #0x10\n"
-      "zip2 v28.4s, v6.4s, v13.4s\n"
-      "zip2 v22.4s, v10.4s, v31.4s\n"
-      "ldr q8, [x23], #0x10\n"
-      "ldr q1, [x22], #0x10\n"
-      "zip1 v23.4s, v21.4s, v19.4s\n"
-      "zip1 v6.4s, v25.4s, v3.4s\n"
-      "ldr q31, [x21], #0x10\n"
-      "ldr q5, [x20], #0x10\n"
-      "zip2 v20.4s, v21.4s, v19.4s\n"
-      "zip2 v21.4s, v25.4s, v3.4s\n"
-      "ldr q16, [x9], #0x10\n"
+      "zip1 v30.4s, v7.4s, v15.4s\n"
+      "zip1 v20.4s, v16.4s, v3.4s\n"
+      "ldr q17, [x9], #0x10\n"
+      "ldr q9, [x28], #0x10\n"
+      "zip2 v1.4s, v7.4s, v15.4s\n"
+      "zip2 v24.4s, v16.4s, v3.4s\n"
+      "ldr q10, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v0.4s, v17.4s, v10.4s\n"
+      "zip1 v8.4s, v9.4s, v16.4s\n"
+      "ldr q7, [x25], #0x10\n"
+      "ldr q2, [x23], #0x10\n"
+      "zip2 v17.4s, v17.4s, v10.4s\n"
+      "zip2 v3.4s, v9.4s, v16.4s\n"
+      "ldr q9, [x22], #0x10\n"
+      "ldr q10, [x20], #0x10\n"
+      "zip1 v25.4s, v7.4s, v9.4s\n"
+      "zip1 v23.4s, v2.4s, v10.4s\n"
+      "ldr q31, [x9], #0x10\n"
+      "ldr q21, [x28], #0x10\n"
+      "zip2 v16.4s, v7.4s, v9.4s\n"
+      "zip2 v27.4s, v2.4s, v10.4s\n"
+      "ldr q26, [x27], #0x10\n"
+      "ldr q19, [x26], #0x10\n"
+      "zip1 v2.4s, v31.4s, v26.4s\n"
+      "zip1 v7.4s, v21.4s, v19.4s\n"
+      "ldr q29, [x25], #0x10\n"
+      "ldr q13, [x23], #0x10\n"
+      "zip2 v31.4s, v31.4s, v26.4s\n"
+      "zip2 v19.4s, v21.4s, v19.4s\n"
+      "ldr q4, [x22], #0x10\n"
+      "ldr q18, [x20], #0x10\n"
+      "zip1 v26.4s, v29.4s, v4.4s\n"
+      "zip1 v15.4s, v13.4s, v18.4s\n"
+      "ldr q9, [x9], #0x10\n"
+      "ldr q22, [x28], #0x10\n"
+      "zip2 v4.4s, v29.4s, v4.4s\n"
+      "zip2 v18.4s, v13.4s, v18.4s\n"
+      "ldr q29, [x27], #0x10\n"
       "ldr q10, [x26], #0x10\n"
-      "zip1 v14.4s, v27.4s, v0.4s\n"
-      "zip1 v3.4s, v9.4s, v11.4s\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q13, [x24], #0x10\n"
-      "zip2 v25.4s, v27.4s, v0.4s\n"
-      "zip2 v12.4s, v9.4s, v11.4s\n"
-      "ldr q11, [x23], #0x10\n"
-      "ldr q0, [x22], #0x10\n"
-      "zip1 v27.4s, v8.4s, v31.4s\n"
-      "zip1 v9.4s, v1.4s, v5.4s\n"
-      "zip2 v31.4s, v8.4s, v31.4s\n"
-      "ldr q8, [x21], #0x10\n"
-      "zip2 v5.4s, v1.4s, v5.4s\n"
-      "zip1 v1.4s, v16.4s, v19.4s\n"
-      "zip2 v19.4s, v16.4s, v19.4s\n"
-      "zip1 v16.4s, v10.4s, v13.4s\n"
-      "zip2 v13.4s, v10.4s, v13.4s\n"
-      "zip1 v10.4s, v11.4s, v8.4s\n"
-      "zip2 v11.4s, v11.4s, v8.4s\n"
-      "zip1 v8.4s, v4.4s, v17.4s\n"
-      "zip2 v17.4s, v4.4s, v17.4s\n"
-      "ldr q4, [x20], #0x10\n"
+      "zip1 v21.4s, v9.4s, v29.4s\n"
+      "zip1 v5.4s, v22.4s, v10.4s\n"
+      "ldr q28, [x25], #0x10\n"
+      "ldr q13, [x23], #0x10\n"
+      "zip2 v29.4s, v9.4s, v29.4s\n"
+      "zip2 v9.4s, v22.4s, v10.4s\n"
+      "ldr q22, [x22], #0x10\n"
+      "zip1 v10.4s, v28.4s, v22.4s\n"
+      "zip2 v28.4s, v28.4s, v22.4s\n"
+      "zip1 v22.4s, v14.4s, v11.4s\n"
+      ".inst 0x0ea16ad6  // bfcvtn v22.4h, v22.4s\n"
+      "zip2 v11.4s, v14.4s, v11.4s\n"
+      "ldr q14, [x20], #0x10\n"
+      ".inst 0x4ea16976  // bfcvtn2 v22.8h, v11.4s\n"
+      "str q22, [x21, #0x0]\n"
+      "zip1 v22.4s, v13.4s, v14.4s\n"
+      "zip2 v14.4s, v13.4s, v14.4s\n"
+      "zip1 v13.4s, v12.4s, v6.4s\n"
+      "zip1 v11.4s, v0.4s, v8.4s\n"
+      ".inst 0x0ea169ad  // bfcvtn v13.4h, v13.4s\n"
+      "zip2 v12.4s, v12.4s, v6.4s\n"
+      "zip1 v6.4s, v17.4s, v3.4s\n"
+      ".inst 0x0ea1696b  // bfcvtn v11.4h, v11.4s\n"
+      "zip2 v0.4s, v0.4s, v8.4s\n"
+      "zip1 v8.4s, v2.4s, v7.4s\n"
+      ".inst 0x0ea168c6  // bfcvtn v6.4h, v6.4s\n"
+      "zip2 v3.4s, v17.4s, v3.4s\n"
+      "zip1 v17.4s, v31.4s, v19.4s\n"
       ".inst 0x0ea16908  // bfcvtn v8.4h, v8.4s\n"
-      ".inst 0x4ea16a28  // bfcvtn2 v8.8h, v17.4s\n"
-      "zip1 v17.4s, v29.4s, v26.4s\n"
-      "zip2 v29.4s, v29.4s, v26.4s\n"
-      "zip1 v26.4s, v0.4s, v4.4s\n"
-      "zip2 v0.4s, v0.4s, v4.4s\n"
-      "zip1 v4.4s, v2.4s, v30.4s\n"
-      "zip2 v2.4s, v2.4s, v30.4s\n"
-      "str q8, [x27, #0x0]\n"
-      "zip1 v8.4s, v28.4s, v22.4s\n"
-      "zip1 v30.4s, v14.4s, v3.4s\n"
+      "zip2 v2.4s, v2.4s, v7.4s\n"
+      "zip1 v7.4s, v21.4s, v5.4s\n"
       ".inst 0x0ea16a31  // bfcvtn v17.4h, v17.4s\n"
-      "zip2 v22.4s, v28.4s, v22.4s\n"
-      "zip1 v28.4s, v25.4s, v12.4s\n"
-      ".inst 0x0ea16884  // bfcvtn v4.4h, v4.4s\n"
-      ".inst 0x0ea16908  // bfcvtn v8.4h, v8.4s\n"
-      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
-      "zip2 v14.4s, v14.4s, v3.4s\n"
-      "zip1 v3.4s, v1.4s, v16.4s\n"
-      ".inst 0x0ea16b9c  // bfcvtn v28.4h, v28.4s\n"
-      "zip2 v25.4s, v25.4s, v12.4s\n"
-      "zip1 v12.4s, v19.4s, v13.4s\n"
-      "zip2 v16.4s, v1.4s, v16.4s\n"
-      "zip1 v1.4s, v18.4s, v7.4s\n"
-      ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
-      "zip2 v13.4s, v19.4s, v13.4s\n"
-      "zip1 v19.4s, v24.4s, v15.4s\n"
-      ".inst 0x0ea1698c  // bfcvtn v12.4h, v12.4s\n"
-      "zip2 v7.4s, v18.4s, v7.4s\n"
-      "zip1 v18.4s, v23.4s, v6.4s\n"
-      ".inst 0x0ea16821  // bfcvtn v1.4h, v1.4s\n"
+      "zip2 v31.4s, v31.4s, v19.4s\n"
+      "zip1 v19.4s, v29.4s, v9.4s\n"
+      ".inst 0x0ea168e7  // bfcvtn v7.4h, v7.4s\n"
+      "zip2 v21.4s, v21.4s, v5.4s\n"
+      "zip1 v5.4s, v30.4s, v20.4s\n"
       ".inst 0x0ea16a73  // bfcvtn v19.4h, v19.4s\n"
-      "zip2 v15.4s, v24.4s, v15.4s\n"
-      "zip1 v24.4s, v20.4s, v21.4s\n"
-      "zip2 v23.4s, v23.4s, v6.4s\n"
-      "zip1 v6.4s, v27.4s, v9.4s\n"
-      ".inst 0x0ea16a52  // bfcvtn v18.4h, v18.4s\n"
-      "zip2 v21.4s, v20.4s, v21.4s\n"
-      "zip1 v20.4s, v31.4s, v5.4s\n"
+      "zip2 v29.4s, v29.4s, v9.4s\n"
+      "zip1 v9.4s, v1.4s, v24.4s\n"
+      ".inst 0x0ea168a5  // bfcvtn v5.4h, v5.4s\n"
+      "zip2 v20.4s, v30.4s, v20.4s\n"
+      "zip1 v30.4s, v25.4s, v23.4s\n"
+      ".inst 0x0ea16929  // bfcvtn v9.4h, v9.4s\n"
+      "zip2 v1.4s, v1.4s, v24.4s\n"
+      "zip1 v24.4s, v16.4s, v27.4s\n"
+      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
+      "zip2 v23.4s, v25.4s, v23.4s\n"
+      "zip1 v25.4s, v26.4s, v15.4s\n"
       ".inst 0x0ea16b18  // bfcvtn v24.4h, v24.4s\n"
-      "zip2 v27.4s, v27.4s, v9.4s\n"
-      "zip1 v9.4s, v10.4s, v26.4s\n"
-      ".inst 0x0ea168c6  // bfcvtn v6.4h, v6.4s\n"
-      ".inst 0x0ea16a94  // bfcvtn v20.4h, v20.4s\n"
-      "zip2 v31.4s, v31.4s, v5.4s\n"
-      "zip1 v5.4s, v11.4s, v0.4s\n"
-      "zip2 v26.4s, v10.4s, v26.4s\n"
-      ".inst 0x0ea1692a  // bfcvtn v10.4h, v9.4s\n"
-      "zip2 v0.4s, v11.4s, v0.4s\n"
-      ".inst 0x4ea16bb1  // bfcvtn2 v17.8h, v29.4s\n"
-      ".inst 0x4ea16844  // bfcvtn2 v4.8h, v2.4s\n"
-      ".inst 0x0ea168a9  // bfcvtn v9.4h, v5.4s\n"
-      ".inst 0x4ea16ac8  // bfcvtn2 v8.8h, v22.4s\n"
-      ".inst 0x4ea169de  // bfcvtn2 v30.8h, v14.4s\n"
-      ".inst 0x4ea16b3c  // bfcvtn2 v28.8h, v25.4s\n"
-      ".inst 0x4ea16a03  // bfcvtn2 v3.8h, v16.4s\n"
-      ".inst 0x4ea169ac  // bfcvtn2 v12.8h, v13.4s\n"
-      "str q17, [x27, #0x10]\n"
-      ".inst 0x4ea168e1  // bfcvtn2 v1.8h, v7.4s\n"
-      ".inst 0x4ea169f3  // bfcvtn2 v19.8h, v15.4s\n"
-      "str q4, [x27, #0x20]\n"
-      ".inst 0x4ea16af2  // bfcvtn2 v18.8h, v23.4s\n"
-      ".inst 0x4ea16ab8  // bfcvtn2 v24.8h, v21.4s\n"
-      "str q8, [x27, #0x30]\n"
-      ".inst 0x4ea16b66  // bfcvtn2 v6.8h, v27.4s\n"
-      ".inst 0x4ea16bf4  // bfcvtn2 v20.8h, v31.4s\n"
-      "str q30, [x27, #0x40]\n"
-      ".inst 0x4ea16b4a  // bfcvtn2 v10.8h, v26.4s\n"
-      ".inst 0x4ea16809  // bfcvtn2 v9.8h, v0.4s\n"
-      "str q28, [x27, #0x50]\n"
-      "str q3, [x27, #0x60]\n"
-      "str q12, [x27, #0x70]\n"
-      "str q1, [x27, #0xc0]\n"
-      "str q19, [x27, #0xd0]\n"
-      "str q18, [x27, #0xe0]\n"
-      "str q24, [x27, #0xf0]\n"
-      "str q6, [x27, #0x100]\n"
-      "str q20, [x27, #0x110]\n"
-      "str q10, [x27, #0x120]\n"
-      "str q9, [x27, #0x130]\n"
-      "add x27, x27, #0x80\n"
+      "zip2 v27.4s, v16.4s, v27.4s\n"
+      "zip1 v16.4s, v4.4s, v18.4s\n"
+      ".inst 0x0ea16b39  // bfcvtn v25.4h, v25.4s\n"
+      "zip2 v15.4s, v26.4s, v15.4s\n"
+      "zip1 v26.4s, v10.4s, v22.4s\n"
+      ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
+      "zip2 v18.4s, v4.4s, v18.4s\n"
+      "zip1 v4.4s, v28.4s, v14.4s\n"
+      ".inst 0x0ea16b5a  // bfcvtn v26.4h, v26.4s\n"
+      "zip2 v10.4s, v10.4s, v22.4s\n"
+      ".inst 0x0ea16896  // bfcvtn v22.4h, v4.4s\n"
+      "zip2 v4.4s, v28.4s, v14.4s\n"
+      ".inst 0x4ea1698d  // bfcvtn2 v13.8h, v12.4s\n"
+      "str q13, [x21, #0x10]\n"
+      ".inst 0x4ea1680b  // bfcvtn2 v11.8h, v0.4s\n"
+      ".inst 0x4ea16866  // bfcvtn2 v6.8h, v3.4s\n"
+      "str q11, [x21, #0x20]\n"
+      ".inst 0x4ea16848  // bfcvtn2 v8.8h, v2.4s\n"
+      ".inst 0x4ea16bf1  // bfcvtn2 v17.8h, v31.4s\n"
+      "str q6, [x21, #0x30]\n"
+      ".inst 0x4ea16aa7  // bfcvtn2 v7.8h, v21.4s\n"
+      ".inst 0x4ea16bb3  // bfcvtn2 v19.8h, v29.4s\n"
+      "str q8, [x21, #0x40]\n"
+      ".inst 0x4ea16a85  // bfcvtn2 v5.8h, v20.4s\n"
+      ".inst 0x4ea16829  // bfcvtn2 v9.8h, v1.4s\n"
+      "str q17, [x21, #0x50]\n"
+      ".inst 0x4ea16afe  // bfcvtn2 v30.8h, v23.4s\n"
+      ".inst 0x4ea16b78  // bfcvtn2 v24.8h, v27.4s\n"
+      "str q7, [x21, #0x60]\n"
+      ".inst 0x4ea169f9  // bfcvtn2 v25.8h, v15.4s\n"
+      ".inst 0x4ea16a50  // bfcvtn2 v16.8h, v18.4s\n"
+      "str q19, [x21, #0x70]\n"
+      ".inst 0x4ea1695a  // bfcvtn2 v26.8h, v10.4s\n"
+      ".inst 0x4ea16896  // bfcvtn2 v22.8h, v4.4s\n"
+      "str q5, [x21, #0xc0]\n"
+      "str q9, [x21, #0xd0]\n"
+      "str q30, [x21, #0xe0]\n"
+      "str q24, [x21, #0xf0]\n"
+      "str q25, [x21, #0x100]\n"
+      "str q16, [x21, #0x110]\n"
+      "str q26, [x21, #0x120]\n"
+      "str q22, [x21, #0x130]\n"
+      "add x21, x21, #0x80\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 16 loop: skip
-      "cmp x28, #0x4\n"
+      "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr q25, [x9], #0x10\n"
-      "ldr q24, [x26], #0x10\n"
-      "sub x28, x28, #0x4\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "cmp x28, #0x4\n"
-      "ldr q23, [x23], #0x10\n"
-      "ldr q19, [x22], #0x10\n"
-      "ldr q18, [x21], #0x10\n"
-      "ldr q17, [x20], #0x10\n"
-      "zip1 v22.4s, v25.4s, v21.4s\n"
-      "zip1 v16.4s, v24.4s, v20.4s\n"
-      "zip2 v21.4s, v25.4s, v21.4s\n"
-      "zip2 v20.4s, v24.4s, v20.4s\n"
-      "zip1 v27.4s, v23.4s, v18.4s\n"
-      "zip1 v26.4s, v19.4s, v17.4s\n"
-      "zip2 v25.4s, v23.4s, v18.4s\n"
-      "zip2 v24.4s, v19.4s, v17.4s\n"
-      "zip1 v19.4s, v22.4s, v16.4s\n"
-      "zip1 v18.4s, v21.4s, v20.4s\n"
+      "ldr q23, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.4s, v23.4s, v17.4s\n"
+      "zip1 v21.4s, v20.4s, v16.4s\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v28.4s, v23.4s, v17.4s\n"
+      "zip2 v20.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "zip1 v27.4s, v19.4s, v17.4s\n"
+      "zip1 v26.4s, v18.4s, v16.4s\n"
+      "zip2 v25.4s, v19.4s, v17.4s\n"
+      "zip2 v24.4s, v18.4s, v16.4s\n"
+      "zip1 v19.4s, v22.4s, v21.4s\n"
+      "zip1 v18.4s, v28.4s, v20.4s\n"
       "zip1 v17.4s, v27.4s, v26.4s\n"
-      "zip2 v23.4s, v22.4s, v16.4s\n"
       "zip1 v16.4s, v25.4s, v24.4s\n"
-      "zip2 v22.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a75  // bfcvtn v21.4h, v19.4s\n"
-      ".inst 0x0ea16a54  // bfcvtn v20.4h, v18.4s\n"
+      ".inst 0x0ea16a77  // bfcvtn v23.4h, v19.4s\n"
+      "zip2 v22.4s, v22.4s, v21.4s\n"
+      ".inst 0x0ea16a55  // bfcvtn v21.4h, v18.4s\n"
+      "zip2 v20.4s, v28.4s, v20.4s\n"
       ".inst 0x0ea16a33  // bfcvtn v19.4h, v17.4s\n"
       "zip2 v18.4s, v27.4s, v26.4s\n"
       ".inst 0x0ea16a11  // bfcvtn v17.4h, v16.4s\n"
       "zip2 v16.4s, v25.4s, v24.4s\n"
-      ".inst 0x4ea16af5  // bfcvtn2 v21.8h, v23.4s\n"
-      ".inst 0x4ea16ad4  // bfcvtn2 v20.8h, v22.4s\n"
+      ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
+      ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
+      "str q23, [x21, #0x0]\n"
       ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
       ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
-      "str q21, [x27, #0x0]\n"
-      "str q20, [x27, #0x10]\n"
-      "str q19, [x27, #0xc0]\n"
-      "str q17, [x27, #0xd0]\n"
-      "add x27, x27, #0x20\n"
+      "str q21, [x21, #0x10]\n"
+      "str q19, [x21, #0xc0]\n"
+      "str q17, [x21, #0xd0]\n"
+      "add x21, x21, #0x20\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x28, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr s23, [x9], #0x4\n"
-      "ldr s22, [x26], #0x4\n"
-      "sub x28, x28, #0x1\n"
-      "ldr s19, [x25], #0x4\n"
-      "ldr s17, [x24], #0x4\n"
-      "cmp x28, #0x1\n"
-      "ldr s21, [x23], #0x4\n"
-      "ldr s20, [x22], #0x4\n"
-      "ldr s18, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
-      "zip1 v19.4s, v23.4s, v19.4s\n"
-      "zip1 v17.4s, v22.4s, v17.4s\n"
-      "zip1 v18.4s, v21.4s, v18.4s\n"
-      "zip1 v16.4s, v20.4s, v16.4s\n"
+      "ldr s19, [x9], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
       "zip1 v17.4s, v19.4s, v17.4s\n"
       "zip1 v16.4s, v18.4s, v16.4s\n"
-      ".inst 0x0ea16a31  // bfcvtn v17.4h, v17.4s\n"
+      "ldr s20, [x25], #0x4\n"
+      "ldr s19, [x23], #0x4\n"
+      "zip1 v16.4s, v17.4s, v16.4s\n"
+      ".inst 0x0ea16a12  // bfcvtn v18.4h, v16.4s\n"
+      "ldr s17, [x22], #0x4\n"
+      "ldr s16, [x20], #0x4\n"
+      "zip1 v17.4s, v20.4s, v17.4s\n"
+      "zip1 v16.4s, v19.4s, v16.4s\n"
+      "zip1 v16.4s, v17.4s, v16.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      "str d17, [x27, #0x0]\n"
-      "str d16, [x27, #0xc0]\n"
-      "add x27, x27, #0x8\n"
+      "str d18, [x21, #0x0]\n"
+      "str d16, [x21, #0xc0]\n"
+      "add x21, x21, #0x8\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x8\n"
       "add %x[out], %x[out], #0x180\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x9, %x[in]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x26, x27, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x9, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "csel x25, x25, %x[pad_row], GE\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x26, %x[in_stride]\n"
       "csel x26, x26, %x[pad_row], GT\n"
+      "csel x27, x27, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x28, x28, %x[pad_row], GT\n"
       "cmp x20, #0x18\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Column loop
-      "ldr q24, [x9], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Column loop
+      "ldr q20, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
       "sub x20, x20, #0x18\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x18\n"
-      "ldr q26, [x9], #0x10\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v9.4s, v20.4s, v17.4s\n"
+      "zip1 v30.4s, v18.4s, v16.4s\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q19, [x28], #0x10\n"
+      "zip2 v17.4s, v20.4s, v17.4s\n"
+      "zip2 v5.4s, v18.4s, v16.4s\n"
+      "ldr q18, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v0.4s, v21.4s, v18.4s\n"
+      "zip1 v3.4s, v19.4s, v16.4s\n"
+      "ldr q23, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v1.4s, v21.4s, v18.4s\n"
+      "zip2 v16.4s, v19.4s, v16.4s\n"
+      "ldr q19, [x27], #0x10\n"
+      "ldr q18, [x26], #0x10\n"
+      "zip1 v4.4s, v23.4s, v19.4s\n"
+      "zip1 v2.4s, v20.4s, v18.4s\n"
+      "ldr q22, [x9], #0x10\n"
+      "ldr q21, [x28], #0x10\n"
+      "zip2 v27.4s, v23.4s, v19.4s\n"
+      "zip2 v28.4s, v20.4s, v18.4s\n"
+      "ldr q20, [x27], #0x10\n"
+      "ldr q24, [x26], #0x10\n"
+      "zip1 v25.4s, v22.4s, v20.4s\n"
+      "zip1 v26.4s, v21.4s, v24.4s\n"
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v14.4s, v22.4s, v20.4s\n"
+      "zip2 v12.4s, v21.4s, v24.4s\n"
+      "ldr q31, [x27], #0x10\n"
+      "ldr q24, [x26], #0x10\n"
+      "zip1 v15.4s, v19.4s, v31.4s\n"
+      "zip1 v13.4s, v18.4s, v24.4s\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q11, [x28], #0x10\n"
+      "zip2 v20.4s, v19.4s, v31.4s\n"
+      "zip2 v10.4s, v18.4s, v24.4s\n"
+      "ldr q22, [x27], #0x10\n"
       "ldr q23, [x26], #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "ldr q25, [x9], #0x10\n"
-      "zip1 v6.4s, v24.4s, v19.4s\n"
-      "zip1 v11.4s, v22.4s, v17.4s\n"
-      "ldr q21, [x26], #0x10\n"
-      "ldr q20, [x25], #0x10\n"
-      "zip2 v2.4s, v24.4s, v19.4s\n"
-      "zip2 v22.4s, v22.4s, v17.4s\n"
-      "ldr q17, [x24], #0x10\n"
-      "ldr q7, [x9], #0x10\n"
-      "zip1 v19.4s, v26.4s, v18.4s\n"
-      "zip1 v5.4s, v23.4s, v16.4s\n"
-      "ldr q3, [x26], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
-      "zip2 v0.4s, v26.4s, v18.4s\n"
-      "zip2 v4.4s, v23.4s, v16.4s\n"
-      "ldr q16, [x24], #0x10\n"
-      "ldr q1, [x9], #0x10\n"
-      "zip1 v31.4s, v25.4s, v20.4s\n"
-      "zip1 v23.4s, v21.4s, v17.4s\n"
-      "ldr q30, [x26], #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "zip2 v29.4s, v25.4s, v20.4s\n"
-      "zip2 v28.4s, v21.4s, v17.4s\n"
-      "ldr q25, [x24], #0x10\n"
-      "ldr q20, [x9], #0x10\n"
-      "zip1 v21.4s, v7.4s, v24.4s\n"
-      "zip1 v26.4s, v3.4s, v16.4s\n"
-      "ldr q17, [x26], #0x10\n"
-      "ldr q27, [x25], #0x10\n"
-      "zip2 v24.4s, v7.4s, v24.4s\n"
-      "zip2 v14.4s, v3.4s, v16.4s\n"
-      "ldr q10, [x24], #0x10\n"
-      "zip1 v16.4s, v1.4s, v18.4s\n"
-      "zip1 v13.4s, v30.4s, v25.4s\n"
-      "zip2 v18.4s, v1.4s, v18.4s\n"
-      "zip2 v12.4s, v30.4s, v25.4s\n"
-      "zip1 v15.4s, v20.4s, v27.4s\n"
-      "zip1 v9.4s, v17.4s, v10.4s\n"
-      "zip2 v8.4s, v20.4s, v27.4s\n"
-      "zip2 v10.4s, v17.4s, v10.4s\n"
-      "zip1 v17.4s, v6.4s, v11.4s\n"
-      "zip1 v7.4s, v2.4s, v22.4s\n"
-      "zip1 v3.4s, v19.4s, v5.4s\n"
-      "zip1 v1.4s, v0.4s, v4.4s\n"
-      "zip1 v30.4s, v31.4s, v23.4s\n"
-      "zip1 v25.4s, v29.4s, v28.4s\n"
-      "zip1 v27.4s, v21.4s, v26.4s\n"
-      "zip1 v20.4s, v24.4s, v14.4s\n"
-      ".inst 0x0ea16a31  // bfcvtn v17.4h, v17.4s\n"
-      "zip2 v6.4s, v6.4s, v11.4s\n"
-      "zip1 v11.4s, v16.4s, v13.4s\n"
+      "zip1 v19.4s, v21.4s, v22.4s\n"
+      "zip1 v18.4s, v11.4s, v23.4s\n"
+      "zip2 v6.4s, v21.4s, v22.4s\n"
+      "zip2 v11.4s, v11.4s, v23.4s\n"
+      "zip1 v8.4s, v9.4s, v30.4s\n"
+      "zip1 v21.4s, v17.4s, v5.4s\n"
+      "zip1 v7.4s, v0.4s, v3.4s\n"
+      "zip1 v31.4s, v1.4s, v16.4s\n"
+      "zip1 v29.4s, v4.4s, v2.4s\n"
+      "zip1 v22.4s, v27.4s, v28.4s\n"
+      "zip1 v24.4s, v25.4s, v26.4s\n"
+      "zip1 v23.4s, v14.4s, v12.4s\n"
+      ".inst 0x0ea16908  // bfcvtn v8.4h, v8.4s\n"
+      "zip2 v9.4s, v9.4s, v30.4s\n"
+      "zip1 v30.4s, v15.4s, v13.4s\n"
+      ".inst 0x0ea16ab5  // bfcvtn v21.4h, v21.4s\n"
+      "zip2 v5.4s, v17.4s, v5.4s\n"
+      "zip1 v17.4s, v20.4s, v10.4s\n"
       ".inst 0x0ea168e7  // bfcvtn v7.4h, v7.4s\n"
-      "zip2 v22.4s, v2.4s, v22.4s\n"
-      "zip1 v2.4s, v18.4s, v12.4s\n"
-      ".inst 0x0ea16863  // bfcvtn v3.4h, v3.4s\n"
-      "zip2 v5.4s, v19.4s, v5.4s\n"
-      "zip1 v19.4s, v15.4s, v9.4s\n"
-      ".inst 0x0ea16821  // bfcvtn v1.4h, v1.4s\n"
-      "zip2 v4.4s, v0.4s, v4.4s\n"
-      "zip1 v0.4s, v8.4s, v10.4s\n"
-      ".inst 0x0ea16bde  // bfcvtn v30.4h, v30.4s\n"
-      "zip2 v31.4s, v31.4s, v23.4s\n"
-      ".inst 0x0ea16b37  // bfcvtn v23.4h, v25.4s\n"
-      "zip2 v29.4s, v29.4s, v28.4s\n"
-      ".inst 0x0ea16b7c  // bfcvtn v28.4h, v27.4s\n"
-      "zip2 v27.4s, v21.4s, v26.4s\n"
-      ".inst 0x0ea16a9a  // bfcvtn v26.4h, v20.4s\n"
-      "zip2 v25.4s, v24.4s, v14.4s\n"
-      ".inst 0x0ea16978  // bfcvtn v24.4h, v11.4s\n"
-      "zip2 v14.4s, v16.4s, v13.4s\n"
-      ".inst 0x0ea16855  // bfcvtn v21.4h, v2.4s\n"
-      "zip2 v20.4s, v18.4s, v12.4s\n"
-      ".inst 0x0ea16a73  // bfcvtn v19.4h, v19.4s\n"
-      "zip2 v18.4s, v15.4s, v9.4s\n"
-      ".inst 0x0ea1680b  // bfcvtn v11.4h, v0.4s\n"
-      "zip2 v16.4s, v8.4s, v10.4s\n"
-      ".inst 0x4ea168d1  // bfcvtn2 v17.8h, v6.4s\n"
-      ".inst 0x4ea16ac7  // bfcvtn2 v7.8h, v22.4s\n"
-      ".inst 0x4ea168a3  // bfcvtn2 v3.8h, v5.4s\n"
-      ".inst 0x4ea16881  // bfcvtn2 v1.8h, v4.4s\n"
-      ".inst 0x4ea16bfe  // bfcvtn2 v30.8h, v31.4s\n"
-      ".inst 0x4ea16bb7  // bfcvtn2 v23.8h, v29.4s\n"
-      ".inst 0x4ea16b7c  // bfcvtn2 v28.8h, v27.4s\n"
-      ".inst 0x4ea16b3a  // bfcvtn2 v26.8h, v25.4s\n"
-      "str q17, [x27, #0x0]\n"
-      ".inst 0x4ea169d8  // bfcvtn2 v24.8h, v14.4s\n"
-      ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
-      "str q7, [x27, #0x10]\n"
-      ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
-      ".inst 0x4ea16a0b  // bfcvtn2 v11.8h, v16.4s\n"
-      "str q3, [x27, #0x20]\n"
-      "str q1, [x27, #0x30]\n"
-      "str q30, [x27, #0x40]\n"
-      "str q23, [x27, #0x50]\n"
-      "str q28, [x27, #0x60]\n"
-      "str q26, [x27, #0x70]\n"
-      "str q24, [x27, #0x80]\n"
-      "str q21, [x27, #0x90]\n"
-      "str q19, [x27, #0xa0]\n"
-      "str q11, [x27, #0xb0]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "zip2 v0.4s, v0.4s, v3.4s\n"
+      "zip1 v3.4s, v19.4s, v18.4s\n"
+      ".inst 0x0ea16bff  // bfcvtn v31.4h, v31.4s\n"
+      "zip2 v16.4s, v1.4s, v16.4s\n"
+      "zip1 v1.4s, v6.4s, v11.4s\n"
+      ".inst 0x0ea16bbd  // bfcvtn v29.4h, v29.4s\n"
+      "zip2 v2.4s, v4.4s, v2.4s\n"
+      ".inst 0x0ea16ac4  // bfcvtn v4.4h, v22.4s\n"
+      "zip2 v27.4s, v27.4s, v28.4s\n"
+      ".inst 0x0ea16b1c  // bfcvtn v28.4h, v24.4s\n"
+      "zip2 v25.4s, v25.4s, v26.4s\n"
+      ".inst 0x0ea16afa  // bfcvtn v26.4h, v23.4s\n"
+      "zip2 v14.4s, v14.4s, v12.4s\n"
+      ".inst 0x0ea16bd8  // bfcvtn v24.4h, v30.4s\n"
+      "zip2 v13.4s, v15.4s, v13.4s\n"
+      ".inst 0x0ea16a2f  // bfcvtn v15.4h, v17.4s\n"
+      "zip2 v12.4s, v20.4s, v10.4s\n"
+      ".inst 0x0ea16874  // bfcvtn v20.4h, v3.4s\n"
+      "zip2 v10.4s, v19.4s, v18.4s\n"
+      ".inst 0x0ea16831  // bfcvtn v17.4h, v1.4s\n"
+      "zip2 v18.4s, v6.4s, v11.4s\n"
+      ".inst 0x4ea16928  // bfcvtn2 v8.8h, v9.4s\n"
+      ".inst 0x4ea168b5  // bfcvtn2 v21.8h, v5.4s\n"
+      "str q8, [x21, #0x0]\n"
+      ".inst 0x4ea16807  // bfcvtn2 v7.8h, v0.4s\n"
+      ".inst 0x4ea16a1f  // bfcvtn2 v31.8h, v16.4s\n"
+      "str q21, [x21, #0x10]\n"
+      ".inst 0x4ea1685d  // bfcvtn2 v29.8h, v2.4s\n"
+      ".inst 0x4ea16b64  // bfcvtn2 v4.8h, v27.4s\n"
+      "str q7, [x21, #0x20]\n"
+      ".inst 0x4ea16b3c  // bfcvtn2 v28.8h, v25.4s\n"
+      ".inst 0x4ea169da  // bfcvtn2 v26.8h, v14.4s\n"
+      "str q31, [x21, #0x30]\n"
+      ".inst 0x4ea169b8  // bfcvtn2 v24.8h, v13.4s\n"
+      ".inst 0x4ea1698f  // bfcvtn2 v15.8h, v12.4s\n"
+      "str q29, [x21, #0x40]\n"
+      ".inst 0x4ea16954  // bfcvtn2 v20.8h, v10.4s\n"
+      ".inst 0x4ea16a51  // bfcvtn2 v17.8h, v18.4s\n"
+      "str q4, [x21, #0x50]\n"
+      "str q28, [x21, #0x60]\n"
+      "str q26, [x21, #0x70]\n"
+      "str q24, [x21, #0x80]\n"
+      "str q15, [x21, #0x90]\n"
+      "str q20, [x21, #0xa0]\n"
+      "str q17, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Column loop skip
       "cmp x20, #0x10\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "str q16, [x27, #0x20]\n"
-      "str q16, [x27, #0x30]\n"
-      "str q16, [x27, #0x40]\n"
-      "str q16, [x27, #0x50]\n"
-      "str q16, [x27, #0x60]\n"
-      "str q16, [x27, #0x70]\n"
-      "str q16, [x27, #0x80]\n"
-      "str q16, [x27, #0x90]\n"
-      "str q16, [x27, #0xa0]\n"
-      "str q16, [x27, #0xb0]\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 16 loop: loop
-      "ldr q20, [x9], #0x10\n"
-      "ldr q19, [x26], #0x10\n"
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 16 loop: loop
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
       "sub x20, x20, #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x10\n"
-      "ldr q0, [x9], #0x10\n"
-      "ldr q31, [x26], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
-      "ldr q16, [x24], #0x10\n"
-      "ldr q23, [x9], #0x10\n"
-      "zip1 v30.4s, v20.4s, v18.4s\n"
-      "zip1 v29.4s, v19.4s, v17.4s\n"
-      "ldr q22, [x26], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
-      "zip2 v28.4s, v20.4s, v18.4s\n"
-      "zip2 v27.4s, v19.4s, v17.4s\n"
-      "ldr q20, [x24], #0x10\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v30.4s, v21.4s, v17.4s\n"
+      "zip1 v29.4s, v20.4s, v16.4s\n"
       "ldr q19, [x9], #0x10\n"
-      "zip1 v26.4s, v0.4s, v24.4s\n"
-      "zip1 v25.4s, v31.4s, v16.4s\n"
-      "ldr q18, [x26], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
-      "zip2 v8.4s, v0.4s, v24.4s\n"
-      "zip2 v24.4s, v31.4s, v16.4s\n"
-      "ldr q16, [x24], #0x10\n"
-      "zip1 v7.4s, v23.4s, v21.4s\n"
-      "zip1 v6.4s, v22.4s, v20.4s\n"
-      "zip2 v5.4s, v23.4s, v21.4s\n"
-      "zip2 v4.4s, v22.4s, v20.4s\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v28.4s, v21.4s, v17.4s\n"
+      "zip2 v27.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v26.4s, v19.4s, v17.4s\n"
+      "zip1 v25.4s, v18.4s, v16.4s\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v8.4s, v19.4s, v17.4s\n"
+      "zip2 v24.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v7.4s, v21.4s, v17.4s\n"
+      "zip1 v6.4s, v20.4s, v16.4s\n"
+      "ldr q19, [x9], #0x10\n"
+      "ldr q18, [x28], #0x10\n"
+      "zip2 v5.4s, v21.4s, v17.4s\n"
+      "zip2 v4.4s, v20.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
       "zip1 v3.4s, v19.4s, v17.4s\n"
       "zip1 v2.4s, v18.4s, v16.4s\n"
       "zip2 v1.4s, v19.4s, v17.4s\n"
@@ -739,71 +698,70 @@ void a64_transpose_interleave_24_2x4_fp32bf16(bfloat16 *out, const float *in, si
       "zip2 v16.4s, v1.4s, v0.4s\n"
       ".inst 0x4ea16bdf  // bfcvtn2 v31.8h, v30.4s\n"
       ".inst 0x4ea16b9d  // bfcvtn2 v29.8h, v28.4s\n"
+      "str q31, [x21, #0x0]\n"
       ".inst 0x4ea16b5b  // bfcvtn2 v27.8h, v26.4s\n"
       ".inst 0x4ea16b19  // bfcvtn2 v25.8h, v24.4s\n"
+      "str q29, [x21, #0x10]\n"
       ".inst 0x4ea16ad7  // bfcvtn2 v23.8h, v22.4s\n"
       ".inst 0x4ea16a95  // bfcvtn2 v21.8h, v20.4s\n"
+      "str q27, [x21, #0x20]\n"
       ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
       ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
-      "str q31, [x27, #0x0]\n"
-      "str q29, [x27, #0x10]\n"
-      "str q27, [x27, #0x20]\n"
-      "str q25, [x27, #0x30]\n"
-      "str q23, [x27, #0x40]\n"
-      "str q21, [x27, #0x50]\n"
-      "str q19, [x27, #0x60]\n"
-      "str q17, [x27, #0x70]\n"
-      "add x27, x27, #0x80\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 16 loop: skip
+      "str q25, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q21, [x21, #0x50]\n"
+      "str q19, [x21, #0x60]\n"
+      "str q17, [x21, #0x70]\n"
+      "add x21, x21, #0x80\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 16 loop: skip
       "cmp x20, #0x4\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
-      "ldr q21, [x9], #0x10\n"
-      "ldr q20, [x26], #0x10\n"
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
+      "ldr q20, [x9], #0x10\n"
+      "ldr q19, [x28], #0x10\n"
       "sub x20, x20, #0x4\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q17, [x24], #0x10\n"
       "cmp x20, #0x4\n"
-      "zip1 v18.4s, v21.4s, v19.4s\n"
-      "zip1 v16.4s, v20.4s, v17.4s\n"
-      "zip2 v21.4s, v21.4s, v19.4s\n"
-      "zip2 v20.4s, v20.4s, v17.4s\n"
-      "zip1 v17.4s, v18.4s, v16.4s\n"
-      "zip2 v19.4s, v18.4s, v16.4s\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.4s, v20.4s, v17.4s\n"
+      "zip1 v18.4s, v19.4s, v16.4s\n"
+      "zip2 v21.4s, v20.4s, v17.4s\n"
+      "zip2 v20.4s, v19.4s, v16.4s\n"
+      "zip1 v17.4s, v22.4s, v18.4s\n"
       "zip1 v16.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a32  // bfcvtn v18.4h, v17.4s\n"
-      "zip2 v17.4s, v21.4s, v20.4s\n"
-      ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      ".inst 0x4ea16a72  // bfcvtn2 v18.8h, v19.4s\n"
-      ".inst 0x4ea16a30  // bfcvtn2 v16.8h, v17.4s\n"
-      "str q18, [x27, #0x0]\n"
-      "str q16, [x27, #0x10]\n"
-      "add x27, x27, #0x20\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      ".inst 0x0ea16a33  // bfcvtn v19.4h, v17.4s\n"
+      "zip2 v18.4s, v22.4s, v18.4s\n"
+      ".inst 0x0ea16a11  // bfcvtn v17.4h, v16.4s\n"
+      "zip2 v16.4s, v21.4s, v20.4s\n"
+      ".inst 0x4ea16a53  // bfcvtn2 v19.8h, v18.4s\n"
+      ".inst 0x4ea16a11  // bfcvtn2 v17.8h, v16.4s\n"
+      "str q19, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "add x21, x21, #0x20\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr s19, [x9], #0x4\n"
-      "ldr s18, [x26], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
       "sub x20, x20, #0x1\n"
-      "ldr s17, [x25], #0x4\n"
-      "ldr s16, [x24], #0x4\n"
       "cmp x20, #0x1\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
       "zip1 v17.4s, v19.4s, v17.4s\n"
       "zip1 v16.4s, v18.4s, v16.4s\n"
       "zip1 v16.4s, v17.4s, v16.4s\n"
       ".inst 0x0ea16a10  // bfcvtn v16.4h, v16.4s\n"
-      "str d16, [x27, #0x0]\n"
-      "add x27, x27, #0x8\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0xc0\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_bf16fp32.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_bf16fp32.hpp
index 94c442b772..e2884ef80b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_bf16fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_bf16fp32.hpp
@@ -34,266 +34,245 @@ void a64_transpose_interleave_24_bf16fp32(float *out, const bfloat16 *in, size_t
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
+      "ldr q18, [x25], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
       "sub x24, x24, #0x18\n"
-      "ldr q17, [x21], #0x10\n"
-      "ldr q27, [x20], #0x10\n"
+      "shll v26.4s, v18.4h, #0x10\n"
+      "ldr q16, [x22], #0x10\n"
+      "ldr q25, [x20], #0x10\n"
+      "shll2 v24.4s, v18.8h, #0x10\n"
+      "shll v5.4s, v17.4h, #0x10\n"
+      "ldr q23, [x25], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "shll v21.4s, v23.4h, #0x10\n"
+      "shll2 v4.4s, v17.8h, #0x10\n"
+      "ldr q20, [x22], #0x10\n"
+      "ldr q3, [x20], #0x10\n"
+      "shll v2.4s, v22.4h, #0x10\n"
+      "shll v1.4s, v16.4h, #0x10\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "shll2 v0.4s, v16.8h, #0x10\n"
+      "shll v31.4s, v20.4h, #0x10\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "shll v30.4s, v25.4h, #0x10\n"
+      "shll2 v29.4s, v25.8h, #0x10\n"
+      "shll v28.4s, v3.4h, #0x10\n"
+      "str q26, [x21, #0x0]\n"
       "cmp x24, #0x18\n"
-      "ldr q26, [x25], #0x10\n"
-      "ldr q3, [x22], #0x10\n"
-      "ldr q2, [x21], #0x10\n"
-      "shll v16.4s, v19.4h, #0x10\n"
+      "shll2 v27.4s, v23.8h, #0x10\n"
+      "str q24, [x21, #0x10]\n"
+      "shll v26.4s, v19.4h, #0x10\n"
       "shll2 v25.4s, v19.8h, #0x10\n"
-      "ldr q1, [x20], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
+      "str q21, [x21, #0x20]\n"
+      "shll2 v24.4s, v22.8h, #0x10\n"
       "shll v23.4s, v18.4h, #0x10\n"
+      "str q5, [x21, #0x30]\n"
       "shll2 v22.4s, v18.8h, #0x10\n"
-      "ldr q21, [x22], #0x10\n"
-      "ldr q0, [x21], #0x10\n"
-      "shll v20.4s, v26.4h, #0x10\n"
-      "shll v19.4s, v3.4h, #0x10\n"
-      "ldr q31, [x20], #0x10\n"
-      "shll v18.4s, v17.4h, #0x10\n"
-      "shll2 v17.4s, v17.8h, #0x10\n"
-      "str q16, [x23, #0x0]\n"
-      "shll v16.4s, v2.4h, #0x10\n"
-      "shll v30.4s, v27.4h, #0x10\n"
-      "str q25, [x23, #0x10]\n"
-      "shll2 v29.4s, v27.8h, #0x10\n"
-      "shll v28.4s, v1.4h, #0x10\n"
-      "str q20, [x23, #0x20]\n"
-      "str q23, [x23, #0x30]\n"
-      "shll2 v27.4s, v26.8h, #0x10\n"
-      "shll v26.4s, v24.4h, #0x10\n"
-      "str q22, [x23, #0x40]\n"
-      "shll2 v25.4s, v24.8h, #0x10\n"
-      "shll2 v24.4s, v3.8h, #0x10\n"
-      "str q19, [x23, #0x50]\n"
-      "shll v23.4s, v21.4h, #0x10\n"
-      "shll2 v22.4s, v21.8h, #0x10\n"
-      "str q18, [x23, #0x60]\n"
-      "shll2 v21.4s, v2.8h, #0x10\n"
-      "shll v20.4s, v0.4h, #0x10\n"
-      "str q17, [x23, #0x70]\n"
-      "shll2 v19.4s, v0.8h, #0x10\n"
-      "shll2 v18.4s, v1.8h, #0x10\n"
-      "str q16, [x23, #0x80]\n"
-      "shll v17.4s, v31.4h, #0x10\n"
-      "shll2 v16.4s, v31.8h, #0x10\n"
-      "str q30, [x23, #0x90]\n"
-      "str q29, [x23, #0xa0]\n"
-      "str q28, [x23, #0xb0]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q27, [x23, #0x0]\n"
-      "str q26, [x23, #0x10]\n"
-      "str q25, [x23, #0x20]\n"
-      "str q24, [x23, #0x30]\n"
-      "str q23, [x23, #0x40]\n"
-      "str q22, [x23, #0x50]\n"
-      "str q21, [x23, #0x60]\n"
-      "str q20, [x23, #0x70]\n"
-      "str q19, [x23, #0x80]\n"
-      "str q18, [x23, #0x90]\n"
-      "str q17, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "shll2 v21.4s, v20.8h, #0x10\n"
+      "str q4, [x21, #0x40]\n"
+      "shll v20.4s, v17.4h, #0x10\n"
+      "shll2 v19.4s, v17.8h, #0x10\n"
+      "str q2, [x21, #0x50]\n"
+      "shll2 v18.4s, v3.8h, #0x10\n"
+      "shll v17.4s, v16.4h, #0x10\n"
+      "str q1, [x21, #0x60]\n"
+      "shll2 v16.4s, v16.8h, #0x10\n"
+      "str q0, [x21, #0x70]\n"
+      "str q31, [x21, #0x80]\n"
+      "str q30, [x21, #0x90]\n"
+      "str q29, [x21, #0xa0]\n"
+      "str q28, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q27, [x21, #0x0]\n"
+      "str q26, [x21, #0x10]\n"
+      "str q25, [x21, #0x20]\n"
+      "str q24, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q22, [x21, #0x50]\n"
+      "str q21, [x21, #0x60]\n"
+      "str q20, [x21, #0x70]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q17, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
       "ldr q16, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
+      "ldr q21, [x23], #0x10\n"
       "sub x24, x24, #0xc\n"
-      "ldr q27, [x21], #0x10\n"
-      "ldr q26, [x20], #0x10\n"
       "cmp x24, #0xc\n"
-      "ldr d21, [x25], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "ldr d19, [x21], #0x8\n"
-      "shll v18.4s, v16.4h, #0x10\n"
-      "shll2 v17.4s, v16.8h, #0x10\n"
+      "ldr q20, [x22], #0x10\n"
+      "ldr q27, [x20], #0x10\n"
+      "shll v19.4s, v16.4h, #0x10\n"
+      "shll2 v26.4s, v16.8h, #0x10\n"
+      "ldr d16, [x25], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
+      "shll v25.4s, v16.4h, #0x10\n"
+      "shll v24.4s, v21.4h, #0x10\n"
+      "ldr d17, [x22], #0x8\n"
       "ldr d16, [x20], #0x8\n"
-      "shll v25.4s, v22.4h, #0x10\n"
-      "shll2 v24.4s, v22.8h, #0x10\n"
-      "shll v23.4s, v21.4h, #0x10\n"
-      "shll v22.4s, v20.4h, #0x10\n"
-      "shll v21.4s, v27.4h, #0x10\n"
-      "shll2 v20.4s, v27.8h, #0x10\n"
-      "str q18, [x23, #0x0]\n"
-      "shll v19.4s, v19.4h, #0x10\n"
-      "shll v18.4s, v26.4h, #0x10\n"
-      "str q17, [x23, #0x10]\n"
-      "shll2 v17.4s, v26.8h, #0x10\n"
+      "shll2 v23.4s, v21.8h, #0x10\n"
+      "shll v22.4s, v18.4h, #0x10\n"
+      "shll v21.4s, v20.4h, #0x10\n"
+      "shll2 v20.4s, v20.8h, #0x10\n"
+      "str q19, [x21, #0x0]\n"
+      "shll v19.4s, v17.4h, #0x10\n"
+      "shll v18.4s, v27.4h, #0x10\n"
+      "str q26, [x21, #0x10]\n"
+      "shll2 v17.4s, v27.8h, #0x10\n"
       "shll v16.4s, v16.4h, #0x10\n"
-      "str q23, [x23, #0x20]\n"
-      "str q25, [x23, #0x30]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q22, [x23, #0x50]\n"
-      "str q21, [x23, #0x60]\n"
-      "str q20, [x23, #0x70]\n"
-      "str q19, [x23, #0x80]\n"
-      "str q18, [x23, #0x90]\n"
-      "str q17, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "str q25, [x21, #0x20]\n"
+      "str q24, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q22, [x21, #0x50]\n"
+      "str q21, [x21, #0x60]\n"
+      "str q20, [x21, #0x70]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q17, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "str q16, [x23, #0x80]\n"
-      "str q16, [x23, #0x90]\n"
-      "str q16, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr d19, [x25], #0x8\n"
-      "ldr d18, [x22], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "sub x24, x24, #0x4\n"
-      "ldr d17, [x21], #0x8\n"
-      "ldr d16, [x20], #0x8\n"
       "cmp x24, #0x4\n"
+      "ldr d17, [x22], #0x8\n"
+      "ldr d16, [x20], #0x8\n"
       "shll v19.4s, v19.4h, #0x10\n"
       "shll v18.4s, v18.4h, #0x10\n"
       "shll v17.4s, v17.4h, #0x10\n"
       "shll v16.4s, v16.4h, #0x10\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x30]\n"
-      "str q17, [x23, #0x60]\n"
-      "str q16, [x23, #0x90]\n"
-      "add x23, x23, #0x10\n"
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x30]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x90]\n"
+      "add x21, x21, #0x10\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr h19, [x25], #0x2\n"
-      "ldr h18, [x22], #0x2\n"
+      "ldr h18, [x23], #0x2\n"
       "sub x24, x24, #0x1\n"
-      "ldr h17, [x21], #0x2\n"
-      "ldr h16, [x20], #0x2\n"
       "cmp x24, #0x1\n"
+      "ldr h17, [x22], #0x2\n"
+      "ldr h16, [x20], #0x2\n"
       "shll v19.4s, v19.4h, #0x10\n"
       "shll v18.4s, v18.4h, #0x10\n"
       "shll v17.4s, v17.4h, #0x10\n"
       "shll v16.4s, v16.4h, #0x10\n"
-      "str s19, [x23, #0x0]\n"
-      "str s18, [x23, #0x30]\n"
-      "str s17, [x23, #0x60]\n"
-      "str s16, [x23, #0x90]\n"
-      "add x23, x23, #0x4\n"
+      "str s19, [x21, #0x0]\n"
+      "str s18, [x21, #0x30]\n"
+      "str s17, [x21, #0x60]\n"
+      "str s16, [x21, #0x90]\n"
+      "add x21, x21, #0x4\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0xc0\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x18\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
       "ldr q16, [x25], #0x10\n"
-      "sub x20, x20, #0x18\n"
-      "ldr q18, [x25], #0x10\n"
       "ldr q20, [x25], #0x10\n"
+      "sub x20, x20, #0x18\n"
+      "shll v18.4s, v16.4h, #0x10\n"
+      "ldr q19, [x25], #0x10\n"
+      "shll2 v17.4s, v16.8h, #0x10\n"
+      "shll v16.4s, v20.4h, #0x10\n"
+      "str q18, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
       "cmp x20, #0x18\n"
-      "shll v17.4s, v16.4h, #0x10\n"
-      "shll2 v16.4s, v16.8h, #0x10\n"
-      "shll v19.4s, v18.4h, #0x10\n"
-      "shll2 v18.4s, v18.8h, #0x10\n"
-      "str q17, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "shll v17.4s, v20.4h, #0x10\n"
-      "shll2 v16.4s, v20.8h, #0x10\n"
-      "str q19, [x23, #0x20]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q18, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "shll2 v18.4s, v20.8h, #0x10\n"
+      "shll v17.4s, v19.4h, #0x10\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "shll2 v16.4s, v19.8h, #0x10\n"
+      "str q18, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
       "ldr q17, [x25], #0x10\n"
+      "ldr d18, [x25], #0x8\n"
       "sub x20, x20, #0xc\n"
-      "ldr d16, [x25], #0x8\n"
       "cmp x20, #0xc\n"
-      "shll v18.4s, v17.4h, #0x10\n"
+      "shll v16.4s, v17.4h, #0x10\n"
       "shll2 v17.4s, v17.8h, #0x10\n"
-      "shll v16.4s, v16.4h, #0x10\n"
-      "str q18, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q16, [x21, #0x0]\n"
+      "shll v16.4s, v18.4h, #0x10\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr d16, [x25], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
       "shll v16.4s, v16.4h, #0x10\n"
-      "str q16, [x23, #0x0]\n"
-      "add x23, x23, #0x10\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr h16, [x25], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
       "shll v16.4s, v16.4h, #0x10\n"
-      "str s16, [x23, #0x0]\n"
-      "add x23, x23, #0x4\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x30\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_fp16fp32.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_fp16fp32.hpp
index 4d106f4e6d..5e64f812e1 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_fp16fp32.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_fp16fp32.hpp
@@ -34,266 +34,245 @@ void a64_transpose_interleave_24_fp16fp32(float *out, const __fp16 *in, size_t w
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
+      "ldr q18, [x25], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
       "sub x24, x24, #0x18\n"
-      "ldr q17, [x21], #0x10\n"
-      "ldr q27, [x20], #0x10\n"
+      "fcvtl v26.4s, v18.4h\n"
+      "ldr q16, [x22], #0x10\n"
+      "ldr q25, [x20], #0x10\n"
+      "fcvtl2 v24.4s, v18.8h\n"
+      "fcvtl v5.4s, v17.4h\n"
+      "ldr q23, [x25], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "fcvtl v21.4s, v23.4h\n"
+      "fcvtl2 v4.4s, v17.8h\n"
+      "ldr q20, [x22], #0x10\n"
+      "ldr q3, [x20], #0x10\n"
+      "fcvtl v2.4s, v22.4h\n"
+      "fcvtl v1.4s, v16.4h\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "fcvtl2 v0.4s, v16.8h\n"
+      "fcvtl v31.4s, v20.4h\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "fcvtl v30.4s, v25.4h\n"
+      "fcvtl2 v29.4s, v25.8h\n"
+      "fcvtl v28.4s, v3.4h\n"
+      "str q26, [x21, #0x0]\n"
       "cmp x24, #0x18\n"
-      "ldr q26, [x25], #0x10\n"
-      "ldr q3, [x22], #0x10\n"
-      "ldr q2, [x21], #0x10\n"
-      "fcvtl v16.4s, v19.4h\n"
+      "fcvtl2 v27.4s, v23.8h\n"
+      "str q24, [x21, #0x10]\n"
+      "fcvtl v26.4s, v19.4h\n"
       "fcvtl2 v25.4s, v19.8h\n"
-      "ldr q1, [x20], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
+      "str q21, [x21, #0x20]\n"
+      "fcvtl2 v24.4s, v22.8h\n"
       "fcvtl v23.4s, v18.4h\n"
+      "str q5, [x21, #0x30]\n"
       "fcvtl2 v22.4s, v18.8h\n"
-      "ldr q21, [x22], #0x10\n"
-      "ldr q0, [x21], #0x10\n"
-      "fcvtl v20.4s, v26.4h\n"
-      "fcvtl v19.4s, v3.4h\n"
-      "ldr q31, [x20], #0x10\n"
-      "fcvtl v18.4s, v17.4h\n"
-      "fcvtl2 v17.4s, v17.8h\n"
-      "str q16, [x23, #0x0]\n"
-      "fcvtl v16.4s, v2.4h\n"
-      "fcvtl v30.4s, v27.4h\n"
-      "str q25, [x23, #0x10]\n"
-      "fcvtl2 v29.4s, v27.8h\n"
-      "fcvtl v28.4s, v1.4h\n"
-      "str q20, [x23, #0x20]\n"
-      "str q23, [x23, #0x30]\n"
-      "fcvtl2 v27.4s, v26.8h\n"
-      "fcvtl v26.4s, v24.4h\n"
-      "str q22, [x23, #0x40]\n"
-      "fcvtl2 v25.4s, v24.8h\n"
-      "fcvtl2 v24.4s, v3.8h\n"
-      "str q19, [x23, #0x50]\n"
-      "fcvtl v23.4s, v21.4h\n"
-      "fcvtl2 v22.4s, v21.8h\n"
-      "str q18, [x23, #0x60]\n"
-      "fcvtl2 v21.4s, v2.8h\n"
-      "fcvtl v20.4s, v0.4h\n"
-      "str q17, [x23, #0x70]\n"
-      "fcvtl2 v19.4s, v0.8h\n"
-      "fcvtl2 v18.4s, v1.8h\n"
-      "str q16, [x23, #0x80]\n"
-      "fcvtl v17.4s, v31.4h\n"
-      "fcvtl2 v16.4s, v31.8h\n"
-      "str q30, [x23, #0x90]\n"
-      "str q29, [x23, #0xa0]\n"
-      "str q28, [x23, #0xb0]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q27, [x23, #0x0]\n"
-      "str q26, [x23, #0x10]\n"
-      "str q25, [x23, #0x20]\n"
-      "str q24, [x23, #0x30]\n"
-      "str q23, [x23, #0x40]\n"
-      "str q22, [x23, #0x50]\n"
-      "str q21, [x23, #0x60]\n"
-      "str q20, [x23, #0x70]\n"
-      "str q19, [x23, #0x80]\n"
-      "str q18, [x23, #0x90]\n"
-      "str q17, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "fcvtl2 v21.4s, v20.8h\n"
+      "str q4, [x21, #0x40]\n"
+      "fcvtl v20.4s, v17.4h\n"
+      "fcvtl2 v19.4s, v17.8h\n"
+      "str q2, [x21, #0x50]\n"
+      "fcvtl2 v18.4s, v3.8h\n"
+      "fcvtl v17.4s, v16.4h\n"
+      "str q1, [x21, #0x60]\n"
+      "fcvtl2 v16.4s, v16.8h\n"
+      "str q0, [x21, #0x70]\n"
+      "str q31, [x21, #0x80]\n"
+      "str q30, [x21, #0x90]\n"
+      "str q29, [x21, #0xa0]\n"
+      "str q28, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q27, [x21, #0x0]\n"
+      "str q26, [x21, #0x10]\n"
+      "str q25, [x21, #0x20]\n"
+      "str q24, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q22, [x21, #0x50]\n"
+      "str q21, [x21, #0x60]\n"
+      "str q20, [x21, #0x70]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q17, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cmp x24, #0xc\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
       "ldr q16, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
+      "ldr q21, [x23], #0x10\n"
       "sub x24, x24, #0xc\n"
-      "ldr q27, [x21], #0x10\n"
-      "ldr q26, [x20], #0x10\n"
       "cmp x24, #0xc\n"
-      "ldr d21, [x25], #0x8\n"
-      "ldr d20, [x22], #0x8\n"
-      "ldr d19, [x21], #0x8\n"
-      "fcvtl v18.4s, v16.4h\n"
-      "fcvtl2 v17.4s, v16.8h\n"
+      "ldr q20, [x22], #0x10\n"
+      "ldr q27, [x20], #0x10\n"
+      "fcvtl v19.4s, v16.4h\n"
+      "fcvtl2 v26.4s, v16.8h\n"
+      "ldr d16, [x25], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
+      "fcvtl v25.4s, v16.4h\n"
+      "fcvtl v24.4s, v21.4h\n"
+      "ldr d17, [x22], #0x8\n"
       "ldr d16, [x20], #0x8\n"
-      "fcvtl v25.4s, v22.4h\n"
-      "fcvtl2 v24.4s, v22.8h\n"
-      "fcvtl v23.4s, v21.4h\n"
-      "fcvtl v22.4s, v20.4h\n"
-      "fcvtl v21.4s, v27.4h\n"
-      "fcvtl2 v20.4s, v27.8h\n"
-      "str q18, [x23, #0x0]\n"
-      "fcvtl v19.4s, v19.4h\n"
-      "fcvtl v18.4s, v26.4h\n"
-      "str q17, [x23, #0x10]\n"
-      "fcvtl2 v17.4s, v26.8h\n"
+      "fcvtl2 v23.4s, v21.8h\n"
+      "fcvtl v22.4s, v18.4h\n"
+      "fcvtl v21.4s, v20.4h\n"
+      "fcvtl2 v20.4s, v20.8h\n"
+      "str q19, [x21, #0x0]\n"
+      "fcvtl v19.4s, v17.4h\n"
+      "fcvtl v18.4s, v27.4h\n"
+      "str q26, [x21, #0x10]\n"
+      "fcvtl2 v17.4s, v27.8h\n"
       "fcvtl v16.4s, v16.4h\n"
-      "str q23, [x23, #0x20]\n"
-      "str q25, [x23, #0x30]\n"
-      "str q24, [x23, #0x40]\n"
-      "str q22, [x23, #0x50]\n"
-      "str q21, [x23, #0x60]\n"
-      "str q20, [x23, #0x70]\n"
-      "str q19, [x23, #0x80]\n"
-      "str q18, [x23, #0x90]\n"
-      "str q17, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "str q25, [x21, #0x20]\n"
+      "str q24, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q22, [x21, #0x50]\n"
+      "str q21, [x21, #0x60]\n"
+      "str q20, [x21, #0x70]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q17, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "str q16, [x23, #0x80]\n"
-      "str q16, [x23, #0x90]\n"
-      "str q16, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr d19, [x25], #0x8\n"
-      "ldr d18, [x22], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "sub x24, x24, #0x4\n"
-      "ldr d17, [x21], #0x8\n"
-      "ldr d16, [x20], #0x8\n"
       "cmp x24, #0x4\n"
+      "ldr d17, [x22], #0x8\n"
+      "ldr d16, [x20], #0x8\n"
       "fcvtl v19.4s, v19.4h\n"
       "fcvtl v18.4s, v18.4h\n"
       "fcvtl v17.4s, v17.4h\n"
       "fcvtl v16.4s, v16.4h\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x30]\n"
-      "str q17, [x23, #0x60]\n"
-      "str q16, [x23, #0x90]\n"
-      "add x23, x23, #0x10\n"
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x30]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x90]\n"
+      "add x21, x21, #0x10\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr h19, [x25], #0x2\n"
-      "ldr h18, [x22], #0x2\n"
+      "ldr h18, [x23], #0x2\n"
       "sub x24, x24, #0x1\n"
-      "ldr h17, [x21], #0x2\n"
-      "ldr h16, [x20], #0x2\n"
       "cmp x24, #0x1\n"
+      "ldr h17, [x22], #0x2\n"
+      "ldr h16, [x20], #0x2\n"
       "fcvtl v19.4s, v19.4h\n"
       "fcvtl v18.4s, v18.4h\n"
       "fcvtl v17.4s, v17.4h\n"
       "fcvtl v16.4s, v16.4h\n"
-      "str s19, [x23, #0x0]\n"
-      "str s18, [x23, #0x30]\n"
-      "str s17, [x23, #0x60]\n"
-      "str s16, [x23, #0x90]\n"
-      "add x23, x23, #0x4\n"
+      "str s19, [x21, #0x0]\n"
+      "str s18, [x21, #0x30]\n"
+      "str s17, [x21, #0x60]\n"
+      "str s16, [x21, #0x90]\n"
+      "add x21, x21, #0x4\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0xc0\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x18\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Unroll column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Unroll column loop
       "ldr q16, [x25], #0x10\n"
-      "sub x20, x20, #0x18\n"
-      "ldr q18, [x25], #0x10\n"
       "ldr q20, [x25], #0x10\n"
+      "sub x20, x20, #0x18\n"
+      "fcvtl v18.4s, v16.4h\n"
+      "ldr q19, [x25], #0x10\n"
+      "fcvtl2 v17.4s, v16.8h\n"
+      "fcvtl v16.4s, v20.4h\n"
+      "str q18, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
       "cmp x20, #0x18\n"
-      "fcvtl v17.4s, v16.4h\n"
-      "fcvtl2 v16.4s, v16.8h\n"
-      "fcvtl v19.4s, v18.4h\n"
-      "fcvtl2 v18.4s, v18.8h\n"
-      "str q17, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "fcvtl v17.4s, v20.4h\n"
-      "fcvtl2 v16.4s, v20.8h\n"
-      "str q19, [x23, #0x20]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q18, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Unroll column loop skip
+      "fcvtl2 v18.4s, v20.8h\n"
+      "fcvtl v17.4s, v19.4h\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "fcvtl2 v16.4s, v19.8h\n"
+      "str q18, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0xc\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Column loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: Column loop
       "ldr q17, [x25], #0x10\n"
+      "ldr d18, [x25], #0x8\n"
       "sub x20, x20, #0xc\n"
-      "ldr d16, [x25], #0x8\n"
       "cmp x20, #0xc\n"
-      "fcvtl v18.4s, v17.4h\n"
+      "fcvtl v16.4s, v17.4h\n"
       "fcvtl2 v17.4s, v17.8h\n"
-      "fcvtl v16.4s, v16.4h\n"
-      "str q18, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q16, [x21, #0x0]\n"
+      "fcvtl v16.4s, v18.4h\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Column loop skip
       "cmp x20, #0x4\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr d16, [x25], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
       "fcvtl v16.4s, v16.4h\n"
-      "str q16, [x23, #0x0]\n"
-      "add x23, x23, #0x10\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr h16, [x25], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
       "fcvtl v16.4s, v16.4h\n"
-      "str s16, [x23, #0x0]\n"
-      "add x23, x23, #0x4\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x30\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_32_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_32_1x4.hpp
index 8c16e5ba46..918d3ffaa3 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_32_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_32_1x4.hpp
@@ -40,16 +40,14 @@ void a64_transpose_interleave_32_1x4(uint8_t *out, const uint8_t *in, size_t wid
 
     __asm__ __volatile__(
       "cmp %x[height], #0x10\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x17, %x[in]\n"
-      "mov x16, %x[width]\n"
-      "mov x15, %x[out]\n"
-      "sub %x[height], %x[height], #0x10\n"
-      "add x14, x17, %x[in_stride]\n"
+      "add x16, x17, %x[in_stride]\n"
+      "add x15, x16, %x[in_stride]\n"
+      "add x14, x15, %x[in_stride]\n"
       "add x13, x14, %x[in_stride]\n"
       "add x12, x13, %x[in_stride]\n"
-      "cmp x16, #0x20\n"
       "add x11, x12, %x[in_stride]\n"
       "add x10, x11, %x[in_stride]\n"
       "add x9, x10, %x[in_stride]\n"
@@ -57,464 +55,420 @@ void a64_transpose_interleave_32_1x4(uint8_t *out, const uint8_t *in, size_t wid
       "add x27, x28, %x[in_stride]\n"
       "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x20\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x10\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
-      "ldr q28, [x17], #0x10\n"
-      "ldr q6, [x14], #0x10\n"
-      "sub x16, x16, #0x20\n"
-      "ldr q1, [x13], #0x10\n"
-      "ldr q30, [x12], #0x10\n"
-      "cmp x16, #0x20\n"
-      "ldr q21, [x11], #0x10\n"
-      "ldr q9, [x10], #0x10\n"
-      "ldr q15, [x9], #0x10\n"
-      "ldr q17, [x28], #0x10\n"
-      "ldr q19, [x27], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
-      "zip1 v4.16b, v28.16b, v1.16b\n"
-      "zip1 v3.16b, v6.16b, v30.16b\n"
-      "ldr q26, [x25], #0x10\n"
-      "ldr q11, [x24], #0x10\n"
-      "zip2 v29.16b, v28.16b, v1.16b\n"
-      "zip2 v16.16b, v6.16b, v30.16b\n"
-      "ldr q25, [x23], #0x10\n"
-      "ldr q7, [x22], #0x10\n"
-      "zip1 v0.16b, v21.16b, v15.16b\n"
-      "zip1 v1.16b, v9.16b, v17.16b\n"
-      "ldr q23, [x21], #0x10\n"
-      "ldr q28, [x20], #0x10\n"
-      "zip2 v20.16b, v21.16b, v15.16b\n"
-      "zip2 v24.16b, v9.16b, v17.16b\n"
+      "ldr q6, [x17], #0x10\n"
+      "ldr q31, [x16], #0x10\n"
+      "sub x24, x24, #0x20\n"
+      "cmp x24, #0x20\n"
+      "ldr q7, [x15], #0x10\n"
+      "ldr q0, [x14], #0x10\n"
+      "zip1 v9.16b, v6.16b, v7.16b\n"
+      "zip1 v20.16b, v31.16b, v0.16b\n"
+      "ldr q24, [x13], #0x10\n"
+      "ldr q19, [x12], #0x10\n"
+      "zip2 v30.16b, v6.16b, v7.16b\n"
+      "zip2 v12.16b, v31.16b, v0.16b\n"
+      "ldr q23, [x11], #0x10\n"
+      "ldr q17, [x10], #0x10\n"
+      "zip1 v13.16b, v24.16b, v23.16b\n"
+      "zip1 v16.16b, v19.16b, v17.16b\n"
+      "ldr q0, [x9], #0x10\n"
+      "ldr q31, [x28], #0x10\n"
+      "zip2 v15.16b, v24.16b, v23.16b\n"
+      "zip2 v11.16b, v19.16b, v17.16b\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q4, [x26], #0x10\n"
+      "zip1 v1.16b, v0.16b, v17.16b\n"
+      "zip1 v21.16b, v31.16b, v4.16b\n"
+      "ldr q28, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v0.16b, v0.16b, v17.16b\n"
+      "zip2 v26.16b, v31.16b, v4.16b\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q19, [x20], #0x10\n"
+      "zip1 v23.16b, v28.16b, v17.16b\n"
+      "zip1 v25.16b, v18.16b, v19.16b\n"
       "ldr q2, [x17], #0x10\n"
+      "ldr q3, [x16], #0x10\n"
+      "zip2 v7.16b, v28.16b, v17.16b\n"
+      "zip2 v8.16b, v18.16b, v19.16b\n"
+      "ldr q22, [x15], #0x10\n"
       "ldr q27, [x14], #0x10\n"
-      "zip1 v15.16b, v19.16b, v26.16b\n"
-      "zip1 v21.16b, v22.16b, v11.16b\n"
-      "ldr q18, [x13], #0x10\n"
-      "ldr q8, [x12], #0x10\n"
-      "zip2 v14.16b, v19.16b, v26.16b\n"
-      "zip2 v13.16b, v22.16b, v11.16b\n"
-      "ldr q19, [x11], #0x10\n"
-      "ldr q10, [x10], #0x10\n"
-      "zip1 v12.16b, v25.16b, v23.16b\n"
-      "zip1 v5.16b, v7.16b, v28.16b\n"
-      "ldr q6, [x9], #0x10\n"
-      "ldr q30, [x28], #0x10\n"
-      "zip2 v31.16b, v25.16b, v23.16b\n"
-      "zip2 v11.16b, v7.16b, v28.16b\n"
-      "ldr q25, [x27], #0x10\n"
-      "ldr q7, [x26], #0x10\n"
-      "zip1 v23.16b, v2.16b, v18.16b\n"
-      "zip1 v17.16b, v27.16b, v8.16b\n"
-      "ldr q26, [x25], #0x10\n"
-      "ldr q9, [x24], #0x10\n"
-      "zip2 v18.16b, v2.16b, v18.16b\n"
-      "zip2 v22.16b, v27.16b, v8.16b\n"
-      "ldr q28, [x23], #0x10\n"
-      "ldr q8, [x22], #0x10\n"
-      "zip1 v2.16b, v19.16b, v6.16b\n"
-      "zip1 v27.16b, v10.16b, v30.16b\n"
-      "zip2 v6.16b, v19.16b, v6.16b\n"
-      "ldr q19, [x21], #0x10\n"
-      "zip2 v30.16b, v10.16b, v30.16b\n"
-      "zip1 v10.16b, v25.16b, v26.16b\n"
-      "zip2 v26.16b, v25.16b, v26.16b\n"
-      "zip1 v25.16b, v7.16b, v9.16b\n"
-      "zip2 v9.16b, v7.16b, v9.16b\n"
-      "zip1 v7.16b, v28.16b, v19.16b\n"
-      "zip2 v19.16b, v28.16b, v19.16b\n"
-      "zip1 v28.16b, v4.16b, v3.16b\n"
-      "zip2 v3.16b, v4.16b, v3.16b\n"
-      "ldr q4, [x20], #0x10\n"
-      "str q28, [x15, #0x0]\n"
-      "zip1 v28.16b, v29.16b, v16.16b\n"
-      "zip2 v16.16b, v29.16b, v16.16b\n"
-      "zip1 v29.16b, v23.16b, v17.16b\n"
-      "zip2 v23.16b, v23.16b, v17.16b\n"
-      "zip1 v17.16b, v8.16b, v4.16b\n"
-      "zip2 v4.16b, v8.16b, v4.16b\n"
-      "str q3, [x15, #0x10]\n"
-      "zip1 v8.16b, v18.16b, v22.16b\n"
-      "zip2 v22.16b, v18.16b, v22.16b\n"
-      "str q28, [x15, #0x20]\n"
-      "zip1 v28.16b, v0.16b, v1.16b\n"
-      "zip2 v3.16b, v0.16b, v1.16b\n"
-      "str q16, [x15, #0x30]\n"
-      "zip1 v16.16b, v20.16b, v24.16b\n"
-      "zip2 v18.16b, v20.16b, v24.16b\n"
-      "str q29, [x15, #0x40]\n"
-      "zip1 v1.16b, v2.16b, v27.16b\n"
-      "zip2 v24.16b, v2.16b, v27.16b\n"
-      "str q23, [x15, #0x50]\n"
-      "zip1 v2.16b, v6.16b, v30.16b\n"
-      "zip2 v6.16b, v6.16b, v30.16b\n"
-      "str q8, [x15, #0x60]\n"
-      "zip1 v0.16b, v15.16b, v21.16b\n"
-      "zip2 v30.16b, v15.16b, v21.16b\n"
-      "str q22, [x15, #0x70]\n"
-      "zip1 v29.16b, v14.16b, v13.16b\n"
-      "zip2 v23.16b, v14.16b, v13.16b\n"
-      "str q28, [x15, #0x80]\n"
-      "zip1 v27.16b, v10.16b, v25.16b\n"
-      "zip2 v8.16b, v10.16b, v25.16b\n"
-      "str q3, [x15, #0x90]\n"
-      "zip1 v15.16b, v26.16b, v9.16b\n"
-      "zip2 v9.16b, v26.16b, v9.16b\n"
-      "str q16, [x15, #0xa0]\n"
-      "zip1 v28.16b, v12.16b, v5.16b\n"
-      "zip2 v22.16b, v12.16b, v5.16b\n"
-      "str q18, [x15, #0xb0]\n"
-      "zip1 v21.16b, v31.16b, v11.16b\n"
-      "zip2 v20.16b, v31.16b, v11.16b\n"
-      "str q1, [x15, #0xc0]\n"
-      "zip1 v25.16b, v7.16b, v17.16b\n"
-      "zip2 v18.16b, v7.16b, v17.16b\n"
-      "str q24, [x15, #0xd0]\n"
-      "zip1 v17.16b, v19.16b, v4.16b\n"
-      "zip2 v16.16b, v19.16b, v4.16b\n"
-      "str q2, [x15, #0xe0]\n"
-      "str q6, [x15, #0xf0]\n"
-      "str q0, [x15, #0x100]\n"
-      "str q30, [x15, #0x110]\n"
-      "str q29, [x15, #0x120]\n"
-      "str q23, [x15, #0x130]\n"
-      "str q27, [x15, #0x140]\n"
-      "str q8, [x15, #0x150]\n"
-      "str q15, [x15, #0x160]\n"
-      "str q9, [x15, #0x170]\n"
-      "str q28, [x15, #0x180]\n"
-      "str q22, [x15, #0x190]\n"
-      "str q21, [x15, #0x1a0]\n"
-      "str q20, [x15, #0x1b0]\n"
-      "str q25, [x15, #0x1c0]\n"
-      "str q18, [x15, #0x1d0]\n"
-      "str q17, [x15, #0x1e0]\n"
-      "str q16, [x15, #0x1f0]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "zip1 v19.16b, v2.16b, v22.16b\n"
+      "zip1 v17.16b, v3.16b, v27.16b\n"
+      "ldr q6, [x13], #0x10\n"
+      "ldr q4, [x12], #0x10\n"
+      "zip2 v24.16b, v2.16b, v22.16b\n"
+      "zip2 v22.16b, v3.16b, v27.16b\n"
+      "ldr q14, [x11], #0x10\n"
+      "ldr q18, [x10], #0x10\n"
+      "zip1 v29.16b, v6.16b, v14.16b\n"
+      "zip1 v31.16b, v4.16b, v18.16b\n"
+      "ldr q2, [x9], #0x10\n"
+      "ldr q10, [x28], #0x10\n"
+      "zip2 v28.16b, v6.16b, v14.16b\n"
+      "zip2 v27.16b, v4.16b, v18.16b\n"
+      "ldr q6, [x27], #0x10\n"
+      "ldr q5, [x26], #0x10\n"
+      "zip1 v14.16b, v2.16b, v6.16b\n"
+      "zip1 v4.16b, v10.16b, v5.16b\n"
+      "ldr q3, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v6.16b, v2.16b, v6.16b\n"
+      "zip2 v10.16b, v10.16b, v5.16b\n"
+      "ldr q5, [x22], #0x10\n"
+      "zip1 v2.16b, v3.16b, v5.16b\n"
+      "zip2 v3.16b, v3.16b, v5.16b\n"
+      "zip1 v5.16b, v9.16b, v20.16b\n"
+      "zip2 v20.16b, v9.16b, v20.16b\n"
+      "ldr q9, [x20], #0x10\n"
+      "str q5, [x21, #0x0]\n"
+      "zip1 v5.16b, v18.16b, v9.16b\n"
+      "zip2 v9.16b, v18.16b, v9.16b\n"
+      "str q20, [x21, #0x10]\n"
+      "zip1 v18.16b, v30.16b, v12.16b\n"
+      "zip2 v30.16b, v30.16b, v12.16b\n"
+      "str q18, [x21, #0x20]\n"
+      "zip1 v20.16b, v19.16b, v17.16b\n"
+      "zip2 v12.16b, v19.16b, v17.16b\n"
+      "str q30, [x21, #0x30]\n"
+      "zip1 v18.16b, v24.16b, v22.16b\n"
+      "zip2 v17.16b, v24.16b, v22.16b\n"
+      "str q20, [x21, #0x40]\n"
+      "zip1 v30.16b, v13.16b, v16.16b\n"
+      "zip2 v24.16b, v13.16b, v16.16b\n"
+      "str q12, [x21, #0x50]\n"
+      "zip1 v22.16b, v15.16b, v11.16b\n"
+      "zip2 v20.16b, v15.16b, v11.16b\n"
+      "str q18, [x21, #0x60]\n"
+      "zip1 v19.16b, v29.16b, v31.16b\n"
+      "zip2 v18.16b, v29.16b, v31.16b\n"
+      "str q17, [x21, #0x70]\n"
+      "zip1 v17.16b, v28.16b, v27.16b\n"
+      "zip2 v16.16b, v28.16b, v27.16b\n"
+      "str q30, [x21, #0x80]\n"
+      "zip1 v31.16b, v1.16b, v21.16b\n"
+      "zip2 v1.16b, v1.16b, v21.16b\n"
+      "str q24, [x21, #0x90]\n"
+      "zip1 v30.16b, v0.16b, v26.16b\n"
+      "zip2 v29.16b, v0.16b, v26.16b\n"
+      "str q22, [x21, #0xa0]\n"
+      "zip1 v28.16b, v14.16b, v4.16b\n"
+      "zip2 v27.16b, v14.16b, v4.16b\n"
+      "str q20, [x21, #0xb0]\n"
+      "zip1 v26.16b, v6.16b, v10.16b\n"
+      "zip2 v24.16b, v6.16b, v10.16b\n"
+      "str q19, [x21, #0xc0]\n"
+      "zip1 v14.16b, v23.16b, v25.16b\n"
+      "zip2 v22.16b, v23.16b, v25.16b\n"
+      "str q18, [x21, #0xd0]\n"
+      "zip1 v21.16b, v7.16b, v8.16b\n"
+      "zip2 v20.16b, v7.16b, v8.16b\n"
+      "str q17, [x21, #0xe0]\n"
+      "zip1 v19.16b, v2.16b, v5.16b\n"
+      "zip2 v18.16b, v2.16b, v5.16b\n"
+      "str q16, [x21, #0xf0]\n"
+      "zip1 v17.16b, v3.16b, v9.16b\n"
+      "zip2 v16.16b, v3.16b, v9.16b\n"
+      "str q31, [x21, #0x100]\n"
+      "str q1, [x21, #0x110]\n"
+      "str q30, [x21, #0x120]\n"
+      "str q29, [x21, #0x130]\n"
+      "str q28, [x21, #0x140]\n"
+      "str q27, [x21, #0x150]\n"
+      "str q26, [x21, #0x160]\n"
+      "str q24, [x21, #0x170]\n"
+      "str q14, [x21, #0x180]\n"
+      "str q22, [x21, #0x190]\n"
+      "str q21, [x21, #0x1a0]\n"
+      "str q20, [x21, #0x1b0]\n"
+      "str q19, [x21, #0x1c0]\n"
+      "str q18, [x21, #0x1d0]\n"
+      "str q17, [x21, #0x1e0]\n"
+      "str q16, [x21, #0x1f0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x16, 10f\n"
-      "cmp x16, #0x10\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x15, #0x0]\n"
-      "str q16, [x15, #0x10]\n"
-      "str q16, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "str q16, [x15, #0x40]\n"
-      "str q16, [x15, #0x50]\n"
-      "str q16, [x15, #0x60]\n"
-      "str q16, [x15, #0x70]\n"
-      "str q16, [x15, #0x80]\n"
-      "str q16, [x15, #0x90]\n"
-      "str q16, [x15, #0xa0]\n"
-      "str q16, [x15, #0xb0]\n"
-      "str q16, [x15, #0xc0]\n"
-      "str q16, [x15, #0xd0]\n"
-      "str q16, [x15, #0xe0]\n"
-      "str q16, [x15, #0xf0]\n"
-      "str q16, [x15, #0x100]\n"
-      "str q16, [x15, #0x110]\n"
-      "str q16, [x15, #0x120]\n"
-      "str q16, [x15, #0x130]\n"
-      "str q16, [x15, #0x140]\n"
-      "str q16, [x15, #0x150]\n"
-      "str q16, [x15, #0x160]\n"
-      "str q16, [x15, #0x170]\n"
-      "str q16, [x15, #0x180]\n"
-      "str q16, [x15, #0x190]\n"
-      "str q16, [x15, #0x1a0]\n"
-      "str q16, [x15, #0x1b0]\n"
-      "str q16, [x15, #0x1c0]\n"
-      "str q16, [x15, #0x1d0]\n"
-      "str q16, [x15, #0x1e0]\n"
-      "str q16, [x15, #0x1f0]\n"
+      "cmp x24, #0x10\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 16 loop: loop
-      "ldr q19, [x17], #0x10\n"
-      "ldr q18, [x14], #0x10\n"
-      "sub x16, x16, #0x10\n"
-      "ldr q17, [x13], #0x10\n"
-      "ldr q16, [x12], #0x10\n"
-      "cmp x16, #0x10\n"
-      "ldr q24, [x11], #0x10\n"
-      "ldr q28, [x10], #0x10\n"
-      "ldr q23, [x9], #0x10\n"
-      "ldr q22, [x28], #0x10\n"
-      "ldr q27, [x27], #0x10\n"
-      "ldr q26, [x26], #0x10\n"
-      "zip1 v5.16b, v19.16b, v17.16b\n"
-      "zip1 v4.16b, v18.16b, v16.16b\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "zip2 v3.16b, v19.16b, v17.16b\n"
-      "zip2 v2.16b, v18.16b, v16.16b\n"
-      "ldr q19, [x23], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "zip1 v1.16b, v24.16b, v23.16b\n"
-      "zip1 v25.16b, v28.16b, v22.16b\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q21, [x17], #0x10\n"
+      "ldr q20, [x16], #0x10\n"
+      "sub x24, x24, #0x10\n"
+      "cmp x24, #0x10\n"
+      "ldr q17, [x15], #0x10\n"
+      "ldr q16, [x14], #0x10\n"
+      "zip1 v3.16b, v21.16b, v17.16b\n"
+      "zip1 v2.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x13], #0x10\n"
+      "ldr q18, [x12], #0x10\n"
+      "zip2 v1.16b, v21.16b, v17.16b\n"
+      "zip2 v0.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x11], #0x10\n"
+      "ldr q16, [x10], #0x10\n"
+      "zip1 v31.16b, v19.16b, v17.16b\n"
+      "zip1 v30.16b, v18.16b, v16.16b\n"
+      "ldr q24, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v29.16b, v19.16b, v17.16b\n"
+      "zip2 v23.16b, v18.16b, v16.16b\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v22.16b, v24.16b, v17.16b\n"
+      "zip1 v21.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v28.16b, v24.16b, v17.16b\n"
+      "zip2 v20.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
-      "zip2 v23.16b, v28.16b, v22.16b\n"
-      "zip1 v0.16b, v27.16b, v21.16b\n"
-      "zip1 v22.16b, v26.16b, v20.16b\n"
-      "zip2 v31.16b, v27.16b, v21.16b\n"
-      "zip2 v30.16b, v26.16b, v20.16b\n"
-      "zip1 v29.16b, v19.16b, v17.16b\n"
-      "zip1 v28.16b, v18.16b, v16.16b\n"
-      "zip2 v27.16b, v19.16b, v17.16b\n"
-      "zip2 v26.16b, v18.16b, v16.16b\n"
-      "zip1 v21.16b, v5.16b, v4.16b\n"
-      "zip2 v20.16b, v5.16b, v4.16b\n"
-      "zip1 v19.16b, v3.16b, v2.16b\n"
+      "zip1 v27.16b, v19.16b, v17.16b\n"
+      "zip1 v26.16b, v18.16b, v16.16b\n"
+      "zip2 v25.16b, v19.16b, v17.16b\n"
+      "zip2 v24.16b, v18.16b, v16.16b\n"
+      "zip1 v16.16b, v3.16b, v2.16b\n"
       "zip2 v18.16b, v3.16b, v2.16b\n"
-      "zip1 v17.16b, v1.16b, v25.16b\n"
-      "zip2 v16.16b, v1.16b, v25.16b\n"
-      "str q21, [x15, #0x0]\n"
-      "str q20, [x15, #0x10]\n"
-      "zip1 v25.16b, v24.16b, v23.16b\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
-      "str q19, [x15, #0x20]\n"
-      "zip1 v23.16b, v0.16b, v22.16b\n"
-      "zip2 v22.16b, v0.16b, v22.16b\n"
-      "str q18, [x15, #0x30]\n"
-      "zip1 v21.16b, v31.16b, v30.16b\n"
-      "zip2 v20.16b, v31.16b, v30.16b\n"
-      "str q17, [x15, #0x80]\n"
-      "zip1 v19.16b, v29.16b, v28.16b\n"
-      "zip2 v18.16b, v29.16b, v28.16b\n"
-      "str q16, [x15, #0x90]\n"
-      "zip1 v17.16b, v27.16b, v26.16b\n"
-      "zip2 v16.16b, v27.16b, v26.16b\n"
-      "str q25, [x15, #0xa0]\n"
-      "str q24, [x15, #0xb0]\n"
-      "str q23, [x15, #0x100]\n"
-      "str q22, [x15, #0x110]\n"
-      "str q21, [x15, #0x120]\n"
-      "str q20, [x15, #0x130]\n"
-      "str q19, [x15, #0x180]\n"
-      "str q18, [x15, #0x190]\n"
-      "str q17, [x15, #0x1a0]\n"
-      "str q16, [x15, #0x1b0]\n"
-      "add x15, x15, #0x40\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v17.16b, v1.16b, v0.16b\n"
+      "zip2 v16.16b, v1.16b, v0.16b\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "zip1 v19.16b, v31.16b, v30.16b\n"
+      "zip2 v18.16b, v31.16b, v30.16b\n"
+      "str q16, [x21, #0x30]\n"
+      "zip1 v17.16b, v29.16b, v23.16b\n"
+      "zip2 v16.16b, v29.16b, v23.16b\n"
+      "str q19, [x21, #0x80]\n"
+      "zip1 v23.16b, v22.16b, v21.16b\n"
+      "zip2 v22.16b, v22.16b, v21.16b\n"
+      "str q18, [x21, #0x90]\n"
+      "zip1 v21.16b, v28.16b, v20.16b\n"
+      "zip2 v20.16b, v28.16b, v20.16b\n"
+      "str q17, [x21, #0xa0]\n"
+      "zip1 v19.16b, v27.16b, v26.16b\n"
+      "zip2 v18.16b, v27.16b, v26.16b\n"
+      "str q16, [x21, #0xb0]\n"
+      "zip1 v17.16b, v25.16b, v24.16b\n"
+      "zip2 v16.16b, v25.16b, v24.16b\n"
+      "str q23, [x21, #0x100]\n"
+      "str q22, [x21, #0x110]\n"
+      "str q21, [x21, #0x120]\n"
+      "str q20, [x21, #0x130]\n"
+      "str q19, [x21, #0x180]\n"
+      "str q18, [x21, #0x190]\n"
+      "str q17, [x21, #0x1a0]\n"
+      "str q16, [x21, #0x1b0]\n"
+      "add x21, x21, #0x40\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 16 loop: skip
-      "cmp x16, #0x4\n"
+      "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
-      "ldr s23, [x17], #0x4\n"
-      "ldr s21, [x14], #0x4\n"
-      "sub x16, x16, #0x4\n"
-      "ldr s20, [x13], #0x4\n"
-      "ldr s19, [x12], #0x4\n"
-      "cmp x16, #0x4\n"
-      "ldr s22, [x11], #0x4\n"
-      "ldr s18, [x10], #0x4\n"
-      "ldr s17, [x9], #0x4\n"
-      "ldr s16, [x28], #0x4\n"
-      "ldr s27, [x27], #0x4\n"
-      "ldr s26, [x26], #0x4\n"
-      "zip1 v25.16b, v23.16b, v20.16b\n"
-      "zip1 v21.16b, v21.16b, v19.16b\n"
+      "ldr s19, [x17], #0x4\n"
+      "ldr s18, [x16], #0x4\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr s17, [x15], #0x4\n"
+      "ldr s16, [x14], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr s19, [x13], #0x4\n"
+      "ldr s18, [x12], #0x4\n"
+      "zip1 v22.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x11], #0x4\n"
+      "ldr s16, [x10], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr s19, [x9], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
+      "zip1 v21.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
       "ldr s20, [x25], #0x4\n"
-      "ldr s19, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s23, [x22], #0x4\n"
-      "zip1 v22.16b, v22.16b, v17.16b\n"
-      "zip1 v17.16b, v18.16b, v16.16b\n"
-      "ldr s18, [x21], #0x4\n"
+      "ldr s19, [x23], #0x4\n"
+      "zip1 v18.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x22], #0x4\n"
       "ldr s16, [x20], #0x4\n"
-      "zip1 v21.16b, v25.16b, v21.16b\n"
-      "zip1 v20.16b, v27.16b, v20.16b\n"
-      "zip1 v19.16b, v26.16b, v19.16b\n"
-      "zip1 v17.16b, v22.16b, v17.16b\n"
-      "zip1 v18.16b, v24.16b, v18.16b\n"
-      "zip1 v16.16b, v23.16b, v16.16b\n"
-      "str q21, [x15, #0x0]\n"
-      "str q17, [x15, #0x80]\n"
-      "zip1 v17.16b, v20.16b, v19.16b\n"
-      "zip1 v16.16b, v18.16b, v16.16b\n"
-      "str q17, [x15, #0x100]\n"
-      "str q16, [x15, #0x180]\n"
-      "add x15, x15, #0x10\n"
+      "zip1 v17.16b, v20.16b, v17.16b\n"
+      "zip1 v16.16b, v19.16b, v16.16b\n"
+      "str q22, [x21, #0x0]\n"
+      "zip1 v16.16b, v17.16b, v16.16b\n"
+      "str q21, [x21, #0x80]\n"
+      "str q18, [x21, #0x100]\n"
+      "str q16, [x21, #0x180]\n"
+      "add x21, x21, #0x10\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
-      "cmp x16, #0x1\n"
+      "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
-      "ldr b23, [x17], #0x1\n"
-      "ldr b21, [x14], #0x1\n"
-      "sub x16, x16, #0x1\n"
-      "ldr b20, [x13], #0x1\n"
-      "ldr b19, [x12], #0x1\n"
-      "cmp x16, #0x1\n"
-      "ldr b22, [x11], #0x1\n"
-      "ldr b18, [x10], #0x1\n"
-      "ldr b17, [x9], #0x1\n"
-      "ldr b16, [x28], #0x1\n"
-      "ldr b27, [x27], #0x1\n"
-      "ldr b26, [x26], #0x1\n"
-      "zip1 v25.16b, v23.16b, v20.16b\n"
-      "zip1 v21.16b, v21.16b, v19.16b\n"
+      "ldr b19, [x17], #0x1\n"
+      "ldr b18, [x16], #0x1\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr b17, [x15], #0x1\n"
+      "ldr b16, [x14], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr b19, [x13], #0x1\n"
+      "ldr b18, [x12], #0x1\n"
+      "zip1 v22.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x11], #0x1\n"
+      "ldr b16, [x10], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr b19, [x9], #0x1\n"
+      "ldr b18, [x28], #0x1\n"
+      "zip1 v21.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x27], #0x1\n"
+      "ldr b16, [x26], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
       "ldr b20, [x25], #0x1\n"
-      "ldr b19, [x24], #0x1\n"
-      "ldr b24, [x23], #0x1\n"
-      "ldr b23, [x22], #0x1\n"
-      "zip1 v22.16b, v22.16b, v17.16b\n"
-      "zip1 v17.16b, v18.16b, v16.16b\n"
-      "ldr b18, [x21], #0x1\n"
+      "ldr b19, [x23], #0x1\n"
+      "zip1 v18.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x22], #0x1\n"
       "ldr b16, [x20], #0x1\n"
-      "zip1 v21.16b, v25.16b, v21.16b\n"
-      "zip1 v20.16b, v27.16b, v20.16b\n"
-      "zip1 v19.16b, v26.16b, v19.16b\n"
-      "zip1 v17.16b, v22.16b, v17.16b\n"
-      "zip1 v18.16b, v24.16b, v18.16b\n"
-      "zip1 v16.16b, v23.16b, v16.16b\n"
-      "str s21, [x15, #0x0]\n"
-      "str s17, [x15, #0x80]\n"
-      "zip1 v17.16b, v20.16b, v19.16b\n"
-      "zip1 v16.16b, v18.16b, v16.16b\n"
-      "str s17, [x15, #0x100]\n"
-      "str s16, [x15, #0x180]\n"
-      "add x15, x15, #0x4\n"
+      "zip1 v17.16b, v20.16b, v17.16b\n"
+      "zip1 v16.16b, v19.16b, v16.16b\n"
+      "str s22, [x21, #0x0]\n"
+      "zip1 v16.16b, v17.16b, v16.16b\n"
+      "str s21, [x21, #0x80]\n"
+      "str s18, [x21, #0x100]\n"
+      "str s16, [x21, #0x180]\n"
+      "add x21, x21, #0x4\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x10\n"
       "add %x[out], %x[out], #0x200\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x17, %x[in]\n"
+      "add x16, x17, %x[in_stride]\n"
+      "add x15, x16, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x14, x15, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x15, %x[out]\n"
-      "add x14, x17, %x[in_stride]\n"
-      "add x13, x14, %x[in_stride]\n"
-      "add x12, x13, %x[in_stride]\n"
-      "csel x13, x13, %x[pad_row], GE\n"
-      "add %x[in], x12, %x[in_stride]\n"
-      "csel x12, x12, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x14, %x[in_stride]\n"
       "csel x14, x14, %x[pad_row], GT\n"
+      "csel x15, x15, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x16, x16, %x[pad_row], GT\n"
       "cmp x20, #0x20\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Column loop
-      "ldr q20, [x17], #0x10\n"
-      "ldr q19, [x14], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Column loop
+      "ldr q19, [x17], #0x10\n"
+      "ldr q18, [x16], #0x10\n"
       "sub x20, x20, #0x20\n"
-      "ldr q18, [x13], #0x10\n"
-      "ldr q17, [x12], #0x10\n"
       "cmp x20, #0x20\n"
-      "ldr q24, [x17], #0x10\n"
-      "ldr q25, [x14], #0x10\n"
-      "ldr q23, [x13], #0x10\n"
-      "ldr q16, [x12], #0x10\n"
-      "zip1 v22.16b, v20.16b, v18.16b\n"
-      "zip1 v21.16b, v19.16b, v17.16b\n"
-      "zip2 v20.16b, v20.16b, v18.16b\n"
-      "zip2 v19.16b, v19.16b, v17.16b\n"
-      "zip1 v18.16b, v24.16b, v23.16b\n"
-      "zip1 v17.16b, v25.16b, v16.16b\n"
-      "zip2 v24.16b, v24.16b, v23.16b\n"
-      "zip2 v16.16b, v25.16b, v16.16b\n"
-      "zip1 v23.16b, v22.16b, v21.16b\n"
-      "zip2 v22.16b, v22.16b, v21.16b\n"
-      "zip1 v21.16b, v20.16b, v19.16b\n"
+      "ldr q17, [x15], #0x10\n"
+      "ldr q16, [x14], #0x10\n"
+      "zip1 v25.16b, v19.16b, v17.16b\n"
+      "zip1 v24.16b, v18.16b, v16.16b\n"
+      "ldr q22, [x17], #0x10\n"
+      "ldr q21, [x16], #0x10\n"
+      "zip2 v20.16b, v19.16b, v17.16b\n"
+      "zip2 v19.16b, v18.16b, v16.16b\n"
+      "ldr q17, [x15], #0x10\n"
+      "ldr q16, [x14], #0x10\n"
+      "zip1 v23.16b, v22.16b, v17.16b\n"
+      "zip1 v18.16b, v21.16b, v16.16b\n"
+      "zip2 v22.16b, v22.16b, v17.16b\n"
+      "zip2 v21.16b, v21.16b, v16.16b\n"
+      "zip1 v16.16b, v25.16b, v24.16b\n"
+      "zip2 v17.16b, v25.16b, v24.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v16.16b, v20.16b, v19.16b\n"
       "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip1 v19.16b, v18.16b, v17.16b\n"
-      "zip2 v18.16b, v18.16b, v17.16b\n"
-      "zip1 v17.16b, v24.16b, v16.16b\n"
-      "zip2 v16.16b, v24.16b, v16.16b\n"
-      "str q23, [x15, #0x0]\n"
-      "str q22, [x15, #0x10]\n"
-      "str q21, [x15, #0x20]\n"
-      "str q20, [x15, #0x30]\n"
-      "str q19, [x15, #0x40]\n"
-      "str q18, [x15, #0x50]\n"
-      "str q17, [x15, #0x60]\n"
-      "str q16, [x15, #0x70]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q17, [x21, #0x10]\n"
+      "zip1 v19.16b, v23.16b, v18.16b\n"
+      "zip2 v18.16b, v23.16b, v18.16b\n"
+      "str q16, [x21, #0x20]\n"
+      "zip1 v17.16b, v22.16b, v21.16b\n"
+      "zip2 v16.16b, v22.16b, v21.16b\n"
+      "str q20, [x21, #0x30]\n"
+      "str q19, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Column loop skip
       "cmp x20, #0x10\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x15, #0x0]\n"
-      "str q16, [x15, #0x10]\n"
-      "str q16, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "str q16, [x15, #0x40]\n"
-      "str q16, [x15, #0x50]\n"
-      "str q16, [x15, #0x60]\n"
-      "str q16, [x15, #0x70]\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 16 loop: loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 16 loop: loop
       "ldr q20, [x17], #0x10\n"
-      "ldr q21, [x14], #0x10\n"
+      "ldr q21, [x16], #0x10\n"
       "sub x20, x20, #0x10\n"
-      "ldr q19, [x13], #0x10\n"
-      "ldr q16, [x12], #0x10\n"
       "cmp x20, #0x10\n"
+      "ldr q19, [x15], #0x10\n"
+      "ldr q16, [x14], #0x10\n"
       "zip1 v18.16b, v20.16b, v19.16b\n"
       "zip1 v17.16b, v21.16b, v16.16b\n"
       "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip2 v16.16b, v21.16b, v16.16b\n"
-      "zip1 v19.16b, v18.16b, v17.16b\n"
+      "zip2 v19.16b, v21.16b, v16.16b\n"
+      "zip1 v16.16b, v18.16b, v17.16b\n"
       "zip2 v18.16b, v18.16b, v17.16b\n"
-      "zip1 v17.16b, v20.16b, v16.16b\n"
-      "zip2 v16.16b, v20.16b, v16.16b\n"
-      "str q19, [x15, #0x0]\n"
-      "str q18, [x15, #0x10]\n"
-      "str q17, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "add x15, x15, #0x40\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 16 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "zip1 v17.16b, v20.16b, v19.16b\n"
+      "zip2 v16.16b, v20.16b, v19.16b\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, #0x40\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 16 loop: skip
       "cmp x20, #0x4\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr s19, [x17], #0x4\n"
-      "ldr s18, [x14], #0x4\n"
+      "ldr s18, [x16], #0x4\n"
       "sub x20, x20, #0x4\n"
-      "ldr s17, [x13], #0x4\n"
-      "ldr s16, [x12], #0x4\n"
       "cmp x20, #0x4\n"
+      "ldr s17, [x15], #0x4\n"
+      "ldr s16, [x14], #0x4\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str q16, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr b19, [x17], #0x1\n"
-      "ldr b18, [x14], #0x1\n"
+      "ldr b18, [x16], #0x1\n"
       "sub x20, x20, #0x1\n"
-      "ldr b17, [x13], #0x1\n"
-      "ldr b16, [x12], #0x1\n"
       "cmp x20, #0x1\n"
+      "ldr b17, [x15], #0x1\n"
+      "ldr b16, [x14], #0x1\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str s16, [x15, #0x0]\n"
-      "add x15, x15, #0x4\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x80\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_32_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_32_2x2.hpp
index f2c6b692a7..747d4538bd 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_32_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_32_2x2.hpp
@@ -40,422 +40,392 @@ void a64_transpose_interleave_32_2x2(uint16_t *out, const uint16_t *in, size_t w
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 13f\n"
+      "blt 12f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x40\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q8, [x25], #0x10\n"
-      "ldr q6, [x22], #0x10\n"
+      "ldr q14, [x25], #0x10\n"
+      "ldr q10, [x23], #0x10\n"
       "sub x24, x24, #0x40\n"
-      "ldr q2, [x21], #0x10\n"
+      "zip1 v12.8h, v14.8h, v10.8h\n"
+      "ldr q5, [x22], #0x10\n"
+      "ldr q3, [x20], #0x10\n"
+      "zip2 v31.8h, v14.8h, v10.8h\n"
+      "zip1 v19.8h, v5.8h, v3.8h\n"
+      "ldr q27, [x25], #0x10\n"
+      "ldr q25, [x23], #0x10\n"
+      "zip1 v11.8h, v27.8h, v25.8h\n"
+      "zip2 v24.8h, v27.8h, v25.8h\n"
+      "ldr q6, [x22], #0x10\n"
+      "ldr q29, [x20], #0x10\n"
+      "zip2 v15.8h, v5.8h, v3.8h\n"
+      "zip1 v18.8h, v6.8h, v29.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q9, [x23], #0x10\n"
+      "zip1 v0.8h, v17.8h, v9.8h\n"
+      "zip2 v9.8h, v17.8h, v9.8h\n"
+      "ldr q21, [x22], #0x10\n"
       "ldr q20, [x20], #0x10\n"
+      "zip2 v8.8h, v6.8h, v29.8h\n"
+      "zip1 v30.8h, v21.8h, v20.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q5, [x23], #0x10\n"
+      "zip1 v13.8h, v17.8h, v5.8h\n"
+      "zip2 v25.8h, v17.8h, v5.8h\n"
+      "ldr q7, [x22], #0x10\n"
+      "ldr q29, [x20], #0x10\n"
+      "zip2 v27.8h, v21.8h, v20.8h\n"
+      "zip1 v14.8h, v7.8h, v29.8h\n"
+      "ldr q28, [x25], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
+      "zip2 v1.8h, v7.8h, v29.8h\n"
       "cmp x24, #0x40\n"
+      "ldr q10, [x22], #0x10\n"
+      "ldr q21, [x20], #0x10\n"
+      "zip1 v16.8h, v28.8h, v17.8h\n"
+      "zip2 v17.8h, v28.8h, v17.8h\n"
+      "ldr q5, [x25], #0x10\n"
+      "ldr q20, [x23], #0x10\n"
+      "zip1 v3.8h, v5.8h, v20.8h\n"
+      "zip2 v7.8h, v5.8h, v20.8h\n"
+      "ldr q22, [x22], #0x10\n"
+      "ldr q29, [x20], #0x10\n"
+      "zip1 v2.8h, v10.8h, v21.8h\n"
+      "zip2 v5.8h, v10.8h, v21.8h\n"
       "ldr q21, [x25], #0x10\n"
-      "ldr q28, [x22], #0x10\n"
-      "ldr q23, [x21], #0x10\n"
-      "zip1 v11.8h, v8.8h, v6.8h\n"
-      "zip2 v8.8h, v8.8h, v6.8h\n"
-      "ldr q1, [x20], #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "zip1 v12.8h, v2.8h, v20.8h\n"
-      "zip2 v14.8h, v2.8h, v20.8h\n"
-      "ldr q24, [x22], #0x10\n"
-      "ldr q20, [x21], #0x10\n"
-      "zip1 v13.8h, v21.8h, v28.8h\n"
-      "zip2 v15.8h, v21.8h, v28.8h\n"
-      "ldr q17, [x20], #0x10\n"
-      "ldr q19, [x25], #0x10\n"
-      "zip1 v6.8h, v23.8h, v1.8h\n"
-      "zip2 v28.8h, v23.8h, v1.8h\n"
-      "ldr q16, [x22], #0x10\n"
-      "ldr q30, [x21], #0x10\n"
-      "zip1 v7.8h, v18.8h, v24.8h\n"
-      "zip2 v9.8h, v18.8h, v24.8h\n"
-      "ldr q18, [x20], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
-      "zip1 v5.8h, v20.8h, v17.8h\n"
-      "zip2 v3.8h, v20.8h, v17.8h\n"
-      "ldr q23, [x22], #0x10\n"
-      "ldr q26, [x21], #0x10\n"
-      "zip1 v25.8h, v19.8h, v16.8h\n"
-      "zip2 v10.8h, v19.8h, v16.8h\n"
-      "ldr q16, [x20], #0x10\n"
-      "ldr q19, [x25], #0x10\n"
-      "zip1 v2.8h, v30.8h, v18.8h\n"
-      "zip2 v4.8h, v30.8h, v18.8h\n"
+      "ldr q20, [x23], #0x10\n"
+      "zip1 v4.8h, v21.8h, v20.8h\n"
+      "zip2 v28.8h, v21.8h, v20.8h\n"
+      "ldr q6, [x22], #0x10\n"
+      "ldr q10, [x20], #0x10\n"
+      "zip1 v26.8h, v22.8h, v29.8h\n"
+      "zip2 v20.8h, v22.8h, v29.8h\n"
+      "ldr q29, [x25], #0x10\n"
+      "ldr q23, [x23], #0x10\n"
+      "zip1 v21.8h, v29.8h, v23.8h\n"
+      "zip2 v23.8h, v29.8h, v23.8h\n"
       "ldr q22, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
-      "zip1 v27.8h, v21.8h, v23.8h\n"
-      "zip2 v20.8h, v21.8h, v23.8h\n"
-      "ldr q1, [x20], #0x10\n"
-      "ldr q23, [x25], #0x10\n"
-      "zip1 v0.8h, v26.8h, v16.8h\n"
-      "zip2 v31.8h, v26.8h, v16.8h\n"
-      "ldr q16, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
-      "zip1 v30.8h, v19.8h, v22.8h\n"
-      "zip2 v29.8h, v19.8h, v22.8h\n"
-      "ldr q19, [x20], #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "zip1 v24.8h, v17.8h, v1.8h\n"
-      "zip2 v26.8h, v17.8h, v1.8h\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q1, [x21], #0x10\n"
-      "zip1 v22.8h, v23.8h, v16.8h\n"
-      "zip2 v23.8h, v23.8h, v16.8h\n"
-      "ldr q16, [x20], #0x10\n"
-      "str q11, [x23, #0x0]\n"
-      "zip1 v11.8h, v21.8h, v19.8h\n"
-      "zip2 v21.8h, v21.8h, v19.8h\n"
-      "str q8, [x23, #0x10]\n"
-      "zip1 v19.8h, v18.8h, v17.8h\n"
-      "zip2 v18.8h, v18.8h, v17.8h\n"
-      "str q13, [x23, #0x20]\n"
-      "zip1 v17.8h, v1.8h, v16.8h\n"
-      "zip2 v16.8h, v1.8h, v16.8h\n"
-      "str q15, [x23, #0x30]\n"
-      "str q7, [x23, #0x40]\n"
-      "str q9, [x23, #0x50]\n"
-      "str q25, [x23, #0x60]\n"
-      "str q10, [x23, #0x70]\n"
-      "str q12, [x23, #0x80]\n"
-      "str q14, [x23, #0x90]\n"
-      "str q6, [x23, #0xa0]\n"
-      "str q28, [x23, #0xb0]\n"
-      "str q5, [x23, #0xc0]\n"
-      "str q3, [x23, #0xd0]\n"
-      "str q2, [x23, #0xe0]\n"
-      "str q4, [x23, #0xf0]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q27, [x23, #0x0]\n"
-      "str q20, [x23, #0x10]\n"
-      "str q30, [x23, #0x20]\n"
-      "str q29, [x23, #0x30]\n"
-      "str q22, [x23, #0x40]\n"
-      "str q23, [x23, #0x50]\n"
-      "str q19, [x23, #0x60]\n"
-      "str q18, [x23, #0x70]\n"
-      "str q0, [x23, #0x80]\n"
-      "str q31, [x23, #0x90]\n"
-      "str q24, [x23, #0xa0]\n"
-      "str q26, [x23, #0xb0]\n"
-      "str q11, [x23, #0xc0]\n"
-      "str q21, [x23, #0xd0]\n"
-      "str q17, [x23, #0xe0]\n"
-      "str q16, [x23, #0xf0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "ldr q29, [x20], #0x10\n"
+      "str q12, [x21, #0x0]\n"
+      "zip1 v12.8h, v6.8h, v10.8h\n"
+      "str q31, [x21, #0x10]\n"
+      "zip2 v6.8h, v6.8h, v10.8h\n"
+      "zip1 v31.8h, v22.8h, v29.8h\n"
+      "str q11, [x21, #0x20]\n"
+      "zip2 v11.8h, v22.8h, v29.8h\n"
+      "str q24, [x21, #0x30]\n"
+      "str q0, [x21, #0x40]\n"
+      "str q9, [x21, #0x50]\n"
+      "str q13, [x21, #0x60]\n"
+      "str q25, [x21, #0x70]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q15, [x21, #0x90]\n"
+      "str q18, [x21, #0xa0]\n"
+      "str q8, [x21, #0xb0]\n"
+      "str q30, [x21, #0xc0]\n"
+      "str q27, [x21, #0xd0]\n"
+      "str q14, [x21, #0xe0]\n"
+      "str q1, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q16, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "str q3, [x21, #0x20]\n"
+      "str q7, [x21, #0x30]\n"
+      "str q4, [x21, #0x40]\n"
+      "str q28, [x21, #0x50]\n"
+      "str q21, [x21, #0x60]\n"
+      "str q23, [x21, #0x70]\n"
+      "str q2, [x21, #0x80]\n"
+      "str q5, [x21, #0x90]\n"
+      "str q26, [x21, #0xa0]\n"
+      "str q20, [x21, #0xb0]\n"
+      "str q12, [x21, #0xc0]\n"
+      "str q6, [x21, #0xd0]\n"
+      "str q31, [x21, #0xe0]\n"
+      "str q11, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cmp x24, #0x20\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr q21, [x25], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
       "sub x24, x24, #0x20\n"
-      "ldr q20, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
       "cmp x24, #0x20\n"
-      "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q22, [x21], #0x10\n"
-      "zip1 v0.8h, v21.8h, v17.8h\n"
-      "zip2 v31.8h, v21.8h, v17.8h\n"
-      "ldr q17, [x20], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
-      "zip1 v30.8h, v20.8h, v16.8h\n"
-      "zip2 v29.8h, v20.8h, v16.8h\n"
-      "ldr q16, [x22], #0x10\n"
-      "ldr q20, [x21], #0x10\n"
-      "zip1 v28.8h, v19.8h, v18.8h\n"
-      "zip2 v27.8h, v19.8h, v18.8h\n"
+      "ldr q21, [x22], #0x10\n"
+      "ldr q18, [x20], #0x10\n"
+      "zip1 v1.8h, v17.8h, v16.8h\n"
+      "zip2 v0.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip1 v31.8h, v17.8h, v16.8h\n"
+      "zip2 v30.8h, v17.8h, v16.8h\n"
+      "ldr q20, [x22], #0x10\n"
       "ldr q19, [x20], #0x10\n"
-      "ldr q18, [x25], #0x10\n"
-      "zip1 v26.8h, v22.8h, v17.8h\n"
-      "zip2 v25.8h, v22.8h, v17.8h\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q24, [x21], #0x10\n"
-      "zip1 v23.8h, v21.8h, v16.8h\n"
-      "zip2 v22.8h, v21.8h, v16.8h\n"
+      "zip1 v29.8h, v21.8h, v18.8h\n"
+      "zip2 v28.8h, v21.8h, v18.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip1 v27.8h, v17.8h, v16.8h\n"
+      "zip2 v26.8h, v17.8h, v16.8h\n"
+      "ldr q25, [x22], #0x10\n"
+      "ldr q18, [x20], #0x10\n"
+      "zip1 v24.8h, v20.8h, v19.8h\n"
+      "zip2 v23.8h, v20.8h, v19.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip1 v22.8h, v17.8h, v16.8h\n"
+      "zip2 v21.8h, v17.8h, v16.8h\n"
+      "ldr q20, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "str q0, [x23, #0x0]\n"
-      "zip1 v19.8h, v18.8h, v17.8h\n"
-      "zip2 v18.8h, v18.8h, v17.8h\n"
-      "str q31, [x23, #0x10]\n"
-      "zip1 v17.8h, v24.8h, v16.8h\n"
-      "zip2 v16.8h, v24.8h, v16.8h\n"
-      "str q28, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q23, [x23, #0x40]\n"
-      "str q22, [x23, #0x50]\n"
-      "str q19, [x23, #0x60]\n"
-      "str q18, [x23, #0x70]\n"
-      "str q30, [x23, #0x80]\n"
-      "str q29, [x23, #0x90]\n"
-      "str q26, [x23, #0xa0]\n"
-      "str q25, [x23, #0xb0]\n"
-      "str q21, [x23, #0xc0]\n"
-      "str q20, [x23, #0xd0]\n"
-      "str q17, [x23, #0xe0]\n"
-      "str q16, [x23, #0xf0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "zip1 v19.8h, v25.8h, v18.8h\n"
+      "zip2 v18.8h, v25.8h, v18.8h\n"
+      "zip1 v17.8h, v20.8h, v16.8h\n"
+      "zip2 v16.8h, v20.8h, v16.8h\n"
+      "str q1, [x21, #0x0]\n"
+      "str q0, [x21, #0x10]\n"
+      "str q31, [x21, #0x20]\n"
+      "str q30, [x21, #0x30]\n"
+      "str q27, [x21, #0x40]\n"
+      "str q26, [x21, #0x50]\n"
+      "str q22, [x21, #0x60]\n"
+      "str q21, [x21, #0x70]\n"
+      "str q29, [x21, #0x80]\n"
+      "str q28, [x21, #0x90]\n"
+      "str q24, [x21, #0xa0]\n"
+      "str q23, [x21, #0xb0]\n"
+      "str q19, [x21, #0xc0]\n"
+      "str q18, [x21, #0xd0]\n"
+      "str q17, [x21, #0xe0]\n"
+      "str q16, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x24, 12f\n"
       "cmp x24, #0x10\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "str q16, [x23, #0x80]\n"
-      "str q16, [x23, #0x90]\n"
-      "str q16, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "str q16, [x23, #0xc0]\n"
-      "str q16, [x23, #0xd0]\n"
-      "str q16, [x23, #0xe0]\n"
-      "str q16, [x23, #0xf0]\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 16 loop: loop
       "ldr q17, [x25], #0x10\n"
-      "ldr q16, [x22], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
       "sub x24, x24, #0x10\n"
-      "ldr q20, [x21], #0x10\n"
-      "ldr q19, [x20], #0x10\n"
       "cmp x24, #0x10\n"
-      "ldr q24, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q23, [x21], #0x10\n"
+      "ldr q24, [x22], #0x10\n"
+      "ldr q23, [x20], #0x10\n"
+      "zip1 v19.8h, v17.8h, v16.8h\n"
+      "zip2 v18.8h, v17.8h, v16.8h\n"
+      "ldr q17, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
       "zip1 v22.8h, v17.8h, v16.8h\n"
-      "zip2 v17.8h, v17.8h, v16.8h\n"
+      "zip2 v21.8h, v17.8h, v16.8h\n"
+      "ldr q20, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "zip1 v19.8h, v24.8h, v18.8h\n"
-      "zip2 v18.8h, v24.8h, v18.8h\n"
-      "str q22, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "zip1 v17.8h, v23.8h, v16.8h\n"
-      "zip2 v16.8h, v23.8h, v16.8h\n"
-      "str q19, [x23, #0x20]\n"
-      "str q18, [x23, #0x30]\n"
-      "str q21, [x23, #0x80]\n"
-      "str q20, [x23, #0x90]\n"
-      "str q17, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "add x23, x23, #0x40\n"
+      "str q19, [x21, #0x0]\n"
+      "zip1 v19.8h, v24.8h, v23.8h\n"
+      "str q18, [x21, #0x10]\n"
+      "zip2 v18.8h, v24.8h, v23.8h\n"
+      "zip1 v17.8h, v20.8h, v16.8h\n"
+      "str q22, [x21, #0x20]\n"
+      "zip2 v16.8h, v20.8h, v16.8h\n"
+      "str q21, [x21, #0x30]\n"
+      "str q19, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q17, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, #0x40\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 16 loop: skip
       "cmp x24, #0x4\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 4 loop: loop
       "ldr d19, [x25], #0x8\n"
-      "ldr d16, [x22], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "sub x24, x24, #0x4\n"
-      "ldr d18, [x21], #0x8\n"
-      "ldr d17, [x20], #0x8\n"
       "cmp x24, #0x4\n"
+      "ldr d18, [x22], #0x8\n"
+      "ldr d17, [x20], #0x8\n"
       "zip1 v16.8h, v19.8h, v16.8h\n"
-      "str q16, [x23, #0x0]\n"
+      "str q16, [x21, #0x0]\n"
       "zip1 v16.8h, v18.8h, v17.8h\n"
-      "str q16, [x23, #0x80]\n"
-      "add x23, x23, #0x10\n"
+      "str q16, [x21, #0x80]\n"
+      "add x21, x21, #0x10\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 11f\n"
       "10:"  // Main row loop: width 1 loop: loop
       "ldr h19, [x25], #0x2\n"
-      "ldr h16, [x22], #0x2\n"
+      "ldr h16, [x23], #0x2\n"
       "sub x24, x24, #0x1\n"
-      "ldr h18, [x21], #0x2\n"
-      "ldr h17, [x20], #0x2\n"
       "cmp x24, #0x1\n"
+      "ldr h18, [x22], #0x2\n"
+      "ldr h17, [x20], #0x2\n"
       "zip1 v16.8h, v19.8h, v16.8h\n"
-      "str s16, [x23, #0x0]\n"
+      "str s16, [x21, #0x0]\n"
       "zip1 v16.8h, v18.8h, v17.8h\n"
-      "str s16, [x23, #0x80]\n"
-      "add x23, x23, #0x4\n"
+      "str s16, [x21, #0x80]\n"
+      "add x21, x21, #0x4\n"
       "bge 10b\n"
       "11:"  // Main row loop: width 1 loop: skip
-      "12:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0x100\n"
       "bge 1b\n"
-      "cbz %x[height], 26f\n"
-      "13:"  // Main loop skip
-      "14:"  // Tail row loop: Head
+      "cbz %x[height], 24f\n"
+      "12:"  // Main loop skip
+      "13:"  // Tail row loop: Head
       "mov x25, %x[in]\n"
       "mov x20, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x2\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
-      "csel x22, x22, %x[pad_row], GT\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "csel x23, x23, %x[pad_row], GT\n"
       "cmp x20, #0x40\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: Unroll column loop
-      "ldr q21, [x25], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x2\n"
+      "blt 15f\n"
+      "14:"  // Tail row loop: Unroll column loop
+      "ldr q18, [x25], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
       "sub x20, x20, #0x40\n"
-      "ldr q20, [x25], #0x10\n"
+      "zip1 v0.8h, v18.8h, v17.8h\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip2 v31.8h, v18.8h, v17.8h\n"
+      "zip1 v30.8h, v19.8h, v16.8h\n"
+      "ldr q18, [x25], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
+      "zip2 v29.8h, v19.8h, v16.8h\n"
+      "zip1 v28.8h, v18.8h, v17.8h\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip2 v27.8h, v18.8h, v17.8h\n"
+      "zip1 v26.8h, v19.8h, v16.8h\n"
+      "ldr q18, [x25], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
+      "zip2 v25.8h, v19.8h, v16.8h\n"
       "cmp x20, #0x40\n"
-      "ldr q16, [x22], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q22, [x25], #0x10\n"
-      "zip1 v0.8h, v21.8h, v17.8h\n"
-      "zip2 v31.8h, v21.8h, v17.8h\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q21, [x25], #0x10\n"
-      "zip1 v30.8h, v20.8h, v16.8h\n"
-      "zip2 v29.8h, v20.8h, v16.8h\n"
-      "ldr q16, [x22], #0x10\n"
-      "ldr q20, [x25], #0x10\n"
-      "zip1 v28.8h, v19.8h, v18.8h\n"
-      "zip2 v27.8h, v19.8h, v18.8h\n"
-      "ldr q19, [x22], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip1 v24.8h, v18.8h, v17.8h\n"
+      "zip2 v23.8h, v18.8h, v17.8h\n"
       "ldr q18, [x25], #0x10\n"
-      "zip1 v26.8h, v22.8h, v17.8h\n"
-      "zip2 v25.8h, v22.8h, v17.8h\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
-      "zip1 v23.8h, v21.8h, v16.8h\n"
-      "zip2 v22.8h, v21.8h, v16.8h\n"
-      "ldr q16, [x22], #0x10\n"
-      "str q0, [x23, #0x0]\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
-      "str q31, [x23, #0x10]\n"
+      "ldr q17, [x23], #0x10\n"
+      "zip1 v22.8h, v19.8h, v16.8h\n"
+      "zip2 v21.8h, v19.8h, v16.8h\n"
+      "ldr q20, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "str q0, [x21, #0x0]\n"
       "zip1 v19.8h, v18.8h, v17.8h\n"
+      "str q31, [x21, #0x10]\n"
       "zip2 v18.8h, v18.8h, v17.8h\n"
-      "str q30, [x23, #0x20]\n"
-      "zip1 v17.8h, v24.8h, v16.8h\n"
-      "zip2 v16.8h, v24.8h, v16.8h\n"
-      "str q29, [x23, #0x30]\n"
-      "str q28, [x23, #0x40]\n"
-      "str q27, [x23, #0x50]\n"
-      "str q26, [x23, #0x60]\n"
-      "str q25, [x23, #0x70]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "str q23, [x23, #0x0]\n"
-      "str q22, [x23, #0x10]\n"
-      "str q21, [x23, #0x20]\n"
-      "str q20, [x23, #0x30]\n"
-      "str q19, [x23, #0x40]\n"
-      "str q18, [x23, #0x50]\n"
-      "str q17, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: Unroll column loop skip
+      "zip1 v17.8h, v20.8h, v16.8h\n"
+      "str q30, [x21, #0x20]\n"
+      "zip2 v16.8h, v20.8h, v16.8h\n"
+      "str q29, [x21, #0x30]\n"
+      "str q28, [x21, #0x40]\n"
+      "str q27, [x21, #0x50]\n"
+      "str q26, [x21, #0x60]\n"
+      "str q25, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q24, [x21, #0x0]\n"
+      "str q23, [x21, #0x10]\n"
+      "str q22, [x21, #0x20]\n"
+      "str q21, [x21, #0x30]\n"
+      "str q19, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0x20\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: Column loop
-      "ldr q21, [x25], #0x10\n"
-      "ldr q16, [x22], #0x10\n"
+      "blt 17f\n"
+      "16:"  // Tail row loop: Column loop
+      "ldr q18, [x25], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
       "sub x20, x20, #0x20\n"
-      "ldr q20, [x25], #0x10\n"
       "cmp x20, #0x20\n"
-      "ldr q19, [x22], #0x10\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
+      "zip1 v24.8h, v18.8h, v17.8h\n"
+      "zip2 v23.8h, v18.8h, v17.8h\n"
       "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
-      "ldr q24, [x25], #0x10\n"
-      "zip1 v23.8h, v21.8h, v16.8h\n"
-      "zip2 v22.8h, v21.8h, v16.8h\n"
-      "ldr q16, [x22], #0x10\n"
-      "zip1 v21.8h, v20.8h, v19.8h\n"
-      "zip2 v20.8h, v20.8h, v19.8h\n"
+      "ldr q17, [x23], #0x10\n"
+      "zip1 v22.8h, v19.8h, v16.8h\n"
+      "zip2 v21.8h, v19.8h, v16.8h\n"
+      "ldr q20, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
       "zip1 v19.8h, v18.8h, v17.8h\n"
       "zip2 v18.8h, v18.8h, v17.8h\n"
-      "zip1 v17.8h, v24.8h, v16.8h\n"
-      "zip2 v16.8h, v24.8h, v16.8h\n"
-      "str q23, [x23, #0x0]\n"
-      "str q22, [x23, #0x10]\n"
-      "str q21, [x23, #0x20]\n"
-      "str q20, [x23, #0x30]\n"
-      "str q19, [x23, #0x40]\n"
-      "str q18, [x23, #0x50]\n"
-      "str q17, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: Column loop skip
-      "cbz x20, 25f\n"
+      "zip1 v17.8h, v20.8h, v16.8h\n"
+      "zip2 v16.8h, v20.8h, v16.8h\n"
+      "str q24, [x21, #0x0]\n"
+      "str q23, [x21, #0x10]\n"
+      "str q22, [x21, #0x20]\n"
+      "str q21, [x21, #0x30]\n"
+      "str q19, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "str q17, [x21, #0x60]\n"
+      "str q16, [x21, #0x70]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: Column loop skip
       "cmp x20, #0x10\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 16 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 16 loop: loop
       "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
+      "ldr q17, [x23], #0x10\n"
       "sub x20, x20, #0x10\n"
-      "ldr q20, [x25], #0x10\n"
       "cmp x20, #0x10\n"
-      "ldr q16, [x22], #0x10\n"
+      "ldr q20, [x25], #0x10\n"
+      "ldr q16, [x23], #0x10\n"
       "zip1 v19.8h, v18.8h, v17.8h\n"
       "zip2 v18.8h, v18.8h, v17.8h\n"
       "zip1 v17.8h, v20.8h, v16.8h\n"
       "zip2 v16.8h, v20.8h, v16.8h\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x10]\n"
-      "str q17, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "add x23, x23, #0x40\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 16 loop: skip
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, #0x40\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 16 loop: skip
       "cmp x20, #0x4\n"
-      "blt 22f\n"
-      "21:"  // Tail row loop: width 4 loop: loop
+      "blt 21f\n"
+      "20:"  // Tail row loop: width 4 loop: loop
       "ldr d17, [x25], #0x8\n"
-      "ldr d16, [x22], #0x8\n"
+      "ldr d16, [x23], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
       "zip1 v16.8h, v17.8h, v16.8h\n"
-      "str q16, [x23, #0x0]\n"
-      "add x23, x23, #0x10\n"
-      "bge 21b\n"
-      "22:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 20b\n"
+      "21:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 24f\n"
-      "23:"  // Tail row loop: width 1 loop: loop
+      "blt 23f\n"
+      "22:"  // Tail row loop: width 1 loop: loop
       "ldr h17, [x25], #0x2\n"
-      "ldr h16, [x22], #0x2\n"
+      "ldr h16, [x23], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
       "zip1 v16.8h, v17.8h, v16.8h\n"
-      "str s16, [x23, #0x0]\n"
-      "add x23, x23, #0x4\n"
-      "bge 23b\n"
-      "24:"  // Tail row loop: width 1 loop: skip
-      "25:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 22b\n"
+      "23:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x80\n"
-      "bge 14b\n"
-      "26:"  // Done
+      "bge 13b\n"
+      "24:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_48.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_48.hpp
index 44b6e7c9d4..222551909b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_48.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_48.hpp
@@ -34,186 +34,165 @@ void a64_transpose_interleave_48(uint16_t *out, const uint16_t *in, size_t width
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
       "ldr q27, [x25], #0x10\n"
-      "ldr q26, [x22], #0x10\n"
+      "ldr q26, [x23], #0x10\n"
       "sub x24, x24, #0x18\n"
-      "ldr q25, [x21], #0x10\n"
-      "ldr q24, [x20], #0x10\n"
       "cmp x24, #0x18\n"
+      "ldr q25, [x22], #0x10\n"
+      "ldr q24, [x20], #0x10\n"
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "ldr q21, [x22], #0x10\n"
       "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q27, [x23, #0x0]\n"
-      "str q23, [x23, #0x10]\n"
-      "str q19, [x23, #0x20]\n"
-      "str q26, [x23, #0x30]\n"
-      "str q22, [x23, #0x40]\n"
-      "str q18, [x23, #0x50]\n"
-      "str q25, [x23, #0x60]\n"
-      "str q21, [x23, #0x70]\n"
-      "str q17, [x23, #0x80]\n"
-      "str q24, [x23, #0x90]\n"
-      "str q20, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "str q27, [x21, #0x0]\n"
+      "str q23, [x21, #0x10]\n"
+      "str q19, [x21, #0x20]\n"
+      "str q26, [x21, #0x30]\n"
+      "str q22, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "str q25, [x21, #0x60]\n"
+      "str q21, [x21, #0x70]\n"
+      "str q17, [x21, #0x80]\n"
+      "str q24, [x21, #0x90]\n"
+      "str q20, [x21, #0xa0]\n"
+      "str q16, [x21, #0xb0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x10\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "str q16, [x23, #0x80]\n"
-      "str q16, [x23, #0x90]\n"
-      "str q16, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 16 loop: loop
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
       "sub x24, x24, #0x10\n"
-      "ldr q21, [x21], #0x10\n"
-      "ldr q20, [x20], #0x10\n"
       "cmp x24, #0x10\n"
+      "ldr q21, [x22], #0x10\n"
+      "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q23, [x23, #0x0]\n"
-      "str q19, [x23, #0x10]\n"
-      "str q22, [x23, #0x30]\n"
-      "str q18, [x23, #0x40]\n"
-      "str q21, [x23, #0x60]\n"
-      "str q17, [x23, #0x70]\n"
-      "str q20, [x23, #0x90]\n"
-      "str q16, [x23, #0xa0]\n"
-      "add x23, x23, #0x20\n"
+      "str q23, [x21, #0x0]\n"
+      "str q19, [x21, #0x10]\n"
+      "str q22, [x21, #0x30]\n"
+      "str q18, [x21, #0x40]\n"
+      "str q21, [x21, #0x60]\n"
+      "str q17, [x21, #0x70]\n"
+      "str q20, [x21, #0x90]\n"
+      "str q16, [x21, #0xa0]\n"
+      "add x21, x21, #0x20\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 16 loop: skip
       "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr d19, [x25], #0x8\n"
-      "ldr d18, [x22], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "sub x24, x24, #0x4\n"
-      "ldr d17, [x21], #0x8\n"
-      "ldr d16, [x20], #0x8\n"
       "cmp x24, #0x4\n"
-      "str d19, [x23, #0x0]\n"
-      "str d18, [x23, #0x30]\n"
-      "str d17, [x23, #0x60]\n"
-      "str d16, [x23, #0x90]\n"
-      "add x23, x23, #0x8\n"
+      "ldr d17, [x22], #0x8\n"
+      "ldr d16, [x20], #0x8\n"
+      "str d19, [x21, #0x0]\n"
+      "str d18, [x21, #0x30]\n"
+      "str d17, [x21, #0x60]\n"
+      "str d16, [x21, #0x90]\n"
+      "add x21, x21, #0x8\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr h19, [x25], #0x2\n"
-      "ldr h18, [x22], #0x2\n"
+      "ldr h18, [x23], #0x2\n"
       "sub x24, x24, #0x1\n"
-      "ldr h17, [x21], #0x2\n"
-      "ldr h16, [x20], #0x2\n"
       "cmp x24, #0x1\n"
-      "str h19, [x23, #0x0]\n"
-      "str h18, [x23, #0x30]\n"
-      "str h17, [x23, #0x60]\n"
-      "str h16, [x23, #0x90]\n"
-      "add x23, x23, #0x2\n"
+      "ldr h17, [x22], #0x2\n"
+      "ldr h16, [x20], #0x2\n"
+      "str h19, [x21, #0x0]\n"
+      "str h18, [x21, #0x30]\n"
+      "str h17, [x21, #0x60]\n"
+      "str h16, [x21, #0x90]\n"
+      "add x21, x21, #0x2\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0xc0\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x18\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Column loop
       "ldr q18, [x25], #0x10\n"
-      "sub x20, x20, #0x18\n"
       "ldr q17, [x25], #0x10\n"
-      "ldr q16, [x25], #0x10\n"
+      "sub x20, x20, #0x18\n"
       "cmp x20, #0x18\n"
-      "str q18, [x23, #0x0]\n"
-      "str q17, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "ldr q16, [x25], #0x10\n"
+      "str q18, [x21, #0x0]\n"
+      "str q17, [x21, #0x10]\n"
+      "str q16, [x21, #0x20]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Column loop skip
       "cmp x20, #0x10\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 16 loop: loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 16 loop: loop
       "ldr q17, [x25], #0x10\n"
-      "sub x20, x20, #0x10\n"
       "ldr q16, [x25], #0x10\n"
+      "sub x20, x20, #0x10\n"
       "cmp x20, #0x10\n"
-      "str q17, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "add x23, x23, #0x20\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 16 loop: skip
+      "str q17, [x21, #0x0]\n"
+      "str q16, [x21, #0x10]\n"
+      "add x21, x21, #0x20\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 16 loop: skip
       "cmp x20, #0x4\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr d16, [x25], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
-      "str d16, [x23, #0x0]\n"
-      "add x23, x23, #0x8\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr h16, [x25], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
-      "str h16, [x23, #0x0]\n"
-      "add x23, x23, #0x2\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str h16, [x21, #0x0]\n"
+      "add x21, x21, #0x2\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x30\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
       : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "x20", "x21", "x22", "x23", "x24", "x25"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_1x16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_1x16.hpp
index ed12f5dfa9..7b9c7ecb30 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_1x16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_1x16.hpp
@@ -41,10 +41,9 @@ void a64_transpose_interleave_4_1x16(uint8_t *out, const uint8_t *in, size_t wid
     __asm__ __volatile__(
       "1:"  // Main row loop: Head
       "mov x17, %x[in]\n"
-      "cmp %x[height], #0xf\n"
-      "mov x16, %x[width]\n"
-      "mov x15, %x[out]\n"
-      "add x14, x17, %x[in_stride]\n"
+      "add x16, x17, %x[in_stride]\n"
+      "add x15, x16, %x[in_stride]\n"
+      "add x14, x15, %x[in_stride]\n"
       "add x13, x14, %x[in_stride]\n"
       "add x12, x13, %x[in_stride]\n"
       "add x11, x12, %x[in_stride]\n"
@@ -57,231 +56,228 @@ void a64_transpose_interleave_4_1x16(uint8_t *out, const uint8_t *in, size_t wid
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GE\n"
-      "add %x[in], x20, %x[in_stride]\n"
-      "csel x20, x20, %x[pad_row], GT\n"
-      "cmp %x[height], #0xd\n"
+      "cmp %x[height], #0xf\n"
+      "add %x[in], x22, %x[in_stride]\n"
       "csel x22, x22, %x[pad_row], GT\n"
       "csel x23, x23, %x[pad_row], GE\n"
-      "cmp %x[height], #0xb\n"
+      "cmp %x[height], #0xd\n"
       "csel x24, x24, %x[pad_row], GT\n"
       "csel x25, x25, %x[pad_row], GE\n"
-      "cmp %x[height], #0x9\n"
+      "cmp %x[height], #0xb\n"
       "csel x26, x26, %x[pad_row], GT\n"
       "csel x27, x27, %x[pad_row], GE\n"
-      "cmp %x[height], #0x7\n"
+      "cmp %x[height], #0x9\n"
       "csel x28, x28, %x[pad_row], GT\n"
       "csel x9, x9, %x[pad_row], GE\n"
-      "cmp %x[height], #0x5\n"
+      "cmp %x[height], #0x7\n"
       "csel x10, x10, %x[pad_row], GT\n"
       "csel x11, x11, %x[pad_row], GE\n"
-      "cmp %x[height], #0x3\n"
+      "cmp %x[height], #0x5\n"
+      "mov x21, %x[width]\n"
       "csel x12, x12, %x[pad_row], GT\n"
       "csel x13, x13, %x[pad_row], GE\n"
+      "cmp %x[height], #0x3\n"
+      "csel x14, x14, %x[pad_row], GT\n"
+      "csel x15, x15, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
+      "csel x16, x16, %x[pad_row], GT\n"
+      "cmp x21, #0x10\n"
+      "mov x20, %x[out]\n"
       "sub %x[height], %x[height], #0x10\n"
-      "csel x14, x14, %x[pad_row], GT\n"
-      "cmp x16, #0x10\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
       "ldr q3, [x17], #0x10\n"
-      "ldr q9, [x14], #0x10\n"
-      "sub x16, x16, #0x10\n"
-      "ldr q2, [x13], #0x10\n"
-      "ldr q8, [x12], #0x10\n"
-      "cmp x16, #0x10\n"
-      "ldr q1, [x11], #0x10\n"
+      "ldr q9, [x16], #0x10\n"
+      "sub x21, x21, #0x10\n"
+      "cmp x21, #0x10\n"
+      "ldr q2, [x15], #0x10\n"
+      "ldr q8, [x14], #0x10\n"
+      "ldr q0, [x13], #0x10\n"
+      "ldr q31, [x12], #0x10\n"
+      "ldr q30, [x11], #0x10\n"
       "ldr q7, [x10], #0x10\n"
-      "ldr q0, [x9], #0x10\n"
-      "ldr q6, [x28], #0x10\n"
-      "ldr q31, [x27], #0x10\n"
-      "ldr q30, [x26], #0x10\n"
-      "ldr q29, [x25], #0x10\n"
-      "ldr q28, [x24], #0x10\n"
-      "ldr q26, [x23], #0x10\n"
-      "ldr q25, [x22], #0x10\n"
-      "ldr q24, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
-      "zip1 v23.16b, v3.16b, v31.16b\n"
-      "zip1 v27.16b, v9.16b, v30.16b\n"
-      "zip1 v22.16b, v2.16b, v29.16b\n"
-      "zip1 v21.16b, v8.16b, v28.16b\n"
-      "zip1 v20.16b, v1.16b, v26.16b\n"
-      "zip1 v19.16b, v7.16b, v25.16b\n"
-      "zip1 v18.16b, v0.16b, v24.16b\n"
-      "zip1 v17.16b, v6.16b, v16.16b\n"
-      "zip2 v5.16b, v3.16b, v31.16b\n"
-      "zip2 v1.16b, v1.16b, v26.16b\n"
-      "zip2 v4.16b, v2.16b, v29.16b\n"
-      "zip2 v3.16b, v0.16b, v24.16b\n"
-      "zip2 v2.16b, v9.16b, v30.16b\n"
-      "zip2 v0.16b, v7.16b, v25.16b\n"
-      "zip2 v31.16b, v8.16b, v28.16b\n"
-      "zip2 v30.16b, v6.16b, v16.16b\n"
-      "zip1 v26.16b, v23.16b, v20.16b\n"
-      "zip1 v25.16b, v22.16b, v18.16b\n"
-      "zip1 v24.16b, v27.16b, v19.16b\n"
-      "zip1 v16.16b, v21.16b, v17.16b\n"
-      "zip2 v29.16b, v23.16b, v20.16b\n"
-      "zip2 v23.16b, v22.16b, v18.16b\n"
-      "zip2 v22.16b, v27.16b, v19.16b\n"
-      "zip2 v21.16b, v21.16b, v17.16b\n"
-      "zip1 v28.16b, v5.16b, v1.16b\n"
-      "zip1 v27.16b, v4.16b, v3.16b\n"
-      "zip1 v20.16b, v2.16b, v0.16b\n"
-      "zip1 v19.16b, v31.16b, v30.16b\n"
-      "zip1 v18.16b, v26.16b, v25.16b\n"
-      "zip1 v17.16b, v24.16b, v16.16b\n"
-      "zip2 v26.16b, v26.16b, v25.16b\n"
-      "zip2 v16.16b, v24.16b, v16.16b\n"
-      "zip2 v1.16b, v5.16b, v1.16b\n"
-      "zip2 v25.16b, v4.16b, v3.16b\n"
-      "zip2 v0.16b, v2.16b, v0.16b\n"
-      "zip2 v24.16b, v31.16b, v30.16b\n"
-      "zip1 v31.16b, v29.16b, v23.16b\n"
-      "zip1 v30.16b, v22.16b, v21.16b\n"
-      "zip2 v29.16b, v29.16b, v23.16b\n"
-      "zip2 v23.16b, v22.16b, v21.16b\n"
-      "zip1 v22.16b, v28.16b, v27.16b\n"
-      "zip1 v21.16b, v20.16b, v19.16b\n"
+      "ldr q29, [x9], #0x10\n"
+      "ldr q28, [x28], #0x10\n"
+      "zip1 v27.16b, v3.16b, v29.16b\n"
+      "zip1 v6.16b, v9.16b, v28.16b\n"
+      "ldr q25, [x27], #0x10\n"
+      "ldr q24, [x26], #0x10\n"
+      "zip1 v26.16b, v2.16b, v25.16b\n"
+      "zip1 v1.16b, v8.16b, v24.16b\n"
+      "ldr q23, [x25], #0x10\n"
+      "ldr q22, [x24], #0x10\n"
+      "zip1 v21.16b, v0.16b, v23.16b\n"
+      "zip1 v20.16b, v31.16b, v22.16b\n"
+      "ldr q19, [x23], #0x10\n"
+      "ldr q18, [x22], #0x10\n"
+      "zip1 v17.16b, v30.16b, v19.16b\n"
+      "zip1 v16.16b, v7.16b, v18.16b\n"
+      "zip2 v5.16b, v3.16b, v29.16b\n"
+      "zip2 v0.16b, v0.16b, v23.16b\n"
+      "zip2 v4.16b, v2.16b, v25.16b\n"
+      "zip2 v3.16b, v30.16b, v19.16b\n"
+      "zip2 v2.16b, v9.16b, v28.16b\n"
+      "zip2 v31.16b, v31.16b, v22.16b\n"
+      "zip2 v30.16b, v8.16b, v24.16b\n"
+      "zip2 v29.16b, v7.16b, v18.16b\n"
+      "zip1 v25.16b, v27.16b, v21.16b\n"
+      "zip1 v24.16b, v26.16b, v17.16b\n"
+      "zip1 v23.16b, v6.16b, v20.16b\n"
+      "zip1 v22.16b, v1.16b, v16.16b\n"
+      "zip2 v28.16b, v27.16b, v21.16b\n"
+      "zip2 v27.16b, v26.16b, v17.16b\n"
+      "zip2 v26.16b, v6.16b, v20.16b\n"
+      "zip2 v21.16b, v1.16b, v16.16b\n"
+      "zip1 v1.16b, v5.16b, v0.16b\n"
+      "zip1 v20.16b, v4.16b, v3.16b\n"
+      "zip1 v19.16b, v2.16b, v31.16b\n"
+      "zip1 v16.16b, v30.16b, v29.16b\n"
+      "zip1 v18.16b, v25.16b, v24.16b\n"
+      "zip1 v17.16b, v23.16b, v22.16b\n"
+      "zip2 v25.16b, v25.16b, v24.16b\n"
+      "zip2 v24.16b, v23.16b, v22.16b\n"
+      "zip2 v0.16b, v5.16b, v0.16b\n"
+      "zip2 v23.16b, v4.16b, v3.16b\n"
+      "zip2 v31.16b, v2.16b, v31.16b\n"
+      "zip2 v22.16b, v30.16b, v29.16b\n"
+      "zip1 v30.16b, v28.16b, v27.16b\n"
+      "zip1 v29.16b, v26.16b, v21.16b\n"
       "zip2 v28.16b, v28.16b, v27.16b\n"
-      "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip1 v19.16b, v18.16b, v17.16b\n"
+      "zip2 v27.16b, v26.16b, v21.16b\n"
+      "zip1 v26.16b, v1.16b, v20.16b\n"
+      "zip1 v21.16b, v19.16b, v16.16b\n"
+      "zip2 v20.16b, v1.16b, v20.16b\n"
+      "zip2 v19.16b, v19.16b, v16.16b\n"
+      "zip1 v16.16b, v18.16b, v17.16b\n"
       "zip2 v18.16b, v18.16b, v17.16b\n"
-      "zip1 v17.16b, v26.16b, v16.16b\n"
-      "zip2 v16.16b, v26.16b, v16.16b\n"
-      "zip1 v27.16b, v1.16b, v25.16b\n"
-      "zip1 v26.16b, v0.16b, v24.16b\n"
-      "str q19, [x15, #0x0]\n"
-      "str q18, [x15, #0x10]\n"
-      "zip2 v25.16b, v1.16b, v25.16b\n"
-      "zip2 v24.16b, v0.16b, v24.16b\n"
-      "str q17, [x15, #0x20]\n"
-      "zip1 v19.16b, v31.16b, v30.16b\n"
-      "zip2 v18.16b, v31.16b, v30.16b\n"
-      "str q16, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "zip1 v17.16b, v29.16b, v23.16b\n"
-      "zip2 v16.16b, v29.16b, v23.16b\n"
-      "zip1 v23.16b, v22.16b, v21.16b\n"
-      "zip2 v22.16b, v22.16b, v21.16b\n"
-      "str q19, [x15, #0x0]\n"
-      "str q18, [x15, #0x10]\n"
-      "zip1 v21.16b, v28.16b, v20.16b\n"
-      "zip2 v20.16b, v28.16b, v20.16b\n"
-      "str q17, [x15, #0x20]\n"
-      "zip1 v19.16b, v27.16b, v26.16b\n"
-      "zip2 v18.16b, v27.16b, v26.16b\n"
-      "str q16, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "str q16, [x20, #0x0]\n"
       "zip1 v17.16b, v25.16b, v24.16b\n"
       "zip2 v16.16b, v25.16b, v24.16b\n"
-      "str q23, [x15, #0x0]\n"
-      "str q22, [x15, #0x10]\n"
-      "str q21, [x15, #0x20]\n"
-      "str q20, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "str q19, [x15, #0x0]\n"
-      "str q18, [x15, #0x10]\n"
-      "str q17, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "str q18, [x20, #0x10]\n"
+      "str q17, [x20, #0x20]\n"
+      "zip1 v25.16b, v0.16b, v23.16b\n"
+      "zip1 v24.16b, v31.16b, v22.16b\n"
+      "str q16, [x20, #0x30]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "zip2 v23.16b, v0.16b, v23.16b\n"
+      "zip2 v22.16b, v31.16b, v22.16b\n"
+      "zip1 v16.16b, v30.16b, v29.16b\n"
+      "zip2 v17.16b, v30.16b, v29.16b\n"
+      "str q16, [x20, #0x0]\n"
+      "zip1 v16.16b, v28.16b, v27.16b\n"
+      "zip2 v18.16b, v28.16b, v27.16b\n"
+      "str q17, [x20, #0x10]\n"
+      "str q16, [x20, #0x20]\n"
+      "zip1 v17.16b, v26.16b, v21.16b\n"
+      "zip2 v16.16b, v26.16b, v21.16b\n"
+      "str q18, [x20, #0x30]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "zip1 v21.16b, v20.16b, v19.16b\n"
+      "zip2 v20.16b, v20.16b, v19.16b\n"
+      "str q17, [x20, #0x0]\n"
+      "zip1 v19.16b, v25.16b, v24.16b\n"
+      "zip2 v18.16b, v25.16b, v24.16b\n"
+      "str q16, [x20, #0x10]\n"
+      "zip1 v17.16b, v23.16b, v22.16b\n"
+      "zip2 v16.16b, v23.16b, v22.16b\n"
+      "str q21, [x20, #0x20]\n"
+      "str q20, [x20, #0x30]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "str q19, [x20, #0x0]\n"
+      "str q18, [x20, #0x10]\n"
+      "str q17, [x20, #0x20]\n"
+      "str q16, [x20, #0x30]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x16, #0x4\n"
+      "cmp x21, #0x4\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr s31, [x17], #0x4\n"
-      "ldr s30, [x14], #0x4\n"
-      "sub x16, x16, #0x4\n"
-      "ldr s29, [x13], #0x4\n"
-      "ldr s28, [x12], #0x4\n"
-      "cmp x16, #0x4\n"
-      "ldr s27, [x11], #0x4\n"
-      "ldr s26, [x10], #0x4\n"
-      "ldr s25, [x9], #0x4\n"
-      "ldr s24, [x28], #0x4\n"
-      "ldr s20, [x27], #0x4\n"
-      "ldr s21, [x26], #0x4\n"
-      "ldr s23, [x25], #0x4\n"
-      "ldr s22, [x24], #0x4\n"
-      "ldr s19, [x23], #0x4\n"
-      "ldr s18, [x22], #0x4\n"
-      "ldr s17, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
-      "zip1 v20.16b, v31.16b, v20.16b\n"
-      "zip1 v21.16b, v30.16b, v21.16b\n"
-      "zip1 v23.16b, v29.16b, v23.16b\n"
-      "zip1 v22.16b, v28.16b, v22.16b\n"
-      "zip1 v19.16b, v27.16b, v19.16b\n"
-      "zip1 v18.16b, v26.16b, v18.16b\n"
+      "ldr s21, [x17], #0x4\n"
+      "ldr s23, [x16], #0x4\n"
+      "sub x21, x21, #0x4\n"
+      "cmp x21, #0x4\n"
+      "ldr s20, [x15], #0x4\n"
+      "ldr s22, [x14], #0x4\n"
+      "ldr s19, [x13], #0x4\n"
+      "ldr s18, [x12], #0x4\n"
+      "ldr s25, [x11], #0x4\n"
+      "ldr s24, [x10], #0x4\n"
+      "ldr s17, [x9], #0x4\n"
+      "ldr s16, [x28], #0x4\n"
+      "zip1 v21.16b, v21.16b, v17.16b\n"
+      "zip1 v23.16b, v23.16b, v16.16b\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
+      "zip1 v20.16b, v20.16b, v17.16b\n"
+      "zip1 v22.16b, v22.16b, v16.16b\n"
+      "ldr s17, [x25], #0x4\n"
+      "ldr s16, [x24], #0x4\n"
+      "zip1 v19.16b, v19.16b, v17.16b\n"
+      "zip1 v18.16b, v18.16b, v16.16b\n"
+      "ldr s17, [x23], #0x4\n"
+      "ldr s16, [x22], #0x4\n"
       "zip1 v17.16b, v25.16b, v17.16b\n"
       "zip1 v16.16b, v24.16b, v16.16b\n"
-      "zip1 v20.16b, v20.16b, v19.16b\n"
-      "zip1 v21.16b, v21.16b, v18.16b\n"
-      "zip1 v19.16b, v23.16b, v17.16b\n"
+      "zip1 v21.16b, v21.16b, v19.16b\n"
+      "zip1 v20.16b, v20.16b, v17.16b\n"
+      "zip1 v19.16b, v23.16b, v18.16b\n"
       "zip1 v16.16b, v22.16b, v16.16b\n"
-      "zip1 v18.16b, v20.16b, v19.16b\n"
-      "zip1 v17.16b, v21.16b, v16.16b\n"
-      "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip2 v16.16b, v21.16b, v16.16b\n"
-      "zip1 v19.16b, v18.16b, v17.16b\n"
+      "zip1 v18.16b, v21.16b, v20.16b\n"
+      "zip1 v17.16b, v19.16b, v16.16b\n"
+      "zip2 v20.16b, v21.16b, v20.16b\n"
+      "zip2 v19.16b, v19.16b, v16.16b\n"
+      "zip1 v16.16b, v18.16b, v17.16b\n"
       "zip2 v18.16b, v18.16b, v17.16b\n"
-      "zip1 v17.16b, v20.16b, v16.16b\n"
-      "zip2 v16.16b, v20.16b, v16.16b\n"
-      "str q19, [x15, #0x0]\n"
-      "str q18, [x15, #0x10]\n"
-      "str q17, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "str q16, [x20, #0x0]\n"
+      "zip1 v17.16b, v20.16b, v19.16b\n"
+      "zip2 v16.16b, v20.16b, v19.16b\n"
+      "str q18, [x20, #0x10]\n"
+      "str q17, [x20, #0x20]\n"
+      "str q16, [x20, #0x30]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x16, 7f\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x15, #0x0]\n"
-      "str q16, [x15, #0x10]\n"
-      "str q16, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
+      "cmp x21, #0x1\n"
+      "blt 7f\n"
       "6:"  // Main row loop: width 1 loop: loop
-      "ldr b31, [x17], #0x1\n"
-      "ldr b30, [x14], #0x1\n"
-      "sub x16, x16, #0x1\n"
-      "ldr b29, [x13], #0x1\n"
-      "ldr b28, [x12], #0x1\n"
-      "cmp x16, #0x1\n"
-      "ldr b27, [x11], #0x1\n"
-      "ldr b26, [x10], #0x1\n"
-      "ldr b25, [x9], #0x1\n"
-      "ldr b24, [x28], #0x1\n"
-      "ldr b23, [x27], #0x1\n"
-      "ldr b22, [x26], #0x1\n"
-      "ldr b21, [x25], #0x1\n"
-      "ldr b20, [x24], #0x1\n"
-      "ldr b19, [x23], #0x1\n"
-      "ldr b18, [x22], #0x1\n"
-      "ldr b17, [x21], #0x1\n"
-      "ldr b16, [x20], #0x1\n"
-      "zip1 v23.16b, v31.16b, v23.16b\n"
-      "zip1 v22.16b, v30.16b, v22.16b\n"
-      "zip1 v21.16b, v29.16b, v21.16b\n"
-      "zip1 v20.16b, v28.16b, v20.16b\n"
-      "zip1 v19.16b, v27.16b, v19.16b\n"
-      "zip1 v18.16b, v26.16b, v18.16b\n"
+      "ldr b23, [x17], #0x1\n"
+      "ldr b22, [x16], #0x1\n"
+      "sub x21, x21, #0x1\n"
+      "cmp x21, #0x1\n"
+      "ldr b21, [x15], #0x1\n"
+      "ldr b20, [x14], #0x1\n"
+      "ldr b19, [x13], #0x1\n"
+      "ldr b18, [x12], #0x1\n"
+      "ldr b25, [x11], #0x1\n"
+      "ldr b24, [x10], #0x1\n"
+      "ldr b17, [x9], #0x1\n"
+      "ldr b16, [x28], #0x1\n"
+      "zip1 v23.16b, v23.16b, v17.16b\n"
+      "zip1 v22.16b, v22.16b, v16.16b\n"
+      "ldr b17, [x27], #0x1\n"
+      "ldr b16, [x26], #0x1\n"
+      "zip1 v21.16b, v21.16b, v17.16b\n"
+      "zip1 v20.16b, v20.16b, v16.16b\n"
+      "ldr b17, [x25], #0x1\n"
+      "ldr b16, [x24], #0x1\n"
+      "zip1 v19.16b, v19.16b, v17.16b\n"
+      "zip1 v18.16b, v18.16b, v16.16b\n"
+      "ldr b17, [x23], #0x1\n"
+      "ldr b16, [x22], #0x1\n"
       "zip1 v17.16b, v25.16b, v17.16b\n"
       "zip1 v16.16b, v24.16b, v16.16b\n"
       "zip1 v19.16b, v23.16b, v19.16b\n"
-      "zip1 v18.16b, v22.16b, v18.16b\n"
       "zip1 v17.16b, v21.16b, v17.16b\n"
+      "zip1 v18.16b, v22.16b, v18.16b\n"
       "zip1 v16.16b, v20.16b, v16.16b\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str q16, [x15, #0x0]\n"
-      "add x15, x15, #0x10\n"
+      "str q16, [x20, #0x0]\n"
+      "add x20, x20, #0x10\n"
       "bge 6b\n"
-      "7:"  // Main row loop: odd col skip
+      "7:"  // Main row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x40\n"
       "bge 1b\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_1x4.hpp
index 282375b3a3..94a4b5d07f 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_1x4.hpp
@@ -43,13 +43,11 @@ void a64_transpose_interleave_4_1x4(uint8_t *out, const uint8_t *in, size_t widt
       "blt 8f\n"
       "1:"  // Main row loop: Head
       "mov x17, %x[in]\n"
-      "mov x16, %x[width]\n"
-      "mov x15, %x[out]\n"
-      "sub %x[height], %x[height], #0x10\n"
-      "add x14, x17, %x[in_stride]\n"
+      "add x16, x17, %x[in_stride]\n"
+      "add x15, x16, %x[in_stride]\n"
+      "add x14, x15, %x[in_stride]\n"
       "add x13, x14, %x[in_stride]\n"
       "add x12, x13, %x[in_stride]\n"
-      "cmp x16, #0x10\n"
       "add x11, x12, %x[in_stride]\n"
       "add x10, x11, %x[in_stride]\n"
       "add x9, x10, %x[in_stride]\n"
@@ -57,170 +55,168 @@ void a64_transpose_interleave_4_1x4(uint8_t *out, const uint8_t *in, size_t widt
       "add x27, x28, %x[in_stride]\n"
       "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add x23, x24, %x[in_stride]\n"
+      "mov x24, %x[width]\n"
+      "add x23, x25, %x[in_stride]\n"
       "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
+      "cmp x24, #0x10\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x10\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ldr q19, [x17], #0x10\n"
-      "ldr q18, [x14], #0x10\n"
-      "sub x16, x16, #0x10\n"
-      "ldr q17, [x13], #0x10\n"
-      "ldr q16, [x12], #0x10\n"
-      "cmp x16, #0x10\n"
-      "ldr q27, [x11], #0x10\n"
-      "ldr q26, [x10], #0x10\n"
-      "ldr q25, [x9], #0x10\n"
-      "ldr q24, [x28], #0x10\n"
-      "ldr q23, [x27], #0x10\n"
-      "ldr q22, [x26], #0x10\n"
-      "zip1 v5.16b, v19.16b, v17.16b\n"
-      "zip1 v4.16b, v18.16b, v16.16b\n"
-      "ldr q21, [x25], #0x10\n"
-      "ldr q20, [x24], #0x10\n"
-      "zip2 v3.16b, v19.16b, v17.16b\n"
-      "zip2 v2.16b, v18.16b, v16.16b\n"
-      "ldr q19, [x23], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "zip1 v1.16b, v27.16b, v25.16b\n"
-      "zip1 v0.16b, v26.16b, v24.16b\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q21, [x17], #0x10\n"
+      "ldr q20, [x16], #0x10\n"
+      "sub x24, x24, #0x10\n"
+      "cmp x24, #0x10\n"
+      "ldr q17, [x15], #0x10\n"
+      "ldr q16, [x14], #0x10\n"
+      "zip1 v3.16b, v21.16b, v17.16b\n"
+      "zip1 v2.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x13], #0x10\n"
+      "ldr q18, [x12], #0x10\n"
+      "zip2 v1.16b, v21.16b, v17.16b\n"
+      "zip2 v0.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x11], #0x10\n"
+      "ldr q16, [x10], #0x10\n"
+      "zip1 v31.16b, v19.16b, v17.16b\n"
+      "zip1 v30.16b, v18.16b, v16.16b\n"
+      "ldr q21, [x9], #0x10\n"
+      "ldr q20, [x28], #0x10\n"
+      "zip2 v29.16b, v19.16b, v17.16b\n"
+      "zip2 v28.16b, v18.16b, v16.16b\n"
+      "ldr q17, [x27], #0x10\n"
+      "ldr q16, [x26], #0x10\n"
+      "zip1 v23.16b, v21.16b, v17.16b\n"
+      "zip1 v22.16b, v20.16b, v16.16b\n"
+      "ldr q19, [x25], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "zip2 v27.16b, v21.16b, v17.16b\n"
+      "zip2 v26.16b, v20.16b, v16.16b\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "zip2 v31.16b, v27.16b, v25.16b\n"
-      "zip2 v30.16b, v26.16b, v24.16b\n"
-      "zip1 v25.16b, v23.16b, v21.16b\n"
-      "zip1 v24.16b, v22.16b, v20.16b\n"
-      "zip2 v29.16b, v23.16b, v21.16b\n"
-      "zip2 v28.16b, v22.16b, v20.16b\n"
-      "zip1 v23.16b, v19.16b, v17.16b\n"
-      "zip1 v22.16b, v18.16b, v16.16b\n"
-      "zip2 v27.16b, v19.16b, v17.16b\n"
-      "zip2 v26.16b, v18.16b, v16.16b\n"
-      "zip1 v21.16b, v5.16b, v4.16b\n"
-      "zip1 v20.16b, v1.16b, v0.16b\n"
-      "zip1 v19.16b, v25.16b, v24.16b\n"
-      "zip1 v18.16b, v23.16b, v22.16b\n"
-      "zip2 v17.16b, v5.16b, v4.16b\n"
-      "zip2 v16.16b, v1.16b, v0.16b\n"
-      "str q21, [x15, #0x0]\n"
-      "str q20, [x15, #0x10]\n"
-      "zip2 v25.16b, v25.16b, v24.16b\n"
-      "zip2 v24.16b, v23.16b, v22.16b\n"
-      "str q19, [x15, #0x20]\n"
-      "zip1 v23.16b, v3.16b, v2.16b\n"
-      "zip1 v22.16b, v31.16b, v30.16b\n"
-      "str q18, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "zip1 v21.16b, v29.16b, v28.16b\n"
-      "zip1 v20.16b, v27.16b, v26.16b\n"
-      "str q17, [x15, #0x0]\n"
+      "zip1 v21.16b, v19.16b, v17.16b\n"
+      "zip1 v20.16b, v18.16b, v16.16b\n"
+      "zip2 v25.16b, v19.16b, v17.16b\n"
+      "zip2 v24.16b, v18.16b, v16.16b\n"
+      "zip1 v16.16b, v3.16b, v2.16b\n"
+      "zip1 v18.16b, v31.16b, v30.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "zip1 v17.16b, v23.16b, v22.16b\n"
+      "zip1 v16.16b, v21.16b, v20.16b\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
       "zip2 v19.16b, v3.16b, v2.16b\n"
       "zip2 v18.16b, v31.16b, v30.16b\n"
-      "str q16, [x15, #0x10]\n"
-      "zip2 v17.16b, v29.16b, v28.16b\n"
-      "zip2 v16.16b, v27.16b, v26.16b\n"
-      "str q25, [x15, #0x20]\n"
-      "str q24, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "str q23, [x15, #0x0]\n"
-      "str q22, [x15, #0x10]\n"
-      "str q21, [x15, #0x20]\n"
-      "str q20, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "str q19, [x15, #0x0]\n"
-      "str q18, [x15, #0x10]\n"
-      "str q17, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip2 v17.16b, v23.16b, v22.16b\n"
+      "zip2 v16.16b, v21.16b, v20.16b\n"
+      "str q19, [x21, #0x0]\n"
+      "zip1 v23.16b, v1.16b, v0.16b\n"
+      "zip1 v22.16b, v29.16b, v28.16b\n"
+      "str q18, [x21, #0x10]\n"
+      "zip1 v21.16b, v27.16b, v26.16b\n"
+      "zip1 v20.16b, v25.16b, v24.16b\n"
+      "str q17, [x21, #0x20]\n"
+      "zip2 v19.16b, v1.16b, v0.16b\n"
+      "zip2 v18.16b, v29.16b, v28.16b\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip2 v17.16b, v27.16b, v26.16b\n"
+      "zip2 v16.16b, v25.16b, v24.16b\n"
+      "str q23, [x21, #0x0]\n"
+      "str q22, [x21, #0x10]\n"
+      "str q21, [x21, #0x20]\n"
+      "str q20, [x21, #0x30]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cmp x16, #0x4\n"
+      "cmp x24, #0x4\n"
       "blt 5f\n"
       "4:"  // Main row loop: Column loop
-      "ldr s23, [x17], #0x4\n"
-      "ldr s21, [x14], #0x4\n"
-      "sub x16, x16, #0x4\n"
-      "ldr s20, [x13], #0x4\n"
-      "ldr s17, [x12], #0x4\n"
-      "cmp x16, #0x4\n"
-      "ldr s22, [x11], #0x4\n"
-      "ldr s19, [x10], #0x4\n"
-      "ldr s18, [x9], #0x4\n"
-      "ldr s16, [x28], #0x4\n"
-      "ldr s27, [x27], #0x4\n"
-      "ldr s26, [x26], #0x4\n"
-      "zip1 v25.16b, v23.16b, v20.16b\n"
-      "zip1 v21.16b, v21.16b, v17.16b\n"
+      "ldr s19, [x17], #0x4\n"
+      "ldr s18, [x16], #0x4\n"
+      "sub x24, x24, #0x4\n"
+      "cmp x24, #0x4\n"
+      "ldr s17, [x15], #0x4\n"
+      "ldr s16, [x14], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr s19, [x13], #0x4\n"
+      "ldr s18, [x12], #0x4\n"
+      "zip1 v22.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x11], #0x4\n"
+      "ldr s16, [x10], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr s19, [x9], #0x4\n"
+      "ldr s18, [x28], #0x4\n"
+      "zip1 v21.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x27], #0x4\n"
+      "ldr s16, [x26], #0x4\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
       "ldr s20, [x25], #0x4\n"
-      "ldr s17, [x24], #0x4\n"
-      "ldr s24, [x23], #0x4\n"
-      "ldr s23, [x22], #0x4\n"
-      "zip1 v22.16b, v22.16b, v18.16b\n"
-      "zip1 v19.16b, v19.16b, v16.16b\n"
-      "ldr s18, [x21], #0x4\n"
+      "ldr s19, [x23], #0x4\n"
+      "zip1 v18.16b, v17.16b, v16.16b\n"
+      "ldr s17, [x22], #0x4\n"
       "ldr s16, [x20], #0x4\n"
-      "zip1 v21.16b, v25.16b, v21.16b\n"
-      "zip1 v20.16b, v27.16b, v20.16b\n"
-      "zip1 v17.16b, v26.16b, v17.16b\n"
-      "zip1 v19.16b, v22.16b, v19.16b\n"
-      "zip1 v18.16b, v24.16b, v18.16b\n"
-      "zip1 v16.16b, v23.16b, v16.16b\n"
-      "str q21, [x15, #0x0]\n"
       "zip1 v17.16b, v20.16b, v17.16b\n"
-      "str q19, [x15, #0x10]\n"
-      "zip1 v16.16b, v18.16b, v16.16b\n"
-      "str q17, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "zip1 v16.16b, v19.16b, v16.16b\n"
+      "zip1 v16.16b, v17.16b, v16.16b\n"
+      "str q22, [x21, #0x0]\n"
+      "str q21, [x21, #0x10]\n"
+      "str q18, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 4b\n"
       "5:"  // Main row loop: Column loop skip
-      "cbz x16, 7f\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x15, #0x0]\n"
-      "str q16, [x15, #0x10]\n"
-      "str q16, [x15, #0x20]\n"
-      "str q16, [x15, #0x30]\n"
+      "cmp x24, #0x1\n"
+      "blt 7f\n"
       "6:"  // Main row loop: width 1 loop: loop
-      "ldr b23, [x17], #0x1\n"
-      "ldr b21, [x14], #0x1\n"
-      "sub x16, x16, #0x1\n"
-      "ldr b20, [x13], #0x1\n"
-      "ldr b19, [x12], #0x1\n"
-      "cmp x16, #0x1\n"
-      "ldr b22, [x11], #0x1\n"
-      "ldr b18, [x10], #0x1\n"
-      "ldr b17, [x9], #0x1\n"
-      "ldr b16, [x28], #0x1\n"
-      "ldr b27, [x27], #0x1\n"
-      "ldr b26, [x26], #0x1\n"
-      "zip1 v25.16b, v23.16b, v20.16b\n"
-      "zip1 v21.16b, v21.16b, v19.16b\n"
+      "ldr b19, [x17], #0x1\n"
+      "ldr b18, [x16], #0x1\n"
+      "sub x24, x24, #0x1\n"
+      "cmp x24, #0x1\n"
+      "ldr b17, [x15], #0x1\n"
+      "ldr b16, [x14], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr b19, [x13], #0x1\n"
+      "ldr b18, [x12], #0x1\n"
+      "zip1 v22.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x11], #0x1\n"
+      "ldr b16, [x10], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
+      "ldr b19, [x9], #0x1\n"
+      "ldr b18, [x28], #0x1\n"
+      "zip1 v21.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x27], #0x1\n"
+      "ldr b16, [x26], #0x1\n"
+      "zip1 v17.16b, v19.16b, v17.16b\n"
+      "zip1 v16.16b, v18.16b, v16.16b\n"
       "ldr b20, [x25], #0x1\n"
-      "ldr b19, [x24], #0x1\n"
-      "ldr b24, [x23], #0x1\n"
-      "ldr b23, [x22], #0x1\n"
-      "zip1 v22.16b, v22.16b, v17.16b\n"
-      "zip1 v17.16b, v18.16b, v16.16b\n"
-      "ldr b18, [x21], #0x1\n"
+      "ldr b19, [x23], #0x1\n"
+      "zip1 v18.16b, v17.16b, v16.16b\n"
+      "ldr b17, [x22], #0x1\n"
       "ldr b16, [x20], #0x1\n"
-      "zip1 v21.16b, v25.16b, v21.16b\n"
-      "zip1 v20.16b, v27.16b, v20.16b\n"
-      "zip1 v19.16b, v26.16b, v19.16b\n"
-      "zip1 v17.16b, v22.16b, v17.16b\n"
-      "zip1 v18.16b, v24.16b, v18.16b\n"
-      "zip1 v16.16b, v23.16b, v16.16b\n"
-      "str s21, [x15, #0x0]\n"
-      "str s17, [x15, #0x10]\n"
-      "zip1 v17.16b, v20.16b, v19.16b\n"
-      "zip1 v16.16b, v18.16b, v16.16b\n"
-      "str s17, [x15, #0x20]\n"
-      "str s16, [x15, #0x30]\n"
-      "add x15, x15, #0x4\n"
+      "zip1 v17.16b, v20.16b, v17.16b\n"
+      "zip1 v16.16b, v19.16b, v16.16b\n"
+      "str s22, [x21, #0x0]\n"
+      "zip1 v16.16b, v17.16b, v16.16b\n"
+      "str s21, [x21, #0x10]\n"
+      "str s18, [x21, #0x20]\n"
+      "str s16, [x21, #0x30]\n"
+      "add x21, x21, #0x4\n"
       "bge 6b\n"
-      "7:"  // Main row loop: odd col skip
+      "7:"  // Main row loop: width 1 loop: skip
       "cmp %x[height], #0x10\n"
       "add %x[out], %x[out], #0x40\n"
       "bge 1b\n"
@@ -228,85 +224,84 @@ void a64_transpose_interleave_4_1x4(uint8_t *out, const uint8_t *in, size_t widt
       "8:"  // Main loop skip
       "9:"  // Tail row loop: Head
       "mov x17, %x[in]\n"
+      "add x16, x17, %x[in_stride]\n"
+      "add x15, x16, %x[in_stride]\n"
       "mov x20, %x[width]\n"
+      "add x14, x15, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x15, %x[out]\n"
-      "add x14, x17, %x[in_stride]\n"
-      "add x13, x14, %x[in_stride]\n"
-      "add x12, x13, %x[in_stride]\n"
-      "csel x13, x13, %x[pad_row], GE\n"
-      "add %x[in], x12, %x[in_stride]\n"
-      "csel x12, x12, %x[pad_row], GT\n"
-      "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add %x[in], x14, %x[in_stride]\n"
       "csel x14, x14, %x[pad_row], GT\n"
+      "csel x15, x15, %x[pad_row], GE\n"
+      "cmp %x[height], #0x1\n"
+      "csel x16, x16, %x[pad_row], GT\n"
       "cmp x20, #0x10\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 11f\n"
       "10:"  // Tail row loop: Unroll column loop
-      "ldr q20, [x17], #0x10\n"
-      "ldr q21, [x14], #0x10\n"
+      "ldr q19, [x17], #0x10\n"
+      "ldr q21, [x16], #0x10\n"
       "sub x20, x20, #0x10\n"
-      "ldr q19, [x13], #0x10\n"
-      "ldr q16, [x12], #0x10\n"
       "cmp x20, #0x10\n"
-      "zip1 v18.16b, v20.16b, v19.16b\n"
+      "ldr q18, [x15], #0x10\n"
+      "ldr q16, [x14], #0x10\n"
+      "zip1 v20.16b, v19.16b, v18.16b\n"
       "zip1 v17.16b, v21.16b, v16.16b\n"
-      "zip2 v20.16b, v20.16b, v19.16b\n"
-      "zip2 v19.16b, v21.16b, v16.16b\n"
-      "zip1 v16.16b, v18.16b, v17.16b\n"
-      "zip2 v18.16b, v18.16b, v17.16b\n"
-      "str q16, [x15, #0x0]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "zip1 v17.16b, v20.16b, v19.16b\n"
-      "zip2 v16.16b, v20.16b, v19.16b\n"
-      "str q18, [x15, #0x0]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "str q17, [x15, #0x0]\n"
-      "add x15, x15, %x[out_stride]\n"
-      "str q16, [x15, #0x0]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "zip2 v19.16b, v19.16b, v18.16b\n"
+      "zip2 v18.16b, v21.16b, v16.16b\n"
+      "zip1 v16.16b, v20.16b, v17.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip2 v16.16b, v20.16b, v17.16b\n"
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip1 v17.16b, v19.16b, v18.16b\n"
+      "zip2 v16.16b, v19.16b, v18.16b\n"
+      "str q17, [x21, #0x0]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 10b\n"
       "11:"  // Tail row loop: Unroll column loop skip
       "cmp x20, #0x4\n"
       "blt 13f\n"
       "12:"  // Tail row loop: Column loop
       "ldr s19, [x17], #0x4\n"
-      "ldr s18, [x14], #0x4\n"
+      "ldr s18, [x16], #0x4\n"
       "sub x20, x20, #0x4\n"
-      "ldr s17, [x13], #0x4\n"
-      "ldr s16, [x12], #0x4\n"
       "cmp x20, #0x4\n"
+      "ldr s17, [x15], #0x4\n"
+      "ldr s16, [x14], #0x4\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str q16, [x15, #0x0]\n"
-      "add x15, x15, %x[out_stride]\n"
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 12b\n"
       "13:"  // Tail row loop: Column loop skip
-      "cbz x20, 15f\n"
-      "movi v16.16b, #0x0\n"
-      "str q16, [x15, #0x0]\n"
+      "cmp x20, #0x1\n"
+      "blt 15f\n"
       "14:"  // Tail row loop: width 1 loop: loop
       "ldr b19, [x17], #0x1\n"
-      "ldr b18, [x14], #0x1\n"
+      "ldr b18, [x16], #0x1\n"
       "sub x20, x20, #0x1\n"
-      "ldr b17, [x13], #0x1\n"
-      "ldr b16, [x12], #0x1\n"
       "cmp x20, #0x1\n"
+      "ldr b17, [x15], #0x1\n"
+      "ldr b16, [x14], #0x1\n"
       "zip1 v17.16b, v19.16b, v17.16b\n"
       "zip1 v16.16b, v18.16b, v16.16b\n"
       "zip1 v16.16b, v17.16b, v16.16b\n"
-      "str s16, [x15, #0x0]\n"
-      "add x15, x15, #0x4\n"
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
       "bge 14b\n"
-      "15:"  // Tail row loop: odd col skip
+      "15:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x10\n"
       "bge 9b\n"
       "16:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp
index 387c6adabd..03b134e422 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp
@@ -34,201 +34,175 @@ void a64_transpose_interleave_64(uint16_t *out, const uint16_t *in, size_t width
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x20\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
       "ldr q31, [x25], #0x10\n"
-      "ldr q30, [x22], #0x10\n"
+      "ldr q30, [x23], #0x10\n"
       "sub x24, x24, #0x20\n"
-      "ldr q29, [x21], #0x10\n"
-      "ldr q28, [x20], #0x10\n"
       "cmp x24, #0x20\n"
+      "ldr q29, [x22], #0x10\n"
+      "ldr q28, [x20], #0x10\n"
       "ldr q27, [x25], #0x10\n"
-      "ldr q26, [x22], #0x10\n"
-      "ldr q25, [x21], #0x10\n"
+      "ldr q26, [x23], #0x10\n"
+      "ldr q25, [x22], #0x10\n"
       "ldr q24, [x20], #0x10\n"
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "ldr q21, [x22], #0x10\n"
       "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q31, [x23, #0x0]\n"
-      "str q27, [x23, #0x10]\n"
-      "str q23, [x23, #0x20]\n"
-      "str q19, [x23, #0x30]\n"
-      "str q30, [x23, #0x40]\n"
-      "str q26, [x23, #0x50]\n"
-      "str q22, [x23, #0x60]\n"
-      "str q18, [x23, #0x70]\n"
-      "str q29, [x23, #0x80]\n"
-      "str q25, [x23, #0x90]\n"
-      "str q21, [x23, #0xa0]\n"
-      "str q17, [x23, #0xb0]\n"
-      "str q28, [x23, #0xc0]\n"
-      "str q24, [x23, #0xd0]\n"
-      "str q20, [x23, #0xe0]\n"
-      "str q16, [x23, #0xf0]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "str q31, [x21, #0x0]\n"
+      "str q27, [x21, #0x10]\n"
+      "str q23, [x21, #0x20]\n"
+      "str q19, [x21, #0x30]\n"
+      "str q30, [x21, #0x40]\n"
+      "str q26, [x21, #0x50]\n"
+      "str q22, [x21, #0x60]\n"
+      "str q18, [x21, #0x70]\n"
+      "str q29, [x21, #0x80]\n"
+      "str q25, [x21, #0x90]\n"
+      "str q21, [x21, #0xa0]\n"
+      "str q17, [x21, #0xb0]\n"
+      "str q28, [x21, #0xc0]\n"
+      "str q24, [x21, #0xd0]\n"
+      "str q20, [x21, #0xe0]\n"
+      "str q16, [x21, #0xf0]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x10\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "str q16, [x23, #0x80]\n"
-      "str q16, [x23, #0x90]\n"
-      "str q16, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "str q16, [x23, #0xc0]\n"
-      "str q16, [x23, #0xd0]\n"
-      "str q16, [x23, #0xe0]\n"
-      "str q16, [x23, #0xf0]\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 16 loop: loop
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
       "sub x24, x24, #0x10\n"
-      "ldr q21, [x21], #0x10\n"
-      "ldr q20, [x20], #0x10\n"
       "cmp x24, #0x10\n"
+      "ldr q21, [x22], #0x10\n"
+      "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q23, [x23, #0x0]\n"
-      "str q19, [x23, #0x10]\n"
-      "str q22, [x23, #0x40]\n"
-      "str q18, [x23, #0x50]\n"
-      "str q21, [x23, #0x80]\n"
-      "str q17, [x23, #0x90]\n"
-      "str q20, [x23, #0xc0]\n"
-      "str q16, [x23, #0xd0]\n"
-      "add x23, x23, #0x20\n"
+      "str q23, [x21, #0x0]\n"
+      "str q19, [x21, #0x10]\n"
+      "str q22, [x21, #0x40]\n"
+      "str q18, [x21, #0x50]\n"
+      "str q21, [x21, #0x80]\n"
+      "str q17, [x21, #0x90]\n"
+      "str q20, [x21, #0xc0]\n"
+      "str q16, [x21, #0xd0]\n"
+      "add x21, x21, #0x20\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 16 loop: skip
       "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr d19, [x25], #0x8\n"
-      "ldr d18, [x22], #0x8\n"
+      "ldr d18, [x23], #0x8\n"
       "sub x24, x24, #0x4\n"
-      "ldr d17, [x21], #0x8\n"
-      "ldr d16, [x20], #0x8\n"
       "cmp x24, #0x4\n"
-      "str d19, [x23, #0x0]\n"
-      "str d18, [x23, #0x40]\n"
-      "str d17, [x23, #0x80]\n"
-      "str d16, [x23, #0xc0]\n"
-      "add x23, x23, #0x8\n"
+      "ldr d17, [x22], #0x8\n"
+      "ldr d16, [x20], #0x8\n"
+      "str d19, [x21, #0x0]\n"
+      "str d18, [x21, #0x40]\n"
+      "str d17, [x21, #0x80]\n"
+      "str d16, [x21, #0xc0]\n"
+      "add x21, x21, #0x8\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr h19, [x25], #0x2\n"
-      "ldr h18, [x22], #0x2\n"
+      "ldr h18, [x23], #0x2\n"
       "sub x24, x24, #0x1\n"
-      "ldr h17, [x21], #0x2\n"
-      "ldr h16, [x20], #0x2\n"
       "cmp x24, #0x1\n"
-      "str h19, [x23, #0x0]\n"
-      "str h18, [x23, #0x40]\n"
-      "str h17, [x23, #0x80]\n"
-      "str h16, [x23, #0xc0]\n"
-      "add x23, x23, #0x2\n"
+      "ldr h17, [x22], #0x2\n"
+      "ldr h16, [x20], #0x2\n"
+      "str h19, [x21, #0x0]\n"
+      "str h18, [x21, #0x40]\n"
+      "str h17, [x21, #0x80]\n"
+      "str h16, [x21, #0xc0]\n"
+      "add x21, x21, #0x2\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0x100\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x20\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Column loop
       "ldr q19, [x25], #0x10\n"
-      "sub x20, x20, #0x20\n"
       "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
+      "sub x20, x20, #0x20\n"
       "cmp x20, #0x20\n"
+      "ldr q17, [x25], #0x10\n"
       "ldr q16, [x25], #0x10\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x10]\n"
-      "str q17, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Column loop skip
       "cmp x20, #0x10\n"
-      "movi v16.8h, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 16 loop: loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 16 loop: loop
       "ldr q17, [x25], #0x10\n"
-      "sub x20, x20, #0x10\n"
       "ldr q16, [x25], #0x10\n"
+      "sub x20, x20, #0x10\n"
       "cmp x20, #0x10\n"
-      "str q17, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "add x23, x23, #0x20\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 16 loop: skip
+      "str q17, [x21, #0x0]\n"
+      "str q16, [x21, #0x10]\n"
+      "add x21, x21, #0x20\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 16 loop: skip
       "cmp x20, #0x4\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr d16, [x25], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
-      "str d16, [x23, #0x0]\n"
-      "add x23, x23, #0x8\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str d16, [x21, #0x0]\n"
+      "add x21, x21, #0x8\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr h16, [x25], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
-      "str h16, [x23, #0x0]\n"
-      "add x23, x23, #0x2\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str h16, [x21, #0x0]\n"
+      "add x21, x21, #0x2\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x40\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
       : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_96.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_96.hpp
index e5778860a4..2719d24750 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_96.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_96.hpp
@@ -34,251 +34,215 @@ void a64_transpose_interleave_96(uint32_t *out, const uint32_t *in, size_t width
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 11f\n"
+      "blt 10f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x25, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
+      "add x23, x25, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "add x20, x22, %x[in_stride]\n"
       "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
       "ldr q7, [x25], #0x10\n"
-      "ldr q6, [x22], #0x10\n"
+      "ldr q6, [x23], #0x10\n"
       "sub x24, x24, #0x18\n"
-      "ldr q5, [x21], #0x10\n"
-      "ldr q4, [x20], #0x10\n"
       "cmp x24, #0x18\n"
+      "ldr q5, [x22], #0x10\n"
+      "ldr q4, [x20], #0x10\n"
       "ldr q3, [x25], #0x10\n"
-      "ldr q2, [x22], #0x10\n"
-      "ldr q1, [x21], #0x10\n"
+      "ldr q2, [x23], #0x10\n"
+      "ldr q1, [x22], #0x10\n"
       "ldr q0, [x20], #0x10\n"
       "ldr q31, [x25], #0x10\n"
-      "ldr q30, [x22], #0x10\n"
-      "ldr q29, [x21], #0x10\n"
+      "ldr q30, [x23], #0x10\n"
+      "ldr q29, [x22], #0x10\n"
       "ldr q28, [x20], #0x10\n"
       "ldr q27, [x25], #0x10\n"
-      "ldr q26, [x22], #0x10\n"
-      "ldr q25, [x21], #0x10\n"
+      "ldr q26, [x23], #0x10\n"
+      "ldr q25, [x22], #0x10\n"
       "ldr q24, [x20], #0x10\n"
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "ldr q21, [x22], #0x10\n"
       "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q7, [x23, #0x0]\n"
-      "str q3, [x23, #0x10]\n"
-      "str q31, [x23, #0x20]\n"
-      "str q27, [x23, #0x30]\n"
-      "str q23, [x23, #0x40]\n"
-      "str q19, [x23, #0x50]\n"
-      "str q6, [x23, #0x60]\n"
-      "str q2, [x23, #0x70]\n"
-      "str q30, [x23, #0x80]\n"
-      "str q26, [x23, #0x90]\n"
-      "str q22, [x23, #0xa0]\n"
-      "str q18, [x23, #0xb0]\n"
-      "str q5, [x23, #0xc0]\n"
-      "str q1, [x23, #0xd0]\n"
-      "str q29, [x23, #0xe0]\n"
-      "str q25, [x23, #0xf0]\n"
-      "str q21, [x23, #0x100]\n"
-      "str q17, [x23, #0x110]\n"
-      "str q4, [x23, #0x120]\n"
-      "str q0, [x23, #0x130]\n"
-      "str q28, [x23, #0x140]\n"
-      "str q24, [x23, #0x150]\n"
-      "str q20, [x23, #0x160]\n"
-      "str q16, [x23, #0x170]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "str q7, [x21, #0x0]\n"
+      "str q3, [x21, #0x10]\n"
+      "str q31, [x21, #0x20]\n"
+      "str q27, [x21, #0x30]\n"
+      "str q23, [x21, #0x40]\n"
+      "str q19, [x21, #0x50]\n"
+      "str q6, [x21, #0x60]\n"
+      "str q2, [x21, #0x70]\n"
+      "str q30, [x21, #0x80]\n"
+      "str q26, [x21, #0x90]\n"
+      "str q22, [x21, #0xa0]\n"
+      "str q18, [x21, #0xb0]\n"
+      "str q5, [x21, #0xc0]\n"
+      "str q1, [x21, #0xd0]\n"
+      "str q29, [x21, #0xe0]\n"
+      "str q25, [x21, #0xf0]\n"
+      "str q21, [x21, #0x100]\n"
+      "str q17, [x21, #0x110]\n"
+      "str q4, [x21, #0x120]\n"
+      "str q0, [x21, #0x130]\n"
+      "str q28, [x21, #0x140]\n"
+      "str q24, [x21, #0x150]\n"
+      "str q20, [x21, #0x160]\n"
+      "str q16, [x21, #0x170]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
-      "cbz x24, 10f\n"
       "cmp x24, #0x10\n"
-      "movi v16.4s, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "str q16, [x23, #0x60]\n"
-      "str q16, [x23, #0x70]\n"
-      "str q16, [x23, #0x80]\n"
-      "str q16, [x23, #0x90]\n"
-      "str q16, [x23, #0xa0]\n"
-      "str q16, [x23, #0xb0]\n"
-      "str q16, [x23, #0xc0]\n"
-      "str q16, [x23, #0xd0]\n"
-      "str q16, [x23, #0xe0]\n"
-      "str q16, [x23, #0xf0]\n"
-      "str q16, [x23, #0x100]\n"
-      "str q16, [x23, #0x110]\n"
-      "str q16, [x23, #0x120]\n"
-      "str q16, [x23, #0x130]\n"
-      "str q16, [x23, #0x140]\n"
-      "str q16, [x23, #0x150]\n"
-      "str q16, [x23, #0x160]\n"
-      "str q16, [x23, #0x170]\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 16 loop: loop
       "ldr q31, [x25], #0x10\n"
-      "ldr q30, [x22], #0x10\n"
+      "ldr q30, [x23], #0x10\n"
       "sub x24, x24, #0x10\n"
-      "ldr q29, [x21], #0x10\n"
-      "ldr q28, [x20], #0x10\n"
       "cmp x24, #0x10\n"
+      "ldr q29, [x22], #0x10\n"
+      "ldr q28, [x20], #0x10\n"
       "ldr q27, [x25], #0x10\n"
-      "ldr q26, [x22], #0x10\n"
-      "ldr q25, [x21], #0x10\n"
+      "ldr q26, [x23], #0x10\n"
+      "ldr q25, [x22], #0x10\n"
       "ldr q24, [x20], #0x10\n"
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x22], #0x10\n"
-      "ldr q21, [x21], #0x10\n"
+      "ldr q22, [x23], #0x10\n"
+      "ldr q21, [x22], #0x10\n"
       "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
-      "ldr q17, [x21], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
+      "ldr q17, [x22], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q31, [x23, #0x0]\n"
-      "str q27, [x23, #0x10]\n"
-      "str q23, [x23, #0x20]\n"
-      "str q19, [x23, #0x30]\n"
-      "str q30, [x23, #0x60]\n"
-      "str q26, [x23, #0x70]\n"
-      "str q22, [x23, #0x80]\n"
-      "str q18, [x23, #0x90]\n"
-      "str q29, [x23, #0xc0]\n"
-      "str q25, [x23, #0xd0]\n"
-      "str q21, [x23, #0xe0]\n"
-      "str q17, [x23, #0xf0]\n"
-      "str q28, [x23, #0x120]\n"
-      "str q24, [x23, #0x130]\n"
-      "str q20, [x23, #0x140]\n"
-      "str q16, [x23, #0x150]\n"
-      "add x23, x23, #0x40\n"
+      "str q31, [x21, #0x0]\n"
+      "str q27, [x21, #0x10]\n"
+      "str q23, [x21, #0x20]\n"
+      "str q19, [x21, #0x30]\n"
+      "str q30, [x21, #0x60]\n"
+      "str q26, [x21, #0x70]\n"
+      "str q22, [x21, #0x80]\n"
+      "str q18, [x21, #0x90]\n"
+      "str q29, [x21, #0xc0]\n"
+      "str q25, [x21, #0xd0]\n"
+      "str q21, [x21, #0xe0]\n"
+      "str q17, [x21, #0xf0]\n"
+      "str q28, [x21, #0x120]\n"
+      "str q24, [x21, #0x130]\n"
+      "str q20, [x21, #0x140]\n"
+      "str q16, [x21, #0x150]\n"
+      "add x21, x21, #0x40\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 16 loop: skip
       "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x22], #0x10\n"
+      "ldr q18, [x23], #0x10\n"
       "sub x24, x24, #0x4\n"
-      "ldr q17, [x21], #0x10\n"
-      "ldr q16, [x20], #0x10\n"
       "cmp x24, #0x4\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x60]\n"
-      "str q17, [x23, #0xc0]\n"
-      "str q16, [x23, #0x120]\n"
-      "add x23, x23, #0x10\n"
+      "ldr q17, [x22], #0x10\n"
+      "ldr q16, [x20], #0x10\n"
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x60]\n"
+      "str q17, [x21, #0xc0]\n"
+      "str q16, [x21, #0x120]\n"
+      "add x21, x21, #0x10\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr s19, [x25], #0x4\n"
-      "ldr s18, [x22], #0x4\n"
+      "ldr s18, [x23], #0x4\n"
       "sub x24, x24, #0x1\n"
-      "ldr s17, [x21], #0x4\n"
-      "ldr s16, [x20], #0x4\n"
       "cmp x24, #0x1\n"
-      "str s19, [x23, #0x0]\n"
-      "str s18, [x23, #0x60]\n"
-      "str s17, [x23, #0xc0]\n"
-      "str s16, [x23, #0x120]\n"
-      "add x23, x23, #0x4\n"
+      "ldr s17, [x22], #0x4\n"
+      "ldr s16, [x20], #0x4\n"
+      "str s19, [x21, #0x0]\n"
+      "str s18, [x21, #0x60]\n"
+      "str s17, [x21, #0xc0]\n"
+      "str s16, [x21, #0x120]\n"
+      "add x21, x21, #0x4\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
-      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0x180\n"
       "bge 1b\n"
-      "cbz %x[height], 22f\n"
-      "11:"  // Main loop skip
-      "12:"  // Tail row loop: Head
+      "cbz %x[height], 20f\n"
+      "10:"  // Main loop skip
+      "11:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x18\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "blt 14f\n"
-      "13:"  // Tail row loop: Column loop
+      "mov x21, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
+      "blt 13f\n"
+      "12:"  // Tail row loop: Column loop
       "ldr q21, [x25], #0x10\n"
-      "sub x20, x20, #0x18\n"
       "ldr q20, [x25], #0x10\n"
-      "ldr q19, [x25], #0x10\n"
+      "sub x20, x20, #0x18\n"
       "cmp x20, #0x18\n"
+      "ldr q19, [x25], #0x10\n"
       "ldr q18, [x25], #0x10\n"
       "ldr q17, [x25], #0x10\n"
       "ldr q16, [x25], #0x10\n"
-      "str q21, [x23, #0x0]\n"
-      "str q20, [x23, #0x10]\n"
-      "str q19, [x23, #0x20]\n"
-      "str q18, [x23, #0x30]\n"
-      "str q17, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "bge 13b\n"
-      "14:"  // Tail row loop: Column loop skip
-      "cbz x20, 21f\n"
+      "str q21, [x21, #0x0]\n"
+      "str q20, [x21, #0x10]\n"
+      "str q19, [x21, #0x20]\n"
+      "str q18, [x21, #0x30]\n"
+      "str q17, [x21, #0x40]\n"
+      "str q16, [x21, #0x50]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "bge 12b\n"
+      "13:"  // Tail row loop: Column loop skip
       "cmp x20, #0x10\n"
-      "movi v16.4s, #0x0\n"
-      "str q16, [x23, #0x0]\n"
-      "str q16, [x23, #0x10]\n"
-      "str q16, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "str q16, [x23, #0x40]\n"
-      "str q16, [x23, #0x50]\n"
-      "blt 16f\n"
-      "15:"  // Tail row loop: width 16 loop: loop
+      "blt 15f\n"
+      "14:"  // Tail row loop: width 16 loop: loop
       "ldr q19, [x25], #0x10\n"
-      "sub x20, x20, #0x10\n"
       "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
+      "sub x20, x20, #0x10\n"
       "cmp x20, #0x10\n"
+      "ldr q17, [x25], #0x10\n"
       "ldr q16, [x25], #0x10\n"
-      "str q19, [x23, #0x0]\n"
-      "str q18, [x23, #0x10]\n"
-      "str q17, [x23, #0x20]\n"
-      "str q16, [x23, #0x30]\n"
-      "add x23, x23, #0x40\n"
-      "bge 15b\n"
-      "16:"  // Tail row loop: width 16 loop: skip
+      "str q19, [x21, #0x0]\n"
+      "str q18, [x21, #0x10]\n"
+      "str q17, [x21, #0x20]\n"
+      "str q16, [x21, #0x30]\n"
+      "add x21, x21, #0x40\n"
+      "bge 14b\n"
+      "15:"  // Tail row loop: width 16 loop: skip
       "cmp x20, #0x4\n"
-      "blt 18f\n"
-      "17:"  // Tail row loop: width 4 loop: loop
+      "blt 17f\n"
+      "16:"  // Tail row loop: width 4 loop: loop
       "ldr q16, [x25], #0x10\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
-      "str q16, [x23, #0x0]\n"
-      "add x23, x23, #0x10\n"
-      "bge 17b\n"
-      "18:"  // Tail row loop: width 4 loop: skip
+      "str q16, [x21, #0x0]\n"
+      "add x21, x21, #0x10\n"
+      "bge 16b\n"
+      "17:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 20f\n"
-      "19:"  // Tail row loop: width 1 loop: loop
+      "blt 19f\n"
+      "18:"  // Tail row loop: width 1 loop: loop
       "ldr s16, [x25], #0x4\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
-      "str s16, [x23, #0x0]\n"
-      "add x23, x23, #0x4\n"
-      "bge 19b\n"
-      "20:"  // Tail row loop: width 1 loop: skip
-      "21:"  // Tail row loop: odd col skip
+      "str s16, [x21, #0x0]\n"
+      "add x21, x21, #0x4\n"
+      "bge 18b\n"
+      "19:"  // Tail row loop: width 1 loop: skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x60\n"
-      "bge 12b\n"
-      "22:"  // Done
+      "bge 11b\n"
+      "20:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL.hpp
index 4c7b367ed9..768719b0de 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL.hpp
@@ -28,7 +28,7 @@
 
 namespace {
 
-void sme_transpose_interleave_16VL(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height)
+void sme_transpose_interleave_16VL(uint32_t *out, const uint32_t *in, size_t width, size_t in_stride, size_t height)
 {
     size_t out_stride = 16 * height * sme::get_vector_length<uint8_t>();
 
@@ -36,82 +36,82 @@ void sme_transpose_interleave_16VL(uint16_t *out, const uint16_t *in, size_t wid
       ".inst 0xd503477f  // SMSTART ZA\n"
       "ptrue p7.b\n"
       "1:"  // Main row loop: Head
-      "mov x24, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "add %x[in], x24, %x[in_stride]\n"
+      "mov x23, %x[in]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x1\n"
-      "mov x22, %x[width]\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x21, x22\n"
-      "mov x20, x23\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z31.h }, p0/Z, [x24]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z30.h }, p1/Z, [x24, #1, MUL VL]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z29.h }, p0/Z, [x24, #2, MUL VL]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z28.h }, p1/Z, [x24, #3, MUL VL]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z27.h }, p0/Z, [x24, #4, MUL VL]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z26.h }, p1/Z, [x24, #5, MUL VL]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z25.h }, p0/Z, [x24, #6, MUL VL]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z24.h }, p1/Z, [x24, #7, MUL VL]\n"
-      "whilelt p6.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p5.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p4.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p3.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p2.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "addvl x24, x24, #16\n"
-      "dech x22, ALL, MUL #16\n"
-      "ld1h { z23.h }, p0/Z, [x24, #-8, MUL VL]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "cmp x22, #0x0\n"
-      "ld1h { z22.h }, p6/Z, [x24, #-7, MUL VL]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "ld1h { z21.h }, p5/Z, [x24, #-6, MUL VL]\n"
-      "ld1h { z20.h }, p4/Z, [x24, #-5, MUL VL]\n"
-      "ld1h { z19.h }, p3/Z, [x24, #-4, MUL VL]\n"
-      "ld1h { z18.h }, p2/Z, [x24, #-3, MUL VL]\n"
-      "ld1h { z17.h }, p1/Z, [x24, #-2, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x24, #-1, MUL VL]\n"
-      "st1h { z31.h }, p7, [x20]\n"
-      "st1h { z30.h }, p7, [x20, #1, MUL VL]\n"
-      "st1h { z29.h }, p7, [x20, #2, MUL VL]\n"
-      "st1h { z28.h }, p7, [x20, #3, MUL VL]\n"
-      "st1h { z27.h }, p7, [x20, #4, MUL VL]\n"
-      "st1h { z26.h }, p7, [x20, #5, MUL VL]\n"
-      "st1h { z25.h }, p7, [x20, #6, MUL VL]\n"
-      "st1h { z24.h }, p7, [x20, #7, MUL VL]\n"
+      "mov x20, x21\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z31.s }, p0/Z, [x23]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z30.s }, p0/Z, [x23, #1, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z29.s }, p0/Z, [x23, #2, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z28.s }, p0/Z, [x23, #3, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z27.s }, p0/Z, [x23, #4, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z26.s }, p0/Z, [x23, #5, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z25.s }, p0/Z, [x23, #6, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z24.s }, p0/Z, [x23, #7, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "decw x20\n"
+      "whilelt p6.s, XZR, x20\n"
+      "decw x20\n"
+      "whilelt p5.s, XZR, x20\n"
+      "decw x20\n"
+      "whilelt p4.s, XZR, x20\n"
+      "decw x20\n"
+      "whilelt p3.s, XZR, x20\n"
+      "decw x20\n"
+      "whilelt p2.s, XZR, x20\n"
+      "decw x20\n"
+      "whilelt p1.s, XZR, x20\n"
+      "decw x20\n"
+      "addvl x23, x23, #16\n"
+      "ld1w { z23.s }, p0/Z, [x23, #-8, MUL VL]\n"
+      "whilelt p0.s, XZR, x20\n"
+      "mov x20, x22\n"
+      "ld1w { z22.s }, p6/Z, [x23, #-7, MUL VL]\n"
+      "decw x21, ALL, MUL #16\n"
+      "ld1w { z21.s }, p5/Z, [x23, #-6, MUL VL]\n"
+      "cmp x21, #0x0\n"
+      "ld1w { z20.s }, p4/Z, [x23, #-5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1w { z19.s }, p3/Z, [x23, #-4, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x23, #-3, MUL VL]\n"
+      "ld1w { z17.s }, p1/Z, [x23, #-2, MUL VL]\n"
+      "ld1w { z16.s }, p0/Z, [x23, #-1, MUL VL]\n"
+      "st1w { z31.s }, p7, [x20]\n"
+      "st1w { z30.s }, p7, [x20, #1, MUL VL]\n"
+      "st1w { z29.s }, p7, [x20, #2, MUL VL]\n"
+      "st1w { z28.s }, p7, [x20, #3, MUL VL]\n"
+      "st1w { z27.s }, p7, [x20, #4, MUL VL]\n"
+      "st1w { z26.s }, p7, [x20, #5, MUL VL]\n"
+      "st1w { z25.s }, p7, [x20, #6, MUL VL]\n"
+      "st1w { z24.s }, p7, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
-      "st1h { z23.h }, p7, [x20, #-8, MUL VL]\n"
-      "st1h { z22.h }, p7, [x20, #-7, MUL VL]\n"
-      "st1h { z21.h }, p7, [x20, #-6, MUL VL]\n"
-      "st1h { z20.h }, p7, [x20, #-5, MUL VL]\n"
-      "st1h { z19.h }, p7, [x20, #-4, MUL VL]\n"
-      "st1h { z18.h }, p7, [x20, #-3, MUL VL]\n"
-      "st1h { z17.h }, p7, [x20, #-2, MUL VL]\n"
-      "st1h { z16.h }, p7, [x20, #-1, MUL VL]\n"
+      "st1w { z23.s }, p7, [x20, #-8, MUL VL]\n"
+      "st1w { z22.s }, p7, [x20, #-7, MUL VL]\n"
+      "st1w { z21.s }, p7, [x20, #-6, MUL VL]\n"
+      "st1w { z20.s }, p7, [x20, #-5, MUL VL]\n"
+      "st1w { z19.s }, p7, [x20, #-4, MUL VL]\n"
+      "st1w { z18.s }, p7, [x20, #-3, MUL VL]\n"
+      "st1w { z17.s }, p7, [x20, #-2, MUL VL]\n"
+      "st1w { z16.s }, p7, [x20, #-1, MUL VL]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -120,7 +120,7 @@ void sme_transpose_interleave_16VL(uint16_t *out, const uint16_t *in, size_t wid
       ".inst 0xd503467f  // SMSTOP\n"
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x20", "x21", "x22", "x23", "x24", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x20", "x21", "x22", "x23", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
@@ -131,26 +131,13 @@ void Transform<16, 1, true, VLType::SME>(
     float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
 {
     sme_transpose_interleave_16VL(
-        reinterpret_cast<uint16_t *>(out),
-        reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
-        (xmax-x0) * sizeof(float) / 2,
+        reinterpret_cast<uint32_t *>(out),
+        reinterpret_cast<const uint32_t *>(in + k0 * stride + x0),
+        (xmax-x0) * sizeof(float) / 4,
         stride * sizeof(float),
         (kmax-k0)
     );
 }
 
-template<>
-void Transform<16, 1, true, VLType::SME>(
-    __fp16 *out, const __fp16 *in, int stride, int x0, int xmax, int k0, int kmax)
-{
-    sme_transpose_interleave_16VL(
-        reinterpret_cast<uint16_t *>(out),
-        reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
-        (xmax-x0) * sizeof(__fp16) / 2,
-        stride * sizeof(__fp16),
-        (kmax-k0)
-    );
-}
-
 
 #endif  // defined(ARM_COMPUTE_ENABLE_SME)
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_1x4.hpp
index dca0031b55..bb866b2983 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_1x4.hpp
@@ -42,92 +42,92 @@ void sme_transpose_interleave_16VL_1x4(uint8_t *out, const uint8_t *in, size_t w
       ".inst 0xd503477f  // SMSTART ZA\n"
       "ptrue p4.b\n"
       "1:"  // Main row loop: Head
-      "mov x27, %x[in]\n"
+      "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "add x23, x24, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "add x26, x27, %x[in_stride]\n"
-      "mov x25, %x[out]\n"
-      "add x24, x26, %x[in_stride]\n"
-      "mov x23, %x[width]\n"
-      "add x22, x24, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "csel x23, x23, %x[pad_row], GT\n"
       "csel x24, x24, %x[pad_row], GE\n"
-      "add %x[in], x22, %x[in_stride]\n"
-      "csel x22, x22, %x[pad_row], GT\n"
       "cmp %x[height], #0x1\n"
+      "mov x22, %x[out]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "sub %x[height], %x[height], #0x4\n"
-      "csel x26, x26, %x[pad_row], GT\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x21, x23\n"
-      "mov x20, x25\n"
-      "whilelt p3.b, XZR, x21\n"
-      "decb x21\n"
-      "whilelt p2.b, XZR, x21\n"
-      "decb x21\n"
-      "ld1b { z21.b }, p3/Z, [x27]\n"
-      "whilelt p1.b, XZR, x21\n"
-      "decb x21\n"
-      "ld1b { z24.b }, p2/Z, [x27, #1, MUL VL]\n"
-      "whilelt p0.b, XZR, x21\n"
-      "ld1b { z23.b }, p3/Z, [x26]\n"
-      "decw x23, ALL, MUL #16\n"
-      "ld1b { z20.b }, p2/Z, [x26, #1, MUL VL]\n"
-      "cmp x23, #0x0\n"
-      "add x25, x25, %x[out_stride]\n"
-      "ld1b { z19.b }, p3/Z, [x24]\n"
-      "ld1b { z17.b }, p2/Z, [x24, #1, MUL VL]\n"
-      "ld1b { z16.b }, p3/Z, [x22]\n"
-      "ld1b { z18.b }, p2/Z, [x22, #1, MUL VL]\n"
-      "zip1 z22.b, z21.b, z19.b\n"
-      "zip2 z21.b, z21.b, z19.b\n"
-      "ld1b { z28.b }, p1/Z, [x27, #2, MUL VL]\n"
-      "zip1 z1.b, z24.b, z17.b\n"
-      "zip2 z0.b, z24.b, z17.b\n"
-      "ld1b { z27.b }, p0/Z, [x27, #3, MUL VL]\n"
-      "zip1 z17.b, z23.b, z16.b\n"
-      "zip2 z16.b, z23.b, z16.b\n"
-      "addvl x27, x27, #4\n"
-      "ld1b { z26.b }, p1/Z, [x26, #2, MUL VL]\n"
-      "zip1 z31.b, z20.b, z18.b\n"
-      "zip2 z30.b, z20.b, z18.b\n"
-      "ld1b { z25.b }, p0/Z, [x26, #3, MUL VL]\n"
+      "mov x20, x21\n"
+      "whilelt p3.b, XZR, x20\n"
+      "ld1b { z20.b }, p3/Z, [x26]\n"
+      "decb x20\n"
+      "whilelt p2.b, XZR, x20\n"
+      "ld1b { z18.b }, p2/Z, [x26, #1, MUL VL]\n"
+      "decb x20\n"
+      "whilelt p1.b, XZR, x20\n"
+      "ld1b { z17.b }, p3/Z, [x25]\n"
+      "decb x20\n"
+      "whilelt p0.b, XZR, x20\n"
+      "ld1b { z19.b }, p2/Z, [x25, #1, MUL VL]\n"
+      "ld1b { z16.b }, p3/Z, [x24]\n"
+      "zip1 z25.b, z20.b, z16.b\n"
+      "zip2 z24.b, z20.b, z16.b\n"
+      "mov x20, x22\n"
+      "ld1b { z16.b }, p2/Z, [x24, #1, MUL VL]\n"
+      "zip1 z22.b, z18.b, z16.b\n"
+      "zip2 z21.b, z18.b, z16.b\n"
+      "decw x21, ALL, MUL #16\n"
+      "ld1b { z16.b }, p3/Z, [x23]\n"
+      "zip1 z18.b, z17.b, z16.b\n"
+      "zip2 z17.b, z17.b, z16.b\n"
+      "cmp x21, #0x0\n"
+      "ld1b { z16.b }, p2/Z, [x23, #1, MUL VL]\n"
+      "zip1 z20.b, z19.b, z16.b\n"
+      "zip2 z16.b, z19.b, z16.b\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1b { z19.b }, p1/Z, [x26, #2, MUL VL]\n"
+      "zip1 z23.b, z25.b, z18.b\n"
+      "zip2 z0.b, z25.b, z18.b\n"
+      "ld1b { z18.b }, p0/Z, [x26, #3, MUL VL]\n"
+      "zip1 z31.b, z24.b, z17.b\n"
+      "zip2 z30.b, z24.b, z17.b\n"
       "addvl x26, x26, #4\n"
-      "ld1b { z20.b }, p1/Z, [x24, #2, MUL VL]\n"
-      "ld1b { z19.b }, p0/Z, [x24, #3, MUL VL]\n"
-      "zip1 z18.b, z22.b, z17.b\n"
-      "zip2 z24.b, z22.b, z17.b\n"
+      "ld1b { z17.b }, p1/Z, [x25, #2, MUL VL]\n"
+      "zip1 z29.b, z22.b, z20.b\n"
+      "zip2 z28.b, z22.b, z20.b\n"
+      "ld1b { z22.b }, p0/Z, [x25, #3, MUL VL]\n"
+      "zip1 z27.b, z21.b, z16.b\n"
+      "zip2 z26.b, z21.b, z16.b\n"
+      "addvl x25, x25, #4\n"
+      "ld1b { z16.b }, p1/Z, [x24, #2, MUL VL]\n"
+      "zip1 z21.b, z19.b, z16.b\n"
+      "zip2 z20.b, z19.b, z16.b\n"
+      "ld1b { z16.b }, p0/Z, [x24, #3, MUL VL]\n"
+      "zip1 z25.b, z18.b, z16.b\n"
+      "zip2 z24.b, z18.b, z16.b\n"
       "addvl x24, x24, #4\n"
-      "ld1b { z17.b }, p1/Z, [x22, #2, MUL VL]\n"
-      "zip1 z23.b, z21.b, z16.b\n"
-      "zip2 z22.b, z21.b, z16.b\n"
-      "ld1b { z16.b }, p0/Z, [x22, #3, MUL VL]\n"
-      "zip1 z21.b, z28.b, z20.b\n"
-      "zip2 z29.b, z28.b, z20.b\n"
-      "addvl x22, x22, #4\n"
-      "zip1 z28.b, z27.b, z19.b\n"
-      "zip2 z27.b, z27.b, z19.b\n"
-      "zip1 z20.b, z26.b, z17.b\n"
-      "zip2 z19.b, z26.b, z17.b\n"
-      "st1b { z18.b }, p4, [x20]\n"
-      "zip1 z18.b, z25.b, z16.b\n"
-      "zip2 z26.b, z25.b, z16.b\n"
-      "st1b { z24.b }, p4, [x20, #1, MUL VL]\n"
-      "zip1 z17.b, z1.b, z31.b\n"
-      "zip2 z16.b, z1.b, z31.b\n"
-      "st1b { z23.b }, p4, [x20, #2, MUL VL]\n"
-      "zip1 z25.b, z0.b, z30.b\n"
-      "zip2 z24.b, z0.b, z30.b\n"
-      "st1b { z22.b }, p4, [x20, #3, MUL VL]\n"
-      "zip1 z23.b, z21.b, z20.b\n"
-      "zip2 z22.b, z21.b, z20.b\n"
-      "zip1 z21.b, z29.b, z19.b\n"
-      "zip2 z20.b, z29.b, z19.b\n"
-      "st1b { z17.b }, p4, [x20, #4, MUL VL]\n"
-      "zip1 z19.b, z28.b, z18.b\n"
-      "zip2 z18.b, z28.b, z18.b\n"
-      "st1b { z16.b }, p4, [x20, #5, MUL VL]\n"
-      "zip1 z17.b, z27.b, z26.b\n"
-      "zip2 z16.b, z27.b, z26.b\n"
-      "st1b { z25.b }, p4, [x20, #6, MUL VL]\n"
-      "st1b { z24.b }, p4, [x20, #7, MUL VL]\n"
+      "ld1b { z16.b }, p1/Z, [x23, #2, MUL VL]\n"
+      "zip1 z19.b, z17.b, z16.b\n"
+      "zip2 z18.b, z17.b, z16.b\n"
+      "ld1b { z16.b }, p0/Z, [x23, #3, MUL VL]\n"
+      "zip1 z17.b, z22.b, z16.b\n"
+      "zip2 z16.b, z22.b, z16.b\n"
+      "addvl x23, x23, #4\n"
+      "st1b { z23.b }, p4, [x20]\n"
+      "zip1 z23.b, z21.b, z19.b\n"
+      "zip2 z22.b, z21.b, z19.b\n"
+      "st1b { z0.b }, p4, [x20, #1, MUL VL]\n"
+      "zip1 z21.b, z20.b, z18.b\n"
+      "zip2 z20.b, z20.b, z18.b\n"
+      "st1b { z31.b }, p4, [x20, #2, MUL VL]\n"
+      "zip1 z19.b, z25.b, z17.b\n"
+      "zip2 z18.b, z25.b, z17.b\n"
+      "st1b { z30.b }, p4, [x20, #3, MUL VL]\n"
+      "zip1 z17.b, z24.b, z16.b\n"
+      "zip2 z16.b, z24.b, z16.b\n"
+      "st1b { z29.b }, p4, [x20, #4, MUL VL]\n"
+      "st1b { z28.b }, p4, [x20, #5, MUL VL]\n"
+      "st1b { z27.b }, p4, [x20, #6, MUL VL]\n"
+      "st1b { z26.b }, p4, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
       "st1b { z23.b }, p4, [x20, #-8, MUL VL]\n"
       "st1b { z22.b }, p4, [x20, #-7, MUL VL]\n"
@@ -145,7 +145,7 @@ void sme_transpose_interleave_16VL_1x4(uint8_t *out, const uint8_t *in, size_t w
       ".inst 0xd503467f  // SMSTOP\n"
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp
index 896288cdda..0e34bf143b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_16VL_2x2_fp32bf16.hpp
@@ -43,133 +43,133 @@ void sme_transpose_interleave_16VL_2x2_fp32bf16(bfloat16 *out, const float *in,
       "ptrue p7.b\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
-      "cmp %x[height], #0x1\n"
       "add x24, x25, %x[in_stride]\n"
-      "mov x23, %x[out]\n"
+      "cmp %x[height], #0x1\n"
       "add %x[in], x24, %x[in_stride]\n"
+      "mov x23, %x[out]\n"
       "csel x24, x24, %x[pad_row], GT\n"
       "sub %x[height], %x[height], #0x2\n"
       "mov x22, %x[width]\n"
       "2:"  // Main row loop: Column loop
       "mov x21, x22\n"
-      "mov x20, x23\n"
       "whilelt p1.s, XZR, x21\n"
+      "ld1w { z16.s }, p1/Z, [x25]\n"
+      ".inst 0x658abe00  // bfcvt z0.h, p7/M, z16.s\n"
       "decw x21\n"
       "whilelt p0.s, XZR, x21\n"
+      "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
+      ".inst 0x658abe1f  // bfcvt z31.h, p7/M, z16.s\n"
       "decw x21\n"
-      "ld1w { z16.s }, p1/Z, [x25]\n"
       "whilelt p6.s, XZR, x21\n"
+      "ld1w { z16.s }, p6/Z, [x25, #2, MUL VL]\n"
+      ".inst 0x658abe1e  // bfcvt z30.h, p7/M, z16.s\n"
       "decw x21\n"
-      "ld1w { z18.s }, p0/Z, [x25, #1, MUL VL]\n"
       "whilelt p5.s, XZR, x21\n"
+      "ld1w { z16.s }, p5/Z, [x25, #3, MUL VL]\n"
+      ".inst 0x658abe1d  // bfcvt z29.h, p7/M, z16.s\n"
       "decw x21\n"
-      "ld1w { z17.s }, p6/Z, [x25, #2, MUL VL]\n"
       "whilelt p4.s, XZR, x21\n"
+      "ld1w { z16.s }, p4/Z, [x25, #4, MUL VL]\n"
+      ".inst 0x658abe1c  // bfcvt z28.h, p7/M, z16.s\n"
       "decw x21\n"
-      "ld1w { z19.s }, p5/Z, [x25, #3, MUL VL]\n"
-      ".inst 0x658abe03  // bfcvt z3.h, p7/M, z16.s\n"
       "whilelt p3.s, XZR, x21\n"
+      "ld1w { z16.s }, p3/Z, [x25, #5, MUL VL]\n"
+      ".inst 0x658abe1b  // bfcvt z27.h, p7/M, z16.s\n"
       "decw x21\n"
-      "ld1w { z16.s }, p4/Z, [x25, #4, MUL VL]\n"
-      ".inst 0x658abe42  // bfcvt z2.h, p7/M, z18.s\n"
       "whilelt p2.s, XZR, x21\n"
+      "ld1w { z16.s }, p2/Z, [x25, #6, MUL VL]\n"
+      ".inst 0x658abe1a  // bfcvt z26.h, p7/M, z16.s\n"
       "decw x21\n"
-      "ld1w { z18.s }, p3/Z, [x25, #5, MUL VL]\n"
-      ".inst 0x658abe21  // bfcvt z1.h, p7/M, z17.s\n"
-      "ld1w { z17.s }, p2/Z, [x25, #6, MUL VL]\n"
-      ".inst 0x658abe60  // bfcvt z0.h, p7/M, z19.s\n"
-      "decw x22, ALL, MUL #16\n"
-      "add x23, x23, %x[out_stride]\n"
-      ".inst 0x658abe1f  // bfcvt z31.h, p7/M, z16.s\n"
-      "ld1w { z19.s }, p1/Z, [x24]\n"
+      "ld1w { z16.s }, p1/Z, [x24]\n"
       "whilelt p1.s, XZR, x21\n"
+      ".inst 0x648abe00  // bfcvtnt z0.h, p7/M, z16.s\n"
       "decw x21\n"
       "ld1w { z16.s }, p1/Z, [x25, #7, MUL VL]\n"
       "addvl x25, x25, #16\n"
-      ".inst 0x658abe5e  // bfcvt z30.h, p7/M, z18.s\n"
-      ".inst 0x658abe3d  // bfcvt z29.h, p7/M, z17.s\n"
-      "ld1w { z18.s }, p0/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x658abe19  // bfcvt z25.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n"
       "whilelt p0.s, XZR, x21\n"
       "decw x21\n"
-      "ld1w { z17.s }, p0/Z, [x25, #-8, MUL VL]\n"
-      ".inst 0x648abe63  // bfcvtnt z3.h, p7/M, z19.s\n"
-      ".inst 0x658abe1c  // bfcvt z28.h, p7/M, z16.s\n"
-      "ld1w { z19.s }, p6/Z, [x24, #2, MUL VL]\n"
+      ".inst 0x648abe1f  // bfcvtnt z31.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x25, #-8, MUL VL]\n"
+      ".inst 0x658abe18  // bfcvt z24.h, p7/M, z16.s\n"
+      "mov x20, x23\n"
+      "decw x22, ALL, MUL #16\n"
+      "ld1w { z16.s }, p6/Z, [x24, #2, MUL VL]\n"
       "whilelt p6.s, XZR, x21\n"
       "decw x21\n"
+      ".inst 0x648abe1e  // bfcvtnt z30.h, p7/M, z16.s\n"
       "ld1w { z16.s }, p6/Z, [x25, #-7, MUL VL]\n"
-      ".inst 0x648abe42  // bfcvtnt z2.h, p7/M, z18.s\n"
-      "ld1w { z18.s }, p5/Z, [x24, #3, MUL VL]\n"
+      ".inst 0x658abe17  // bfcvt z23.h, p7/M, z16.s\n"
+      "add x23, x23, %x[out_stride]\n"
+      "ld1w { z16.s }, p5/Z, [x24, #3, MUL VL]\n"
       "whilelt p5.s, XZR, x21\n"
       "decw x21\n"
-      ".inst 0x658abe3b  // bfcvt z27.h, p7/M, z17.s\n"
-      "ld1w { z17.s }, p5/Z, [x25, #-6, MUL VL]\n"
-      ".inst 0x648abe61  // bfcvtnt z1.h, p7/M, z19.s\n"
-      "ld1w { z19.s }, p4/Z, [x24, #4, MUL VL]\n"
+      ".inst 0x648abe1d  // bfcvtnt z29.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p5/Z, [x25, #-6, MUL VL]\n"
+      ".inst 0x658abe16  // bfcvt z22.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p4/Z, [x24, #4, MUL VL]\n"
       "whilelt p4.s, XZR, x21\n"
       "decw x21\n"
-      ".inst 0x658abe1a  // bfcvt z26.h, p7/M, z16.s\n"
+      ".inst 0x648abe1c  // bfcvtnt z28.h, p7/M, z16.s\n"
       "ld1w { z16.s }, p4/Z, [x25, #-5, MUL VL]\n"
-      ".inst 0x648abe40  // bfcvtnt z0.h, p7/M, z18.s\n"
-      "ld1w { z18.s }, p3/Z, [x24, #5, MUL VL]\n"
+      ".inst 0x658abe15  // bfcvt z21.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p3/Z, [x24, #5, MUL VL]\n"
       "whilelt p3.s, XZR, x21\n"
       "decw x21\n"
-      ".inst 0x658abe39  // bfcvt z25.h, p7/M, z17.s\n"
-      "ld1w { z17.s }, p3/Z, [x25, #-4, MUL VL]\n"
-      ".inst 0x648abe7f  // bfcvtnt z31.h, p7/M, z19.s\n"
-      "ld1w { z19.s }, p2/Z, [x24, #6, MUL VL]\n"
+      ".inst 0x648abe1b  // bfcvtnt z27.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p3/Z, [x25, #-4, MUL VL]\n"
+      ".inst 0x658abe14  // bfcvt z20.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x24, #6, MUL VL]\n"
       "whilelt p2.s, XZR, x21\n"
       "decw x21\n"
-      ".inst 0x658abe18  // bfcvt z24.h, p7/M, z16.s\n"
+      ".inst 0x648abe1a  // bfcvtnt z26.h, p7/M, z16.s\n"
       "ld1w { z16.s }, p2/Z, [x25, #-3, MUL VL]\n"
-      ".inst 0x648abe5e  // bfcvtnt z30.h, p7/M, z18.s\n"
-      "ld1w { z18.s }, p1/Z, [x24, #7, MUL VL]\n"
+      ".inst 0x658abe13  // bfcvt z19.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x24, #7, MUL VL]\n"
       "whilelt p1.s, XZR, x21\n"
       "decw x21\n"
-      ".inst 0x658abe37  // bfcvt z23.h, p7/M, z17.s\n"
-      "ld1w { z17.s }, p1/Z, [x25, #-2, MUL VL]\n"
+      ".inst 0x648abe19  // bfcvtnt z25.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #-2, MUL VL]\n"
       "addvl x24, x24, #16\n"
-      ".inst 0x648abe7d  // bfcvtnt z29.h, p7/M, z19.s\n"
-      ".inst 0x658abe16  // bfcvt z22.h, p7/M, z16.s\n"
-      "ld1w { z19.s }, p0/Z, [x24, #-8, MUL VL]\n"
+      ".inst 0x658abe12  // bfcvt z18.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x24, #-8, MUL VL]\n"
       "whilelt p0.s, XZR, x21\n"
       "cmp x22, #0x0\n"
+      ".inst 0x648abe18  // bfcvtnt z24.h, p7/M, z16.s\n"
       "ld1w { z16.s }, p0/Z, [x25, #-1, MUL VL]\n"
-      ".inst 0x648abe5c  // bfcvtnt z28.h, p7/M, z18.s\n"
-      ".inst 0x658abe35  // bfcvt z21.h, p7/M, z17.s\n"
-      "ld1w { z18.s }, p6/Z, [x24, #-7, MUL VL]\n"
-      "ld1w { z17.s }, p5/Z, [x24, #-6, MUL VL]\n"
-      ".inst 0x648abe7b  // bfcvtnt z27.h, p7/M, z19.s\n"
-      ".inst 0x658abe14  // bfcvt z20.h, p7/M, z16.s\n"
+      ".inst 0x658abe11  // bfcvt z17.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p6/Z, [x24, #-7, MUL VL]\n"
+      ".inst 0x648abe17  // bfcvtnt z23.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p5/Z, [x24, #-6, MUL VL]\n"
+      ".inst 0x648abe16  // bfcvtnt z22.h, p7/M, z16.s\n"
       "ld1w { z16.s }, p4/Z, [x24, #-5, MUL VL]\n"
-      "ld1w { z19.s }, p3/Z, [x24, #-4, MUL VL]\n"
-      ".inst 0x648abe5a  // bfcvtnt z26.h, p7/M, z18.s\n"
-      "ld1w { z18.s }, p2/Z, [x24, #-3, MUL VL]\n"
-      ".inst 0x648abe39  // bfcvtnt z25.h, p7/M, z17.s\n"
-      "ld1w { z17.s }, p1/Z, [x24, #-2, MUL VL]\n"
-      ".inst 0x648abe18  // bfcvtnt z24.h, p7/M, z16.s\n"
-      "ld1w { z16.s }, p0/Z, [x24, #-1, MUL VL]\n"
-      "st1h { z3.h }, p7, [x20]\n"
-      ".inst 0x648abe77  // bfcvtnt z23.h, p7/M, z19.s\n"
-      "st1h { z2.h }, p7, [x20, #1, MUL VL]\n"
-      ".inst 0x648abe56  // bfcvtnt z22.h, p7/M, z18.s\n"
-      "st1h { z1.h }, p7, [x20, #2, MUL VL]\n"
-      ".inst 0x648abe35  // bfcvtnt z21.h, p7/M, z17.s\n"
-      "st1h { z0.h }, p7, [x20, #3, MUL VL]\n"
+      ".inst 0x648abe15  // bfcvtnt z21.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p3/Z, [x24, #-4, MUL VL]\n"
       ".inst 0x648abe14  // bfcvtnt z20.h, p7/M, z16.s\n"
-      "st1h { z31.h }, p7, [x20, #4, MUL VL]\n"
-      "st1h { z30.h }, p7, [x20, #5, MUL VL]\n"
-      "st1h { z29.h }, p7, [x20, #6, MUL VL]\n"
-      "st1h { z28.h }, p7, [x20, #7, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x24, #-3, MUL VL]\n"
+      ".inst 0x648abe13  // bfcvtnt z19.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x24, #-2, MUL VL]\n"
+      ".inst 0x648abe12  // bfcvtnt z18.h, p7/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x24, #-1, MUL VL]\n"
+      "st1h { z0.h }, p7, [x20]\n"
+      ".inst 0x648abe11  // bfcvtnt z17.h, p7/M, z16.s\n"
+      "st1h { z31.h }, p7, [x20, #1, MUL VL]\n"
+      "st1h { z30.h }, p7, [x20, #2, MUL VL]\n"
+      "st1h { z29.h }, p7, [x20, #3, MUL VL]\n"
+      "st1h { z28.h }, p7, [x20, #4, MUL VL]\n"
+      "st1h { z27.h }, p7, [x20, #5, MUL VL]\n"
+      "st1h { z26.h }, p7, [x20, #6, MUL VL]\n"
+      "st1h { z25.h }, p7, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
-      "st1h { z27.h }, p7, [x20, #-8, MUL VL]\n"
-      "st1h { z26.h }, p7, [x20, #-7, MUL VL]\n"
-      "st1h { z25.h }, p7, [x20, #-6, MUL VL]\n"
-      "st1h { z24.h }, p7, [x20, #-5, MUL VL]\n"
-      "st1h { z23.h }, p7, [x20, #-4, MUL VL]\n"
-      "st1h { z22.h }, p7, [x20, #-3, MUL VL]\n"
-      "st1h { z21.h }, p7, [x20, #-2, MUL VL]\n"
-      "st1h { z20.h }, p7, [x20, #-1, MUL VL]\n"
+      "st1h { z24.h }, p7, [x20, #-8, MUL VL]\n"
+      "st1h { z23.h }, p7, [x20, #-7, MUL VL]\n"
+      "st1h { z22.h }, p7, [x20, #-6, MUL VL]\n"
+      "st1h { z21.h }, p7, [x20, #-5, MUL VL]\n"
+      "st1h { z20.h }, p7, [x20, #-4, MUL VL]\n"
+      "st1h { z19.h }, p7, [x20, #-3, MUL VL]\n"
+      "st1h { z18.h }, p7, [x20, #-2, MUL VL]\n"
+      "st1h { z17.h }, p7, [x20, #-1, MUL VL]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL.hpp
index 1ece4005d6..36b364a57b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL.hpp
@@ -39,34 +39,34 @@ void sme_transpose_interleave_1VL(uint16_t *out, const uint16_t *in, size_t widt
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
-      "add x24, x26, %x[in_stride]\n"
-      "cnth x23, ALL, MUL #4\n"
-      "add x21, x24, %x[in_stride]\n"
-      "cmp x25, x23\n"
-      "add x20, x21, %x[in_stride]\n"
-      "mov x22, %x[out]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
+      "cnth x21, ALL, MUL #4\n"
+      "add x20, x24, %x[in_stride]\n"
+      "cmp x23, x21\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
       "ld1h { z31.h }, p1/Z, [x26]\n"
-      "sub x25, x25, x23\n"
+      "sub x23, x23, x21\n"
+      "cmp x23, x21\n"
       "ld1h { z30.h }, p1/Z, [x26, #1, MUL VL]\n"
-      "cmp x25, x23\n"
       "ld1h { z29.h }, p1/Z, [x26, #2, MUL VL]\n"
       "ld1h { z28.h }, p1/Z, [x26, #3, MUL VL]\n"
       "addvl x26, x26, #4\n"
-      "ld1h { z27.h }, p1/Z, [x24]\n"
-      "ld1h { z26.h }, p1/Z, [x24, #1, MUL VL]\n"
-      "ld1h { z25.h }, p1/Z, [x24, #2, MUL VL]\n"
-      "ld1h { z24.h }, p1/Z, [x24, #3, MUL VL]\n"
+      "ld1h { z27.h }, p1/Z, [x25]\n"
+      "ld1h { z26.h }, p1/Z, [x25, #1, MUL VL]\n"
+      "ld1h { z25.h }, p1/Z, [x25, #2, MUL VL]\n"
+      "ld1h { z24.h }, p1/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      "ld1h { z23.h }, p1/Z, [x24]\n"
+      "ld1h { z22.h }, p1/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z21.h }, p1/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z20.h }, p1/Z, [x24, #3, MUL VL]\n"
       "addvl x24, x24, #4\n"
-      "ld1h { z23.h }, p1/Z, [x21]\n"
-      "ld1h { z22.h }, p1/Z, [x21, #1, MUL VL]\n"
-      "ld1h { z21.h }, p1/Z, [x21, #2, MUL VL]\n"
-      "ld1h { z20.h }, p1/Z, [x21, #3, MUL VL]\n"
-      "addvl x21, x21, #4\n"
       "ld1h { z19.h }, p1/Z, [x20]\n"
       "ld1h { z18.h }, p1/Z, [x20, #1, MUL VL]\n"
       "ld1h { z17.h }, p1/Z, [x20, #2, MUL VL]\n"
@@ -94,17 +94,17 @@ void sme_transpose_interleave_1VL(uint16_t *out, const uint16_t *in, size_t widt
       "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.h, XZR, x25\n"
-      "dech x25\n"
-      "cmp x25, #0x0\n"
+      "whilelt p0.h, XZR, x23\n"
+      "dech x23\n"
       "ld1h { z19.h }, p0/Z, [x26]\n"
+      "cmp x23, #0x0\n"
       "addvl x26, x26, #1\n"
-      "ld1h { z18.h }, p0/Z, [x24]\n"
+      "ld1h { z18.h }, p0/Z, [x25]\n"
+      "addvl x25, x25, #1\n"
+      "ld1h { z17.h }, p0/Z, [x24]\n"
       "addvl x24, x24, #1\n"
-      "ld1h { z17.h }, p0/Z, [x21]\n"
-      "addvl x21, x21, #1\n"
       "ld1h { z16.h }, p0/Z, [x20]\n"
       "addvl x20, x20, #1\n"
       "st1h { z19.h }, p1, [x22]\n"
@@ -131,8 +131,8 @@ void sme_transpose_interleave_1VL(uint16_t *out, const uint16_t *in, size_t widt
       "8:"  // Tail row loop: Unroll column loop
       "ld1h { z19.h }, p1/Z, [x26]\n"
       "sub x21, x21, x20\n"
-      "ld1h { z18.h }, p1/Z, [x26, #1, MUL VL]\n"
       "cmp x21, x20\n"
+      "ld1h { z18.h }, p1/Z, [x26, #1, MUL VL]\n"
       "ld1h { z17.h }, p1/Z, [x26, #2, MUL VL]\n"
       "ld1h { z16.h }, p1/Z, [x26, #3, MUL VL]\n"
       "st1h { z19.h }, p1, [x22]\n"
@@ -150,10 +150,10 @@ void sme_transpose_interleave_1VL(uint16_t *out, const uint16_t *in, size_t widt
       "10:"  // Tail row loop: Column loop
       "whilelt p0.h, XZR, x21\n"
       "dech x21\n"
-      "cmp x21, #0x0\n"
       "ld1h { z16.h }, p0/Z, [x26]\n"
-      "addvl x26, x26, #1\n"
       "st1h { z16.h }, p1, [x22]\n"
+      "cmp x21, #0x0\n"
+      "addvl x26, x26, #1\n"
       "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_1x4.hpp
index ff8cfc7efe..d67e353f18 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_1x4.hpp
@@ -43,65 +43,65 @@ void sme_transpose_interleave_1VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "ptrue p1.b\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "cmp %x[height], #0x3\n"
       "add x25, x26, %x[in_stride]\n"
-      "mov x24, %x[width]\n"
-      "add x23, x25, %x[in_stride]\n"
-      "cntb x22\n"
-      "add x21, x23, %x[in_stride]\n"
-      "csel x23, x23, %x[pad_row], GE\n"
-      "add %x[in], x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GT\n"
+      "add x24, x25, %x[in_stride]\n"
+      "add x23, x24, %x[in_stride]\n"
+      "cmp %x[height], #0x3\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "csel x23, x23, %x[pad_row], GT\n"
+      "csel x24, x24, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "mov x20, %x[out]\n"
+      "mov x22, %x[width]\n"
+      "cntb x21\n"
       "csel x25, x25, %x[pad_row], GT\n"
-      "cmp x24, x22\n"
+      "cmp x22, x21\n"
+      "mov x20, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1b { z20.b }, p1/Z, [x26]\n"
-      "sub x24, x24, x22\n"
+      "ld1b { z17.b }, p1/Z, [x26]\n"
+      "sub x22, x22, x21\n"
+      "cmp x22, x21\n"
+      "ld1b { z18.b }, p1/Z, [x25]\n"
       "addvl x26, x26, #1\n"
-      "ld1b { z19.b }, p1/Z, [x25]\n"
-      "cmp x24, x22\n"
       "addvl x25, x25, #1\n"
-      "ld1b { z17.b }, p1/Z, [x23]\n"
+      "ld1b { z16.b }, p1/Z, [x24]\n"
+      "zip1 z20.b, z17.b, z16.b\n"
+      "zip2 z19.b, z17.b, z16.b\n"
+      "addvl x24, x24, #1\n"
+      "ld1b { z16.b }, p1/Z, [x23]\n"
+      "zip1 z17.b, z18.b, z16.b\n"
+      "zip2 z18.b, z18.b, z16.b\n"
       "addvl x23, x23, #1\n"
-      "ld1b { z16.b }, p1/Z, [x21]\n"
-      "addvl x21, x21, #1\n"
-      "zip1 z18.b, z20.b, z17.b\n"
-      "zip2 z20.b, z20.b, z17.b\n"
-      "zip1 z17.b, z19.b, z16.b\n"
-      "zip2 z16.b, z19.b, z16.b\n"
-      "zip1 z19.b, z18.b, z17.b\n"
-      "zip2 z18.b, z18.b, z17.b\n"
-      "zip1 z17.b, z20.b, z16.b\n"
-      "zip2 z16.b, z20.b, z16.b\n"
-      "st1b { z19.b }, p1, [x20]\n"
+      "zip1 z16.b, z20.b, z17.b\n"
+      "st1b { z16.b }, p1, [x20]\n"
       "add x20, x20, %x[out_stride]\n"
-      "st1b { z18.b }, p1, [x20]\n"
+      "zip2 z16.b, z20.b, z17.b\n"
+      "st1b { z16.b }, p1, [x20]\n"
       "add x20, x20, %x[out_stride]\n"
+      "zip1 z17.b, z19.b, z18.b\n"
+      "zip2 z16.b, z19.b, z18.b\n"
       "st1b { z17.b }, p1, [x20]\n"
       "add x20, x20, %x[out_stride]\n"
       "st1b { z16.b }, p1, [x20]\n"
       "add x20, x20, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x24, 5f\n"
+      "cbz x22, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.b, XZR, x24\n"
-      "decw x24\n"
-      "ld1b { z19.b }, p0/Z, [x26]\n"
-      "cmp x24, #0x0\n"
-      "incd x26, ALL, MUL #2\n"
+      "whilelt p0.b, XZR, x22\n"
+      "ld1b { z17.b }, p0/Z, [x26]\n"
+      "decw x22\n"
       "ld1b { z18.b }, p0/Z, [x25]\n"
+      "cmp x22, #0x0\n"
+      "incd x26, ALL, MUL #2\n"
+      "ld1b { z16.b }, p0/Z, [x24]\n"
+      "zip1 z17.b, z17.b, z16.b\n"
       "incd x25, ALL, MUL #2\n"
-      "ld1b { z17.b }, p0/Z, [x23]\n"
-      "incd x23, ALL, MUL #2\n"
-      "ld1b { z16.b }, p0/Z, [x21]\n"
-      "incd x21, ALL, MUL #2\n"
-      "zip1 z17.b, z19.b, z17.b\n"
+      "incd x24, ALL, MUL #2\n"
+      "ld1b { z16.b }, p0/Z, [x23]\n"
       "zip1 z16.b, z18.b, z16.b\n"
+      "incd x23, ALL, MUL #2\n"
       "zip1 z16.b, z17.b, z16.b\n"
       "st1b { z16.b }, p1, [x20]\n"
       "add x20, x20, %x[out_stride]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2.hpp
index 54c2af1a84..f8980d25f6 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2.hpp
@@ -45,69 +45,69 @@ void sme_transpose_interleave_1VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
-      "add x24, x26, %x[in_stride]\n"
-      "cnth x23, ALL, MUL #2\n"
-      "add x21, x24, %x[in_stride]\n"
-      "cmp x25, x23\n"
-      "add x20, x21, %x[in_stride]\n"
-      "mov x22, %x[out]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
+      "cnth x21, ALL, MUL #2\n"
+      "add x20, x24, %x[in_stride]\n"
+      "cmp x23, x21\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1h { z18.h }, p1/Z, [x26]\n"
-      "sub x25, x25, x23\n"
+      "ld1h { z17.h }, p1/Z, [x26]\n"
+      "sub x23, x23, x21\n"
+      "cmp x23, x21\n"
+      "ld1h { z16.h }, p1/Z, [x25]\n"
+      "zip1 z24.h, z17.h, z16.h\n"
+      "zip2 z23.h, z17.h, z16.h\n"
       "ld1h { z17.h }, p1/Z, [x24]\n"
-      "cmp x25, x23\n"
-      "ld1h { z20.h }, p1/Z, [x21]\n"
       "ld1h { z16.h }, p1/Z, [x20]\n"
-      "ld1h { z23.h }, p1/Z, [x26, #1, MUL VL]\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z22.h, z18.h, z17.h\n"
+      "zip1 z22.h, z17.h, z16.h\n"
+      "zip2 z21.h, z17.h, z16.h\n"
+      "ld1h { z17.h }, p1/Z, [x26, #1, MUL VL]\n"
       "addvl x26, x26, #2\n"
+      "ld1h { z16.h }, p1/Z, [x25, #1, MUL VL]\n"
+      "zip1 z20.h, z17.h, z16.h\n"
+      "addvl x25, x25, #2\n"
+      "zip2 z19.h, z17.h, z16.h\n"
       "ld1h { z18.h }, p1/Z, [x24, #1, MUL VL]\n"
       "addvl x24, x24, #2\n"
-      "ld1h { z21.h }, p1/Z, [x21, #1, MUL VL]\n"
-      "zip1 z17.h, z20.h, z16.h\n"
-      "zip2 z20.h, z20.h, z16.h\n"
-      "addvl x21, x21, #2\n"
       "ld1h { z16.h }, p1/Z, [x20, #1, MUL VL]\n"
+      "st1h { z24.h }, p1, [x22]\n"
+      "zip1 z17.h, z18.h, z16.h\n"
       "addvl x20, x20, #2\n"
-      "st1h { z19.h }, p1, [x22]\n"
-      "zip1 z19.h, z23.h, z18.h\n"
-      "zip2 z18.h, z23.h, z18.h\n"
-      "st1h { z17.h }, p1, [x22, #1, MUL VL]\n"
+      "st1h { z22.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      "zip1 z17.h, z21.h, z16.h\n"
-      "zip2 z16.h, z21.h, z16.h\n"
-      "st1h { z22.h }, p1, [x22]\n"
-      "st1h { z20.h }, p1, [x22, #1, MUL VL]\n"
+      "zip2 z16.h, z18.h, z16.h\n"
+      "st1h { z23.h }, p1, [x22]\n"
+      "st1h { z21.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      "st1h { z19.h }, p1, [x22]\n"
+      "st1h { z20.h }, p1, [x22]\n"
       "st1h { z17.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      "st1h { z18.h }, p1, [x22]\n"
+      "st1h { z19.h }, p1, [x22]\n"
       "st1h { z16.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.h, XZR, x25\n"
-      "decw x25\n"
-      "ld1h { z19.h }, p0/Z, [x26]\n"
-      "cmp x25, #0x0\n"
+      "whilelt p0.h, XZR, x23\n"
+      "ld1h { z17.h }, p0/Z, [x26]\n"
+      "decw x23\n"
+      "ld1h { z16.h }, p0/Z, [x25]\n"
+      "cmp x23, #0x0\n"
       "incd x26, ALL, MUL #4\n"
+      "zip1 z18.h, z17.h, z16.h\n"
       "ld1h { z17.h }, p0/Z, [x24]\n"
+      "incd x25, ALL, MUL #4\n"
       "incd x24, ALL, MUL #4\n"
-      "ld1h { z18.h }, p0/Z, [x21]\n"
-      "incd x21, ALL, MUL #4\n"
       "ld1h { z16.h }, p0/Z, [x20]\n"
       "incd x20, ALL, MUL #4\n"
-      "zip1 z17.h, z19.h, z17.h\n"
-      "zip1 z16.h, z18.h, z16.h\n"
-      "st1h { z17.h }, p1, [x22]\n"
+      "zip1 z16.h, z17.h, z16.h\n"
+      "st1h { z18.h }, p1, [x22]\n"
       "st1h { z16.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bgt 4b\n"
@@ -119,12 +119,12 @@ void sme_transpose_interleave_1VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "add x24, x26, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #2\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
+      "add %x[in], x25, %x[in_stride]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "cmp x21, x20\n"
       "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x2\n"
@@ -132,20 +132,20 @@ void sme_transpose_interleave_1VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "8:"  // Tail row loop: Unroll column loop
       "ld1h { z18.h }, p1/Z, [x26]\n"
       "sub x21, x21, x20\n"
-      "ld1h { z17.h }, p1/Z, [x24]\n"
       "cmp x21, x20\n"
-      "ld1h { z20.h }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1h { z16.h }, p1/Z, [x25]\n"
+      "zip1 z17.h, z18.h, z16.h\n"
+      "zip2 z19.h, z18.h, z16.h\n"
+      "ld1h { z18.h }, p1/Z, [x26, #1, MUL VL]\n"
       "addvl x26, x26, #2\n"
-      "ld1h { z16.h }, p1/Z, [x24, #1, MUL VL]\n"
-      "addvl x24, x24, #2\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "zip1 z17.h, z20.h, z16.h\n"
-      "zip2 z16.h, z20.h, z16.h\n"
-      "st1h { z19.h }, p1, [x22]\n"
+      "ld1h { z16.h }, p1/Z, [x25, #1, MUL VL]\n"
+      "st1h { z17.h }, p1, [x22]\n"
       "add x22, x22, %x[out_stride]\n"
-      "st1h { z18.h }, p1, [x22]\n"
+      "zip1 z17.h, z18.h, z16.h\n"
+      "st1h { z19.h }, p1, [x22]\n"
       "add x22, x22, %x[out_stride]\n"
+      "addvl x25, x25, #2\n"
+      "zip2 z16.h, z18.h, z16.h\n"
       "st1h { z17.h }, p1, [x22]\n"
       "add x22, x22, %x[out_stride]\n"
       "st1h { z16.h }, p1, [x22]\n"
@@ -155,13 +155,13 @@ void sme_transpose_interleave_1VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "whilelt p0.h, XZR, x21\n"
-      "decw x21\n"
       "ld1h { z17.h }, p0/Z, [x26]\n"
+      "decw x21\n"
+      "ld1h { z16.h }, p0/Z, [x25]\n"
       "cmp x21, #0x0\n"
       "incd x26, ALL, MUL #4\n"
-      "ld1h { z16.h }, p0/Z, [x24]\n"
-      "incd x24, ALL, MUL #4\n"
       "zip1 z16.h, z17.h, z16.h\n"
+      "incd x25, ALL, MUL #4\n"
       "st1h { z16.h }, p1, [x22]\n"
       "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
@@ -205,5 +205,4 @@ void Transform<1, 2, true, VLType::SME>(
     );
 }
 
-
 #endif  // defined(ARM_COMPUTE_ENABLE_SME)
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp
index 2fafefbbc5..c740a9c64b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp
@@ -45,88 +45,88 @@ void sme_transpose_interleave_1VL_2x2_fp32bf16(bfloat16 *out, const float *in, s
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
-      "add x24, x26, %x[in_stride]\n"
-      "cnth x23, ALL, MUL #2\n"
-      "add x21, x24, %x[in_stride]\n"
-      "cmp x25, x23\n"
-      "add x20, x21, %x[in_stride]\n"
-      "mov x22, %x[out]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
+      "cnth x21, ALL, MUL #2\n"
+      "add x20, x24, %x[in_stride]\n"
+      "cmp x23, x21\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1w { z19.s }, p1/Z, [x26]\n"
-      "sub x25, x25, x23\n"
-      "ld1w { z18.s }, p1/Z, [x21]\n"
-      "cmp x25, x23\n"
-      "ld1w { z17.s }, p1/Z, [x26, #1, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x21, #1, MUL VL]\n"
-      ".inst 0x658aa67b  // bfcvt z27.h, p1/M, z19.s\n"
-      "ld1w { z19.s }, p1/Z, [x26, #2, MUL VL]\n"
-      ".inst 0x658aa65a  // bfcvt z26.h, p1/M, z18.s\n"
-      "ld1w { z18.s }, p1/Z, [x21, #2, MUL VL]\n"
-      ".inst 0x658aa639  // bfcvt z25.h, p1/M, z17.s\n"
-      "ld1w { z17.s }, p1/Z, [x26, #3, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x26]\n"
       ".inst 0x658aa618  // bfcvt z24.h, p1/M, z16.s\n"
-      "addvl x26, x26, #4\n"
-      "ld1w { z16.s }, p1/Z, [x21, #3, MUL VL]\n"
-      ".inst 0x658aa677  // bfcvt z23.h, p1/M, z19.s\n"
-      "addvl x21, x21, #4\n"
-      "ld1w { z19.s }, p1/Z, [x24]\n"
-      ".inst 0x658aa656  // bfcvt z22.h, p1/M, z18.s\n"
-      "ld1w { z18.s }, p1/Z, [x20]\n"
-      ".inst 0x658aa635  // bfcvt z21.h, p1/M, z17.s\n"
-      "ld1w { z17.s }, p1/Z, [x24, #1, MUL VL]\n"
+      "sub x23, x23, x21\n"
+      "cmp x23, x21\n"
+      "ld1w { z16.s }, p1/Z, [x24]\n"
+      ".inst 0x658aa617  // bfcvt z23.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x658aa616  // bfcvt z22.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x658aa615  // bfcvt z21.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n"
       ".inst 0x658aa614  // bfcvt z20.h, p1/M, z16.s\n"
-      "ld1w { z16.s }, p1/Z, [x20, #1, MUL VL]\n"
-      ".inst 0x648aa67b  // bfcvtnt z27.h, p1/M, z19.s\n"
-      "ld1w { z19.s }, p1/Z, [x24, #2, MUL VL]\n"
-      ".inst 0x648aa65a  // bfcvtnt z26.h, p1/M, z18.s\n"
-      "ld1w { z18.s }, p1/Z, [x20, #2, MUL VL]\n"
-      ".inst 0x648aa639  // bfcvtnt z25.h, p1/M, z17.s\n"
-      "ld1w { z17.s }, p1/Z, [x24, #3, MUL VL]\n"
-      ".inst 0x648aa618  // bfcvtnt z24.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x24, #2, MUL VL]\n"
+      ".inst 0x658aa613  // bfcvt z19.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x26, #3, MUL VL]\n"
+      ".inst 0x658aa612  // bfcvt z18.h, p1/M, z16.s\n"
+      "addvl x26, x26, #4\n"
+      "ld1w { z16.s }, p1/Z, [x24, #3, MUL VL]\n"
+      ".inst 0x658aa611  // bfcvt z17.h, p1/M, z16.s\n"
       "addvl x24, x24, #4\n"
+      "ld1w { z16.s }, p1/Z, [x25]\n"
+      ".inst 0x648aa618  // bfcvtnt z24.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x20]\n"
+      ".inst 0x648aa617  // bfcvtnt z23.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #1, MUL VL]\n"
+      ".inst 0x648aa616  // bfcvtnt z22.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x20, #1, MUL VL]\n"
+      ".inst 0x648aa615  // bfcvtnt z21.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n"
+      ".inst 0x648aa614  // bfcvtnt z20.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x20, #2, MUL VL]\n"
+      ".inst 0x648aa613  // bfcvtnt z19.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0x648aa612  // bfcvtnt z18.h, p1/M, z16.s\n"
       "ld1w { z16.s }, p1/Z, [x20, #3, MUL VL]\n"
-      "st1h { z27.h }, p1, [x22]\n"
-      ".inst 0x648aa677  // bfcvtnt z23.h, p1/M, z19.s\n"
+      "st1h { z24.h }, p1, [x22]\n"
       "addvl x20, x20, #4\n"
-      "st1h { z26.h }, p1, [x22, #1, MUL VL]\n"
+      ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
+      "st1h { z23.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      ".inst 0x648aa656  // bfcvtnt z22.h, p1/M, z18.s\n"
-      "st1h { z25.h }, p1, [x22]\n"
-      ".inst 0x648aa635  // bfcvtnt z21.h, p1/M, z17.s\n"
-      "st1h { z24.h }, p1, [x22, #1, MUL VL]\n"
+      "st1h { z22.h }, p1, [x22]\n"
+      "st1h { z21.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      ".inst 0x648aa614  // bfcvtnt z20.h, p1/M, z16.s\n"
-      "st1h { z23.h }, p1, [x22]\n"
-      "st1h { z22.h }, p1, [x22, #1, MUL VL]\n"
+      "st1h { z20.h }, p1, [x22]\n"
+      "st1h { z19.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      "st1h { z21.h }, p1, [x22]\n"
-      "st1h { z20.h }, p1, [x22, #1, MUL VL]\n"
+      "st1h { z18.h }, p1, [x22]\n"
+      "st1h { z17.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.s, XZR, x25\n"
-      "decw x25\n"
-      "ld1w { z17.s }, p0/Z, [x26]\n"
-      "cmp x25, #0x0\n"
+      "whilelt p0.s, XZR, x23\n"
+      "ld1w { z16.s }, p0/Z, [x26]\n"
+      ".inst 0x658aa612  // bfcvt z18.h, p1/M, z16.s\n"
+      "decw x23\n"
+      "ld1w { z16.s }, p0/Z, [x24]\n"
+      ".inst 0x658aa611  // bfcvt z17.h, p1/M, z16.s\n"
+      "cmp x23, #0x0\n"
       "addvl x26, x26, #1\n"
-      "ld1w { z16.s }, p0/Z, [x21]\n"
-      "addvl x21, x21, #1\n"
-      "ld1w { z19.s }, p0/Z, [x24]\n"
+      "ld1w { z16.s }, p0/Z, [x25]\n"
+      "addvl x25, x25, #1\n"
       "addvl x24, x24, #1\n"
-      ".inst 0x658aa632  // bfcvt z18.h, p1/M, z17.s\n"
-      "ld1w { z17.s }, p0/Z, [x20]\n"
+      ".inst 0x648aa612  // bfcvtnt z18.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x20]\n"
       "addvl x20, x20, #1\n"
-      ".inst 0x658aa610  // bfcvt z16.h, p1/M, z16.s\n"
-      ".inst 0x648aa672  // bfcvtnt z18.h, p1/M, z19.s\n"
-      ".inst 0x648aa630  // bfcvtnt z16.h, p1/M, z17.s\n"
+      ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
       "st1h { z18.h }, p1, [x22]\n"
-      "st1h { z16.h }, p1, [x22, #1, MUL VL]\n"
+      "st1h { z17.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
@@ -137,43 +137,43 @@ void sme_transpose_interleave_1VL_2x2_fp32bf16(bfloat16 *out, const float *in, s
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "add x24, x26, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #2\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
+      "add %x[in], x25, %x[in_stride]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "cmp x21, x20\n"
       "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x2\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1w { z19.s }, p1/Z, [x26]\n"
+      "ld1w { z16.s }, p1/Z, [x26]\n"
+      ".inst 0x658aa614  // bfcvt z20.h, p1/M, z16.s\n"
       "sub x21, x21, x20\n"
-      "ld1w { z18.s }, p1/Z, [x26, #1, MUL VL]\n"
       "cmp x21, x20\n"
-      "ld1w { z17.s }, p1/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x658aa613  // bfcvt z19.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n"
+      ".inst 0x658aa612  // bfcvt z18.h, p1/M, z16.s\n"
       "ld1w { z16.s }, p1/Z, [x26, #3, MUL VL]\n"
-      ".inst 0x658aa677  // bfcvt z23.h, p1/M, z19.s\n"
-      "addvl x26, x26, #4\n"
-      "ld1w { z22.s }, p1/Z, [x24]\n"
-      ".inst 0x658aa655  // bfcvt z21.h, p1/M, z18.s\n"
-      "ld1w { z20.s }, p1/Z, [x24, #1, MUL VL]\n"
-      ".inst 0x658aa633  // bfcvt z19.h, p1/M, z17.s\n"
-      "ld1w { z18.s }, p1/Z, [x24, #2, MUL VL]\n"
       ".inst 0x658aa611  // bfcvt z17.h, p1/M, z16.s\n"
-      "ld1w { z16.s }, p1/Z, [x24, #3, MUL VL]\n"
-      ".inst 0x648aa6d7  // bfcvtnt z23.h, p1/M, z22.s\n"
-      "addvl x24, x24, #4\n"
-      ".inst 0x648aa695  // bfcvtnt z21.h, p1/M, z20.s\n"
-      ".inst 0x648aa653  // bfcvtnt z19.h, p1/M, z18.s\n"
-      ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
-      "st1h { z23.h }, p1, [x22]\n"
-      "add x22, x22, %x[out_stride]\n"
-      "st1h { z21.h }, p1, [x22]\n"
+      "addvl x26, x26, #4\n"
+      "ld1w { z16.s }, p1/Z, [x25]\n"
+      ".inst 0x648aa614  // bfcvtnt z20.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #1, MUL VL]\n"
+      ".inst 0x648aa613  // bfcvtnt z19.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n"
+      ".inst 0x648aa612  // bfcvtnt z18.h, p1/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #3, MUL VL]\n"
+      "st1h { z20.h }, p1, [x22]\n"
       "add x22, x22, %x[out_stride]\n"
+      "addvl x25, x25, #4\n"
       "st1h { z19.h }, p1, [x22]\n"
       "add x22, x22, %x[out_stride]\n"
+      ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
+      "st1h { z18.h }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
       "st1h { z17.h }, p1, [x22]\n"
       "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
@@ -181,15 +181,15 @@ void sme_transpose_interleave_1VL_2x2_fp32bf16(bfloat16 *out, const float *in, s
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "whilelt p0.s, XZR, x21\n"
-      "decw x21\n"
       "ld1w { z16.s }, p0/Z, [x26]\n"
+      ".inst 0x658aa611  // bfcvt z17.h, p1/M, z16.s\n"
+      "decw x21\n"
+      "ld1w { z16.s }, p0/Z, [x25]\n"
       "cmp x21, #0x0\n"
       "addvl x26, x26, #1\n"
-      "ld1w { z17.s }, p0/Z, [x24]\n"
-      "addvl x24, x24, #1\n"
-      ".inst 0x658aa610  // bfcvt z16.h, p1/M, z16.s\n"
-      ".inst 0x648aa630  // bfcvtnt z16.h, p1/M, z17.s\n"
-      "st1h { z16.h }, p1, [x22]\n"
+      ".inst 0x648aa611  // bfcvtnt z17.h, p1/M, z16.s\n"
+      "addvl x25, x25, #1\n"
+      "st1h { z17.h }, p1, [x22]\n"
       "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp
index f981624a1d..f7d29a9f01 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp
@@ -39,34 +39,34 @@ void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t widt
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
-      "add x24, x26, %x[in_stride]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
       "cnth x20, ALL, MUL #4\n"
-      "add x23, x24, %x[in_stride]\n"
-      "cmp x25, x20\n"
-      "add x21, x23, %x[in_stride]\n"
-      "mov x22, %x[out]\n"
+      "add x21, x24, %x[in_stride]\n"
+      "cmp x23, x20\n"
       "add %x[in], x21, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "sub x25, x25, x20\n"
+      "sub x23, x23, x20\n"
       "ld1h { z31.h }, p2/Z, [x26]\n"
+      "cmp x23, x20\n"
       "ld1h { z30.h }, p2/Z, [x26, #1, MUL VL]\n"
-      "cmp x25, x20\n"
       "ld1h { z29.h }, p2/Z, [x26, #2, MUL VL]\n"
       "ld1h { z28.h }, p2/Z, [x26, #3, MUL VL]\n"
       "addvl x26, x26, #4\n"
-      "ld1h { z27.h }, p2/Z, [x24]\n"
-      "ld1h { z26.h }, p2/Z, [x24, #1, MUL VL]\n"
-      "ld1h { z25.h }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1h { z24.h }, p2/Z, [x24, #3, MUL VL]\n"
+      "ld1h { z27.h }, p2/Z, [x25]\n"
+      "ld1h { z26.h }, p2/Z, [x25, #1, MUL VL]\n"
+      "ld1h { z25.h }, p2/Z, [x25, #2, MUL VL]\n"
+      "ld1h { z24.h }, p2/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      "ld1h { z23.h }, p2/Z, [x24]\n"
+      "ld1h { z22.h }, p2/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z21.h }, p2/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z20.h }, p2/Z, [x24, #3, MUL VL]\n"
       "addvl x24, x24, #4\n"
-      "ld1h { z23.h }, p2/Z, [x23]\n"
-      "ld1h { z22.h }, p2/Z, [x23, #1, MUL VL]\n"
-      "ld1h { z21.h }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z20.h }, p2/Z, [x23, #3, MUL VL]\n"
-      "addvl x23, x23, #4\n"
       "ld1h { z19.h }, p2/Z, [x21]\n"
       "ld1h { z18.h }, p2/Z, [x21, #1, MUL VL]\n"
       "ld1h { z17.h }, p2/Z, [x21, #2, MUL VL]\n"
@@ -92,29 +92,29 @@ void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t widt
       "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x20, x25\n"
-      "dech x25, ALL, MUL #2\n"
+      "mov x20, x23\n"
       "whilelt p1.h, XZR, x20\n"
+      "ld1h { z23.h }, p1/Z, [x26]\n"
       "dech x20\n"
+      "dech x23, ALL, MUL #2\n"
+      "ld1h { z22.h }, p1/Z, [x25]\n"
       "whilelt p0.h, XZR, x20\n"
-      "cmp x25, #0x0\n"
-      "ld1h { z23.h }, p1/Z, [x26]\n"
-      "ld1h { z22.h }, p0/Z, [x26, #1, MUL VL]\n"
+      "cmp x23, #0x0\n"
+      "ld1h { z21.h }, p0/Z, [x26, #1, MUL VL]\n"
       "addvl x26, x26, #2\n"
-      "ld1h { z21.h }, p1/Z, [x24]\n"
-      "ld1h { z20.h }, p0/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z20.h }, p0/Z, [x25, #1, MUL VL]\n"
+      "addvl x25, x25, #2\n"
+      "ld1h { z19.h }, p1/Z, [x24]\n"
+      "ld1h { z18.h }, p0/Z, [x24, #1, MUL VL]\n"
       "addvl x24, x24, #2\n"
-      "ld1h { z19.h }, p1/Z, [x23]\n"
-      "ld1h { z18.h }, p0/Z, [x23, #1, MUL VL]\n"
-      "addvl x23, x23, #2\n"
       "ld1h { z17.h }, p1/Z, [x21]\n"
       "ld1h { z16.h }, p0/Z, [x21, #1, MUL VL]\n"
       "addvl x21, x21, #2\n"
       "st1h { z23.h }, p2, [x22]\n"
-      "st1h { z22.h }, p2, [x22, #1, MUL VL]\n"
-      "st1h { z21.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z21.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x22, #2, MUL VL]\n"
       "st1h { z20.h }, p2, [x22, #3, MUL VL]\n"
       "st1h { z19.h }, p2, [x22, #4, MUL VL]\n"
       "st1h { z18.h }, p2, [x22, #5, MUL VL]\n"
@@ -140,8 +140,8 @@ void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t widt
       "8:"  // Tail row loop: Unroll column loop
       "sub x21, x21, x20\n"
       "ld1h { z19.h }, p2/Z, [x26]\n"
-      "ld1h { z18.h }, p2/Z, [x26, #1, MUL VL]\n"
       "cmp x21, x20\n"
+      "ld1h { z18.h }, p2/Z, [x26, #1, MUL VL]\n"
       "ld1h { z17.h }, p2/Z, [x26, #2, MUL VL]\n"
       "ld1h { z16.h }, p2/Z, [x26, #3, MUL VL]\n"
       "st1h { z19.h }, p2, [x22]\n"
@@ -156,15 +156,15 @@ void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t widt
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "dech x21, ALL, MUL #2\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z17.h }, p0/Z, [x26]\n"
       "dech x20\n"
+      "dech x21, ALL, MUL #2\n"
       "whilelt p0.h, XZR, x20\n"
       "cmp x21, #0x0\n"
-      "ld1h { z17.h }, p1/Z, [x26]\n"
       "ld1h { z16.h }, p0/Z, [x26, #1, MUL VL]\n"
-      "addvl x26, x26, #2\n"
       "st1h { z17.h }, p2, [x22]\n"
+      "addvl x26, x26, #2\n"
       "st1h { z16.h }, p2, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_1x4.hpp
index a7a384c85f..f07d34f46c 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_1x4.hpp
@@ -43,63 +43,63 @@ void sme_transpose_interleave_2VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "ptrue p1.b\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "cmp %x[height], #0x3\n"
       "add x25, x26, %x[in_stride]\n"
-      "mov x24, %x[width]\n"
-      "add x23, x25, %x[in_stride]\n"
-      "cntb x22\n"
-      "add x21, x23, %x[in_stride]\n"
-      "csel x23, x23, %x[pad_row], GE\n"
-      "add %x[in], x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GT\n"
+      "add x24, x25, %x[in_stride]\n"
+      "add x23, x24, %x[in_stride]\n"
+      "cmp %x[height], #0x3\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "csel x23, x23, %x[pad_row], GT\n"
+      "csel x24, x24, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "mov x20, %x[out]\n"
+      "mov x22, %x[width]\n"
+      "cntb x21\n"
       "csel x25, x25, %x[pad_row], GT\n"
-      "cmp x24, x22\n"
+      "cmp x22, x21\n"
+      "mov x20, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1b { z20.b }, p1/Z, [x26]\n"
-      "sub x24, x24, x22\n"
+      "ld1b { z17.b }, p1/Z, [x26]\n"
+      "sub x22, x22, x21\n"
+      "cmp x22, x21\n"
+      "ld1b { z18.b }, p1/Z, [x25]\n"
       "addvl x26, x26, #1\n"
-      "ld1b { z19.b }, p1/Z, [x25]\n"
-      "cmp x24, x22\n"
       "addvl x25, x25, #1\n"
+      "ld1b { z16.b }, p1/Z, [x24]\n"
+      "zip1 z20.b, z17.b, z16.b\n"
+      "zip2 z19.b, z17.b, z16.b\n"
+      "addvl x24, x24, #1\n"
       "ld1b { z17.b }, p1/Z, [x23]\n"
-      "addvl x23, x23, #1\n"
-      "ld1b { z16.b }, p1/Z, [x21]\n"
-      "addvl x21, x21, #1\n"
-      "zip1 z18.b, z20.b, z17.b\n"
-      "zip2 z20.b, z20.b, z17.b\n"
-      "zip1 z17.b, z19.b, z16.b\n"
-      "zip2 z16.b, z19.b, z16.b\n"
-      "zip1 z19.b, z18.b, z17.b\n"
+      "zip1 z16.b, z18.b, z17.b\n"
       "zip2 z18.b, z18.b, z17.b\n"
+      "addvl x23, x23, #1\n"
       "zip1 z17.b, z20.b, z16.b\n"
       "zip2 z16.b, z20.b, z16.b\n"
-      "st1b { z19.b }, p1, [x20]\n"
-      "st1b { z18.b }, p1, [x20, #1, MUL VL]\n"
+      "st1b { z17.b }, p1, [x20]\n"
+      "st1b { z16.b }, p1, [x20, #1, MUL VL]\n"
       "add x20, x20, %x[out_stride]\n"
+      "zip1 z17.b, z19.b, z18.b\n"
+      "zip2 z16.b, z19.b, z18.b\n"
       "st1b { z17.b }, p1, [x20]\n"
       "st1b { z16.b }, p1, [x20, #1, MUL VL]\n"
       "add x20, x20, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x24, 5f\n"
+      "cbz x22, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.b, XZR, x24\n"
-      "decw x24, ALL, MUL #2\n"
+      "whilelt p0.b, XZR, x22\n"
       "ld1b { z18.b }, p0/Z, [x26]\n"
-      "cmp x24, #0x0\n"
+      "decw x22, ALL, MUL #2\n"
+      "ld1b { z17.b }, p0/Z, [x25]\n"
+      "cmp x22, #0x0\n"
       "incd x26, ALL, MUL #4\n"
-      "ld1b { z19.b }, p0/Z, [x25]\n"
+      "ld1b { z16.b }, p0/Z, [x24]\n"
+      "zip1 z18.b, z18.b, z16.b\n"
       "incd x25, ALL, MUL #4\n"
-      "ld1b { z17.b }, p0/Z, [x23]\n"
+      "incd x24, ALL, MUL #4\n"
+      "ld1b { z16.b }, p0/Z, [x23]\n"
+      "zip1 z16.b, z17.b, z16.b\n"
       "incd x23, ALL, MUL #4\n"
-      "ld1b { z16.b }, p0/Z, [x21]\n"
-      "incd x21, ALL, MUL #4\n"
-      "zip1 z18.b, z18.b, z17.b\n"
-      "zip1 z16.b, z19.b, z16.b\n"
       "zip1 z17.b, z18.b, z16.b\n"
       "zip2 z16.b, z18.b, z16.b\n"
       "st1b { z17.b }, p1, [x20]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2.hpp
index 651ae5f061..35d74e727b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2.hpp
@@ -45,69 +45,69 @@ void sme_transpose_interleave_2VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
-      "add x24, x26, %x[in_stride]\n"
-      "cnth x23, ALL, MUL #2\n"
-      "add x21, x24, %x[in_stride]\n"
-      "cmp x25, x23\n"
-      "add x20, x21, %x[in_stride]\n"
-      "mov x22, %x[out]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
+      "cnth x21, ALL, MUL #2\n"
+      "add x20, x24, %x[in_stride]\n"
+      "cmp x23, x21\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1h { z18.h }, p1/Z, [x26]\n"
-      "sub x25, x25, x23\n"
+      "ld1h { z17.h }, p1/Z, [x26]\n"
+      "sub x23, x23, x21\n"
+      "cmp x23, x21\n"
+      "ld1h { z16.h }, p1/Z, [x25]\n"
+      "zip1 z24.h, z17.h, z16.h\n"
+      "zip2 z23.h, z17.h, z16.h\n"
       "ld1h { z17.h }, p1/Z, [x24]\n"
-      "cmp x25, x23\n"
-      "ld1h { z20.h }, p1/Z, [x21]\n"
       "ld1h { z16.h }, p1/Z, [x20]\n"
-      "ld1h { z23.h }, p1/Z, [x26, #1, MUL VL]\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z22.h, z18.h, z17.h\n"
+      "zip1 z22.h, z17.h, z16.h\n"
+      "zip2 z21.h, z17.h, z16.h\n"
+      "ld1h { z17.h }, p1/Z, [x26, #1, MUL VL]\n"
       "addvl x26, x26, #2\n"
+      "ld1h { z16.h }, p1/Z, [x25, #1, MUL VL]\n"
+      "addvl x25, x25, #2\n"
+      "zip1 z20.h, z17.h, z16.h\n"
+      "zip2 z19.h, z17.h, z16.h\n"
       "ld1h { z18.h }, p1/Z, [x24, #1, MUL VL]\n"
       "addvl x24, x24, #2\n"
-      "ld1h { z21.h }, p1/Z, [x21, #1, MUL VL]\n"
-      "zip1 z17.h, z20.h, z16.h\n"
-      "zip2 z20.h, z20.h, z16.h\n"
-      "addvl x21, x21, #2\n"
       "ld1h { z16.h }, p1/Z, [x20, #1, MUL VL]\n"
+      "st1h { z24.h }, p1, [x22]\n"
       "addvl x20, x20, #2\n"
-      "st1h { z19.h }, p1, [x22]\n"
-      "zip1 z19.h, z23.h, z18.h\n"
-      "zip2 z18.h, z23.h, z18.h\n"
-      "st1h { z22.h }, p1, [x22, #1, MUL VL]\n"
-      "st1h { z17.h }, p1, [x22, #2, MUL VL]\n"
-      "zip1 z17.h, z21.h, z16.h\n"
-      "zip2 z16.h, z21.h, z16.h\n"
-      "st1h { z20.h }, p1, [x22, #3, MUL VL]\n"
+      "zip1 z17.h, z18.h, z16.h\n"
+      "st1h { z23.h }, p1, [x22, #1, MUL VL]\n"
+      "zip2 z16.h, z18.h, z16.h\n"
+      "st1h { z22.h }, p1, [x22, #2, MUL VL]\n"
+      "st1h { z21.h }, p1, [x22, #3, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      "st1h { z19.h }, p1, [x22]\n"
-      "st1h { z18.h }, p1, [x22, #1, MUL VL]\n"
+      "st1h { z20.h }, p1, [x22]\n"
+      "st1h { z19.h }, p1, [x22, #1, MUL VL]\n"
       "st1h { z17.h }, p1, [x22, #2, MUL VL]\n"
       "st1h { z16.h }, p1, [x22, #3, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.h, XZR, x25\n"
-      "decw x25, ALL, MUL #2\n"
-      "ld1h { z18.h }, p0/Z, [x26]\n"
-      "cmp x25, #0x0\n"
+      "whilelt p0.h, XZR, x23\n"
+      "ld1h { z17.h }, p0/Z, [x26]\n"
+      "decw x23, ALL, MUL #2\n"
+      "ld1h { z16.h }, p0/Z, [x25]\n"
+      "cmp x23, #0x0\n"
       "addvl x26, x26, #1\n"
-      "ld1h { z17.h }, p0/Z, [x24]\n"
+      "zip1 z20.h, z17.h, z16.h\n"
+      "ld1h { z19.h }, p0/Z, [x24]\n"
+      "addvl x25, x25, #1\n"
       "addvl x24, x24, #1\n"
-      "ld1h { z20.h }, p0/Z, [x21]\n"
-      "addvl x21, x21, #1\n"
+      "zip2 z18.h, z17.h, z16.h\n"
       "ld1h { z16.h }, p0/Z, [x20]\n"
       "addvl x20, x20, #1\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "zip1 z17.h, z20.h, z16.h\n"
-      "zip2 z16.h, z20.h, z16.h\n"
-      "st1h { z19.h }, p1, [x22]\n"
+      "zip1 z17.h, z19.h, z16.h\n"
+      "zip2 z16.h, z19.h, z16.h\n"
+      "st1h { z20.h }, p1, [x22]\n"
       "st1h { z18.h }, p1, [x22, #1, MUL VL]\n"
       "st1h { z17.h }, p1, [x22, #2, MUL VL]\n"
       "st1h { z16.h }, p1, [x22, #3, MUL VL]\n"
@@ -121,12 +121,12 @@ void sme_transpose_interleave_2VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "add x24, x26, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #2\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
+      "add %x[in], x25, %x[in_stride]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "cmp x21, x20\n"
       "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x2\n"
@@ -134,19 +134,19 @@ void sme_transpose_interleave_2VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "8:"  // Tail row loop: Unroll column loop
       "ld1h { z18.h }, p1/Z, [x26]\n"
       "sub x21, x21, x20\n"
-      "ld1h { z17.h }, p1/Z, [x24]\n"
       "cmp x21, x20\n"
-      "ld1h { z20.h }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1h { z16.h }, p1/Z, [x25]\n"
+      "zip1 z17.h, z18.h, z16.h\n"
+      "zip2 z19.h, z18.h, z16.h\n"
+      "ld1h { z18.h }, p1/Z, [x26, #1, MUL VL]\n"
       "addvl x26, x26, #2\n"
-      "ld1h { z16.h }, p1/Z, [x24, #1, MUL VL]\n"
-      "addvl x24, x24, #2\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "zip1 z17.h, z20.h, z16.h\n"
-      "zip2 z16.h, z20.h, z16.h\n"
-      "st1h { z19.h }, p1, [x22]\n"
-      "st1h { z18.h }, p1, [x22, #1, MUL VL]\n"
+      "ld1h { z16.h }, p1/Z, [x25, #1, MUL VL]\n"
+      "st1h { z17.h }, p1, [x22]\n"
+      "addvl x25, x25, #2\n"
+      "zip1 z17.h, z18.h, z16.h\n"
+      "st1h { z19.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
+      "zip2 z16.h, z18.h, z16.h\n"
       "st1h { z17.h }, p1, [x22]\n"
       "st1h { z16.h }, p1, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
@@ -155,13 +155,13 @@ void sme_transpose_interleave_2VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "whilelt p0.h, XZR, x21\n"
-      "decw x21, ALL, MUL #2\n"
       "ld1h { z18.h }, p0/Z, [x26]\n"
+      "decw x21, ALL, MUL #2\n"
+      "ld1h { z16.h }, p0/Z, [x25]\n"
       "cmp x21, #0x0\n"
       "addvl x26, x26, #1\n"
-      "ld1h { z16.h }, p0/Z, [x24]\n"
-      "addvl x24, x24, #1\n"
       "zip1 z17.h, z18.h, z16.h\n"
+      "addvl x25, x25, #1\n"
       "zip2 z16.h, z18.h, z16.h\n"
       "st1h { z17.h }, p1, [x22]\n"
       "st1h { z16.h }, p1, [x22, #1, MUL VL]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp
index 382d4af314..284216a337 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp
@@ -45,99 +45,99 @@ void sme_transpose_interleave_2VL_2x2_fp32bf16(bfloat16 *out, const float *in, s
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
-      "add x24, x26, %x[in_stride]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
       "cnth x20, ALL, MUL #2\n"
-      "add x23, x24, %x[in_stride]\n"
-      "cmp x25, x20\n"
-      "add x21, x23, %x[in_stride]\n"
-      "mov x22, %x[out]\n"
+      "add x21, x24, %x[in_stride]\n"
+      "cmp x23, x20\n"
       "add %x[in], x21, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1w { z19.s }, p2/Z, [x26]\n"
-      "sub x25, x25, x20\n"
-      "ld1w { z18.s }, p2/Z, [x26, #1, MUL VL]\n"
-      "cmp x25, x20\n"
-      "ld1w { z17.s }, p2/Z, [x23]\n"
-      "ld1w { z16.s }, p2/Z, [x23, #1, MUL VL]\n"
-      ".inst 0x658aaa7b  // bfcvt z27.h, p2/M, z19.s\n"
-      "ld1w { z19.s }, p2/Z, [x26, #2, MUL VL]\n"
-      ".inst 0x658aaa5a  // bfcvt z26.h, p2/M, z18.s\n"
-      "ld1w { z18.s }, p2/Z, [x26, #3, MUL VL]\n"
-      ".inst 0x658aaa39  // bfcvt z25.h, p2/M, z17.s\n"
-      "addvl x26, x26, #4\n"
-      "ld1w { z17.s }, p2/Z, [x23, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x26]\n"
       ".inst 0x658aaa18  // bfcvt z24.h, p2/M, z16.s\n"
-      "ld1w { z16.s }, p2/Z, [x23, #3, MUL VL]\n"
-      ".inst 0x658aaa77  // bfcvt z23.h, p2/M, z19.s\n"
-      "addvl x23, x23, #4\n"
-      "ld1w { z19.s }, p2/Z, [x24]\n"
-      ".inst 0x658aaa56  // bfcvt z22.h, p2/M, z18.s\n"
-      "ld1w { z18.s }, p2/Z, [x24, #1, MUL VL]\n"
-      ".inst 0x658aaa35  // bfcvt z21.h, p2/M, z17.s\n"
-      "ld1w { z17.s }, p2/Z, [x21]\n"
+      "sub x23, x23, x20\n"
+      "cmp x23, x20\n"
+      "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x658aaa17  // bfcvt z23.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x24]\n"
+      ".inst 0x658aaa16  // bfcvt z22.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x658aaa15  // bfcvt z21.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x26, #2, MUL VL]\n"
       ".inst 0x658aaa14  // bfcvt z20.h, p2/M, z16.s\n"
-      "ld1w { z16.s }, p2/Z, [x21, #1, MUL VL]\n"
-      ".inst 0x648aaa7b  // bfcvtnt z27.h, p2/M, z19.s\n"
-      "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n"
-      ".inst 0x648aaa5a  // bfcvtnt z26.h, p2/M, z18.s\n"
-      "ld1w { z18.s }, p2/Z, [x24, #3, MUL VL]\n"
-      ".inst 0x648aaa39  // bfcvtnt z25.h, p2/M, z17.s\n"
+      "ld1w { z16.s }, p2/Z, [x26, #3, MUL VL]\n"
+      ".inst 0x658aaa13  // bfcvt z19.h, p2/M, z16.s\n"
+      "addvl x26, x26, #4\n"
+      "ld1w { z16.s }, p2/Z, [x24, #2, MUL VL]\n"
+      ".inst 0x658aaa12  // bfcvt z18.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n"
+      ".inst 0x658aaa11  // bfcvt z17.h, p2/M, z16.s\n"
       "addvl x24, x24, #4\n"
-      "ld1w { z17.s }, p2/Z, [x21, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x25]\n"
       ".inst 0x648aaa18  // bfcvtnt z24.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
+      ".inst 0x648aaa17  // bfcvtnt z23.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x21]\n"
+      ".inst 0x648aaa16  // bfcvtnt z22.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x21, #1, MUL VL]\n"
+      ".inst 0x648aaa15  // bfcvtnt z21.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
+      ".inst 0x648aaa14  // bfcvtnt z20.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0x648aaa13  // bfcvtnt z19.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x21, #2, MUL VL]\n"
+      ".inst 0x648aaa12  // bfcvtnt z18.h, p2/M, z16.s\n"
       "ld1w { z16.s }, p2/Z, [x21, #3, MUL VL]\n"
-      "st1h { z27.h }, p2, [x22]\n"
+      "st1h { z24.h }, p2, [x22]\n"
       "addvl x21, x21, #4\n"
-      ".inst 0x648aaa77  // bfcvtnt z23.h, p2/M, z19.s\n"
-      "st1h { z26.h }, p2, [x22, #1, MUL VL]\n"
-      ".inst 0x648aaa56  // bfcvtnt z22.h, p2/M, z18.s\n"
-      "st1h { z25.h }, p2, [x22, #2, MUL VL]\n"
-      ".inst 0x648aaa35  // bfcvtnt z21.h, p2/M, z17.s\n"
-      "st1h { z24.h }, p2, [x22, #3, MUL VL]\n"
+      ".inst 0x648aaa11  // bfcvtnt z17.h, p2/M, z16.s\n"
+      "st1h { z23.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z21.h }, p2, [x22, #3, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      ".inst 0x648aaa14  // bfcvtnt z20.h, p2/M, z16.s\n"
-      "st1h { z23.h }, p2, [x22]\n"
-      "st1h { z22.h }, p2, [x22, #1, MUL VL]\n"
-      "st1h { z21.h }, p2, [x22, #2, MUL VL]\n"
-      "st1h { z20.h }, p2, [x22, #3, MUL VL]\n"
+      "st1h { z20.h }, p2, [x22]\n"
+      "st1h { z19.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #3, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x20, x25\n"
-      "decw x25, ALL, MUL #2\n"
+      "mov x20, x23\n"
       "whilelt p1.s, XZR, x20\n"
+      "ld1w { z16.s }, p1/Z, [x26]\n"
+      ".inst 0x658aaa14  // bfcvt z20.h, p2/M, z16.s\n"
       "decw x20\n"
       "whilelt p0.s, XZR, x20\n"
-      "ld1w { z19.s }, p1/Z, [x26]\n"
-      "cmp x25, #0x0\n"
-      "ld1w { z18.s }, p0/Z, [x26, #1, MUL VL]\n"
+      "ld1w { z16.s }, p0/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x658aaa13  // bfcvt z19.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x24]\n"
+      ".inst 0x658aaa12  // bfcvt z18.h, p2/M, z16.s\n"
+      "decw x23, ALL, MUL #2\n"
+      "cmp x23, #0x0\n"
+      "ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x658aaa11  // bfcvt z17.h, p2/M, z16.s\n"
       "addvl x26, x26, #2\n"
-      "ld1w { z17.s }, p1/Z, [x23]\n"
-      "ld1w { z16.s }, p0/Z, [x23, #1, MUL VL]\n"
-      ".inst 0x658aaa77  // bfcvt z23.h, p2/M, z19.s\n"
-      "addvl x23, x23, #2\n"
-      ".inst 0x658aaa56  // bfcvt z22.h, p2/M, z18.s\n"
-      "ld1w { z21.s }, p1/Z, [x24]\n"
-      ".inst 0x658aaa34  // bfcvt z20.h, p2/M, z17.s\n"
-      "ld1w { z19.s }, p0/Z, [x24, #1, MUL VL]\n"
       "addvl x24, x24, #2\n"
-      ".inst 0x658aaa12  // bfcvt z18.h, p2/M, z16.s\n"
-      "ld1w { z17.s }, p1/Z, [x21]\n"
+      "ld1w { z16.s }, p1/Z, [x25]\n"
+      ".inst 0x648aaa14  // bfcvtnt z20.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
+      "addvl x25, x25, #2\n"
+      ".inst 0x648aaa13  // bfcvtnt z19.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x21]\n"
+      ".inst 0x648aaa12  // bfcvtnt z18.h, p2/M, z16.s\n"
       "ld1w { z16.s }, p0/Z, [x21, #1, MUL VL]\n"
       "addvl x21, x21, #2\n"
-      ".inst 0x648aaab7  // bfcvtnt z23.h, p2/M, z21.s\n"
-      ".inst 0x648aaa76  // bfcvtnt z22.h, p2/M, z19.s\n"
-      ".inst 0x648aaa34  // bfcvtnt z20.h, p2/M, z17.s\n"
-      ".inst 0x648aaa12  // bfcvtnt z18.h, p2/M, z16.s\n"
-      "st1h { z23.h }, p2, [x22]\n"
-      "st1h { z22.h }, p2, [x22, #1, MUL VL]\n"
-      "st1h { z20.h }, p2, [x22, #2, MUL VL]\n"
-      "st1h { z18.h }, p2, [x22, #3, MUL VL]\n"
+      ".inst 0x648aaa11  // bfcvtnt z17.h, p2/M, z16.s\n"
+      "st1h { z20.h }, p2, [x22]\n"
+      "st1h { z19.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #3, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
@@ -148,65 +148,65 @@ void sme_transpose_interleave_2VL_2x2_fp32bf16(bfloat16 *out, const float *in, s
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "add x24, x26, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #2\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
+      "add %x[in], x25, %x[in_stride]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "cmp x21, x20\n"
       "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x2\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1w { z19.s }, p2/Z, [x26]\n"
+      "ld1w { z16.s }, p2/Z, [x26]\n"
+      ".inst 0x658aaa14  // bfcvt z20.h, p2/M, z16.s\n"
       "sub x21, x21, x20\n"
-      "ld1w { z18.s }, p2/Z, [x26, #1, MUL VL]\n"
       "cmp x21, x20\n"
-      "ld1w { z17.s }, p2/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x658aaa13  // bfcvt z19.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x26, #2, MUL VL]\n"
+      ".inst 0x658aaa12  // bfcvt z18.h, p2/M, z16.s\n"
       "ld1w { z16.s }, p2/Z, [x26, #3, MUL VL]\n"
-      ".inst 0x658aaa77  // bfcvt z23.h, p2/M, z19.s\n"
+      ".inst 0x658aaa11  // bfcvt z17.h, p2/M, z16.s\n"
       "addvl x26, x26, #4\n"
-      ".inst 0x658aaa56  // bfcvt z22.h, p2/M, z18.s\n"
-      "ld1w { z21.s }, p2/Z, [x24]\n"
-      "ld1w { z20.s }, p2/Z, [x24, #1, MUL VL]\n"
-      ".inst 0x658aaa33  // bfcvt z19.h, p2/M, z17.s\n"
-      ".inst 0x658aaa12  // bfcvt z18.h, p2/M, z16.s\n"
-      "ld1w { z17.s }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n"
-      ".inst 0x648aaab7  // bfcvtnt z23.h, p2/M, z21.s\n"
-      "addvl x24, x24, #4\n"
-      ".inst 0x648aaa96  // bfcvtnt z22.h, p2/M, z20.s\n"
-      ".inst 0x648aaa33  // bfcvtnt z19.h, p2/M, z17.s\n"
+      "ld1w { z16.s }, p2/Z, [x25]\n"
+      ".inst 0x648aaa14  // bfcvtnt z20.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
+      ".inst 0x648aaa13  // bfcvtnt z19.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n"
       ".inst 0x648aaa12  // bfcvtnt z18.h, p2/M, z16.s\n"
-      "st1h { z23.h }, p2, [x22]\n"
-      "st1h { z22.h }, p2, [x22, #1, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n"
+      "st1h { z20.h }, p2, [x22]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0x648aaa11  // bfcvtnt z17.h, p2/M, z16.s\n"
+      "st1h { z19.h }, p2, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
-      "st1h { z19.h }, p2, [x22]\n"
-      "st1h { z18.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22]\n"
+      "st1h { z17.h }, p2, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decw x21, ALL, MUL #2\n"
       "whilelt p1.s, XZR, x20\n"
+      "ld1w { z16.s }, p1/Z, [x26]\n"
+      ".inst 0x658aaa12  // bfcvt z18.h, p2/M, z16.s\n"
       "decw x20\n"
       "whilelt p0.s, XZR, x20\n"
-      "ld1w { z17.s }, p1/Z, [x26]\n"
-      "cmp x21, #0x0\n"
       "ld1w { z16.s }, p0/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x658aaa11  // bfcvt z17.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25]\n"
+      "decw x21, ALL, MUL #2\n"
+      "cmp x21, #0x0\n"
+      ".inst 0x648aaa12  // bfcvtnt z18.h, p2/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n"
       "addvl x26, x26, #2\n"
-      "ld1w { z19.s }, p1/Z, [x24]\n"
-      ".inst 0x658aaa32  // bfcvt z18.h, p2/M, z17.s\n"
-      "ld1w { z17.s }, p0/Z, [x24, #1, MUL VL]\n"
-      "addvl x24, x24, #2\n"
-      ".inst 0x658aaa10  // bfcvt z16.h, p2/M, z16.s\n"
-      ".inst 0x648aaa72  // bfcvtnt z18.h, p2/M, z19.s\n"
-      ".inst 0x648aaa30  // bfcvtnt z16.h, p2/M, z17.s\n"
+      "addvl x25, x25, #2\n"
+      ".inst 0x648aaa11  // bfcvtnt z17.h, p2/M, z16.s\n"
       "st1h { z18.h }, p2, [x22]\n"
-      "st1h { z16.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #1, MUL VL]\n"
       "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL.hpp
index 8d3aa59d13..9677ea2016 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL.hpp
@@ -38,35 +38,35 @@ void sme_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt
       "ptrue p4.b\n"
       "blt 4f\n"
       "1:"  // Main row loop: Head
-      "mov x27, %x[in]\n"
-      "mov x26, %x[out]\n"
-      "add x25, x27, %x[in_stride]\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
-      "mov x23, %x[width]\n"
-      "add x22, x24, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "add x23, x24, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x21, x23\n"
-      "mov x20, x26\n"
-      "whilelt p3.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p2.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z31.h }, p3/Z, [x27]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z30.h }, p2/Z, [x27, #1, MUL VL]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x23, ALL, MUL #4\n"
-      "ld1h { z29.h }, p1/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z28.h }, p0/Z, [x27, #3, MUL VL]\n"
-      "cmp x23, #0x0\n"
-      "addvl x27, x27, #4\n"
+      "mov x20, x21\n"
+      "whilelt p3.h, XZR, x20\n"
+      "ld1h { z31.h }, p3/Z, [x26]\n"
+      "dech x20\n"
+      "whilelt p2.h, XZR, x20\n"
+      "ld1h { z30.h }, p2/Z, [x26, #1, MUL VL]\n"
+      "dech x20\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z29.h }, p1/Z, [x26, #2, MUL VL]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z28.h }, p0/Z, [x26, #3, MUL VL]\n"
+      "mov x20, x22\n"
+      "dech x21, ALL, MUL #4\n"
       "ld1h { z27.h }, p3/Z, [x25]\n"
-      "add x26, x26, %x[out_stride]\n"
       "ld1h { z26.h }, p2/Z, [x25, #1, MUL VL]\n"
+      "cmp x21, #0x0\n"
+      "addvl x26, x26, #4\n"
       "ld1h { z25.h }, p1/Z, [x25, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "ld1h { z24.h }, p0/Z, [x25, #3, MUL VL]\n"
       "addvl x25, x25, #4\n"
       "ld1h { z23.h }, p3/Z, [x24]\n"
@@ -74,12 +74,12 @@ void sme_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt
       "ld1h { z21.h }, p1/Z, [x24, #2, MUL VL]\n"
       "ld1h { z20.h }, p0/Z, [x24, #3, MUL VL]\n"
       "addvl x24, x24, #4\n"
-      "ld1h { z19.h }, p3/Z, [x22]\n"
-      "ld1h { z18.h }, p2/Z, [x22, #1, MUL VL]\n"
-      "ld1h { z17.h }, p1/Z, [x22, #2, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x22, #3, MUL VL]\n"
+      "ld1h { z19.h }, p3/Z, [x23]\n"
+      "ld1h { z18.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "ld1h { z17.h }, p1/Z, [x23, #2, MUL VL]\n"
+      "ld1h { z16.h }, p0/Z, [x23, #3, MUL VL]\n"
       "st1h { z31.h }, p4, [x20]\n"
-      "addvl x22, x22, #4\n"
+      "addvl x23, x23, #4\n"
       "st1h { z30.h }, p4, [x20, #1, MUL VL]\n"
       "st1h { z29.h }, p4, [x20, #2, MUL VL]\n"
       "st1h { z28.h }, p4, [x20, #3, MUL VL]\n"
@@ -104,32 +104,32 @@ void sme_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt
       "cbz %x[height], 8f\n"
       "4:"  // Main loop skip
       "5:"  // Tail row loop: Head
-      "mov x27, %x[in]\n"
-      "mov x26, %x[out]\n"
-      "add %x[in], x27, %x[in_stride]\n"
+      "mov x26, %x[in]\n"
+      "add %x[in], x26, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x1\n"
       "mov x21, %x[width]\n"
       "6:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "dech x21, ALL, MUL #4\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z19.h }, p0/Z, [x26]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z18.h }, p0/Z, [x26, #1, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z19.h }, p1/Z, [x27]\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z17.h }, p0/Z, [x26, #2, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z18.h }, p0/Z, [x27, #1, MUL VL]\n"
+      "dech x21, ALL, MUL #4\n"
       "whilelt p0.h, XZR, x20\n"
       "cmp x21, #0x0\n"
-      "ld1h { z17.h }, p1/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x27, #3, MUL VL]\n"
-      "addvl x27, x27, #4\n"
-      "st1h { z19.h }, p4, [x26]\n"
-      "st1h { z18.h }, p4, [x26, #1, MUL VL]\n"
-      "st1h { z17.h }, p4, [x26, #2, MUL VL]\n"
-      "st1h { z16.h }, p4, [x26, #3, MUL VL]\n"
-      "add x26, x26, %x[out_stride]\n"
+      "ld1h { z16.h }, p0/Z, [x26, #3, MUL VL]\n"
+      "st1h { z19.h }, p4, [x22]\n"
+      "addvl x26, x26, #4\n"
+      "st1h { z18.h }, p4, [x22, #1, MUL VL]\n"
+      "st1h { z17.h }, p4, [x22, #2, MUL VL]\n"
+      "st1h { z16.h }, p4, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 6b\n"
       "7:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -139,7 +139,7 @@ void sme_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt
       ".inst 0xd503467f  // SMSTOP\n"
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
@@ -147,19 +147,6 @@ void sme_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt
 
 template<>
 void Transform<4, 1, true, VLType::SME>(
-    double *out, const double *in, int stride, int x0, int xmax, int k0, int kmax)
-{
-    sme_transpose_interleave_4VL(
-        reinterpret_cast<uint16_t *>(out),
-        reinterpret_cast<const uint16_t *>(in + k0 * stride + x0),
-        (xmax-x0) * sizeof(double) / 2,
-        stride * sizeof(double),
-        (kmax-k0)
-    );
-}
-
-template<>
-void Transform<4, 1, true, VLType::SME>(
     float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax)
 {
     sme_transpose_interleave_4VL(
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_1x4.hpp
index c7d4882b42..94d1c0840a 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_1x4.hpp
@@ -43,43 +43,43 @@ void sme_transpose_interleave_4VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "ptrue p1.b\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
-      "cmp %x[height], #0x3\n"
       "add x24, x25, %x[in_stride]\n"
-      "mov x23, %x[out]\n"
-      "add x22, x24, %x[in_stride]\n"
-      "mov x21, %x[width]\n"
-      "add x20, x22, %x[in_stride]\n"
-      "csel x22, x22, %x[pad_row], GE\n"
-      "add %x[in], x20, %x[in_stride]\n"
-      "csel x20, x20, %x[pad_row], GT\n"
+      "add x23, x24, %x[in_stride]\n"
+      "add x22, x23, %x[in_stride]\n"
+      "cmp %x[height], #0x3\n"
+      "add %x[in], x22, %x[in_stride]\n"
+      "csel x22, x22, %x[pad_row], GT\n"
+      "csel x23, x23, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "mov x21, %x[out]\n"
       "csel x24, x24, %x[pad_row], GT\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "mov x20, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "whilelt p0.b, XZR, x21\n"
-      "decw x21, ALL, MUL #4\n"
-      "ld1b { z20.b }, p0/Z, [x25]\n"
-      "cmp x21, #0x0\n"
-      "addvl x25, x25, #1\n"
+      "whilelt p0.b, XZR, x20\n"
+      "ld1b { z17.b }, p0/Z, [x25]\n"
+      "decw x20, ALL, MUL #4\n"
       "ld1b { z19.b }, p0/Z, [x24]\n"
+      "cmp x20, #0x0\n"
+      "addvl x25, x25, #1\n"
+      "ld1b { z16.b }, p0/Z, [x23]\n"
+      "zip1 z18.b, z17.b, z16.b\n"
+      "zip2 z20.b, z17.b, z16.b\n"
       "addvl x24, x24, #1\n"
-      "ld1b { z17.b }, p0/Z, [x22]\n"
-      "addvl x22, x22, #1\n"
-      "ld1b { z16.b }, p0/Z, [x20]\n"
-      "addvl x20, x20, #1\n"
-      "zip1 z18.b, z20.b, z17.b\n"
-      "zip2 z20.b, z20.b, z17.b\n"
+      "ld1b { z16.b }, p0/Z, [x22]\n"
       "zip1 z17.b, z19.b, z16.b\n"
-      "zip2 z16.b, z19.b, z16.b\n"
-      "zip1 z19.b, z18.b, z17.b\n"
+      "zip2 z19.b, z19.b, z16.b\n"
+      "addvl x23, x23, #1\n"
+      "addvl x22, x22, #1\n"
+      "zip1 z16.b, z18.b, z17.b\n"
       "zip2 z18.b, z18.b, z17.b\n"
-      "zip1 z17.b, z20.b, z16.b\n"
-      "zip2 z16.b, z20.b, z16.b\n"
-      "st1b { z19.b }, p1, [x23]\n"
-      "st1b { z18.b }, p1, [x23, #1, MUL VL]\n"
-      "st1b { z17.b }, p1, [x23, #2, MUL VL]\n"
-      "st1b { z16.b }, p1, [x23, #3, MUL VL]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "st1b { z16.b }, p1, [x21]\n"
+      "zip1 z17.b, z20.b, z19.b\n"
+      "zip2 z16.b, z20.b, z19.b\n"
+      "st1b { z18.b }, p1, [x21, #1, MUL VL]\n"
+      "st1b { z17.b }, p1, [x21, #2, MUL VL]\n"
+      "st1b { z16.b }, p1, [x21, #3, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp
index f070d7d322..4327466387 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp
@@ -45,49 +45,49 @@ void sme_transpose_interleave_4VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "blt 4f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[out]\n"
-      "add x24, x26, %x[in_stride]\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "mov x22, %x[width]\n"
-      "add x21, x23, %x[in_stride]\n"
-      "add %x[in], x21, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x20, x22\n"
-      "decw x22, ALL, MUL #4\n"
+      "mov x20, x21\n"
       "whilelt p1.h, XZR, x20\n"
+      "ld1h { z19.h }, p1/Z, [x26]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
-      "ld1h { z20.h }, p1/Z, [x26]\n"
-      "cmp x22, #0x0\n"
-      "ld1h { z19.h }, p0/Z, [x26, #1, MUL VL]\n"
+      "ld1h { z18.h }, p0/Z, [x26, #1, MUL VL]\n"
+      "ld1h { z17.h }, p1/Z, [x25]\n"
+      "decw x21, ALL, MUL #4\n"
+      "cmp x21, #0x0\n"
+      "zip1 z24.h, z19.h, z17.h\n"
+      "ld1h { z16.h }, p0/Z, [x25, #1, MUL VL]\n"
       "addvl x26, x26, #2\n"
+      "addvl x25, x25, #2\n"
+      "zip2 z23.h, z19.h, z17.h\n"
       "ld1h { z17.h }, p1/Z, [x24]\n"
-      "ld1h { z16.h }, p0/Z, [x24, #1, MUL VL]\n"
+      "zip1 z22.h, z18.h, z16.h\n"
+      "zip2 z21.h, z18.h, z16.h\n"
+      "ld1h { z20.h }, p0/Z, [x24, #1, MUL VL]\n"
       "addvl x24, x24, #2\n"
-      "ld1h { z18.h }, p1/Z, [x23]\n"
-      "ld1h { z24.h }, p0/Z, [x23, #1, MUL VL]\n"
+      "ld1h { z16.h }, p1/Z, [x23]\n"
+      "zip1 z19.h, z17.h, z16.h\n"
+      "zip2 z18.h, z17.h, z16.h\n"
+      "ld1h { z16.h }, p0/Z, [x23, #1, MUL VL]\n"
       "addvl x23, x23, #2\n"
-      "zip1 z23.h, z20.h, z17.h\n"
-      "zip2 z22.h, z20.h, z17.h\n"
-      "ld1h { z17.h }, p1/Z, [x21]\n"
-      "zip1 z21.h, z19.h, z16.h\n"
-      "zip2 z20.h, z19.h, z16.h\n"
-      "ld1h { z16.h }, p0/Z, [x21, #1, MUL VL]\n"
-      "addvl x21, x21, #2\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "st1h { z23.h }, p2, [x25]\n"
-      "zip1 z17.h, z24.h, z16.h\n"
-      "zip2 z16.h, z24.h, z16.h\n"
-      "st1h { z22.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z21.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z20.h }, p2, [x25, #3, MUL VL]\n"
-      "st1h { z19.h }, p2, [x25, #4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #5, MUL VL]\n"
-      "st1h { z17.h }, p2, [x25, #6, MUL VL]\n"
-      "st1h { z16.h }, p2, [x25, #7, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
+      "zip1 z17.h, z20.h, z16.h\n"
+      "zip2 z16.h, z20.h, z16.h\n"
+      "st1h { z24.h }, p2, [x22]\n"
+      "st1h { z23.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z21.h }, p2, [x22, #3, MUL VL]\n"
+      "st1h { z19.h }, p2, [x22, #4, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22, #5, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #6, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x4\n"
@@ -97,35 +97,35 @@ void sme_transpose_interleave_4VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "4:"  // Main loop skip
       "5:"  // Tail row loop: Head
       "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "add x24, x26, %x[in_stride]\n"
-      "mov x25, %x[out]\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
+      "add %x[in], x25, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "sub %x[height], %x[height], #0x2\n"
       "mov x21, %x[width]\n"
       "6:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decw x21, ALL, MUL #4\n"
       "whilelt p1.h, XZR, x20\n"
+      "ld1h { z18.h }, p1/Z, [x26]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
-      "ld1h { z18.h }, p1/Z, [x26]\n"
-      "cmp x21, #0x0\n"
       "ld1h { z20.h }, p0/Z, [x26, #1, MUL VL]\n"
-      "addvl x26, x26, #2\n"
-      "ld1h { z17.h }, p1/Z, [x24]\n"
-      "ld1h { z16.h }, p0/Z, [x24, #1, MUL VL]\n"
-      "addvl x24, x24, #2\n"
+      "ld1h { z17.h }, p1/Z, [x25]\n"
+      "decw x21, ALL, MUL #4\n"
+      "cmp x21, #0x0\n"
       "zip1 z19.h, z18.h, z17.h\n"
+      "ld1h { z16.h }, p0/Z, [x25, #1, MUL VL]\n"
+      "addvl x26, x26, #2\n"
+      "addvl x25, x25, #2\n"
       "zip2 z18.h, z18.h, z17.h\n"
       "zip1 z17.h, z20.h, z16.h\n"
       "zip2 z16.h, z20.h, z16.h\n"
-      "st1h { z19.h }, p2, [x25]\n"
-      "st1h { z18.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z17.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x25, #3, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
+      "st1h { z19.h }, p2, [x22]\n"
+      "st1h { z18.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 6b\n"
       "7:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp
index 44305f0513..1c97bed317 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp
@@ -45,69 +45,69 @@ void sme_transpose_interleave_4VL_2x2_fp32bf16(bfloat16 *out, const float *in, s
       "blt 4f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[out]\n"
-      "add x24, x26, %x[in_stride]\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "mov x22, %x[width]\n"
-      "add x21, x23, %x[in_stride]\n"
-      "add %x[in], x21, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x20, x22\n"
-      "decw x22, ALL, MUL #4\n"
+      "mov x20, x21\n"
       "whilelt p3.s, XZR, x20\n"
+      "ld1w { z16.s }, p3/Z, [x26]\n"
+      ".inst 0x658ab218  // bfcvt z24.h, p4/M, z16.s\n"
       "decw x20\n"
       "whilelt p2.s, XZR, x20\n"
+      "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x658ab217  // bfcvt z23.h, p4/M, z16.s\n"
       "decw x20\n"
-      "ld1w { z19.s }, p3/Z, [x26]\n"
       "whilelt p1.s, XZR, x20\n"
+      "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n"
+      ".inst 0x658ab216  // bfcvt z22.h, p4/M, z16.s\n"
       "decw x20\n"
-      "ld1w { z18.s }, p2/Z, [x26, #1, MUL VL]\n"
       "whilelt p0.s, XZR, x20\n"
-      "ld1w { z17.s }, p1/Z, [x26, #2, MUL VL]\n"
-      "cmp x22, #0x0\n"
       "ld1w { z16.s }, p0/Z, [x26, #3, MUL VL]\n"
-      ".inst 0x658ab27b  // bfcvt z27.h, p4/M, z19.s\n"
-      "addvl x26, x26, #4\n"
-      "ld1w { z19.s }, p3/Z, [x23]\n"
-      ".inst 0x658ab25a  // bfcvt z26.h, p4/M, z18.s\n"
-      "ld1w { z18.s }, p2/Z, [x23, #1, MUL VL]\n"
-      ".inst 0x658ab239  // bfcvt z25.h, p4/M, z17.s\n"
-      "ld1w { z17.s }, p1/Z, [x23, #2, MUL VL]\n"
-      ".inst 0x658ab218  // bfcvt z24.h, p4/M, z16.s\n"
-      "ld1w { z16.s }, p0/Z, [x23, #3, MUL VL]\n"
-      ".inst 0x658ab277  // bfcvt z23.h, p4/M, z19.s\n"
-      "addvl x23, x23, #4\n"
-      ".inst 0x658ab256  // bfcvt z22.h, p4/M, z18.s\n"
-      "ld1w { z19.s }, p3/Z, [x24]\n"
-      ".inst 0x658ab235  // bfcvt z21.h, p4/M, z17.s\n"
-      "ld1w { z18.s }, p2/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x658ab215  // bfcvt z21.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p3/Z, [x24]\n"
       ".inst 0x658ab214  // bfcvt z20.h, p4/M, z16.s\n"
-      "ld1w { z17.s }, p1/Z, [x24, #2, MUL VL]\n"
+      "decw x21, ALL, MUL #4\n"
+      "cmp x21, #0x0\n"
+      "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
+      ".inst 0x658ab213  // bfcvt z19.h, p4/M, z16.s\n"
+      "addvl x26, x26, #4\n"
+      "ld1w { z16.s }, p1/Z, [x24, #2, MUL VL]\n"
+      ".inst 0x658ab212  // bfcvt z18.h, p4/M, z16.s\n"
       "ld1w { z16.s }, p0/Z, [x24, #3, MUL VL]\n"
+      ".inst 0x658ab211  // bfcvt z17.h, p4/M, z16.s\n"
       "addvl x24, x24, #4\n"
-      ".inst 0x648ab27b  // bfcvtnt z27.h, p4/M, z19.s\n"
-      "ld1w { z19.s }, p3/Z, [x21]\n"
-      ".inst 0x648ab25a  // bfcvtnt z26.h, p4/M, z18.s\n"
-      "ld1w { z18.s }, p2/Z, [x21, #1, MUL VL]\n"
-      ".inst 0x648ab239  // bfcvtnt z25.h, p4/M, z17.s\n"
-      "ld1w { z17.s }, p1/Z, [x21, #2, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x25]\n"
       ".inst 0x648ab218  // bfcvtnt z24.h, p4/M, z16.s\n"
-      "ld1w { z16.s }, p0/Z, [x21, #3, MUL VL]\n"
-      "addvl x21, x21, #4\n"
-      ".inst 0x648ab277  // bfcvtnt z23.h, p4/M, z19.s\n"
-      "st1h { z27.h }, p4, [x25]\n"
-      ".inst 0x648ab256  // bfcvtnt z22.h, p4/M, z18.s\n"
-      "st1h { z26.h }, p4, [x25, #1, MUL VL]\n"
-      ".inst 0x648ab235  // bfcvtnt z21.h, p4/M, z17.s\n"
-      "st1h { z25.h }, p4, [x25, #2, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
+      ".inst 0x648ab217  // bfcvtnt z23.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n"
+      ".inst 0x648ab216  // bfcvtnt z22.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0x648ab215  // bfcvtnt z21.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p3/Z, [x23]\n"
       ".inst 0x648ab214  // bfcvtnt z20.h, p4/M, z16.s\n"
-      "st1h { z24.h }, p4, [x25, #3, MUL VL]\n"
-      "st1h { z23.h }, p4, [x25, #4, MUL VL]\n"
-      "st1h { z22.h }, p4, [x25, #5, MUL VL]\n"
-      "st1h { z21.h }, p4, [x25, #6, MUL VL]\n"
-      "st1h { z20.h }, p4, [x25, #7, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
+      "ld1w { z16.s }, p2/Z, [x23, #1, MUL VL]\n"
+      ".inst 0x648ab213  // bfcvtnt z19.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x23, #2, MUL VL]\n"
+      ".inst 0x648ab212  // bfcvtnt z18.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p0/Z, [x23, #3, MUL VL]\n"
+      "addvl x23, x23, #4\n"
+      ".inst 0x648ab211  // bfcvtnt z17.h, p4/M, z16.s\n"
+      "st1h { z24.h }, p4, [x22]\n"
+      "st1h { z23.h }, p4, [x22, #1, MUL VL]\n"
+      "st1h { z22.h }, p4, [x22, #2, MUL VL]\n"
+      "st1h { z21.h }, p4, [x22, #3, MUL VL]\n"
+      "st1h { z20.h }, p4, [x22, #4, MUL VL]\n"
+      "st1h { z19.h }, p4, [x22, #5, MUL VL]\n"
+      "st1h { z18.h }, p4, [x22, #6, MUL VL]\n"
+      "st1h { z17.h }, p4, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x4\n"
@@ -117,47 +117,47 @@ void sme_transpose_interleave_4VL_2x2_fp32bf16(bfloat16 *out, const float *in, s
       "4:"  // Main loop skip
       "5:"  // Tail row loop: Head
       "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "add x24, x26, %x[in_stride]\n"
-      "mov x25, %x[out]\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
+      "add %x[in], x25, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "sub %x[height], %x[height], #0x2\n"
       "mov x21, %x[width]\n"
       "6:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decw x21, ALL, MUL #4\n"
       "whilelt p3.s, XZR, x20\n"
+      "ld1w { z16.s }, p3/Z, [x26]\n"
+      ".inst 0x658ab214  // bfcvt z20.h, p4/M, z16.s\n"
       "decw x20\n"
       "whilelt p2.s, XZR, x20\n"
+      "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n"
+      ".inst 0x658ab213  // bfcvt z19.h, p4/M, z16.s\n"
       "decw x20\n"
-      "ld1w { z19.s }, p3/Z, [x26]\n"
       "whilelt p1.s, XZR, x20\n"
+      "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n"
+      ".inst 0x658ab212  // bfcvt z18.h, p4/M, z16.s\n"
       "decw x20\n"
-      "ld1w { z18.s }, p2/Z, [x26, #1, MUL VL]\n"
       "whilelt p0.s, XZR, x20\n"
-      "ld1w { z17.s }, p1/Z, [x26, #2, MUL VL]\n"
-      "cmp x21, #0x0\n"
       "ld1w { z16.s }, p0/Z, [x26, #3, MUL VL]\n"
-      ".inst 0x658ab277  // bfcvt z23.h, p4/M, z19.s\n"
+      ".inst 0x658ab211  // bfcvt z17.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p3/Z, [x25]\n"
+      "decw x21, ALL, MUL #4\n"
+      "cmp x21, #0x0\n"
+      ".inst 0x648ab214  // bfcvtnt z20.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n"
       "addvl x26, x26, #4\n"
-      ".inst 0x658ab256  // bfcvt z22.h, p4/M, z18.s\n"
-      "ld1w { z21.s }, p3/Z, [x24]\n"
-      ".inst 0x658ab234  // bfcvt z20.h, p4/M, z17.s\n"
-      "ld1w { z19.s }, p2/Z, [x24, #1, MUL VL]\n"
-      ".inst 0x658ab212  // bfcvt z18.h, p4/M, z16.s\n"
-      "ld1w { z17.s }, p1/Z, [x24, #2, MUL VL]\n"
-      "ld1w { z16.s }, p0/Z, [x24, #3, MUL VL]\n"
-      "addvl x24, x24, #4\n"
-      ".inst 0x648ab2b7  // bfcvtnt z23.h, p4/M, z21.s\n"
-      ".inst 0x648ab276  // bfcvtnt z22.h, p4/M, z19.s\n"
-      ".inst 0x648ab234  // bfcvtnt z20.h, p4/M, z17.s\n"
+      ".inst 0x648ab213  // bfcvtnt z19.h, p4/M, z16.s\n"
+      "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n"
       ".inst 0x648ab212  // bfcvtnt z18.h, p4/M, z16.s\n"
-      "st1h { z23.h }, p4, [x25]\n"
-      "st1h { z22.h }, p4, [x25, #1, MUL VL]\n"
-      "st1h { z20.h }, p4, [x25, #2, MUL VL]\n"
-      "st1h { z18.h }, p4, [x25, #3, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
+      "ld1w { z16.s }, p0/Z, [x25, #3, MUL VL]\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0x648ab211  // bfcvtnt z17.h, p4/M, z16.s\n"
+      "st1h { z20.h }, p4, [x22]\n"
+      "st1h { z19.h }, p4, [x22, #1, MUL VL]\n"
+      "st1h { z18.h }, p4, [x22, #2, MUL VL]\n"
+      "st1h { z17.h }, p4, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 6b\n"
       "7:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL.hpp
index c49c5ba433..98e8bb20a5 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL.hpp
@@ -39,51 +39,51 @@ void sme_transpose_interleave_8VL(uint16_t *out, const uint16_t *in, size_t widt
       "blt 4f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
-      "mov x24, %x[out]\n"
-      "add x23, x25, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "add %x[in], x24, %x[in_stride]\n"
+      "mov x23, %x[out]\n"
       "sub %x[height], %x[height], #0x2\n"
-      "add %x[in], x23, %x[in_stride]\n"
       "mov x22, %x[width]\n"
       "2:"  // Main row loop: Column loop
       "mov x21, x22\n"
-      "mov x20, x24\n"
       "whilelt p0.h, XZR, x21\n"
+      "ld1h { z31.h }, p0/Z, [x25]\n"
       "dech x21\n"
       "whilelt p6.h, XZR, x21\n"
+      "ld1h { z30.h }, p6/Z, [x25, #1, MUL VL]\n"
       "dech x21\n"
-      "ld1h { z31.h }, p0/Z, [x25]\n"
       "whilelt p5.h, XZR, x21\n"
+      "ld1h { z29.h }, p5/Z, [x25, #2, MUL VL]\n"
       "dech x21\n"
-      "ld1h { z30.h }, p6/Z, [x25, #1, MUL VL]\n"
       "whilelt p4.h, XZR, x21\n"
+      "ld1h { z28.h }, p4/Z, [x25, #3, MUL VL]\n"
       "dech x21\n"
-      "ld1h { z29.h }, p5/Z, [x25, #2, MUL VL]\n"
       "whilelt p3.h, XZR, x21\n"
+      "ld1h { z27.h }, p3/Z, [x25, #4, MUL VL]\n"
       "dech x21\n"
-      "ld1h { z28.h }, p4/Z, [x25, #3, MUL VL]\n"
       "whilelt p2.h, XZR, x21\n"
+      "ld1h { z26.h }, p2/Z, [x25, #5, MUL VL]\n"
       "dech x21\n"
-      "ld1h { z27.h }, p3/Z, [x25, #4, MUL VL]\n"
       "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z26.h }, p2/Z, [x25, #5, MUL VL]\n"
-      "dech x22, ALL, MUL #8\n"
       "ld1h { z25.h }, p1/Z, [x25, #6, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
-      "ld1h { z24.h }, p0/Z, [x23]\n"
+      "dech x21\n"
+      "mov x20, x23\n"
+      "ld1h { z24.h }, p0/Z, [x24]\n"
       "whilelt p0.h, XZR, x21\n"
-      "cmp x22, #0x0\n"
+      "dech x22, ALL, MUL #8\n"
       "ld1h { z23.h }, p0/Z, [x25, #7, MUL VL]\n"
+      "ld1h { z22.h }, p6/Z, [x24, #1, MUL VL]\n"
+      "cmp x22, #0x0\n"
       "addvl x25, x25, #8\n"
-      "ld1h { z22.h }, p6/Z, [x23, #1, MUL VL]\n"
-      "ld1h { z21.h }, p5/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z20.h }, p4/Z, [x23, #3, MUL VL]\n"
-      "ld1h { z19.h }, p3/Z, [x23, #4, MUL VL]\n"
-      "ld1h { z18.h }, p2/Z, [x23, #5, MUL VL]\n"
-      "ld1h { z17.h }, p1/Z, [x23, #6, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x23, #7, MUL VL]\n"
+      "ld1h { z21.h }, p5/Z, [x24, #2, MUL VL]\n"
+      "add x23, x23, %x[out_stride]\n"
+      "ld1h { z20.h }, p4/Z, [x24, #3, MUL VL]\n"
+      "ld1h { z19.h }, p3/Z, [x24, #4, MUL VL]\n"
+      "ld1h { z18.h }, p2/Z, [x24, #5, MUL VL]\n"
+      "ld1h { z17.h }, p1/Z, [x24, #6, MUL VL]\n"
+      "ld1h { z16.h }, p0/Z, [x24, #7, MUL VL]\n"
       "st1h { z31.h }, p7, [x20]\n"
-      "addvl x23, x23, #8\n"
+      "addvl x24, x24, #8\n"
       "st1h { z30.h }, p7, [x20, #1, MUL VL]\n"
       "st1h { z29.h }, p7, [x20, #2, MUL VL]\n"
       "st1h { z28.h }, p7, [x20, #3, MUL VL]\n"
@@ -109,47 +109,47 @@ void sme_transpose_interleave_8VL(uint16_t *out, const uint16_t *in, size_t widt
       "4:"  // Main loop skip
       "5:"  // Tail row loop: Head
       "mov x25, %x[in]\n"
-      "mov x24, %x[out]\n"
       "add %x[in], x25, %x[in_stride]\n"
+      "mov x23, %x[out]\n"
       "sub %x[height], %x[height], #0x1\n"
       "mov x21, %x[width]\n"
       "6:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "dech x21, ALL, MUL #8\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z23.h }, p0/Z, [x25]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z22.h }, p0/Z, [x25, #1, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z23.h }, p1/Z, [x25]\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z21.h }, p0/Z, [x25, #2, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z22.h }, p0/Z, [x25, #1, MUL VL]\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z20.h }, p0/Z, [x25, #3, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z21.h }, p1/Z, [x25, #2, MUL VL]\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z19.h }, p0/Z, [x25, #4, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z20.h }, p0/Z, [x25, #3, MUL VL]\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z18.h }, p0/Z, [x25, #5, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z19.h }, p1/Z, [x25, #4, MUL VL]\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z17.h }, p0/Z, [x25, #6, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z18.h }, p0/Z, [x25, #5, MUL VL]\n"
+      "dech x21, ALL, MUL #8\n"
       "whilelt p0.h, XZR, x20\n"
       "cmp x21, #0x0\n"
-      "ld1h { z17.h }, p1/Z, [x25, #6, MUL VL]\n"
       "ld1h { z16.h }, p0/Z, [x25, #7, MUL VL]\n"
+      "st1h { z23.h }, p7, [x23]\n"
       "addvl x25, x25, #8\n"
-      "st1h { z23.h }, p7, [x24]\n"
-      "st1h { z22.h }, p7, [x24, #1, MUL VL]\n"
-      "st1h { z21.h }, p7, [x24, #2, MUL VL]\n"
-      "st1h { z20.h }, p7, [x24, #3, MUL VL]\n"
-      "st1h { z19.h }, p7, [x24, #4, MUL VL]\n"
-      "st1h { z18.h }, p7, [x24, #5, MUL VL]\n"
-      "st1h { z17.h }, p7, [x24, #6, MUL VL]\n"
-      "st1h { z16.h }, p7, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
+      "st1h { z22.h }, p7, [x23, #1, MUL VL]\n"
+      "st1h { z21.h }, p7, [x23, #2, MUL VL]\n"
+      "st1h { z20.h }, p7, [x23, #3, MUL VL]\n"
+      "st1h { z19.h }, p7, [x23, #4, MUL VL]\n"
+      "st1h { z18.h }, p7, [x23, #5, MUL VL]\n"
+      "st1h { z17.h }, p7, [x23, #6, MUL VL]\n"
+      "st1h { z16.h }, p7, [x23, #7, MUL VL]\n"
+      "add x23, x23, %x[out_stride]\n"
       "bgt 6b\n"
       "7:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL_1x4.hpp
index 30a9dc3e9c..5d9c123835 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL_1x4.hpp
@@ -43,62 +43,62 @@ void sme_transpose_interleave_8VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "ptrue p2.b\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "cmp %x[height], #0x3\n"
       "add x25, x26, %x[in_stride]\n"
-      "mov x24, %x[out]\n"
-      "add x23, x25, %x[in_stride]\n"
-      "mov x22, %x[width]\n"
-      "add x21, x23, %x[in_stride]\n"
-      "csel x23, x23, %x[pad_row], GE\n"
-      "add %x[in], x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GT\n"
+      "add x24, x25, %x[in_stride]\n"
+      "add x23, x24, %x[in_stride]\n"
+      "cmp %x[height], #0x3\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "csel x23, x23, %x[pad_row], GT\n"
+      "csel x24, x24, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "sub %x[height], %x[height], #0x4\n"
+      "mov x22, %x[out]\n"
       "csel x25, x25, %x[pad_row], GT\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x20, x22\n"
-      "decw x22, ALL, MUL #8\n"
+      "mov x20, x21\n"
       "whilelt p1.b, XZR, x20\n"
+      "ld1b { z19.b }, p1/Z, [x26]\n"
       "decb x20\n"
       "whilelt p0.b, XZR, x20\n"
-      "ld1b { z20.b }, p1/Z, [x26]\n"
-      "cmp x22, #0x0\n"
-      "ld1b { z24.b }, p0/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z17.b }, p0/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z18.b }, p1/Z, [x25]\n"
+      "decw x21, ALL, MUL #8\n"
+      "cmp x21, #0x0\n"
+      "ld1b { z21.b }, p0/Z, [x25, #1, MUL VL]\n"
       "addvl x26, x26, #2\n"
-      "ld1b { z23.b }, p1/Z, [x25]\n"
-      "ld1b { z22.b }, p0/Z, [x25, #1, MUL VL]\n"
       "addvl x25, x25, #2\n"
-      "ld1b { z19.b }, p1/Z, [x23]\n"
-      "ld1b { z18.b }, p0/Z, [x23, #1, MUL VL]\n"
+      "ld1b { z16.b }, p1/Z, [x24]\n"
+      "zip1 z24.b, z19.b, z16.b\n"
+      "zip2 z20.b, z19.b, z16.b\n"
+      "ld1b { z16.b }, p0/Z, [x24, #1, MUL VL]\n"
+      "zip1 z23.b, z17.b, z16.b\n"
+      "zip2 z22.b, z17.b, z16.b\n"
+      "addvl x24, x24, #2\n"
+      "ld1b { z16.b }, p1/Z, [x23]\n"
+      "zip1 z17.b, z18.b, z16.b\n"
+      "zip2 z19.b, z18.b, z16.b\n"
+      "ld1b { z16.b }, p0/Z, [x23, #1, MUL VL]\n"
+      "zip1 z18.b, z21.b, z16.b\n"
+      "zip2 z21.b, z21.b, z16.b\n"
       "addvl x23, x23, #2\n"
-      "ld1b { z17.b }, p1/Z, [x21]\n"
-      "ld1b { z16.b }, p0/Z, [x21, #1, MUL VL]\n"
-      "zip1 z21.b, z20.b, z19.b\n"
+      "zip1 z16.b, z24.b, z17.b\n"
+      "zip2 z17.b, z24.b, z17.b\n"
+      "st1b { z16.b }, p2, [x22]\n"
+      "zip1 z16.b, z20.b, z19.b\n"
       "zip2 z20.b, z20.b, z19.b\n"
-      "addvl x21, x21, #2\n"
-      "zip1 z25.b, z24.b, z18.b\n"
-      "zip2 z24.b, z24.b, z18.b\n"
-      "zip1 z19.b, z23.b, z17.b\n"
-      "zip2 z18.b, z23.b, z17.b\n"
-      "zip1 z17.b, z22.b, z16.b\n"
-      "zip2 z16.b, z22.b, z16.b\n"
-      "zip1 z23.b, z21.b, z19.b\n"
-      "zip2 z22.b, z21.b, z19.b\n"
-      "zip1 z21.b, z20.b, z18.b\n"
-      "zip2 z20.b, z20.b, z18.b\n"
-      "zip1 z19.b, z25.b, z17.b\n"
-      "zip2 z18.b, z25.b, z17.b\n"
-      "zip1 z17.b, z24.b, z16.b\n"
-      "zip2 z16.b, z24.b, z16.b\n"
-      "st1b { z23.b }, p2, [x24]\n"
-      "st1b { z22.b }, p2, [x24, #1, MUL VL]\n"
-      "st1b { z21.b }, p2, [x24, #2, MUL VL]\n"
-      "st1b { z20.b }, p2, [x24, #3, MUL VL]\n"
-      "st1b { z19.b }, p2, [x24, #4, MUL VL]\n"
-      "st1b { z18.b }, p2, [x24, #5, MUL VL]\n"
-      "st1b { z17.b }, p2, [x24, #6, MUL VL]\n"
-      "st1b { z16.b }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
+      "st1b { z17.b }, p2, [x22, #1, MUL VL]\n"
+      "zip1 z19.b, z23.b, z18.b\n"
+      "zip2 z18.b, z23.b, z18.b\n"
+      "st1b { z16.b }, p2, [x22, #2, MUL VL]\n"
+      "zip1 z17.b, z22.b, z21.b\n"
+      "zip2 z16.b, z22.b, z21.b\n"
+      "st1b { z20.b }, p2, [x22, #3, MUL VL]\n"
+      "st1b { z19.b }, p2, [x22, #4, MUL VL]\n"
+      "st1b { z18.b }, p2, [x22, #5, MUL VL]\n"
+      "st1b { z17.b }, p2, [x22, #6, MUL VL]\n"
+      "st1b { z16.b }, p2, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL_2x2.hpp
index 75bc57a649..9e1b2dca3e 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_8VL_2x2.hpp
@@ -43,40 +43,40 @@ void sme_transpose_interleave_8VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "ptrue p4.b\n"
       "1:"  // Main row loop: Head
       "mov x24, %x[in]\n"
-      "cmp %x[height], #0x1\n"
       "add x23, x24, %x[in_stride]\n"
-      "mov x22, %x[out]\n"
+      "cmp %x[height], #0x1\n"
       "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "csel x23, x23, %x[pad_row], GT\n"
       "sub %x[height], %x[height], #0x2\n"
       "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
       "mov x20, x21\n"
-      "decw x21, ALL, MUL #8\n"
       "whilelt p3.h, XZR, x20\n"
+      "ld1h { z20.h }, p3/Z, [x24]\n"
       "dech x20\n"
       "whilelt p2.h, XZR, x20\n"
+      "ld1h { z19.h }, p2/Z, [x24, #1, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z21.h }, p3/Z, [x24]\n"
       "whilelt p1.h, XZR, x20\n"
+      "ld1h { z18.h }, p1/Z, [x24, #2, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z20.h }, p2/Z, [x24, #1, MUL VL]\n"
       "whilelt p0.h, XZR, x20\n"
-      "ld1h { z25.h }, p1/Z, [x24, #2, MUL VL]\n"
-      "cmp x21, #0x0\n"
       "ld1h { z24.h }, p0/Z, [x24, #3, MUL VL]\n"
+      "ld1h { z17.h }, p3/Z, [x23]\n"
+      "decw x21, ALL, MUL #8\n"
+      "cmp x21, #0x0\n"
+      "zip1 z23.h, z20.h, z17.h\n"
+      "ld1h { z16.h }, p2/Z, [x23, #1, MUL VL]\n"
       "addvl x24, x24, #4\n"
-      "ld1h { z19.h }, p3/Z, [x23]\n"
-      "ld1h { z18.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "zip2 z22.h, z20.h, z17.h\n"
+      "zip1 z21.h, z19.h, z16.h\n"
       "ld1h { z17.h }, p1/Z, [x23, #2, MUL VL]\n"
+      "zip2 z20.h, z19.h, z16.h\n"
+      "zip1 z19.h, z18.h, z17.h\n"
       "ld1h { z16.h }, p0/Z, [x23, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
-      "zip1 z23.h, z21.h, z19.h\n"
-      "zip2 z22.h, z21.h, z19.h\n"
-      "zip1 z21.h, z20.h, z18.h\n"
-      "zip2 z20.h, z20.h, z18.h\n"
-      "zip1 z19.h, z25.h, z17.h\n"
-      "zip2 z18.h, z25.h, z17.h\n"
+      "zip2 z18.h, z18.h, z17.h\n"
       "zip1 z17.h, z24.h, z16.h\n"
       "zip2 z16.h, z24.h, z16.h\n"
       "st1h { z23.h }, p4, [x22]\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_12VL_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_12VL_2x4_fp32bf16.hpp
index 86dcddd07b..7e9b40b0d0 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_12VL_2x4_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_12VL_2x4_fp32bf16.hpp
@@ -39,314 +39,314 @@ void sve_transpose_interleave_12VL_2x4_fp32bf16(bfloat16 *out, const float *in,
     size_t out_stride = 12 * roundup<size_t>(height, 4) * get_vector_length<uint32_t>();
 
     __asm__ __volatile__(
-      "ptrue p2.b\n"
+      "ptrue p6.b\n"
       "1:"  // Main row loop: Head
       "mov x28, %x[in]\n"
-      "mov x27, %x[width]\n"
-      "cnth x26, ALL, MUL #6\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
+      "mov x25, %x[width]\n"
+      "cnth x24, ALL, MUL #6\n"
+      "add x23, x26, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x25, %x[out]\n"
-      "add x24, x28, %x[in_stride]\n"
-      "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
-      "csel x22, x22, %x[pad_row], GT\n"
-      "csel x23, x23, %x[pad_row], GE\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "csel x23, x23, %x[pad_row], GT\n"
+      "csel x26, x26, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "cmp x27, x26\n"
+      "csel x27, x27, %x[pad_row], GT\n"
+      "cmp x25, x24\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1w { z16.s }, p2/Z, [x28]\n"
-      "ld1w { z22.s }, p2/Z, [x28, #1, MUL VL]\n"
-      "mov x21, x25\n"
-      "add x25, x25, %x[out_stride]\n"
-      "ld1w { z30.s }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1w { z11.s }, p2/Z, [x28, #3, MUL VL]\n"
-      "mov x20, x25\n"
-      "sub x27, x27, x26\n"
-      "ld1w { z23.s }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x28, #5, MUL VL]\n"
-      "cmp x27, x26\n"
-      "add x25, x25, %x[out_stride]\n"
-      "ld1w { z17.s }, p2/Z, [x28, #6, MUL VL]\n"
-      "ld1w { z0.s }, p2/Z, [x28, #7, MUL VL]\n"
+      "ld1w { z22.s }, p6/Z, [x28]\n"
+      "ld1w { z7.s }, p6/Z, [x28, #1, MUL VL]\n"
+      "mov x21, x22\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1w { z19.s }, p6/Z, [x28, #2, MUL VL]\n"
+      "ld1w { z18.s }, p6/Z, [x28, #3, MUL VL]\n"
+      "mov x20, x22\n"
+      "sub x25, x25, x24\n"
+      "ld1w { z5.s }, p6/Z, [x28, #4, MUL VL]\n"
+      "ld1w { z25.s }, p6/Z, [x28, #5, MUL VL]\n"
+      "cmp x25, x24\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1w { z20.s }, p6/Z, [x28, #6, MUL VL]\n"
+      "ld1w { z23.s }, p6/Z, [x28, #7, MUL VL]\n"
       "addvl x28, x28, #12\n"
-      "ld1w { z10.s }, p2/Z, [x23]\n"
-      "ld1w { z14.s }, p2/Z, [x23, #1, MUL VL]\n"
-      "ld1w { z12.s }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z13.s }, p2/Z, [x23, #3, MUL VL]\n"
-      "ld1w { z29.s }, p2/Z, [x23, #4, MUL VL]\n"
-      "ld1w { z31.s }, p2/Z, [x23, #5, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x23, #6, MUL VL]\n"
-      "ld1w { z1.s }, p2/Z, [x23, #7, MUL VL]\n"
+      "ld1w { z4.s }, p6/Z, [x26]\n"
+      "ld1w { z10.s }, p6/Z, [x26, #1, MUL VL]\n"
+      "zip1 z14.s, z22.s, z4.s\n"
+      "zip2 z22.s, z22.s, z4.s\n"
+      "ld1w { z28.s }, p6/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z27.s }, p6/Z, [x26, #3, MUL VL]\n"
+      "zip1 z24.s, z7.s, z10.s\n"
+      "zip2 z15.s, z7.s, z10.s\n"
+      "ld1w { z7.s }, p6/Z, [x26, #4, MUL VL]\n"
+      "ld1w { z2.s }, p6/Z, [x26, #5, MUL VL]\n"
+      "zip1 z9.s, z19.s, z28.s\n"
+      "zip2 z0.s, z19.s, z28.s\n"
+      "ld1w { z19.s }, p6/Z, [x26, #6, MUL VL]\n"
+      "ld1w { z16.s }, p6/Z, [x26, #7, MUL VL]\n"
+      "addvl x26, x26, #12\n"
+      "zip1 z1.s, z18.s, z27.s\n"
+      "ld1w { z30.s }, p6/Z, [x28, #-4, MUL VL]\n"
+      "ld1w { z29.s }, p6/Z, [x28, #-3, MUL VL]\n"
+      "zip2 z17.s, z18.s, z27.s\n"
+      ".inst 0x658ab9d5  // bfcvt z21.h, p6/M, z14.s\n"
+      "ld1w { z31.s }, p6/Z, [x27]\n"
+      "ld1w { z8.s }, p6/Z, [x27, #1, MUL VL]\n"
+      ".inst 0x658abacc  // bfcvt z12.h, p6/M, z22.s\n"
+      ".inst 0x658abb0e  // bfcvt z14.h, p6/M, z24.s\n"
+      "ld1w { z22.s }, p6/Z, [x27, #2, MUL VL]\n"
+      "ld1w { z28.s }, p6/Z, [x27, #3, MUL VL]\n"
+      ".inst 0x658ab9ea  // bfcvt z10.h, p6/M, z15.s\n"
+      ".inst 0x658ab92f  // bfcvt z15.h, p6/M, z9.s\n"
+      "ld1w { z27.s }, p6/Z, [x27, #4, MUL VL]\n"
+      "ld1w { z13.s }, p6/Z, [x27, #5, MUL VL]\n"
+      ".inst 0x658ab803  // bfcvt z3.h, p6/M, z0.s\n"
+      ".inst 0x658ab832  // bfcvt z18.h, p6/M, z1.s\n"
+      "ld1w { z26.s }, p6/Z, [x27, #6, MUL VL]\n"
+      "ld1w { z9.s }, p6/Z, [x27, #7, MUL VL]\n"
+      "addvl x27, x27, #12\n"
+      ".inst 0x658aba26  // bfcvt z6.h, p6/M, z17.s\n"
+      "ld1w { z1.s }, p6/Z, [x26, #-4, MUL VL]\n"
+      "ld1w { z0.s }, p6/Z, [x26, #-3, MUL VL]\n"
+      "zip1 z17.s, z5.s, z7.s\n"
+      "zip2 z5.s, z5.s, z7.s\n"
+      "ld1w { z24.s }, p6/Z, [x23]\n"
+      "ld1w { z11.s }, p6/Z, [x23, #1, MUL VL]\n"
+      "zip1 z7.s, z31.s, z24.s\n"
+      "zip2 z31.s, z31.s, z24.s\n"
+      "ld1w { z4.s }, p6/Z, [x23, #2, MUL VL]\n"
+      "ld1w { z24.s }, p6/Z, [x23, #3, MUL VL]\n"
+      ".inst 0x648ab8f5  // bfcvtnt z21.h, p6/M, z7.s\n"
+      "zip1 z7.s, z8.s, z11.s\n"
+      "zip2 z11.s, z8.s, z11.s\n"
+      "ld1w { z8.s }, p6/Z, [x23, #4, MUL VL]\n"
+      ".inst 0x648abbec  // bfcvtnt z12.h, p6/M, z31.s\n"
+      "ld1w { z31.s }, p6/Z, [x23, #5, MUL VL]\n"
+      ".inst 0x648ab8ee  // bfcvtnt z14.h, p6/M, z7.s\n"
+      "ld1w { z7.s }, p6/Z, [x23, #6, MUL VL]\n"
+      ".inst 0x648ab96a  // bfcvtnt z10.h, p6/M, z11.s\n"
+      "zip1 z11.s, z22.s, z4.s\n"
+      "zip2 z4.s, z22.s, z4.s\n"
+      "ld1w { z22.s }, p6/Z, [x23, #7, MUL VL]\n"
       "addvl x23, x23, #12\n"
-      "zip1 z26.s, z16.s, z10.s\n"
-      "ld1w { z2.s }, p2/Z, [x28, #-4, MUL VL]\n"
-      "ld1w { z24.s }, p2/Z, [x28, #-3, MUL VL]\n"
-      "zip2 z15.s, z16.s, z10.s\n"
-      "zip1 z6.s, z22.s, z14.s\n"
-      "ld1w { z27.s }, p2/Z, [x24]\n"
-      "ld1w { z18.s }, p2/Z, [x24, #1, MUL VL]\n"
-      "zip2 z28.s, z22.s, z14.s\n"
-      "zip1 z25.s, z30.s, z12.s\n"
-      "ld1w { z21.s }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1w { z3.s }, p2/Z, [x24, #3, MUL VL]\n"
-      "zip2 z7.s, z30.s, z12.s\n"
-      "zip1 z9.s, z11.s, z13.s\n"
-      "ld1w { z4.s }, p2/Z, [x24, #4, MUL VL]\n"
-      "ld1w { z22.s }, p2/Z, [x24, #5, MUL VL]\n"
-      "zip2 z16.s, z11.s, z13.s\n"
-      ".inst 0x658aab4c  // bfcvt z12.h, p2/M, z26.s\n"
-      "ld1w { z14.s }, p2/Z, [x24, #6, MUL VL]\n"
-      "ld1w { z30.s }, p2/Z, [x24, #7, MUL VL]\n"
-      "addvl x24, x24, #12\n"
-      ".inst 0x658aa9ef  // bfcvt z15.h, p2/M, z15.s\n"
-      "ld1w { z26.s }, p2/Z, [x23, #-4, MUL VL]\n"
-      "ld1w { z5.s }, p2/Z, [x23, #-3, MUL VL]\n"
-      ".inst 0x658aa8cd  // bfcvt z13.h, p2/M, z6.s\n"
-      ".inst 0x658aab8a  // bfcvt z10.h, p2/M, z28.s\n"
-      "ld1w { z28.s }, p2/Z, [x22]\n"
-      "ld1w { z8.s }, p2/Z, [x22, #1, MUL VL]\n"
-      ".inst 0x658aab39  // bfcvt z25.h, p2/M, z25.s\n"
-      ".inst 0x658aa8e6  // bfcvt z6.h, p2/M, z7.s\n"
-      "ld1w { z11.s }, p2/Z, [x22, #2, MUL VL]\n"
-      ".inst 0x658aa927  // bfcvt z7.h, p2/M, z9.s\n"
-      ".inst 0x658aaa10  // bfcvt z16.h, p2/M, z16.s\n"
-      "zip1 z9.s, z23.s, z29.s\n"
-      "zip2 z23.s, z23.s, z29.s\n"
-      "zip1 z29.s, z27.s, z28.s\n"
-      "zip2 z27.s, z27.s, z28.s\n"
-      "ld1w { z28.s }, p2/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x658aa929  // bfcvt z9.h, p2/M, z9.s\n"
-      ".inst 0x658aaaf7  // bfcvt z23.h, p2/M, z23.s\n"
-      ".inst 0x648aabac  // bfcvtnt z12.h, p2/M, z29.s\n"
-      "ld1w { z29.s }, p2/Z, [x22, #4, MUL VL]\n"
-      ".inst 0x648aab6f  // bfcvtnt z15.h, p2/M, z27.s\n"
-      "zip1 z27.s, z18.s, z8.s\n"
-      "zip2 z8.s, z18.s, z8.s\n"
-      "ld1w { z18.s }, p2/Z, [x22, #5, MUL VL]\n"
-      ".inst 0x648aab6d  // bfcvtnt z13.h, p2/M, z27.s\n"
-      "ld1w { z27.s }, p2/Z, [x22, #6, MUL VL]\n"
-      ".inst 0x648aa90a  // bfcvtnt z10.h, p2/M, z8.s\n"
-      "zip1 z8.s, z21.s, z11.s\n"
-      "zip2 z21.s, z21.s, z11.s\n"
-      "ld1w { z11.s }, p2/Z, [x22, #7, MUL VL]\n"
-      "addvl x22, x22, #12\n"
-      ".inst 0x648aa919  // bfcvtnt z25.h, p2/M, z8.s\n"
-      "ld1w { z8.s }, p2/Z, [x28, #-2, MUL VL]\n"
-      ".inst 0x648aaaa6  // bfcvtnt z6.h, p2/M, z21.s\n"
-      "zip1 z21.s, z3.s, z28.s\n"
-      "zip2 z3.s, z3.s, z28.s\n"
-      "ld1w { z28.s }, p2/Z, [x28, #-1, MUL VL]\n"
-      ".inst 0x648aaaa7  // bfcvtnt z7.h, p2/M, z21.s\n"
-      "ld1w { z21.s }, p2/Z, [x24, #-4, MUL VL]\n"
-      ".inst 0x648aa870  // bfcvtnt z16.h, p2/M, z3.s\n"
-      "zip1 z3.s, z20.s, z31.s\n"
-      "zip2 z31.s, z20.s, z31.s\n"
-      "zip1 z20.s, z17.s, z19.s\n"
-      "zip2 z17.s, z17.s, z19.s\n"
-      "zip1 z19.s, z0.s, z1.s\n"
-      "zip2 z1.s, z0.s, z1.s\n"
-      "zip1 z0.s, z2.s, z26.s\n"
-      "zip2 z2.s, z2.s, z26.s\n"
-      "zip1 z26.s, z24.s, z5.s\n"
-      "zip2 z24.s, z24.s, z5.s\n"
-      "zip1 z5.s, z4.s, z29.s\n"
-      "zip2 z4.s, z4.s, z29.s\n"
-      "ld1w { z29.s }, p2/Z, [x24, #-3, MUL VL]\n"
-      ".inst 0x658aa863  // bfcvt z3.h, p2/M, z3.s\n"
-      ".inst 0x658aabff  // bfcvt z31.h, p2/M, z31.s\n"
-      ".inst 0x658aaa94  // bfcvt z20.h, p2/M, z20.s\n"
-      ".inst 0x658aaa31  // bfcvt z17.h, p2/M, z17.s\n"
-      ".inst 0x658aaa73  // bfcvt z19.h, p2/M, z19.s\n"
-      ".inst 0x658aa821  // bfcvt z1.h, p2/M, z1.s\n"
-      ".inst 0x658aa800  // bfcvt z0.h, p2/M, z0.s\n"
-      ".inst 0x658aa842  // bfcvt z2.h, p2/M, z2.s\n"
-      ".inst 0x658aab5a  // bfcvt z26.h, p2/M, z26.s\n"
-      ".inst 0x658aab18  // bfcvt z24.h, p2/M, z24.s\n"
-      ".inst 0x648aa8a9  // bfcvtnt z9.h, p2/M, z5.s\n"
-      "ld1w { z5.s }, p2/Z, [x23, #-2, MUL VL]\n"
-      ".inst 0x648aa897  // bfcvtnt z23.h, p2/M, z4.s\n"
-      "zip1 z4.s, z22.s, z18.s\n"
-      "zip2 z22.s, z22.s, z18.s\n"
-      "ld1w { z18.s }, p2/Z, [x23, #-1, MUL VL]\n"
-      ".inst 0x648aa883  // bfcvtnt z3.h, p2/M, z4.s\n"
-      "ld1w { z4.s }, p2/Z, [x22, #-4, MUL VL]\n"
-      ".inst 0x648aaadf  // bfcvtnt z31.h, p2/M, z22.s\n"
-      "zip1 z22.s, z14.s, z27.s\n"
-      "zip2 z14.s, z14.s, z27.s\n"
-      "ld1w { z27.s }, p2/Z, [x22, #-3, MUL VL]\n"
-      ".inst 0x648aaad4  // bfcvtnt z20.h, p2/M, z22.s\n"
-      "ld1w { z22.s }, p2/Z, [x24, #-2, MUL VL]\n"
-      ".inst 0x648aa9d1  // bfcvtnt z17.h, p2/M, z14.s\n"
-      "zip1 z14.s, z30.s, z11.s\n"
-      "zip2 z11.s, z30.s, z11.s\n"
-      "ld1w { z30.s }, p2/Z, [x24, #-1, MUL VL]\n"
-      ".inst 0x648aa9d3  // bfcvtnt z19.h, p2/M, z14.s\n"
-      "ld1w { z14.s }, p2/Z, [x22, #-2, MUL VL]\n"
-      ".inst 0x648aa961  // bfcvtnt z1.h, p2/M, z11.s\n"
-      "ld1w { z11.s }, p2/Z, [x22, #-1, MUL VL]\n"
-      "st1h { z12.h }, p2, [x21]\n"
-      "zip1 z12.s, z21.s, z4.s\n"
-      "zip2 z21.s, z21.s, z4.s\n"
-      "zip1 z4.s, z29.s, z27.s\n"
-      "zip2 z29.s, z29.s, z27.s\n"
-      "st1h { z15.h }, p2, [x21, #1, MUL VL]\n"
-      "zip1 z27.s, z8.s, z5.s\n"
-      "zip2 z8.s, z8.s, z5.s\n"
-      "st1h { z13.h }, p2, [x21, #2, MUL VL]\n"
-      "zip1 z5.s, z28.s, z18.s\n"
-      "zip2 z28.s, z28.s, z18.s\n"
-      "st1h { z10.h }, p2, [x21, #3, MUL VL]\n"
-      "st1h { z25.h }, p2, [x21, #4, MUL VL]\n"
-      ".inst 0x648aa980  // bfcvtnt z0.h, p2/M, z12.s\n"
-      ".inst 0x648aaaa2  // bfcvtnt z2.h, p2/M, z21.s\n"
-      "st1h { z6.h }, p2, [x21, #5, MUL VL]\n"
-      ".inst 0x648aa89a  // bfcvtnt z26.h, p2/M, z4.s\n"
-      ".inst 0x648aabb8  // bfcvtnt z24.h, p2/M, z29.s\n"
-      "st1h { z7.h }, p2, [x21, #6, MUL VL]\n"
-      ".inst 0x658aab7b  // bfcvt z27.h, p2/M, z27.s\n"
-      "zip1 z25.s, z22.s, z14.s\n"
-      "st1h { z16.h }, p2, [x21, #7, MUL VL]\n"
+      ".inst 0x648ab96f  // bfcvtnt z15.h, p6/M, z11.s\n"
+      "ld1w { z11.s }, p6/Z, [x28, #-2, MUL VL]\n"
+      ".inst 0x648ab883  // bfcvtnt z3.h, p6/M, z4.s\n"
+      "zip1 z4.s, z28.s, z24.s\n"
+      "zip2 z24.s, z28.s, z24.s\n"
+      "ld1w { z28.s }, p6/Z, [x28, #-1, MUL VL]\n"
+      ".inst 0x648ab892  // bfcvtnt z18.h, p6/M, z4.s\n"
+      "ld1w { z4.s }, p6/Z, [x27, #-4, MUL VL]\n"
+      ".inst 0x648abb06  // bfcvtnt z6.h, p6/M, z24.s\n"
+      "zip1 z24.s, z25.s, z2.s\n"
+      "zip2 z25.s, z25.s, z2.s\n"
+      "zip1 z2.s, z20.s, z19.s\n"
+      "zip2 z20.s, z20.s, z19.s\n"
+      "zip1 z19.s, z23.s, z16.s\n"
+      "zip2 z16.s, z23.s, z16.s\n"
+      "zip1 z23.s, z30.s, z1.s\n"
+      "zip2 z30.s, z30.s, z1.s\n"
+      "zip1 z1.s, z29.s, z0.s\n"
+      "zip2 z0.s, z29.s, z0.s\n"
+      ".inst 0x658aba31  // bfcvt z17.h, p6/M, z17.s\n"
+      "zip1 z29.s, z27.s, z8.s\n"
+      ".inst 0x658ab8a5  // bfcvt z5.h, p6/M, z5.s\n"
+      "zip2 z27.s, z27.s, z8.s\n"
+      "ld1w { z8.s }, p6/Z, [x27, #-3, MUL VL]\n"
+      ".inst 0x658abb18  // bfcvt z24.h, p6/M, z24.s\n"
+      ".inst 0x658abb39  // bfcvt z25.h, p6/M, z25.s\n"
+      ".inst 0x658ab842  // bfcvt z2.h, p6/M, z2.s\n"
+      ".inst 0x658aba94  // bfcvt z20.h, p6/M, z20.s\n"
+      ".inst 0x658aba73  // bfcvt z19.h, p6/M, z19.s\n"
+      ".inst 0x658aba10  // bfcvt z16.h, p6/M, z16.s\n"
+      ".inst 0x658abaf7  // bfcvt z23.h, p6/M, z23.s\n"
+      ".inst 0x658abbde  // bfcvt z30.h, p6/M, z30.s\n"
+      ".inst 0x658ab821  // bfcvt z1.h, p6/M, z1.s\n"
+      ".inst 0x658ab800  // bfcvt z0.h, p6/M, z0.s\n"
+      ".inst 0x648abbb1  // bfcvtnt z17.h, p6/M, z29.s\n"
+      "ld1w { z29.s }, p6/Z, [x26, #-2, MUL VL]\n"
+      ".inst 0x648abb65  // bfcvtnt z5.h, p6/M, z27.s\n"
+      "zip1 z27.s, z13.s, z31.s\n"
+      "zip2 z31.s, z13.s, z31.s\n"
+      "ld1w { z13.s }, p6/Z, [x26, #-1, MUL VL]\n"
+      ".inst 0x648abb78  // bfcvtnt z24.h, p6/M, z27.s\n"
+      "ld1w { z27.s }, p6/Z, [x23, #-4, MUL VL]\n"
+      ".inst 0x648abbf9  // bfcvtnt z25.h, p6/M, z31.s\n"
+      "zip1 z31.s, z26.s, z7.s\n"
+      "zip2 z26.s, z26.s, z7.s\n"
+      "ld1w { z7.s }, p6/Z, [x23, #-3, MUL VL]\n"
+      ".inst 0x648abbe2  // bfcvtnt z2.h, p6/M, z31.s\n"
+      "ld1w { z31.s }, p6/Z, [x27, #-2, MUL VL]\n"
+      ".inst 0x648abb54  // bfcvtnt z20.h, p6/M, z26.s\n"
+      "zip1 z26.s, z9.s, z22.s\n"
+      "zip2 z9.s, z9.s, z22.s\n"
+      "ld1w { z22.s }, p6/Z, [x27, #-1, MUL VL]\n"
+      ".inst 0x648abb53  // bfcvtnt z19.h, p6/M, z26.s\n"
+      "ld1w { z26.s }, p6/Z, [x23, #-2, MUL VL]\n"
+      ".inst 0x648ab930  // bfcvtnt z16.h, p6/M, z9.s\n"
+      "ld1w { z9.s }, p6/Z, [x23, #-1, MUL VL]\n"
+      "st1h { z21.h }, p6, [x21]\n"
+      "zip1 z21.s, z4.s, z27.s\n"
+      "zip2 z27.s, z4.s, z27.s\n"
+      "zip1 z4.s, z8.s, z7.s\n"
+      "zip2 z8.s, z8.s, z7.s\n"
+      "st1h { z12.h }, p6, [x21, #1, MUL VL]\n"
+      "zip1 z7.s, z11.s, z29.s\n"
+      "zip2 z11.s, z11.s, z29.s\n"
+      "st1h { z14.h }, p6, [x21, #2, MUL VL]\n"
+      "zip1 z29.s, z28.s, z13.s\n"
+      "zip2 z12.s, z28.s, z13.s\n"
+      "st1h { z10.h }, p6, [x21, #3, MUL VL]\n"
+      "st1h { z15.h }, p6, [x21, #4, MUL VL]\n"
+      ".inst 0x648abab7  // bfcvtnt z23.h, p6/M, z21.s\n"
+      ".inst 0x648abb7e  // bfcvtnt z30.h, p6/M, z27.s\n"
+      "st1h { z3.h }, p6, [x21, #5, MUL VL]\n"
+      ".inst 0x648ab881  // bfcvtnt z1.h, p6/M, z4.s\n"
+      ".inst 0x648ab900  // bfcvtnt z0.h, p6/M, z8.s\n"
+      "st1h { z18.h }, p6, [x21, #6, MUL VL]\n"
+      ".inst 0x658ab8e8  // bfcvt z8.h, p6/M, z7.s\n"
+      "zip1 z27.s, z31.s, z26.s\n"
+      "st1h { z6.h }, p6, [x21, #7, MUL VL]\n"
       "addvl x21, x21, #12\n"
-      ".inst 0x658aa906  // bfcvt z6.h, p2/M, z8.s\n"
-      "zip2 z4.s, z22.s, z14.s\n"
-      ".inst 0x658aa8b2  // bfcvt z18.h, p2/M, z5.s\n"
-      "zip1 z22.s, z30.s, z11.s\n"
-      ".inst 0x658aab95  // bfcvt z21.h, p2/M, z28.s\n"
-      "zip2 z16.s, z30.s, z11.s\n"
-      "st1h { z9.h }, p2, [x21, #-4, MUL VL]\n"
-      "st1h { z23.h }, p2, [x21, #-3, MUL VL]\n"
-      ".inst 0x648aab3b  // bfcvtnt z27.h, p2/M, z25.s\n"
-      ".inst 0x648aa886  // bfcvtnt z6.h, p2/M, z4.s\n"
-      "st1h { z3.h }, p2, [x21, #-2, MUL VL]\n"
-      ".inst 0x648aaad2  // bfcvtnt z18.h, p2/M, z22.s\n"
-      "st1h { z31.h }, p2, [x21, #-1, MUL VL]\n"
-      ".inst 0x648aaa15  // bfcvtnt z21.h, p2/M, z16.s\n"
-      "st1h { z20.h }, p2, [x20]\n"
-      "st1h { z17.h }, p2, [x20, #1, MUL VL]\n"
-      "st1h { z19.h }, p2, [x20, #2, MUL VL]\n"
-      "st1h { z1.h }, p2, [x20, #3, MUL VL]\n"
-      "st1h { z0.h }, p2, [x20, #4, MUL VL]\n"
-      "st1h { z2.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z26.h }, p2, [x20, #6, MUL VL]\n"
-      "st1h { z24.h }, p2, [x20, #7, MUL VL]\n"
+      ".inst 0x658ab96e  // bfcvt z14.h, p6/M, z11.s\n"
+      "zip2 z28.s, z31.s, z26.s\n"
+      ".inst 0x658abbbd  // bfcvt z29.h, p6/M, z29.s\n"
+      "zip1 z21.s, z22.s, z9.s\n"
+      "st1h { z17.h }, p6, [x21, #-4, MUL VL]\n"
+      ".inst 0x658ab992  // bfcvt z18.h, p6/M, z12.s\n"
+      "zip2 z17.s, z22.s, z9.s\n"
+      "st1h { z5.h }, p6, [x21, #-3, MUL VL]\n"
+      "st1h { z24.h }, p6, [x21, #-2, MUL VL]\n"
+      ".inst 0x648abb68  // bfcvtnt z8.h, p6/M, z27.s\n"
+      ".inst 0x648abb8e  // bfcvtnt z14.h, p6/M, z28.s\n"
+      "st1h { z25.h }, p6, [x21, #-1, MUL VL]\n"
+      ".inst 0x648ababd  // bfcvtnt z29.h, p6/M, z21.s\n"
+      ".inst 0x648aba32  // bfcvtnt z18.h, p6/M, z17.s\n"
+      "st1h { z2.h }, p6, [x20]\n"
+      "st1h { z20.h }, p6, [x20, #1, MUL VL]\n"
+      "st1h { z19.h }, p6, [x20, #2, MUL VL]\n"
+      "st1h { z16.h }, p6, [x20, #3, MUL VL]\n"
+      "st1h { z23.h }, p6, [x20, #4, MUL VL]\n"
+      "st1h { z30.h }, p6, [x20, #5, MUL VL]\n"
+      "st1h { z1.h }, p6, [x20, #6, MUL VL]\n"
+      "st1h { z0.h }, p6, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
-      "st1h { z27.h }, p2, [x20, #-4, MUL VL]\n"
-      "st1h { z6.h }, p2, [x20, #-3, MUL VL]\n"
-      "st1h { z18.h }, p2, [x20, #-2, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #-1, MUL VL]\n"
+      "st1h { z8.h }, p6, [x20, #-4, MUL VL]\n"
+      "st1h { z14.h }, p6, [x20, #-3, MUL VL]\n"
+      "st1h { z29.h }, p6, [x20, #-2, MUL VL]\n"
+      "st1h { z18.h }, p6, [x20, #-1, MUL VL]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x27, 5f\n"
+      "cbz x25, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x21, x27\n"
       "mov x20, x25\n"
-      "decd x27, ALL, MUL #12\n"
-      "add x25, x25, %x[out_stride]\n"
-      "whilelt p1.s, XZR, x21\n"
-      "decw x21\n"
-      "whilelt p0.s, XZR, x21\n"
-      "decw x21\n"
-      "ld1w { z19.s }, p1/Z, [x28]\n"
-      "ld1w { z18.s }, p1/Z, [x23]\n"
-      "ld1w { z30.s }, p1/Z, [x24]\n"
-      "ld1w { z29.s }, p1/Z, [x22]\n"
-      "whilelt p1.s, XZR, x21\n"
-      "decw x21\n"
-      "ld1w { z21.s }, p0/Z, [x28, #1, MUL VL]\n"
-      "ld1w { z17.s }, p0/Z, [x23, #1, MUL VL]\n"
-      "ld1w { z28.s }, p0/Z, [x24, #1, MUL VL]\n"
-      "ld1w { z27.s }, p0/Z, [x22, #1, MUL VL]\n"
-      "zip1 z16.s, z19.s, z18.s\n"
-      "zip2 z26.s, z19.s, z18.s\n"
-      "whilelt p0.s, XZR, x21\n"
-      "decw x21\n"
-      "ld1w { z20.s }, p1/Z, [x28, #2, MUL VL]\n"
-      "ld1w { z19.s }, p1/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z25.s }, p1/Z, [x24, #2, MUL VL]\n"
-      "ld1w { z24.s }, p1/Z, [x22, #2, MUL VL]\n"
-      "zip1 z18.s, z21.s, z17.s\n"
-      "zip2 z23.s, z21.s, z17.s\n"
-      ".inst 0x658aaa0a  // bfcvt z10.h, p2/M, z16.s\n"
-      "zip1 z9.s, z30.s, z29.s\n"
-      "whilelt p1.s, XZR, x21\n"
-      "decw x21\n"
-      "ld1w { z17.s }, p0/Z, [x28, #3, MUL VL]\n"
-      "ld1w { z16.s }, p0/Z, [x23, #3, MUL VL]\n"
-      "zip1 z22.s, z20.s, z19.s\n"
-      "zip2 z21.s, z20.s, z19.s\n"
-      "ld1w { z20.s }, p0/Z, [x24, #3, MUL VL]\n"
-      "ld1w { z19.s }, p0/Z, [x22, #3, MUL VL]\n"
-      ".inst 0x658aab48  // bfcvt z8.h, p2/M, z26.s\n"
-      "zip2 z7.s, z30.s, z29.s\n"
-      "whilelt p0.s, XZR, x21\n"
-      "ld1w { z6.s }, p1/Z, [x28, #4, MUL VL]\n"
-      "ld1w { z5.s }, p1/Z, [x23, #4, MUL VL]\n"
-      ".inst 0x658aaa44  // bfcvt z4.h, p2/M, z18.s\n"
-      "zip1 z18.s, z17.s, z16.s\n"
-      "zip2 z17.s, z17.s, z16.s\n"
-      "ld1w { z3.s }, p1/Z, [x24, #4, MUL VL]\n"
-      "ld1w { z2.s }, p1/Z, [x22, #4, MUL VL]\n"
-      "zip1 z1.s, z28.s, z27.s\n"
-      ".inst 0x658aaae0  // bfcvt z0.h, p2/M, z23.s\n"
-      "cmp x27, #0x0\n"
-      "ld1w { z31.s }, p0/Z, [x28, #5, MUL VL]\n"
-      "ld1w { z16.s }, p0/Z, [x23, #5, MUL VL]\n"
-      "ld1w { z30.s }, p0/Z, [x24, #5, MUL VL]\n"
-      "zip2 z29.s, z28.s, z27.s\n"
-      ".inst 0x658aaadc  // bfcvt z28.h, p2/M, z22.s\n"
-      "ld1w { z27.s }, p0/Z, [x22, #5, MUL VL]\n"
-      "zip1 z23.s, z25.s, z24.s\n"
-      ".inst 0x658aaaba  // bfcvt z26.h, p2/M, z21.s\n"
+      "whilelt p5.s, XZR, x20\n"
+      "ld1w { z22.s }, p5/Z, [x28]\n"
+      "ld1w { z21.s }, p5/Z, [x26]\n"
+      "decw x20\n"
+      "whilelt p4.s, XZR, x20\n"
+      "ld1w { z20.s }, p4/Z, [x28, #1, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x26, #1, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p3.s, XZR, x20\n"
+      "ld1w { z18.s }, p3/Z, [x28, #2, MUL VL]\n"
+      "ld1w { z17.s }, p3/Z, [x26, #2, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p2.s, XZR, x20\n"
+      "ld1w { z30.s }, p2/Z, [x28, #3, MUL VL]\n"
+      "ld1w { z16.s }, p2/Z, [x26, #3, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p1.s, XZR, x20\n"
+      "ld1w { z13.s }, p1/Z, [x28, #4, MUL VL]\n"
+      "ld1w { z29.s }, p5/Z, [x27]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z12.s }, p0/Z, [x28, #5, MUL VL]\n"
+      "ld1w { z28.s }, p4/Z, [x27, #1, MUL VL]\n"
+      "ld1w { z11.s }, p3/Z, [x27, #2, MUL VL]\n"
+      "ld1w { z10.s }, p2/Z, [x27, #3, MUL VL]\n"
+      "zip1 z27.s, z22.s, z21.s\n"
+      "zip2 z26.s, z22.s, z21.s\n"
+      "ld1w { z9.s }, p1/Z, [x26, #4, MUL VL]\n"
+      "ld1w { z8.s }, p0/Z, [x26, #5, MUL VL]\n"
+      "zip1 z25.s, z20.s, z19.s\n"
+      "zip2 z24.s, z20.s, z19.s\n"
+      "ld1w { z23.s }, p5/Z, [x23]\n"
+      "ld1w { z22.s }, p4/Z, [x23, #1, MUL VL]\n"
+      "zip1 z21.s, z18.s, z17.s\n"
+      "zip2 z20.s, z18.s, z17.s\n"
+      "ld1w { z19.s }, p3/Z, [x23, #2, MUL VL]\n"
+      "ld1w { z18.s }, p2/Z, [x23, #3, MUL VL]\n"
+      "zip1 z17.s, z30.s, z16.s\n"
+      "zip2 z16.s, z30.s, z16.s\n"
+      "ld1w { z7.s }, p1/Z, [x27, #4, MUL VL]\n"
+      "ld1w { z6.s }, p0/Z, [x27, #5, MUL VL]\n"
+      ".inst 0x658abb65  // bfcvt z5.h, p6/M, z27.s\n"
+      "zip1 z4.s, z29.s, z23.s\n"
+      "ld1w { z3.s }, p1/Z, [x23, #4, MUL VL]\n"
+      "ld1w { z2.s }, p0/Z, [x23, #5, MUL VL]\n"
+      ".inst 0x658abb41  // bfcvt z1.h, p6/M, z26.s\n"
+      "zip2 z0.s, z29.s, z23.s\n"
+      ".inst 0x658abb3f  // bfcvt z31.h, p6/M, z25.s\n"
+      "zip1 z30.s, z28.s, z22.s\n"
+      "mov x20, x22\n"
+      "decd x25, ALL, MUL #12\n"
+      ".inst 0x658abb1d  // bfcvt z29.h, p6/M, z24.s\n"
+      "zip2 z28.s, z28.s, z22.s\n"
+      "cmp x25, #0x0\n"
       "addvl x28, x28, #6\n"
-      "zip2 z22.s, z25.s, z24.s\n"
-      ".inst 0x658aaa59  // bfcvt z25.h, p2/M, z18.s\n"
-      "addvl x24, x24, #6\n"
+      ".inst 0x658ababb  // bfcvt z27.h, p6/M, z21.s\n"
+      "zip1 z23.s, z11.s, z19.s\n"
+      "addvl x27, x27, #6\n"
+      "addvl x26, x26, #6\n"
+      ".inst 0x658aba9a  // bfcvt z26.h, p6/M, z20.s\n"
+      "zip2 z22.s, z11.s, z19.s\n"
       "addvl x23, x23, #6\n"
-      "zip1 z21.s, z20.s, z19.s\n"
-      ".inst 0x658aaa38  // bfcvt z24.h, p2/M, z17.s\n"
-      "addvl x22, x22, #6\n"
-      "zip2 z20.s, z20.s, z19.s\n"
-      "zip1 z19.s, z6.s, z5.s\n"
-      "zip2 z18.s, z6.s, z5.s\n"
-      "zip1 z17.s, z31.s, z16.s\n"
-      "zip2 z16.s, z31.s, z16.s\n"
-      ".inst 0x648aa92a  // bfcvtnt z10.h, p2/M, z9.s\n"
-      ".inst 0x648aa8e8  // bfcvtnt z8.h, p2/M, z7.s\n"
-      ".inst 0x648aa824  // bfcvtnt z4.h, p2/M, z1.s\n"
-      ".inst 0x648aaba0  // bfcvtnt z0.h, p2/M, z29.s\n"
-      ".inst 0x648aaafc  // bfcvtnt z28.h, p2/M, z23.s\n"
-      ".inst 0x648aaada  // bfcvtnt z26.h, p2/M, z22.s\n"
-      ".inst 0x648aaab9  // bfcvtnt z25.h, p2/M, z21.s\n"
-      "st1h { z10.h }, p2, [x20]\n"
-      ".inst 0x648aaa98  // bfcvtnt z24.h, p2/M, z20.s\n"
-      ".inst 0x658aaa77  // bfcvt z23.h, p2/M, z19.s\n"
-      "st1h { z8.h }, p2, [x20, #1, MUL VL]\n"
-      "zip1 z22.s, z3.s, z2.s\n"
-      ".inst 0x658aaa55  // bfcvt z21.h, p2/M, z18.s\n"
-      "st1h { z4.h }, p2, [x20, #2, MUL VL]\n"
-      "zip2 z20.s, z3.s, z2.s\n"
-      ".inst 0x658aaa33  // bfcvt z19.h, p2/M, z17.s\n"
-      "st1h { z0.h }, p2, [x20, #3, MUL VL]\n"
-      "zip1 z18.s, z30.s, z27.s\n"
-      ".inst 0x658aaa11  // bfcvt z17.h, p2/M, z16.s\n"
-      "st1h { z28.h }, p2, [x20, #4, MUL VL]\n"
-      "zip2 z16.s, z30.s, z27.s\n"
-      "st1h { z26.h }, p2, [x20, #5, MUL VL]\n"
-      ".inst 0x648aaad7  // bfcvtnt z23.h, p2/M, z22.s\n"
-      "st1h { z25.h }, p2, [x20, #6, MUL VL]\n"
-      ".inst 0x648aaa95  // bfcvtnt z21.h, p2/M, z20.s\n"
-      "st1h { z24.h }, p2, [x20, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      ".inst 0x658aba39  // bfcvt z25.h, p6/M, z17.s\n"
+      "zip1 z21.s, z10.s, z18.s\n"
+      ".inst 0x658aba18  // bfcvt z24.h, p6/M, z16.s\n"
+      "zip2 z20.s, z10.s, z18.s\n"
+      "zip1 z19.s, z13.s, z9.s\n"
+      "zip2 z18.s, z13.s, z9.s\n"
+      "zip1 z17.s, z12.s, z8.s\n"
+      "zip2 z16.s, z12.s, z8.s\n"
+      ".inst 0x648ab885  // bfcvtnt z5.h, p6/M, z4.s\n"
+      ".inst 0x648ab801  // bfcvtnt z1.h, p6/M, z0.s\n"
+      "st1h { z5.h }, p6, [x20]\n"
+      ".inst 0x648abbdf  // bfcvtnt z31.h, p6/M, z30.s\n"
+      ".inst 0x648abb9d  // bfcvtnt z29.h, p6/M, z28.s\n"
+      "st1h { z1.h }, p6, [x20, #1, MUL VL]\n"
+      ".inst 0x648abafb  // bfcvtnt z27.h, p6/M, z23.s\n"
+      ".inst 0x648abada  // bfcvtnt z26.h, p6/M, z22.s\n"
+      "st1h { z31.h }, p6, [x20, #2, MUL VL]\n"
+      ".inst 0x648abab9  // bfcvtnt z25.h, p6/M, z21.s\n"
+      ".inst 0x648aba98  // bfcvtnt z24.h, p6/M, z20.s\n"
+      "st1h { z29.h }, p6, [x20, #3, MUL VL]\n"
+      ".inst 0x658aba77  // bfcvt z23.h, p6/M, z19.s\n"
+      "zip1 z22.s, z7.s, z3.s\n"
+      "st1h { z27.h }, p6, [x20, #4, MUL VL]\n"
+      ".inst 0x658aba55  // bfcvt z21.h, p6/M, z18.s\n"
+      "zip2 z20.s, z7.s, z3.s\n"
+      "st1h { z26.h }, p6, [x20, #5, MUL VL]\n"
+      ".inst 0x658aba33  // bfcvt z19.h, p6/M, z17.s\n"
+      "zip1 z18.s, z6.s, z2.s\n"
+      "st1h { z25.h }, p6, [x20, #6, MUL VL]\n"
+      ".inst 0x658aba11  // bfcvt z17.h, p6/M, z16.s\n"
+      "zip2 z16.s, z6.s, z2.s\n"
+      "st1h { z24.h }, p6, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
-      ".inst 0x648aaa53  // bfcvtnt z19.h, p2/M, z18.s\n"
-      ".inst 0x648aaa11  // bfcvtnt z17.h, p2/M, z16.s\n"
-      "st1h { z23.h }, p2, [x20, #-4, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #-3, MUL VL]\n"
-      "st1h { z19.h }, p2, [x20, #-2, MUL VL]\n"
-      "st1h { z17.h }, p2, [x20, #-1, MUL VL]\n"
+      ".inst 0x648abad7  // bfcvtnt z23.h, p6/M, z22.s\n"
+      ".inst 0x648aba95  // bfcvtnt z21.h, p6/M, z20.s\n"
+      "st1h { z23.h }, p6, [x20, #-4, MUL VL]\n"
+      ".inst 0x648aba53  // bfcvtnt z19.h, p6/M, z18.s\n"
+      ".inst 0x648aba11  // bfcvtnt z17.h, p6/M, z16.s\n"
+      "st1h { z21.h }, p6, [x20, #-3, MUL VL]\n"
+      "st1h { z19.h }, p6, [x20, #-2, MUL VL]\n"
+      "st1h { z17.h }, p6, [x20, #-1, MUL VL]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -354,7 +354,7 @@ void sve_transpose_interleave_12VL_2x4_fp32bf16(bfloat16 *out, const float *in,
       "bge 1b\n"
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp
index e37879e19b..beddf76c5b 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp
@@ -40,59 +40,59 @@ void sve_transpose_interleave_1VL(uint32_t *out, const uint32_t *in, size_t widt
       "mov x26, %x[in]\n"
       "mov x25, %x[width]\n"
       "cntw x24, ALL, MUL #2\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x22, x26, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
+      "add x23, x26, %x[in_stride]\n"
+      "add x21, x23, %x[in_stride]\n"
       "add x20, x21, %x[in_stride]\n"
       "cmp x25, x24\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
       "sub x25, x25, x24\n"
       "ld1w { z23.s }, p1/Z, [x26]\n"
       "ld1w { z22.s }, p1/Z, [x26, #1, MUL VL]\n"
-      "addvl x26, x26, #2\n"
-      "ld1w { z21.s }, p1/Z, [x22]\n"
-      "ld1w { z20.s }, p1/Z, [x22, #1, MUL VL]\n"
       "cmp x25, x24\n"
-      "addvl x22, x22, #2\n"
+      "ld1w { z21.s }, p1/Z, [x23]\n"
+      "ld1w { z20.s }, p1/Z, [x23, #1, MUL VL]\n"
+      "addvl x26, x26, #2\n"
+      "addvl x23, x23, #2\n"
       "ld1w { z19.s }, p1/Z, [x21]\n"
       "ld1w { z18.s }, p1/Z, [x21, #1, MUL VL]\n"
       "addvl x21, x21, #2\n"
       "ld1w { z17.s }, p1/Z, [x20]\n"
       "ld1w { z16.s }, p1/Z, [x20, #1, MUL VL]\n"
-      "st1w { z23.s }, p1, [x23]\n"
+      "st1w { z23.s }, p1, [x22]\n"
       "addvl x20, x20, #2\n"
-      "st1w { z21.s }, p1, [x23, #1, MUL VL]\n"
-      "st1w { z19.s }, p1, [x23, #2, MUL VL]\n"
-      "st1w { z17.s }, p1, [x23, #3, MUL VL]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "st1w { z22.s }, p1, [x23]\n"
-      "st1w { z20.s }, p1, [x23, #1, MUL VL]\n"
-      "st1w { z18.s }, p1, [x23, #2, MUL VL]\n"
-      "st1w { z16.s }, p1, [x23, #3, MUL VL]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "st1w { z21.s }, p1, [x22, #1, MUL VL]\n"
+      "st1w { z19.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z17.s }, p1, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1w { z22.s }, p1, [x22]\n"
+      "st1w { z20.s }, p1, [x22, #1, MUL VL]\n"
+      "st1w { z18.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z16.s }, p1, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cbz x25, 5f\n"
       "4:"  // Main row loop: Column loop
       "whilelt p0.s, XZR, x25\n"
       "decw x25\n"
-      "cmp x25, #0x0\n"
       "ld1w { z19.s }, p0/Z, [x26]\n"
+      "ld1w { z18.s }, p0/Z, [x23]\n"
+      "cmp x25, #0x0\n"
       "addvl x26, x26, #1\n"
-      "ld1w { z18.s }, p0/Z, [x22]\n"
-      "addvl x22, x22, #1\n"
       "ld1w { z17.s }, p0/Z, [x21]\n"
-      "addvl x21, x21, #1\n"
       "ld1w { z16.s }, p0/Z, [x20]\n"
+      "addvl x23, x23, #1\n"
+      "addvl x21, x21, #1\n"
+      "st1w { z19.s }, p1, [x22]\n"
       "addvl x20, x20, #1\n"
-      "st1w { z19.s }, p1, [x23]\n"
-      "st1w { z18.s }, p1, [x23, #1, MUL VL]\n"
-      "st1w { z17.s }, p1, [x23, #2, MUL VL]\n"
-      "st1w { z16.s }, p1, [x23, #3, MUL VL]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "st1w { z18.s }, p1, [x22, #1, MUL VL]\n"
+      "st1w { z17.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z16.s }, p1, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x4\n"
@@ -104,32 +104,32 @@ void sve_transpose_interleave_1VL(uint32_t *out, const uint32_t *in, size_t widt
       "mov x21, %x[width]\n"
       "cntw x20, ALL, MUL #2\n"
       "mov x26, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x21, x20\n"
       "add %x[in], x26, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
       "sub x21, x21, x20\n"
       "ld1w { z17.s }, p1/Z, [x26]\n"
       "ld1w { z16.s }, p1/Z, [x26, #1, MUL VL]\n"
-      "addvl x26, x26, #2\n"
+      "st1w { z17.s }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
       "cmp x21, x20\n"
-      "st1w { z17.s }, p1, [x23]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "st1w { z16.s }, p1, [x23]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "st1w { z16.s }, p1, [x22]\n"
+      "addvl x26, x26, #2\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "whilelt p0.s, XZR, x21\n"
       "decw x21\n"
-      "cmp x21, #0x0\n"
       "ld1w { z16.s }, p0/Z, [x26]\n"
+      "st1w { z16.s }, p1, [x22]\n"
+      "cmp x21, #0x0\n"
       "addvl x26, x26, #1\n"
-      "st1w { z16.s }, p1, [x23]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL_1x4.hpp
index 60ac125bff..1103008fe2 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL_1x4.hpp
@@ -44,135 +44,135 @@ void sve_transpose_interleave_1VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x10, %x[in]\n"
-      "mov x9, %x[width]\n"
-      "cntb x28, ALL, MUL #2\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x10, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add x23, x24, %x[in_stride]\n"
-      "cmp x9, x28\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
+      "mov x25, %x[width]\n"
+      "cntb x24, ALL, MUL #2\n"
+      "add x23, x26, %x[in_stride]\n"
+      "add x21, x23, %x[in_stride]\n"
       "add x20, x21, %x[in_stride]\n"
+      "cmp x25, x24\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1b { z25.b }, p1/Z, [x10]\n"
-      "ld1b { z24.b }, p1/Z, [x26]\n"
-      "sub x9, x9, x28\n"
-      "ld1b { z20.b }, p1/Z, [x25]\n"
-      "ld1b { z17.b }, p1/Z, [x24]\n"
-      "cmp x9, x28\n"
-      "ld1b { z23.b }, p1/Z, [x23]\n"
-      "ld1b { z22.b }, p1/Z, [x22]\n"
-      "ld1b { z21.b }, p1/Z, [x21]\n"
-      "ld1b { z19.b }, p1/Z, [x20]\n"
-      "zip1 z18.b, z25.b, z20.b\n"
-      "zip1 z16.b, z24.b, z17.b\n"
-      "ld1b { z4.b }, p1/Z, [x10, #1, MUL VL]\n"
-      "ld1b { z3.b }, p1/Z, [x26, #1, MUL VL]\n"
-      "zip2 z2.b, z25.b, z20.b\n"
-      "zip2 z1.b, z24.b, z17.b\n"
-      "ld1b { z25.b }, p1/Z, [x25, #1, MUL VL]\n"
-      "ld1b { z0.b }, p1/Z, [x24, #1, MUL VL]\n"
+      "ld1b { z20.b }, p1/Z, [x10]\n"
+      "ld1b { z18.b }, p1/Z, [x9]\n"
+      "sub x25, x25, x24\n"
+      "cmp x25, x24\n"
+      "ld1b { z17.b }, p1/Z, [x28]\n"
+      "ld1b { z16.b }, p1/Z, [x27]\n"
+      "zip1 z25.b, z20.b, z17.b\n"
+      "zip1 z24.b, z18.b, z16.b\n"
+      "ld1b { z21.b }, p1/Z, [x26]\n"
+      "ld1b { z19.b }, p1/Z, [x23]\n"
+      "zip2 z2.b, z20.b, z17.b\n"
+      "zip2 z1.b, z18.b, z16.b\n"
+      "ld1b { z18.b }, p1/Z, [x21]\n"
+      "ld1b { z17.b }, p1/Z, [x20]\n"
+      "zip1 z20.b, z21.b, z18.b\n"
+      "zip1 z16.b, z19.b, z17.b\n"
+      "ld1b { z0.b }, p1/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z31.b }, p1/Z, [x9, #1, MUL VL]\n"
+      "zip2 z30.b, z21.b, z18.b\n"
+      "zip2 z29.b, z19.b, z17.b\n"
+      "ld1b { z23.b }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z22.b }, p1/Z, [x27, #1, MUL VL]\n"
+      "zip1 z19.b, z25.b, z24.b\n"
+      "zip1 z18.b, z20.b, z16.b\n"
+      "ld1b { z28.b }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z27.b }, p1/Z, [x23, #1, MUL VL]\n"
+      "zip2 z17.b, z25.b, z24.b\n"
+      "zip2 z16.b, z20.b, z16.b\n"
+      "ld1b { z21.b }, p1/Z, [x21, #1, MUL VL]\n"
+      "ld1b { z20.b }, p1/Z, [x20, #1, MUL VL]\n"
+      "st1b { z19.b }, p1, [x22]\n"
+      "zip1 z26.b, z0.b, z23.b\n"
+      "st1b { z18.b }, p1, [x22, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z25.b, z31.b, z22.b\n"
+      "zip1 z24.b, z28.b, z21.b\n"
+      "st1b { z17.b }, p1, [x22]\n"
+      "zip1 z19.b, z27.b, z20.b\n"
+      "zip1 z17.b, z2.b, z1.b\n"
       "addvl x10, x10, #2\n"
+      "st1b { z16.b }, p1, [x22, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z18.b, z30.b, z29.b\n"
+      "zip2 z16.b, z2.b, z1.b\n"
+      "st1b { z17.b }, p1, [x22]\n"
+      "zip2 z17.b, z30.b, z29.b\n"
+      "zip2 z23.b, z0.b, z23.b\n"
+      "addvl x9, x9, #2\n"
+      "st1b { z18.b }, p1, [x22, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z22.b, z31.b, z22.b\n"
+      "zip2 z21.b, z28.b, z21.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "zip2 z20.b, z27.b, z20.b\n"
+      "zip1 z16.b, z26.b, z25.b\n"
+      "addvl x28, x28, #2\n"
+      "st1b { z17.b }, p1, [x22, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z18.b, z24.b, z19.b\n"
+      "zip2 z17.b, z26.b, z25.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "zip2 z16.b, z24.b, z19.b\n"
+      "zip1 z19.b, z23.b, z22.b\n"
+      "addvl x27, x27, #2\n"
+      "st1b { z18.b }, p1, [x22, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z18.b, z21.b, z20.b\n"
       "addvl x26, x26, #2\n"
-      "zip1 z20.b, z23.b, z21.b\n"
-      "zip1 z17.b, z22.b, z19.b\n"
-      "ld1b { z31.b }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1b { z30.b }, p1/Z, [x22, #1, MUL VL]\n"
-      "ld1b { z24.b }, p1/Z, [x21, #1, MUL VL]\n"
-      "ld1b { z29.b }, p1/Z, [x20, #1, MUL VL]\n"
-      "zip2 z23.b, z23.b, z21.b\n"
-      "zip2 z22.b, z22.b, z19.b\n"
-      "zip1 z19.b, z18.b, z16.b\n"
-      "zip2 z18.b, z18.b, z16.b\n"
-      "addvl x25, x25, #2\n"
-      "addvl x24, x24, #2\n"
-      "zip1 z16.b, z20.b, z17.b\n"
-      "zip2 z17.b, z20.b, z17.b\n"
+      "st1b { z17.b }, p1, [x22]\n"
       "addvl x23, x23, #2\n"
-      "addvl x22, x22, #2\n"
-      "zip1 z28.b, z4.b, z25.b\n"
-      "zip1 z21.b, z3.b, z0.b\n"
       "addvl x21, x21, #2\n"
+      "zip2 z17.b, z23.b, z22.b\n"
+      "st1b { z16.b }, p1, [x22, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "addvl x20, x20, #2\n"
-      "st1b { z19.b }, p1, [x27]\n"
-      "zip1 z27.b, z31.b, z24.b\n"
-      "zip1 z26.b, z30.b, z29.b\n"
-      "st1b { z16.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z20.b, z2.b, z1.b\n"
-      "zip1 z16.b, z23.b, z22.b\n"
-      "st1b { z18.b }, p1, [x27]\n"
-      "zip2 z19.b, z2.b, z1.b\n"
-      "zip2 z18.b, z23.b, z22.b\n"
-      "st1b { z17.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip2 z25.b, z4.b, z25.b\n"
-      "zip2 z17.b, z3.b, z0.b\n"
-      "st1b { z20.b }, p1, [x27]\n"
-      "zip2 z24.b, z31.b, z24.b\n"
-      "zip2 z23.b, z30.b, z29.b\n"
-      "st1b { z16.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z16.b, z28.b, z21.b\n"
-      "zip1 z22.b, z27.b, z26.b\n"
-      "st1b { z19.b }, p1, [x27]\n"
-      "zip2 z21.b, z28.b, z21.b\n"
-      "zip2 z20.b, z27.b, z26.b\n"
-      "st1b { z18.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z19.b, z25.b, z17.b\n"
-      "zip1 z18.b, z24.b, z23.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "zip2 z17.b, z25.b, z17.b\n"
-      "zip2 z16.b, z24.b, z23.b\n"
-      "st1b { z22.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z19.b }, p1, [x27]\n"
-      "st1b { z18.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z17.b }, p1, [x27]\n"
-      "st1b { z16.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip2 z16.b, z21.b, z20.b\n"
+      "st1b { z19.b }, p1, [x22]\n"
+      "st1b { z18.b }, p1, [x22, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1b { z17.b }, p1, [x22]\n"
+      "st1b { z16.b }, p1, [x22, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x9, 5f\n"
+      "cbz x25, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.b, XZR, x9\n"
-      "decw x9\n"
-      "ld1b { z23.b }, p0/Z, [x10]\n"
+      "whilelt p0.b, XZR, x25\n"
+      "ld1b { z19.b }, p0/Z, [x10]\n"
+      "ld1b { z18.b }, p0/Z, [x9]\n"
+      "decw x25\n"
+      "ld1b { z17.b }, p0/Z, [x28]\n"
+      "ld1b { z16.b }, p0/Z, [x27]\n"
+      "zip1 z21.b, z19.b, z17.b\n"
+      "zip1 z20.b, z18.b, z16.b\n"
+      "ld1b { z18.b }, p0/Z, [x26]\n"
+      "ld1b { z19.b }, p0/Z, [x23]\n"
+      "cmp x25, #0x0\n"
       "incd x10, ALL, MUL #2\n"
-      "ld1b { z22.b }, p0/Z, [x26]\n"
+      "ld1b { z17.b }, p0/Z, [x21]\n"
+      "ld1b { z16.b }, p0/Z, [x20]\n"
+      "zip1 z18.b, z18.b, z17.b\n"
+      "zip1 z16.b, z19.b, z16.b\n"
+      "incd x9, ALL, MUL #2\n"
+      "incd x28, ALL, MUL #2\n"
+      "zip1 z17.b, z21.b, z20.b\n"
+      "zip1 z16.b, z18.b, z16.b\n"
+      "incd x27, ALL, MUL #2\n"
       "incd x26, ALL, MUL #2\n"
-      "ld1b { z19.b }, p0/Z, [x25]\n"
-      "incd x25, ALL, MUL #2\n"
-      "ld1b { z17.b }, p0/Z, [x24]\n"
-      "incd x24, ALL, MUL #2\n"
-      "ld1b { z21.b }, p0/Z, [x23]\n"
-      "ld1b { z20.b }, p0/Z, [x22]\n"
-      "ld1b { z18.b }, p0/Z, [x21]\n"
-      "cmp x9, #0x0\n"
+      "st1b { z17.b }, p1, [x22]\n"
       "incd x23, ALL, MUL #2\n"
-      "ld1b { z16.b }, p0/Z, [x20]\n"
-      "zip1 z19.b, z23.b, z19.b\n"
-      "incd x22, ALL, MUL #2\n"
       "incd x21, ALL, MUL #2\n"
-      "zip1 z17.b, z22.b, z17.b\n"
+      "st1b { z16.b }, p1, [x22, #1, MUL VL]\n"
       "incd x20, ALL, MUL #2\n"
-      "zip1 z18.b, z21.b, z18.b\n"
-      "zip1 z16.b, z20.b, z16.b\n"
-      "zip1 z17.b, z19.b, z17.b\n"
-      "zip1 z16.b, z18.b, z16.b\n"
-      "st1b { z17.b }, p1, [x27]\n"
-      "st1b { z16.b }, p1, [x27, #1, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x8\n"
@@ -182,88 +182,88 @@ void sve_transpose_interleave_1VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x10, %x[in]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cntb x20, ALL, MUL #2\n"
+      "add x27, x28, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x10, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "csel x25, x25, %x[pad_row], GE\n"
+      "add %x[in], x27, %x[in_stride]\n"
+      "csel x27, x27, %x[pad_row], GT\n"
+      "csel x28, x28, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x26, x26, %x[pad_row], GT\n"
+      "csel x9, x9, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
       "ld1b { z21.b }, p1/Z, [x10]\n"
-      "ld1b { z19.b }, p1/Z, [x26]\n"
+      "ld1b { z18.b }, p1/Z, [x9]\n"
       "sub x21, x21, x20\n"
-      "ld1b { z17.b }, p1/Z, [x25]\n"
-      "ld1b { z16.b }, p1/Z, [x24]\n"
       "cmp x21, x20\n"
-      "ld1b { z27.b }, p1/Z, [x10, #1, MUL VL]\n"
-      "ld1b { z26.b }, p1/Z, [x26, #1, MUL VL]\n"
-      "addvl x10, x10, #2\n"
-      "addvl x26, x26, #2\n"
-      "ld1b { z25.b }, p1/Z, [x25, #1, MUL VL]\n"
-      "ld1b { z24.b }, p1/Z, [x24, #1, MUL VL]\n"
-      "addvl x25, x25, #2\n"
-      "addvl x24, x24, #2\n"
+      "ld1b { z17.b }, p1/Z, [x28]\n"
+      "ld1b { z16.b }, p1/Z, [x27]\n"
       "zip1 z20.b, z21.b, z17.b\n"
-      "zip1 z18.b, z19.b, z16.b\n"
-      "zip2 z17.b, z21.b, z17.b\n"
-      "zip2 z16.b, z19.b, z16.b\n"
-      "zip1 z23.b, z27.b, z25.b\n"
-      "zip1 z22.b, z26.b, z24.b\n"
-      "zip1 z19.b, z20.b, z18.b\n"
-      "zip2 z18.b, z20.b, z18.b\n"
-      "zip1 z21.b, z17.b, z16.b\n"
-      "zip2 z17.b, z17.b, z16.b\n"
-      "zip2 z20.b, z27.b, z25.b\n"
-      "zip2 z16.b, z26.b, z24.b\n"
-      "st1b { z19.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z18.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z19.b, z23.b, z22.b\n"
-      "zip2 z18.b, z23.b, z22.b\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z17.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z17.b, z20.b, z16.b\n"
-      "zip2 z16.b, z20.b, z16.b\n"
-      "st1b { z19.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z18.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z17.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 z19.b, z18.b, z16.b\n"
+      "ld1b { z24.b }, p1/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z23.b }, p1/Z, [x9, #1, MUL VL]\n"
+      "zip2 z22.b, z21.b, z17.b\n"
+      "zip2 z21.b, z18.b, z16.b\n"
+      "ld1b { z18.b }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z17.b }, p1/Z, [x27, #1, MUL VL]\n"
+      "zip1 z16.b, z20.b, z19.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z16.b, z20.b, z19.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z20.b, z24.b, z18.b\n"
+      "zip1 z19.b, z23.b, z17.b\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      "zip1 z16.b, z22.b, z21.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z16.b, z22.b, z21.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z18.b, z24.b, z18.b\n"
+      "zip2 z17.b, z23.b, z17.b\n"
+      "zip1 z16.b, z20.b, z19.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z16.b, z20.b, z19.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z16.b, z18.b, z17.b\n"
+      "addvl x28, x28, #2\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "addvl x27, x27, #2\n"
+      "zip2 z16.b, z18.b, z17.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "whilelt p0.b, XZR, x21\n"
-      "decw x21\n"
       "ld1b { z19.b }, p0/Z, [x10]\n"
-      "incd x10, ALL, MUL #2\n"
-      "ld1b { z18.b }, p0/Z, [x26]\n"
-      "incd x26, ALL, MUL #2\n"
-      "ld1b { z17.b }, p0/Z, [x25]\n"
-      "incd x25, ALL, MUL #2\n"
-      "ld1b { z16.b }, p0/Z, [x24]\n"
-      "incd x24, ALL, MUL #2\n"
-      "cmp x21, #0x0\n"
+      "ld1b { z18.b }, p0/Z, [x9]\n"
+      "decw x21\n"
+      "ld1b { z17.b }, p0/Z, [x28]\n"
+      "ld1b { z16.b }, p0/Z, [x27]\n"
       "zip1 z17.b, z19.b, z17.b\n"
       "zip1 z16.b, z18.b, z16.b\n"
+      "cmp x21, #0x0\n"
+      "incd x10, ALL, MUL #2\n"
       "zip1 z16.b, z17.b, z16.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "incd x9, ALL, MUL #2\n"
+      "incd x28, ALL, MUL #2\n"
+      "incd x27, ALL, MUL #2\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -272,7 +272,7 @@ void sve_transpose_interleave_1VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "12:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL.hpp
index e10f818ba2..0e138e4422 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL.hpp
@@ -34,57 +34,57 @@ void sve_transpose_interleave_3VL(uint16_t *out, const uint16_t *in, size_t widt
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "ptrue p2.b\n"
+      "ptrue p3.b\n"
       "blt 4f\n"
       "1:"  // Main row loop: Head
-      "mov x27, %x[in]\n"
-      "mov x26, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "mov x25, %x[width]\n"
-      "add x24, x27, %x[in_stride]\n"
+      "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x21, x25\n"
-      "mov x20, x26\n"
-      "dech x25, ALL, MUL #3\n"
-      "add x26, x26, %x[out_stride]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z27.h }, p0/Z, [x27]\n"
-      "ld1h { z26.h }, p0/Z, [x24]\n"
-      "ld1h { z25.h }, p0/Z, [x23]\n"
-      "ld1h { z24.h }, p0/Z, [x22]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "cmp x25, #0x0\n"
-      "ld1h { z23.h }, p1/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z22.h }, p1/Z, [x24, #1, MUL VL]\n"
-      "ld1h { z21.h }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1h { z20.h }, p1/Z, [x22, #1, MUL VL]\n"
-      "ld1h { z19.h }, p0/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z18.h }, p0/Z, [x24, #2, MUL VL]\n"
-      "addvl x27, x27, #3\n"
+      "mov x20, x21\n"
+      "whilelt p2.h, XZR, x20\n"
+      "ld1h { z27.h }, p2/Z, [x26]\n"
+      "ld1h { z26.h }, p2/Z, [x25]\n"
+      "dech x20\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z25.h }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1h { z24.h }, p1/Z, [x25, #1, MUL VL]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z23.h }, p0/Z, [x26, #2, MUL VL]\n"
+      "ld1h { z22.h }, p0/Z, [x25, #2, MUL VL]\n"
+      "mov x20, x22\n"
+      "dech x21, ALL, MUL #3\n"
+      "ld1h { z21.h }, p2/Z, [x24]\n"
+      "ld1h { z20.h }, p1/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z19.h }, p0/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z18.h }, p2/Z, [x23]\n"
+      "cmp x21, #0x0\n"
+      "addvl x26, x26, #3\n"
+      "ld1h { z17.h }, p1/Z, [x23, #1, MUL VL]\n"
+      "ld1h { z16.h }, p0/Z, [x23, #2, MUL VL]\n"
+      "st1h { z27.h }, p3, [x20]\n"
+      "addvl x25, x25, #3\n"
+      "st1h { z25.h }, p3, [x20, #1, MUL VL]\n"
       "addvl x24, x24, #3\n"
-      "ld1h { z17.h }, p0/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x22, #2, MUL VL]\n"
-      "st1h { z27.h }, p2, [x20]\n"
       "addvl x23, x23, #3\n"
-      "st1h { z23.h }, p2, [x20, #1, MUL VL]\n"
-      "addvl x22, x22, #3\n"
-      "st1h { z19.h }, p2, [x20, #2, MUL VL]\n"
-      "st1h { z26.h }, p2, [x20, #3, MUL VL]\n"
-      "st1h { z22.h }, p2, [x20, #4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z25.h }, p2, [x20, #6, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #7, MUL VL]\n"
+      "st1h { z23.h }, p3, [x20, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z26.h }, p3, [x20, #3, MUL VL]\n"
+      "st1h { z24.h }, p3, [x20, #4, MUL VL]\n"
+      "st1h { z22.h }, p3, [x20, #5, MUL VL]\n"
+      "st1h { z21.h }, p3, [x20, #6, MUL VL]\n"
+      "st1h { z20.h }, p3, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
-      "st1h { z17.h }, p2, [x20, #-4, MUL VL]\n"
-      "st1h { z24.h }, p2, [x20, #-3, MUL VL]\n"
-      "st1h { z20.h }, p2, [x20, #-2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
+      "st1h { z19.h }, p3, [x20, #-4, MUL VL]\n"
+      "st1h { z18.h }, p3, [x20, #-3, MUL VL]\n"
+      "st1h { z17.h }, p3, [x20, #-2, MUL VL]\n"
+      "st1h { z16.h }, p3, [x20, #-1, MUL VL]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x4\n"
@@ -93,28 +93,28 @@ void sve_transpose_interleave_3VL(uint16_t *out, const uint16_t *in, size_t widt
       "cbz %x[height], 8f\n"
       "4:"  // Main loop skip
       "5:"  // Tail row loop: Head
-      "mov x27, %x[in]\n"
-      "mov x26, %x[out]\n"
+      "mov x26, %x[in]\n"
+      "add %x[in], x26, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x1\n"
       "mov x21, %x[width]\n"
-      "add %x[in], x27, %x[in_stride]\n"
       "6:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "dech x21, ALL, MUL #3\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z18.h }, p0/Z, [x26]\n"
       "dech x20\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z17.h }, p0/Z, [x26, #1, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z18.h }, p0/Z, [x27]\n"
+      "dech x21, ALL, MUL #3\n"
       "whilelt p0.h, XZR, x20\n"
       "cmp x21, #0x0\n"
-      "ld1h { z17.h }, p1/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x27, #2, MUL VL]\n"
-      "addvl x27, x27, #3\n"
-      "st1h { z18.h }, p2, [x26]\n"
-      "st1h { z17.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z16.h }, p2, [x26, #2, MUL VL]\n"
-      "add x26, x26, %x[out_stride]\n"
+      "ld1h { z16.h }, p0/Z, [x26, #2, MUL VL]\n"
+      "st1h { z18.h }, p3, [x22]\n"
+      "addvl x26, x26, #3\n"
+      "st1h { z17.h }, p3, [x22, #1, MUL VL]\n"
+      "st1h { z16.h }, p3, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 6b\n"
       "7:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -123,7 +123,7 @@ void sve_transpose_interleave_3VL(uint16_t *out, const uint16_t *in, size_t widt
       "8:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL_1x4.hpp
index e5bfb7bb7c..bc462414be 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL_1x4.hpp
@@ -44,175 +44,175 @@ void sve_transpose_interleave_3VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x10, %x[in]\n"
-      "mov x9, %x[width]\n"
-      "cntb x28, ALL, MUL #3\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x10, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add x23, x24, %x[in_stride]\n"
-      "cmp x9, x28\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
+      "mov x25, %x[width]\n"
+      "cntb x24, ALL, MUL #3\n"
+      "add x23, x26, %x[in_stride]\n"
+      "add x21, x23, %x[in_stride]\n"
       "add x20, x21, %x[in_stride]\n"
+      "cmp x25, x24\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1b { z19.b }, p1/Z, [x10]\n"
-      "ld1b { z18.b }, p1/Z, [x26]\n"
-      "sub x9, x9, x28\n"
-      "ld1b { z17.b }, p1/Z, [x25]\n"
-      "ld1b { z16.b }, p1/Z, [x24]\n"
-      "cmp x9, x28\n"
-      "ld1b { z27.b }, p1/Z, [x23]\n"
-      "ld1b { z26.b }, p1/Z, [x22]\n"
-      "ld1b { z25.b }, p1/Z, [x21]\n"
-      "ld1b { z24.b }, p1/Z, [x20]\n"
-      "ld1b { z23.b }, p1/Z, [x10, #1, MUL VL]\n"
-      "ld1b { z22.b }, p1/Z, [x26, #1, MUL VL]\n"
-      "zip1 z1.b, z19.b, z17.b\n"
-      "zip1 z0.b, z18.b, z16.b\n"
-      "ld1b { z21.b }, p1/Z, [x25, #1, MUL VL]\n"
-      "ld1b { z20.b }, p1/Z, [x24, #1, MUL VL]\n"
-      "zip2 z15.b, z19.b, z17.b\n"
-      "zip2 z14.b, z18.b, z16.b\n"
-      "ld1b { z19.b }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1b { z18.b }, p1/Z, [x22, #1, MUL VL]\n"
-      "zip1 z13.b, z27.b, z25.b\n"
-      "zip1 z12.b, z26.b, z24.b\n"
+      "ld1b { z21.b }, p1/Z, [x10]\n"
+      "ld1b { z20.b }, p1/Z, [x9]\n"
+      "sub x25, x25, x24\n"
+      "cmp x25, x24\n"
+      "ld1b { z17.b }, p1/Z, [x28]\n"
+      "ld1b { z16.b }, p1/Z, [x27]\n"
+      "zip1 z31.b, z21.b, z17.b\n"
+      "zip1 z22.b, z20.b, z16.b\n"
+      "ld1b { z19.b }, p1/Z, [x26]\n"
+      "ld1b { z18.b }, p1/Z, [x23]\n"
+      "zip2 z14.b, z21.b, z17.b\n"
+      "zip2 z13.b, z20.b, z16.b\n"
+      "ld1b { z17.b }, p1/Z, [x21]\n"
+      "ld1b { z16.b }, p1/Z, [x20]\n"
+      "zip1 z30.b, z19.b, z17.b\n"
+      "zip1 z29.b, z18.b, z16.b\n"
+      "ld1b { z21.b }, p1/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z20.b }, p1/Z, [x9, #1, MUL VL]\n"
+      "zip2 z12.b, z19.b, z17.b\n"
+      "zip2 z11.b, z18.b, z16.b\n"
+      "ld1b { z17.b }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z16.b }, p1/Z, [x27, #1, MUL VL]\n"
+      "zip1 z10.b, z21.b, z17.b\n"
+      "zip1 z9.b, z20.b, z16.b\n"
+      "ld1b { z19.b }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z18.b }, p1/Z, [x23, #1, MUL VL]\n"
+      "zip2 z8.b, z21.b, z17.b\n"
+      "zip2 z7.b, z20.b, z16.b\n"
       "ld1b { z17.b }, p1/Z, [x21, #1, MUL VL]\n"
       "ld1b { z16.b }, p1/Z, [x20, #1, MUL VL]\n"
-      "zip2 z11.b, z27.b, z25.b\n"
-      "zip2 z10.b, z26.b, z24.b\n"
-      "ld1b { z9.b }, p1/Z, [x10, #2, MUL VL]\n"
-      "ld1b { z8.b }, p1/Z, [x26, #2, MUL VL]\n"
-      "zip1 z7.b, z23.b, z21.b\n"
-      "zip1 z6.b, z22.b, z20.b\n"
-      "ld1b { z31.b }, p1/Z, [x25, #2, MUL VL]\n"
-      "ld1b { z30.b }, p1/Z, [x24, #2, MUL VL]\n"
-      "zip2 z5.b, z23.b, z21.b\n"
-      "zip2 z4.b, z22.b, z20.b\n"
-      "ld1b { z29.b }, p1/Z, [x23, #2, MUL VL]\n"
-      "ld1b { z28.b }, p1/Z, [x22, #2, MUL VL]\n"
-      "zip1 z27.b, z19.b, z17.b\n"
-      "zip1 z26.b, z18.b, z16.b\n"
-      "ld1b { z25.b }, p1/Z, [x21, #2, MUL VL]\n"
-      "ld1b { z24.b }, p1/Z, [x20, #2, MUL VL]\n"
-      "zip2 z23.b, z19.b, z17.b\n"
-      "zip2 z22.b, z18.b, z16.b\n"
-      "zip1 z3.b, z9.b, z31.b\n"
-      "zip1 z2.b, z8.b, z30.b\n"
+      "zip1 z6.b, z19.b, z17.b\n"
+      "zip1 z5.b, z18.b, z16.b\n"
+      "ld1b { z28.b }, p1/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z27.b }, p1/Z, [x9, #2, MUL VL]\n"
+      "zip2 z4.b, z19.b, z17.b\n"
+      "zip2 z3.b, z18.b, z16.b\n"
+      "ld1b { z26.b }, p1/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z25.b }, p1/Z, [x27, #2, MUL VL]\n"
+      "zip1 z2.b, z28.b, z26.b\n"
+      "zip1 z1.b, z27.b, z25.b\n"
+      "ld1b { z24.b }, p1/Z, [x26, #2, MUL VL]\n"
+      "ld1b { z23.b }, p1/Z, [x23, #2, MUL VL]\n"
+      "zip1 z16.b, z31.b, z22.b\n"
+      "zip2 z22.b, z31.b, z22.b\n"
+      "ld1b { z21.b }, p1/Z, [x21, #2, MUL VL]\n"
+      "ld1b { z20.b }, p1/Z, [x20, #2, MUL VL]\n"
+      "zip1 z0.b, z24.b, z21.b\n"
+      "zip1 z31.b, z23.b, z20.b\n"
+      "zip1 z19.b, z14.b, z13.b\n"
+      "zip1 z18.b, z30.b, z29.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
       "addvl x10, x10, #3\n"
+      "zip2 z16.b, z30.b, z29.b\n"
+      "zip1 z17.b, z12.b, z11.b\n"
+      "st1b { z22.b }, p1, [x22, #1, MUL VL]\n"
+      "addvl x9, x9, #3\n"
+      "st1b { z19.b }, p1, [x22, #2, MUL VL]\n"
+      "zip2 z30.b, z28.b, z26.b\n"
+      "zip2 z29.b, z27.b, z25.b\n"
+      "addvl x28, x28, #3\n"
+      "st1b { z18.b }, p1, [x22, #3, MUL VL]\n"
+      "zip2 z28.b, z24.b, z21.b\n"
+      "zip2 z27.b, z23.b, z20.b\n"
+      "addvl x27, x27, #3\n"
+      "st1b { z16.b }, p1, [x22, #4, MUL VL]\n"
+      "zip2 z21.b, z14.b, z13.b\n"
+      "zip1 z16.b, z10.b, z9.b\n"
       "addvl x26, x26, #3\n"
-      "zip1 z21.b, z1.b, z0.b\n"
-      "zip2 z20.b, z1.b, z0.b\n"
-      "addvl x25, x25, #3\n"
-      "addvl x24, x24, #3\n"
-      "zip1 z1.b, z29.b, z25.b\n"
-      "zip1 z0.b, z28.b, z24.b\n"
+      "st1b { z17.b }, p1, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z20.b, z10.b, z9.b\n"
+      "zip2 z19.b, z12.b, z11.b\n"
+      "zip1 z18.b, z6.b, z5.b\n"
+      "zip2 z17.b, z6.b, z5.b\n"
+      "st1b { z21.b }, p1, [x22]\n"
       "addvl x23, x23, #3\n"
-      "addvl x22, x22, #3\n"
-      "zip1 z19.b, z15.b, z14.b\n"
-      "zip1 z18.b, z13.b, z12.b\n"
+      "st1b { z16.b }, p1, [x22, #1, MUL VL]\n"
+      "zip1 z16.b, z8.b, z7.b\n"
+      "zip2 z26.b, z8.b, z7.b\n"
       "addvl x21, x21, #3\n"
+      "st1b { z20.b }, p1, [x22, #2, MUL VL]\n"
+      "zip1 z25.b, z2.b, z1.b\n"
+      "zip1 z24.b, z4.b, z3.b\n"
       "addvl x20, x20, #3\n"
-      "zip2 z17.b, z13.b, z12.b\n"
-      "zip1 z16.b, z11.b, z10.b\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "zip2 z31.b, z9.b, z31.b\n"
-      "zip2 z30.b, z8.b, z30.b\n"
-      "st1b { z19.b }, p1, [x27, #2, MUL VL]\n"
-      "zip2 z29.b, z29.b, z25.b\n"
-      "zip2 z28.b, z28.b, z24.b\n"
-      "st1b { z18.b }, p1, [x27, #3, MUL VL]\n"
-      "zip2 z21.b, z15.b, z14.b\n"
-      "zip1 z20.b, z7.b, z6.b\n"
-      "st1b { z17.b }, p1, [x27, #4, MUL VL]\n"
-      "zip2 z19.b, z7.b, z6.b\n"
-      "zip2 z18.b, z11.b, z10.b\n"
-      "st1b { z16.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z17.b, z27.b, z26.b\n"
-      "zip2 z16.b, z27.b, z26.b\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "zip1 z27.b, z5.b, z4.b\n"
-      "zip2 z26.b, z5.b, z4.b\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "zip1 z25.b, z3.b, z2.b\n"
-      "zip1 z24.b, z23.b, z22.b\n"
-      "st1b { z19.b }, p1, [x27, #2, MUL VL]\n"
-      "zip2 z23.b, z23.b, z22.b\n"
-      "zip1 z22.b, z1.b, z0.b\n"
-      "st1b { z18.b }, p1, [x27, #3, MUL VL]\n"
-      "zip2 z21.b, z3.b, z2.b\n"
-      "zip1 z20.b, z31.b, z30.b\n"
-      "st1b { z17.b }, p1, [x27, #4, MUL VL]\n"
-      "zip2 z19.b, z31.b, z30.b\n"
-      "zip2 z18.b, z1.b, z0.b\n"
-      "st1b { z16.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z17.b, z29.b, z28.b\n"
-      "zip2 z16.b, z29.b, z28.b\n"
-      "st1b { z27.b }, p1, [x27]\n"
-      "st1b { z26.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z25.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z24.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z23.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z22.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z18.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "st1b { z19.b }, p1, [x22, #3, MUL VL]\n"
+      "zip2 z23.b, z4.b, z3.b\n"
+      "zip1 z22.b, z0.b, z31.b\n"
+      "st1b { z18.b }, p1, [x22, #4, MUL VL]\n"
+      "zip2 z21.b, z2.b, z1.b\n"
+      "zip1 z20.b, z30.b, z29.b\n"
+      "st1b { z17.b }, p1, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z19.b, z30.b, z29.b\n"
+      "zip2 z18.b, z0.b, z31.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "zip1 z17.b, z28.b, z27.b\n"
+      "zip2 z16.b, z28.b, z27.b\n"
+      "st1b { z26.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z25.b }, p1, [x22, #2, MUL VL]\n"
+      "st1b { z24.b }, p1, [x22, #3, MUL VL]\n"
+      "st1b { z23.b }, p1, [x22, #4, MUL VL]\n"
+      "st1b { z22.b }, p1, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1b { z21.b }, p1, [x22]\n"
+      "st1b { z20.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z19.b }, p1, [x22, #2, MUL VL]\n"
+      "st1b { z18.b }, p1, [x22, #3, MUL VL]\n"
+      "st1b { z17.b }, p1, [x22, #4, MUL VL]\n"
+      "st1b { z16.b }, p1, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x9, 5f\n"
+      "cbz x25, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.b, XZR, x9\n"
-      "decw x9, ALL, MUL #3\n"
-      "ld1b { z24.b }, p0/Z, [x10]\n"
+      "whilelt p0.b, XZR, x25\n"
+      "ld1b { z19.b }, p0/Z, [x10]\n"
+      "ld1b { z18.b }, p0/Z, [x9]\n"
+      "decw x25, ALL, MUL #3\n"
+      "ld1b { z17.b }, p0/Z, [x28]\n"
+      "ld1b { z16.b }, p0/Z, [x27]\n"
+      "zip1 z26.b, z19.b, z17.b\n"
+      "zip1 z25.b, z18.b, z16.b\n"
+      "ld1b { z21.b }, p0/Z, [x26]\n"
+      "ld1b { z20.b }, p0/Z, [x23]\n"
+      "zip2 z24.b, z19.b, z17.b\n"
+      "zip2 z19.b, z18.b, z16.b\n"
+      "ld1b { z18.b }, p0/Z, [x21]\n"
+      "ld1b { z16.b }, p0/Z, [x20]\n"
+      "zip1 z23.b, z21.b, z18.b\n"
+      "zip1 z17.b, z20.b, z16.b\n"
+      "zip2 z22.b, z21.b, z18.b\n"
+      "zip2 z16.b, z20.b, z16.b\n"
+      "cmp x25, #0x0\n"
       "incd x10, ALL, MUL #6\n"
-      "ld1b { z23.b }, p0/Z, [x26]\n"
+      "incd x9, ALL, MUL #6\n"
+      "incd x28, ALL, MUL #6\n"
+      "zip1 z21.b, z26.b, z25.b\n"
+      "zip2 z20.b, z26.b, z25.b\n"
+      "incd x27, ALL, MUL #6\n"
       "incd x26, ALL, MUL #6\n"
-      "ld1b { z19.b }, p0/Z, [x25]\n"
-      "incd x25, ALL, MUL #6\n"
-      "ld1b { z18.b }, p0/Z, [x24]\n"
-      "incd x24, ALL, MUL #6\n"
-      "ld1b { z22.b }, p0/Z, [x23]\n"
-      "ld1b { z21.b }, p0/Z, [x22]\n"
-      "ld1b { z17.b }, p0/Z, [x21]\n"
-      "cmp x9, #0x0\n"
+      "zip1 z19.b, z24.b, z19.b\n"
+      "zip1 z18.b, z23.b, z17.b\n"
       "incd x23, ALL, MUL #6\n"
-      "ld1b { z16.b }, p0/Z, [x20]\n"
-      "zip1 z20.b, z24.b, z19.b\n"
-      "zip2 z24.b, z24.b, z19.b\n"
-      "incd x22, ALL, MUL #6\n"
-      "zip1 z19.b, z23.b, z18.b\n"
-      "zip2 z18.b, z23.b, z18.b\n"
       "incd x21, ALL, MUL #6\n"
-      "incd x20, ALL, MUL #6\n"
-      "zip1 z23.b, z22.b, z17.b\n"
-      "zip2 z22.b, z22.b, z17.b\n"
-      "zip1 z17.b, z21.b, z16.b\n"
-      "zip2 z16.b, z21.b, z16.b\n"
-      "zip1 z21.b, z20.b, z19.b\n"
-      "zip2 z20.b, z20.b, z19.b\n"
-      "zip1 z19.b, z24.b, z18.b\n"
-      "zip1 z18.b, z23.b, z17.b\n"
       "zip2 z17.b, z23.b, z17.b\n"
       "zip1 z16.b, z22.b, z16.b\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z18.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "incd x20, ALL, MUL #6\n"
+      "st1b { z21.b }, p1, [x22]\n"
+      "st1b { z20.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z19.b }, p1, [x22, #2, MUL VL]\n"
+      "st1b { z18.b }, p1, [x22, #3, MUL VL]\n"
+      "st1b { z17.b }, p1, [x22, #4, MUL VL]\n"
+      "st1b { z16.b }, p1, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x8\n"
@@ -222,106 +222,106 @@ void sve_transpose_interleave_3VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x10, %x[in]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cntb x20, ALL, MUL #3\n"
+      "add x27, x28, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x10, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "csel x25, x25, %x[pad_row], GE\n"
+      "add %x[in], x27, %x[in_stride]\n"
+      "csel x27, x27, %x[pad_row], GT\n"
+      "csel x28, x28, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x26, x26, %x[pad_row], GT\n"
+      "csel x9, x9, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1b { z27.b }, p1/Z, [x10]\n"
-      "ld1b { z22.b }, p1/Z, [x26]\n"
+      "ld1b { z21.b }, p1/Z, [x10]\n"
+      "ld1b { z20.b }, p1/Z, [x9]\n"
       "sub x21, x21, x20\n"
-      "ld1b { z21.b }, p1/Z, [x25]\n"
-      "ld1b { z17.b }, p1/Z, [x24]\n"
       "cmp x21, x20\n"
-      "ld1b { z26.b }, p1/Z, [x10, #1, MUL VL]\n"
-      "ld1b { z25.b }, p1/Z, [x26, #1, MUL VL]\n"
-      "ld1b { z20.b }, p1/Z, [x25, #1, MUL VL]\n"
-      "ld1b { z19.b }, p1/Z, [x24, #1, MUL VL]\n"
-      "ld1b { z30.b }, p1/Z, [x10, #2, MUL VL]\n"
-      "ld1b { z29.b }, p1/Z, [x26, #2, MUL VL]\n"
-      "zip1 z18.b, z27.b, z21.b\n"
-      "zip1 z16.b, z22.b, z17.b\n"
-      "ld1b { z24.b }, p1/Z, [x25, #2, MUL VL]\n"
-      "ld1b { z23.b }, p1/Z, [x24, #2, MUL VL]\n"
-      "zip2 z28.b, z27.b, z21.b\n"
-      "zip2 z17.b, z22.b, z17.b\n"
-      "zip1 z22.b, z26.b, z20.b\n"
-      "zip1 z21.b, z25.b, z19.b\n"
+      "ld1b { z17.b }, p1/Z, [x28]\n"
+      "ld1b { z16.b }, p1/Z, [x27]\n"
+      "zip1 z31.b, z21.b, z17.b\n"
+      "zip1 z30.b, z20.b, z16.b\n"
+      "ld1b { z19.b }, p1/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z18.b }, p1/Z, [x9, #1, MUL VL]\n"
+      "zip2 z29.b, z21.b, z17.b\n"
+      "zip2 z28.b, z20.b, z16.b\n"
+      "ld1b { z17.b }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z16.b }, p1/Z, [x27, #1, MUL VL]\n"
+      "zip1 z27.b, z19.b, z17.b\n"
+      "zip1 z26.b, z18.b, z16.b\n"
+      "ld1b { z22.b }, p1/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z21.b }, p1/Z, [x9, #2, MUL VL]\n"
+      "zip2 z25.b, z19.b, z17.b\n"
+      "zip2 z20.b, z18.b, z16.b\n"
+      "ld1b { z19.b }, p1/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z18.b }, p1/Z, [x27, #2, MUL VL]\n"
+      "zip1 z24.b, z22.b, z19.b\n"
+      "zip1 z23.b, z21.b, z18.b\n"
+      "zip1 z16.b, z31.b, z30.b\n"
+      "zip2 z17.b, z31.b, z30.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
       "addvl x10, x10, #3\n"
-      "addvl x26, x26, #3\n"
-      "zip2 z27.b, z26.b, z20.b\n"
-      "zip2 z20.b, z25.b, z19.b\n"
-      "addvl x25, x25, #3\n"
-      "addvl x24, x24, #3\n"
-      "zip1 z26.b, z30.b, z24.b\n"
-      "zip1 z25.b, z29.b, z23.b\n"
-      "zip1 z19.b, z18.b, z16.b\n"
-      "zip2 z16.b, z18.b, z16.b\n"
-      "zip1 z18.b, z28.b, z17.b\n"
-      "zip2 z24.b, z30.b, z24.b\n"
-      "zip2 z23.b, z29.b, z23.b\n"
-      "zip2 z17.b, z28.b, z17.b\n"
-      "st1b { z19.b }, p1, [x27]\n"
-      "st1b { z16.b }, p1, [x27, #1, MUL VL]\n"
-      "zip1 z16.b, z22.b, z21.b\n"
-      "zip2 z22.b, z22.b, z21.b\n"
-      "st1b { z18.b }, p1, [x27, #2, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z21.b, z27.b, z20.b\n"
-      "zip2 z20.b, z27.b, z20.b\n"
-      "st1b { z17.b }, p1, [x27]\n"
-      "zip1 z19.b, z26.b, z25.b\n"
-      "zip2 z18.b, z26.b, z25.b\n"
-      "st1b { z16.b }, p1, [x27, #1, MUL VL]\n"
-      "zip1 z17.b, z24.b, z23.b\n"
-      "zip2 z16.b, z24.b, z23.b\n"
-      "st1b { z22.b }, p1, [x27, #2, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #2, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z18.b }, p1, [x27]\n"
-      "st1b { z17.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #2, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 z16.b, z29.b, z28.b\n"
+      "st1b { z17.b }, p1, [x22, #1, MUL VL]\n"
+      "zip2 z22.b, z22.b, z19.b\n"
+      "addvl x9, x9, #3\n"
+      "st1b { z16.b }, p1, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z21.b, z21.b, z18.b\n"
+      "zip2 z18.b, z29.b, z28.b\n"
+      "zip1 z16.b, z27.b, z26.b\n"
+      "zip2 z17.b, z27.b, z26.b\n"
+      "st1b { z18.b }, p1, [x22]\n"
+      "addvl x28, x28, #3\n"
+      "st1b { z16.b }, p1, [x22, #1, MUL VL]\n"
+      "zip1 z16.b, z25.b, z20.b\n"
+      "zip2 z20.b, z25.b, z20.b\n"
+      "addvl x27, x27, #3\n"
+      "st1b { z17.b }, p1, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z19.b, z24.b, z23.b\n"
+      "zip2 z18.b, z24.b, z23.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "zip1 z17.b, z22.b, z21.b\n"
+      "zip2 z16.b, z22.b, z21.b\n"
+      "st1b { z20.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z19.b }, p1, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1b { z18.b }, p1, [x22]\n"
+      "st1b { z17.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z16.b }, p1, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "whilelt p0.b, XZR, x21\n"
-      "decw x21, ALL, MUL #3\n"
       "ld1b { z19.b }, p0/Z, [x10]\n"
-      "incd x10, ALL, MUL #6\n"
-      "ld1b { z18.b }, p0/Z, [x26]\n"
-      "incd x26, ALL, MUL #6\n"
-      "ld1b { z17.b }, p0/Z, [x25]\n"
-      "incd x25, ALL, MUL #6\n"
-      "ld1b { z16.b }, p0/Z, [x24]\n"
-      "incd x24, ALL, MUL #6\n"
+      "ld1b { z21.b }, p0/Z, [x9]\n"
+      "decw x21, ALL, MUL #3\n"
+      "ld1b { z18.b }, p0/Z, [x28]\n"
+      "ld1b { z16.b }, p0/Z, [x27]\n"
+      "zip1 z20.b, z19.b, z18.b\n"
+      "zip1 z17.b, z21.b, z16.b\n"
+      "zip2 z19.b, z19.b, z18.b\n"
+      "zip2 z16.b, z21.b, z16.b\n"
       "cmp x21, #0x0\n"
-      "zip1 z20.b, z19.b, z17.b\n"
-      "zip2 z19.b, z19.b, z17.b\n"
-      "zip1 z17.b, z18.b, z16.b\n"
-      "zip2 z16.b, z18.b, z16.b\n"
+      "incd x10, ALL, MUL #6\n"
+      "incd x9, ALL, MUL #6\n"
+      "incd x28, ALL, MUL #6\n"
       "zip1 z18.b, z20.b, z17.b\n"
       "zip2 z17.b, z20.b, z17.b\n"
+      "incd x27, ALL, MUL #6\n"
       "zip1 z16.b, z19.b, z16.b\n"
-      "st1b { z18.b }, p1, [x27]\n"
-      "st1b { z17.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #2, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "st1b { z18.b }, p1, [x22]\n"
+      "st1b { z17.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z16.b }, p1, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -330,7 +330,7 @@ void sve_transpose_interleave_3VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "12:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL_2x2.hpp
index 70eb77ebe5..5cf7139fe4 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_3VL_2x2.hpp
@@ -44,104 +44,104 @@ void sve_transpose_interleave_3VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x12, %x[in]\n"
-      "mov x11, %x[width]\n"
-      "cnth x10, ALL, MUL #3\n"
-      "mov x9, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x28, x12, %x[in_stride]\n"
-      "add x27, x28, %x[in_stride]\n"
-      "add x26, x27, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "cmp x11, x10\n"
+      "add x11, x12, %x[in_stride]\n"
+      "add x10, x11, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "mov x27, %x[width]\n"
+      "cnth x26, ALL, MUL #3\n"
+      "add x25, x28, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "cmp x27, x26\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1h { z18.h }, p2/Z, [x12]\n"
-      "ld1h { z29.h }, p2/Z, [x12, #1, MUL VL]\n"
-      "mov x21, x9\n"
-      "add x9, x9, %x[out_stride]\n"
-      "ld1h { z17.h }, p2/Z, [x28]\n"
-      "ld1h { z16.h }, p2/Z, [x28, #1, MUL VL]\n"
-      "mov x20, x9\n"
-      "sub x11, x11, x10\n"
-      "ld1h { z21.h }, p2/Z, [x27]\n"
-      "ld1h { z28.h }, p2/Z, [x27, #1, MUL VL]\n"
-      "cmp x11, x10\n"
-      "add x9, x9, %x[out_stride]\n"
-      "ld1h { z20.h }, p2/Z, [x26]\n"
-      "ld1h { z27.h }, p2/Z, [x26, #1, MUL VL]\n"
-      "ld1h { z26.h }, p2/Z, [x25]\n"
-      "ld1h { z19.h }, p2/Z, [x24]\n"
-      "zip1 z25.h, z18.h, z17.h\n"
-      "zip2 z24.h, z18.h, z17.h\n"
-      "ld1h { z23.h }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z18.h }, p2/Z, [x28, #2, MUL VL]\n"
-      "zip1 z22.h, z29.h, z16.h\n"
-      "zip2 z8.h, z29.h, z16.h\n"
-      "ld1h { z17.h }, p2/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x26, #2, MUL VL]\n"
-      "zip1 z7.h, z21.h, z20.h\n"
-      "zip2 z6.h, z21.h, z20.h\n"
-      "ld1h { z21.h }, p2/Z, [x25, #1, MUL VL]\n"
-      "ld1h { z5.h }, p2/Z, [x25, #2, MUL VL]\n"
-      "zip1 z4.h, z28.h, z27.h\n"
-      "zip1 z3.h, z26.h, z19.h\n"
-      "ld1h { z20.h }, p2/Z, [x24, #1, MUL VL]\n"
-      "ld1h { z2.h }, p2/Z, [x24, #2, MUL VL]\n"
-      "zip2 z1.h, z26.h, z19.h\n"
-      "zip1 z0.h, z23.h, z18.h\n"
-      "ld1h { z19.h }, p2/Z, [x23]\n"
-      "ld1h { z31.h }, p2/Z, [x23, #1, MUL VL]\n"
-      "zip2 z30.h, z23.h, z18.h\n"
-      "zip2 z29.h, z28.h, z27.h\n"
-      "ld1h { z28.h }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z18.h }, p2/Z, [x22]\n"
-      "zip1 z27.h, z17.h, z16.h\n"
-      "zip2 z26.h, z17.h, z16.h\n"
-      "ld1h { z17.h }, p2/Z, [x22, #1, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x22, #2, MUL VL]\n"
-      "st1h { z25.h }, p2, [x21]\n"
-      "zip1 z25.h, z21.h, z20.h\n"
-      "st1h { z24.h }, p2, [x21, #1, MUL VL]\n"
-      "zip2 z24.h, z21.h, z20.h\n"
-      "zip1 z23.h, z5.h, z2.h\n"
+      "ld1h { z17.h }, p2/Z, [x12]\n"
+      "ld1h { z23.h }, p2/Z, [x12, #1, MUL VL]\n"
+      "mov x21, x22\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1h { z16.h }, p2/Z, [x11]\n"
+      "ld1h { z20.h }, p2/Z, [x11, #1, MUL VL]\n"
+      "zip1 z9.h, z17.h, z16.h\n"
+      "zip2 z8.h, z17.h, z16.h\n"
+      "ld1h { z17.h }, p2/Z, [x10]\n"
+      "ld1h { z22.h }, p2/Z, [x10, #1, MUL VL]\n"
+      "zip1 z7.h, z23.h, z20.h\n"
+      "mov x20, x22\n"
+      "ld1h { z16.h }, p2/Z, [x9]\n"
+      "ld1h { z21.h }, p2/Z, [x9, #1, MUL VL]\n"
+      "zip1 z6.h, z17.h, z16.h\n"
+      "zip2 z5.h, z17.h, z16.h\n"
+      "ld1h { z18.h }, p2/Z, [x28]\n"
+      "ld1h { z17.h }, p2/Z, [x25]\n"
+      "zip1 z4.h, z22.h, z21.h\n"
+      "zip1 z3.h, z18.h, z17.h\n"
+      "ld1h { z19.h }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x11, #2, MUL VL]\n"
+      "zip2 z2.h, z18.h, z17.h\n"
+      "zip2 z1.h, z23.h, z20.h\n"
+      "ld1h { z18.h }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z17.h }, p2/Z, [x9, #2, MUL VL]\n"
+      "zip1 z0.h, z19.h, z16.h\n"
+      "zip2 z31.h, z19.h, z16.h\n"
+      "ld1h { z20.h }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1h { z30.h }, p2/Z, [x28, #2, MUL VL]\n"
+      "zip2 z29.h, z22.h, z21.h\n"
+      "zip1 z28.h, z18.h, z17.h\n"
+      "ld1h { z16.h }, p2/Z, [x25, #1, MUL VL]\n"
+      "ld1h { z19.h }, p2/Z, [x25, #2, MUL VL]\n"
+      "zip1 z27.h, z20.h, z16.h\n"
+      "zip2 z26.h, z18.h, z17.h\n"
+      "ld1h { z17.h }, p2/Z, [x24]\n"
+      "ld1h { z18.h }, p2/Z, [x24, #1, MUL VL]\n"
+      "zip2 z25.h, z20.h, z16.h\n"
+      "zip1 z24.h, z30.h, z19.h\n"
+      "ld1h { z23.h }, p2/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x23]\n"
+      "zip1 z22.h, z17.h, z16.h\n"
+      "zip2 z21.h, z17.h, z16.h\n"
+      "ld1h { z17.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x23, #2, MUL VL]\n"
+      "st1h { z9.h }, p2, [x21]\n"
+      "zip1 z20.h, z18.h, z17.h\n"
+      "st1h { z8.h }, p2, [x21, #1, MUL VL]\n"
+      "sub x27, x27, x26\n"
+      "cmp x27, x26\n"
+      "zip2 z19.h, z30.h, z19.h\n"
+      "st1h { z7.h }, p2, [x21, #2, MUL VL]\n"
       "addvl x12, x12, #3\n"
-      "st1h { z22.h }, p2, [x21, #2, MUL VL]\n"
-      "zip1 z22.h, z19.h, z18.h\n"
-      "zip2 z21.h, z19.h, z18.h\n"
+      "addvl x11, x11, #3\n"
+      "zip2 z18.h, z18.h, z17.h\n"
+      "st1h { z6.h }, p2, [x21, #3, MUL VL]\n"
+      "addvl x10, x10, #3\n"
+      "addvl x9, x9, #3\n"
+      "zip1 z17.h, z23.h, z16.h\n"
+      "st1h { z5.h }, p2, [x21, #4, MUL VL]\n"
       "addvl x28, x28, #3\n"
-      "st1h { z7.h }, p2, [x21, #3, MUL VL]\n"
-      "zip1 z20.h, z31.h, z17.h\n"
-      "addvl x27, x27, #3\n"
-      "addvl x26, x26, #3\n"
-      "st1h { z6.h }, p2, [x21, #4, MUL VL]\n"
       "addvl x25, x25, #3\n"
-      "addvl x24, x24, #3\n"
-      "zip2 z19.h, z5.h, z2.h\n"
+      "zip2 z16.h, z23.h, z16.h\n"
       "st1h { z4.h }, p2, [x21, #5, MUL VL]\n"
+      "addvl x24, x24, #3\n"
       "addvl x23, x23, #3\n"
-      "addvl x22, x22, #3\n"
-      "zip2 z18.h, z31.h, z17.h\n"
       "st1h { z3.h }, p2, [x21, #6, MUL VL]\n"
-      "zip1 z17.h, z28.h, z16.h\n"
-      "zip2 z16.h, z28.h, z16.h\n"
-      "st1h { z1.h }, p2, [x21, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z2.h }, p2, [x21, #7, MUL VL]\n"
       "addvl x21, x21, #12\n"
-      "st1h { z25.h }, p2, [x21, #-4, MUL VL]\n"
+      "st1h { z27.h }, p2, [x21, #-4, MUL VL]\n"
       "st1h { z22.h }, p2, [x21, #-3, MUL VL]\n"
       "st1h { z21.h }, p2, [x21, #-2, MUL VL]\n"
       "st1h { z20.h }, p2, [x21, #-1, MUL VL]\n"
-      "st1h { z8.h }, p2, [x20]\n"
+      "st1h { z1.h }, p2, [x20]\n"
       "st1h { z0.h }, p2, [x20, #1, MUL VL]\n"
-      "st1h { z30.h }, p2, [x20, #2, MUL VL]\n"
+      "st1h { z31.h }, p2, [x20, #2, MUL VL]\n"
       "st1h { z29.h }, p2, [x20, #3, MUL VL]\n"
-      "st1h { z27.h }, p2, [x20, #4, MUL VL]\n"
+      "st1h { z28.h }, p2, [x20, #4, MUL VL]\n"
       "st1h { z26.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z24.h }, p2, [x20, #6, MUL VL]\n"
-      "st1h { z23.h }, p2, [x20, #7, MUL VL]\n"
+      "st1h { z25.h }, p2, [x20, #6, MUL VL]\n"
+      "st1h { z24.h }, p2, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
       "st1h { z19.h }, p2, [x20, #-4, MUL VL]\n"
       "st1h { z18.h }, p2, [x20, #-3, MUL VL]\n"
@@ -149,67 +149,67 @@ void sve_transpose_interleave_3VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x11, 5f\n"
+      "cbz x27, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x21, x11\n"
-      "mov x20, x9\n"
-      "decw x11, ALL, MUL #3\n"
-      "add x9, x9, %x[out_stride]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p0.h, XZR, x21\n"
-      "cmp x11, #0x0\n"
-      "ld1h { z19.h }, p1/Z, [x12]\n"
-      "ld1h { z18.h }, p1/Z, [x28]\n"
-      "ld1h { z0.h }, p1/Z, [x27]\n"
-      "ld1h { z31.h }, p1/Z, [x26]\n"
-      "ld1h { z30.h }, p1/Z, [x25]\n"
-      "ld1h { z29.h }, p1/Z, [x24]\n"
-      "ld1h { z17.h }, p0/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x28, #1, MUL VL]\n"
+      "mov x20, x27\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z0.h }, p1/Z, [x12]\n"
+      "ld1h { z16.h }, p1/Z, [x11]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z21.h }, p0/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z19.h }, p0/Z, [x11, #1, MUL VL]\n"
+      "ld1h { z31.h }, p1/Z, [x10]\n"
+      "ld1h { z30.h }, p0/Z, [x10, #1, MUL VL]\n"
+      "mov x20, x22\n"
+      "decw x27, ALL, MUL #3\n"
+      "ld1h { z18.h }, p1/Z, [x9]\n"
+      "ld1h { z29.h }, p0/Z, [x9, #1, MUL VL]\n"
       "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "ld1h { z28.h }, p1/Z, [x28]\n"
+      "ld1h { z20.h }, p1/Z, [x25]\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
+      "ld1h { z27.h }, p0/Z, [x28, #1, MUL VL]\n"
       "addvl x28, x28, #1\n"
-      "ld1h { z22.h }, p0/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z21.h }, p0/Z, [x26, #1, MUL VL]\n"
-      "addvl x27, x27, #1\n"
-      "addvl x26, x26, #1\n"
-      "ld1h { z28.h }, p0/Z, [x25, #1, MUL VL]\n"
+      "ld1h { z26.h }, p0/Z, [x25, #1, MUL VL]\n"
       "addvl x25, x25, #1\n"
-      "ld1h { z27.h }, p0/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z25.h }, p1/Z, [x24]\n"
+      "ld1h { z24.h }, p0/Z, [x24, #1, MUL VL]\n"
       "addvl x24, x24, #1\n"
-      "ld1h { z26.h }, p1/Z, [x23]\n"
-      "ld1h { z25.h }, p0/Z, [x23, #1, MUL VL]\n"
+      "zip1 z17.h, z0.h, z16.h\n"
+      "ld1h { z23.h }, p1/Z, [x23]\n"
+      "ld1h { z22.h }, p0/Z, [x23, #1, MUL VL]\n"
       "addvl x23, x23, #1\n"
-      "zip1 z20.h, z19.h, z18.h\n"
-      "ld1h { z24.h }, p1/Z, [x22]\n"
-      "ld1h { z23.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "addvl x22, x22, #1\n"
-      "zip2 z19.h, z19.h, z18.h\n"
-      "zip1 z18.h, z17.h, z16.h\n"
-      "zip1 z17.h, z0.h, z31.h\n"
+      "zip2 z16.h, z0.h, z16.h\n"
+      "zip1 z21.h, z21.h, z19.h\n"
+      "zip1 z19.h, z31.h, z18.h\n"
+      "st1h { z17.h }, p2, [x20]\n"
+      "cmp x27, #0x0\n"
+      "zip2 z18.h, z31.h, z18.h\n"
+      "zip1 z17.h, z30.h, z29.h\n"
+      "st1h { z16.h }, p2, [x20, #1, MUL VL]\n"
       "incd x12, ALL, MUL #4\n"
+      "zip1 z16.h, z28.h, z20.h\n"
+      "zip2 z20.h, z28.h, z20.h\n"
+      "st1h { z21.h }, p2, [x20, #2, MUL VL]\n"
+      "incd x11, ALL, MUL #4\n"
+      "st1h { z19.h }, p2, [x20, #3, MUL VL]\n"
+      "incd x10, ALL, MUL #4\n"
+      "incd x9, ALL, MUL #4\n"
+      "zip1 z19.h, z27.h, z26.h\n"
+      "st1h { z18.h }, p2, [x20, #4, MUL VL]\n"
       "incd x28, ALL, MUL #4\n"
-      "zip2 z16.h, z0.h, z31.h\n"
-      "zip1 z22.h, z22.h, z21.h\n"
-      "st1h { z20.h }, p2, [x20]\n"
-      "incd x27, ALL, MUL #4\n"
-      "zip1 z21.h, z30.h, z29.h\n"
-      "zip2 z20.h, z30.h, z29.h\n"
-      "st1h { z19.h }, p2, [x20, #1, MUL VL]\n"
-      "incd x26, ALL, MUL #4\n"
-      "st1h { z18.h }, p2, [x20, #2, MUL VL]\n"
       "incd x25, ALL, MUL #4\n"
+      "zip1 z18.h, z25.h, z23.h\n"
+      "st1h { z17.h }, p2, [x20, #5, MUL VL]\n"
       "incd x24, ALL, MUL #4\n"
-      "zip1 z19.h, z28.h, z27.h\n"
-      "st1h { z17.h }, p2, [x20, #3, MUL VL]\n"
       "incd x23, ALL, MUL #4\n"
-      "incd x22, ALL, MUL #4\n"
-      "zip1 z18.h, z26.h, z24.h\n"
-      "st1h { z16.h }, p2, [x20, #4, MUL VL]\n"
-      "zip2 z17.h, z26.h, z24.h\n"
-      "zip1 z16.h, z25.h, z23.h\n"
-      "st1h { z22.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #6, MUL VL]\n"
+      "zip2 z17.h, z25.h, z23.h\n"
+      "st1h { z16.h }, p2, [x20, #6, MUL VL]\n"
+      "zip1 z16.h, z24.h, z22.h\n"
+      "add x22, x22, %x[out_stride]\n"
       "st1h { z20.h }, p2, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
       "st1h { z19.h }, p2, [x20, #-4, MUL VL]\n"
@@ -227,64 +227,64 @@ void sve_transpose_interleave_3VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "mov x12, %x[in]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #3\n"
+      "add x11, x12, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "mov x9, %x[out]\n"
-      "sub %x[height], %x[height], #0x2\n"
-      "add x28, x12, %x[in_stride]\n"
-      "add %x[in], x28, %x[in_stride]\n"
-      "csel x28, x28, %x[pad_row], GT\n"
+      "add %x[in], x11, %x[in_stride]\n"
+      "csel x11, x11, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x2\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1h { z19.h }, p2/Z, [x12]\n"
-      "ld1h { z23.h }, p2/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z17.h }, p2/Z, [x12]\n"
+      "ld1h { z22.h }, p2/Z, [x12, #1, MUL VL]\n"
       "sub x21, x21, x20\n"
-      "ld1h { z18.h }, p2/Z, [x28]\n"
-      "ld1h { z17.h }, p2/Z, [x28, #1, MUL VL]\n"
       "cmp x21, x20\n"
-      "ld1h { z22.h }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x11]\n"
+      "ld1h { z21.h }, p2/Z, [x11, #1, MUL VL]\n"
+      "zip1 z18.h, z17.h, z16.h\n"
+      "zip2 z17.h, z17.h, z16.h\n"
+      "ld1h { z20.h }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z19.h }, p2/Z, [x11, #2, MUL VL]\n"
+      "zip1 z16.h, z22.h, z21.h\n"
+      "st1h { z18.h }, p2, [x22]\n"
+      "st1h { z17.h }, p2, [x22, #1, MUL VL]\n"
       "addvl x12, x12, #3\n"
-      "addvl x28, x28, #3\n"
-      "zip1 z21.h, z19.h, z18.h\n"
-      "zip2 z20.h, z19.h, z18.h\n"
-      "zip1 z19.h, z23.h, z17.h\n"
-      "zip2 z18.h, z23.h, z17.h\n"
-      "zip1 z17.h, z22.h, z16.h\n"
-      "zip2 z16.h, z22.h, z16.h\n"
-      "st1h { z21.h }, p2, [x9]\n"
-      "st1h { z20.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z19.h }, p2, [x9, #2, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
-      "st1h { z18.h }, p2, [x9]\n"
-      "st1h { z17.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #2, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
+      "addvl x11, x11, #3\n"
+      "zip2 z18.h, z22.h, z21.h\n"
+      "st1h { z16.h }, p2, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z17.h, z20.h, z19.h\n"
+      "zip2 z16.h, z20.h, z19.h\n"
+      "st1h { z18.h }, p2, [x22]\n"
+      "st1h { z17.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decw x21, ALL, MUL #3\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z20.h }, p0/Z, [x12]\n"
+      "ld1h { z17.h }, p0/Z, [x11]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
-      "cmp x21, #0x0\n"
-      "ld1h { z20.h }, p1/Z, [x12]\n"
-      "ld1h { z17.h }, p1/Z, [x28]\n"
       "ld1h { z19.h }, p0/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z16.h }, p0/Z, [x11, #1, MUL VL]\n"
+      "decw x21, ALL, MUL #3\n"
       "addvl x12, x12, #1\n"
-      "ld1h { z16.h }, p0/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #1\n"
       "zip1 z18.h, z20.h, z17.h\n"
       "zip2 z17.h, z20.h, z17.h\n"
-      "incd x12, ALL, MUL #4\n"
-      "incd x28, ALL, MUL #4\n"
+      "addvl x11, x11, #1\n"
+      "cmp x21, #0x0\n"
       "zip1 z16.h, z19.h, z16.h\n"
-      "st1h { z18.h }, p2, [x9]\n"
-      "st1h { z17.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #2, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
+      "st1h { z18.h }, p2, [x22]\n"
+      "incd x12, ALL, MUL #4\n"
+      "incd x11, ALL, MUL #4\n"
+      "st1h { z17.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #2, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -293,7 +293,7 @@ void sve_transpose_interleave_3VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "12:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp
index 539b4946b4..ae2ae8c310 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp
@@ -34,67 +34,67 @@ void sve_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "ptrue p2.b\n"
+      "ptrue p4.b\n"
       "blt 4f\n"
       "1:"  // Main row loop: Head
-      "mov x27, %x[in]\n"
-      "mov x26, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "mov x25, %x[width]\n"
-      "add x24, x27, %x[in_stride]\n"
+      "mov x26, %x[in]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "mov x21, %x[width]\n"
       "2:"  // Main row loop: Column loop
-      "mov x21, x25\n"
-      "mov x20, x26\n"
-      "dech x25, ALL, MUL #4\n"
-      "add x26, x26, %x[out_stride]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z31.h }, p1/Z, [x27]\n"
-      "ld1h { z30.h }, p1/Z, [x24]\n"
-      "ld1h { z29.h }, p1/Z, [x23]\n"
-      "ld1h { z28.h }, p1/Z, [x22]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z27.h }, p0/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z26.h }, p0/Z, [x24, #1, MUL VL]\n"
-      "ld1h { z25.h }, p0/Z, [x23, #1, MUL VL]\n"
-      "ld1h { z24.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "whilelt p0.h, XZR, x21\n"
-      "cmp x25, #0x0\n"
-      "ld1h { z23.h }, p1/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z22.h }, p1/Z, [x24, #2, MUL VL]\n"
-      "ld1h { z21.h }, p1/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z20.h }, p1/Z, [x22, #2, MUL VL]\n"
-      "ld1h { z19.h }, p0/Z, [x27, #3, MUL VL]\n"
-      "ld1h { z18.h }, p0/Z, [x24, #3, MUL VL]\n"
-      "addvl x27, x27, #4\n"
+      "mov x20, x21\n"
+      "whilelt p3.h, XZR, x20\n"
+      "ld1h { z31.h }, p3/Z, [x26]\n"
+      "ld1h { z30.h }, p3/Z, [x25]\n"
+      "dech x20\n"
+      "whilelt p2.h, XZR, x20\n"
+      "ld1h { z29.h }, p2/Z, [x26, #1, MUL VL]\n"
+      "ld1h { z28.h }, p2/Z, [x25, #1, MUL VL]\n"
+      "dech x20\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z27.h }, p1/Z, [x26, #2, MUL VL]\n"
+      "ld1h { z26.h }, p1/Z, [x25, #2, MUL VL]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z25.h }, p0/Z, [x26, #3, MUL VL]\n"
+      "ld1h { z24.h }, p0/Z, [x25, #3, MUL VL]\n"
+      "mov x20, x22\n"
+      "dech x21, ALL, MUL #4\n"
+      "ld1h { z23.h }, p3/Z, [x24]\n"
+      "ld1h { z22.h }, p2/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z21.h }, p1/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z20.h }, p0/Z, [x24, #3, MUL VL]\n"
+      "cmp x21, #0x0\n"
+      "addvl x26, x26, #4\n"
+      "ld1h { z19.h }, p3/Z, [x23]\n"
+      "ld1h { z18.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "addvl x25, x25, #4\n"
       "addvl x24, x24, #4\n"
-      "ld1h { z17.h }, p0/Z, [x23, #3, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x22, #3, MUL VL]\n"
-      "st1h { z31.h }, p2, [x20]\n"
+      "ld1h { z17.h }, p1/Z, [x23, #2, MUL VL]\n"
+      "ld1h { z16.h }, p0/Z, [x23, #3, MUL VL]\n"
+      "st1h { z31.h }, p4, [x20]\n"
       "addvl x23, x23, #4\n"
-      "st1h { z27.h }, p2, [x20, #1, MUL VL]\n"
-      "addvl x22, x22, #4\n"
-      "st1h { z23.h }, p2, [x20, #2, MUL VL]\n"
-      "st1h { z19.h }, p2, [x20, #3, MUL VL]\n"
-      "st1h { z30.h }, p2, [x20, #4, MUL VL]\n"
-      "st1h { z26.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z22.h }, p2, [x20, #6, MUL VL]\n"
-      "st1h { z18.h }, p2, [x20, #7, MUL VL]\n"
+      "st1h { z29.h }, p4, [x20, #1, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z27.h }, p4, [x20, #2, MUL VL]\n"
+      "st1h { z25.h }, p4, [x20, #3, MUL VL]\n"
+      "st1h { z30.h }, p4, [x20, #4, MUL VL]\n"
+      "st1h { z28.h }, p4, [x20, #5, MUL VL]\n"
+      "st1h { z26.h }, p4, [x20, #6, MUL VL]\n"
+      "st1h { z24.h }, p4, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
-      "st1h { z29.h }, p2, [x20, #-8, MUL VL]\n"
-      "st1h { z25.h }, p2, [x20, #-7, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #-6, MUL VL]\n"
-      "st1h { z17.h }, p2, [x20, #-5, MUL VL]\n"
-      "st1h { z28.h }, p2, [x20, #-4, MUL VL]\n"
-      "st1h { z24.h }, p2, [x20, #-3, MUL VL]\n"
-      "st1h { z20.h }, p2, [x20, #-2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
+      "st1h { z23.h }, p4, [x20, #-8, MUL VL]\n"
+      "st1h { z22.h }, p4, [x20, #-7, MUL VL]\n"
+      "st1h { z21.h }, p4, [x20, #-6, MUL VL]\n"
+      "st1h { z20.h }, p4, [x20, #-5, MUL VL]\n"
+      "st1h { z19.h }, p4, [x20, #-4, MUL VL]\n"
+      "st1h { z18.h }, p4, [x20, #-3, MUL VL]\n"
+      "st1h { z17.h }, p4, [x20, #-2, MUL VL]\n"
+      "st1h { z16.h }, p4, [x20, #-1, MUL VL]\n"
       "bgt 2b\n"
       "3:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x4\n"
@@ -103,32 +103,32 @@ void sve_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt
       "cbz %x[height], 8f\n"
       "4:"  // Main loop skip
       "5:"  // Tail row loop: Head
-      "mov x27, %x[in]\n"
-      "mov x26, %x[out]\n"
+      "mov x26, %x[in]\n"
+      "add %x[in], x26, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x1\n"
       "mov x21, %x[width]\n"
-      "add %x[in], x27, %x[in_stride]\n"
       "6:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "dech x21, ALL, MUL #4\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z19.h }, p0/Z, [x26]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z18.h }, p0/Z, [x26, #1, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z19.h }, p1/Z, [x27]\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z17.h }, p0/Z, [x26, #2, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z18.h }, p0/Z, [x27, #1, MUL VL]\n"
+      "dech x21, ALL, MUL #4\n"
       "whilelt p0.h, XZR, x20\n"
       "cmp x21, #0x0\n"
-      "ld1h { z17.h }, p1/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x27, #3, MUL VL]\n"
-      "addvl x27, x27, #4\n"
-      "st1h { z19.h }, p2, [x26]\n"
-      "st1h { z18.h }, p2, [x26, #1, MUL VL]\n"
-      "st1h { z17.h }, p2, [x26, #2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x26, #3, MUL VL]\n"
-      "add x26, x26, %x[out_stride]\n"
+      "ld1h { z16.h }, p0/Z, [x26, #3, MUL VL]\n"
+      "st1h { z19.h }, p4, [x22]\n"
+      "addvl x26, x26, #4\n"
+      "st1h { z18.h }, p4, [x22, #1, MUL VL]\n"
+      "st1h { z17.h }, p4, [x22, #2, MUL VL]\n"
+      "st1h { z16.h }, p4, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 6b\n"
       "7:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -137,7 +137,7 @@ void sve_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt
       "8:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL_1x4.hpp
index 5f0b4ea8d6..e87c602b54 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL_1x4.hpp
@@ -44,145 +44,145 @@ void sve_transpose_interleave_4VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x10, %x[in]\n"
-      "mov x9, %x[width]\n"
-      "cntb x28, ALL, MUL #2\n"
-      "mov x27, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x26, x10, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add x23, x24, %x[in_stride]\n"
-      "cmp x9, x28\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
+      "mov x25, %x[width]\n"
+      "cntb x24, ALL, MUL #2\n"
+      "add x23, x26, %x[in_stride]\n"
+      "add x21, x23, %x[in_stride]\n"
       "add x20, x21, %x[in_stride]\n"
+      "cmp x25, x24\n"
       "add %x[in], x20, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1b { z19.b }, p1/Z, [x10]\n"
-      "ld1b { z18.b }, p1/Z, [x26]\n"
-      "sub x9, x9, x28\n"
-      "ld1b { z17.b }, p1/Z, [x25]\n"
-      "ld1b { z16.b }, p1/Z, [x24]\n"
-      "cmp x9, x28\n"
-      "ld1b { z24.b }, p1/Z, [x23]\n"
-      "ld1b { z23.b }, p1/Z, [x22]\n"
-      "ld1b { z22.b }, p1/Z, [x21]\n"
-      "ld1b { z21.b }, p1/Z, [x20]\n"
-      "ld1b { z29.b }, p1/Z, [x10, #1, MUL VL]\n"
-      "ld1b { z28.b }, p1/Z, [x26, #1, MUL VL]\n"
-      "zip1 z4.b, z19.b, z17.b\n"
-      "zip1 z3.b, z18.b, z16.b\n"
-      "ld1b { z27.b }, p1/Z, [x25, #1, MUL VL]\n"
-      "ld1b { z20.b }, p1/Z, [x24, #1, MUL VL]\n"
-      "zip2 z2.b, z19.b, z17.b\n"
-      "zip2 z1.b, z18.b, z16.b\n"
-      "ld1b { z19.b }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1b { z18.b }, p1/Z, [x22, #1, MUL VL]\n"
-      "zip1 z26.b, z24.b, z22.b\n"
-      "zip1 z25.b, z23.b, z21.b\n"
+      "ld1b { z21.b }, p1/Z, [x10]\n"
+      "ld1b { z20.b }, p1/Z, [x9]\n"
+      "sub x25, x25, x24\n"
+      "cmp x25, x24\n"
+      "ld1b { z17.b }, p1/Z, [x28]\n"
+      "ld1b { z16.b }, p1/Z, [x27]\n"
+      "zip1 z4.b, z21.b, z17.b\n"
+      "zip1 z3.b, z20.b, z16.b\n"
+      "ld1b { z19.b }, p1/Z, [x26]\n"
+      "ld1b { z18.b }, p1/Z, [x23]\n"
+      "zip2 z2.b, z21.b, z17.b\n"
+      "zip2 z1.b, z20.b, z16.b\n"
+      "ld1b { z17.b }, p1/Z, [x21]\n"
+      "ld1b { z16.b }, p1/Z, [x20]\n"
+      "zip1 z0.b, z19.b, z17.b\n"
+      "zip1 z31.b, z18.b, z16.b\n"
+      "ld1b { z24.b }, p1/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z20.b }, p1/Z, [x9, #1, MUL VL]\n"
+      "zip2 z30.b, z19.b, z17.b\n"
+      "zip2 z23.b, z18.b, z16.b\n"
+      "ld1b { z17.b }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z16.b }, p1/Z, [x27, #1, MUL VL]\n"
+      "zip1 z22.b, z24.b, z17.b\n"
+      "zip1 z21.b, z20.b, z16.b\n"
+      "ld1b { z19.b }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z18.b }, p1/Z, [x23, #1, MUL VL]\n"
+      "zip2 z29.b, z24.b, z17.b\n"
+      "zip2 z28.b, z20.b, z16.b\n"
       "ld1b { z17.b }, p1/Z, [x21, #1, MUL VL]\n"
       "ld1b { z16.b }, p1/Z, [x20, #1, MUL VL]\n"
-      "zip2 z24.b, z24.b, z22.b\n"
-      "zip2 z23.b, z23.b, z21.b\n"
-      "zip1 z22.b, z29.b, z27.b\n"
-      "zip1 z21.b, z28.b, z20.b\n"
+      "zip1 z27.b, z19.b, z17.b\n"
+      "zip1 z26.b, z18.b, z16.b\n"
+      "zip2 z25.b, z19.b, z17.b\n"
+      "zip2 z24.b, z18.b, z16.b\n"
       "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      "zip1 z16.b, z4.b, z3.b\n"
+      "zip2 z17.b, z4.b, z3.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "addvl x28, x28, #2\n"
+      "zip1 z16.b, z2.b, z1.b\n"
+      "zip2 z20.b, z2.b, z1.b\n"
+      "st1b { z17.b }, p1, [x22, #1, MUL VL]\n"
+      "addvl x27, x27, #2\n"
+      "zip1 z19.b, z0.b, z31.b\n"
+      "zip2 z18.b, z0.b, z31.b\n"
+      "st1b { z16.b }, p1, [x22, #2, MUL VL]\n"
       "addvl x26, x26, #2\n"
-      "zip2 z0.b, z29.b, z27.b\n"
-      "zip2 z31.b, z28.b, z20.b\n"
-      "addvl x25, x25, #2\n"
-      "addvl x24, x24, #2\n"
-      "zip1 z30.b, z19.b, z17.b\n"
-      "zip1 z29.b, z18.b, z16.b\n"
+      "zip1 z17.b, z30.b, z23.b\n"
+      "zip2 z16.b, z30.b, z23.b\n"
+      "st1b { z20.b }, p1, [x22, #3, MUL VL]\n"
       "addvl x23, x23, #2\n"
-      "addvl x22, x22, #2\n"
-      "zip2 z28.b, z19.b, z17.b\n"
-      "zip2 z27.b, z18.b, z16.b\n"
+      "st1b { z19.b }, p1, [x22, #4, MUL VL]\n"
       "addvl x21, x21, #2\n"
       "addvl x20, x20, #2\n"
-      "zip1 z20.b, z4.b, z3.b\n"
-      "zip2 z19.b, z4.b, z3.b\n"
-      "zip1 z18.b, z2.b, z1.b\n"
-      "zip2 z17.b, z2.b, z1.b\n"
-      "zip1 z16.b, z26.b, z25.b\n"
-      "zip2 z26.b, z26.b, z25.b\n"
-      "zip1 z25.b, z24.b, z23.b\n"
-      "zip2 z24.b, z24.b, z23.b\n"
-      "st1b { z20.b }, p1, [x27]\n"
-      "st1b { z19.b }, p1, [x27, #1, MUL VL]\n"
       "zip1 z23.b, z22.b, z21.b\n"
+      "st1b { z18.b }, p1, [x22, #5, MUL VL]\n"
       "zip2 z22.b, z22.b, z21.b\n"
-      "st1b { z18.b }, p1, [x27, #2, MUL VL]\n"
-      "zip1 z21.b, z0.b, z31.b\n"
-      "zip2 z20.b, z0.b, z31.b\n"
-      "st1b { z17.b }, p1, [x27, #3, MUL VL]\n"
-      "zip1 z19.b, z30.b, z29.b\n"
-      "zip2 z18.b, z30.b, z29.b\n"
-      "st1b { z16.b }, p1, [x27, #4, MUL VL]\n"
-      "zip1 z17.b, z28.b, z27.b\n"
-      "zip2 z16.b, z28.b, z27.b\n"
-      "st1b { z26.b }, p1, [x27, #5, MUL VL]\n"
-      "st1b { z25.b }, p1, [x27, #6, MUL VL]\n"
-      "st1b { z24.b }, p1, [x27, #7, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z23.b }, p1, [x27]\n"
-      "st1b { z22.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z21.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z20.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z18.b }, p1, [x27, #5, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #6, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #7, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 z21.b, z29.b, z28.b\n"
+      "st1b { z17.b }, p1, [x22, #6, MUL VL]\n"
+      "zip2 z20.b, z29.b, z28.b\n"
+      "zip1 z19.b, z27.b, z26.b\n"
+      "st1b { z16.b }, p1, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z18.b, z27.b, z26.b\n"
+      "zip1 z17.b, z25.b, z24.b\n"
+      "zip2 z16.b, z25.b, z24.b\n"
+      "st1b { z23.b }, p1, [x22]\n"
+      "st1b { z22.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z21.b }, p1, [x22, #2, MUL VL]\n"
+      "st1b { z20.b }, p1, [x22, #3, MUL VL]\n"
+      "st1b { z19.b }, p1, [x22, #4, MUL VL]\n"
+      "st1b { z18.b }, p1, [x22, #5, MUL VL]\n"
+      "st1b { z17.b }, p1, [x22, #6, MUL VL]\n"
+      "st1b { z16.b }, p1, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x9, 5f\n"
+      "cbz x25, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.b, XZR, x9\n"
-      "decw x9, ALL, MUL #4\n"
-      "ld1b { z20.b }, p0/Z, [x10]\n"
+      "whilelt p0.b, XZR, x25\n"
+      "ld1b { z19.b }, p0/Z, [x10]\n"
+      "ld1b { z18.b }, p0/Z, [x9]\n"
+      "decw x25, ALL, MUL #4\n"
+      "ld1b { z17.b }, p0/Z, [x28]\n"
+      "ld1b { z16.b }, p0/Z, [x27]\n"
+      "zip1 z27.b, z19.b, z17.b\n"
+      "zip1 z26.b, z18.b, z16.b\n"
+      "ld1b { z22.b }, p0/Z, [x26]\n"
+      "ld1b { z21.b }, p0/Z, [x23]\n"
+      "zip2 z25.b, z19.b, z17.b\n"
+      "zip2 z20.b, z18.b, z16.b\n"
+      "ld1b { z19.b }, p0/Z, [x21]\n"
+      "ld1b { z16.b }, p0/Z, [x20]\n"
+      "zip1 z18.b, z22.b, z19.b\n"
+      "zip1 z17.b, z21.b, z16.b\n"
+      "zip2 z24.b, z22.b, z19.b\n"
+      "zip2 z16.b, z21.b, z16.b\n"
+      "cmp x25, #0x0\n"
       "addvl x10, x10, #1\n"
-      "ld1b { z24.b }, p0/Z, [x26]\n"
+      "addvl x9, x9, #1\n"
+      "addvl x28, x28, #1\n"
+      "zip1 z23.b, z27.b, z26.b\n"
+      "zip2 z22.b, z27.b, z26.b\n"
+      "addvl x27, x27, #1\n"
       "addvl x26, x26, #1\n"
-      "ld1b { z19.b }, p0/Z, [x25]\n"
-      "addvl x25, x25, #1\n"
-      "ld1b { z18.b }, p0/Z, [x24]\n"
-      "addvl x24, x24, #1\n"
-      "ld1b { z23.b }, p0/Z, [x23]\n"
-      "ld1b { z22.b }, p0/Z, [x22]\n"
-      "ld1b { z17.b }, p0/Z, [x21]\n"
-      "cmp x9, #0x0\n"
+      "zip1 z21.b, z25.b, z20.b\n"
+      "zip2 z20.b, z25.b, z20.b\n"
       "addvl x23, x23, #1\n"
-      "ld1b { z16.b }, p0/Z, [x20]\n"
-      "zip1 z21.b, z20.b, z19.b\n"
-      "zip2 z25.b, z20.b, z19.b\n"
-      "addvl x22, x22, #1\n"
-      "zip1 z20.b, z24.b, z18.b\n"
-      "zip2 z19.b, z24.b, z18.b\n"
       "addvl x21, x21, #1\n"
-      "addvl x20, x20, #1\n"
-      "zip1 z18.b, z23.b, z17.b\n"
-      "zip2 z24.b, z23.b, z17.b\n"
-      "zip1 z17.b, z22.b, z16.b\n"
-      "zip2 z16.b, z22.b, z16.b\n"
-      "zip1 z23.b, z21.b, z20.b\n"
-      "zip2 z22.b, z21.b, z20.b\n"
-      "zip1 z21.b, z25.b, z19.b\n"
-      "zip2 z20.b, z25.b, z19.b\n"
       "zip1 z19.b, z18.b, z17.b\n"
       "zip2 z18.b, z18.b, z17.b\n"
+      "addvl x20, x20, #1\n"
       "zip1 z17.b, z24.b, z16.b\n"
       "zip2 z16.b, z24.b, z16.b\n"
-      "st1b { z23.b }, p1, [x27]\n"
-      "st1b { z22.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z21.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z20.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z18.b }, p1, [x27, #5, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #6, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #7, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "st1b { z23.b }, p1, [x22]\n"
+      "st1b { z22.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z21.b }, p1, [x22, #2, MUL VL]\n"
+      "st1b { z20.b }, p1, [x22, #3, MUL VL]\n"
+      "st1b { z19.b }, p1, [x22, #4, MUL VL]\n"
+      "st1b { z18.b }, p1, [x22, #5, MUL VL]\n"
+      "st1b { z17.b }, p1, [x22, #6, MUL VL]\n"
+      "st1b { z16.b }, p1, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x8\n"
@@ -192,90 +192,90 @@ void sve_transpose_interleave_4VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x10, %x[in]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cntb x20, ALL, MUL #2\n"
+      "add x27, x28, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x10, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "add x24, x25, %x[in_stride]\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "csel x25, x25, %x[pad_row], GE\n"
+      "add %x[in], x27, %x[in_stride]\n"
+      "csel x27, x27, %x[pad_row], GT\n"
+      "csel x28, x28, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x26, x26, %x[pad_row], GT\n"
+      "csel x9, x9, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1b { z20.b }, p1/Z, [x10]\n"
-      "ld1b { z24.b }, p1/Z, [x26]\n"
+      "ld1b { z21.b }, p1/Z, [x10]\n"
+      "ld1b { z19.b }, p1/Z, [x9]\n"
       "sub x21, x21, x20\n"
-      "ld1b { z19.b }, p1/Z, [x25]\n"
-      "ld1b { z18.b }, p1/Z, [x24]\n"
       "cmp x21, x20\n"
-      "ld1b { z23.b }, p1/Z, [x10, #1, MUL VL]\n"
-      "ld1b { z25.b }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z17.b }, p1/Z, [x28]\n"
+      "ld1b { z16.b }, p1/Z, [x27]\n"
+      "zip1 z26.b, z21.b, z17.b\n"
+      "zip1 z25.b, z19.b, z16.b\n"
+      "ld1b { z20.b }, p1/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z18.b }, p1/Z, [x9, #1, MUL VL]\n"
+      "zip2 z24.b, z21.b, z17.b\n"
+      "zip2 z19.b, z19.b, z16.b\n"
+      "ld1b { z17.b }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z16.b }, p1/Z, [x27, #1, MUL VL]\n"
+      "zip1 z23.b, z20.b, z17.b\n"
+      "zip1 z22.b, z18.b, z16.b\n"
+      "zip2 z21.b, z20.b, z17.b\n"
+      "zip2 z20.b, z18.b, z16.b\n"
       "addvl x10, x10, #2\n"
-      "addvl x26, x26, #2\n"
-      "ld1b { z22.b }, p1/Z, [x25, #1, MUL VL]\n"
-      "ld1b { z17.b }, p1/Z, [x24, #1, MUL VL]\n"
-      "addvl x25, x25, #2\n"
-      "addvl x24, x24, #2\n"
-      "zip1 z21.b, z20.b, z19.b\n"
-      "zip1 z16.b, z24.b, z18.b\n"
-      "zip2 z20.b, z20.b, z19.b\n"
-      "zip2 z19.b, z24.b, z18.b\n"
-      "zip1 z24.b, z23.b, z22.b\n"
-      "zip1 z18.b, z25.b, z17.b\n"
-      "zip2 z23.b, z23.b, z22.b\n"
-      "zip2 z22.b, z25.b, z17.b\n"
-      "zip1 z17.b, z21.b, z16.b\n"
-      "zip2 z16.b, z21.b, z16.b\n"
-      "zip1 z21.b, z20.b, z19.b\n"
-      "zip2 z20.b, z20.b, z19.b\n"
-      "zip1 z19.b, z24.b, z18.b\n"
-      "zip2 z18.b, z24.b, z18.b\n"
-      "st1b { z17.b }, p1, [x27]\n"
-      "st1b { z16.b }, p1, [x27, #1, MUL VL]\n"
-      "zip1 z17.b, z23.b, z22.b\n"
-      "zip2 z16.b, z23.b, z22.b\n"
-      "st1b { z21.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z20.b }, p1, [x27, #3, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z19.b }, p1, [x27]\n"
-      "st1b { z18.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #3, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "addvl x9, x9, #2\n"
+      "zip1 z16.b, z26.b, z25.b\n"
+      "zip2 z18.b, z26.b, z25.b\n"
+      "st1b { z16.b }, p1, [x22]\n"
+      "addvl x28, x28, #2\n"
+      "zip1 z17.b, z24.b, z19.b\n"
+      "zip2 z16.b, z24.b, z19.b\n"
+      "st1b { z18.b }, p1, [x22, #1, MUL VL]\n"
+      "addvl x27, x27, #2\n"
+      "st1b { z17.b }, p1, [x22, #2, MUL VL]\n"
+      "zip1 z19.b, z23.b, z22.b\n"
+      "zip2 z18.b, z23.b, z22.b\n"
+      "st1b { z16.b }, p1, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z17.b, z21.b, z20.b\n"
+      "zip2 z16.b, z21.b, z20.b\n"
+      "st1b { z19.b }, p1, [x22]\n"
+      "st1b { z18.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z17.b }, p1, [x22, #2, MUL VL]\n"
+      "st1b { z16.b }, p1, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "whilelt p0.b, XZR, x21\n"
-      "decw x21, ALL, MUL #4\n"
       "ld1b { z20.b }, p0/Z, [x10]\n"
-      "addvl x10, x10, #1\n"
-      "ld1b { z19.b }, p0/Z, [x26]\n"
-      "addvl x26, x26, #1\n"
-      "ld1b { z17.b }, p0/Z, [x25]\n"
-      "addvl x25, x25, #1\n"
-      "ld1b { z16.b }, p0/Z, [x24]\n"
-      "addvl x24, x24, #1\n"
+      "ld1b { z21.b }, p0/Z, [x9]\n"
+      "decw x21, ALL, MUL #4\n"
+      "ld1b { z19.b }, p0/Z, [x28]\n"
+      "ld1b { z16.b }, p0/Z, [x27]\n"
+      "zip1 z18.b, z20.b, z19.b\n"
+      "zip1 z17.b, z21.b, z16.b\n"
+      "zip2 z20.b, z20.b, z19.b\n"
+      "zip2 z16.b, z21.b, z16.b\n"
       "cmp x21, #0x0\n"
-      "zip1 z18.b, z20.b, z17.b\n"
-      "zip2 z20.b, z20.b, z17.b\n"
-      "zip1 z17.b, z19.b, z16.b\n"
-      "zip2 z16.b, z19.b, z16.b\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
+      "addvl x28, x28, #1\n"
       "zip1 z19.b, z18.b, z17.b\n"
       "zip2 z18.b, z18.b, z17.b\n"
+      "addvl x27, x27, #1\n"
       "zip1 z17.b, z20.b, z16.b\n"
       "zip2 z16.b, z20.b, z16.b\n"
-      "st1b { z19.b }, p1, [x27]\n"
-      "st1b { z18.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #3, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "st1b { z19.b }, p1, [x22]\n"
+      "st1b { z18.b }, p1, [x22, #1, MUL VL]\n"
+      "st1b { z17.b }, p1, [x22, #2, MUL VL]\n"
+      "st1b { z16.b }, p1, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL_2x2.hpp
index 56d1c9accc..b753c22b67 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL_2x2.hpp
@@ -44,126 +44,126 @@ void sve_transpose_interleave_4VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x12, %x[in]\n"
-      "mov x11, %x[width]\n"
-      "cnth x10, ALL, MUL #4\n"
-      "mov x9, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x28, x12, %x[in_stride]\n"
-      "add x27, x28, %x[in_stride]\n"
-      "add x26, x27, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "cmp x11, x10\n"
+      "add x11, x12, %x[in_stride]\n"
+      "add x10, x11, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "mov x27, %x[width]\n"
+      "cnth x26, ALL, MUL #4\n"
+      "add x25, x28, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "cmp x27, x26\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1h { z22.h }, p2/Z, [x12]\n"
-      "ld1h { z28.h }, p2/Z, [x12, #1, MUL VL]\n"
-      "mov x21, x9\n"
-      "add x9, x9, %x[out_stride]\n"
-      "ld1h { z26.h }, p2/Z, [x28]\n"
-      "ld1h { z17.h }, p2/Z, [x28, #1, MUL VL]\n"
-      "mov x20, x9\n"
-      "sub x11, x11, x10\n"
-      "ld1h { z13.h }, p2/Z, [x27]\n"
-      "ld1h { z8.h }, p2/Z, [x27, #1, MUL VL]\n"
-      "cmp x11, x10\n"
-      "add x9, x9, %x[out_stride]\n"
-      "ld1h { z30.h }, p2/Z, [x26]\n"
-      "ld1h { z31.h }, p2/Z, [x26, #1, MUL VL]\n"
-      "ld1h { z15.h }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z0.h }, p2/Z, [x12, #3, MUL VL]\n"
-      "zip1 z27.h, z22.h, z26.h\n"
-      "zip2 z26.h, z22.h, z26.h\n"
-      "ld1h { z4.h }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x28, #3, MUL VL]\n"
-      "zip1 z24.h, z28.h, z17.h\n"
-      "zip2 z19.h, z28.h, z17.h\n"
-      "ld1h { z25.h }, p2/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z21.h }, p2/Z, [x27, #3, MUL VL]\n"
-      "zip1 z11.h, z13.h, z30.h\n"
-      "zip2 z20.h, z13.h, z30.h\n"
-      "ld1h { z18.h }, p2/Z, [x26, #2, MUL VL]\n"
-      "ld1h { z1.h }, p2/Z, [x26, #3, MUL VL]\n"
-      "zip1 z12.h, z8.h, z31.h\n"
-      "zip2 z14.h, z8.h, z31.h\n"
-      "ld1h { z23.h }, p2/Z, [x25]\n"
-      "ld1h { z28.h }, p2/Z, [x25, #1, MUL VL]\n"
-      "zip1 z13.h, z15.h, z4.h\n"
-      "zip2 z15.h, z15.h, z4.h\n"
-      "ld1h { z2.h }, p2/Z, [x25, #2, MUL VL]\n"
-      "ld1h { z9.h }, p2/Z, [x25, #3, MUL VL]\n"
-      "zip1 z8.h, z0.h, z16.h\n"
-      "zip2 z10.h, z0.h, z16.h\n"
+      "ld1h { z18.h }, p2/Z, [x12]\n"
+      "ld1h { z20.h }, p2/Z, [x12, #1, MUL VL]\n"
+      "mov x21, x22\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1h { z17.h }, p2/Z, [x11]\n"
+      "ld1h { z16.h }, p2/Z, [x11, #1, MUL VL]\n"
+      "zip1 z25.h, z18.h, z17.h\n"
+      "zip2 z24.h, z18.h, z17.h\n"
+      "ld1h { z19.h }, p2/Z, [x10]\n"
+      "ld1h { z18.h }, p2/Z, [x10, #1, MUL VL]\n"
+      "zip1 z23.h, z20.h, z16.h\n"
+      "zip2 z15.h, z20.h, z16.h\n"
+      "ld1h { z17.h }, p2/Z, [x9]\n"
+      "ld1h { z16.h }, p2/Z, [x9, #1, MUL VL]\n"
+      "zip1 z14.h, z19.h, z17.h\n"
+      "zip2 z13.h, z19.h, z17.h\n"
+      "ld1h { z17.h }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z19.h }, p2/Z, [x12, #3, MUL VL]\n"
+      "zip1 z12.h, z18.h, z16.h\n"
+      "zip2 z11.h, z18.h, z16.h\n"
+      "ld1h { z16.h }, p2/Z, [x11, #2, MUL VL]\n"
+      "ld1h { z18.h }, p2/Z, [x11, #3, MUL VL]\n"
+      "mov x20, x22\n"
+      "zip1 z10.h, z17.h, z16.h\n"
+      "ld1h { z21.h }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z20.h }, p2/Z, [x10, #3, MUL VL]\n"
+      "zip2 z9.h, z17.h, z16.h\n"
+      "zip1 z8.h, z19.h, z18.h\n"
+      "ld1h { z17.h }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x9, #3, MUL VL]\n"
+      "zip2 z7.h, z19.h, z18.h\n"
+      "zip1 z6.h, z21.h, z17.h\n"
+      "ld1h { z19.h }, p2/Z, [x28]\n"
+      "ld1h { z18.h }, p2/Z, [x28, #1, MUL VL]\n"
+      "zip2 z5.h, z21.h, z17.h\n"
+      "zip1 z4.h, z20.h, z16.h\n"
+      "ld1h { z22.h }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z3.h }, p2/Z, [x28, #3, MUL VL]\n"
+      "zip2 z2.h, z20.h, z16.h\n"
+      "sub x27, x27, x26\n"
+      "ld1h { z17.h }, p2/Z, [x25]\n"
+      "ld1h { z16.h }, p2/Z, [x25, #1, MUL VL]\n"
+      "zip1 z1.h, z19.h, z17.h\n"
+      "zip2 z0.h, z19.h, z17.h\n"
+      "ld1h { z21.h }, p2/Z, [x25, #2, MUL VL]\n"
+      "ld1h { z20.h }, p2/Z, [x25, #3, MUL VL]\n"
+      "zip1 z31.h, z18.h, z16.h\n"
+      "zip2 z30.h, z18.h, z16.h\n"
       "ld1h { z17.h }, p2/Z, [x24]\n"
-      "ld1h { z16.h }, p2/Z, [x24, #1, MUL VL]\n"
-      "zip1 z6.h, z25.h, z18.h\n"
-      "zip2 z5.h, z25.h, z18.h\n"
-      "ld1h { z4.h }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1h { z7.h }, p2/Z, [x24, #3, MUL VL]\n"
-      "zip1 z30.h, z21.h, z1.h\n"
-      "zip2 z3.h, z21.h, z1.h\n"
-      "ld1h { z22.h }, p2/Z, [x23]\n"
-      "ld1h { z21.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "ld1h { z19.h }, p2/Z, [x24, #1, MUL VL]\n"
+      "cmp x27, x26\n"
       "addvl x12, x12, #4\n"
+      "ld1h { z29.h }, p2/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z28.h }, p2/Z, [x24, #3, MUL VL]\n"
+      "addvl x11, x11, #4\n"
+      "addvl x10, x10, #4\n"
+      "ld1h { z16.h }, p2/Z, [x23]\n"
+      "ld1h { z18.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "zip1 z27.h, z17.h, z16.h\n"
+      "zip2 z26.h, z17.h, z16.h\n"
+      "ld1h { z17.h }, p2/Z, [x23, #2, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x23, #3, MUL VL]\n"
+      "st1h { z25.h }, p2, [x21]\n"
+      "zip1 z25.h, z19.h, z18.h\n"
+      "st1h { z24.h }, p2, [x21, #1, MUL VL]\n"
+      "zip2 z24.h, z19.h, z18.h\n"
+      "addvl x9, x9, #4\n"
       "addvl x28, x28, #4\n"
-      "ld1h { z18.h }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z31.h }, p2/Z, [x23, #3, MUL VL]\n"
-      "zip1 z1.h, z23.h, z17.h\n"
-      "zip2 z0.h, z23.h, z17.h\n"
-      "ld1h { z25.h }, p2/Z, [x22]\n"
-      "ld1h { z23.h }, p2/Z, [x22, #1, MUL VL]\n"
-      "zip1 z29.h, z28.h, z16.h\n"
-      "zip2 z28.h, z28.h, z16.h\n"
-      "ld1h { z17.h }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x22, #3, MUL VL]\n"
-      "st1h { z27.h }, p2, [x21]\n"
-      "addvl x27, x27, #4\n"
-      "st1h { z26.h }, p2, [x21, #1, MUL VL]\n"
-      "addvl x26, x26, #4\n"
+      "st1h { z23.h }, p2, [x21, #2, MUL VL]\n"
       "addvl x25, x25, #4\n"
-      "zip1 z27.h, z2.h, z4.h\n"
-      "st1h { z24.h }, p2, [x21, #2, MUL VL]\n"
-      "zip1 z26.h, z22.h, z25.h\n"
-      "zip2 z25.h, z22.h, z25.h\n"
       "addvl x24, x24, #4\n"
-      "st1h { z19.h }, p2, [x21, #3, MUL VL]\n"
-      "zip1 z24.h, z21.h, z23.h\n"
-      "zip2 z23.h, z21.h, z23.h\n"
+      "zip1 z23.h, z22.h, z21.h\n"
+      "st1h { z15.h }, p2, [x21, #3, MUL VL]\n"
       "addvl x23, x23, #4\n"
-      "st1h { z11.h }, p2, [x21, #4, MUL VL]\n"
-      "addvl x22, x22, #4\n"
-      "zip2 z22.h, z2.h, z4.h\n"
-      "zip1 z21.h, z9.h, z7.h\n"
-      "st1h { z20.h }, p2, [x21, #5, MUL VL]\n"
-      "zip2 z20.h, z9.h, z7.h\n"
-      "zip1 z19.h, z18.h, z17.h\n"
+      "zip2 z22.h, z22.h, z21.h\n"
+      "zip1 z21.h, z3.h, z20.h\n"
+      "st1h { z14.h }, p2, [x21, #4, MUL VL]\n"
+      "zip2 z20.h, z3.h, z20.h\n"
+      "zip1 z19.h, z29.h, z17.h\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z13.h }, p2, [x21, #5, MUL VL]\n"
+      "zip2 z18.h, z29.h, z17.h\n"
+      "zip1 z17.h, z28.h, z16.h\n"
       "st1h { z12.h }, p2, [x21, #6, MUL VL]\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "zip1 z17.h, z31.h, z16.h\n"
-      "st1h { z14.h }, p2, [x21, #7, MUL VL]\n"
+      "zip2 z16.h, z28.h, z16.h\n"
+      "st1h { z11.h }, p2, [x21, #7, MUL VL]\n"
       "addvl x21, x21, #16\n"
-      "zip2 z16.h, z31.h, z16.h\n"
       "st1h { z1.h }, p2, [x21, #-8, MUL VL]\n"
       "st1h { z0.h }, p2, [x21, #-7, MUL VL]\n"
-      "st1h { z29.h }, p2, [x21, #-6, MUL VL]\n"
-      "st1h { z28.h }, p2, [x21, #-5, MUL VL]\n"
-      "st1h { z26.h }, p2, [x21, #-4, MUL VL]\n"
-      "st1h { z25.h }, p2, [x21, #-3, MUL VL]\n"
-      "st1h { z24.h }, p2, [x21, #-2, MUL VL]\n"
-      "st1h { z23.h }, p2, [x21, #-1, MUL VL]\n"
-      "st1h { z13.h }, p2, [x20]\n"
-      "st1h { z15.h }, p2, [x20, #1, MUL VL]\n"
+      "st1h { z31.h }, p2, [x21, #-6, MUL VL]\n"
+      "st1h { z30.h }, p2, [x21, #-5, MUL VL]\n"
+      "st1h { z27.h }, p2, [x21, #-4, MUL VL]\n"
+      "st1h { z26.h }, p2, [x21, #-3, MUL VL]\n"
+      "st1h { z25.h }, p2, [x21, #-2, MUL VL]\n"
+      "st1h { z24.h }, p2, [x21, #-1, MUL VL]\n"
+      "st1h { z10.h }, p2, [x20]\n"
+      "st1h { z9.h }, p2, [x20, #1, MUL VL]\n"
       "st1h { z8.h }, p2, [x20, #2, MUL VL]\n"
-      "st1h { z10.h }, p2, [x20, #3, MUL VL]\n"
+      "st1h { z7.h }, p2, [x20, #3, MUL VL]\n"
       "st1h { z6.h }, p2, [x20, #4, MUL VL]\n"
       "st1h { z5.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z30.h }, p2, [x20, #6, MUL VL]\n"
-      "st1h { z3.h }, p2, [x20, #7, MUL VL]\n"
+      "st1h { z4.h }, p2, [x20, #6, MUL VL]\n"
+      "st1h { z2.h }, p2, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
-      "st1h { z27.h }, p2, [x20, #-8, MUL VL]\n"
+      "st1h { z23.h }, p2, [x20, #-8, MUL VL]\n"
       "st1h { z22.h }, p2, [x20, #-7, MUL VL]\n"
       "st1h { z21.h }, p2, [x20, #-6, MUL VL]\n"
       "st1h { z20.h }, p2, [x20, #-5, MUL VL]\n"
@@ -173,63 +173,63 @@ void sve_transpose_interleave_4VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x11, 5f\n"
+      "cbz x27, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x21, x11\n"
-      "mov x20, x9\n"
-      "decw x11, ALL, MUL #4\n"
-      "add x9, x9, %x[out_stride]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p0.h, XZR, x21\n"
-      "cmp x11, #0x0\n"
-      "ld1h { z21.h }, p1/Z, [x12]\n"
-      "ld1h { z20.h }, p1/Z, [x28]\n"
-      "ld1h { z19.h }, p1/Z, [x27]\n"
-      "ld1h { z18.h }, p1/Z, [x26]\n"
-      "ld1h { z1.h }, p1/Z, [x25]\n"
-      "ld1h { z0.h }, p1/Z, [x24]\n"
-      "ld1h { z17.h }, p0/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x28, #1, MUL VL]\n"
-      "zip1 z23.h, z21.h, z20.h\n"
-      "zip2 z31.h, z21.h, z20.h\n"
-      "ld1h { z22.h }, p0/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z21.h }, p0/Z, [x26, #1, MUL VL]\n"
-      "zip1 z30.h, z19.h, z18.h\n"
-      "zip2 z29.h, z19.h, z18.h\n"
-      "ld1h { z28.h }, p0/Z, [x25, #1, MUL VL]\n"
-      "ld1h { z20.h }, p0/Z, [x24, #1, MUL VL]\n"
+      "mov x20, x27\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z22.h }, p1/Z, [x12]\n"
+      "ld1h { z21.h }, p1/Z, [x11]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z20.h }, p0/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z19.h }, p0/Z, [x11, #1, MUL VL]\n"
+      "ld1h { z18.h }, p1/Z, [x10]\n"
+      "ld1h { z24.h }, p0/Z, [x10, #1, MUL VL]\n"
+      "mov x20, x22\n"
+      "decw x27, ALL, MUL #4\n"
+      "ld1h { z17.h }, p1/Z, [x9]\n"
+      "ld1h { z16.h }, p0/Z, [x9, #1, MUL VL]\n"
+      "zip1 z31.h, z22.h, z21.h\n"
+      "zip2 z23.h, z22.h, z21.h\n"
+      "ld1h { z30.h }, p1/Z, [x28]\n"
+      "ld1h { z29.h }, p0/Z, [x28, #1, MUL VL]\n"
+      "zip1 z22.h, z20.h, z19.h\n"
+      "zip2 z28.h, z20.h, z19.h\n"
+      "ld1h { z21.h }, p1/Z, [x25]\n"
+      "ld1h { z27.h }, p0/Z, [x25, #1, MUL VL]\n"
+      "zip1 z20.h, z18.h, z17.h\n"
+      "zip2 z19.h, z18.h, z17.h\n"
+      "ld1h { z18.h }, p1/Z, [x24]\n"
+      "ld1h { z26.h }, p0/Z, [x24, #1, MUL VL]\n"
+      "zip1 z25.h, z24.h, z16.h\n"
+      "zip2 z24.h, z24.h, z16.h\n"
+      "ld1h { z17.h }, p1/Z, [x23]\n"
+      "ld1h { z16.h }, p0/Z, [x23, #1, MUL VL]\n"
+      "st1h { z31.h }, p2, [x20]\n"
+      "cmp x27, #0x0\n"
+      "st1h { z23.h }, p2, [x20, #1, MUL VL]\n"
       "addvl x12, x12, #2\n"
+      "addvl x11, x11, #2\n"
+      "zip1 z23.h, z30.h, z21.h\n"
+      "st1h { z22.h }, p2, [x20, #2, MUL VL]\n"
+      "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      "zip2 z22.h, z30.h, z21.h\n"
+      "st1h { z28.h }, p2, [x20, #3, MUL VL]\n"
       "addvl x28, x28, #2\n"
-      "ld1h { z27.h }, p1/Z, [x23]\n"
-      "ld1h { z26.h }, p0/Z, [x23, #1, MUL VL]\n"
-      "zip1 z19.h, z17.h, z16.h\n"
-      "zip2 z18.h, z17.h, z16.h\n"
-      "ld1h { z17.h }, p1/Z, [x22]\n"
-      "ld1h { z16.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "zip1 z25.h, z22.h, z21.h\n"
-      "zip2 z24.h, z22.h, z21.h\n"
-      "st1h { z23.h }, p2, [x20]\n"
-      "addvl x27, x27, #2\n"
-      "addvl x26, x26, #2\n"
-      "zip1 z23.h, z1.h, z0.h\n"
-      "st1h { z31.h }, p2, [x20, #1, MUL VL]\n"
       "addvl x25, x25, #2\n"
+      "zip1 z21.h, z29.h, z27.h\n"
+      "st1h { z20.h }, p2, [x20, #4, MUL VL]\n"
       "addvl x24, x24, #2\n"
-      "zip2 z22.h, z1.h, z0.h\n"
-      "st1h { z19.h }, p2, [x20, #2, MUL VL]\n"
       "addvl x23, x23, #2\n"
-      "addvl x22, x22, #2\n"
-      "zip1 z21.h, z28.h, z20.h\n"
-      "st1h { z18.h }, p2, [x20, #3, MUL VL]\n"
-      "zip2 z20.h, z28.h, z20.h\n"
-      "zip1 z19.h, z27.h, z17.h\n"
-      "st1h { z30.h }, p2, [x20, #4, MUL VL]\n"
-      "zip2 z18.h, z27.h, z17.h\n"
+      "zip2 z20.h, z29.h, z27.h\n"
+      "st1h { z19.h }, p2, [x20, #5, MUL VL]\n"
+      "zip1 z19.h, z18.h, z17.h\n"
+      "zip2 z18.h, z18.h, z17.h\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z25.h }, p2, [x20, #6, MUL VL]\n"
       "zip1 z17.h, z26.h, z16.h\n"
-      "st1h { z29.h }, p2, [x20, #5, MUL VL]\n"
       "zip2 z16.h, z26.h, z16.h\n"
-      "st1h { z25.h }, p2, [x20, #6, MUL VL]\n"
       "st1h { z24.h }, p2, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
       "st1h { z23.h }, p2, [x20, #-8, MUL VL]\n"
@@ -251,70 +251,70 @@ void sve_transpose_interleave_4VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "mov x12, %x[in]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #4\n"
+      "add x11, x12, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "mov x9, %x[out]\n"
-      "sub %x[height], %x[height], #0x2\n"
-      "add x28, x12, %x[in_stride]\n"
-      "add %x[in], x28, %x[in_stride]\n"
-      "csel x28, x28, %x[pad_row], GT\n"
+      "add %x[in], x11, %x[in_stride]\n"
+      "csel x11, x11, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x2\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1h { z21.h }, p2/Z, [x12]\n"
+      "ld1h { z18.h }, p2/Z, [x12]\n"
       "ld1h { z20.h }, p2/Z, [x12, #1, MUL VL]\n"
       "sub x21, x21, x20\n"
-      "ld1h { z16.h }, p2/Z, [x28]\n"
-      "ld1h { z19.h }, p2/Z, [x28, #1, MUL VL]\n"
       "cmp x21, x20\n"
-      "ld1h { z24.h }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z23.h }, p2/Z, [x12, #3, MUL VL]\n"
+      "ld1h { z17.h }, p2/Z, [x11]\n"
+      "ld1h { z16.h }, p2/Z, [x11, #1, MUL VL]\n"
+      "zip1 z23.h, z18.h, z17.h\n"
+      "zip2 z19.h, z18.h, z17.h\n"
+      "ld1h { z18.h }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z22.h }, p2/Z, [x12, #3, MUL VL]\n"
+      "zip1 z21.h, z20.h, z16.h\n"
+      "zip2 z20.h, z20.h, z16.h\n"
+      "ld1h { z17.h }, p2/Z, [x11, #2, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x11, #3, MUL VL]\n"
+      "st1h { z23.h }, p2, [x22]\n"
       "addvl x12, x12, #4\n"
-      "ld1h { z18.h }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z22.h }, p2/Z, [x28, #3, MUL VL]\n"
-      "addvl x28, x28, #4\n"
-      "zip1 z17.h, z21.h, z16.h\n"
-      "zip2 z21.h, z21.h, z16.h\n"
-      "zip1 z16.h, z20.h, z19.h\n"
-      "zip2 z20.h, z20.h, z19.h\n"
-      "st1h { z17.h }, p2, [x9]\n"
-      "zip1 z19.h, z24.h, z18.h\n"
-      "zip2 z18.h, z24.h, z18.h\n"
-      "zip1 z17.h, z23.h, z22.h\n"
-      "st1h { z21.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #2, MUL VL]\n"
-      "zip2 z16.h, z23.h, z22.h\n"
-      "st1h { z20.h }, p2, [x9, #3, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
-      "st1h { z19.h }, p2, [x9]\n"
-      "st1h { z18.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z17.h }, p2, [x9, #2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #3, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
+      "st1h { z19.h }, p2, [x22, #1, MUL VL]\n"
+      "addvl x11, x11, #4\n"
+      "zip1 z19.h, z18.h, z17.h\n"
+      "zip2 z18.h, z18.h, z17.h\n"
+      "st1h { z21.h }, p2, [x22, #2, MUL VL]\n"
+      "zip1 z17.h, z22.h, z16.h\n"
+      "zip2 z16.h, z22.h, z16.h\n"
+      "st1h { z20.h }, p2, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z19.h }, p2, [x22]\n"
+      "st1h { z18.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decw x21, ALL, MUL #4\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z18.h }, p0/Z, [x12]\n"
+      "ld1h { z17.h }, p0/Z, [x11]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
-      "cmp x21, #0x0\n"
-      "ld1h { z18.h }, p1/Z, [x12]\n"
-      "ld1h { z17.h }, p1/Z, [x28]\n"
       "ld1h { z20.h }, p0/Z, [x12, #1, MUL VL]\n"
-      "addvl x12, x12, #2\n"
-      "ld1h { z16.h }, p0/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #2\n"
+      "ld1h { z16.h }, p0/Z, [x11, #1, MUL VL]\n"
+      "decw x21, ALL, MUL #4\n"
+      "cmp x21, #0x0\n"
       "zip1 z19.h, z18.h, z17.h\n"
       "zip2 z18.h, z18.h, z17.h\n"
+      "addvl x12, x12, #2\n"
+      "addvl x11, x11, #2\n"
       "zip1 z17.h, z20.h, z16.h\n"
       "zip2 z16.h, z20.h, z16.h\n"
-      "st1h { z19.h }, p2, [x9]\n"
-      "st1h { z18.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z17.h }, p2, [x9, #2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #3, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
+      "st1h { z19.h }, p2, [x22]\n"
+      "st1h { z18.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #3, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_1x8.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_1x8.hpp
index 69eccad912..bba8f107d8 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_1x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_1x8.hpp
@@ -42,216 +42,216 @@ void sve_transpose_interleave_6VL_1x8(uint8_t *out, const uint8_t *in, size_t wi
       "ptrue p1.b\n"
       "1:"  // Main row loop: Head
       "mov x10, %x[in]\n"
-      "cmp %x[height], #0x7\n"
-      "mov x9, %x[width]\n"
-      "cntb x28, ALL, MUL #3\n"
-      "mov x27, %x[out]\n"
-      "add x26, x10, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
-      "add %x[in], x20, %x[in_stride]\n"
-      "csel x20, x20, %x[pad_row], GT\n"
-      "csel x21, x21, %x[pad_row], GE\n"
+      "cmp %x[height], #0x7\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "csel x23, x23, %x[pad_row], GT\n"
+      "csel x24, x24, %x[pad_row], GE\n"
       "cmp %x[height], #0x5\n"
-      "csel x22, x22, %x[pad_row], GT\n"
-      "csel x23, x23, %x[pad_row], GE\n"
+      "mov x22, %x[width]\n"
+      "cntb x21, ALL, MUL #3\n"
+      "csel x25, x25, %x[pad_row], GT\n"
+      "csel x26, x26, %x[pad_row], GE\n"
       "cmp %x[height], #0x3\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "csel x25, x25, %x[pad_row], GE\n"
+      "csel x27, x27, %x[pad_row], GT\n"
+      "csel x28, x28, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x26, x26, %x[pad_row], GT\n"
-      "cmp x9, x28\n"
+      "csel x9, x9, %x[pad_row], GT\n"
+      "cmp x22, x21\n"
+      "mov x20, %x[out]\n"
       "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1b { z1.b }, p1/Z, [x10]\n"
-      "ld1b { z0.b }, p1/Z, [x26]\n"
-      "sub x9, x9, x28\n"
-      "ld1b { z31.b }, p1/Z, [x25]\n"
-      "ld1b { z28.b }, p1/Z, [x24]\n"
-      "cmp x9, x28\n"
-      "ld1b { z27.b }, p1/Z, [x23]\n"
-      "ld1b { z26.b }, p1/Z, [x22]\n"
-      "ld1b { z25.b }, p1/Z, [x21]\n"
-      "ld1b { z24.b }, p1/Z, [x20]\n"
+      "ld1b { z21.b }, p1/Z, [x10]\n"
+      "ld1b { z25.b }, p1/Z, [x9]\n"
+      "sub x22, x22, x21\n"
+      "cmp x22, x21\n"
+      "ld1b { z20.b }, p1/Z, [x28]\n"
+      "ld1b { z24.b }, p1/Z, [x27]\n"
+      "ld1b { z19.b }, p1/Z, [x26]\n"
+      "ld1b { z18.b }, p1/Z, [x25]\n"
+      "zip1 z7.b, z21.b, z19.b\n"
+      "zip1 z6.b, z25.b, z18.b\n"
+      "ld1b { z17.b }, p1/Z, [x24]\n"
+      "ld1b { z16.b }, p1/Z, [x23]\n"
+      "zip1 z28.b, z20.b, z17.b\n"
+      "zip1 z27.b, z24.b, z16.b\n"
       "ld1b { z23.b }, p1/Z, [x10, #1, MUL VL]\n"
-      "ld1b { z30.b }, p1/Z, [x26, #1, MUL VL]\n"
-      "ld1b { z22.b }, p1/Z, [x25, #1, MUL VL]\n"
-      "ld1b { z21.b }, p1/Z, [x24, #1, MUL VL]\n"
-      "zip1 z15.b, z1.b, z27.b\n"
-      "zip1 z9.b, z0.b, z26.b\n"
-      "ld1b { z20.b }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1b { z19.b }, p1/Z, [x22, #1, MUL VL]\n"
-      "zip1 z18.b, z31.b, z25.b\n"
-      "zip1 z29.b, z28.b, z24.b\n"
-      "ld1b { z17.b }, p1/Z, [x21, #1, MUL VL]\n"
-      "ld1b { z16.b }, p1/Z, [x20, #1, MUL VL]\n"
-      "zip2 z14.b, z1.b, z27.b\n"
-      "zip2 z13.b, z31.b, z25.b\n"
-      "ld1b { z8.b }, p1/Z, [x10, #2, MUL VL]\n"
-      "ld1b { z7.b }, p1/Z, [x26, #2, MUL VL]\n"
-      "zip2 z12.b, z0.b, z26.b\n"
-      "zip2 z6.b, z28.b, z24.b\n"
-      "ld1b { z5.b }, p1/Z, [x25, #2, MUL VL]\n"
-      "ld1b { z4.b }, p1/Z, [x24, #2, MUL VL]\n"
-      "zip1 z3.b, z23.b, z20.b\n"
-      "zip1 z11.b, z30.b, z19.b\n"
-      "ld1b { z28.b }, p1/Z, [x23, #2, MUL VL]\n"
-      "ld1b { z27.b }, p1/Z, [x22, #2, MUL VL]\n"
-      "zip1 z2.b, z22.b, z17.b\n"
-      "zip1 z1.b, z21.b, z16.b\n"
-      "ld1b { z26.b }, p1/Z, [x21, #2, MUL VL]\n"
-      "ld1b { z25.b }, p1/Z, [x20, #2, MUL VL]\n"
-      "zip2 z24.b, z23.b, z20.b\n"
-      "zip2 z23.b, z22.b, z17.b\n"
-      "zip2 z22.b, z30.b, z19.b\n"
-      "zip2 z21.b, z21.b, z16.b\n"
+      "ld1b { z22.b }, p1/Z, [x9, #1, MUL VL]\n"
+      "zip2 z5.b, z21.b, z19.b\n"
+      "zip2 z4.b, z20.b, z17.b\n"
+      "ld1b { z21.b }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z20.b }, p1/Z, [x27, #1, MUL VL]\n"
+      "zip2 z3.b, z25.b, z18.b\n"
+      "zip2 z2.b, z24.b, z16.b\n"
+      "ld1b { z19.b }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z18.b }, p1/Z, [x25, #1, MUL VL]\n"
+      "zip1 z1.b, z23.b, z19.b\n"
+      "zip1 z15.b, z22.b, z18.b\n"
+      "ld1b { z17.b }, p1/Z, [x24, #1, MUL VL]\n"
+      "ld1b { z16.b }, p1/Z, [x23, #1, MUL VL]\n"
+      "zip1 z0.b, z21.b, z17.b\n"
+      "zip1 z31.b, z20.b, z16.b\n"
+      "ld1b { z26.b }, p1/Z, [x10, #2, MUL VL]\n"
+      "ld1b { z30.b }, p1/Z, [x9, #2, MUL VL]\n"
+      "zip2 z14.b, z23.b, z19.b\n"
+      "zip2 z13.b, z21.b, z17.b\n"
+      "ld1b { z25.b }, p1/Z, [x28, #2, MUL VL]\n"
+      "ld1b { z24.b }, p1/Z, [x27, #2, MUL VL]\n"
+      "zip2 z12.b, z22.b, z18.b\n"
+      "zip2 z11.b, z20.b, z16.b\n"
+      "ld1b { z23.b }, p1/Z, [x26, #2, MUL VL]\n"
+      "ld1b { z22.b }, p1/Z, [x25, #2, MUL VL]\n"
+      "zip1 z10.b, z26.b, z23.b\n"
+      "zip1 z9.b, z30.b, z22.b\n"
+      "ld1b { z21.b }, p1/Z, [x24, #2, MUL VL]\n"
+      "ld1b { z17.b }, p1/Z, [x23, #2, MUL VL]\n"
+      "zip1 z29.b, z25.b, z21.b\n"
+      "zip1 z8.b, z24.b, z17.b\n"
+      "zip1 z19.b, z7.b, z28.b\n"
+      "zip1 z16.b, z6.b, z27.b\n"
       "addvl x10, x10, #3\n"
+      "addvl x9, x9, #3\n"
+      "zip2 z28.b, z7.b, z28.b\n"
+      "zip2 z18.b, z6.b, z27.b\n"
+      "addvl x28, x28, #3\n"
+      "addvl x27, x27, #3\n"
+      "zip1 z27.b, z5.b, z4.b\n"
+      "zip1 z20.b, z3.b, z2.b\n"
       "addvl x26, x26, #3\n"
-      "zip1 z0.b, z8.b, z28.b\n"
-      "zip1 z10.b, z7.b, z27.b\n"
       "addvl x25, x25, #3\n"
+      "zip2 z7.b, z26.b, z23.b\n"
+      "zip2 z26.b, z25.b, z21.b\n"
       "addvl x24, x24, #3\n"
-      "zip1 z31.b, z5.b, z26.b\n"
-      "zip1 z30.b, z4.b, z25.b\n"
       "addvl x23, x23, #3\n"
-      "addvl x22, x22, #3\n"
-      "zip1 z20.b, z15.b, z18.b\n"
-      "zip1 z19.b, z9.b, z29.b\n"
-      "addvl x21, x21, #3\n"
-      "addvl x20, x20, #3\n"
-      "zip2 z18.b, z15.b, z18.b\n"
-      "zip2 z16.b, z9.b, z29.b\n"
-      "zip1 z29.b, z14.b, z13.b\n"
-      "zip1 z17.b, z12.b, z6.b\n"
-      "zip2 z9.b, z8.b, z28.b\n"
-      "zip2 z28.b, z5.b, z26.b\n"
-      "zip2 z8.b, z7.b, z27.b\n"
-      "zip2 z27.b, z4.b, z25.b\n"
-      "zip2 z7.b, z14.b, z13.b\n"
-      "zip2 z6.b, z12.b, z6.b\n"
-      "zip1 z5.b, z3.b, z2.b\n"
-      "zip1 z4.b, z11.b, z1.b\n"
-      "zip2 z3.b, z3.b, z2.b\n"
-      "zip2 z2.b, z11.b, z1.b\n"
-      "zip1 z26.b, z24.b, z23.b\n"
-      "zip1 z25.b, z22.b, z21.b\n"
-      "zip2 z24.b, z24.b, z23.b\n"
-      "zip2 z23.b, z22.b, z21.b\n"
-      "zip1 z1.b, z0.b, z31.b\n"
-      "zip1 z22.b, z10.b, z30.b\n"
-      "zip1 z21.b, z20.b, z19.b\n"
-      "zip2 z20.b, z20.b, z19.b\n"
-      "zip1 z19.b, z18.b, z16.b\n"
-      "zip2 z18.b, z18.b, z16.b\n"
-      "zip1 z16.b, z29.b, z17.b\n"
-      "zip2 z17.b, z29.b, z17.b\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "zip2 z0.b, z0.b, z31.b\n"
-      "zip2 z31.b, z10.b, z30.b\n"
-      "st1b { z19.b }, p1, [x27, #2, MUL VL]\n"
-      "zip1 z30.b, z9.b, z28.b\n"
-      "zip1 z29.b, z8.b, z27.b\n"
-      "st1b { z18.b }, p1, [x27, #3, MUL VL]\n"
-      "zip2 z28.b, z9.b, z28.b\n"
-      "zip2 z27.b, z8.b, z27.b\n"
-      "st1b { z16.b }, p1, [x27, #4, MUL VL]\n"
-      "zip1 z21.b, z7.b, z6.b\n"
-      "zip2 z16.b, z7.b, z6.b\n"
-      "st1b { z17.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z20.b, z5.b, z4.b\n"
-      "zip2 z19.b, z5.b, z4.b\n"
+      "zip2 z6.b, z30.b, z22.b\n"
+      "zip2 z25.b, z24.b, z17.b\n"
+      "zip2 z5.b, z5.b, z4.b\n"
+      "zip2 z4.b, z3.b, z2.b\n"
+      "zip1 z3.b, z1.b, z0.b\n"
+      "zip1 z2.b, z15.b, z31.b\n"
+      "zip2 z1.b, z1.b, z0.b\n"
+      "zip2 z0.b, z15.b, z31.b\n"
+      "zip1 z31.b, z14.b, z13.b\n"
+      "zip1 z30.b, z12.b, z11.b\n"
+      "zip2 z24.b, z14.b, z13.b\n"
+      "zip2 z23.b, z12.b, z11.b\n"
+      "zip1 z22.b, z10.b, z29.b\n"
+      "zip1 z21.b, z9.b, z8.b\n"
+      "zip1 z17.b, z19.b, z16.b\n"
+      "zip2 z16.b, z19.b, z16.b\n"
+      "st1b { z17.b }, p1, [x20]\n"
+      "zip1 z19.b, z28.b, z18.b\n"
+      "zip2 z18.b, z28.b, z18.b\n"
+      "st1b { z16.b }, p1, [x20, #1, MUL VL]\n"
+      "zip1 z17.b, z27.b, z20.b\n"
+      "zip2 z16.b, z27.b, z20.b\n"
+      "st1b { z19.b }, p1, [x20, #2, MUL VL]\n"
+      "st1b { z18.b }, p1, [x20, #3, MUL VL]\n"
+      "zip2 z29.b, z10.b, z29.b\n"
+      "zip2 z20.b, z9.b, z8.b\n"
+      "st1b { z17.b }, p1, [x20, #4, MUL VL]\n"
+      "zip1 z28.b, z7.b, z26.b\n"
+      "zip1 z27.b, z6.b, z25.b\n"
+      "st1b { z16.b }, p1, [x20, #5, MUL VL]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "zip2 z26.b, z7.b, z26.b\n"
+      "zip2 z25.b, z6.b, z25.b\n"
+      "zip1 z17.b, z5.b, z4.b\n"
+      "zip2 z16.b, z5.b, z4.b\n"
+      "st1b { z17.b }, p1, [x20]\n"
       "zip1 z18.b, z3.b, z2.b\n"
       "zip2 z17.b, z3.b, z2.b\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z16.b }, p1, [x27, #1, MUL VL]\n"
-      "zip1 z16.b, z26.b, z25.b\n"
-      "zip2 z26.b, z26.b, z25.b\n"
-      "st1b { z20.b }, p1, [x27, #2, MUL VL]\n"
-      "zip1 z25.b, z24.b, z23.b\n"
+      "st1b { z16.b }, p1, [x20, #1, MUL VL]\n"
+      "zip1 z16.b, z1.b, z0.b\n"
+      "zip2 z19.b, z1.b, z0.b\n"
+      "st1b { z18.b }, p1, [x20, #2, MUL VL]\n"
+      "st1b { z17.b }, p1, [x20, #3, MUL VL]\n"
+      "zip1 z18.b, z31.b, z30.b\n"
+      "zip2 z17.b, z31.b, z30.b\n"
+      "st1b { z16.b }, p1, [x20, #4, MUL VL]\n"
+      "zip1 z16.b, z24.b, z23.b\n"
       "zip2 z24.b, z24.b, z23.b\n"
-      "st1b { z19.b }, p1, [x27, #3, MUL VL]\n"
-      "zip1 z23.b, z1.b, z22.b\n"
-      "zip2 z22.b, z1.b, z22.b\n"
-      "st1b { z18.b }, p1, [x27, #4, MUL VL]\n"
-      "zip1 z21.b, z0.b, z31.b\n"
-      "zip2 z20.b, z0.b, z31.b\n"
-      "st1b { z17.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "zip1 z19.b, z30.b, z29.b\n"
-      "zip2 z18.b, z30.b, z29.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "zip1 z17.b, z28.b, z27.b\n"
-      "zip2 z16.b, z28.b, z27.b\n"
-      "st1b { z26.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z25.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z24.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z23.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z22.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z18.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "st1b { z19.b }, p1, [x20, #5, MUL VL]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "zip1 z23.b, z22.b, z21.b\n"
+      "zip2 z22.b, z22.b, z21.b\n"
+      "st1b { z18.b }, p1, [x20]\n"
+      "zip1 z21.b, z29.b, z20.b\n"
+      "zip2 z20.b, z29.b, z20.b\n"
+      "st1b { z17.b }, p1, [x20, #1, MUL VL]\n"
+      "zip1 z19.b, z28.b, z27.b\n"
+      "zip2 z18.b, z28.b, z27.b\n"
+      "st1b { z16.b }, p1, [x20, #2, MUL VL]\n"
+      "zip1 z17.b, z26.b, z25.b\n"
+      "zip2 z16.b, z26.b, z25.b\n"
+      "st1b { z24.b }, p1, [x20, #3, MUL VL]\n"
+      "st1b { z23.b }, p1, [x20, #4, MUL VL]\n"
+      "st1b { z22.b }, p1, [x20, #5, MUL VL]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "st1b { z21.b }, p1, [x20]\n"
+      "st1b { z20.b }, p1, [x20, #1, MUL VL]\n"
+      "st1b { z19.b }, p1, [x20, #2, MUL VL]\n"
+      "st1b { z18.b }, p1, [x20, #3, MUL VL]\n"
+      "st1b { z17.b }, p1, [x20, #4, MUL VL]\n"
+      "st1b { z16.b }, p1, [x20, #5, MUL VL]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x9, 5f\n"
+      "cbz x22, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.b, XZR, x9\n"
-      "decd x9, ALL, MUL #6\n"
-      "ld1b { z21.b }, p0/Z, [x10]\n"
+      "whilelt p0.b, XZR, x22\n"
+      "ld1b { z23.b }, p0/Z, [x10]\n"
+      "ld1b { z27.b }, p0/Z, [x9]\n"
+      "decd x22, ALL, MUL #6\n"
+      "ld1b { z21.b }, p0/Z, [x28]\n"
+      "ld1b { z26.b }, p0/Z, [x27]\n"
+      "cmp x22, #0x0\n"
       "incd x10, ALL, MUL #6\n"
-      "ld1b { z26.b }, p0/Z, [x26]\n"
+      "ld1b { z20.b }, p0/Z, [x26]\n"
+      "ld1b { z19.b }, p0/Z, [x25]\n"
+      "zip1 z25.b, z23.b, z20.b\n"
+      "zip1 z24.b, z27.b, z19.b\n"
+      "ld1b { z17.b }, p0/Z, [x24]\n"
+      "ld1b { z16.b }, p0/Z, [x23]\n"
+      "zip1 z22.b, z21.b, z17.b\n"
+      "zip1 z18.b, z26.b, z16.b\n"
+      "zip2 z23.b, z23.b, z20.b\n"
+      "zip2 z21.b, z21.b, z17.b\n"
+      "incd x9, ALL, MUL #6\n"
+      "incd x28, ALL, MUL #6\n"
+      "zip2 z20.b, z27.b, z19.b\n"
+      "zip2 z17.b, z26.b, z16.b\n"
+      "incd x27, ALL, MUL #6\n"
       "incd x26, ALL, MUL #6\n"
-      "ld1b { z20.b }, p0/Z, [x25]\n"
+      "zip1 z19.b, z25.b, z22.b\n"
+      "zip1 z16.b, z24.b, z18.b\n"
       "incd x25, ALL, MUL #6\n"
-      "ld1b { z25.b }, p0/Z, [x24]\n"
       "incd x24, ALL, MUL #6\n"
-      "ld1b { z18.b }, p0/Z, [x23]\n"
-      "ld1b { z19.b }, p0/Z, [x22]\n"
-      "ld1b { z17.b }, p0/Z, [x21]\n"
-      "cmp x9, #0x0\n"
+      "zip2 z22.b, z25.b, z22.b\n"
+      "zip2 z18.b, z24.b, z18.b\n"
       "incd x23, ALL, MUL #6\n"
-      "ld1b { z16.b }, p0/Z, [x20]\n"
-      "incd x22, ALL, MUL #6\n"
-      "incd x21, ALL, MUL #6\n"
-      "zip1 z23.b, z21.b, z18.b\n"
-      "zip2 z24.b, z21.b, z18.b\n"
-      "incd x20, ALL, MUL #6\n"
-      "zip1 z22.b, z20.b, z17.b\n"
-      "zip1 z21.b, z26.b, z19.b\n"
-      "zip2 z18.b, z20.b, z17.b\n"
-      "zip1 z17.b, z25.b, z16.b\n"
-      "zip2 z20.b, z26.b, z19.b\n"
-      "zip2 z16.b, z25.b, z16.b\n"
-      "zip1 z19.b, z23.b, z22.b\n"
-      "zip2 z23.b, z23.b, z22.b\n"
-      "zip1 z22.b, z24.b, z18.b\n"
-      "zip1 z18.b, z21.b, z17.b\n"
-      "zip2 z17.b, z21.b, z17.b\n"
-      "zip1 z16.b, z20.b, z16.b\n"
-      "zip1 z21.b, z19.b, z18.b\n"
-      "zip2 z20.b, z19.b, z18.b\n"
-      "zip1 z19.b, z23.b, z17.b\n"
-      "zip2 z18.b, z23.b, z17.b\n"
-      "zip1 z17.b, z22.b, z16.b\n"
-      "zip2 z16.b, z22.b, z16.b\n"
-      "st1b { z21.b }, p1, [x27]\n"
-      "st1b { z20.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z18.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #5, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "zip1 z21.b, z23.b, z21.b\n"
+      "zip1 z20.b, z20.b, z17.b\n"
+      "zip1 z17.b, z19.b, z16.b\n"
+      "zip2 z16.b, z19.b, z16.b\n"
+      "st1b { z17.b }, p1, [x20]\n"
+      "zip1 z19.b, z22.b, z18.b\n"
+      "zip2 z18.b, z22.b, z18.b\n"
+      "st1b { z16.b }, p1, [x20, #1, MUL VL]\n"
+      "zip1 z17.b, z21.b, z20.b\n"
+      "zip2 z16.b, z21.b, z20.b\n"
+      "st1b { z19.b }, p1, [x20, #2, MUL VL]\n"
+      "st1b { z18.b }, p1, [x20, #3, MUL VL]\n"
+      "st1b { z17.b }, p1, [x20, #4, MUL VL]\n"
+      "st1b { z16.b }, p1, [x20, #5, MUL VL]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_2x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_2x4.hpp
index 6c0a5c029b..ebfc65be34 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_2x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_2x4.hpp
@@ -44,125 +44,125 @@ void sve_transpose_interleave_6VL_2x4(uint16_t *out, const uint16_t *in, size_t
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x12, %x[in]\n"
-      "mov x11, %x[width]\n"
-      "cnth x10, ALL, MUL #3\n"
-      "mov x9, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x28, x12, %x[in_stride]\n"
-      "add x27, x28, %x[in_stride]\n"
-      "add x26, x27, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "cmp x11, x10\n"
+      "add x11, x12, %x[in_stride]\n"
+      "add x10, x11, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "mov x27, %x[width]\n"
+      "cnth x26, ALL, MUL #3\n"
+      "add x25, x28, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "cmp x27, x26\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1h { z19.h }, p2/Z, [x12]\n"
+      "ld1h { z18.h }, p2/Z, [x12]\n"
       "ld1h { z13.h }, p2/Z, [x12, #1, MUL VL]\n"
-      "mov x21, x9\n"
-      "add x9, x9, %x[out_stride]\n"
-      "ld1h { z18.h }, p2/Z, [x28]\n"
-      "ld1h { z12.h }, p2/Z, [x28, #1, MUL VL]\n"
-      "mov x20, x9\n"
-      "sub x11, x11, x10\n"
-      "ld1h { z16.h }, p2/Z, [x27]\n"
-      "ld1h { z11.h }, p2/Z, [x27, #1, MUL VL]\n"
-      "cmp x11, x10\n"
-      "add x9, x9, %x[out_stride]\n"
-      "ld1h { z17.h }, p2/Z, [x26]\n"
-      "ld1h { z10.h }, p2/Z, [x26, #1, MUL VL]\n"
-      "ld1h { z9.h }, p2/Z, [x25]\n"
-      "ld1h { z8.h }, p2/Z, [x24]\n"
-      "ld1h { z28.h }, p2/Z, [x23]\n"
-      "ld1h { z26.h }, p2/Z, [x22]\n"
-      "zip1 z24.h, z19.h, z16.h\n"
-      "zip2 z23.h, z19.h, z16.h\n"
-      "ld1h { z7.h }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z6.h }, p2/Z, [x28, #2, MUL VL]\n"
-      "zip1 z16.h, z18.h, z17.h\n"
-      "zip2 z22.h, z18.h, z17.h\n"
-      "ld1h { z5.h }, p2/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z4.h }, p2/Z, [x26, #2, MUL VL]\n"
+      "mov x21, x22\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1h { z17.h }, p2/Z, [x11]\n"
+      "ld1h { z12.h }, p2/Z, [x11, #1, MUL VL]\n"
+      "mov x20, x22\n"
+      "sub x27, x27, x26\n"
+      "ld1h { z16.h }, p2/Z, [x10]\n"
+      "ld1h { z11.h }, p2/Z, [x10, #1, MUL VL]\n"
+      "zip1 z23.h, z18.h, z16.h\n"
+      "zip2 z29.h, z18.h, z16.h\n"
+      "ld1h { z16.h }, p2/Z, [x9]\n"
+      "ld1h { z10.h }, p2/Z, [x9, #1, MUL VL]\n"
+      "zip1 z22.h, z17.h, z16.h\n"
+      "zip2 z28.h, z17.h, z16.h\n"
+      "ld1h { z27.h }, p2/Z, [x28]\n"
+      "ld1h { z26.h }, p2/Z, [x25]\n"
       "zip1 z21.h, z13.h, z11.h\n"
       "zip1 z20.h, z12.h, z10.h\n"
-      "ld1h { z3.h }, p2/Z, [x25, #1, MUL VL]\n"
-      "ld1h { z2.h }, p2/Z, [x24, #1, MUL VL]\n"
-      "zip1 z19.h, z9.h, z28.h\n"
-      "zip1 z18.h, z8.h, z26.h\n"
-      "ld1h { z27.h }, p2/Z, [x23, #1, MUL VL]\n"
-      "ld1h { z1.h }, p2/Z, [x22, #1, MUL VL]\n"
-      "zip1 z17.h, z24.h, z16.h\n"
-      "zip2 z16.h, z24.h, z16.h\n"
-      "ld1h { z0.h }, p2/Z, [x25, #2, MUL VL]\n"
-      "ld1h { z31.h }, p2/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z18.h }, p2/Z, [x24]\n"
+      "ld1h { z19.h }, p2/Z, [x23]\n"
+      "zip1 z17.h, z27.h, z18.h\n"
+      "zip1 z16.h, z26.h, z19.h\n"
+      "ld1h { z9.h }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z8.h }, p2/Z, [x11, #2, MUL VL]\n"
       "zip1 z25.h, z23.h, z22.h\n"
       "zip2 z24.h, z23.h, z22.h\n"
-      "ld1h { z30.h }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z29.h }, p2/Z, [x22, #2, MUL VL]\n"
-      "zip1 z23.h, z21.h, z20.h\n"
-      "zip2 z22.h, z21.h, z20.h\n"
-      "zip1 z21.h, z19.h, z18.h\n"
-      "zip2 z20.h, z19.h, z18.h\n"
-      "st1h { z17.h }, p2, [x21]\n"
+      "ld1h { z23.h }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z7.h }, p2/Z, [x9, #2, MUL VL]\n"
+      "zip1 z22.h, z29.h, z28.h\n"
+      "zip2 z6.h, z29.h, z28.h\n"
+      "ld1h { z28.h }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1h { z5.h }, p2/Z, [x25, #1, MUL VL]\n"
+      "zip1 z4.h, z21.h, z20.h\n"
+      "zip2 z3.h, z21.h, z20.h\n"
+      "ld1h { z21.h }, p2/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z20.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "zip1 z2.h, z17.h, z16.h\n"
+      "zip2 z1.h, z17.h, z16.h\n"
+      "ld1h { z0.h }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z31.h }, p2/Z, [x25, #2, MUL VL]\n"
+      "zip2 z18.h, z27.h, z18.h\n"
+      "zip2 z17.h, z26.h, z19.h\n"
+      "ld1h { z30.h }, p2/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z29.h }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip1 z19.h, z28.h, z21.h\n"
+      "zip1 z16.h, z5.h, z20.h\n"
+      "st1h { z25.h }, p2, [x21]\n"
+      "zip2 z27.h, z13.h, z11.h\n"
+      "zip2 z26.h, z12.h, z10.h\n"
+      "cmp x27, x26\n"
+      "st1h { z24.h }, p2, [x21, #1, MUL VL]\n"
+      "zip1 z25.h, z9.h, z23.h\n"
+      "zip1 z24.h, z8.h, z7.h\n"
       "addvl x12, x12, #3\n"
-      "zip2 z19.h, z9.h, z28.h\n"
-      "zip2 z18.h, z8.h, z26.h\n"
-      "st1h { z16.h }, p2, [x21, #1, MUL VL]\n"
+      "st1h { z22.h }, p2, [x21, #2, MUL VL]\n"
+      "zip2 z23.h, z9.h, z23.h\n"
+      "zip2 z22.h, z8.h, z7.h\n"
+      "addvl x11, x11, #3\n"
+      "st1h { z6.h }, p2, [x21, #3, MUL VL]\n"
+      "zip2 z28.h, z28.h, z21.h\n"
+      "zip2 z21.h, z5.h, z20.h\n"
+      "addvl x10, x10, #3\n"
+      "st1h { z4.h }, p2, [x21, #4, MUL VL]\n"
+      "zip1 z20.h, z18.h, z17.h\n"
+      "zip2 z18.h, z18.h, z17.h\n"
+      "addvl x9, x9, #3\n"
+      "st1h { z3.h }, p2, [x21, #5, MUL VL]\n"
+      "zip1 z17.h, z19.h, z16.h\n"
+      "zip2 z16.h, z19.h, z16.h\n"
       "addvl x28, x28, #3\n"
-      "zip1 z17.h, z3.h, z27.h\n"
-      "zip1 z16.h, z2.h, z1.h\n"
-      "st1h { z25.h }, p2, [x21, #2, MUL VL]\n"
-      "addvl x27, x27, #3\n"
-      "st1h { z24.h }, p2, [x21, #3, MUL VL]\n"
-      "zip2 z26.h, z13.h, z11.h\n"
-      "zip2 z25.h, z12.h, z10.h\n"
-      "addvl x26, x26, #3\n"
-      "st1h { z23.h }, p2, [x21, #4, MUL VL]\n"
-      "zip1 z28.h, z7.h, z5.h\n"
-      "zip1 z24.h, z6.h, z4.h\n"
+      "st1h { z2.h }, p2, [x21, #6, MUL VL]\n"
+      "zip1 z19.h, z27.h, z26.h\n"
+      "zip2 z27.h, z27.h, z26.h\n"
       "addvl x25, x25, #3\n"
-      "st1h { z22.h }, p2, [x21, #5, MUL VL]\n"
-      "zip2 z23.h, z7.h, z5.h\n"
-      "zip2 z22.h, z6.h, z4.h\n"
-      "addvl x24, x24, #3\n"
-      "st1h { z21.h }, p2, [x21, #6, MUL VL]\n"
-      "zip2 z27.h, z3.h, z27.h\n"
-      "zip2 z21.h, z2.h, z1.h\n"
-      "addvl x23, x23, #3\n"
-      "st1h { z20.h }, p2, [x21, #7, MUL VL]\n"
+      "st1h { z1.h }, p2, [x21, #7, MUL VL]\n"
       "addvl x21, x21, #12\n"
-      "zip1 z20.h, z19.h, z18.h\n"
-      "zip2 z19.h, z19.h, z18.h\n"
-      "zip1 z18.h, z17.h, z16.h\n"
-      "zip2 z17.h, z17.h, z16.h\n"
-      "addvl x22, x22, #3\n"
-      "zip1 z16.h, z26.h, z25.h\n"
-      "zip2 z26.h, z26.h, z25.h\n"
+      "zip1 z26.h, z25.h, z24.h\n"
+      "zip2 z25.h, z25.h, z24.h\n"
       "st1h { z20.h }, p2, [x21, #-4, MUL VL]\n"
-      "st1h { z19.h }, p2, [x21, #-3, MUL VL]\n"
-      "zip1 z25.h, z28.h, z24.h\n"
-      "zip2 z19.h, z28.h, z24.h\n"
-      "st1h { z18.h }, p2, [x21, #-2, MUL VL]\n"
       "zip1 z24.h, z23.h, z22.h\n"
       "zip2 z23.h, z23.h, z22.h\n"
-      "st1h { z17.h }, p2, [x21, #-1, MUL VL]\n"
-      "zip1 z22.h, z27.h, z21.h\n"
-      "zip2 z21.h, z27.h, z21.h\n"
+      "addvl x24, x24, #3\n"
+      "st1h { z18.h }, p2, [x21, #-3, MUL VL]\n"
+      "zip1 z22.h, z28.h, z21.h\n"
+      "zip2 z21.h, z28.h, z21.h\n"
+      "addvl x23, x23, #3\n"
+      "st1h { z17.h }, p2, [x21, #-2, MUL VL]\n"
       "zip1 z18.h, z0.h, z30.h\n"
       "zip1 z17.h, z31.h, z29.h\n"
-      "st1h { z16.h }, p2, [x20]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z16.h }, p2, [x21, #-1, MUL VL]\n"
       "zip2 z20.h, z0.h, z30.h\n"
       "zip2 z16.h, z31.h, z29.h\n"
-      "st1h { z26.h }, p2, [x20, #1, MUL VL]\n"
-      "st1h { z25.h }, p2, [x20, #2, MUL VL]\n"
-      "st1h { z19.h }, p2, [x20, #3, MUL VL]\n"
+      "st1h { z19.h }, p2, [x20]\n"
       "zip1 z19.h, z18.h, z17.h\n"
       "zip2 z18.h, z18.h, z17.h\n"
-      "st1h { z24.h }, p2, [x20, #4, MUL VL]\n"
+      "st1h { z27.h }, p2, [x20, #1, MUL VL]\n"
       "zip1 z17.h, z20.h, z16.h\n"
       "zip2 z16.h, z20.h, z16.h\n"
+      "st1h { z26.h }, p2, [x20, #2, MUL VL]\n"
+      "st1h { z25.h }, p2, [x20, #3, MUL VL]\n"
+      "st1h { z24.h }, p2, [x20, #4, MUL VL]\n"
       "st1h { z23.h }, p2, [x20, #5, MUL VL]\n"
       "st1h { z22.h }, p2, [x20, #6, MUL VL]\n"
       "st1h { z21.h }, p2, [x20, #7, MUL VL]\n"
@@ -173,79 +173,79 @@ void sve_transpose_interleave_6VL_2x4(uint16_t *out, const uint16_t *in, size_t
       "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x11, 5f\n"
+      "cbz x27, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x21, x11\n"
-      "mov x20, x9\n"
-      "decd x11, ALL, MUL #6\n"
-      "add x9, x9, %x[out_stride]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p0.h, XZR, x21\n"
-      "cmp x11, #0x0\n"
-      "ld1h { z23.h }, p1/Z, [x12]\n"
-      "ld1h { z25.h }, p1/Z, [x28]\n"
-      "ld1h { z22.h }, p1/Z, [x27]\n"
-      "ld1h { z19.h }, p1/Z, [x26]\n"
-      "ld1h { z1.h }, p1/Z, [x25]\n"
-      "ld1h { z0.h }, p1/Z, [x24]\n"
-      "ld1h { z21.h }, p0/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z20.h }, p0/Z, [x28, #1, MUL VL]\n"
+      "mov x20, x27\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z17.h }, p1/Z, [x12]\n"
+      "ld1h { z19.h }, p1/Z, [x11]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z22.h }, p0/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z21.h }, p0/Z, [x11, #1, MUL VL]\n"
+      "ld1h { z16.h }, p1/Z, [x10]\n"
+      "ld1h { z20.h }, p0/Z, [x10, #1, MUL VL]\n"
+      "zip1 z25.h, z17.h, z16.h\n"
+      "zip2 z24.h, z17.h, z16.h\n"
+      "ld1h { z18.h }, p1/Z, [x9]\n"
+      "ld1h { z17.h }, p0/Z, [x9, #1, MUL VL]\n"
+      "zip1 z16.h, z19.h, z18.h\n"
+      "zip2 z19.h, z19.h, z18.h\n"
+      "ld1h { z0.h }, p1/Z, [x28]\n"
+      "ld1h { z31.h }, p1/Z, [x25]\n"
+      "zip1 z23.h, z22.h, z20.h\n"
+      "zip1 z22.h, z21.h, z17.h\n"
+      "ld1h { z30.h }, p1/Z, [x24]\n"
+      "ld1h { z29.h }, p1/Z, [x23]\n"
+      "zip1 z21.h, z0.h, z30.h\n"
+      "zip1 z18.h, z31.h, z29.h\n"
+      "ld1h { z28.h }, p0/Z, [x28, #1, MUL VL]\n"
+      "ld1h { z27.h }, p0/Z, [x25, #1, MUL VL]\n"
+      "mov x20, x22\n"
+      "decd x27, ALL, MUL #6\n"
+      "ld1h { z20.h }, p0/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z26.h }, p0/Z, [x23, #1, MUL VL]\n"
       "addvl x12, x12, #1\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
+      "zip1 z17.h, z25.h, z16.h\n"
+      "zip2 z16.h, z25.h, z16.h\n"
       "addvl x28, x28, #1\n"
-      "ld1h { z18.h }, p0/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z17.h }, p0/Z, [x26, #1, MUL VL]\n"
-      "zip1 z24.h, z23.h, z22.h\n"
-      "zip1 z16.h, z25.h, z19.h\n"
-      "ld1h { z31.h }, p1/Z, [x23]\n"
-      "ld1h { z30.h }, p1/Z, [x22]\n"
-      "zip2 z23.h, z23.h, z22.h\n"
-      "zip2 z19.h, z25.h, z19.h\n"
-      "ld1h { z29.h }, p0/Z, [x25, #1, MUL VL]\n"
-      "ld1h { z28.h }, p0/Z, [x24, #1, MUL VL]\n"
-      "addvl x27, x27, #1\n"
-      "addvl x26, x26, #1\n"
-      "ld1h { z27.h }, p0/Z, [x23, #1, MUL VL]\n"
-      "ld1h { z26.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "zip1 z22.h, z21.h, z18.h\n"
-      "zip1 z21.h, z20.h, z17.h\n"
-      "zip1 z20.h, z1.h, z31.h\n"
-      "zip1 z18.h, z0.h, z30.h\n"
       "addvl x25, x25, #1\n"
+      "zip1 z25.h, z24.h, z19.h\n"
+      "zip2 z19.h, z24.h, z19.h\n"
       "addvl x24, x24, #1\n"
       "addvl x23, x23, #1\n"
-      "addvl x22, x22, #1\n"
-      "zip1 z17.h, z24.h, z16.h\n"
-      "zip2 z16.h, z24.h, z16.h\n"
-      "zip1 z25.h, z23.h, z19.h\n"
-      "zip2 z24.h, z23.h, z19.h\n"
-      "incd x12, ALL, MUL #4\n"
-      "incd x28, ALL, MUL #4\n"
-      "zip1 z19.h, z22.h, z21.h\n"
-      "zip2 z23.h, z22.h, z21.h\n"
-      "incd x27, ALL, MUL #4\n"
-      "incd x26, ALL, MUL #4\n"
-      "zip1 z22.h, z20.h, z18.h\n"
-      "zip2 z21.h, z20.h, z18.h\n"
+      "zip1 z24.h, z23.h, z22.h\n"
+      "zip2 z23.h, z23.h, z22.h\n"
+      "zip1 z22.h, z21.h, z18.h\n"
+      "zip2 z21.h, z21.h, z18.h\n"
       "st1h { z17.h }, p2, [x20]\n"
-      "incd x25, ALL, MUL #4\n"
-      "zip2 z18.h, z1.h, z31.h\n"
-      "zip2 z17.h, z0.h, z30.h\n"
+      "cmp x27, #0x0\n"
+      "zip2 z18.h, z0.h, z30.h\n"
+      "zip2 z17.h, z31.h, z29.h\n"
       "st1h { z16.h }, p2, [x20, #1, MUL VL]\n"
-      "incd x24, ALL, MUL #4\n"
-      "zip1 z20.h, z29.h, z27.h\n"
-      "zip1 z16.h, z28.h, z26.h\n"
+      "incd x12, ALL, MUL #4\n"
+      "zip1 z20.h, z28.h, z20.h\n"
+      "zip1 z16.h, z27.h, z26.h\n"
       "st1h { z25.h }, p2, [x20, #2, MUL VL]\n"
-      "incd x23, ALL, MUL #4\n"
-      "st1h { z24.h }, p2, [x20, #3, MUL VL]\n"
-      "incd x22, ALL, MUL #4\n"
-      "st1h { z19.h }, p2, [x20, #4, MUL VL]\n"
+      "incd x11, ALL, MUL #4\n"
+      "st1h { z19.h }, p2, [x20, #3, MUL VL]\n"
+      "incd x10, ALL, MUL #4\n"
+      "incd x9, ALL, MUL #4\n"
       "zip1 z19.h, z18.h, z17.h\n"
+      "st1h { z24.h }, p2, [x20, #4, MUL VL]\n"
+      "incd x28, ALL, MUL #4\n"
+      "incd x25, ALL, MUL #4\n"
       "zip2 z18.h, z18.h, z17.h\n"
       "st1h { z23.h }, p2, [x20, #5, MUL VL]\n"
+      "incd x24, ALL, MUL #4\n"
+      "incd x23, ALL, MUL #4\n"
       "zip1 z17.h, z20.h, z16.h\n"
-      "zip2 z16.h, z20.h, z16.h\n"
       "st1h { z22.h }, p2, [x20, #6, MUL VL]\n"
+      "zip2 z16.h, z20.h, z16.h\n"
+      "add x22, x22, %x[out_stride]\n"
       "st1h { z21.h }, p2, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
       "st1h { z19.h }, p2, [x20, #-4, MUL VL]\n"
@@ -261,123 +261,123 @@ void sve_transpose_interleave_6VL_2x4(uint16_t *out, const uint16_t *in, size_t
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x12, %x[in]\n"
+      "add x11, x12, %x[in_stride]\n"
+      "add x10, x11, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #3\n"
+      "add x9, x10, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x9, %x[out]\n"
-      "add x28, x12, %x[in_stride]\n"
-      "add x27, x28, %x[in_stride]\n"
-      "add x26, x27, %x[in_stride]\n"
-      "add %x[in], x26, %x[in_stride]\n"
-      "csel x26, x26, %x[pad_row], GT\n"
-      "csel x27, x27, %x[pad_row], GE\n"
+      "add %x[in], x9, %x[in_stride]\n"
+      "csel x9, x9, %x[pad_row], GT\n"
+      "csel x10, x10, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x28, x28, %x[pad_row], GT\n"
+      "csel x11, x11, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1h { z22.h }, p2/Z, [x12]\n"
-      "ld1h { z28.h }, p2/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z18.h }, p2/Z, [x12]\n"
+      "ld1h { z24.h }, p2/Z, [x12, #1, MUL VL]\n"
       "sub x21, x21, x20\n"
-      "ld1h { z21.h }, p2/Z, [x28]\n"
-      "ld1h { z27.h }, p2/Z, [x28, #1, MUL VL]\n"
       "cmp x21, x20\n"
-      "ld1h { z17.h }, p2/Z, [x27]\n"
-      "ld1h { z26.h }, p2/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x26]\n"
-      "ld1h { z20.h }, p2/Z, [x26, #1, MUL VL]\n"
-      "ld1h { z25.h }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z29.h }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z17.h }, p2/Z, [x11]\n"
+      "ld1h { z23.h }, p2/Z, [x11, #1, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x10]\n"
+      "ld1h { z22.h }, p2/Z, [x10, #1, MUL VL]\n"
+      "zip1 z31.h, z18.h, z16.h\n"
+      "zip2 z30.h, z18.h, z16.h\n"
+      "ld1h { z16.h }, p2/Z, [x9]\n"
+      "ld1h { z20.h }, p2/Z, [x9, #1, MUL VL]\n"
+      "zip1 z29.h, z17.h, z16.h\n"
+      "zip2 z28.h, z17.h, z16.h\n"
+      "ld1h { z19.h }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z18.h }, p2/Z, [x11, #2, MUL VL]\n"
+      "zip1 z27.h, z24.h, z22.h\n"
+      "zip1 z21.h, z23.h, z20.h\n"
+      "ld1h { z17.h }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x9, #2, MUL VL]\n"
+      "zip2 z26.h, z24.h, z22.h\n"
+      "zip2 z20.h, z23.h, z20.h\n"
+      "zip1 z25.h, z19.h, z17.h\n"
+      "zip1 z24.h, z18.h, z16.h\n"
       "addvl x12, x12, #3\n"
-      "addvl x28, x28, #3\n"
-      "ld1h { z24.h }, p2/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z19.h }, p2/Z, [x26, #2, MUL VL]\n"
-      "zip1 z18.h, z22.h, z17.h\n"
-      "zip2 z23.h, z22.h, z17.h\n"
-      "zip1 z17.h, z21.h, z16.h\n"
-      "zip2 z16.h, z21.h, z16.h\n"
-      "addvl x27, x27, #3\n"
-      "addvl x26, x26, #3\n"
-      "zip1 z22.h, z28.h, z26.h\n"
-      "zip1 z21.h, z27.h, z20.h\n"
-      "zip2 z28.h, z28.h, z26.h\n"
-      "zip2 z20.h, z27.h, z20.h\n"
-      "zip1 z27.h, z25.h, z24.h\n"
-      "zip1 z26.h, z29.h, z19.h\n"
-      "zip2 z25.h, z25.h, z24.h\n"
-      "zip2 z24.h, z29.h, z19.h\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "zip1 z17.h, z23.h, z16.h\n"
-      "zip2 z16.h, z23.h, z16.h\n"
-      "zip1 z23.h, z22.h, z21.h\n"
-      "zip2 z22.h, z22.h, z21.h\n"
-      "st1h { z19.h }, p2, [x9]\n"
-      "st1h { z18.h }, p2, [x9, #1, MUL VL]\n"
-      "zip1 z21.h, z28.h, z20.h\n"
-      "zip2 z20.h, z28.h, z20.h\n"
-      "st1h { z17.h }, p2, [x9, #2, MUL VL]\n"
-      "zip1 z19.h, z27.h, z26.h\n"
-      "zip2 z18.h, z27.h, z26.h\n"
-      "st1h { z16.h }, p2, [x9, #3, MUL VL]\n"
-      "zip1 z17.h, z25.h, z24.h\n"
-      "zip2 z16.h, z25.h, z24.h\n"
-      "st1h { z23.h }, p2, [x9, #4, MUL VL]\n"
-      "st1h { z22.h }, p2, [x9, #5, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
-      "st1h { z21.h }, p2, [x9]\n"
-      "st1h { z20.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z19.h }, p2, [x9, #2, MUL VL]\n"
-      "st1h { z18.h }, p2, [x9, #3, MUL VL]\n"
-      "st1h { z17.h }, p2, [x9, #4, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #5, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
+      "addvl x11, x11, #3\n"
+      "zip2 z23.h, z19.h, z17.h\n"
+      "zip2 z22.h, z18.h, z16.h\n"
+      "addvl x10, x10, #3\n"
+      "addvl x9, x9, #3\n"
+      "zip1 z17.h, z31.h, z29.h\n"
+      "zip2 z16.h, z31.h, z29.h\n"
+      "st1h { z17.h }, p2, [x22]\n"
+      "zip1 z19.h, z30.h, z28.h\n"
+      "zip2 z18.h, z30.h, z28.h\n"
+      "st1h { z16.h }, p2, [x22, #1, MUL VL]\n"
+      "zip1 z17.h, z27.h, z21.h\n"
+      "zip2 z16.h, z27.h, z21.h\n"
+      "st1h { z19.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22, #3, MUL VL]\n"
+      "zip1 z21.h, z26.h, z20.h\n"
+      "zip2 z20.h, z26.h, z20.h\n"
+      "st1h { z17.h }, p2, [x22, #4, MUL VL]\n"
+      "zip1 z19.h, z25.h, z24.h\n"
+      "zip2 z18.h, z25.h, z24.h\n"
+      "st1h { z16.h }, p2, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z17.h, z23.h, z22.h\n"
+      "zip2 z16.h, z23.h, z22.h\n"
+      "st1h { z21.h }, p2, [x22]\n"
+      "st1h { z20.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z19.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22, #3, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #4, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decd x21, ALL, MUL #6\n"
       "whilelt p1.h, XZR, x20\n"
+      "ld1h { z22.h }, p1/Z, [x12]\n"
+      "ld1h { z25.h }, p1/Z, [x11]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
-      "cmp x21, #0x0\n"
-      "ld1h { z25.h }, p1/Z, [x12]\n"
-      "ld1h { z24.h }, p1/Z, [x28]\n"
-      "ld1h { z18.h }, p1/Z, [x27]\n"
-      "ld1h { z17.h }, p1/Z, [x26]\n"
-      "ld1h { z22.h }, p0/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z24.h }, p0/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z23.h }, p0/Z, [x11, #1, MUL VL]\n"
+      "ld1h { z21.h }, p1/Z, [x10]\n"
+      "ld1h { z20.h }, p0/Z, [x10, #1, MUL VL]\n"
+      "decd x21, ALL, MUL #6\n"
       "addvl x12, x12, #1\n"
-      "ld1h { z23.h }, p0/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #1\n"
-      "ld1h { z21.h }, p0/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x26, #1, MUL VL]\n"
-      "addvl x27, x27, #1\n"
-      "addvl x26, x26, #1\n"
-      "zip1 z20.h, z25.h, z18.h\n"
-      "zip1 z19.h, z24.h, z17.h\n"
+      "ld1h { z18.h }, p1/Z, [x9]\n"
+      "ld1h { z17.h }, p0/Z, [x9, #1, MUL VL]\n"
+      "addvl x11, x11, #1\n"
+      "addvl x10, x10, #1\n"
+      "addvl x9, x9, #1\n"
+      "zip1 z19.h, z22.h, z21.h\n"
+      "zip1 z16.h, z25.h, z18.h\n"
+      "cmp x21, #0x0\n"
+      "zip2 z22.h, z22.h, z21.h\n"
       "zip2 z18.h, z25.h, z18.h\n"
-      "zip2 z17.h, z24.h, z17.h\n"
       "incd x12, ALL, MUL #4\n"
-      "incd x28, ALL, MUL #4\n"
-      "zip1 z22.h, z22.h, z21.h\n"
-      "zip1 z16.h, z23.h, z16.h\n"
-      "incd x27, ALL, MUL #4\n"
-      "incd x26, ALL, MUL #4\n"
-      "zip1 z21.h, z20.h, z19.h\n"
-      "zip2 z20.h, z20.h, z19.h\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "zip1 z17.h, z22.h, z16.h\n"
-      "zip2 z16.h, z22.h, z16.h\n"
-      "st1h { z21.h }, p2, [x9]\n"
-      "st1h { z20.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z19.h }, p2, [x9, #2, MUL VL]\n"
-      "st1h { z18.h }, p2, [x9, #3, MUL VL]\n"
-      "st1h { z17.h }, p2, [x9, #4, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #5, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
+      "incd x11, ALL, MUL #4\n"
+      "zip1 z21.h, z24.h, z20.h\n"
+      "zip1 z20.h, z23.h, z17.h\n"
+      "incd x10, ALL, MUL #4\n"
+      "incd x9, ALL, MUL #4\n"
+      "zip1 z17.h, z19.h, z16.h\n"
+      "zip2 z16.h, z19.h, z16.h\n"
+      "st1h { z17.h }, p2, [x22]\n"
+      "zip1 z19.h, z22.h, z18.h\n"
+      "zip2 z18.h, z22.h, z18.h\n"
+      "st1h { z16.h }, p2, [x22, #1, MUL VL]\n"
+      "zip1 z17.h, z21.h, z20.h\n"
+      "zip2 z16.h, z21.h, z20.h\n"
+      "st1h { z19.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22, #3, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #4, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_2x4_fp32bf16.hpp
index 8ff9551c19..f23b9011d0 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_2x4_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_2x4_fp32bf16.hpp
@@ -39,176 +39,176 @@ void sve_transpose_interleave_6VL_2x4_fp32bf16(bfloat16 *out, const float *in, s
     size_t out_stride = 6 * roundup<size_t>(height, 4) * get_vector_length<uint32_t>();
 
     __asm__ __volatile__(
-      "ptrue p2.b\n"
+      "ptrue p3.b\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
       "cnth x20, ALL, MUL #3\n"
+      "add x22, x24, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x24, %x[out]\n"
-      "add x23, x26, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add %x[in], x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GT\n"
-      "csel x22, x22, %x[pad_row], GE\n"
+      "add %x[in], x22, %x[in_stride]\n"
+      "csel x22, x22, %x[pad_row], GT\n"
+      "csel x24, x24, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x23, x23, %x[pad_row], GT\n"
-      "cmp x25, x20\n"
+      "csel x25, x25, %x[pad_row], GT\n"
+      "cmp x23, x20\n"
+      "mov x21, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1w { z17.s }, p2/Z, [x26]\n"
-      "ld1w { z24.s }, p2/Z, [x26, #1, MUL VL]\n"
-      "sub x25, x25, x20\n"
-      "ld1w { z23.s }, p2/Z, [x26, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x22]\n"
-      "cmp x25, x20\n"
-      "ld1w { z22.s }, p2/Z, [x22, #1, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x26, #3, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x26, #4, MUL VL]\n"
-      "ld1w { z5.s }, p2/Z, [x26, #5, MUL VL]\n"
-      "ld1w { z18.s }, p2/Z, [x22, #3, MUL VL]\n"
-      "zip1 z4.s, z17.s, z16.s\n"
-      "zip2 z3.s, z17.s, z16.s\n"
-      "ld1w { z17.s }, p2/Z, [x22, #4, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x22, #5, MUL VL]\n"
-      "zip1 z2.s, z24.s, z22.s\n"
-      "zip2 z1.s, z24.s, z22.s\n"
-      "ld1w { z0.s }, p2/Z, [x23]\n"
-      "ld1w { z31.s }, p2/Z, [x23, #1, MUL VL]\n"
-      "zip1 z30.s, z23.s, z21.s\n"
-      "zip2 z29.s, z23.s, z21.s\n"
-      "ld1w { z28.s }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z27.s }, p2/Z, [x21]\n"
-      "zip1 z26.s, z20.s, z18.s\n"
-      "zip2 z25.s, z20.s, z18.s\n"
-      "ld1w { z24.s }, p2/Z, [x21, #1, MUL VL]\n"
-      "ld1w { z23.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "zip1 z22.s, z19.s, z17.s\n"
-      "zip2 z10.s, z19.s, z17.s\n"
-      "ld1w { z21.s }, p2/Z, [x23, #3, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x23, #4, MUL VL]\n"
-      "zip1 z19.s, z5.s, z16.s\n"
-      "zip2 z9.s, z5.s, z16.s\n"
-      "ld1w { z8.s }, p2/Z, [x23, #5, MUL VL]\n"
-      "ld1w { z18.s }, p2/Z, [x21, #3, MUL VL]\n"
-      ".inst 0x658aa887  // bfcvt z7.h, p2/M, z4.s\n"
-      "zip1 z6.s, z0.s, z27.s\n"
-      "ld1w { z17.s }, p2/Z, [x21, #4, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x21, #5, MUL VL]\n"
-      ".inst 0x658aa865  // bfcvt z5.h, p2/M, z3.s\n"
-      "zip2 z4.s, z0.s, z27.s\n"
-      ".inst 0x658aa843  // bfcvt z3.h, p2/M, z2.s\n"
-      "zip1 z2.s, z31.s, z24.s\n"
+      "ld1w { z17.s }, p3/Z, [x26]\n"
+      "ld1w { z18.s }, p3/Z, [x26, #1, MUL VL]\n"
+      "sub x23, x23, x20\n"
+      "cmp x23, x20\n"
+      "ld1w { z19.s }, p3/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x24]\n"
+      "zip1 z21.s, z17.s, z16.s\n"
+      "zip2 z20.s, z17.s, z16.s\n"
+      "ld1w { z17.s }, p3/Z, [x24, #1, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x24, #2, MUL VL]\n"
+      "zip1 z29.s, z18.s, z17.s\n"
+      "zip2 z28.s, z18.s, z17.s\n"
+      "ld1w { z17.s }, p3/Z, [x26, #3, MUL VL]\n"
+      "ld1w { z18.s }, p3/Z, [x26, #4, MUL VL]\n"
+      "zip1 z27.s, z19.s, z16.s\n"
+      "zip2 z26.s, z19.s, z16.s\n"
+      "ld1w { z19.s }, p3/Z, [x26, #5, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x24, #3, MUL VL]\n"
+      "zip1 z25.s, z17.s, z16.s\n"
+      "zip2 z24.s, z17.s, z16.s\n"
+      "ld1w { z17.s }, p3/Z, [x24, #4, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x24, #5, MUL VL]\n"
+      "zip1 z12.s, z18.s, z17.s\n"
+      "zip2 z11.s, z18.s, z17.s\n"
+      "ld1w { z18.s }, p3/Z, [x25]\n"
+      "ld1w { z23.s }, p3/Z, [x25, #1, MUL VL]\n"
+      "zip1 z10.s, z19.s, z16.s\n"
+      "zip2 z9.s, z19.s, z16.s\n"
+      "ld1w { z22.s }, p3/Z, [x25, #2, MUL VL]\n"
+      "ld1w { z17.s }, p3/Z, [x22]\n"
+      ".inst 0x658aaea8  // bfcvt z8.h, p3/M, z21.s\n"
+      "zip1 z7.s, z18.s, z17.s\n"
+      "ld1w { z16.s }, p3/Z, [x22, #1, MUL VL]\n"
+      "ld1w { z21.s }, p3/Z, [x22, #2, MUL VL]\n"
+      ".inst 0x658aae86  // bfcvt z6.h, p3/M, z20.s\n"
+      "zip2 z5.s, z18.s, z17.s\n"
+      "ld1w { z20.s }, p3/Z, [x25, #3, MUL VL]\n"
+      "ld1w { z19.s }, p3/Z, [x25, #4, MUL VL]\n"
+      ".inst 0x658aafa4  // bfcvt z4.h, p3/M, z29.s\n"
+      "zip1 z3.s, z23.s, z16.s\n"
+      "ld1w { z2.s }, p3/Z, [x25, #5, MUL VL]\n"
+      "ld1w { z18.s }, p3/Z, [x22, #3, MUL VL]\n"
+      ".inst 0x658aaf81  // bfcvt z1.h, p3/M, z28.s\n"
+      "zip2 z0.s, z23.s, z16.s\n"
+      "ld1w { z17.s }, p3/Z, [x22, #4, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x22, #5, MUL VL]\n"
+      ".inst 0x658aaf7f  // bfcvt z31.h, p3/M, z27.s\n"
+      "zip1 z30.s, z22.s, z21.s\n"
+      ".inst 0x658aaf5d  // bfcvt z29.h, p3/M, z26.s\n"
+      "zip2 z28.s, z22.s, z21.s\n"
       "addvl x26, x26, #6\n"
-      "addvl x23, x23, #6\n"
-      ".inst 0x658aa821  // bfcvt z1.h, p2/M, z1.s\n"
-      "zip2 z0.s, z31.s, z24.s\n"
+      "addvl x25, x25, #6\n"
+      ".inst 0x658aaf3b  // bfcvt z27.h, p3/M, z25.s\n"
+      "zip1 z26.s, z20.s, z18.s\n"
+      "addvl x24, x24, #6\n"
       "addvl x22, x22, #6\n"
-      "addvl x21, x21, #6\n"
-      ".inst 0x658aabdf  // bfcvt z31.h, p2/M, z30.s\n"
-      "zip1 z30.s, z28.s, z23.s\n"
-      ".inst 0x658aabbd  // bfcvt z29.h, p2/M, z29.s\n"
-      "zip2 z28.s, z28.s, z23.s\n"
-      ".inst 0x658aab5b  // bfcvt z27.h, p2/M, z26.s\n"
-      "zip1 z26.s, z21.s, z18.s\n"
-      ".inst 0x658aab39  // bfcvt z25.h, p2/M, z25.s\n"
-      "zip2 z24.s, z21.s, z18.s\n"
-      ".inst 0x658aaad7  // bfcvt z23.h, p2/M, z22.s\n"
-      "zip1 z22.s, z20.s, z17.s\n"
-      ".inst 0x658aa955  // bfcvt z21.h, p2/M, z10.s\n"
-      "zip2 z20.s, z20.s, z17.s\n"
-      ".inst 0x658aaa73  // bfcvt z19.h, p2/M, z19.s\n"
-      "zip1 z18.s, z8.s, z16.s\n"
-      ".inst 0x658aa931  // bfcvt z17.h, p2/M, z9.s\n"
-      "zip2 z16.s, z8.s, z16.s\n"
-      ".inst 0x648aa8c7  // bfcvtnt z7.h, p2/M, z6.s\n"
-      ".inst 0x648aa885  // bfcvtnt z5.h, p2/M, z4.s\n"
-      ".inst 0x648aa843  // bfcvtnt z3.h, p2/M, z2.s\n"
-      ".inst 0x648aa801  // bfcvtnt z1.h, p2/M, z0.s\n"
-      ".inst 0x648aabdf  // bfcvtnt z31.h, p2/M, z30.s\n"
-      ".inst 0x648aab9d  // bfcvtnt z29.h, p2/M, z28.s\n"
-      "st1h { z7.h }, p2, [x24]\n"
-      "st1h { z5.h }, p2, [x24, #1, MUL VL]\n"
-      ".inst 0x648aab5b  // bfcvtnt z27.h, p2/M, z26.s\n"
-      ".inst 0x648aab19  // bfcvtnt z25.h, p2/M, z24.s\n"
-      "st1h { z3.h }, p2, [x24, #2, MUL VL]\n"
-      ".inst 0x648aaad7  // bfcvtnt z23.h, p2/M, z22.s\n"
-      ".inst 0x648aaa95  // bfcvtnt z21.h, p2/M, z20.s\n"
-      "st1h { z1.h }, p2, [x24, #3, MUL VL]\n"
-      ".inst 0x648aaa53  // bfcvtnt z19.h, p2/M, z18.s\n"
-      ".inst 0x648aaa11  // bfcvtnt z17.h, p2/M, z16.s\n"
-      "st1h { z31.h }, p2, [x24, #4, MUL VL]\n"
-      "st1h { z29.h }, p2, [x24, #5, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
-      "st1h { z27.h }, p2, [x24]\n"
-      "st1h { z25.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z23.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z21.h }, p2, [x24, #3, MUL VL]\n"
-      "st1h { z19.h }, p2, [x24, #4, MUL VL]\n"
-      "st1h { z17.h }, p2, [x24, #5, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
+      ".inst 0x658aaf19  // bfcvt z25.h, p3/M, z24.s\n"
+      "zip2 z24.s, z20.s, z18.s\n"
+      ".inst 0x658aad97  // bfcvt z23.h, p3/M, z12.s\n"
+      "zip1 z22.s, z19.s, z17.s\n"
+      ".inst 0x658aad75  // bfcvt z21.h, p3/M, z11.s\n"
+      "zip2 z20.s, z19.s, z17.s\n"
+      ".inst 0x658aad53  // bfcvt z19.h, p3/M, z10.s\n"
+      "zip1 z18.s, z2.s, z16.s\n"
+      ".inst 0x658aad31  // bfcvt z17.h, p3/M, z9.s\n"
+      "zip2 z16.s, z2.s, z16.s\n"
+      ".inst 0x648aace8  // bfcvtnt z8.h, p3/M, z7.s\n"
+      ".inst 0x648aaca6  // bfcvtnt z6.h, p3/M, z5.s\n"
+      "st1h { z8.h }, p3, [x21]\n"
+      ".inst 0x648aac64  // bfcvtnt z4.h, p3/M, z3.s\n"
+      ".inst 0x648aac01  // bfcvtnt z1.h, p3/M, z0.s\n"
+      "st1h { z6.h }, p3, [x21, #1, MUL VL]\n"
+      ".inst 0x648aafdf  // bfcvtnt z31.h, p3/M, z30.s\n"
+      ".inst 0x648aaf9d  // bfcvtnt z29.h, p3/M, z28.s\n"
+      "st1h { z4.h }, p3, [x21, #2, MUL VL]\n"
+      "st1h { z1.h }, p3, [x21, #3, MUL VL]\n"
+      ".inst 0x648aaf5b  // bfcvtnt z27.h, p3/M, z26.s\n"
+      ".inst 0x648aaf19  // bfcvtnt z25.h, p3/M, z24.s\n"
+      "st1h { z31.h }, p3, [x21, #4, MUL VL]\n"
+      ".inst 0x648aaed7  // bfcvtnt z23.h, p3/M, z22.s\n"
+      ".inst 0x648aae95  // bfcvtnt z21.h, p3/M, z20.s\n"
+      "st1h { z29.h }, p3, [x21, #5, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
+      ".inst 0x648aae53  // bfcvtnt z19.h, p3/M, z18.s\n"
+      ".inst 0x648aae11  // bfcvtnt z17.h, p3/M, z16.s\n"
+      "st1h { z27.h }, p3, [x21]\n"
+      "st1h { z25.h }, p3, [x21, #1, MUL VL]\n"
+      "st1h { z23.h }, p3, [x21, #2, MUL VL]\n"
+      "st1h { z21.h }, p3, [x21, #3, MUL VL]\n"
+      "st1h { z19.h }, p3, [x21, #4, MUL VL]\n"
+      "st1h { z17.h }, p3, [x21, #5, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x20, x25\n"
-      "decd x25, ALL, MUL #6\n"
-      "whilelt p0.s, XZR, x20\n"
+      "mov x20, x23\n"
+      "whilelt p2.s, XZR, x20\n"
+      "ld1w { z20.s }, p2/Z, [x26]\n"
+      "ld1w { z19.s }, p2/Z, [x24]\n"
       "decw x20\n"
       "whilelt p1.s, XZR, x20\n"
+      "ld1w { z18.s }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1w { z17.s }, p1/Z, [x24, #1, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z17.s }, p0/Z, [x26]\n"
-      "ld1w { z16.s }, p0/Z, [x22]\n"
-      "ld1w { z23.s }, p0/Z, [x23]\n"
-      "ld1w { z19.s }, p0/Z, [x21]\n"
       "whilelt p0.s, XZR, x20\n"
-      "cmp x25, #0x0\n"
-      "ld1w { z22.s }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1w { z25.s }, p0/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z16.s }, p0/Z, [x24, #2, MUL VL]\n"
+      "ld1w { z24.s }, p2/Z, [x25]\n"
+      "ld1w { z30.s }, p1/Z, [x25, #1, MUL VL]\n"
+      "zip1 z23.s, z20.s, z19.s\n"
+      "zip2 z22.s, z20.s, z19.s\n"
+      "ld1w { z29.s }, p0/Z, [x25, #2, MUL VL]\n"
+      "ld1w { z21.s }, p2/Z, [x22]\n"
+      "zip1 z20.s, z18.s, z17.s\n"
+      "zip2 z19.s, z18.s, z17.s\n"
       "ld1w { z18.s }, p1/Z, [x22, #1, MUL VL]\n"
-      "ld1w { z31.s }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1w { z30.s }, p1/Z, [x21, #1, MUL VL]\n"
-      "zip1 z21.s, z17.s, z16.s\n"
-      "zip2 z17.s, z17.s, z16.s\n"
-      "ld1w { z20.s }, p0/Z, [x26, #2, MUL VL]\n"
-      "ld1w { z16.s }, p0/Z, [x22, #2, MUL VL]\n"
-      "zip1 z29.s, z23.s, z19.s\n"
-      "zip2 z28.s, z23.s, z19.s\n"
-      "ld1w { z27.s }, p0/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z26.s }, p0/Z, [x21, #2, MUL VL]\n"
-      "zip1 z19.s, z22.s, z18.s\n"
-      "zip2 z18.s, z22.s, z18.s\n"
-      ".inst 0x658aaab9  // bfcvt z25.h, p2/M, z21.s\n"
-      ".inst 0x658aaa38  // bfcvt z24.h, p2/M, z17.s\n"
+      "ld1w { z28.s }, p0/Z, [x22, #2, MUL VL]\n"
+      "zip1 z17.s, z25.s, z16.s\n"
+      "zip2 z16.s, z25.s, z16.s\n"
+      "decd x23, ALL, MUL #6\n"
+      ".inst 0x658aaefb  // bfcvt z27.h, p3/M, z23.s\n"
+      "zip1 z26.s, z24.s, z21.s\n"
+      "cmp x23, #0x0\n"
+      ".inst 0x658aaed9  // bfcvt z25.h, p3/M, z22.s\n"
+      "zip2 z24.s, z24.s, z21.s\n"
       "addvl x26, x26, #3\n"
-      "addvl x23, x23, #3\n"
-      "zip1 z17.s, z20.s, z16.s\n"
-      "zip2 z16.s, z20.s, z16.s\n"
+      "addvl x25, x25, #3\n"
+      ".inst 0x658aae97  // bfcvt z23.h, p3/M, z20.s\n"
+      "zip1 z22.s, z30.s, z18.s\n"
+      "addvl x24, x24, #3\n"
       "addvl x22, x22, #3\n"
-      "addvl x21, x21, #3\n"
-      ".inst 0x658aaa77  // bfcvt z23.h, p2/M, z19.s\n"
-      "zip1 z22.s, z31.s, z30.s\n"
-      ".inst 0x658aaa55  // bfcvt z21.h, p2/M, z18.s\n"
-      "zip2 z20.s, z31.s, z30.s\n"
-      ".inst 0x658aaa33  // bfcvt z19.h, p2/M, z17.s\n"
-      "zip1 z18.s, z27.s, z26.s\n"
-      ".inst 0x658aaa11  // bfcvt z17.h, p2/M, z16.s\n"
-      "zip2 z16.s, z27.s, z26.s\n"
-      ".inst 0x648aabb9  // bfcvtnt z25.h, p2/M, z29.s\n"
-      ".inst 0x648aab98  // bfcvtnt z24.h, p2/M, z28.s\n"
-      ".inst 0x648aaad7  // bfcvtnt z23.h, p2/M, z22.s\n"
-      ".inst 0x648aaa95  // bfcvtnt z21.h, p2/M, z20.s\n"
-      ".inst 0x648aaa53  // bfcvtnt z19.h, p2/M, z18.s\n"
-      ".inst 0x648aaa11  // bfcvtnt z17.h, p2/M, z16.s\n"
-      "st1h { z25.h }, p2, [x24]\n"
-      "st1h { z24.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z23.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z21.h }, p2, [x24, #3, MUL VL]\n"
-      "st1h { z19.h }, p2, [x24, #4, MUL VL]\n"
-      "st1h { z17.h }, p2, [x24, #5, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
+      ".inst 0x658aae75  // bfcvt z21.h, p3/M, z19.s\n"
+      "zip2 z20.s, z30.s, z18.s\n"
+      ".inst 0x658aae33  // bfcvt z19.h, p3/M, z17.s\n"
+      "zip1 z18.s, z29.s, z28.s\n"
+      ".inst 0x658aae11  // bfcvt z17.h, p3/M, z16.s\n"
+      "zip2 z16.s, z29.s, z28.s\n"
+      ".inst 0x648aaf5b  // bfcvtnt z27.h, p3/M, z26.s\n"
+      ".inst 0x648aaf19  // bfcvtnt z25.h, p3/M, z24.s\n"
+      "st1h { z27.h }, p3, [x21]\n"
+      ".inst 0x648aaed7  // bfcvtnt z23.h, p3/M, z22.s\n"
+      ".inst 0x648aae95  // bfcvtnt z21.h, p3/M, z20.s\n"
+      "st1h { z25.h }, p3, [x21, #1, MUL VL]\n"
+      ".inst 0x648aae53  // bfcvtnt z19.h, p3/M, z18.s\n"
+      ".inst 0x648aae11  // bfcvtnt z17.h, p3/M, z16.s\n"
+      "st1h { z23.h }, p3, [x21, #2, MUL VL]\n"
+      "st1h { z21.h }, p3, [x21, #3, MUL VL]\n"
+      "st1h { z19.h }, p3, [x21, #4, MUL VL]\n"
+      "st1h { z17.h }, p3, [x21, #5, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -216,7 +216,7 @@ void sve_transpose_interleave_6VL_2x4_fp32bf16(bfloat16 *out, const float *in, s
       "bge 1b\n"
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_4x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_4x2.hpp
index 314e3aebbd..ac61301ea4 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_4x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_6VL_4x2.hpp
@@ -40,160 +40,160 @@ void sve_transpose_interleave_6VL_4x2(uint32_t *out, const uint32_t *in, size_t
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "ptrue p2.b\n"
+      "ptrue p3.b\n"
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x28, %x[in]\n"
       "mov x27, %x[width]\n"
       "cntw x26, ALL, MUL #6\n"
-      "mov x25, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x24, x28, %x[in_stride]\n"
+      "add x25, x28, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
       "cmp x27, x26\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1w { z18.s }, p2/Z, [x28]\n"
-      "ld1w { z28.s }, p2/Z, [x28, #1, MUL VL]\n"
-      "mov x21, x25\n"
-      "add x25, x25, %x[out_stride]\n"
-      "ld1w { z22.s }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1w { z17.s }, p2/Z, [x24]\n"
-      "mov x20, x25\n"
+      "ld1w { z18.s }, p3/Z, [x28]\n"
+      "ld1w { z17.s }, p3/Z, [x28, #1, MUL VL]\n"
+      "mov x21, x22\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1w { z19.s }, p3/Z, [x28, #2, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x25]\n"
+      "zip1 z9.s, z18.s, z16.s\n"
+      "zip2 z8.s, z18.s, z16.s\n"
+      "ld1w { z16.s }, p3/Z, [x25, #1, MUL VL]\n"
+      "ld1w { z18.s }, p3/Z, [x25, #2, MUL VL]\n"
+      "zip1 z7.s, z17.s, z16.s\n"
+      "zip2 z6.s, z17.s, z16.s\n"
+      "ld1w { z17.s }, p3/Z, [x24]\n"
+      "ld1w { z16.s }, p3/Z, [x23]\n"
+      "zip1 z5.s, z19.s, z18.s\n"
+      "zip2 z4.s, z19.s, z18.s\n"
+      "ld1w { z18.s }, p3/Z, [x28, #3, MUL VL]\n"
+      "ld1w { z21.s }, p3/Z, [x28, #4, MUL VL]\n"
+      "zip1 z3.s, z17.s, z16.s\n"
+      "zip2 z2.s, z17.s, z16.s\n"
+      "ld1w { z20.s }, p3/Z, [x28, #5, MUL VL]\n"
+      "ld1w { z17.s }, p3/Z, [x25, #3, MUL VL]\n"
+      "mov x20, x22\n"
+      "zip1 z1.s, z18.s, z17.s\n"
+      "ld1w { z19.s }, p3/Z, [x25, #4, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x25, #5, MUL VL]\n"
+      "zip2 z0.s, z18.s, z17.s\n"
+      "zip1 z31.s, z21.s, z19.s\n"
+      "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n"
+      "ld1w { z17.s }, p3/Z, [x24, #2, MUL VL]\n"
+      "zip2 z30.s, z21.s, z19.s\n"
+      "zip1 z29.s, z20.s, z16.s\n"
+      "ld1w { z19.s }, p3/Z, [x24, #3, MUL VL]\n"
+      "ld1w { z28.s }, p3/Z, [x24, #4, MUL VL]\n"
+      "zip2 z27.s, z20.s, z16.s\n"
       "sub x27, x27, x26\n"
-      "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x24, #2, MUL VL]\n"
+      "ld1w { z26.s }, p3/Z, [x24, #5, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x23, #1, MUL VL]\n"
+      "zip1 z25.s, z18.s, z16.s\n"
+      "zip2 z24.s, z18.s, z16.s\n"
+      "ld1w { z16.s }, p3/Z, [x23, #2, MUL VL]\n"
+      "ld1w { z18.s }, p3/Z, [x23, #3, MUL VL]\n"
+      "zip1 z23.s, z17.s, z16.s\n"
+      "zip2 z22.s, z17.s, z16.s\n"
+      "ld1w { z17.s }, p3/Z, [x23, #4, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x23, #5, MUL VL]\n"
+      "st1w { z9.s }, p3, [x21]\n"
+      "zip1 z21.s, z19.s, z18.s\n"
+      "st1w { z8.s }, p3, [x21, #1, MUL VL]\n"
+      "zip2 z20.s, z19.s, z18.s\n"
       "cmp x27, x26\n"
-      "add x25, x25, %x[out_stride]\n"
-      "ld1w { z27.s }, p2/Z, [x23]\n"
-      "ld1w { z26.s }, p2/Z, [x22]\n"
-      "ld1w { z20.s }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x28, #4, MUL VL]\n"
-      "zip1 z8.s, z18.s, z17.s\n"
-      "zip2 z25.s, z18.s, z17.s\n"
-      "ld1w { z24.s }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1w { z18.s }, p2/Z, [x24, #3, MUL VL]\n"
-      "zip1 z23.s, z28.s, z16.s\n"
-      "zip2 z7.s, z28.s, z16.s\n"
-      "ld1w { z17.s }, p2/Z, [x24, #4, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x24, #5, MUL VL]\n"
-      "zip1 z6.s, z22.s, z21.s\n"
-      "zip2 z5.s, z22.s, z21.s\n"
-      "ld1w { z22.s }, p2/Z, [x23, #1, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x23, #2, MUL VL]\n"
-      "zip1 z4.s, z27.s, z26.s\n"
-      "zip2 z3.s, z27.s, z26.s\n"
-      "ld1w { z2.s }, p2/Z, [x23, #3, MUL VL]\n"
-      "ld1w { z1.s }, p2/Z, [x23, #4, MUL VL]\n"
-      "zip1 z0.s, z20.s, z18.s\n"
-      "zip2 z31.s, z20.s, z18.s\n"
-      "ld1w { z30.s }, p2/Z, [x23, #5, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x22, #1, MUL VL]\n"
-      "zip1 z29.s, z19.s, z17.s\n"
-      "zip2 z28.s, z19.s, z17.s\n"
-      "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z18.s }, p2/Z, [x22, #3, MUL VL]\n"
-      "zip1 z27.s, z24.s, z16.s\n"
-      "zip2 z26.s, z24.s, z16.s\n"
-      "ld1w { z17.s }, p2/Z, [x22, #4, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x22, #5, MUL VL]\n"
-      "st1w { z8.s }, p2, [x21]\n"
       "addvl x28, x28, #6\n"
-      "st1w { z25.s }, p2, [x21, #1, MUL VL]\n"
-      "zip1 z25.s, z22.s, z20.s\n"
-      "zip2 z24.s, z22.s, z20.s\n"
+      "st1w { z7.s }, p3, [x21, #2, MUL VL]\n"
+      "addvl x25, x25, #6\n"
       "addvl x24, x24, #6\n"
-      "st1w { z23.s }, p2, [x21, #2, MUL VL]\n"
-      "zip1 z23.s, z21.s, z19.s\n"
-      "zip2 z22.s, z21.s, z19.s\n"
+      "zip1 z19.s, z28.s, z17.s\n"
+      "st1w { z6.s }, p3, [x21, #3, MUL VL]\n"
       "addvl x23, x23, #6\n"
-      "st1w { z7.s }, p2, [x21, #3, MUL VL]\n"
-      "zip1 z21.s, z2.s, z18.s\n"
-      "zip2 z20.s, z2.s, z18.s\n"
-      "addvl x22, x22, #6\n"
-      "st1w { z6.s }, p2, [x21, #4, MUL VL]\n"
-      "zip1 z19.s, z1.s, z17.s\n"
-      "zip2 z18.s, z1.s, z17.s\n"
-      "st1w { z5.s }, p2, [x21, #5, MUL VL]\n"
-      "zip1 z17.s, z30.s, z16.s\n"
-      "zip2 z16.s, z30.s, z16.s\n"
-      "st1w { z4.s }, p2, [x21, #6, MUL VL]\n"
-      "st1w { z3.s }, p2, [x21, #7, MUL VL]\n"
+      "zip2 z18.s, z28.s, z17.s\n"
+      "zip1 z17.s, z26.s, z16.s\n"
+      "st1w { z5.s }, p3, [x21, #4, MUL VL]\n"
+      "zip2 z16.s, z26.s, z16.s\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1w { z4.s }, p3, [x21, #5, MUL VL]\n"
+      "st1w { z3.s }, p3, [x21, #6, MUL VL]\n"
+      "st1w { z2.s }, p3, [x21, #7, MUL VL]\n"
       "addvl x21, x21, #12\n"
-      "st1w { z25.s }, p2, [x21, #-4, MUL VL]\n"
-      "st1w { z24.s }, p2, [x21, #-3, MUL VL]\n"
-      "st1w { z23.s }, p2, [x21, #-2, MUL VL]\n"
-      "st1w { z22.s }, p2, [x21, #-1, MUL VL]\n"
-      "st1w { z0.s }, p2, [x20]\n"
-      "st1w { z31.s }, p2, [x20, #1, MUL VL]\n"
-      "st1w { z29.s }, p2, [x20, #2, MUL VL]\n"
-      "st1w { z28.s }, p2, [x20, #3, MUL VL]\n"
-      "st1w { z27.s }, p2, [x20, #4, MUL VL]\n"
-      "st1w { z26.s }, p2, [x20, #5, MUL VL]\n"
-      "st1w { z21.s }, p2, [x20, #6, MUL VL]\n"
-      "st1w { z20.s }, p2, [x20, #7, MUL VL]\n"
+      "st1w { z25.s }, p3, [x21, #-4, MUL VL]\n"
+      "st1w { z24.s }, p3, [x21, #-3, MUL VL]\n"
+      "st1w { z23.s }, p3, [x21, #-2, MUL VL]\n"
+      "st1w { z22.s }, p3, [x21, #-1, MUL VL]\n"
+      "st1w { z1.s }, p3, [x20]\n"
+      "st1w { z0.s }, p3, [x20, #1, MUL VL]\n"
+      "st1w { z31.s }, p3, [x20, #2, MUL VL]\n"
+      "st1w { z30.s }, p3, [x20, #3, MUL VL]\n"
+      "st1w { z29.s }, p3, [x20, #4, MUL VL]\n"
+      "st1w { z27.s }, p3, [x20, #5, MUL VL]\n"
+      "st1w { z21.s }, p3, [x20, #6, MUL VL]\n"
+      "st1w { z20.s }, p3, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
-      "st1w { z19.s }, p2, [x20, #-4, MUL VL]\n"
-      "st1w { z18.s }, p2, [x20, #-3, MUL VL]\n"
-      "st1w { z17.s }, p2, [x20, #-2, MUL VL]\n"
-      "st1w { z16.s }, p2, [x20, #-1, MUL VL]\n"
+      "st1w { z19.s }, p3, [x20, #-4, MUL VL]\n"
+      "st1w { z18.s }, p3, [x20, #-3, MUL VL]\n"
+      "st1w { z17.s }, p3, [x20, #-2, MUL VL]\n"
+      "st1w { z16.s }, p3, [x20, #-1, MUL VL]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cbz x27, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x21, x27\n"
-      "mov x20, x25\n"
+      "mov x20, x27\n"
+      "whilelt p2.s, XZR, x20\n"
+      "ld1w { z19.s }, p2/Z, [x28]\n"
+      "ld1w { z18.s }, p2/Z, [x25]\n"
+      "decw x20\n"
+      "whilelt p1.s, XZR, x20\n"
+      "ld1w { z17.s }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x25, #1, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z22.s }, p0/Z, [x28, #2, MUL VL]\n"
+      "ld1w { z21.s }, p0/Z, [x25, #2, MUL VL]\n"
+      "ld1w { z28.s }, p2/Z, [x24]\n"
+      "ld1w { z27.s }, p2/Z, [x23]\n"
+      "mov x20, x22\n"
       "decd x27, ALL, MUL #6\n"
-      "add x25, x25, %x[out_stride]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "decw x21\n"
-      "whilelt p1.s, XZR, x21\n"
-      "decw x21\n"
-      "ld1w { z19.s }, p0/Z, [x28]\n"
-      "ld1w { z18.s }, p0/Z, [x24]\n"
-      "ld1w { z22.s }, p0/Z, [x23]\n"
-      "ld1w { z17.s }, p0/Z, [x22]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "cmp x27, #0x0\n"
-      "ld1w { z21.s }, p1/Z, [x28, #1, MUL VL]\n"
-      "ld1w { z16.s }, p1/Z, [x24, #1, MUL VL]\n"
-      "ld1w { z28.s }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1w { z27.s }, p1/Z, [x22, #1, MUL VL]\n"
+      "ld1w { z26.s }, p1/Z, [x24, #1, MUL VL]\n"
+      "ld1w { z25.s }, p0/Z, [x24, #2, MUL VL]\n"
       "zip1 z20.s, z19.s, z18.s\n"
-      "zip2 z26.s, z19.s, z18.s\n"
-      "ld1w { z19.s }, p0/Z, [x28, #2, MUL VL]\n"
-      "ld1w { z18.s }, p0/Z, [x24, #2, MUL VL]\n"
-      "zip1 z25.s, z22.s, z17.s\n"
-      "zip2 z24.s, z22.s, z17.s\n"
+      "zip2 z19.s, z19.s, z18.s\n"
+      "ld1w { z24.s }, p1/Z, [x23, #1, MUL VL]\n"
       "ld1w { z23.s }, p0/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z22.s }, p0/Z, [x22, #2, MUL VL]\n"
-      "zip1 z17.s, z21.s, z16.s\n"
-      "zip2 z16.s, z21.s, z16.s\n"
-      "st1w { z20.s }, p2, [x20]\n"
+      "zip1 z18.s, z17.s, z16.s\n"
+      "zip2 z17.s, z17.s, z16.s\n"
+      "zip1 z16.s, z22.s, z21.s\n"
+      "zip2 z22.s, z22.s, z21.s\n"
+      "st1w { z20.s }, p3, [x20]\n"
+      "cmp x27, #0x0\n"
+      "zip1 z21.s, z28.s, z27.s\n"
+      "zip2 z20.s, z28.s, z27.s\n"
+      "st1w { z19.s }, p3, [x20, #1, MUL VL]\n"
       "addvl x28, x28, #3\n"
+      "st1w { z18.s }, p3, [x20, #2, MUL VL]\n"
+      "addvl x25, x25, #3\n"
       "addvl x24, x24, #3\n"
-      "zip1 z21.s, z28.s, z27.s\n"
-      "zip1 z20.s, z19.s, z18.s\n"
-      "zip2 z19.s, z19.s, z18.s\n"
-      "st1w { z26.s }, p2, [x20, #1, MUL VL]\n"
+      "zip1 z19.s, z26.s, z24.s\n"
+      "st1w { z17.s }, p3, [x20, #3, MUL VL]\n"
       "addvl x23, x23, #3\n"
-      "st1w { z17.s }, p2, [x20, #2, MUL VL]\n"
-      "addvl x22, x22, #3\n"
-      "zip2 z18.s, z28.s, z27.s\n"
-      "zip1 z17.s, z23.s, z22.s\n"
-      "st1w { z16.s }, p2, [x20, #3, MUL VL]\n"
-      "zip2 z16.s, z23.s, z22.s\n"
-      "st1w { z20.s }, p2, [x20, #4, MUL VL]\n"
-      "st1w { z19.s }, p2, [x20, #5, MUL VL]\n"
-      "st1w { z25.s }, p2, [x20, #6, MUL VL]\n"
-      "st1w { z24.s }, p2, [x20, #7, MUL VL]\n"
+      "zip2 z18.s, z26.s, z24.s\n"
+      "zip1 z17.s, z25.s, z23.s\n"
+      "st1w { z16.s }, p3, [x20, #4, MUL VL]\n"
+      "zip2 z16.s, z25.s, z23.s\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1w { z22.s }, p3, [x20, #5, MUL VL]\n"
+      "st1w { z21.s }, p3, [x20, #6, MUL VL]\n"
+      "st1w { z20.s }, p3, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #12\n"
-      "st1w { z21.s }, p2, [x20, #-4, MUL VL]\n"
-      "st1w { z18.s }, p2, [x20, #-3, MUL VL]\n"
-      "st1w { z17.s }, p2, [x20, #-2, MUL VL]\n"
-      "st1w { z16.s }, p2, [x20, #-1, MUL VL]\n"
+      "st1w { z19.s }, p3, [x20, #-4, MUL VL]\n"
+      "st1w { z18.s }, p3, [x20, #-3, MUL VL]\n"
+      "st1w { z17.s }, p3, [x20, #-2, MUL VL]\n"
+      "st1w { z16.s }, p3, [x20, #-1, MUL VL]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x4\n"
@@ -205,90 +205,90 @@ void sve_transpose_interleave_6VL_4x2(uint32_t *out, const uint32_t *in, size_t
       "mov x28, %x[in]\n"
       "mov x21, %x[width]\n"
       "cntw x20, ALL, MUL #6\n"
+      "add x25, x28, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "mov x25, %x[out]\n"
-      "sub %x[height], %x[height], #0x2\n"
-      "add x24, x28, %x[in_stride]\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
+      "add %x[in], x25, %x[in_stride]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x2\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1w { z17.s }, p2/Z, [x28]\n"
-      "ld1w { z23.s }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1w { z17.s }, p3/Z, [x28]\n"
+      "ld1w { z19.s }, p3/Z, [x28, #1, MUL VL]\n"
       "sub x21, x21, x20\n"
-      "ld1w { z29.s }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x24]\n"
       "cmp x21, x20\n"
-      "ld1w { z22.s }, p2/Z, [x24, #1, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1w { z28.s }, p2/Z, [x28, #3, MUL VL]\n"
-      "ld1w { z27.s }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1w { z26.s }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x24, #3, MUL VL]\n"
-      "zip1 z19.s, z17.s, z16.s\n"
-      "zip2 z18.s, z17.s, z16.s\n"
-      "ld1w { z17.s }, p2/Z, [x24, #4, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x24, #5, MUL VL]\n"
-      "zip1 z16.s, z23.s, z22.s\n"
-      "zip2 z24.s, z23.s, z22.s\n"
-      "zip1 z23.s, z29.s, z21.s\n"
-      "zip2 z22.s, z29.s, z21.s\n"
+      "ld1w { z18.s }, p3/Z, [x28, #2, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x25]\n"
+      "zip1 z28.s, z17.s, z16.s\n"
+      "zip2 z20.s, z17.s, z16.s\n"
+      "ld1w { z17.s }, p3/Z, [x25, #1, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x25, #2, MUL VL]\n"
+      "zip1 z27.s, z19.s, z17.s\n"
+      "zip2 z26.s, z19.s, z17.s\n"
+      "ld1w { z19.s }, p3/Z, [x28, #3, MUL VL]\n"
+      "ld1w { z25.s }, p3/Z, [x28, #4, MUL VL]\n"
+      "zip1 z24.s, z18.s, z16.s\n"
+      "zip2 z23.s, z18.s, z16.s\n"
+      "ld1w { z22.s }, p3/Z, [x28, #5, MUL VL]\n"
+      "ld1w { z18.s }, p3/Z, [x25, #3, MUL VL]\n"
       "addvl x28, x28, #6\n"
-      "addvl x24, x24, #6\n"
-      "st1w { z19.s }, p2, [x25]\n"
-      "zip1 z21.s, z28.s, z20.s\n"
-      "zip2 z20.s, z28.s, z20.s\n"
-      "st1w { z18.s }, p2, [x25, #1, MUL VL]\n"
-      "zip1 z19.s, z27.s, z17.s\n"
-      "zip2 z18.s, z27.s, z17.s\n"
-      "st1w { z16.s }, p2, [x25, #2, MUL VL]\n"
-      "zip1 z17.s, z26.s, z25.s\n"
-      "zip2 z16.s, z26.s, z25.s\n"
-      "st1w { z24.s }, p2, [x25, #3, MUL VL]\n"
-      "st1w { z23.s }, p2, [x25, #4, MUL VL]\n"
-      "st1w { z22.s }, p2, [x25, #5, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
-      "st1w { z21.s }, p2, [x25]\n"
-      "st1w { z20.s }, p2, [x25, #1, MUL VL]\n"
-      "st1w { z19.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #3, MUL VL]\n"
-      "st1w { z17.s }, p2, [x25, #4, MUL VL]\n"
-      "st1w { z16.s }, p2, [x25, #5, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
+      "zip1 z21.s, z19.s, z18.s\n"
+      "ld1w { z17.s }, p3/Z, [x25, #4, MUL VL]\n"
+      "ld1w { z16.s }, p3/Z, [x25, #5, MUL VL]\n"
+      "st1w { z28.s }, p3, [x22]\n"
+      "addvl x25, x25, #6\n"
+      "st1w { z20.s }, p3, [x22, #1, MUL VL]\n"
+      "zip2 z20.s, z19.s, z18.s\n"
+      "zip1 z19.s, z25.s, z17.s\n"
+      "st1w { z27.s }, p3, [x22, #2, MUL VL]\n"
+      "zip2 z18.s, z25.s, z17.s\n"
+      "zip1 z17.s, z22.s, z16.s\n"
+      "st1w { z26.s }, p3, [x22, #3, MUL VL]\n"
+      "zip2 z16.s, z22.s, z16.s\n"
+      "st1w { z24.s }, p3, [x22, #4, MUL VL]\n"
+      "st1w { z23.s }, p3, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1w { z21.s }, p3, [x22]\n"
+      "st1w { z20.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z19.s }, p3, [x22, #2, MUL VL]\n"
+      "st1w { z18.s }, p3, [x22, #3, MUL VL]\n"
+      "st1w { z17.s }, p3, [x22, #4, MUL VL]\n"
+      "st1w { z16.s }, p3, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decd x21, ALL, MUL #6\n"
       "whilelt p0.s, XZR, x20\n"
+      "ld1w { z20.s }, p0/Z, [x28]\n"
+      "ld1w { z19.s }, p0/Z, [x25]\n"
       "decw x20\n"
-      "whilelt p1.s, XZR, x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z18.s }, p0/Z, [x28, #1, MUL VL]\n"
+      "ld1w { z17.s }, p0/Z, [x25, #1, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z17.s }, p0/Z, [x28]\n"
-      "ld1w { z16.s }, p0/Z, [x24]\n"
       "whilelt p0.s, XZR, x20\n"
+      "ld1w { z22.s }, p0/Z, [x28, #2, MUL VL]\n"
+      "ld1w { z16.s }, p0/Z, [x25, #2, MUL VL]\n"
+      "decd x21, ALL, MUL #6\n"
       "cmp x21, #0x0\n"
-      "ld1w { z22.s }, p1/Z, [x28, #1, MUL VL]\n"
-      "ld1w { z18.s }, p1/Z, [x24, #1, MUL VL]\n"
-      "zip1 z21.s, z17.s, z16.s\n"
-      "zip2 z17.s, z17.s, z16.s\n"
-      "ld1w { z20.s }, p0/Z, [x28, #2, MUL VL]\n"
+      "zip1 z21.s, z20.s, z19.s\n"
+      "zip2 z20.s, z20.s, z19.s\n"
       "addvl x28, x28, #3\n"
-      "ld1w { z16.s }, p0/Z, [x24, #2, MUL VL]\n"
-      "addvl x24, x24, #3\n"
-      "zip1 z19.s, z22.s, z18.s\n"
-      "zip2 z18.s, z22.s, z18.s\n"
-      "st1w { z21.s }, p2, [x25]\n"
-      "st1w { z17.s }, p2, [x25, #1, MUL VL]\n"
-      "zip1 z17.s, z20.s, z16.s\n"
-      "zip2 z16.s, z20.s, z16.s\n"
-      "st1w { z19.s }, p2, [x25, #2, MUL VL]\n"
-      "st1w { z18.s }, p2, [x25, #3, MUL VL]\n"
-      "st1w { z17.s }, p2, [x25, #4, MUL VL]\n"
-      "st1w { z16.s }, p2, [x25, #5, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
+      "addvl x25, x25, #3\n"
+      "zip1 z19.s, z18.s, z17.s\n"
+      "zip2 z18.s, z18.s, z17.s\n"
+      "zip1 z17.s, z22.s, z16.s\n"
+      "zip2 z16.s, z22.s, z16.s\n"
+      "st1w { z21.s }, p3, [x22]\n"
+      "st1w { z20.s }, p3, [x22, #1, MUL VL]\n"
+      "st1w { z19.s }, p3, [x22, #2, MUL VL]\n"
+      "st1w { z18.s }, p3, [x22, #3, MUL VL]\n"
+      "st1w { z17.s }, p3, [x22, #4, MUL VL]\n"
+      "st1w { z16.s }, p3, [x22, #5, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -297,7 +297,7 @@ void sve_transpose_interleave_6VL_4x2(uint32_t *out, const uint32_t *in, size_t
       "12:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL.hpp
index ae228d3916..87d5372b57 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL.hpp
@@ -34,152 +34,152 @@ void sve_transpose_interleave_8VL(uint32_t *out, const uint32_t *in, size_t widt
 
     __asm__ __volatile__(
       "cmp %x[height], #0x2\n"
-      "ptrue p2.b\n"
+      "ptrue p1.b\n"
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
       "mov x25, %x[width]\n"
       "cntw x24, ALL, MUL #16\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x2\n"
-      "add x22, x26, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "add x23, x26, %x[in_stride]\n"
       "cmp x25, x24\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x2\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1w { z15.s }, p2/Z, [x26]\n"
-      "ld1w { z14.s }, p2/Z, [x26, #1, MUL VL]\n"
-      "mov x21, x23\n"
-      "add x23, x23, %x[out_stride]\n"
-      "ld1w { z13.s }, p2/Z, [x26, #2, MUL VL]\n"
-      "ld1w { z12.s }, p2/Z, [x26, #3, MUL VL]\n"
-      "mov x20, x23\n"
+      "ld1w { z15.s }, p1/Z, [x26]\n"
+      "ld1w { z14.s }, p1/Z, [x26, #1, MUL VL]\n"
+      "mov x21, x22\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1w { z13.s }, p1/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z12.s }, p1/Z, [x26, #3, MUL VL]\n"
+      "mov x20, x22\n"
       "sub x25, x25, x24\n"
-      "ld1w { z11.s }, p2/Z, [x26, #4, MUL VL]\n"
-      "ld1w { z10.s }, p2/Z, [x26, #5, MUL VL]\n"
+      "ld1w { z11.s }, p1/Z, [x26, #4, MUL VL]\n"
+      "ld1w { z10.s }, p1/Z, [x26, #5, MUL VL]\n"
       "cmp x25, x24\n"
-      "add x23, x23, %x[out_stride]\n"
-      "ld1w { z9.s }, p2/Z, [x26, #6, MUL VL]\n"
-      "ld1w { z8.s }, p2/Z, [x26, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1w { z9.s }, p1/Z, [x26, #6, MUL VL]\n"
+      "ld1w { z8.s }, p1/Z, [x26, #7, MUL VL]\n"
       "addvl x26, x26, #16\n"
-      "ld1w { z7.s }, p2/Z, [x22]\n"
-      "ld1w { z6.s }, p2/Z, [x22, #1, MUL VL]\n"
-      "ld1w { z5.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z4.s }, p2/Z, [x22, #3, MUL VL]\n"
-      "ld1w { z3.s }, p2/Z, [x22, #4, MUL VL]\n"
-      "ld1w { z2.s }, p2/Z, [x22, #5, MUL VL]\n"
-      "ld1w { z1.s }, p2/Z, [x22, #6, MUL VL]\n"
-      "ld1w { z0.s }, p2/Z, [x22, #7, MUL VL]\n"
-      "addvl x22, x22, #16\n"
-      "ld1w { z31.s }, p2/Z, [x26, #-8, MUL VL]\n"
-      "ld1w { z30.s }, p2/Z, [x26, #-7, MUL VL]\n"
-      "ld1w { z29.s }, p2/Z, [x26, #-6, MUL VL]\n"
-      "ld1w { z28.s }, p2/Z, [x26, #-5, MUL VL]\n"
-      "ld1w { z27.s }, p2/Z, [x26, #-4, MUL VL]\n"
-      "ld1w { z26.s }, p2/Z, [x26, #-3, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x26, #-2, MUL VL]\n"
-      "ld1w { z24.s }, p2/Z, [x26, #-1, MUL VL]\n"
-      "ld1w { z23.s }, p2/Z, [x22, #-8, MUL VL]\n"
-      "ld1w { z22.s }, p2/Z, [x22, #-7, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x22, #-6, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x22, #-5, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x22, #-4, MUL VL]\n"
-      "ld1w { z18.s }, p2/Z, [x22, #-3, MUL VL]\n"
-      "ld1w { z17.s }, p2/Z, [x22, #-2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x22, #-1, MUL VL]\n"
-      "st1w { z15.s }, p2, [x21]\n"
-      "st1w { z14.s }, p2, [x21, #1, MUL VL]\n"
-      "st1w { z13.s }, p2, [x21, #2, MUL VL]\n"
-      "st1w { z12.s }, p2, [x21, #3, MUL VL]\n"
-      "st1w { z11.s }, p2, [x21, #4, MUL VL]\n"
-      "st1w { z10.s }, p2, [x21, #5, MUL VL]\n"
-      "st1w { z9.s }, p2, [x21, #6, MUL VL]\n"
-      "st1w { z8.s }, p2, [x21, #7, MUL VL]\n"
+      "ld1w { z7.s }, p1/Z, [x23]\n"
+      "ld1w { z6.s }, p1/Z, [x23, #1, MUL VL]\n"
+      "ld1w { z5.s }, p1/Z, [x23, #2, MUL VL]\n"
+      "ld1w { z4.s }, p1/Z, [x23, #3, MUL VL]\n"
+      "ld1w { z3.s }, p1/Z, [x23, #4, MUL VL]\n"
+      "ld1w { z2.s }, p1/Z, [x23, #5, MUL VL]\n"
+      "ld1w { z1.s }, p1/Z, [x23, #6, MUL VL]\n"
+      "ld1w { z0.s }, p1/Z, [x23, #7, MUL VL]\n"
+      "addvl x23, x23, #16\n"
+      "ld1w { z31.s }, p1/Z, [x26, #-8, MUL VL]\n"
+      "ld1w { z30.s }, p1/Z, [x26, #-7, MUL VL]\n"
+      "ld1w { z29.s }, p1/Z, [x26, #-6, MUL VL]\n"
+      "ld1w { z28.s }, p1/Z, [x26, #-5, MUL VL]\n"
+      "ld1w { z27.s }, p1/Z, [x26, #-4, MUL VL]\n"
+      "ld1w { z26.s }, p1/Z, [x26, #-3, MUL VL]\n"
+      "ld1w { z25.s }, p1/Z, [x26, #-2, MUL VL]\n"
+      "ld1w { z24.s }, p1/Z, [x26, #-1, MUL VL]\n"
+      "ld1w { z23.s }, p1/Z, [x23, #-8, MUL VL]\n"
+      "ld1w { z22.s }, p1/Z, [x23, #-7, MUL VL]\n"
+      "ld1w { z21.s }, p1/Z, [x23, #-6, MUL VL]\n"
+      "ld1w { z20.s }, p1/Z, [x23, #-5, MUL VL]\n"
+      "ld1w { z19.s }, p1/Z, [x23, #-4, MUL VL]\n"
+      "ld1w { z18.s }, p1/Z, [x23, #-3, MUL VL]\n"
+      "ld1w { z17.s }, p1/Z, [x23, #-2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x23, #-1, MUL VL]\n"
+      "st1w { z15.s }, p1, [x21]\n"
+      "st1w { z14.s }, p1, [x21, #1, MUL VL]\n"
+      "st1w { z13.s }, p1, [x21, #2, MUL VL]\n"
+      "st1w { z12.s }, p1, [x21, #3, MUL VL]\n"
+      "st1w { z11.s }, p1, [x21, #4, MUL VL]\n"
+      "st1w { z10.s }, p1, [x21, #5, MUL VL]\n"
+      "st1w { z9.s }, p1, [x21, #6, MUL VL]\n"
+      "st1w { z8.s }, p1, [x21, #7, MUL VL]\n"
       "addvl x21, x21, #16\n"
-      "st1w { z7.s }, p2, [x21, #-8, MUL VL]\n"
-      "st1w { z6.s }, p2, [x21, #-7, MUL VL]\n"
-      "st1w { z5.s }, p2, [x21, #-6, MUL VL]\n"
-      "st1w { z4.s }, p2, [x21, #-5, MUL VL]\n"
-      "st1w { z3.s }, p2, [x21, #-4, MUL VL]\n"
-      "st1w { z2.s }, p2, [x21, #-3, MUL VL]\n"
-      "st1w { z1.s }, p2, [x21, #-2, MUL VL]\n"
-      "st1w { z0.s }, p2, [x21, #-1, MUL VL]\n"
-      "st1w { z31.s }, p2, [x20]\n"
-      "st1w { z30.s }, p2, [x20, #1, MUL VL]\n"
-      "st1w { z29.s }, p2, [x20, #2, MUL VL]\n"
-      "st1w { z28.s }, p2, [x20, #3, MUL VL]\n"
-      "st1w { z27.s }, p2, [x20, #4, MUL VL]\n"
-      "st1w { z26.s }, p2, [x20, #5, MUL VL]\n"
-      "st1w { z25.s }, p2, [x20, #6, MUL VL]\n"
-      "st1w { z24.s }, p2, [x20, #7, MUL VL]\n"
+      "st1w { z7.s }, p1, [x21, #-8, MUL VL]\n"
+      "st1w { z6.s }, p1, [x21, #-7, MUL VL]\n"
+      "st1w { z5.s }, p1, [x21, #-6, MUL VL]\n"
+      "st1w { z4.s }, p1, [x21, #-5, MUL VL]\n"
+      "st1w { z3.s }, p1, [x21, #-4, MUL VL]\n"
+      "st1w { z2.s }, p1, [x21, #-3, MUL VL]\n"
+      "st1w { z1.s }, p1, [x21, #-2, MUL VL]\n"
+      "st1w { z0.s }, p1, [x21, #-1, MUL VL]\n"
+      "st1w { z31.s }, p1, [x20]\n"
+      "st1w { z30.s }, p1, [x20, #1, MUL VL]\n"
+      "st1w { z29.s }, p1, [x20, #2, MUL VL]\n"
+      "st1w { z28.s }, p1, [x20, #3, MUL VL]\n"
+      "st1w { z27.s }, p1, [x20, #4, MUL VL]\n"
+      "st1w { z26.s }, p1, [x20, #5, MUL VL]\n"
+      "st1w { z25.s }, p1, [x20, #6, MUL VL]\n"
+      "st1w { z24.s }, p1, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
-      "st1w { z23.s }, p2, [x20, #-8, MUL VL]\n"
-      "st1w { z22.s }, p2, [x20, #-7, MUL VL]\n"
-      "st1w { z21.s }, p2, [x20, #-6, MUL VL]\n"
-      "st1w { z20.s }, p2, [x20, #-5, MUL VL]\n"
-      "st1w { z19.s }, p2, [x20, #-4, MUL VL]\n"
-      "st1w { z18.s }, p2, [x20, #-3, MUL VL]\n"
-      "st1w { z17.s }, p2, [x20, #-2, MUL VL]\n"
-      "st1w { z16.s }, p2, [x20, #-1, MUL VL]\n"
+      "st1w { z23.s }, p1, [x20, #-8, MUL VL]\n"
+      "st1w { z22.s }, p1, [x20, #-7, MUL VL]\n"
+      "st1w { z21.s }, p1, [x20, #-6, MUL VL]\n"
+      "st1w { z20.s }, p1, [x20, #-5, MUL VL]\n"
+      "st1w { z19.s }, p1, [x20, #-4, MUL VL]\n"
+      "st1w { z18.s }, p1, [x20, #-3, MUL VL]\n"
+      "st1w { z17.s }, p1, [x20, #-2, MUL VL]\n"
+      "st1w { z16.s }, p1, [x20, #-1, MUL VL]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cbz x25, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x21, x25\n"
-      "mov x20, x23\n"
-      "decw x25, ALL, MUL #8\n"
-      "add x23, x23, %x[out_stride]\n"
-      "whilelt p1.s, XZR, x21\n"
-      "decw x21\n"
-      "whilelt p0.s, XZR, x21\n"
-      "decw x21\n"
-      "ld1w { z31.s }, p1/Z, [x26]\n"
-      "ld1w { z30.s }, p1/Z, [x22]\n"
-      "whilelt p1.s, XZR, x21\n"
-      "decw x21\n"
+      "mov x20, x25\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z31.s }, p0/Z, [x26]\n"
+      "ld1w { z30.s }, p0/Z, [x23]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
       "ld1w { z29.s }, p0/Z, [x26, #1, MUL VL]\n"
-      "ld1w { z28.s }, p0/Z, [x22, #1, MUL VL]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "decw x21\n"
-      "ld1w { z27.s }, p1/Z, [x26, #2, MUL VL]\n"
-      "ld1w { z26.s }, p1/Z, [x22, #2, MUL VL]\n"
-      "whilelt p1.s, XZR, x21\n"
-      "decw x21\n"
+      "ld1w { z28.s }, p0/Z, [x23, #1, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z27.s }, p0/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z26.s }, p0/Z, [x23, #2, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
       "ld1w { z25.s }, p0/Z, [x26, #3, MUL VL]\n"
-      "ld1w { z24.s }, p0/Z, [x22, #3, MUL VL]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "decw x21\n"
-      "ld1w { z23.s }, p1/Z, [x26, #4, MUL VL]\n"
-      "ld1w { z22.s }, p1/Z, [x22, #4, MUL VL]\n"
-      "whilelt p1.s, XZR, x21\n"
-      "decw x21\n"
+      "ld1w { z24.s }, p0/Z, [x23, #3, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z23.s }, p0/Z, [x26, #4, MUL VL]\n"
+      "ld1w { z22.s }, p0/Z, [x23, #4, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
       "ld1w { z21.s }, p0/Z, [x26, #5, MUL VL]\n"
-      "ld1w { z20.s }, p0/Z, [x22, #5, MUL VL]\n"
-      "whilelt p0.s, XZR, x21\n"
-      "cmp x25, #0x0\n"
-      "ld1w { z19.s }, p1/Z, [x26, #6, MUL VL]\n"
-      "ld1w { z18.s }, p1/Z, [x22, #6, MUL VL]\n"
+      "ld1w { z20.s }, p0/Z, [x23, #5, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z19.s }, p0/Z, [x26, #6, MUL VL]\n"
+      "ld1w { z18.s }, p0/Z, [x23, #6, MUL VL]\n"
+      "decw x20\n"
+      "whilelt p0.s, XZR, x20\n"
       "ld1w { z17.s }, p0/Z, [x26, #7, MUL VL]\n"
+      "ld1w { z16.s }, p0/Z, [x23, #7, MUL VL]\n"
+      "mov x20, x22\n"
+      "decw x25, ALL, MUL #8\n"
+      "st1w { z31.s }, p1, [x20]\n"
+      "st1w { z29.s }, p1, [x20, #1, MUL VL]\n"
+      "cmp x25, #0x0\n"
       "addvl x26, x26, #8\n"
-      "ld1w { z16.s }, p0/Z, [x22, #7, MUL VL]\n"
-      "st1w { z31.s }, p2, [x20]\n"
-      "st1w { z29.s }, p2, [x20, #1, MUL VL]\n"
-      "addvl x22, x22, #8\n"
-      "st1w { z27.s }, p2, [x20, #2, MUL VL]\n"
-      "st1w { z25.s }, p2, [x20, #3, MUL VL]\n"
-      "st1w { z23.s }, p2, [x20, #4, MUL VL]\n"
-      "st1w { z21.s }, p2, [x20, #5, MUL VL]\n"
-      "st1w { z19.s }, p2, [x20, #6, MUL VL]\n"
-      "st1w { z17.s }, p2, [x20, #7, MUL VL]\n"
+      "st1w { z27.s }, p1, [x20, #2, MUL VL]\n"
+      "addvl x23, x23, #8\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1w { z25.s }, p1, [x20, #3, MUL VL]\n"
+      "st1w { z23.s }, p1, [x20, #4, MUL VL]\n"
+      "st1w { z21.s }, p1, [x20, #5, MUL VL]\n"
+      "st1w { z19.s }, p1, [x20, #6, MUL VL]\n"
+      "st1w { z17.s }, p1, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
-      "st1w { z30.s }, p2, [x20, #-8, MUL VL]\n"
-      "st1w { z28.s }, p2, [x20, #-7, MUL VL]\n"
-      "st1w { z26.s }, p2, [x20, #-6, MUL VL]\n"
-      "st1w { z24.s }, p2, [x20, #-5, MUL VL]\n"
-      "st1w { z22.s }, p2, [x20, #-4, MUL VL]\n"
-      "st1w { z20.s }, p2, [x20, #-3, MUL VL]\n"
-      "st1w { z18.s }, p2, [x20, #-2, MUL VL]\n"
-      "st1w { z16.s }, p2, [x20, #-1, MUL VL]\n"
+      "st1w { z30.s }, p1, [x20, #-8, MUL VL]\n"
+      "st1w { z28.s }, p1, [x20, #-7, MUL VL]\n"
+      "st1w { z26.s }, p1, [x20, #-6, MUL VL]\n"
+      "st1w { z24.s }, p1, [x20, #-5, MUL VL]\n"
+      "st1w { z22.s }, p1, [x20, #-4, MUL VL]\n"
+      "st1w { z20.s }, p1, [x20, #-3, MUL VL]\n"
+      "st1w { z18.s }, p1, [x20, #-2, MUL VL]\n"
+      "st1w { z16.s }, p1, [x20, #-1, MUL VL]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x2\n"
@@ -191,89 +191,89 @@ void sve_transpose_interleave_8VL(uint32_t *out, const uint32_t *in, size_t widt
       "mov x21, %x[width]\n"
       "cntw x20, ALL, MUL #16\n"
       "mov x26, %x[in]\n"
-      "mov x23, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
       "cmp x21, x20\n"
       "add %x[in], x26, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1w { z31.s }, p2/Z, [x26]\n"
-      "ld1w { z30.s }, p2/Z, [x26, #1, MUL VL]\n"
+      "ld1w { z31.s }, p1/Z, [x26]\n"
+      "ld1w { z30.s }, p1/Z, [x26, #1, MUL VL]\n"
       "sub x21, x21, x20\n"
-      "ld1w { z29.s }, p2/Z, [x26, #2, MUL VL]\n"
-      "ld1w { z28.s }, p2/Z, [x26, #3, MUL VL]\n"
       "cmp x21, x20\n"
-      "ld1w { z27.s }, p2/Z, [x26, #4, MUL VL]\n"
-      "ld1w { z26.s }, p2/Z, [x26, #5, MUL VL]\n"
-      "ld1w { z25.s }, p2/Z, [x26, #6, MUL VL]\n"
-      "ld1w { z24.s }, p2/Z, [x26, #7, MUL VL]\n"
+      "ld1w { z29.s }, p1/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z28.s }, p1/Z, [x26, #3, MUL VL]\n"
+      "ld1w { z27.s }, p1/Z, [x26, #4, MUL VL]\n"
+      "ld1w { z26.s }, p1/Z, [x26, #5, MUL VL]\n"
+      "ld1w { z25.s }, p1/Z, [x26, #6, MUL VL]\n"
+      "ld1w { z24.s }, p1/Z, [x26, #7, MUL VL]\n"
       "addvl x26, x26, #16\n"
-      "ld1w { z23.s }, p2/Z, [x26, #-8, MUL VL]\n"
-      "ld1w { z22.s }, p2/Z, [x26, #-7, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x26, #-6, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x26, #-5, MUL VL]\n"
-      "ld1w { z19.s }, p2/Z, [x26, #-4, MUL VL]\n"
-      "ld1w { z18.s }, p2/Z, [x26, #-3, MUL VL]\n"
-      "ld1w { z17.s }, p2/Z, [x26, #-2, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x26, #-1, MUL VL]\n"
-      "st1w { z31.s }, p2, [x23]\n"
-      "st1w { z30.s }, p2, [x23, #1, MUL VL]\n"
-      "st1w { z29.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z28.s }, p2, [x23, #3, MUL VL]\n"
-      "st1w { z27.s }, p2, [x23, #4, MUL VL]\n"
-      "st1w { z26.s }, p2, [x23, #5, MUL VL]\n"
-      "st1w { z25.s }, p2, [x23, #6, MUL VL]\n"
-      "st1w { z24.s }, p2, [x23, #7, MUL VL]\n"
-      "add x23, x23, %x[out_stride]\n"
-      "st1w { z23.s }, p2, [x23]\n"
-      "st1w { z22.s }, p2, [x23, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z20.s }, p2, [x23, #3, MUL VL]\n"
-      "st1w { z19.s }, p2, [x23, #4, MUL VL]\n"
-      "st1w { z18.s }, p2, [x23, #5, MUL VL]\n"
-      "st1w { z17.s }, p2, [x23, #6, MUL VL]\n"
-      "st1w { z16.s }, p2, [x23, #7, MUL VL]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "ld1w { z23.s }, p1/Z, [x26, #-8, MUL VL]\n"
+      "ld1w { z22.s }, p1/Z, [x26, #-7, MUL VL]\n"
+      "ld1w { z21.s }, p1/Z, [x26, #-6, MUL VL]\n"
+      "ld1w { z20.s }, p1/Z, [x26, #-5, MUL VL]\n"
+      "ld1w { z19.s }, p1/Z, [x26, #-4, MUL VL]\n"
+      "ld1w { z18.s }, p1/Z, [x26, #-3, MUL VL]\n"
+      "ld1w { z17.s }, p1/Z, [x26, #-2, MUL VL]\n"
+      "ld1w { z16.s }, p1/Z, [x26, #-1, MUL VL]\n"
+      "st1w { z31.s }, p1, [x22]\n"
+      "st1w { z30.s }, p1, [x22, #1, MUL VL]\n"
+      "st1w { z29.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z28.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z27.s }, p1, [x22, #4, MUL VL]\n"
+      "st1w { z26.s }, p1, [x22, #5, MUL VL]\n"
+      "st1w { z25.s }, p1, [x22, #6, MUL VL]\n"
+      "st1w { z24.s }, p1, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1w { z23.s }, p1, [x22]\n"
+      "st1w { z22.s }, p1, [x22, #1, MUL VL]\n"
+      "st1w { z21.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z20.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z19.s }, p1, [x22, #4, MUL VL]\n"
+      "st1w { z18.s }, p1, [x22, #5, MUL VL]\n"
+      "st1w { z17.s }, p1, [x22, #6, MUL VL]\n"
+      "st1w { z16.s }, p1, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decw x21, ALL, MUL #8\n"
-      "whilelt p1.s, XZR, x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z23.s }, p0/Z, [x26]\n"
       "decw x20\n"
       "whilelt p0.s, XZR, x20\n"
+      "ld1w { z22.s }, p0/Z, [x26, #1, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z23.s }, p1/Z, [x26]\n"
-      "whilelt p1.s, XZR, x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z21.s }, p0/Z, [x26, #2, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z22.s }, p0/Z, [x26, #1, MUL VL]\n"
       "whilelt p0.s, XZR, x20\n"
+      "ld1w { z20.s }, p0/Z, [x26, #3, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z21.s }, p1/Z, [x26, #2, MUL VL]\n"
-      "whilelt p1.s, XZR, x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z19.s }, p0/Z, [x26, #4, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z20.s }, p0/Z, [x26, #3, MUL VL]\n"
       "whilelt p0.s, XZR, x20\n"
+      "ld1w { z18.s }, p0/Z, [x26, #5, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z19.s }, p1/Z, [x26, #4, MUL VL]\n"
-      "whilelt p1.s, XZR, x20\n"
+      "whilelt p0.s, XZR, x20\n"
+      "ld1w { z17.s }, p0/Z, [x26, #6, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z18.s }, p0/Z, [x26, #5, MUL VL]\n"
+      "decw x21, ALL, MUL #8\n"
       "whilelt p0.s, XZR, x20\n"
       "cmp x21, #0x0\n"
-      "ld1w { z17.s }, p1/Z, [x26, #6, MUL VL]\n"
       "ld1w { z16.s }, p0/Z, [x26, #7, MUL VL]\n"
+      "st1w { z23.s }, p1, [x22]\n"
       "addvl x26, x26, #8\n"
-      "st1w { z23.s }, p2, [x23]\n"
-      "st1w { z22.s }, p2, [x23, #1, MUL VL]\n"
-      "st1w { z21.s }, p2, [x23, #2, MUL VL]\n"
-      "st1w { z20.s }, p2, [x23, #3, MUL VL]\n"
-      "st1w { z19.s }, p2, [x23, #4, MUL VL]\n"
-      "st1w { z18.s }, p2, [x23, #5, MUL VL]\n"
-      "st1w { z17.s }, p2, [x23, #6, MUL VL]\n"
-      "st1w { z16.s }, p2, [x23, #7, MUL VL]\n"
-      "add x23, x23, %x[out_stride]\n"
+      "st1w { z22.s }, p1, [x22, #1, MUL VL]\n"
+      "st1w { z21.s }, p1, [x22, #2, MUL VL]\n"
+      "st1w { z20.s }, p1, [x22, #3, MUL VL]\n"
+      "st1w { z19.s }, p1, [x22, #4, MUL VL]\n"
+      "st1w { z18.s }, p1, [x22, #5, MUL VL]\n"
+      "st1w { z17.s }, p1, [x22, #6, MUL VL]\n"
+      "st1w { z16.s }, p1, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -282,7 +282,7 @@ void sve_transpose_interleave_8VL(uint32_t *out, const uint32_t *in, size_t widt
       "12:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_1x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_1x4.hpp
index 03d192c874..35e60a223c 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_1x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_1x4.hpp
@@ -42,207 +42,207 @@ void sve_transpose_interleave_8VL_1x4(uint8_t *out, const uint8_t *in, size_t wi
       "ptrue p2.b\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
       "cntb x20, ALL, MUL #8\n"
+      "add x22, x24, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x24, %x[out]\n"
-      "add x23, x26, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add %x[in], x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GT\n"
-      "csel x22, x22, %x[pad_row], GE\n"
+      "add %x[in], x22, %x[in_stride]\n"
+      "csel x22, x22, %x[pad_row], GT\n"
+      "csel x24, x24, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x23, x23, %x[pad_row], GT\n"
-      "cmp x25, x20\n"
+      "csel x25, x25, %x[pad_row], GT\n"
+      "cmp x23, x20\n"
+      "mov x21, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1b { z22.b }, p2/Z, [x26]\n"
-      "ld1b { z14.b }, p2/Z, [x26, #1, MUL VL]\n"
-      "sub x25, x25, x20\n"
-      "ld1b { z2.b }, p2/Z, [x23]\n"
-      "ld1b { z10.b }, p2/Z, [x23, #1, MUL VL]\n"
-      "cmp x25, x20\n"
-      "ld1b { z27.b }, p2/Z, [x22]\n"
-      "ld1b { z21.b }, p2/Z, [x22, #1, MUL VL]\n"
-      "ld1b { z13.b }, p2/Z, [x21]\n"
-      "ld1b { z28.b }, p2/Z, [x21, #1, MUL VL]\n"
-      "ld1b { z11.b }, p2/Z, [x26, #2, MUL VL]\n"
-      "ld1b { z9.b }, p2/Z, [x26, #3, MUL VL]\n"
-      "ld1b { z23.b }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x23, #3, MUL VL]\n"
-      "zip1 z24.b, z22.b, z27.b\n"
-      "zip2 z19.b, z22.b, z27.b\n"
-      "ld1b { z5.b }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1b { z3.b }, p2/Z, [x22, #3, MUL VL]\n"
-      "zip1 z6.b, z2.b, z13.b\n"
-      "zip2 z20.b, z2.b, z13.b\n"
-      "ld1b { z25.b }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1b { z30.b }, p2/Z, [x21, #3, MUL VL]\n"
-      "zip1 z18.b, z14.b, z21.b\n"
-      "zip1 z16.b, z10.b, z28.b\n"
-      "ld1b { z26.b }, p2/Z, [x26, #4, MUL VL]\n"
-      "ld1b { z22.b }, p2/Z, [x26, #5, MUL VL]\n"
-      "zip2 z17.b, z14.b, z21.b\n"
-      "zip2 z10.b, z10.b, z28.b\n"
-      "ld1b { z27.b }, p2/Z, [x23, #4, MUL VL]\n"
-      "ld1b { z7.b }, p2/Z, [x23, #5, MUL VL]\n"
-      "zip1 z21.b, z11.b, z5.b\n"
-      "zip2 z12.b, z11.b, z5.b\n"
-      "ld1b { z2.b }, p2/Z, [x22, #4, MUL VL]\n"
-      "ld1b { z5.b }, p2/Z, [x22, #5, MUL VL]\n"
-      "zip1 z13.b, z23.b, z25.b\n"
-      "zip2 z31.b, z23.b, z25.b\n"
-      "ld1b { z25.b }, p2/Z, [x21, #4, MUL VL]\n"
-      "ld1b { z8.b }, p2/Z, [x21, #5, MUL VL]\n"
-      "zip1 z15.b, z9.b, z3.b\n"
-      "zip1 z11.b, z29.b, z30.b\n"
-      "ld1b { z14.b }, p2/Z, [x26, #6, MUL VL]\n"
-      "ld1b { z1.b }, p2/Z, [x26, #7, MUL VL]\n"
-      "zip2 z23.b, z9.b, z3.b\n"
-      "zip2 z0.b, z29.b, z30.b\n"
-      "ld1b { z3.b }, p2/Z, [x23, #6, MUL VL]\n"
-      "ld1b { z29.b }, p2/Z, [x23, #7, MUL VL]\n"
-      "zip1 z9.b, z26.b, z2.b\n"
-      "zip2 z26.b, z26.b, z2.b\n"
-      "ld1b { z2.b }, p2/Z, [x22, #6, MUL VL]\n"
-      "ld1b { z4.b }, p2/Z, [x22, #7, MUL VL]\n"
-      "zip1 z30.b, z27.b, z25.b\n"
-      "zip2 z28.b, z27.b, z25.b\n"
-      "ld1b { z27.b }, p2/Z, [x21, #6, MUL VL]\n"
-      "zip1 z25.b, z22.b, z5.b\n"
-      "zip2 z5.b, z22.b, z5.b\n"
+      "ld1b { z7.b }, p2/Z, [x26]\n"
+      "ld1b { z24.b }, p2/Z, [x26, #1, MUL VL]\n"
+      "sub x23, x23, x20\n"
+      "cmp x23, x20\n"
+      "ld1b { z31.b }, p2/Z, [x25]\n"
+      "ld1b { z18.b }, p2/Z, [x25, #1, MUL VL]\n"
+      "ld1b { z19.b }, p2/Z, [x24]\n"
+      "ld1b { z25.b }, p2/Z, [x24, #1, MUL VL]\n"
+      "zip1 z23.b, z7.b, z19.b\n"
+      "zip2 z20.b, z7.b, z19.b\n"
+      "ld1b { z30.b }, p2/Z, [x22]\n"
+      "ld1b { z3.b }, p2/Z, [x22, #1, MUL VL]\n"
+      "zip1 z21.b, z31.b, z30.b\n"
+      "zip2 z19.b, z31.b, z30.b\n"
+      "ld1b { z16.b }, p2/Z, [x26, #2, MUL VL]\n"
+      "ld1b { z30.b }, p2/Z, [x26, #3, MUL VL]\n"
+      "zip1 z2.b, z24.b, z25.b\n"
+      "zip1 z17.b, z18.b, z3.b\n"
+      "ld1b { z29.b }, p2/Z, [x25, #2, MUL VL]\n"
+      "ld1b { z8.b }, p2/Z, [x25, #3, MUL VL]\n"
+      "zip2 z22.b, z24.b, z25.b\n"
+      "zip2 z4.b, z18.b, z3.b\n"
+      "ld1b { z0.b }, p2/Z, [x24, #2, MUL VL]\n"
+      "ld1b { z3.b }, p2/Z, [x24, #3, MUL VL]\n"
+      "zip1 z9.b, z16.b, z0.b\n"
+      "zip2 z14.b, z16.b, z0.b\n"
+      "ld1b { z18.b }, p2/Z, [x22, #2, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x22, #3, MUL VL]\n"
+      "zip1 z24.b, z29.b, z18.b\n"
+      "zip2 z11.b, z29.b, z18.b\n"
+      "ld1b { z1.b }, p2/Z, [x26, #4, MUL VL]\n"
+      "ld1b { z12.b }, p2/Z, [x26, #5, MUL VL]\n"
+      "zip1 z13.b, z30.b, z3.b\n"
+      "zip1 z15.b, z8.b, z16.b\n"
+      "ld1b { z5.b }, p2/Z, [x25, #4, MUL VL]\n"
+      "ld1b { z29.b }, p2/Z, [x25, #5, MUL VL]\n"
+      "zip2 z31.b, z30.b, z3.b\n"
+      "zip2 z30.b, z8.b, z16.b\n"
+      "ld1b { z16.b }, p2/Z, [x24, #4, MUL VL]\n"
+      "ld1b { z18.b }, p2/Z, [x24, #5, MUL VL]\n"
+      "zip1 z27.b, z1.b, z16.b\n"
+      "zip2 z10.b, z1.b, z16.b\n"
+      "ld1b { z7.b }, p2/Z, [x22, #4, MUL VL]\n"
+      "ld1b { z16.b }, p2/Z, [x22, #5, MUL VL]\n"
+      "zip1 z8.b, z5.b, z7.b\n"
+      "zip2 z26.b, z5.b, z7.b\n"
+      "ld1b { z3.b }, p2/Z, [x26, #6, MUL VL]\n"
+      "ld1b { z25.b }, p2/Z, [x26, #7, MUL VL]\n"
+      "zip1 z6.b, z12.b, z18.b\n"
+      "zip1 z5.b, z29.b, z16.b\n"
+      "ld1b { z0.b }, p2/Z, [x25, #6, MUL VL]\n"
+      "ld1b { z28.b }, p2/Z, [x25, #7, MUL VL]\n"
+      "zip2 z12.b, z12.b, z18.b\n"
+      "zip2 z7.b, z29.b, z16.b\n"
+      "ld1b { z1.b }, p2/Z, [x24, #6, MUL VL]\n"
+      "ld1b { z29.b }, p2/Z, [x24, #7, MUL VL]\n"
+      "zip1 z16.b, z23.b, z21.b\n"
+      "zip2 z18.b, z23.b, z21.b\n"
+      "ld1b { z23.b }, p2/Z, [x22, #6, MUL VL]\n"
+      "ld1b { z21.b }, p2/Z, [x22, #7, MUL VL]\n"
+      "st1b { z16.b }, p2, [x21]\n"
+      "zip1 z16.b, z20.b, z19.b\n"
+      "zip2 z20.b, z20.b, z19.b\n"
+      "zip1 z19.b, z2.b, z17.b\n"
+      "st1b { z18.b }, p2, [x21, #1, MUL VL]\n"
       "addvl x26, x26, #8\n"
-      "zip1 z22.b, z7.b, z8.b\n"
-      "zip2 z7.b, z7.b, z8.b\n"
-      "addvl x23, x23, #8\n"
+      "zip2 z18.b, z2.b, z17.b\n"
+      "zip1 z17.b, z22.b, z4.b\n"
+      "st1b { z16.b }, p2, [x21, #2, MUL VL]\n"
+      "addvl x25, x25, #8\n"
+      "zip2 z16.b, z22.b, z4.b\n"
+      "st1b { z20.b }, p2, [x21, #3, MUL VL]\n"
+      "zip1 z4.b, z3.b, z1.b\n"
+      "addvl x24, x24, #8\n"
+      "st1b { z19.b }, p2, [x21, #4, MUL VL]\n"
+      "zip1 z22.b, z0.b, z23.b\n"
+      "zip2 z3.b, z3.b, z1.b\n"
       "addvl x22, x22, #8\n"
-      "zip1 z8.b, z24.b, z6.b\n"
-      "zip2 z6.b, z24.b, z6.b\n"
-      "ld1b { z24.b }, p2/Z, [x21, #7, MUL VL]\n"
-      "addvl x21, x21, #8\n"
-      "st1b { z8.b }, p2, [x24]\n"
-      "zip1 z8.b, z19.b, z20.b\n"
-      "zip2 z20.b, z19.b, z20.b\n"
-      "zip1 z19.b, z18.b, z16.b\n"
-      "zip2 z18.b, z18.b, z16.b\n"
-      "zip1 z16.b, z17.b, z10.b\n"
-      "zip2 z17.b, z17.b, z10.b\n"
-      "st1b { z6.b }, p2, [x24, #1, MUL VL]\n"
-      "st1b { z8.b }, p2, [x24, #2, MUL VL]\n"
-      "zip1 z8.b, z14.b, z2.b\n"
-      "zip1 z6.b, z3.b, z27.b\n"
-      "st1b { z20.b }, p2, [x24, #3, MUL VL]\n"
-      "zip2 z10.b, z14.b, z2.b\n"
-      "zip2 z14.b, z3.b, z27.b\n"
-      "st1b { z19.b }, p2, [x24, #4, MUL VL]\n"
-      "zip1 z2.b, z1.b, z4.b\n"
-      "zip1 z3.b, z29.b, z24.b\n"
-      "st1b { z18.b }, p2, [x24, #5, MUL VL]\n"
-      "zip2 z1.b, z1.b, z4.b\n"
-      "zip2 z4.b, z29.b, z24.b\n"
-      "st1b { z16.b }, p2, [x24, #6, MUL VL]\n"
-      "zip1 z24.b, z21.b, z13.b\n"
-      "zip2 z16.b, z21.b, z13.b\n"
-      "st1b { z17.b }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
-      "zip1 z29.b, z12.b, z31.b\n"
-      "zip2 z21.b, z12.b, z31.b\n"
-      "zip1 z20.b, z15.b, z11.b\n"
-      "zip2 z19.b, z15.b, z11.b\n"
-      "zip1 z18.b, z23.b, z0.b\n"
-      "zip2 z17.b, z23.b, z0.b\n"
-      "st1b { z24.b }, p2, [x24]\n"
-      "st1b { z16.b }, p2, [x24, #1, MUL VL]\n"
-      "zip1 z16.b, z9.b, z30.b\n"
-      "zip2 z30.b, z9.b, z30.b\n"
-      "st1b { z29.b }, p2, [x24, #2, MUL VL]\n"
-      "zip1 z29.b, z26.b, z28.b\n"
-      "zip2 z28.b, z26.b, z28.b\n"
-      "st1b { z21.b }, p2, [x24, #3, MUL VL]\n"
-      "zip1 z27.b, z25.b, z22.b\n"
-      "zip2 z26.b, z25.b, z22.b\n"
-      "st1b { z20.b }, p2, [x24, #4, MUL VL]\n"
-      "zip1 z25.b, z5.b, z7.b\n"
-      "zip2 z24.b, z5.b, z7.b\n"
-      "st1b { z19.b }, p2, [x24, #5, MUL VL]\n"
-      "zip1 z23.b, z8.b, z6.b\n"
-      "zip2 z22.b, z8.b, z6.b\n"
-      "st1b { z18.b }, p2, [x24, #6, MUL VL]\n"
-      "zip1 z21.b, z10.b, z14.b\n"
-      "zip2 z20.b, z10.b, z14.b\n"
-      "st1b { z17.b }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
-      "zip1 z19.b, z2.b, z3.b\n"
-      "zip2 z18.b, z2.b, z3.b\n"
-      "st1b { z16.b }, p2, [x24]\n"
-      "zip1 z17.b, z1.b, z4.b\n"
-      "zip2 z16.b, z1.b, z4.b\n"
-      "st1b { z30.b }, p2, [x24, #1, MUL VL]\n"
-      "st1b { z29.b }, p2, [x24, #2, MUL VL]\n"
-      "st1b { z28.b }, p2, [x24, #3, MUL VL]\n"
-      "st1b { z27.b }, p2, [x24, #4, MUL VL]\n"
-      "st1b { z26.b }, p2, [x24, #5, MUL VL]\n"
-      "st1b { z25.b }, p2, [x24, #6, MUL VL]\n"
-      "st1b { z24.b }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
-      "st1b { z23.b }, p2, [x24]\n"
-      "st1b { z22.b }, p2, [x24, #1, MUL VL]\n"
-      "st1b { z21.b }, p2, [x24, #2, MUL VL]\n"
-      "st1b { z20.b }, p2, [x24, #3, MUL VL]\n"
-      "st1b { z19.b }, p2, [x24, #4, MUL VL]\n"
-      "st1b { z18.b }, p2, [x24, #5, MUL VL]\n"
-      "st1b { z17.b }, p2, [x24, #6, MUL VL]\n"
-      "st1b { z16.b }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
+      "st1b { z18.b }, p2, [x21, #5, MUL VL]\n"
+      "zip2 z2.b, z0.b, z23.b\n"
+      "zip1 z1.b, z25.b, z29.b\n"
+      "st1b { z17.b }, p2, [x21, #6, MUL VL]\n"
+      "zip1 z0.b, z28.b, z21.b\n"
+      "zip2 z29.b, z25.b, z29.b\n"
+      "st1b { z16.b }, p2, [x21, #7, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip2 z28.b, z28.b, z21.b\n"
+      "zip1 z17.b, z9.b, z24.b\n"
+      "zip2 z16.b, z9.b, z24.b\n"
+      "zip1 z19.b, z14.b, z11.b\n"
+      "st1b { z17.b }, p2, [x21]\n"
+      "zip2 z18.b, z14.b, z11.b\n"
+      "zip1 z17.b, z13.b, z15.b\n"
+      "st1b { z16.b }, p2, [x21, #1, MUL VL]\n"
+      "zip2 z16.b, z13.b, z15.b\n"
+      "zip1 z21.b, z31.b, z30.b\n"
+      "st1b { z19.b }, p2, [x21, #2, MUL VL]\n"
+      "zip2 z20.b, z31.b, z30.b\n"
+      "st1b { z18.b }, p2, [x21, #3, MUL VL]\n"
+      "zip1 z19.b, z27.b, z8.b\n"
+      "st1b { z17.b }, p2, [x21, #4, MUL VL]\n"
+      "zip2 z18.b, z27.b, z8.b\n"
+      "zip1 z17.b, z10.b, z26.b\n"
+      "st1b { z16.b }, p2, [x21, #5, MUL VL]\n"
+      "zip2 z16.b, z10.b, z26.b\n"
+      "zip1 z27.b, z6.b, z5.b\n"
+      "st1b { z21.b }, p2, [x21, #6, MUL VL]\n"
+      "zip2 z26.b, z6.b, z5.b\n"
+      "zip1 z25.b, z12.b, z7.b\n"
+      "st1b { z20.b }, p2, [x21, #7, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "zip2 z24.b, z12.b, z7.b\n"
+      "zip1 z23.b, z4.b, z22.b\n"
+      "st1b { z19.b }, p2, [x21]\n"
+      "zip2 z22.b, z4.b, z22.b\n"
+      "zip1 z21.b, z3.b, z2.b\n"
+      "st1b { z18.b }, p2, [x21, #1, MUL VL]\n"
+      "zip2 z20.b, z3.b, z2.b\n"
+      "zip1 z19.b, z1.b, z0.b\n"
+      "st1b { z17.b }, p2, [x21, #2, MUL VL]\n"
+      "zip2 z18.b, z1.b, z0.b\n"
+      "zip1 z17.b, z29.b, z28.b\n"
+      "st1b { z16.b }, p2, [x21, #3, MUL VL]\n"
+      "zip2 z16.b, z29.b, z28.b\n"
+      "st1b { z27.b }, p2, [x21, #4, MUL VL]\n"
+      "st1b { z26.b }, p2, [x21, #5, MUL VL]\n"
+      "st1b { z25.b }, p2, [x21, #6, MUL VL]\n"
+      "st1b { z24.b }, p2, [x21, #7, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
+      "st1b { z23.b }, p2, [x21]\n"
+      "st1b { z22.b }, p2, [x21, #1, MUL VL]\n"
+      "st1b { z21.b }, p2, [x21, #2, MUL VL]\n"
+      "st1b { z20.b }, p2, [x21, #3, MUL VL]\n"
+      "st1b { z19.b }, p2, [x21, #4, MUL VL]\n"
+      "st1b { z18.b }, p2, [x21, #5, MUL VL]\n"
+      "st1b { z17.b }, p2, [x21, #6, MUL VL]\n"
+      "st1b { z16.b }, p2, [x21, #7, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x20, x25\n"
-      "decw x25, ALL, MUL #8\n"
+      "mov x20, x23\n"
       "whilelt p1.b, XZR, x20\n"
+      "ld1b { z23.b }, p1/Z, [x26]\n"
+      "ld1b { z22.b }, p1/Z, [x25]\n"
       "decb x20\n"
       "whilelt p0.b, XZR, x20\n"
-      "cmp x25, #0x0\n"
-      "ld1b { z20.b }, p1/Z, [x26]\n"
-      "ld1b { z19.b }, p1/Z, [x23]\n"
+      "ld1b { z21.b }, p0/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z25.b }, p0/Z, [x25, #1, MUL VL]\n"
+      "ld1b { z19.b }, p1/Z, [x24]\n"
+      "ld1b { z20.b }, p0/Z, [x24, #1, MUL VL]\n"
+      "decw x23, ALL, MUL #8\n"
+      "zip1 z24.b, z23.b, z19.b\n"
       "ld1b { z18.b }, p1/Z, [x22]\n"
-      "ld1b { z17.b }, p1/Z, [x21]\n"
-      "ld1b { z24.b }, p0/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z16.b }, p0/Z, [x22, #1, MUL VL]\n"
+      "zip1 z17.b, z22.b, z18.b\n"
+      "zip2 z23.b, z23.b, z19.b\n"
+      "zip2 z19.b, z22.b, z18.b\n"
+      "zip1 z22.b, z21.b, z20.b\n"
+      "cmp x23, #0x0\n"
       "addvl x26, x26, #2\n"
-      "ld1b { z25.b }, p0/Z, [x23, #1, MUL VL]\n"
-      "addvl x23, x23, #2\n"
-      "ld1b { z23.b }, p0/Z, [x22, #1, MUL VL]\n"
-      "ld1b { z16.b }, p0/Z, [x21, #1, MUL VL]\n"
+      "zip1 z18.b, z25.b, z16.b\n"
+      "zip2 z21.b, z21.b, z20.b\n"
+      "addvl x25, x25, #2\n"
+      "addvl x24, x24, #2\n"
+      "zip2 z20.b, z25.b, z16.b\n"
       "addvl x22, x22, #2\n"
-      "addvl x21, x21, #2\n"
-      "zip1 z22.b, z20.b, z18.b\n"
-      "zip1 z21.b, z19.b, z17.b\n"
-      "zip2 z20.b, z20.b, z18.b\n"
-      "zip2 z19.b, z19.b, z17.b\n"
-      "zip1 z18.b, z24.b, z23.b\n"
-      "zip1 z17.b, z25.b, z16.b\n"
-      "zip2 z24.b, z24.b, z23.b\n"
-      "zip2 z16.b, z25.b, z16.b\n"
-      "zip1 z23.b, z22.b, z21.b\n"
-      "zip2 z22.b, z22.b, z21.b\n"
-      "zip1 z21.b, z20.b, z19.b\n"
-      "zip2 z20.b, z20.b, z19.b\n"
-      "zip1 z19.b, z18.b, z17.b\n"
-      "zip2 z18.b, z18.b, z17.b\n"
-      "zip1 z17.b, z24.b, z16.b\n"
-      "zip2 z16.b, z24.b, z16.b\n"
-      "st1b { z23.b }, p2, [x24]\n"
-      "st1b { z22.b }, p2, [x24, #1, MUL VL]\n"
-      "st1b { z21.b }, p2, [x24, #2, MUL VL]\n"
-      "st1b { z20.b }, p2, [x24, #3, MUL VL]\n"
-      "st1b { z19.b }, p2, [x24, #4, MUL VL]\n"
-      "st1b { z18.b }, p2, [x24, #5, MUL VL]\n"
-      "st1b { z17.b }, p2, [x24, #6, MUL VL]\n"
-      "st1b { z16.b }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
+      "zip1 z16.b, z24.b, z17.b\n"
+      "st1b { z16.b }, p2, [x21]\n"
+      "zip2 z16.b, z24.b, z17.b\n"
+      "zip1 z17.b, z23.b, z19.b\n"
+      "st1b { z16.b }, p2, [x21, #1, MUL VL]\n"
+      "zip2 z16.b, z23.b, z19.b\n"
+      "zip1 z19.b, z22.b, z18.b\n"
+      "st1b { z17.b }, p2, [x21, #2, MUL VL]\n"
+      "zip2 z18.b, z22.b, z18.b\n"
+      "zip1 z17.b, z21.b, z20.b\n"
+      "st1b { z16.b }, p2, [x21, #3, MUL VL]\n"
+      "zip2 z16.b, z21.b, z20.b\n"
+      "st1b { z19.b }, p2, [x21, #4, MUL VL]\n"
+      "st1b { z18.b }, p2, [x21, #5, MUL VL]\n"
+      "st1b { z17.b }, p2, [x21, #6, MUL VL]\n"
+      "st1b { z16.b }, p2, [x21, #7, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_1x8.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_1x8.hpp
index 091c0e526b..3e20a5882e 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_1x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_1x8.hpp
@@ -42,180 +42,180 @@ void sve_transpose_interleave_8VL_1x8(uint8_t *out, const uint8_t *in, size_t wi
       "ptrue p1.b\n"
       "1:"  // Main row loop: Head
       "mov x10, %x[in]\n"
-      "cmp %x[height], #0x7\n"
-      "mov x9, %x[width]\n"
-      "cntb x28, ALL, MUL #2\n"
-      "mov x27, %x[out]\n"
-      "add x26, x10, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "add x27, x28, %x[in_stride]\n"
+      "add x26, x27, %x[in_stride]\n"
       "add x25, x26, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add x20, x21, %x[in_stride]\n"
-      "add %x[in], x20, %x[in_stride]\n"
-      "csel x20, x20, %x[pad_row], GT\n"
-      "csel x21, x21, %x[pad_row], GE\n"
+      "cmp %x[height], #0x7\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "csel x23, x23, %x[pad_row], GT\n"
+      "csel x24, x24, %x[pad_row], GE\n"
       "cmp %x[height], #0x5\n"
-      "csel x22, x22, %x[pad_row], GT\n"
-      "csel x23, x23, %x[pad_row], GE\n"
+      "mov x22, %x[width]\n"
+      "cntb x21, ALL, MUL #2\n"
+      "csel x25, x25, %x[pad_row], GT\n"
+      "csel x26, x26, %x[pad_row], GE\n"
       "cmp %x[height], #0x3\n"
-      "csel x24, x24, %x[pad_row], GT\n"
-      "csel x25, x25, %x[pad_row], GE\n"
+      "csel x27, x27, %x[pad_row], GT\n"
+      "csel x28, x28, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x26, x26, %x[pad_row], GT\n"
-      "cmp x9, x28\n"
+      "csel x9, x9, %x[pad_row], GT\n"
+      "cmp x22, x21\n"
+      "mov x20, %x[out]\n"
       "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1b { z24.b }, p1/Z, [x10]\n"
-      "ld1b { z0.b }, p1/Z, [x26]\n"
-      "sub x9, x9, x28\n"
-      "ld1b { z31.b }, p1/Z, [x25]\n"
-      "ld1b { z30.b }, p1/Z, [x24]\n"
-      "cmp x9, x28\n"
-      "ld1b { z23.b }, p1/Z, [x23]\n"
-      "ld1b { z29.b }, p1/Z, [x22]\n"
-      "ld1b { z22.b }, p1/Z, [x21]\n"
-      "ld1b { z21.b }, p1/Z, [x20]\n"
-      "ld1b { z28.b }, p1/Z, [x10, #1, MUL VL]\n"
-      "ld1b { z4.b }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z23.b }, p1/Z, [x10]\n"
+      "ld1b { z22.b }, p1/Z, [x9]\n"
+      "sub x22, x22, x21\n"
+      "cmp x22, x21\n"
+      "ld1b { z20.b }, p1/Z, [x28]\n"
+      "ld1b { z21.b }, p1/Z, [x27]\n"
+      "ld1b { z19.b }, p1/Z, [x26]\n"
+      "ld1b { z18.b }, p1/Z, [x25]\n"
+      "zip1 z5.b, z23.b, z19.b\n"
+      "zip1 z4.b, z22.b, z18.b\n"
+      "ld1b { z17.b }, p1/Z, [x24]\n"
+      "ld1b { z16.b }, p1/Z, [x23]\n"
+      "zip1 z3.b, z20.b, z17.b\n"
+      "zip1 z31.b, z21.b, z16.b\n"
+      "ld1b { z25.b }, p1/Z, [x10, #1, MUL VL]\n"
+      "ld1b { z24.b }, p1/Z, [x9, #1, MUL VL]\n"
+      "zip2 z2.b, z23.b, z19.b\n"
+      "zip2 z30.b, z20.b, z17.b\n"
+      "ld1b { z23.b }, p1/Z, [x28, #1, MUL VL]\n"
+      "ld1b { z20.b }, p1/Z, [x27, #1, MUL VL]\n"
+      "zip2 z22.b, z22.b, z18.b\n"
+      "zip2 z21.b, z21.b, z16.b\n"
+      "ld1b { z19.b }, p1/Z, [x26, #1, MUL VL]\n"
+      "ld1b { z18.b }, p1/Z, [x25, #1, MUL VL]\n"
+      "zip1 z29.b, z25.b, z19.b\n"
+      "zip1 z28.b, z24.b, z18.b\n"
+      "ld1b { z17.b }, p1/Z, [x24, #1, MUL VL]\n"
+      "ld1b { z16.b }, p1/Z, [x23, #1, MUL VL]\n"
+      "zip1 z27.b, z23.b, z17.b\n"
+      "zip1 z26.b, z20.b, z16.b\n"
+      "zip2 z1.b, z25.b, z19.b\n"
+      "zip2 z25.b, z23.b, z17.b\n"
       "addvl x10, x10, #2\n"
+      "addvl x9, x9, #2\n"
+      "zip2 z24.b, z24.b, z18.b\n"
+      "zip2 z16.b, z20.b, z16.b\n"
+      "addvl x28, x28, #2\n"
+      "addvl x27, x27, #2\n"
+      "zip1 z0.b, z5.b, z3.b\n"
+      "zip1 z17.b, z4.b, z31.b\n"
       "addvl x26, x26, #2\n"
-      "ld1b { z27.b }, p1/Z, [x25, #1, MUL VL]\n"
-      "ld1b { z20.b }, p1/Z, [x24, #1, MUL VL]\n"
-      "zip1 z3.b, z24.b, z23.b\n"
-      "zip1 z2.b, z0.b, z29.b\n"
-      "ld1b { z19.b }, p1/Z, [x23, #1, MUL VL]\n"
-      "ld1b { z18.b }, p1/Z, [x22, #1, MUL VL]\n"
-      "zip1 z26.b, z31.b, z22.b\n"
-      "zip1 z25.b, z30.b, z21.b\n"
-      "ld1b { z17.b }, p1/Z, [x21, #1, MUL VL]\n"
-      "ld1b { z16.b }, p1/Z, [x20, #1, MUL VL]\n"
-      "zip2 z24.b, z24.b, z23.b\n"
-      "zip2 z23.b, z31.b, z22.b\n"
-      "zip2 z22.b, z0.b, z29.b\n"
-      "zip2 z21.b, z30.b, z21.b\n"
       "addvl x25, x25, #2\n"
+      "zip2 z20.b, z5.b, z3.b\n"
+      "zip2 z19.b, z4.b, z31.b\n"
       "addvl x24, x24, #2\n"
-      "zip1 z0.b, z28.b, z19.b\n"
-      "zip1 z31.b, z4.b, z18.b\n"
       "addvl x23, x23, #2\n"
-      "addvl x22, x22, #2\n"
-      "zip1 z30.b, z27.b, z17.b\n"
-      "zip1 z29.b, z20.b, z16.b\n"
-      "addvl x21, x21, #2\n"
-      "addvl x20, x20, #2\n"
-      "zip2 z1.b, z28.b, z19.b\n"
-      "zip2 z28.b, z27.b, z17.b\n"
-      "zip2 z27.b, z4.b, z18.b\n"
-      "zip2 z20.b, z20.b, z16.b\n"
-      "zip1 z19.b, z3.b, z26.b\n"
-      "zip1 z18.b, z2.b, z25.b\n"
-      "zip2 z17.b, z3.b, z26.b\n"
-      "zip2 z16.b, z2.b, z25.b\n"
-      "zip1 z26.b, z24.b, z23.b\n"
-      "zip1 z25.b, z22.b, z21.b\n"
-      "zip2 z24.b, z24.b, z23.b\n"
+      "zip1 z31.b, z2.b, z30.b\n"
+      "zip1 z18.b, z22.b, z21.b\n"
+      "zip2 z30.b, z2.b, z30.b\n"
       "zip2 z23.b, z22.b, z21.b\n"
-      "zip1 z22.b, z0.b, z30.b\n"
-      "zip1 z21.b, z31.b, z29.b\n"
-      "zip2 z0.b, z0.b, z30.b\n"
-      "zip2 z31.b, z31.b, z29.b\n"
-      "zip1 z30.b, z1.b, z28.b\n"
-      "zip1 z29.b, z27.b, z20.b\n"
-      "zip2 z28.b, z1.b, z28.b\n"
-      "zip2 z27.b, z27.b, z20.b\n"
-      "zip1 z20.b, z19.b, z18.b\n"
-      "zip2 z19.b, z19.b, z18.b\n"
-      "zip1 z18.b, z17.b, z16.b\n"
-      "zip2 z17.b, z17.b, z16.b\n"
-      "zip1 z16.b, z26.b, z25.b\n"
-      "zip2 z26.b, z26.b, z25.b\n"
-      "zip1 z25.b, z24.b, z23.b\n"
-      "zip2 z24.b, z24.b, z23.b\n"
-      "st1b { z20.b }, p1, [x27]\n"
-      "st1b { z19.b }, p1, [x27, #1, MUL VL]\n"
+      "zip1 z22.b, z29.b, z27.b\n"
+      "zip1 z21.b, z28.b, z26.b\n"
+      "zip2 z29.b, z29.b, z27.b\n"
+      "zip2 z28.b, z28.b, z26.b\n"
+      "zip1 z27.b, z1.b, z25.b\n"
+      "zip1 z26.b, z24.b, z16.b\n"
+      "zip2 z25.b, z1.b, z25.b\n"
+      "zip2 z24.b, z24.b, z16.b\n"
+      "zip1 z16.b, z0.b, z17.b\n"
+      "zip2 z17.b, z0.b, z17.b\n"
+      "st1b { z16.b }, p1, [x20]\n"
+      "zip1 z16.b, z20.b, z19.b\n"
+      "zip2 z20.b, z20.b, z19.b\n"
+      "st1b { z17.b }, p1, [x20, #1, MUL VL]\n"
+      "zip1 z19.b, z31.b, z18.b\n"
+      "zip2 z18.b, z31.b, z18.b\n"
+      "st1b { z16.b }, p1, [x20, #2, MUL VL]\n"
+      "zip1 z17.b, z30.b, z23.b\n"
+      "zip2 z16.b, z30.b, z23.b\n"
+      "st1b { z20.b }, p1, [x20, #3, MUL VL]\n"
+      "st1b { z19.b }, p1, [x20, #4, MUL VL]\n"
       "zip1 z23.b, z22.b, z21.b\n"
       "zip2 z22.b, z22.b, z21.b\n"
-      "st1b { z18.b }, p1, [x27, #2, MUL VL]\n"
-      "zip1 z21.b, z0.b, z31.b\n"
-      "zip2 z20.b, z0.b, z31.b\n"
-      "st1b { z17.b }, p1, [x27, #3, MUL VL]\n"
-      "zip1 z19.b, z30.b, z29.b\n"
-      "zip2 z18.b, z30.b, z29.b\n"
-      "st1b { z16.b }, p1, [x27, #4, MUL VL]\n"
-      "zip1 z17.b, z28.b, z27.b\n"
-      "zip2 z16.b, z28.b, z27.b\n"
-      "st1b { z26.b }, p1, [x27, #5, MUL VL]\n"
-      "st1b { z25.b }, p1, [x27, #6, MUL VL]\n"
-      "st1b { z24.b }, p1, [x27, #7, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
-      "st1b { z23.b }, p1, [x27]\n"
-      "st1b { z22.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z21.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z20.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z18.b }, p1, [x27, #5, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #6, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #7, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "st1b { z18.b }, p1, [x20, #5, MUL VL]\n"
+      "zip1 z21.b, z29.b, z28.b\n"
+      "zip2 z20.b, z29.b, z28.b\n"
+      "st1b { z17.b }, p1, [x20, #6, MUL VL]\n"
+      "zip1 z19.b, z27.b, z26.b\n"
+      "zip2 z18.b, z27.b, z26.b\n"
+      "st1b { z16.b }, p1, [x20, #7, MUL VL]\n"
+      "add x20, x20, %x[out_stride]\n"
+      "zip1 z17.b, z25.b, z24.b\n"
+      "zip2 z16.b, z25.b, z24.b\n"
+      "st1b { z23.b }, p1, [x20]\n"
+      "st1b { z22.b }, p1, [x20, #1, MUL VL]\n"
+      "st1b { z21.b }, p1, [x20, #2, MUL VL]\n"
+      "st1b { z20.b }, p1, [x20, #3, MUL VL]\n"
+      "st1b { z19.b }, p1, [x20, #4, MUL VL]\n"
+      "st1b { z18.b }, p1, [x20, #5, MUL VL]\n"
+      "st1b { z17.b }, p1, [x20, #6, MUL VL]\n"
+      "st1b { z16.b }, p1, [x20, #7, MUL VL]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x9, 5f\n"
+      "cbz x22, 5f\n"
       "4:"  // Main row loop: Column loop
-      "whilelt p0.b, XZR, x9\n"
-      "decd x9, ALL, MUL #8\n"
-      "ld1b { z19.b }, p0/Z, [x10]\n"
+      "whilelt p0.b, XZR, x22\n"
+      "ld1b { z25.b }, p0/Z, [x10]\n"
+      "ld1b { z27.b }, p0/Z, [x9]\n"
+      "decd x22, ALL, MUL #8\n"
+      "ld1b { z26.b }, p0/Z, [x28]\n"
+      "ld1b { z24.b }, p0/Z, [x27]\n"
+      "cmp x22, #0x0\n"
       "addvl x10, x10, #1\n"
-      "ld1b { z26.b }, p0/Z, [x26]\n"
+      "ld1b { z22.b }, p0/Z, [x26]\n"
+      "ld1b { z21.b }, p0/Z, [x25]\n"
+      "zip1 z20.b, z25.b, z22.b\n"
+      "zip1 z23.b, z27.b, z21.b\n"
+      "ld1b { z17.b }, p0/Z, [x24]\n"
+      "ld1b { z16.b }, p0/Z, [x23]\n"
+      "zip1 z19.b, z26.b, z17.b\n"
+      "zip1 z18.b, z24.b, z16.b\n"
+      "zip2 z25.b, z25.b, z22.b\n"
+      "zip2 z22.b, z26.b, z17.b\n"
+      "addvl x9, x9, #1\n"
+      "addvl x28, x28, #1\n"
+      "zip2 z21.b, z27.b, z21.b\n"
+      "zip2 z16.b, z24.b, z16.b\n"
+      "addvl x27, x27, #1\n"
       "addvl x26, x26, #1\n"
-      "ld1b { z22.b }, p0/Z, [x25]\n"
+      "zip1 z24.b, z20.b, z19.b\n"
+      "zip1 z17.b, z23.b, z18.b\n"
       "addvl x25, x25, #1\n"
-      "ld1b { z25.b }, p0/Z, [x24]\n"
       "addvl x24, x24, #1\n"
-      "ld1b { z18.b }, p0/Z, [x23]\n"
-      "ld1b { z21.b }, p0/Z, [x22]\n"
-      "ld1b { z17.b }, p0/Z, [x21]\n"
-      "cmp x9, #0x0\n"
+      "zip2 z20.b, z20.b, z19.b\n"
+      "zip2 z19.b, z23.b, z18.b\n"
       "addvl x23, x23, #1\n"
-      "ld1b { z16.b }, p0/Z, [x20]\n"
-      "addvl x22, x22, #1\n"
-      "addvl x21, x21, #1\n"
-      "zip1 z20.b, z19.b, z18.b\n"
-      "zip2 z24.b, z19.b, z18.b\n"
-      "addvl x20, x20, #1\n"
-      "zip1 z19.b, z22.b, z17.b\n"
-      "zip1 z18.b, z26.b, z21.b\n"
-      "zip2 z23.b, z22.b, z17.b\n"
-      "zip1 z17.b, z25.b, z16.b\n"
-      "zip2 z22.b, z26.b, z21.b\n"
-      "zip2 z16.b, z25.b, z16.b\n"
-      "zip1 z21.b, z20.b, z19.b\n"
+      "zip1 z23.b, z25.b, z22.b\n"
+      "zip1 z18.b, z21.b, z16.b\n"
+      "zip2 z22.b, z25.b, z22.b\n"
+      "zip2 z21.b, z21.b, z16.b\n"
+      "zip1 z16.b, z24.b, z17.b\n"
+      "zip2 z17.b, z24.b, z17.b\n"
+      "st1b { z16.b }, p1, [x20]\n"
+      "zip1 z16.b, z20.b, z19.b\n"
       "zip2 z20.b, z20.b, z19.b\n"
-      "zip1 z25.b, z24.b, z23.b\n"
-      "zip1 z19.b, z18.b, z17.b\n"
-      "zip2 z18.b, z18.b, z17.b\n"
-      "zip1 z17.b, z22.b, z16.b\n"
-      "zip2 z24.b, z24.b, z23.b\n"
-      "zip2 z23.b, z22.b, z16.b\n"
-      "zip1 z16.b, z21.b, z19.b\n"
-      "zip2 z22.b, z21.b, z19.b\n"
-      "zip1 z21.b, z20.b, z18.b\n"
-      "zip2 z20.b, z20.b, z18.b\n"
-      "zip1 z19.b, z25.b, z17.b\n"
-      "zip2 z18.b, z25.b, z17.b\n"
-      "zip1 z17.b, z24.b, z23.b\n"
-      "st1b { z16.b }, p1, [x27]\n"
-      "zip2 z16.b, z24.b, z23.b\n"
-      "st1b { z22.b }, p1, [x27, #1, MUL VL]\n"
-      "st1b { z21.b }, p1, [x27, #2, MUL VL]\n"
-      "st1b { z20.b }, p1, [x27, #3, MUL VL]\n"
-      "st1b { z19.b }, p1, [x27, #4, MUL VL]\n"
-      "st1b { z18.b }, p1, [x27, #5, MUL VL]\n"
-      "st1b { z17.b }, p1, [x27, #6, MUL VL]\n"
-      "st1b { z16.b }, p1, [x27, #7, MUL VL]\n"
-      "add x27, x27, %x[out_stride]\n"
+      "st1b { z17.b }, p1, [x20, #1, MUL VL]\n"
+      "zip1 z19.b, z23.b, z18.b\n"
+      "zip2 z18.b, z23.b, z18.b\n"
+      "st1b { z16.b }, p1, [x20, #2, MUL VL]\n"
+      "zip1 z17.b, z22.b, z21.b\n"
+      "zip2 z16.b, z22.b, z21.b\n"
+      "st1b { z20.b }, p1, [x20, #3, MUL VL]\n"
+      "st1b { z19.b }, p1, [x20, #4, MUL VL]\n"
+      "st1b { z18.b }, p1, [x20, #5, MUL VL]\n"
+      "st1b { z17.b }, p1, [x20, #6, MUL VL]\n"
+      "st1b { z16.b }, p1, [x20, #7, MUL VL]\n"
+      "add x20, x20, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -223,7 +223,7 @@ void sve_transpose_interleave_8VL_1x8(uint8_t *out, const uint8_t *in, size_t wi
       "bge 1b\n"
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x2.hpp
index 6d436ebcad..eb7312b2e3 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x2.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x2.hpp
@@ -40,198 +40,198 @@ void sve_transpose_interleave_8VL_2x2(uint16_t *out, const uint16_t *in, size_t
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "ptrue p2.b\n"
+      "ptrue p4.b\n"
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x28, %x[in]\n"
       "mov x27, %x[width]\n"
       "cnth x26, ALL, MUL #8\n"
-      "mov x25, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
-      "add x24, x28, %x[in_stride]\n"
+      "add x25, x28, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
       "cmp x27, x26\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1h { z17.h }, p2/Z, [x28]\n"
-      "ld1h { z30.h }, p2/Z, [x28, #1, MUL VL]\n"
-      "mov x21, x25\n"
-      "add x25, x25, %x[out_stride]\n"
-      "ld1h { z28.h }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z1.h }, p2/Z, [x28, #3, MUL VL]\n"
-      "mov x20, x25\n"
+      "ld1h { z30.h }, p4/Z, [x28]\n"
+      "ld1h { z12.h }, p4/Z, [x28, #1, MUL VL]\n"
+      "mov x21, x22\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1h { z31.h }, p4/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z18.h }, p4/Z, [x28, #3, MUL VL]\n"
+      "mov x20, x22\n"
       "sub x27, x27, x26\n"
-      "ld1h { z16.h }, p2/Z, [x24]\n"
-      "ld1h { z25.h }, p2/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z20.h }, p4/Z, [x25]\n"
+      "ld1h { z17.h }, p4/Z, [x25, #1, MUL VL]\n"
+      "zip1 z3.h, z30.h, z20.h\n"
+      "zip2 z21.h, z30.h, z20.h\n"
+      "ld1h { z26.h }, p4/Z, [x25, #2, MUL VL]\n"
+      "ld1h { z23.h }, p4/Z, [x25, #3, MUL VL]\n"
+      "zip1 z13.h, z12.h, z17.h\n"
+      "zip2 z0.h, z12.h, z17.h\n"
+      "ld1h { z2.h }, p4/Z, [x28, #4, MUL VL]\n"
+      "ld1h { z24.h }, p4/Z, [x28, #5, MUL VL]\n"
+      "zip1 z12.h, z31.h, z26.h\n"
+      "zip2 z14.h, z31.h, z26.h\n"
+      "ld1h { z17.h }, p4/Z, [x28, #6, MUL VL]\n"
+      "ld1h { z29.h }, p4/Z, [x28, #7, MUL VL]\n"
+      "zip1 z16.h, z18.h, z23.h\n"
+      "zip2 z15.h, z18.h, z23.h\n"
+      "ld1h { z9.h }, p4/Z, [x25, #4, MUL VL]\n"
+      "ld1h { z18.h }, p4/Z, [x25, #5, MUL VL]\n"
+      "zip1 z11.h, z2.h, z9.h\n"
+      "zip2 z5.h, z2.h, z9.h\n"
+      "ld1h { z7.h }, p4/Z, [x25, #6, MUL VL]\n"
+      "ld1h { z2.h }, p4/Z, [x25, #7, MUL VL]\n"
+      "zip1 z10.h, z24.h, z18.h\n"
+      "zip2 z6.h, z24.h, z18.h\n"
+      "ld1h { z19.h }, p4/Z, [x24]\n"
+      "ld1h { z18.h }, p4/Z, [x24, #1, MUL VL]\n"
+      "zip1 z9.h, z17.h, z7.h\n"
+      "zip2 z4.h, z17.h, z7.h\n"
+      "ld1h { z24.h }, p4/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z22.h }, p4/Z, [x24, #3, MUL VL]\n"
+      "zip1 z7.h, z29.h, z2.h\n"
+      "zip2 z8.h, z29.h, z2.h\n"
+      "ld1h { z25.h }, p4/Z, [x24, #4, MUL VL]\n"
+      "ld1h { z17.h }, p4/Z, [x24, #5, MUL VL]\n"
       "cmp x27, x26\n"
-      "add x25, x25, %x[out_stride]\n"
-      "ld1h { z24.h }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1h { z22.h }, p2/Z, [x24, #3, MUL VL]\n"
-      "ld1h { z21.h }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1h { z20.h }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1h { z23.h }, p2/Z, [x28, #6, MUL VL]\n"
-      "ld1h { z19.h }, p2/Z, [x28, #7, MUL VL]\n"
-      "zip1 z31.h, z17.h, z16.h\n"
-      "zip2 z29.h, z17.h, z16.h\n"
-      "ld1h { z18.h }, p2/Z, [x24, #4, MUL VL]\n"
-      "ld1h { z17.h }, p2/Z, [x24, #5, MUL VL]\n"
-      "zip1 z27.h, z30.h, z25.h\n"
-      "zip2 z26.h, z30.h, z25.h\n"
-      "ld1h { z16.h }, p2/Z, [x24, #6, MUL VL]\n"
-      "ld1h { z0.h }, p2/Z, [x24, #7, MUL VL]\n"
-      "zip1 z14.h, z28.h, z24.h\n"
-      "zip2 z15.h, z28.h, z24.h\n"
-      "ld1h { z30.h }, p2/Z, [x23]\n"
-      "ld1h { z28.h }, p2/Z, [x23, #1, MUL VL]\n"
-      "zip1 z13.h, z1.h, z22.h\n"
-      "zip2 z12.h, z1.h, z22.h\n"
-      "ld1h { z25.h }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z24.h }, p2/Z, [x23, #3, MUL VL]\n"
-      "zip1 z11.h, z21.h, z18.h\n"
-      "zip2 z10.h, z21.h, z18.h\n"
-      "ld1h { z9.h }, p2/Z, [x23, #4, MUL VL]\n"
-      "ld1h { z8.h }, p2/Z, [x23, #5, MUL VL]\n"
-      "zip1 z7.h, z20.h, z17.h\n"
-      "zip2 z6.h, z20.h, z17.h\n"
-      "ld1h { z5.h }, p2/Z, [x23, #6, MUL VL]\n"
-      "ld1h { z4.h }, p2/Z, [x23, #7, MUL VL]\n"
-      "zip1 z3.h, z23.h, z16.h\n"
-      "zip2 z2.h, z23.h, z16.h\n"
-      "ld1h { z23.h }, p2/Z, [x22]\n"
-      "ld1h { z22.h }, p2/Z, [x22, #1, MUL VL]\n"
-      "zip1 z1.h, z19.h, z0.h\n"
-      "zip2 z0.h, z19.h, z0.h\n"
-      "ld1h { z21.h }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1h { z20.h }, p2/Z, [x22, #3, MUL VL]\n"
       "addvl x28, x28, #8\n"
+      "ld1h { z2.h }, p4/Z, [x24, #6, MUL VL]\n"
+      "ld1h { z30.h }, p4/Z, [x24, #7, MUL VL]\n"
+      "addvl x25, x25, #8\n"
       "addvl x24, x24, #8\n"
-      "ld1h { z19.h }, p2/Z, [x22, #4, MUL VL]\n"
-      "ld1h { z18.h }, p2/Z, [x22, #5, MUL VL]\n"
+      "ld1h { z20.h }, p4/Z, [x23]\n"
+      "ld1h { z27.h }, p4/Z, [x23, #1, MUL VL]\n"
+      "zip1 z31.h, z19.h, z20.h\n"
+      "zip2 z29.h, z19.h, z20.h\n"
+      "ld1h { z26.h }, p4/Z, [x23, #2, MUL VL]\n"
+      "ld1h { z23.h }, p4/Z, [x23, #3, MUL VL]\n"
+      "zip1 z28.h, z18.h, z27.h\n"
+      "zip2 z1.h, z18.h, z27.h\n"
+      "ld1h { z20.h }, p4/Z, [x23, #4, MUL VL]\n"
+      "ld1h { z19.h }, p4/Z, [x23, #5, MUL VL]\n"
+      "zip1 z27.h, z24.h, z26.h\n"
+      "zip2 z26.h, z24.h, z26.h\n"
+      "ld1h { z18.h }, p4/Z, [x23, #6, MUL VL]\n"
+      "ld1h { z24.h }, p4/Z, [x23, #7, MUL VL]\n"
+      "st1h { z3.h }, p4, [x21]\n"
+      "zip1 z3.h, z22.h, z23.h\n"
+      "st1h { z21.h }, p4, [x21, #1, MUL VL]\n"
+      "zip2 z22.h, z22.h, z23.h\n"
       "addvl x23, x23, #8\n"
-      "ld1h { z17.h }, p2/Z, [x22, #6, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x22, #7, MUL VL]\n"
-      "st1h { z31.h }, p2, [x21]\n"
-      "zip1 z31.h, z30.h, z23.h\n"
-      "st1h { z29.h }, p2, [x21, #1, MUL VL]\n"
-      "zip2 z30.h, z30.h, z23.h\n"
-      "zip1 z29.h, z28.h, z22.h\n"
-      "addvl x22, x22, #8\n"
-      "st1h { z27.h }, p2, [x21, #2, MUL VL]\n"
-      "zip2 z28.h, z28.h, z22.h\n"
-      "zip1 z27.h, z25.h, z21.h\n"
-      "st1h { z26.h }, p2, [x21, #3, MUL VL]\n"
-      "zip2 z26.h, z25.h, z21.h\n"
-      "zip1 z25.h, z24.h, z20.h\n"
-      "st1h { z14.h }, p2, [x21, #4, MUL VL]\n"
-      "zip2 z24.h, z24.h, z20.h\n"
-      "zip1 z23.h, z9.h, z19.h\n"
-      "st1h { z15.h }, p2, [x21, #5, MUL VL]\n"
-      "zip2 z22.h, z9.h, z19.h\n"
-      "zip1 z21.h, z8.h, z18.h\n"
-      "st1h { z13.h }, p2, [x21, #6, MUL VL]\n"
-      "zip2 z20.h, z8.h, z18.h\n"
-      "zip1 z19.h, z5.h, z17.h\n"
-      "st1h { z12.h }, p2, [x21, #7, MUL VL]\n"
+      "zip1 z23.h, z25.h, z20.h\n"
+      "st1h { z13.h }, p4, [x21, #2, MUL VL]\n"
+      "zip2 z25.h, z25.h, z20.h\n"
+      "zip1 z21.h, z17.h, z19.h\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z0.h }, p4, [x21, #3, MUL VL]\n"
+      "zip2 z20.h, z17.h, z19.h\n"
+      "zip1 z19.h, z2.h, z18.h\n"
+      "st1h { z12.h }, p4, [x21, #4, MUL VL]\n"
+      "zip2 z18.h, z2.h, z18.h\n"
+      "zip1 z17.h, z30.h, z24.h\n"
+      "st1h { z14.h }, p4, [x21, #5, MUL VL]\n"
+      "zip2 z13.h, z30.h, z24.h\n"
+      "st1h { z16.h }, p4, [x21, #6, MUL VL]\n"
+      "st1h { z15.h }, p4, [x21, #7, MUL VL]\n"
       "addvl x21, x21, #16\n"
-      "zip2 z18.h, z5.h, z17.h\n"
-      "zip1 z17.h, z4.h, z16.h\n"
-      "zip2 z16.h, z4.h, z16.h\n"
-      "st1h { z31.h }, p2, [x21, #-8, MUL VL]\n"
-      "st1h { z30.h }, p2, [x21, #-7, MUL VL]\n"
-      "st1h { z29.h }, p2, [x21, #-6, MUL VL]\n"
-      "st1h { z28.h }, p2, [x21, #-5, MUL VL]\n"
-      "st1h { z27.h }, p2, [x21, #-4, MUL VL]\n"
-      "st1h { z26.h }, p2, [x21, #-3, MUL VL]\n"
-      "st1h { z25.h }, p2, [x21, #-2, MUL VL]\n"
-      "st1h { z24.h }, p2, [x21, #-1, MUL VL]\n"
-      "st1h { z11.h }, p2, [x20]\n"
-      "st1h { z10.h }, p2, [x20, #1, MUL VL]\n"
-      "st1h { z7.h }, p2, [x20, #2, MUL VL]\n"
-      "st1h { z6.h }, p2, [x20, #3, MUL VL]\n"
-      "st1h { z3.h }, p2, [x20, #4, MUL VL]\n"
-      "st1h { z2.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z1.h }, p2, [x20, #6, MUL VL]\n"
-      "st1h { z0.h }, p2, [x20, #7, MUL VL]\n"
+      "st1h { z31.h }, p4, [x21, #-8, MUL VL]\n"
+      "st1h { z29.h }, p4, [x21, #-7, MUL VL]\n"
+      "st1h { z28.h }, p4, [x21, #-6, MUL VL]\n"
+      "st1h { z1.h }, p4, [x21, #-5, MUL VL]\n"
+      "st1h { z27.h }, p4, [x21, #-4, MUL VL]\n"
+      "st1h { z26.h }, p4, [x21, #-3, MUL VL]\n"
+      "st1h { z3.h }, p4, [x21, #-2, MUL VL]\n"
+      "st1h { z22.h }, p4, [x21, #-1, MUL VL]\n"
+      "st1h { z11.h }, p4, [x20]\n"
+      "st1h { z5.h }, p4, [x20, #1, MUL VL]\n"
+      "st1h { z10.h }, p4, [x20, #2, MUL VL]\n"
+      "st1h { z6.h }, p4, [x20, #3, MUL VL]\n"
+      "st1h { z9.h }, p4, [x20, #4, MUL VL]\n"
+      "st1h { z4.h }, p4, [x20, #5, MUL VL]\n"
+      "st1h { z7.h }, p4, [x20, #6, MUL VL]\n"
+      "st1h { z8.h }, p4, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
-      "st1h { z23.h }, p2, [x20, #-8, MUL VL]\n"
-      "st1h { z22.h }, p2, [x20, #-7, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #-6, MUL VL]\n"
-      "st1h { z20.h }, p2, [x20, #-5, MUL VL]\n"
-      "st1h { z19.h }, p2, [x20, #-4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x20, #-3, MUL VL]\n"
-      "st1h { z17.h }, p2, [x20, #-2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
+      "st1h { z23.h }, p4, [x20, #-8, MUL VL]\n"
+      "st1h { z25.h }, p4, [x20, #-7, MUL VL]\n"
+      "st1h { z21.h }, p4, [x20, #-6, MUL VL]\n"
+      "st1h { z20.h }, p4, [x20, #-5, MUL VL]\n"
+      "st1h { z19.h }, p4, [x20, #-4, MUL VL]\n"
+      "st1h { z18.h }, p4, [x20, #-3, MUL VL]\n"
+      "st1h { z17.h }, p4, [x20, #-2, MUL VL]\n"
+      "st1h { z13.h }, p4, [x20, #-1, MUL VL]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
       "cbz x27, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x21, x27\n"
-      "mov x20, x25\n"
+      "mov x20, x27\n"
+      "whilelt p3.h, XZR, x20\n"
+      "ld1h { z20.h }, p3/Z, [x28]\n"
+      "ld1h { z19.h }, p3/Z, [x25]\n"
+      "dech x20\n"
+      "whilelt p2.h, XZR, x20\n"
+      "ld1h { z18.h }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1h { z17.h }, p2/Z, [x25, #1, MUL VL]\n"
+      "dech x20\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z25.h }, p1/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z16.h }, p1/Z, [x25, #2, MUL VL]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z0.h }, p0/Z, [x28, #3, MUL VL]\n"
+      "ld1h { z24.h }, p0/Z, [x25, #3, MUL VL]\n"
+      "mov x20, x22\n"
       "decw x27, ALL, MUL #8\n"
-      "add x25, x25, %x[out_stride]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p0.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z20.h }, p1/Z, [x28]\n"
-      "ld1h { z16.h }, p1/Z, [x24]\n"
-      "ld1h { z23.h }, p1/Z, [x23]\n"
-      "ld1h { z19.h }, p1/Z, [x22]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "ld1h { z18.h }, p0/Z, [x28, #1, MUL VL]\n"
-      "ld1h { z17.h }, p0/Z, [x24, #1, MUL VL]\n"
-      "ld1h { z1.h }, p0/Z, [x23, #1, MUL VL]\n"
-      "ld1h { z0.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "zip1 z22.h, z20.h, z16.h\n"
-      "zip2 z21.h, z20.h, z16.h\n"
-      "whilelt p0.h, XZR, x21\n"
-      "ld1h { z20.h }, p1/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z16.h }, p1/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z31.h }, p3/Z, [x24]\n"
+      "ld1h { z30.h }, p2/Z, [x24, #1, MUL VL]\n"
+      "ld1h { z29.h }, p1/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z28.h }, p0/Z, [x24, #3, MUL VL]\n"
+      "zip1 z23.h, z20.h, z19.h\n"
+      "zip2 z22.h, z20.h, z19.h\n"
+      "ld1h { z21.h }, p3/Z, [x23]\n"
+      "ld1h { z27.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "zip1 z20.h, z18.h, z17.h\n"
+      "zip2 z19.h, z18.h, z17.h\n"
+      "ld1h { z18.h }, p1/Z, [x23, #2, MUL VL]\n"
+      "ld1h { z26.h }, p0/Z, [x23, #3, MUL VL]\n"
+      "zip1 z17.h, z25.h, z16.h\n"
+      "zip2 z16.h, z25.h, z16.h\n"
+      "zip1 z25.h, z0.h, z24.h\n"
+      "zip2 z24.h, z0.h, z24.h\n"
+      "st1h { z23.h }, p4, [x20]\n"
       "cmp x27, #0x0\n"
-      "ld1h { z31.h }, p1/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z30.h }, p1/Z, [x22, #2, MUL VL]\n"
-      "zip1 z29.h, z18.h, z17.h\n"
-      "zip2 z28.h, z18.h, z17.h\n"
-      "zip1 z27.h, z23.h, z19.h\n"
-      "zip2 z26.h, z23.h, z19.h\n"
-      "ld1h { z19.h }, p0/Z, [x28, #3, MUL VL]\n"
+      "st1h { z22.h }, p4, [x20, #1, MUL VL]\n"
       "addvl x28, x28, #4\n"
-      "ld1h { z18.h }, p0/Z, [x24, #3, MUL VL]\n"
-      "ld1h { z25.h }, p0/Z, [x23, #3, MUL VL]\n"
-      "zip1 z17.h, z20.h, z16.h\n"
-      "zip2 z24.h, z20.h, z16.h\n"
-      "ld1h { z16.h }, p0/Z, [x22, #3, MUL VL]\n"
-      "st1h { z22.h }, p2, [x20]\n"
+      "addvl x25, x25, #4\n"
+      "zip1 z23.h, z31.h, z21.h\n"
+      "st1h { z20.h }, p4, [x20, #2, MUL VL]\n"
       "addvl x24, x24, #4\n"
       "addvl x23, x23, #4\n"
-      "st1h { z21.h }, p2, [x20, #1, MUL VL]\n"
-      "addvl x22, x22, #4\n"
-      "zip1 z23.h, z1.h, z0.h\n"
-      "zip2 z22.h, z1.h, z0.h\n"
-      "zip1 z21.h, z19.h, z18.h\n"
-      "zip2 z20.h, z19.h, z18.h\n"
-      "st1h { z29.h }, p2, [x20, #2, MUL VL]\n"
-      "st1h { z28.h }, p2, [x20, #3, MUL VL]\n"
-      "zip1 z19.h, z31.h, z30.h\n"
-      "zip2 z18.h, z31.h, z30.h\n"
-      "st1h { z17.h }, p2, [x20, #4, MUL VL]\n"
-      "zip1 z17.h, z25.h, z16.h\n"
-      "zip2 z16.h, z25.h, z16.h\n"
-      "st1h { z24.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #6, MUL VL]\n"
-      "st1h { z20.h }, p2, [x20, #7, MUL VL]\n"
+      "zip2 z22.h, z31.h, z21.h\n"
+      "st1h { z19.h }, p4, [x20, #3, MUL VL]\n"
+      "zip1 z21.h, z30.h, z27.h\n"
+      "zip2 z20.h, z30.h, z27.h\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z17.h }, p4, [x20, #4, MUL VL]\n"
+      "zip1 z19.h, z29.h, z18.h\n"
+      "zip2 z18.h, z29.h, z18.h\n"
+      "st1h { z16.h }, p4, [x20, #5, MUL VL]\n"
+      "zip1 z17.h, z28.h, z26.h\n"
+      "zip2 z16.h, z28.h, z26.h\n"
+      "st1h { z25.h }, p4, [x20, #6, MUL VL]\n"
+      "st1h { z24.h }, p4, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
-      "st1h { z27.h }, p2, [x20, #-8, MUL VL]\n"
-      "st1h { z26.h }, p2, [x20, #-7, MUL VL]\n"
-      "st1h { z23.h }, p2, [x20, #-6, MUL VL]\n"
-      "st1h { z22.h }, p2, [x20, #-5, MUL VL]\n"
-      "st1h { z19.h }, p2, [x20, #-4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x20, #-3, MUL VL]\n"
-      "st1h { z17.h }, p2, [x20, #-2, MUL VL]\n"
-      "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
+      "st1h { z23.h }, p4, [x20, #-8, MUL VL]\n"
+      "st1h { z22.h }, p4, [x20, #-7, MUL VL]\n"
+      "st1h { z21.h }, p4, [x20, #-6, MUL VL]\n"
+      "st1h { z20.h }, p4, [x20, #-5, MUL VL]\n"
+      "st1h { z19.h }, p4, [x20, #-4, MUL VL]\n"
+      "st1h { z18.h }, p4, [x20, #-3, MUL VL]\n"
+      "st1h { z17.h }, p4, [x20, #-2, MUL VL]\n"
+      "st1h { z16.h }, p4, [x20, #-1, MUL VL]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x4\n"
@@ -243,110 +243,110 @@ void sve_transpose_interleave_8VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "mov x28, %x[in]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #8\n"
+      "add x25, x28, %x[in_stride]\n"
       "cmp %x[height], #0x1\n"
-      "mov x25, %x[out]\n"
-      "sub %x[height], %x[height], #0x2\n"
-      "add x24, x28, %x[in_stride]\n"
-      "add %x[in], x24, %x[in_stride]\n"
-      "csel x24, x24, %x[pad_row], GT\n"
+      "add %x[in], x25, %x[in_stride]\n"
+      "csel x25, x25, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x2\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1h { z19.h }, p2/Z, [x28]\n"
-      "ld1h { z18.h }, p2/Z, [x28, #1, MUL VL]\n"
+      "ld1h { z17.h }, p4/Z, [x28]\n"
+      "ld1h { z20.h }, p4/Z, [x28, #1, MUL VL]\n"
       "sub x21, x21, x20\n"
-      "ld1h { z26.h }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z0.h }, p2/Z, [x28, #3, MUL VL]\n"
       "cmp x21, x20\n"
-      "ld1h { z17.h }, p2/Z, [x24]\n"
-      "ld1h { z16.h }, p2/Z, [x24, #1, MUL VL]\n"
-      "ld1h { z25.h }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1h { z24.h }, p2/Z, [x24, #3, MUL VL]\n"
-      "ld1h { z31.h }, p2/Z, [x28, #4, MUL VL]\n"
-      "ld1h { z30.h }, p2/Z, [x28, #5, MUL VL]\n"
-      "ld1h { z29.h }, p2/Z, [x28, #6, MUL VL]\n"
-      "ld1h { z28.h }, p2/Z, [x28, #7, MUL VL]\n"
-      "zip1 z23.h, z19.h, z17.h\n"
-      "zip2 z22.h, z19.h, z17.h\n"
-      "ld1h { z21.h }, p2/Z, [x24, #4, MUL VL]\n"
-      "ld1h { z20.h }, p2/Z, [x24, #5, MUL VL]\n"
-      "zip1 z19.h, z18.h, z16.h\n"
-      "zip2 z18.h, z18.h, z16.h\n"
-      "ld1h { z17.h }, p2/Z, [x24, #6, MUL VL]\n"
-      "ld1h { z27.h }, p2/Z, [x24, #7, MUL VL]\n"
-      "zip1 z16.h, z26.h, z25.h\n"
-      "zip2 z26.h, z26.h, z25.h\n"
-      "zip1 z25.h, z0.h, z24.h\n"
-      "zip2 z24.h, z0.h, z24.h\n"
-      "st1h { z23.h }, p2, [x25]\n"
+      "ld1h { z23.h }, p4/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z19.h }, p4/Z, [x28, #3, MUL VL]\n"
+      "ld1h { z16.h }, p4/Z, [x25]\n"
+      "ld1h { z18.h }, p4/Z, [x25, #1, MUL VL]\n"
+      "zip1 z0.h, z17.h, z16.h\n"
+      "zip2 z22.h, z17.h, z16.h\n"
+      "ld1h { z17.h }, p4/Z, [x25, #2, MUL VL]\n"
+      "ld1h { z16.h }, p4/Z, [x25, #3, MUL VL]\n"
+      "zip1 z31.h, z20.h, z18.h\n"
+      "zip2 z30.h, z20.h, z18.h\n"
+      "ld1h { z21.h }, p4/Z, [x28, #4, MUL VL]\n"
+      "ld1h { z20.h }, p4/Z, [x28, #5, MUL VL]\n"
+      "zip1 z29.h, z23.h, z17.h\n"
+      "zip2 z28.h, z23.h, z17.h\n"
+      "ld1h { z27.h }, p4/Z, [x28, #6, MUL VL]\n"
+      "ld1h { z26.h }, p4/Z, [x28, #7, MUL VL]\n"
+      "zip1 z25.h, z19.h, z16.h\n"
+      "zip2 z24.h, z19.h, z16.h\n"
+      "ld1h { z19.h }, p4/Z, [x25, #4, MUL VL]\n"
+      "ld1h { z18.h }, p4/Z, [x25, #5, MUL VL]\n"
       "addvl x28, x28, #8\n"
-      "st1h { z22.h }, p2, [x25, #1, MUL VL]\n"
-      "addvl x24, x24, #8\n"
-      "zip1 z23.h, z31.h, z21.h\n"
-      "zip2 z22.h, z31.h, z21.h\n"
-      "st1h { z19.h }, p2, [x25, #2, MUL VL]\n"
-      "zip1 z21.h, z30.h, z20.h\n"
-      "zip2 z20.h, z30.h, z20.h\n"
-      "st1h { z18.h }, p2, [x25, #3, MUL VL]\n"
-      "zip1 z19.h, z29.h, z17.h\n"
-      "zip2 z18.h, z29.h, z17.h\n"
-      "st1h { z16.h }, p2, [x25, #4, MUL VL]\n"
-      "zip1 z17.h, z28.h, z27.h\n"
-      "zip2 z16.h, z28.h, z27.h\n"
-      "st1h { z26.h }, p2, [x25, #5, MUL VL]\n"
-      "st1h { z25.h }, p2, [x25, #6, MUL VL]\n"
-      "st1h { z24.h }, p2, [x25, #7, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
-      "st1h { z23.h }, p2, [x25]\n"
-      "st1h { z22.h }, p2, [x25, #1, MUL VL]\n"
-      "st1h { z21.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z20.h }, p2, [x25, #3, MUL VL]\n"
-      "st1h { z19.h }, p2, [x25, #4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #5, MUL VL]\n"
-      "st1h { z17.h }, p2, [x25, #6, MUL VL]\n"
-      "st1h { z16.h }, p2, [x25, #7, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
+      "zip1 z23.h, z21.h, z19.h\n"
+      "ld1h { z17.h }, p4/Z, [x25, #6, MUL VL]\n"
+      "ld1h { z16.h }, p4/Z, [x25, #7, MUL VL]\n"
+      "st1h { z0.h }, p4, [x22]\n"
+      "addvl x25, x25, #8\n"
+      "st1h { z22.h }, p4, [x22, #1, MUL VL]\n"
+      "zip2 z22.h, z21.h, z19.h\n"
+      "zip1 z21.h, z20.h, z18.h\n"
+      "st1h { z31.h }, p4, [x22, #2, MUL VL]\n"
+      "zip2 z20.h, z20.h, z18.h\n"
+      "zip1 z19.h, z27.h, z17.h\n"
+      "st1h { z30.h }, p4, [x22, #3, MUL VL]\n"
+      "zip2 z18.h, z27.h, z17.h\n"
+      "zip1 z17.h, z26.h, z16.h\n"
+      "st1h { z29.h }, p4, [x22, #4, MUL VL]\n"
+      "zip2 z16.h, z26.h, z16.h\n"
+      "st1h { z28.h }, p4, [x22, #5, MUL VL]\n"
+      "st1h { z25.h }, p4, [x22, #6, MUL VL]\n"
+      "st1h { z24.h }, p4, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z23.h }, p4, [x22]\n"
+      "st1h { z22.h }, p4, [x22, #1, MUL VL]\n"
+      "st1h { z21.h }, p4, [x22, #2, MUL VL]\n"
+      "st1h { z20.h }, p4, [x22, #3, MUL VL]\n"
+      "st1h { z19.h }, p4, [x22, #4, MUL VL]\n"
+      "st1h { z18.h }, p4, [x22, #5, MUL VL]\n"
+      "st1h { z17.h }, p4, [x22, #6, MUL VL]\n"
+      "st1h { z16.h }, p4, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decw x21, ALL, MUL #8\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z22.h }, p0/Z, [x28]\n"
+      "ld1h { z21.h }, p0/Z, [x25]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z20.h }, p0/Z, [x28, #1, MUL VL]\n"
+      "ld1h { z19.h }, p0/Z, [x25, #1, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z18.h }, p1/Z, [x28]\n"
-      "ld1h { z17.h }, p1/Z, [x24]\n"
-      "whilelt p1.h, XZR, x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z18.h }, p0/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z17.h }, p0/Z, [x25, #2, MUL VL]\n"
       "dech x20\n"
-      "ld1h { z20.h }, p0/Z, [x28, #1, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x24, #1, MUL VL]\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z23.h, z18.h, z17.h\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z24.h }, p0/Z, [x28, #3, MUL VL]\n"
+      "ld1h { z23.h }, p0/Z, [x25, #3, MUL VL]\n"
+      "decw x21, ALL, MUL #8\n"
       "cmp x21, #0x0\n"
-      "ld1h { z18.h }, p1/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z17.h }, p1/Z, [x24, #2, MUL VL]\n"
-      "zip1 z22.h, z20.h, z16.h\n"
-      "zip2 z21.h, z20.h, z16.h\n"
-      "ld1h { z20.h }, p0/Z, [x28, #3, MUL VL]\n"
+      "zip1 z16.h, z22.h, z21.h\n"
+      "zip2 z22.h, z22.h, z21.h\n"
       "addvl x28, x28, #4\n"
-      "ld1h { z16.h }, p0/Z, [x24, #3, MUL VL]\n"
-      "addvl x24, x24, #4\n"
-      "st1h { z19.h }, p2, [x25]\n"
+      "addvl x25, x25, #4\n"
+      "zip1 z21.h, z20.h, z19.h\n"
+      "zip2 z20.h, z20.h, z19.h\n"
       "zip1 z19.h, z18.h, z17.h\n"
       "zip2 z18.h, z18.h, z17.h\n"
-      "st1h { z23.h }, p2, [x25, #1, MUL VL]\n"
-      "zip1 z17.h, z20.h, z16.h\n"
-      "zip2 z16.h, z20.h, z16.h\n"
-      "st1h { z22.h }, p2, [x25, #2, MUL VL]\n"
-      "st1h { z21.h }, p2, [x25, #3, MUL VL]\n"
-      "st1h { z19.h }, p2, [x25, #4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x25, #5, MUL VL]\n"
-      "st1h { z17.h }, p2, [x25, #6, MUL VL]\n"
-      "st1h { z16.h }, p2, [x25, #7, MUL VL]\n"
-      "add x25, x25, %x[out_stride]\n"
+      "st1h { z16.h }, p4, [x22]\n"
+      "zip1 z17.h, z24.h, z23.h\n"
+      "zip2 z16.h, z24.h, z23.h\n"
+      "st1h { z22.h }, p4, [x22, #1, MUL VL]\n"
+      "st1h { z21.h }, p4, [x22, #2, MUL VL]\n"
+      "st1h { z20.h }, p4, [x22, #3, MUL VL]\n"
+      "st1h { z19.h }, p4, [x22, #4, MUL VL]\n"
+      "st1h { z18.h }, p4, [x22, #5, MUL VL]\n"
+      "st1h { z17.h }, p4, [x22, #6, MUL VL]\n"
+      "st1h { z16.h }, p4, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -355,7 +355,7 @@ void sve_transpose_interleave_8VL_2x2(uint16_t *out, const uint16_t *in, size_t
       "12:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x4.hpp
index 8ed1879643..1d2c0742ea 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x4.hpp
@@ -44,155 +44,155 @@ void sve_transpose_interleave_8VL_2x4(uint16_t *out, const uint16_t *in, size_t
       "blt 6f\n"
       "1:"  // Main row loop: Head
       "mov x12, %x[in]\n"
-      "mov x11, %x[width]\n"
-      "cnth x10, ALL, MUL #4\n"
-      "mov x9, %x[out]\n"
-      "sub %x[height], %x[height], #0x8\n"
-      "add x28, x12, %x[in_stride]\n"
-      "add x27, x28, %x[in_stride]\n"
-      "add x26, x27, %x[in_stride]\n"
-      "add x25, x26, %x[in_stride]\n"
-      "cmp x11, x10\n"
+      "add x11, x12, %x[in_stride]\n"
+      "add x10, x11, %x[in_stride]\n"
+      "add x9, x10, %x[in_stride]\n"
+      "add x28, x9, %x[in_stride]\n"
+      "mov x27, %x[width]\n"
+      "cnth x26, ALL, MUL #4\n"
+      "add x25, x28, %x[in_stride]\n"
       "add x24, x25, %x[in_stride]\n"
       "add x23, x24, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add %x[in], x22, %x[in_stride]\n"
+      "cmp x27, x26\n"
+      "add %x[in], x23, %x[in_stride]\n"
+      "mov x22, %x[out]\n"
+      "sub %x[height], %x[height], #0x8\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1h { z22.h }, p2/Z, [x12]\n"
-      "ld1h { z2.h }, p2/Z, [x12, #1, MUL VL]\n"
-      "mov x21, x9\n"
-      "add x9, x9, %x[out_stride]\n"
-      "ld1h { z23.h }, p2/Z, [x28]\n"
-      "ld1h { z5.h }, p2/Z, [x28, #1, MUL VL]\n"
-      "mov x20, x9\n"
-      "sub x11, x11, x10\n"
-      "ld1h { z3.h }, p2/Z, [x27]\n"
-      "ld1h { z8.h }, p2/Z, [x27, #1, MUL VL]\n"
-      "cmp x11, x10\n"
-      "add x9, x9, %x[out_stride]\n"
-      "ld1h { z25.h }, p2/Z, [x26]\n"
-      "ld1h { z26.h }, p2/Z, [x26, #1, MUL VL]\n"
-      "ld1h { z6.h }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z12.h }, p2/Z, [x12, #3, MUL VL]\n"
+      "ld1h { z21.h }, p2/Z, [x12]\n"
+      "ld1h { z17.h }, p2/Z, [x12, #1, MUL VL]\n"
+      "mov x21, x22\n"
+      "add x22, x22, %x[out_stride]\n"
+      "ld1h { z31.h }, p2/Z, [x11]\n"
+      "ld1h { z5.h }, p2/Z, [x11, #1, MUL VL]\n"
+      "mov x20, x22\n"
+      "sub x27, x27, x26\n"
+      "ld1h { z15.h }, p2/Z, [x10]\n"
+      "ld1h { z28.h }, p2/Z, [x10, #1, MUL VL]\n"
+      "zip1 z24.h, z21.h, z15.h\n"
+      "zip2 z29.h, z21.h, z15.h\n"
+      "ld1h { z6.h }, p2/Z, [x9]\n"
+      "ld1h { z4.h }, p2/Z, [x9, #1, MUL VL]\n"
+      "zip1 z16.h, z31.h, z6.h\n"
+      "zip2 z18.h, z31.h, z6.h\n"
+      "ld1h { z3.h }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z25.h }, p2/Z, [x12, #3, MUL VL]\n"
+      "zip1 z20.h, z17.h, z28.h\n"
+      "zip1 z7.h, z5.h, z4.h\n"
+      "ld1h { z27.h }, p2/Z, [x11, #2, MUL VL]\n"
+      "ld1h { z22.h }, p2/Z, [x11, #3, MUL VL]\n"
+      "zip2 z2.h, z17.h, z28.h\n"
+      "zip2 z19.h, z5.h, z4.h\n"
+      "ld1h { z28.h }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z17.h }, p2/Z, [x10, #3, MUL VL]\n"
+      "zip1 z21.h, z24.h, z16.h\n"
+      "zip2 z24.h, z24.h, z16.h\n"
+      "ld1h { z5.h }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1h { z1.h }, p2/Z, [x9, #3, MUL VL]\n"
+      "zip1 z14.h, z29.h, z18.h\n"
+      "zip2 z12.h, z29.h, z18.h\n"
+      "ld1h { z18.h }, p2/Z, [x28]\n"
+      "ld1h { z31.h }, p2/Z, [x28, #1, MUL VL]\n"
+      "zip1 z11.h, z20.h, z7.h\n"
+      "zip2 z13.h, z20.h, z7.h\n"
+      "ld1h { z4.h }, p2/Z, [x25]\n"
+      "ld1h { z26.h }, p2/Z, [x25, #1, MUL VL]\n"
+      "zip1 z15.h, z2.h, z19.h\n"
+      "zip2 z10.h, z2.h, z19.h\n"
+      "ld1h { z16.h }, p2/Z, [x24]\n"
+      "ld1h { z30.h }, p2/Z, [x24, #1, MUL VL]\n"
+      "zip1 z19.h, z18.h, z16.h\n"
+      "zip2 z18.h, z18.h, z16.h\n"
+      "ld1h { z8.h }, p2/Z, [x23]\n"
+      "ld1h { z29.h }, p2/Z, [x23, #1, MUL VL]\n"
+      "zip1 z20.h, z4.h, z8.h\n"
+      "zip2 z0.h, z4.h, z8.h\n"
+      "ld1h { z6.h }, p2/Z, [x28, #2, MUL VL]\n"
+      "ld1h { z8.h }, p2/Z, [x28, #3, MUL VL]\n"
+      "zip1 z23.h, z31.h, z30.h\n"
+      "zip1 z16.h, z26.h, z29.h\n"
+      "ld1h { z9.h }, p2/Z, [x25, #2, MUL VL]\n"
+      "ld1h { z7.h }, p2/Z, [x25, #3, MUL VL]\n"
+      "zip2 z31.h, z31.h, z30.h\n"
+      "zip2 z30.h, z26.h, z29.h\n"
+      "ld1h { z2.h }, p2/Z, [x24, #2, MUL VL]\n"
+      "ld1h { z26.h }, p2/Z, [x24, #3, MUL VL]\n"
+      "zip1 z29.h, z3.h, z28.h\n"
+      "zip1 z4.h, z27.h, z5.h\n"
+      "zip2 z28.h, z3.h, z28.h\n"
+      "ld1h { z3.h }, p2/Z, [x23, #2, MUL VL]\n"
+      "zip2 z27.h, z27.h, z5.h\n"
+      "ld1h { z5.h }, p2/Z, [x23, #3, MUL VL]\n"
+      "st1h { z21.h }, p2, [x21]\n"
+      "zip1 z21.h, z25.h, z17.h\n"
+      "zip2 z25.h, z25.h, z17.h\n"
+      "cmp x27, x26\n"
+      "st1h { z24.h }, p2, [x21, #1, MUL VL]\n"
+      "zip1 z24.h, z22.h, z1.h\n"
+      "zip2 z22.h, z22.h, z1.h\n"
       "addvl x12, x12, #4\n"
-      "ld1h { z15.h }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z11.h }, p2/Z, [x28, #3, MUL VL]\n"
-      "zip1 z18.h, z22.h, z3.h\n"
-      "zip2 z24.h, z22.h, z3.h\n"
-      "ld1h { z14.h }, p2/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z13.h }, p2/Z, [x27, #3, MUL VL]\n"
-      "zip1 z4.h, z23.h, z25.h\n"
-      "zip2 z16.h, z23.h, z25.h\n"
-      "ld1h { z7.h }, p2/Z, [x26, #2, MUL VL]\n"
-      "ld1h { z23.h }, p2/Z, [x26, #3, MUL VL]\n"
-      "zip1 z21.h, z2.h, z8.h\n"
-      "zip1 z22.h, z5.h, z26.h\n"
-      "ld1h { z1.h }, p2/Z, [x25]\n"
-      "ld1h { z25.h }, p2/Z, [x25, #1, MUL VL]\n"
-      "zip2 z10.h, z2.h, z8.h\n"
-      "zip2 z5.h, z5.h, z26.h\n"
-      "ld1h { z19.h }, p2/Z, [x24]\n"
-      "ld1h { z31.h }, p2/Z, [x24, #1, MUL VL]\n"
-      "zip1 z17.h, z18.h, z4.h\n"
-      "zip2 z28.h, z18.h, z4.h\n"
-      "ld1h { z3.h }, p2/Z, [x23]\n"
-      "ld1h { z18.h }, p2/Z, [x23, #1, MUL VL]\n"
-      "zip1 z0.h, z24.h, z16.h\n"
-      "zip2 z29.h, z24.h, z16.h\n"
-      "ld1h { z2.h }, p2/Z, [x22]\n"
-      "ld1h { z16.h }, p2/Z, [x22, #1, MUL VL]\n"
-      "zip1 z27.h, z21.h, z22.h\n"
-      "zip2 z26.h, z21.h, z22.h\n"
-      "ld1h { z8.h }, p2/Z, [x25, #2, MUL VL]\n"
-      "ld1h { z4.h }, p2/Z, [x25, #3, MUL VL]\n"
-      "zip1 z24.h, z10.h, z5.h\n"
-      "zip2 z20.h, z10.h, z5.h\n"
-      "ld1h { z30.h }, p2/Z, [x24, #2, MUL VL]\n"
-      "ld1h { z5.h }, p2/Z, [x24, #3, MUL VL]\n"
-      "zip1 z21.h, z1.h, z3.h\n"
-      "zip2 z9.h, z1.h, z3.h\n"
-      "ld1h { z10.h }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1h { z3.h }, p2/Z, [x23, #3, MUL VL]\n"
-      "zip1 z22.h, z19.h, z2.h\n"
-      "zip2 z19.h, z19.h, z2.h\n"
-      "ld1h { z2.h }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1h { z1.h }, p2/Z, [x22, #3, MUL VL]\n"
-      "st1h { z17.h }, p2, [x21]\n"
-      "zip1 z17.h, z25.h, z18.h\n"
-      "zip2 z25.h, z25.h, z18.h\n"
-      "zip1 z18.h, z31.h, z16.h\n"
-      "st1h { z28.h }, p2, [x21, #1, MUL VL]\n"
+      "st1h { z14.h }, p2, [x21, #2, MUL VL]\n"
+      "zip1 z17.h, z19.h, z20.h\n"
+      "zip2 z20.h, z19.h, z20.h\n"
+      "addvl x11, x11, #4\n"
+      "st1h { z12.h }, p2, [x21, #3, MUL VL]\n"
+      "zip1 z19.h, z18.h, z0.h\n"
+      "zip2 z18.h, z18.h, z0.h\n"
+      "addvl x10, x10, #4\n"
+      "st1h { z11.h }, p2, [x21, #4, MUL VL]\n"
+      "zip1 z14.h, z23.h, z16.h\n"
+      "zip2 z16.h, z23.h, z16.h\n"
+      "addvl x9, x9, #4\n"
+      "st1h { z13.h }, p2, [x21, #5, MUL VL]\n"
+      "zip1 z23.h, z31.h, z30.h\n"
+      "zip2 z1.h, z31.h, z30.h\n"
       "addvl x28, x28, #4\n"
-      "zip2 z16.h, z31.h, z16.h\n"
-      "st1h { z0.h }, p2, [x21, #2, MUL VL]\n"
-      "zip1 z0.h, z6.h, z14.h\n"
-      "addvl x27, x27, #4\n"
-      "st1h { z29.h }, p2, [x21, #3, MUL VL]\n"
-      "zip1 z31.h, z15.h, z7.h\n"
-      "zip2 z29.h, z6.h, z14.h\n"
-      "addvl x26, x26, #4\n"
-      "st1h { z27.h }, p2, [x21, #4, MUL VL]\n"
-      "zip2 z28.h, z15.h, z7.h\n"
-      "zip1 z27.h, z12.h, z13.h\n"
+      "st1h { z15.h }, p2, [x21, #6, MUL VL]\n"
+      "zip1 z0.h, z29.h, z4.h\n"
+      "zip2 z31.h, z29.h, z4.h\n"
       "addvl x25, x25, #4\n"
-      "st1h { z26.h }, p2, [x21, #5, MUL VL]\n"
-      "zip1 z26.h, z11.h, z23.h\n"
-      "zip2 z6.h, z12.h, z13.h\n"
+      "st1h { z10.h }, p2, [x21, #7, MUL VL]\n"
+      "addvl x21, x21, #16\n"
+      "zip1 z30.h, z28.h, z27.h\n"
+      "zip2 z29.h, z28.h, z27.h\n"
+      "st1h { z17.h }, p2, [x21, #-8, MUL VL]\n"
+      "zip1 z13.h, z21.h, z24.h\n"
+      "zip2 z27.h, z21.h, z24.h\n"
       "addvl x24, x24, #4\n"
-      "st1h { z24.h }, p2, [x21, #6, MUL VL]\n"
-      "zip2 z24.h, z11.h, z23.h\n"
-      "zip1 z23.h, z21.h, z22.h\n"
+      "st1h { z20.h }, p2, [x21, #-7, MUL VL]\n"
+      "zip1 z28.h, z25.h, z22.h\n"
+      "zip2 z25.h, z25.h, z22.h\n"
       "addvl x23, x23, #4\n"
-      "st1h { z20.h }, p2, [x21, #7, MUL VL]\n"
-      "addvl x21, x21, #16\n"
-      "zip2 z22.h, z21.h, z22.h\n"
-      "zip1 z21.h, z9.h, z19.h\n"
-      "zip2 z20.h, z9.h, z19.h\n"
-      "zip1 z19.h, z17.h, z18.h\n"
-      "addvl x22, x22, #4\n"
-      "zip2 z18.h, z17.h, z18.h\n"
-      "zip1 z17.h, z25.h, z16.h\n"
-      "zip2 z16.h, z25.h, z16.h\n"
-      "st1h { z23.h }, p2, [x21, #-8, MUL VL]\n"
-      "zip1 z23.h, z0.h, z31.h\n"
-      "st1h { z22.h }, p2, [x21, #-7, MUL VL]\n"
-      "zip2 z0.h, z0.h, z31.h\n"
-      "zip1 z31.h, z29.h, z28.h\n"
-      "st1h { z21.h }, p2, [x21, #-6, MUL VL]\n"
-      "zip2 z29.h, z29.h, z28.h\n"
-      "zip1 z28.h, z27.h, z26.h\n"
-      "st1h { z20.h }, p2, [x21, #-5, MUL VL]\n"
-      "zip2 z27.h, z27.h, z26.h\n"
-      "zip1 z26.h, z6.h, z24.h\n"
-      "st1h { z19.h }, p2, [x21, #-4, MUL VL]\n"
-      "zip2 z25.h, z6.h, z24.h\n"
-      "zip1 z22.h, z8.h, z10.h\n"
-      "st1h { z18.h }, p2, [x21, #-3, MUL VL]\n"
-      "zip1 z21.h, z30.h, z2.h\n"
-      "zip2 z20.h, z8.h, z10.h\n"
-      "st1h { z17.h }, p2, [x21, #-2, MUL VL]\n"
-      "zip2 z19.h, z30.h, z2.h\n"
-      "zip1 z18.h, z4.h, z3.h\n"
-      "st1h { z16.h }, p2, [x21, #-1, MUL VL]\n"
-      "zip1 z17.h, z5.h, z1.h\n"
-      "zip2 z24.h, z4.h, z3.h\n"
-      "zip2 z16.h, z5.h, z1.h\n"
-      "st1h { z23.h }, p2, [x20]\n"
+      "st1h { z19.h }, p2, [x21, #-6, MUL VL]\n"
+      "zip1 z22.h, z6.h, z2.h\n"
+      "zip1 z21.h, z9.h, z3.h\n"
+      "add x22, x22, %x[out_stride]\n"
+      "st1h { z18.h }, p2, [x21, #-5, MUL VL]\n"
+      "zip2 z20.h, z6.h, z2.h\n"
+      "zip2 z19.h, z9.h, z3.h\n"
+      "st1h { z14.h }, p2, [x21, #-4, MUL VL]\n"
+      "zip1 z18.h, z8.h, z26.h\n"
+      "zip1 z17.h, z7.h, z5.h\n"
+      "st1h { z16.h }, p2, [x21, #-3, MUL VL]\n"
+      "zip2 z24.h, z8.h, z26.h\n"
+      "zip2 z16.h, z7.h, z5.h\n"
+      "st1h { z23.h }, p2, [x21, #-2, MUL VL]\n"
       "zip1 z23.h, z22.h, z21.h\n"
-      "st1h { z0.h }, p2, [x20, #1, MUL VL]\n"
       "zip2 z22.h, z22.h, z21.h\n"
+      "st1h { z1.h }, p2, [x21, #-1, MUL VL]\n"
       "zip1 z21.h, z20.h, z19.h\n"
-      "st1h { z31.h }, p2, [x20, #2, MUL VL]\n"
       "zip2 z20.h, z20.h, z19.h\n"
+      "st1h { z0.h }, p2, [x20]\n"
       "zip1 z19.h, z18.h, z17.h\n"
-      "st1h { z29.h }, p2, [x20, #3, MUL VL]\n"
       "zip2 z18.h, z18.h, z17.h\n"
+      "st1h { z31.h }, p2, [x20, #1, MUL VL]\n"
       "zip1 z17.h, z24.h, z16.h\n"
-      "st1h { z28.h }, p2, [x20, #4, MUL VL]\n"
       "zip2 z16.h, z24.h, z16.h\n"
+      "st1h { z30.h }, p2, [x20, #2, MUL VL]\n"
+      "st1h { z29.h }, p2, [x20, #3, MUL VL]\n"
+      "st1h { z13.h }, p2, [x20, #4, MUL VL]\n"
       "st1h { z27.h }, p2, [x20, #5, MUL VL]\n"
-      "st1h { z26.h }, p2, [x20, #6, MUL VL]\n"
+      "st1h { z28.h }, p2, [x20, #6, MUL VL]\n"
       "st1h { z25.h }, p2, [x20, #7, MUL VL]\n"
       "addvl x20, x20, #16\n"
       "st1h { z23.h }, p2, [x20, #-8, MUL VL]\n"
@@ -205,84 +205,84 @@ void sve_transpose_interleave_8VL_2x4(uint16_t *out, const uint16_t *in, size_t
       "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x11, 5f\n"
+      "cbz x27, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x21, x11\n"
-      "mov x20, x9\n"
-      "decd x11, ALL, MUL #8\n"
-      "add x9, x9, %x[out_stride]\n"
-      "whilelt p1.h, XZR, x21\n"
-      "dech x21\n"
-      "whilelt p0.h, XZR, x21\n"
-      "cmp x11, #0x0\n"
-      "ld1h { z24.h }, p1/Z, [x12]\n"
-      "ld1h { z23.h }, p1/Z, [x28]\n"
-      "ld1h { z22.h }, p1/Z, [x27]\n"
-      "ld1h { z16.h }, p1/Z, [x26]\n"
-      "ld1h { z2.h }, p1/Z, [x25]\n"
-      "ld1h { z1.h }, p1/Z, [x24]\n"
-      "ld1h { z21.h }, p0/Z, [x12, #1, MUL VL]\n"
-      "ld1h { z26.h }, p0/Z, [x28, #1, MUL VL]\n"
+      "mov x20, x27\n"
+      "whilelt p1.h, XZR, x20\n"
+      "ld1h { z17.h }, p1/Z, [x12]\n"
+      "ld1h { z19.h }, p1/Z, [x11]\n"
+      "dech x20\n"
+      "whilelt p0.h, XZR, x20\n"
+      "ld1h { z24.h }, p0/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z23.h }, p0/Z, [x11, #1, MUL VL]\n"
+      "ld1h { z16.h }, p1/Z, [x10]\n"
+      "ld1h { z20.h }, p0/Z, [x10, #1, MUL VL]\n"
+      "zip1 z1.h, z17.h, z16.h\n"
+      "zip2 z22.h, z17.h, z16.h\n"
+      "ld1h { z18.h }, p1/Z, [x9]\n"
+      "ld1h { z17.h }, p0/Z, [x9, #1, MUL VL]\n"
+      "zip1 z16.h, z19.h, z18.h\n"
+      "zip2 z19.h, z19.h, z18.h\n"
+      "ld1h { z0.h }, p1/Z, [x28]\n"
+      "ld1h { z31.h }, p0/Z, [x28, #1, MUL VL]\n"
+      "zip1 z25.h, z24.h, z20.h\n"
+      "zip1 z21.h, z23.h, z17.h\n"
+      "ld1h { z30.h }, p1/Z, [x25]\n"
+      "ld1h { z29.h }, p0/Z, [x25, #1, MUL VL]\n"
+      "zip2 z28.h, z24.h, z20.h\n"
+      "zip2 z24.h, z23.h, z17.h\n"
+      "ld1h { z20.h }, p1/Z, [x24]\n"
+      "ld1h { z27.h }, p0/Z, [x24, #1, MUL VL]\n"
+      "mov x20, x22\n"
+      "decd x27, ALL, MUL #8\n"
+      "ld1h { z23.h }, p1/Z, [x23]\n"
+      "ld1h { z26.h }, p0/Z, [x23, #1, MUL VL]\n"
+      "zip1 z18.h, z1.h, z16.h\n"
+      "zip2 z17.h, z1.h, z16.h\n"
+      "zip1 z16.h, z22.h, z19.h\n"
+      "zip2 z19.h, z22.h, z19.h\n"
+      "st1h { z18.h }, p2, [x20]\n"
+      "cmp x27, #0x0\n"
+      "zip1 z22.h, z25.h, z21.h\n"
+      "zip2 z21.h, z25.h, z21.h\n"
+      "st1h { z17.h }, p2, [x20, #1, MUL VL]\n"
       "addvl x12, x12, #2\n"
+      "zip1 z25.h, z28.h, z24.h\n"
+      "zip2 z18.h, z28.h, z24.h\n"
+      "st1h { z16.h }, p2, [x20, #2, MUL VL]\n"
+      "addvl x11, x11, #2\n"
+      "zip1 z17.h, z0.h, z20.h\n"
+      "zip1 z16.h, z30.h, z23.h\n"
+      "st1h { z19.h }, p2, [x20, #3, MUL VL]\n"
+      "addvl x10, x10, #2\n"
+      "zip2 z20.h, z0.h, z20.h\n"
+      "zip2 z19.h, z30.h, z23.h\n"
+      "st1h { z22.h }, p2, [x20, #4, MUL VL]\n"
+      "addvl x9, x9, #2\n"
+      "zip1 z24.h, z31.h, z27.h\n"
+      "zip1 z23.h, z29.h, z26.h\n"
+      "st1h { z21.h }, p2, [x20, #5, MUL VL]\n"
       "addvl x28, x28, #2\n"
-      "ld1h { z20.h }, p0/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z19.h }, p0/Z, [x26, #1, MUL VL]\n"
-      "zip1 z18.h, z24.h, z22.h\n"
-      "zip1 z17.h, z23.h, z16.h\n"
-      "ld1h { z0.h }, p0/Z, [x25, #1, MUL VL]\n"
-      "ld1h { z31.h }, p0/Z, [x24, #1, MUL VL]\n"
-      "zip2 z25.h, z24.h, z22.h\n"
-      "zip2 z16.h, z23.h, z16.h\n"
-      "ld1h { z30.h }, p1/Z, [x23]\n"
-      "ld1h { z29.h }, p0/Z, [x23, #1, MUL VL]\n"
-      "addvl x27, x27, #2\n"
-      "addvl x26, x26, #2\n"
-      "ld1h { z24.h }, p1/Z, [x22]\n"
-      "ld1h { z28.h }, p0/Z, [x22, #1, MUL VL]\n"
-      "zip1 z23.h, z21.h, z20.h\n"
-      "zip1 z22.h, z26.h, z19.h\n"
-      "zip2 z21.h, z21.h, z20.h\n"
-      "zip2 z20.h, z26.h, z19.h\n"
+      "zip2 z22.h, z31.h, z27.h\n"
+      "zip2 z21.h, z29.h, z26.h\n"
+      "st1h { z25.h }, p2, [x20, #6, MUL VL]\n"
       "addvl x25, x25, #2\n"
+      "st1h { z18.h }, p2, [x20, #7, MUL VL]\n"
+      "addvl x20, x20, #16\n"
       "addvl x24, x24, #2\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
+      "zip1 z18.h, z17.h, z16.h\n"
       "addvl x23, x23, #2\n"
-      "addvl x22, x22, #2\n"
-      "zip1 z17.h, z25.h, z16.h\n"
-      "zip2 z16.h, z25.h, z16.h\n"
-      "zip1 z27.h, z23.h, z22.h\n"
-      "zip2 z23.h, z23.h, z22.h\n"
-      "zip1 z26.h, z21.h, z20.h\n"
-      "zip2 z25.h, z21.h, z20.h\n"
-      "st1h { z19.h }, p2, [x20]\n"
-      "zip1 z22.h, z2.h, z30.h\n"
-      "zip1 z21.h, z1.h, z24.h\n"
-      "st1h { z18.h }, p2, [x20, #1, MUL VL]\n"
-      "zip2 z20.h, z2.h, z30.h\n"
-      "zip2 z19.h, z1.h, z24.h\n"
-      "st1h { z17.h }, p2, [x20, #2, MUL VL]\n"
-      "zip1 z18.h, z0.h, z29.h\n"
-      "zip1 z17.h, z31.h, z28.h\n"
-      "st1h { z16.h }, p2, [x20, #3, MUL VL]\n"
-      "zip2 z24.h, z0.h, z29.h\n"
-      "zip2 z16.h, z31.h, z28.h\n"
-      "st1h { z27.h }, p2, [x20, #4, MUL VL]\n"
-      "st1h { z23.h }, p2, [x20, #5, MUL VL]\n"
-      "zip1 z23.h, z22.h, z21.h\n"
-      "zip2 z22.h, z22.h, z21.h\n"
-      "st1h { z26.h }, p2, [x20, #6, MUL VL]\n"
-      "zip1 z21.h, z20.h, z19.h\n"
+      "zip2 z17.h, z17.h, z16.h\n"
+      "zip1 z16.h, z20.h, z19.h\n"
+      "st1h { z18.h }, p2, [x20, #-8, MUL VL]\n"
       "zip2 z20.h, z20.h, z19.h\n"
-      "st1h { z25.h }, p2, [x20, #7, MUL VL]\n"
-      "addvl x20, x20, #16\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "zip1 z17.h, z24.h, z16.h\n"
-      "zip2 z16.h, z24.h, z16.h\n"
-      "st1h { z23.h }, p2, [x20, #-8, MUL VL]\n"
-      "st1h { z22.h }, p2, [x20, #-7, MUL VL]\n"
-      "st1h { z21.h }, p2, [x20, #-6, MUL VL]\n"
+      "zip1 z19.h, z24.h, z23.h\n"
+      "st1h { z17.h }, p2, [x20, #-7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip2 z18.h, z24.h, z23.h\n"
+      "zip1 z17.h, z22.h, z21.h\n"
+      "st1h { z16.h }, p2, [x20, #-6, MUL VL]\n"
+      "zip2 z16.h, z22.h, z21.h\n"
       "st1h { z20.h }, p2, [x20, #-5, MUL VL]\n"
       "st1h { z19.h }, p2, [x20, #-4, MUL VL]\n"
       "st1h { z18.h }, p2, [x20, #-3, MUL VL]\n"
@@ -297,141 +297,141 @@ void sve_transpose_interleave_8VL_2x4(uint16_t *out, const uint16_t *in, size_t
       "6:"  // Main loop skip
       "7:"  // Tail row loop: Head
       "mov x12, %x[in]\n"
+      "add x11, x12, %x[in_stride]\n"
+      "add x10, x11, %x[in_stride]\n"
       "mov x21, %x[width]\n"
       "cnth x20, ALL, MUL #4\n"
+      "add x9, x10, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x9, %x[out]\n"
-      "add x28, x12, %x[in_stride]\n"
-      "add x27, x28, %x[in_stride]\n"
-      "add x26, x27, %x[in_stride]\n"
-      "add %x[in], x26, %x[in_stride]\n"
-      "csel x26, x26, %x[pad_row], GT\n"
-      "csel x27, x27, %x[pad_row], GE\n"
+      "add %x[in], x9, %x[in_stride]\n"
+      "csel x9, x9, %x[pad_row], GT\n"
+      "csel x10, x10, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x28, x28, %x[pad_row], GT\n"
+      "csel x11, x11, %x[pad_row], GT\n"
       "cmp x21, x20\n"
+      "mov x22, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 9f\n"
       "8:"  // Tail row loop: Unroll column loop
-      "ld1h { z18.h }, p2/Z, [x12]\n"
-      "ld1h { z30.h }, p2/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z17.h }, p2/Z, [x12]\n"
+      "ld1h { z22.h }, p2/Z, [x12, #1, MUL VL]\n"
       "sub x21, x21, x20\n"
-      "ld1h { z24.h }, p2/Z, [x28]\n"
-      "ld1h { z29.h }, p2/Z, [x28, #1, MUL VL]\n"
       "cmp x21, x20\n"
-      "ld1h { z17.h }, p2/Z, [x27]\n"
-      "ld1h { z23.h }, p2/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x26]\n"
-      "ld1h { z22.h }, p2/Z, [x26, #1, MUL VL]\n"
-      "ld1h { z21.h }, p2/Z, [x12, #2, MUL VL]\n"
-      "ld1h { z4.h }, p2/Z, [x12, #3, MUL VL]\n"
+      "ld1h { z19.h }, p2/Z, [x11]\n"
+      "ld1h { z21.h }, p2/Z, [x11, #1, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x10]\n"
+      "ld1h { z18.h }, p2/Z, [x10, #1, MUL VL]\n"
+      "zip1 z4.h, z17.h, z16.h\n"
+      "zip2 z3.h, z17.h, z16.h\n"
+      "ld1h { z17.h }, p2/Z, [x9]\n"
+      "ld1h { z16.h }, p2/Z, [x9, #1, MUL VL]\n"
+      "zip1 z2.h, z19.h, z17.h\n"
+      "zip2 z1.h, z19.h, z17.h\n"
+      "ld1h { z17.h }, p2/Z, [x12, #2, MUL VL]\n"
+      "ld1h { z24.h }, p2/Z, [x12, #3, MUL VL]\n"
+      "zip1 z0.h, z22.h, z18.h\n"
+      "zip1 z31.h, z21.h, z16.h\n"
+      "ld1h { z20.h }, p2/Z, [x11, #2, MUL VL]\n"
+      "ld1h { z19.h }, p2/Z, [x11, #3, MUL VL]\n"
+      "zip2 z30.h, z22.h, z18.h\n"
+      "zip2 z23.h, z21.h, z16.h\n"
+      "ld1h { z16.h }, p2/Z, [x10, #2, MUL VL]\n"
+      "ld1h { z18.h }, p2/Z, [x10, #3, MUL VL]\n"
+      "zip1 z22.h, z17.h, z16.h\n"
+      "zip2 z29.h, z17.h, z16.h\n"
+      "ld1h { z17.h }, p2/Z, [x9, #2, MUL VL]\n"
+      "ld1h { z16.h }, p2/Z, [x9, #3, MUL VL]\n"
+      "zip1 z21.h, z20.h, z17.h\n"
+      "zip2 z28.h, z20.h, z17.h\n"
+      "zip1 z27.h, z24.h, z18.h\n"
+      "zip1 z26.h, z19.h, z16.h\n"
       "addvl x12, x12, #4\n"
-      "ld1h { z28.h }, p2/Z, [x28, #2, MUL VL]\n"
-      "ld1h { z27.h }, p2/Z, [x28, #3, MUL VL]\n"
-      "zip1 z3.h, z18.h, z17.h\n"
-      "zip2 z2.h, z18.h, z17.h\n"
-      "ld1h { z20.h }, p2/Z, [x27, #2, MUL VL]\n"
-      "ld1h { z19.h }, p2/Z, [x27, #3, MUL VL]\n"
-      "zip1 z18.h, z24.h, z16.h\n"
-      "zip2 z1.h, z24.h, z16.h\n"
-      "ld1h { z17.h }, p2/Z, [x26, #2, MUL VL]\n"
-      "ld1h { z16.h }, p2/Z, [x26, #3, MUL VL]\n"
-      "zip1 z26.h, z30.h, z23.h\n"
-      "zip1 z25.h, z29.h, z22.h\n"
-      "zip2 z24.h, z30.h, z23.h\n"
-      "zip2 z23.h, z29.h, z22.h\n"
-      "addvl x28, x28, #4\n"
-      "addvl x27, x27, #4\n"
-      "zip1 z22.h, z21.h, z20.h\n"
-      "zip2 z0.h, z21.h, z20.h\n"
-      "addvl x26, x26, #4\n"
-      "zip1 z21.h, z28.h, z17.h\n"
-      "zip2 z31.h, z28.h, z17.h\n"
-      "zip1 z30.h, z4.h, z19.h\n"
-      "zip1 z29.h, z27.h, z16.h\n"
-      "zip2 z28.h, z4.h, z19.h\n"
-      "zip2 z27.h, z27.h, z16.h\n"
-      "zip1 z20.h, z3.h, z18.h\n"
-      "zip2 z19.h, z3.h, z18.h\n"
-      "zip1 z18.h, z2.h, z1.h\n"
-      "zip2 z17.h, z2.h, z1.h\n"
-      "zip1 z16.h, z26.h, z25.h\n"
-      "zip2 z26.h, z26.h, z25.h\n"
-      "zip1 z25.h, z24.h, z23.h\n"
-      "zip2 z24.h, z24.h, z23.h\n"
-      "st1h { z20.h }, p2, [x9]\n"
-      "st1h { z19.h }, p2, [x9, #1, MUL VL]\n"
+      "addvl x11, x11, #4\n"
+      "zip2 z25.h, z24.h, z18.h\n"
+      "zip2 z24.h, z19.h, z16.h\n"
+      "addvl x10, x10, #4\n"
+      "addvl x9, x9, #4\n"
+      "zip1 z16.h, z4.h, z2.h\n"
+      "zip2 z17.h, z4.h, z2.h\n"
+      "st1h { z16.h }, p2, [x22]\n"
+      "zip1 z16.h, z3.h, z1.h\n"
+      "zip2 z20.h, z3.h, z1.h\n"
+      "st1h { z17.h }, p2, [x22, #1, MUL VL]\n"
+      "zip1 z19.h, z0.h, z31.h\n"
+      "zip2 z18.h, z0.h, z31.h\n"
+      "st1h { z16.h }, p2, [x22, #2, MUL VL]\n"
+      "zip1 z17.h, z30.h, z23.h\n"
+      "zip2 z16.h, z30.h, z23.h\n"
+      "st1h { z20.h }, p2, [x22, #3, MUL VL]\n"
+      "st1h { z19.h }, p2, [x22, #4, MUL VL]\n"
       "zip1 z23.h, z22.h, z21.h\n"
       "zip2 z22.h, z22.h, z21.h\n"
-      "st1h { z18.h }, p2, [x9, #2, MUL VL]\n"
-      "zip1 z21.h, z0.h, z31.h\n"
-      "zip2 z20.h, z0.h, z31.h\n"
-      "st1h { z17.h }, p2, [x9, #3, MUL VL]\n"
-      "zip1 z19.h, z30.h, z29.h\n"
-      "zip2 z18.h, z30.h, z29.h\n"
-      "st1h { z16.h }, p2, [x9, #4, MUL VL]\n"
-      "zip1 z17.h, z28.h, z27.h\n"
-      "zip2 z16.h, z28.h, z27.h\n"
-      "st1h { z26.h }, p2, [x9, #5, MUL VL]\n"
-      "st1h { z25.h }, p2, [x9, #6, MUL VL]\n"
-      "st1h { z24.h }, p2, [x9, #7, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
-      "st1h { z23.h }, p2, [x9]\n"
-      "st1h { z22.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z21.h }, p2, [x9, #2, MUL VL]\n"
-      "st1h { z20.h }, p2, [x9, #3, MUL VL]\n"
-      "st1h { z19.h }, p2, [x9, #4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x9, #5, MUL VL]\n"
-      "st1h { z17.h }, p2, [x9, #6, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #7, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
+      "st1h { z18.h }, p2, [x22, #5, MUL VL]\n"
+      "zip1 z21.h, z29.h, z28.h\n"
+      "zip2 z20.h, z29.h, z28.h\n"
+      "st1h { z17.h }, p2, [x22, #6, MUL VL]\n"
+      "zip1 z19.h, z27.h, z26.h\n"
+      "zip2 z18.h, z27.h, z26.h\n"
+      "st1h { z16.h }, p2, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
+      "zip1 z17.h, z25.h, z24.h\n"
+      "zip2 z16.h, z25.h, z24.h\n"
+      "st1h { z23.h }, p2, [x22]\n"
+      "st1h { z22.h }, p2, [x22, #1, MUL VL]\n"
+      "st1h { z21.h }, p2, [x22, #2, MUL VL]\n"
+      "st1h { z20.h }, p2, [x22, #3, MUL VL]\n"
+      "st1h { z19.h }, p2, [x22, #4, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22, #5, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #6, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bge 8b\n"
       "9:"  // Tail row loop: Unroll column loop skip
       "cbz x21, 11f\n"
       "10:"  // Tail row loop: Column loop
       "mov x20, x21\n"
-      "decd x21, ALL, MUL #8\n"
       "whilelt p1.h, XZR, x20\n"
+      "ld1h { z23.h }, p1/Z, [x12]\n"
+      "ld1h { z22.h }, p1/Z, [x11]\n"
       "dech x20\n"
       "whilelt p0.h, XZR, x20\n"
+      "ld1h { z21.h }, p0/Z, [x12, #1, MUL VL]\n"
+      "ld1h { z25.h }, p0/Z, [x11, #1, MUL VL]\n"
+      "ld1h { z19.h }, p1/Z, [x10]\n"
+      "ld1h { z20.h }, p0/Z, [x10, #1, MUL VL]\n"
+      "decd x21, ALL, MUL #8\n"
+      "zip1 z24.h, z23.h, z19.h\n"
+      "ld1h { z18.h }, p1/Z, [x9]\n"
+      "ld1h { z16.h }, p0/Z, [x9, #1, MUL VL]\n"
+      "zip1 z17.h, z22.h, z18.h\n"
+      "zip2 z23.h, z23.h, z19.h\n"
+      "zip2 z19.h, z22.h, z18.h\n"
+      "zip1 z22.h, z21.h, z20.h\n"
       "cmp x21, #0x0\n"
-      "ld1h { z20.h }, p1/Z, [x12]\n"
-      "ld1h { z19.h }, p1/Z, [x28]\n"
-      "ld1h { z18.h }, p1/Z, [x27]\n"
-      "ld1h { z17.h }, p1/Z, [x26]\n"
-      "ld1h { z24.h }, p0/Z, [x12, #1, MUL VL]\n"
       "addvl x12, x12, #2\n"
-      "ld1h { z25.h }, p0/Z, [x28, #1, MUL VL]\n"
-      "addvl x28, x28, #2\n"
-      "ld1h { z23.h }, p0/Z, [x27, #1, MUL VL]\n"
-      "ld1h { z16.h }, p0/Z, [x26, #1, MUL VL]\n"
-      "addvl x27, x27, #2\n"
-      "addvl x26, x26, #2\n"
-      "zip1 z22.h, z20.h, z18.h\n"
-      "zip1 z21.h, z19.h, z17.h\n"
-      "zip2 z20.h, z20.h, z18.h\n"
-      "zip2 z19.h, z19.h, z17.h\n"
-      "zip1 z18.h, z24.h, z23.h\n"
-      "zip1 z17.h, z25.h, z16.h\n"
-      "zip2 z24.h, z24.h, z23.h\n"
-      "zip2 z16.h, z25.h, z16.h\n"
-      "zip1 z23.h, z22.h, z21.h\n"
-      "zip2 z22.h, z22.h, z21.h\n"
-      "zip1 z21.h, z20.h, z19.h\n"
-      "zip2 z20.h, z20.h, z19.h\n"
-      "zip1 z19.h, z18.h, z17.h\n"
-      "zip2 z18.h, z18.h, z17.h\n"
-      "zip1 z17.h, z24.h, z16.h\n"
-      "zip2 z16.h, z24.h, z16.h\n"
-      "st1h { z23.h }, p2, [x9]\n"
-      "st1h { z22.h }, p2, [x9, #1, MUL VL]\n"
-      "st1h { z21.h }, p2, [x9, #2, MUL VL]\n"
-      "st1h { z20.h }, p2, [x9, #3, MUL VL]\n"
-      "st1h { z19.h }, p2, [x9, #4, MUL VL]\n"
-      "st1h { z18.h }, p2, [x9, #5, MUL VL]\n"
-      "st1h { z17.h }, p2, [x9, #6, MUL VL]\n"
-      "st1h { z16.h }, p2, [x9, #7, MUL VL]\n"
-      "add x9, x9, %x[out_stride]\n"
+      "zip1 z18.h, z25.h, z16.h\n"
+      "zip2 z21.h, z21.h, z20.h\n"
+      "addvl x11, x11, #2\n"
+      "addvl x10, x10, #2\n"
+      "zip2 z20.h, z25.h, z16.h\n"
+      "addvl x9, x9, #2\n"
+      "zip1 z16.h, z24.h, z17.h\n"
+      "st1h { z16.h }, p2, [x22]\n"
+      "zip2 z16.h, z24.h, z17.h\n"
+      "zip1 z17.h, z23.h, z19.h\n"
+      "st1h { z16.h }, p2, [x22, #1, MUL VL]\n"
+      "zip2 z16.h, z23.h, z19.h\n"
+      "zip1 z19.h, z22.h, z18.h\n"
+      "st1h { z17.h }, p2, [x22, #2, MUL VL]\n"
+      "zip2 z18.h, z22.h, z18.h\n"
+      "zip1 z17.h, z21.h, z20.h\n"
+      "st1h { z16.h }, p2, [x22, #3, MUL VL]\n"
+      "zip2 z16.h, z21.h, z20.h\n"
+      "st1h { z19.h }, p2, [x22, #4, MUL VL]\n"
+      "st1h { z18.h }, p2, [x22, #5, MUL VL]\n"
+      "st1h { z17.h }, p2, [x22, #6, MUL VL]\n"
+      "st1h { z16.h }, p2, [x22, #7, MUL VL]\n"
+      "add x22, x22, %x[out_stride]\n"
       "bgt 10b\n"
       "11:"  // Tail row loop: Column loop skip
       "cmp %x[height], #0x1\n"
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x4_fp32bf16.hpp
index 7160077342..ab3af6f88d 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x4_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_8VL_2x4_fp32bf16.hpp
@@ -39,220 +39,220 @@ void sve_transpose_interleave_8VL_2x4_fp32bf16(bfloat16 *out, const float *in, s
     size_t out_stride = 8 * roundup<size_t>(height, 4) * get_vector_length<uint32_t>();
 
     __asm__ __volatile__(
-      "ptrue p2.b\n"
+      "ptrue p4.b\n"
       "1:"  // Main row loop: Head
       "mov x26, %x[in]\n"
-      "mov x25, %x[width]\n"
+      "add x25, x26, %x[in_stride]\n"
+      "add x24, x25, %x[in_stride]\n"
+      "mov x23, %x[width]\n"
       "cnth x20, ALL, MUL #4\n"
+      "add x22, x24, %x[in_stride]\n"
       "cmp %x[height], #0x3\n"
-      "mov x24, %x[out]\n"
-      "add x23, x26, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x21, x22, %x[in_stride]\n"
-      "add %x[in], x21, %x[in_stride]\n"
-      "csel x21, x21, %x[pad_row], GT\n"
-      "csel x22, x22, %x[pad_row], GE\n"
+      "add %x[in], x22, %x[in_stride]\n"
+      "csel x22, x22, %x[pad_row], GT\n"
+      "csel x24, x24, %x[pad_row], GE\n"
       "cmp %x[height], #0x1\n"
-      "csel x23, x23, %x[pad_row], GT\n"
-      "cmp x25, x20\n"
+      "csel x25, x25, %x[pad_row], GT\n"
+      "cmp x23, x20\n"
+      "mov x21, %x[out]\n"
       "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Unroll column loop
-      "ld1w { z25.s }, p2/Z, [x26]\n"
-      "ld1w { z23.s }, p2/Z, [x26, #1, MUL VL]\n"
-      "sub x25, x25, x20\n"
-      "ld1w { z5.s }, p2/Z, [x26, #2, MUL VL]\n"
-      "ld1w { z31.s }, p2/Z, [x26, #3, MUL VL]\n"
-      "cmp x25, x20\n"
-      "ld1w { z16.s }, p2/Z, [x22]\n"
-      "ld1w { z18.s }, p2/Z, [x22, #1, MUL VL]\n"
-      "ld1w { z21.s }, p2/Z, [x22, #2, MUL VL]\n"
-      "ld1w { z14.s }, p2/Z, [x22, #3, MUL VL]\n"
-      "ld1w { z3.s }, p2/Z, [x26, #4, MUL VL]\n"
-      "ld1w { z30.s }, p2/Z, [x26, #5, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x26, #6, MUL VL]\n"
-      "ld1w { z22.s }, p2/Z, [x26, #7, MUL VL]\n"
-      "zip1 z19.s, z25.s, z16.s\n"
-      "zip2 z17.s, z25.s, z16.s\n"
-      "ld1w { z16.s }, p2/Z, [x22, #4, MUL VL]\n"
-      "ld1w { z13.s }, p2/Z, [x22, #5, MUL VL]\n"
-      "zip1 z11.s, z23.s, z18.s\n"
-      "zip2 z10.s, z23.s, z18.s\n"
-      "ld1w { z18.s }, p2/Z, [x22, #6, MUL VL]\n"
-      "ld1w { z15.s }, p2/Z, [x22, #7, MUL VL]\n"
-      "zip1 z9.s, z5.s, z21.s\n"
-      "zip2 z8.s, z5.s, z21.s\n"
-      "ld1w { z12.s }, p2/Z, [x23]\n"
-      "ld1w { z27.s }, p2/Z, [x23, #1, MUL VL]\n"
-      "zip1 z7.s, z31.s, z14.s\n"
-      "zip2 z6.s, z31.s, z14.s\n"
-      "ld1w { z25.s }, p2/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z23.s }, p2/Z, [x23, #3, MUL VL]\n"
-      "zip1 z5.s, z3.s, z16.s\n"
-      "zip2 z4.s, z3.s, z16.s\n"
-      "ld1w { z16.s }, p2/Z, [x21]\n"
-      "ld1w { z28.s }, p2/Z, [x21, #1, MUL VL]\n"
-      "zip1 z3.s, z30.s, z13.s\n"
-      "zip2 z2.s, z30.s, z13.s\n"
-      "ld1w { z26.s }, p2/Z, [x21, #2, MUL VL]\n"
-      "ld1w { z24.s }, p2/Z, [x21, #3, MUL VL]\n"
-      "zip1 z1.s, z20.s, z18.s\n"
-      "zip2 z0.s, z20.s, z18.s\n"
-      "ld1w { z21.s }, p2/Z, [x23, #4, MUL VL]\n"
-      "ld1w { z20.s }, p2/Z, [x23, #5, MUL VL]\n"
-      "zip1 z31.s, z22.s, z15.s\n"
-      "zip2 z30.s, z22.s, z15.s\n"
-      "ld1w { z14.s }, p2/Z, [x23, #6, MUL VL]\n"
-      "ld1w { z15.s }, p2/Z, [x23, #7, MUL VL]\n"
-      ".inst 0x658aaa76  // bfcvt z22.h, p2/M, z19.s\n"
-      "zip1 z29.s, z12.s, z16.s\n"
-      "ld1w { z19.s }, p2/Z, [x21, #4, MUL VL]\n"
-      "ld1w { z18.s }, p2/Z, [x21, #5, MUL VL]\n"
-      ".inst 0x658aaa2d  // bfcvt z13.h, p2/M, z17.s\n"
-      "zip2 z12.s, z12.s, z16.s\n"
-      "ld1w { z17.s }, p2/Z, [x21, #6, MUL VL]\n"
-      "ld1w { z16.s }, p2/Z, [x21, #7, MUL VL]\n"
-      ".inst 0x658aa96b  // bfcvt z11.h, p2/M, z11.s\n"
-      ".inst 0x658aa94a  // bfcvt z10.h, p2/M, z10.s\n"
-      ".inst 0x658aa929  // bfcvt z9.h, p2/M, z9.s\n"
-      ".inst 0x658aa908  // bfcvt z8.h, p2/M, z8.s\n"
+      "ld1w { z19.s }, p4/Z, [x26]\n"
+      "ld1w { z18.s }, p4/Z, [x26, #1, MUL VL]\n"
+      "sub x23, x23, x20\n"
+      "cmp x23, x20\n"
+      "ld1w { z20.s }, p4/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z24.s }, p4/Z, [x26, #3, MUL VL]\n"
+      "ld1w { z23.s }, p4/Z, [x24]\n"
+      "ld1w { z17.s }, p4/Z, [x24, #1, MUL VL]\n"
+      "zip1 z22.s, z19.s, z23.s\n"
+      "zip2 z21.s, z19.s, z23.s\n"
+      "ld1w { z31.s }, p4/Z, [x24, #2, MUL VL]\n"
+      "ld1w { z16.s }, p4/Z, [x24, #3, MUL VL]\n"
+      "zip1 z9.s, z18.s, z17.s\n"
+      "zip2 z7.s, z18.s, z17.s\n"
+      "ld1w { z19.s }, p4/Z, [x26, #4, MUL VL]\n"
+      "ld1w { z18.s }, p4/Z, [x26, #5, MUL VL]\n"
+      "zip1 z6.s, z20.s, z31.s\n"
+      "zip2 z5.s, z20.s, z31.s\n"
+      "ld1w { z15.s }, p4/Z, [x26, #6, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x26, #7, MUL VL]\n"
+      "zip1 z3.s, z24.s, z16.s\n"
+      "zip2 z2.s, z24.s, z16.s\n"
+      "ld1w { z16.s }, p4/Z, [x24, #4, MUL VL]\n"
+      "ld1w { z17.s }, p4/Z, [x24, #5, MUL VL]\n"
+      "zip1 z1.s, z19.s, z16.s\n"
+      "zip2 z0.s, z19.s, z16.s\n"
+      "ld1w { z16.s }, p4/Z, [x24, #6, MUL VL]\n"
+      "ld1w { z19.s }, p4/Z, [x24, #7, MUL VL]\n"
+      "zip1 z31.s, z18.s, z17.s\n"
+      "zip2 z30.s, z18.s, z17.s\n"
+      "ld1w { z18.s }, p4/Z, [x25]\n"
+      "ld1w { z17.s }, p4/Z, [x25, #1, MUL VL]\n"
+      "zip1 z29.s, z15.s, z16.s\n"
+      "zip2 z28.s, z15.s, z16.s\n"
+      "ld1w { z16.s }, p4/Z, [x25, #2, MUL VL]\n"
+      "ld1w { z23.s }, p4/Z, [x25, #3, MUL VL]\n"
+      "zip1 z27.s, z20.s, z19.s\n"
+      "zip2 z26.s, z20.s, z19.s\n"
+      "ld1w { z11.s }, p4/Z, [x22]\n"
+      "ld1w { z8.s }, p4/Z, [x22, #1, MUL VL]\n"
+      ".inst 0x658ab2d8  // bfcvt z24.h, p4/M, z22.s\n"
+      "zip1 z25.s, z18.s, z11.s\n"
+      "ld1w { z4.s }, p4/Z, [x22, #2, MUL VL]\n"
+      "ld1w { z22.s }, p4/Z, [x22, #3, MUL VL]\n"
+      ".inst 0x658ab2af  // bfcvt z15.h, p4/M, z21.s\n"
+      "zip2 z14.s, z18.s, z11.s\n"
+      "ld1w { z21.s }, p4/Z, [x25, #4, MUL VL]\n"
+      "ld1w { z20.s }, p4/Z, [x25, #5, MUL VL]\n"
+      ".inst 0x658ab12d  // bfcvt z13.h, p4/M, z9.s\n"
+      "zip1 z12.s, z17.s, z8.s\n"
+      "ld1w { z11.s }, p4/Z, [x25, #6, MUL VL]\n"
+      "ld1w { z10.s }, p4/Z, [x25, #7, MUL VL]\n"
+      ".inst 0x658ab0e9  // bfcvt z9.h, p4/M, z7.s\n"
+      "zip2 z8.s, z17.s, z8.s\n"
+      "ld1w { z19.s }, p4/Z, [x22, #4, MUL VL]\n"
+      "ld1w { z18.s }, p4/Z, [x22, #5, MUL VL]\n"
+      ".inst 0x658ab0c7  // bfcvt z7.h, p4/M, z6.s\n"
+      "zip1 z6.s, z16.s, z4.s\n"
+      "ld1w { z17.s }, p4/Z, [x22, #6, MUL VL]\n"
+      ".inst 0x658ab0a5  // bfcvt z5.h, p4/M, z5.s\n"
+      "zip2 z4.s, z16.s, z4.s\n"
+      "ld1w { z16.s }, p4/Z, [x22, #7, MUL VL]\n"
+      ".inst 0x658ab063  // bfcvt z3.h, p4/M, z3.s\n"
+      ".inst 0x658ab042  // bfcvt z2.h, p4/M, z2.s\n"
       "addvl x26, x26, #8\n"
-      "addvl x23, x23, #8\n"
-      ".inst 0x658aa8e7  // bfcvt z7.h, p2/M, z7.s\n"
-      ".inst 0x658aa8c6  // bfcvt z6.h, p2/M, z6.s\n"
+      "addvl x25, x25, #8\n"
+      ".inst 0x658ab021  // bfcvt z1.h, p4/M, z1.s\n"
+      ".inst 0x658ab000  // bfcvt z0.h, p4/M, z0.s\n"
+      "addvl x24, x24, #8\n"
       "addvl x22, x22, #8\n"
-      "addvl x21, x21, #8\n"
-      ".inst 0x658aa8a5  // bfcvt z5.h, p2/M, z5.s\n"
-      ".inst 0x658aa884  // bfcvt z4.h, p2/M, z4.s\n"
-      ".inst 0x658aa863  // bfcvt z3.h, p2/M, z3.s\n"
-      ".inst 0x658aa842  // bfcvt z2.h, p2/M, z2.s\n"
-      ".inst 0x658aa821  // bfcvt z1.h, p2/M, z1.s\n"
-      ".inst 0x658aa800  // bfcvt z0.h, p2/M, z0.s\n"
-      ".inst 0x658aabff  // bfcvt z31.h, p2/M, z31.s\n"
-      ".inst 0x658aabde  // bfcvt z30.h, p2/M, z30.s\n"
-      ".inst 0x648aabb6  // bfcvtnt z22.h, p2/M, z29.s\n"
-      "zip1 z29.s, z27.s, z28.s\n"
-      "zip2 z28.s, z27.s, z28.s\n"
-      "zip1 z27.s, z25.s, z26.s\n"
-      "zip2 z26.s, z25.s, z26.s\n"
-      "zip1 z25.s, z23.s, z24.s\n"
-      "zip2 z24.s, z23.s, z24.s\n"
+      ".inst 0x658ab3ff  // bfcvt z31.h, p4/M, z31.s\n"
+      ".inst 0x658ab3de  // bfcvt z30.h, p4/M, z30.s\n"
+      ".inst 0x658ab3bd  // bfcvt z29.h, p4/M, z29.s\n"
+      ".inst 0x658ab39c  // bfcvt z28.h, p4/M, z28.s\n"
+      ".inst 0x658ab37b  // bfcvt z27.h, p4/M, z27.s\n"
+      ".inst 0x658ab35a  // bfcvt z26.h, p4/M, z26.s\n"
+      ".inst 0x648ab338  // bfcvtnt z24.h, p4/M, z25.s\n"
+      "zip1 z25.s, z23.s, z22.s\n"
+      "st1h { z24.h }, p4, [x21]\n"
+      "zip2 z24.s, z23.s, z22.s\n"
       "zip1 z23.s, z21.s, z19.s\n"
-      "st1h { z22.h }, p2, [x24]\n"
       "zip2 z22.s, z21.s, z19.s\n"
       "zip1 z21.s, z20.s, z18.s\n"
       "zip2 z20.s, z20.s, z18.s\n"
-      "zip1 z19.s, z14.s, z17.s\n"
-      "zip2 z18.s, z14.s, z17.s\n"
-      "zip1 z17.s, z15.s, z16.s\n"
-      "zip2 z16.s, z15.s, z16.s\n"
-      ".inst 0x648aa98d  // bfcvtnt z13.h, p2/M, z12.s\n"
-      ".inst 0x648aabab  // bfcvtnt z11.h, p2/M, z29.s\n"
-      ".inst 0x648aab8a  // bfcvtnt z10.h, p2/M, z28.s\n"
-      ".inst 0x648aab69  // bfcvtnt z9.h, p2/M, z27.s\n"
-      ".inst 0x648aab48  // bfcvtnt z8.h, p2/M, z26.s\n"
-      ".inst 0x648aab27  // bfcvtnt z7.h, p2/M, z25.s\n"
-      ".inst 0x648aab06  // bfcvtnt z6.h, p2/M, z24.s\n"
-      "st1h { z13.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z11.h }, p2, [x24, #2, MUL VL]\n"
-      ".inst 0x648aaae5  // bfcvtnt z5.h, p2/M, z23.s\n"
-      ".inst 0x648aaac4  // bfcvtnt z4.h, p2/M, z22.s\n"
-      "st1h { z10.h }, p2, [x24, #3, MUL VL]\n"
-      ".inst 0x648aaaa3  // bfcvtnt z3.h, p2/M, z21.s\n"
-      ".inst 0x648aaa82  // bfcvtnt z2.h, p2/M, z20.s\n"
-      "st1h { z9.h }, p2, [x24, #4, MUL VL]\n"
-      ".inst 0x648aaa61  // bfcvtnt z1.h, p2/M, z19.s\n"
-      ".inst 0x648aaa40  // bfcvtnt z0.h, p2/M, z18.s\n"
-      "st1h { z8.h }, p2, [x24, #5, MUL VL]\n"
-      ".inst 0x648aaa3f  // bfcvtnt z31.h, p2/M, z17.s\n"
-      ".inst 0x648aaa1e  // bfcvtnt z30.h, p2/M, z16.s\n"
-      "st1h { z7.h }, p2, [x24, #6, MUL VL]\n"
-      "st1h { z6.h }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
-      "st1h { z5.h }, p2, [x24]\n"
-      "st1h { z4.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z3.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z2.h }, p2, [x24, #3, MUL VL]\n"
-      "st1h { z1.h }, p2, [x24, #4, MUL VL]\n"
-      "st1h { z0.h }, p2, [x24, #5, MUL VL]\n"
-      "st1h { z31.h }, p2, [x24, #6, MUL VL]\n"
-      "st1h { z30.h }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
+      "zip1 z19.s, z11.s, z17.s\n"
+      "zip2 z18.s, z11.s, z17.s\n"
+      "zip1 z17.s, z10.s, z16.s\n"
+      "zip2 z16.s, z10.s, z16.s\n"
+      ".inst 0x648ab1cf  // bfcvtnt z15.h, p4/M, z14.s\n"
+      "st1h { z15.h }, p4, [x21, #1, MUL VL]\n"
+      ".inst 0x648ab18d  // bfcvtnt z13.h, p4/M, z12.s\n"
+      ".inst 0x648ab109  // bfcvtnt z9.h, p4/M, z8.s\n"
+      "st1h { z13.h }, p4, [x21, #2, MUL VL]\n"
+      ".inst 0x648ab0c7  // bfcvtnt z7.h, p4/M, z6.s\n"
+      ".inst 0x648ab085  // bfcvtnt z5.h, p4/M, z4.s\n"
+      "st1h { z9.h }, p4, [x21, #3, MUL VL]\n"
+      ".inst 0x648ab323  // bfcvtnt z3.h, p4/M, z25.s\n"
+      ".inst 0x648ab302  // bfcvtnt z2.h, p4/M, z24.s\n"
+      "st1h { z7.h }, p4, [x21, #4, MUL VL]\n"
+      "st1h { z5.h }, p4, [x21, #5, MUL VL]\n"
+      ".inst 0x648ab2e1  // bfcvtnt z1.h, p4/M, z23.s\n"
+      ".inst 0x648ab2c0  // bfcvtnt z0.h, p4/M, z22.s\n"
+      "st1h { z3.h }, p4, [x21, #6, MUL VL]\n"
+      ".inst 0x648ab2bf  // bfcvtnt z31.h, p4/M, z21.s\n"
+      ".inst 0x648ab29e  // bfcvtnt z30.h, p4/M, z20.s\n"
+      "st1h { z2.h }, p4, [x21, #7, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
+      ".inst 0x648ab27d  // bfcvtnt z29.h, p4/M, z19.s\n"
+      ".inst 0x648ab25c  // bfcvtnt z28.h, p4/M, z18.s\n"
+      ".inst 0x648ab23b  // bfcvtnt z27.h, p4/M, z17.s\n"
+      ".inst 0x648ab21a  // bfcvtnt z26.h, p4/M, z16.s\n"
+      "st1h { z1.h }, p4, [x21]\n"
+      "st1h { z0.h }, p4, [x21, #1, MUL VL]\n"
+      "st1h { z31.h }, p4, [x21, #2, MUL VL]\n"
+      "st1h { z30.h }, p4, [x21, #3, MUL VL]\n"
+      "st1h { z29.h }, p4, [x21, #4, MUL VL]\n"
+      "st1h { z28.h }, p4, [x21, #5, MUL VL]\n"
+      "st1h { z27.h }, p4, [x21, #6, MUL VL]\n"
+      "st1h { z26.h }, p4, [x21, #7, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Unroll column loop skip
-      "cbz x25, 5f\n"
+      "cbz x23, 5f\n"
       "4:"  // Main row loop: Column loop
-      "mov x20, x25\n"
-      "decd x25, ALL, MUL #8\n"
-      "whilelt p1.s, XZR, x20\n"
+      "mov x20, x23\n"
+      "whilelt p3.s, XZR, x20\n"
+      "ld1w { z22.s }, p3/Z, [x26]\n"
+      "ld1w { z21.s }, p3/Z, [x24]\n"
       "decw x20\n"
-      "whilelt p0.s, XZR, x20\n"
+      "whilelt p2.s, XZR, x20\n"
+      "ld1w { z20.s }, p2/Z, [x26, #1, MUL VL]\n"
+      "ld1w { z19.s }, p2/Z, [x24, #1, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z18.s }, p1/Z, [x26]\n"
-      "ld1w { z16.s }, p1/Z, [x22]\n"
-      "ld1w { z26.s }, p1/Z, [x23]\n"
-      "ld1w { z25.s }, p1/Z, [x21]\n"
       "whilelt p1.s, XZR, x20\n"
+      "ld1w { z18.s }, p1/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z17.s }, p1/Z, [x24, #2, MUL VL]\n"
       "decw x20\n"
-      "ld1w { z20.s }, p0/Z, [x26, #1, MUL VL]\n"
-      "ld1w { z17.s }, p0/Z, [x22, #1, MUL VL]\n"
-      "ld1w { z24.s }, p0/Z, [x23, #1, MUL VL]\n"
-      "ld1w { z23.s }, p0/Z, [x21, #1, MUL VL]\n"
-      "zip1 z19.s, z18.s, z16.s\n"
-      "zip2 z16.s, z18.s, z16.s\n"
       "whilelt p0.s, XZR, x20\n"
-      "ld1w { z22.s }, p1/Z, [x26, #2, MUL VL]\n"
+      "ld1w { z28.s }, p0/Z, [x26, #3, MUL VL]\n"
+      "ld1w { z16.s }, p0/Z, [x24, #3, MUL VL]\n"
+      "ld1w { z27.s }, p3/Z, [x25]\n"
+      "ld1w { z3.s }, p2/Z, [x25, #1, MUL VL]\n"
+      "zip1 z26.s, z22.s, z21.s\n"
+      "zip2 z25.s, z22.s, z21.s\n"
+      "ld1w { z2.s }, p1/Z, [x25, #2, MUL VL]\n"
+      "ld1w { z1.s }, p0/Z, [x25, #3, MUL VL]\n"
+      "zip1 z24.s, z20.s, z19.s\n"
+      "zip2 z23.s, z20.s, z19.s\n"
+      "ld1w { z22.s }, p3/Z, [x22]\n"
+      "ld1w { z21.s }, p2/Z, [x22, #1, MUL VL]\n"
+      "zip1 z20.s, z18.s, z17.s\n"
+      "zip2 z19.s, z18.s, z17.s\n"
       "ld1w { z18.s }, p1/Z, [x22, #2, MUL VL]\n"
-      "zip1 z3.s, z26.s, z25.s\n"
-      "ld1w { z2.s }, p1/Z, [x23, #2, MUL VL]\n"
-      "ld1w { z1.s }, p1/Z, [x21, #2, MUL VL]\n"
-      "zip1 z21.s, z20.s, z17.s\n"
-      "zip2 z17.s, z20.s, z17.s\n"
-      ".inst 0x658aaa60  // bfcvt z0.h, p2/M, z19.s\n"
-      ".inst 0x658aaa1f  // bfcvt z31.h, p2/M, z16.s\n"
-      "cmp x25, #0x0\n"
-      "ld1w { z20.s }, p0/Z, [x26, #3, MUL VL]\n"
-      "ld1w { z16.s }, p0/Z, [x22, #3, MUL VL]\n"
-      "ld1w { z30.s }, p0/Z, [x23, #3, MUL VL]\n"
-      "zip1 z19.s, z22.s, z18.s\n"
-      "zip2 z18.s, z22.s, z18.s\n"
-      "ld1w { z29.s }, p0/Z, [x21, #3, MUL VL]\n"
-      "zip2 z28.s, z26.s, z25.s\n"
-      ".inst 0x658aaabb  // bfcvt z27.h, p2/M, z21.s\n"
+      "ld1w { z0.s }, p0/Z, [x22, #3, MUL VL]\n"
+      "zip1 z17.s, z28.s, z16.s\n"
+      "zip2 z16.s, z28.s, z16.s\n"
+      "decd x23, ALL, MUL #8\n"
+      ".inst 0x658ab35f  // bfcvt z31.h, p4/M, z26.s\n"
+      "zip1 z30.s, z27.s, z22.s\n"
+      "cmp x23, #0x0\n"
+      ".inst 0x658ab33d  // bfcvt z29.h, p4/M, z25.s\n"
+      "zip2 z28.s, z27.s, z22.s\n"
       "addvl x26, x26, #4\n"
-      "zip1 z26.s, z24.s, z23.s\n"
-      ".inst 0x658aaa39  // bfcvt z25.h, p2/M, z17.s\n"
-      "addvl x23, x23, #4\n"
+      "addvl x25, x25, #4\n"
+      ".inst 0x658ab31b  // bfcvt z27.h, p4/M, z24.s\n"
+      "zip1 z26.s, z3.s, z21.s\n"
+      "addvl x24, x24, #4\n"
       "addvl x22, x22, #4\n"
-      "zip1 z17.s, z20.s, z16.s\n"
-      "zip2 z16.s, z20.s, z16.s\n"
-      "addvl x21, x21, #4\n"
-      "zip2 z24.s, z24.s, z23.s\n"
-      ".inst 0x658aaa77  // bfcvt z23.h, p2/M, z19.s\n"
-      "zip1 z22.s, z2.s, z1.s\n"
-      ".inst 0x658aaa55  // bfcvt z21.h, p2/M, z18.s\n"
-      "zip2 z20.s, z2.s, z1.s\n"
-      ".inst 0x658aaa33  // bfcvt z19.h, p2/M, z17.s\n"
-      "zip1 z18.s, z30.s, z29.s\n"
-      ".inst 0x658aaa11  // bfcvt z17.h, p2/M, z16.s\n"
-      "zip2 z16.s, z30.s, z29.s\n"
-      ".inst 0x648aa860  // bfcvtnt z0.h, p2/M, z3.s\n"
-      ".inst 0x648aab9f  // bfcvtnt z31.h, p2/M, z28.s\n"
-      ".inst 0x648aab5b  // bfcvtnt z27.h, p2/M, z26.s\n"
-      ".inst 0x648aab19  // bfcvtnt z25.h, p2/M, z24.s\n"
-      ".inst 0x648aaad7  // bfcvtnt z23.h, p2/M, z22.s\n"
-      ".inst 0x648aaa95  // bfcvtnt z21.h, p2/M, z20.s\n"
-      ".inst 0x648aaa53  // bfcvtnt z19.h, p2/M, z18.s\n"
-      "st1h { z0.h }, p2, [x24]\n"
-      ".inst 0x648aaa11  // bfcvtnt z17.h, p2/M, z16.s\n"
-      "st1h { z31.h }, p2, [x24, #1, MUL VL]\n"
-      "st1h { z27.h }, p2, [x24, #2, MUL VL]\n"
-      "st1h { z25.h }, p2, [x24, #3, MUL VL]\n"
-      "st1h { z23.h }, p2, [x24, #4, MUL VL]\n"
-      "st1h { z21.h }, p2, [x24, #5, MUL VL]\n"
-      "st1h { z19.h }, p2, [x24, #6, MUL VL]\n"
-      "st1h { z17.h }, p2, [x24, #7, MUL VL]\n"
-      "add x24, x24, %x[out_stride]\n"
+      ".inst 0x658ab2f9  // bfcvt z25.h, p4/M, z23.s\n"
+      "zip2 z24.s, z3.s, z21.s\n"
+      ".inst 0x658ab297  // bfcvt z23.h, p4/M, z20.s\n"
+      "zip1 z22.s, z2.s, z18.s\n"
+      ".inst 0x658ab275  // bfcvt z21.h, p4/M, z19.s\n"
+      "zip2 z20.s, z2.s, z18.s\n"
+      ".inst 0x658ab233  // bfcvt z19.h, p4/M, z17.s\n"
+      "zip1 z18.s, z1.s, z0.s\n"
+      ".inst 0x658ab211  // bfcvt z17.h, p4/M, z16.s\n"
+      "zip2 z16.s, z1.s, z0.s\n"
+      ".inst 0x648ab3df  // bfcvtnt z31.h, p4/M, z30.s\n"
+      ".inst 0x648ab39d  // bfcvtnt z29.h, p4/M, z28.s\n"
+      "st1h { z31.h }, p4, [x21]\n"
+      ".inst 0x648ab35b  // bfcvtnt z27.h, p4/M, z26.s\n"
+      ".inst 0x648ab319  // bfcvtnt z25.h, p4/M, z24.s\n"
+      "st1h { z29.h }, p4, [x21, #1, MUL VL]\n"
+      ".inst 0x648ab2d7  // bfcvtnt z23.h, p4/M, z22.s\n"
+      ".inst 0x648ab295  // bfcvtnt z21.h, p4/M, z20.s\n"
+      "st1h { z27.h }, p4, [x21, #2, MUL VL]\n"
+      ".inst 0x648ab253  // bfcvtnt z19.h, p4/M, z18.s\n"
+      ".inst 0x648ab211  // bfcvtnt z17.h, p4/M, z16.s\n"
+      "st1h { z25.h }, p4, [x21, #3, MUL VL]\n"
+      "st1h { z23.h }, p4, [x21, #4, MUL VL]\n"
+      "st1h { z21.h }, p4, [x21, #5, MUL VL]\n"
+      "st1h { z19.h }, p4, [x21, #6, MUL VL]\n"
+      "st1h { z17.h }, p4, [x21, #7, MUL VL]\n"
+      "add x21, x21, %x[out_stride]\n"
       "bgt 4b\n"
       "5:"  // Main row loop: Column loop skip
       "cmp %x[height], #0x1\n"
@@ -260,7 +260,7 @@ void sve_transpose_interleave_8VL_2x4_fp32bf16(bfloat16 *out, const float *in, s
       "bge 1b\n"
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
-      : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+      : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
     );
 }
 
diff --git a/src/core/NEON/kernels/arm_gemm/utils.hpp b/src/core/NEON/kernels/arm_gemm/utils.hpp
index d0a8635604..9d8e31870d 100644
--- a/src/core/NEON/kernels/arm_gemm/utils.hpp
+++ b/src/core/NEON/kernels/arm_gemm/utils.hpp
@@ -176,6 +176,7 @@ namespace utils {
 // which then calls SVE kernels (compiled accordingly) iff SVE is detected at runtime.
 template <typename T>
 inline unsigned long get_vector_length() {
+#if defined(__aarch64__)
     uint64_t vl;
 
     __asm __volatile (
@@ -187,26 +188,24 @@ inline unsigned long get_vector_length() {
     );
 
     return vl / sizeof(T);
+#else // !defined(__aarch64__)
+    return 16 / sizeof(T);
+#endif // defined(__aarch64__)
 }
 
+#ifdef ARM_COMPUTE_ENABLE_SME
 namespace sme {
 
-template <typename T>
-inline uint64_t get_vector_length() {
-    uint64_t raw_vector_length;
-
-    __asm __volatile (
-        ".inst 0x04bf5821\n" // RDSVL X1, #1
-        "mov %0, X1\n"
-        : "=r" (raw_vector_length)
-        :
-        : "x1"
-    );
+// function from misc-sve.cpp
+extern unsigned int raw_vector_length();
 
-    return raw_vector_length / sizeof(T);
+template <typename T>
+inline unsigned long get_vector_length() {
+    return raw_vector_length() / sizeof(T);
 }
 
 } // namespace sme
+#endif // ARM_COMPUTE_ENABLE_SME
 
 // get_vector_length(VLType): Returns vector length for type "T".
 //
@@ -215,17 +214,48 @@ inline uint64_t get_vector_length() {
 template <typename T>
 inline unsigned long get_vector_length(VLType vl_type) {
   switch (vl_type) {
-#ifdef ARM_COMPUTE_ENABLE_SVE
+#ifdef ARM_COMPUTE_ENABLE_SME
     case VLType::SME:
       return sme::get_vector_length<T>();
+#endif // ARM_COMPUTE_ENABLE_SME
     case VLType::SVE:
       return get_vector_length<T>();
-#endif
     default:
       return 16 / sizeof(T);
   }
 }
 
+// get_default_activation_values(): Returns the default values for activation min and max for integer activation.
+template <typename T>
+inline std::tuple<T, T> get_default_activation_values()
+{
+    const T min = static_cast<T>(std::numeric_limits<T>::min());
+    const T max = static_cast<T>(std::numeric_limits<T>::max());
+
+    return std::make_tuple(min, max);
+}
+
+// get_default_activation_values(): Returns the default values for activation min and max for float activation.
+template <>
+inline std::tuple<float, float> get_default_activation_values()
+{
+    const float min = static_cast<float>(-std::numeric_limits<float>::infinity());
+    const float max = static_cast<float>(std::numeric_limits<float>::infinity());
+
+    return std::make_tuple(min, max);
+}
+
+#if defined(__ARM_FP16_ARGS)
+// get_default_activation_values(): Returns the default values for activation min and max for __fp16 activation.
+template <>
+inline std::tuple<__fp16, __fp16> get_default_activation_values()
+{
+    const __fp16 min = static_cast<__fp16>(-std::numeric_limits<float>::infinity());
+    const __fp16 max = static_cast<__fp16>(std::numeric_limits<float>::infinity());
+
+    return std::make_tuple(min, max);
+}
+#endif  // defined(__ARM_FP16_ARGS)
 } // utils namespace
 } // arm_gemm namespace